Advances in Neural Networks - ISNN 2004

In this paper, a method for trajectory generation in running is proposed with Radial Basis Function Neural Network, which can generate a series of joint trajectories to adjust humanoid robot step length and step time based on the sensor information. Compared with GA, RBFNN use less time to generate new trajectory to deal with sudden obstacles after thorough training. The performance of the proposed method is validated by simulation of a 28 DOF humanoid robot model with ADAMS.

This paper studies coordination control for an uncalibrated single eye-in-hand robotic system to track an object in 3-D space with simultaneous translations and rotations. A set of object features is properly defined to derive an invertible nonlinear visual mapping model from visual feedback to robot control. A novel fuzzy neural network is proposed to realize the nonlinear mapping model effectively, which is essential to implement six-degree-of-freedom control of robot hand. Simulation results show the performance of the proposed method.

A neuro-fuzzy (NF) hybrid position/force control with vibration suppression is developed in this paper for a space robot with flexible dual-arms handling a rigid object. An impedance force control algorithm is derived using force decomposition, and then singular perturbation method is used to construct the composite control, where an adaptive NF inference system is employed to approximate the inverse dynamics of the space robot. Finally, an example is employed to illustrate the validity of the proposed control scheme.

This paper presents an observer for robotic systems using FNN method to estimate the joint velocities of a robot, and then H_∞ approach is embedded to attenuate the effect of external distributes and parametric uncertainties of the robotic systems. Then a simulation example of 2-DOF robotic systems is given at last, from the simulation results, we can see the well performance of the designed observer and the estimation errors of the joint velocities are negligible.

This paper proposes a novel self-learning PD (Proportional- Derivative) control method for mobile robot path-tracking problems. In the self-learning PD control method, a reinforcement-learning (RL) module is used to automatically fine-tune the PD coefficients with only evaluative feedback. The optimization of the PD coefficients is modeled as a Markov decision problem (MDP) with continuous state space. Using an improved AHC (Adaptive Heuristic Critic) learning control method based on recursive least-squares algorithms, the near-optimal control policies of the MDP are approximated efficiently. Besides its simplicity, the self-learning PD controller can be adaptive to uncertainties in the environment as well as the mobile robot dynamics. Simulation and experimental results on a real mobile robot illustrate the effectiveness of the proposed method.

In view of the collision avoidance problem of multi-moving-obstacles in path planning of mobile robot, we present a solution based on reinforcement learning and ART2 (Adaptive Resonance Theory 2) neural network as well as the method of rule-based collision avoidance. The simulation experiment shows that the solution is of good flexibility and can solve the problem on random moving obstacles.

Based on Feedback-Error-Learning (FEL), an adaptive dynamic inverse control approach for single-axis rotational maneuver of spacecraft with flexible appendages by use of on-off thrusters is discussed. This approach uses a conventional feedback controller (CFC) concurrently with a Nonlinear Auto-Regressive Exogenous Input (NARX) neural network, and the NARX neural network can act dynamic adaptive inverse feed-forward controller, which is adapted on-line using the output of the CFC as its error signal, to improve the performance of a conventional non-adaptive feedback controller. The neural network (NN) does not need training in advance, and can utilize input and output on-line information to learn the systematic parameter change and unmodeled dynamics, so that the self-adaptation of control parameter is adjusted. However, the CFC should at least be able to stabilize the system under control when used without the neural network. The numerical simulations have shown that the control strategy is effective and feasible.

The control of an unknown multivariable nonlinear process represents a challenging problem. Model based approaches, like Generalized Minimum Variance, provide a flexible framework for addressing the main issues arising in the control of complex nonlinear systems. However, the final performance will depend heavily on the models representing the system. This work presents a comparative analysis of two modelling approaches for nonlinear systems, namely Artificial Neural Network (ANN) and Gaussian processes. Their advantages and disadvantages as building blocks of a GMV controller are illustrated by simulation.

Many of nonlinear control systems are sampled-data system, i.e. the continuous-time nonlinear plants are controlled by digital controllers. So it is important to investigate that if the solution of the discrete-time output regulation problem is effective to sampled-data nonlinear control systems. Recently a feedforward neural network based approach to solving the discrete regulator equations has been presented. This approach leads to an effective way to practically solve the discrete nonlinear output regulation problem. In this paper the approach is used to sampled-data nonlinear control system. The simulation of the sampled-data system shows that the control law designed by the proposed approach performs much better than the linear control law does.

In this study, we introduce a new neurogenetic approach to the design of fuzzy controller. The development process exploits the key technologies of Computational Intelligence (CI), namely genetic algorithms and neurofuzzy networks. The crux of the design methodology is based on the selection and determination of optimal values of the scaling factors of the fuzzy controllers, which are essential to the entire optimization process. First, the tuning of the scaling factors of the fuzzy controller is carried out, and then the development of a nonlinear mapping for the scaling factors is realized by using GA based Neuro-Fuzzy networks (NFN). The developed approach is applied to a nonlinear system such as an inverted pendulum where we show the results of comprehensive numerical studies and carry out a detailed comparative analysis.

This paper examines explicit rate congestion control for data networks. A neural network (NN) adaptive controller is developed to control traffic where sources regulate their transmission rates in response to the feedback information from network switches. Particularly, the queue length dynamics at a given switch is modeled as an unknown nonlinear discrete time system with cell propagation delay and bounded disturbances. To overcome the effects of delay an iterative transformation is introduced for the future queue length prediction. Then based on the causal form of the dynamics in buffer an adaptive NN controller is designed to regulate the queue length to track a desired value. The convergence of our scheme is derived mathematically. Finally, the performance of the proposed congestion control scheme is also evaluated in the presence of propagation delays and time-vary available bandwidth for robustness considerations.

By using differential neural networks, we present a novel robust adaptive controller for a class of unknown nonlinear systems. First, dead-zone and projection techniques are applied to neural model, such that the identification error is bounded and the weights are different from zero. Then, a linearization controller is designed based on the neuro identifier. Since the approximation capability of the neural networks is limited, four kinds of compensators are addressed.

Proportional-Integral-Derivative (PID) controllers are widely used in process control industry for years. Those plants are in general, slow-response with moderate or low damping characteristics. Zeigler-Nichols(ZN) tuning methods are some of design approaches for finding PID controllers. Basilio and Matos(BM) pointed out a systematic way to design PID to meet transient performance specifications. As for the low-damping, slow-response plants, the BM approach also failed to find the controller. We suggest two approaches to overcome this afore-mentioned difficulty. These two approaches are approach of using approximate 2nd order zeros to cancel the embedded plant poles, and GA-based fine-tuning approach.

Accurate multi-step state predication is very important for the fault tolerant control of nonlinear systems with input delay. Neural network (NN) possesses strong anti-interference ability at multi-step predication, but the predication accuracy is usually not satisfactory. The strong tracking filter (STF) can reduce adaptively estimate bias and has the ability to track changes in nonlinear systems. Thus in this paper the STF and the NN are combined together to provide more accurate multi-step state predication. Based on the state predication an active fault tolerant control law is then proposed against sensor failures of nonlinear time delay systems. Simulation results on a three-tank-system show the effectiveness of the proposed fault tolerant control law.

Recurrent neural network (RNN) is used to model product quality of batch processes from process operational data. Due to model-plant mismatches and unmeasured disturbances, the calculated control policy based on the RNN model may not be optimal when applied to the actual process. Model prediction errors from previous runs are used to improve RNN model predictions for the current run. It is proved that the modified model errors are reduced from run to run. Consequently control trajectory gradually approaches the optimal control policy. The proposed scheme is illustrated on a simulated batch reactor.

A time-delay recurrent neural network (TDRNN) model is presented. TDRNN has a simple structure but far more “depth” and “resolution ratio” in memory. A TDRNN controller for dynamic systems is proposed. A dynamic recurrent back-propagation algorithm is developed and the optimal adaptive learning rates are also proposed to guarantee the global convergence. Numeral experiments for controlling speeds of ultrasonic motors show that the TDRNN has good effectiveness in identification and control for dynamic systems.

A neural network is used as the feedforward controller in a feedforward-feedback combined system. The network is trained by the feedback output that is minimized during training and most control action for disturbance rejection is finally performed by the rapid feedforward action of the network. The neural feedforward controller is independent of the model of plant and self-adaptive to time-variable system. The dynamic architecture of the neural controller is chosen, and the methods for delay time treatment and training network on line are investigated. An application to oxygen replenishment of an underwater plant is used to prove the effectiveness of the scheme and the simulation shows that the dynamic performance of the oxygen control is greatly improved by this neural combined control system.

The cerebella model articulation controller (CMAC) neural network control scheme is a powerful tool for practical real-time nonlinear control applications. The conventional leaning controller based on CMAC can effectively reduce tracking error, but the CMAC control system can suddenly diverge after a long period of stable tracking, due to the influence of accumulative errors when tracking continuous variable signals such as sinusoidal wave. A new self-learning controller based on CMAC is proposed. It uses the dynamic errors of the system as input to the CMAC. This feature helps the controller to avoid the influence of the accumulative errors and the stability of the system is ensured. The simulation results show that the proposed controller is not only effective but also of good robustness. Moreover, it has a high learning rate, which is important to online learning.

This paper extends pole placement control of conventional linear systems to a class of nonlinear dynamical systems via neural networks. An application of typical inverted pendulum illustrates the design method. Multi-layer neural networks are selected to approach nonlinear components arbitrarily, and then are represented by linear difference inclusion (LDI) format. With pole placement regions formed in linear matrix inequalities (LMIs), quadratic stability theory is used as a basic analysis and synthesis methodology. Pole placement controllers via state feedback are derived by numerical solutions of a set of coupled LMIs. Applying common back propagation algorithm (BP) for networks training and interior point computation for LMI solving, some simulation results show the validity of pole placement control.

Based on neural networks, a robust control design method is proposed for strict-feedback block nonlinear systems with mismatched uncertainties. Firstly, Radial-Basis-Function (RBF) neural networks are used to identify the nonlinear parametric uncertainties of the system, and the adaptive tuning rules for updating all the parameters of the RBF neural networks are derived using the Lyapunov stability theorem to improve the approximation ability of RBF neural networks on-line. Considering the known information, neural network and robust control are used to deal with the design problem when control coefficient matrices are unknown and avoid the possible singularities of the controller. For every subsystem, a nonlinear tracking differentiator is introduced to solve the “computer explosion” problem in backstepping design. It is proved that all the signals of the closed-loop system are uniform ultimate bounded.

A model reference control algorithm based on Support Vector Machine (SVM) for discrete nonlinear systems is proposed in this article. It uses SVM as regression tools to learn the feed forward controller from input and output data. Further more, a compensator is used as the feed back controller to make the whole system more robust. Advantages of SVM and the robust compensator help the method perform excellently as shown in the experiments.

The paper provides a new style of PID controller that is based on neural network according to the traditional one’s mathematical formula and neural network’s ability of nonlinear approximation. It also discusses the corresponding learning algorithm and realizing method. This new controller is proven highly practical and effective in the simulation test. This new controller has more advantage than the traditional one, such as more convenient in parameter regulating, better robust, more independence and adaptability on the plant, etc.

A nonlinear predictive control algorithm based on fuzzy model is presented for a family of complex system with severe nonlinearity such as Proton exchange membrane fuel cell (PEMFC). In order to implement nonlinear predictive control of the plant, the fuzzy model is identified by learning offline and rectified online. The model parameters are initialized by fuzzy clustering, and learned using back-propagation algorithm offline. If necessary, it can be rectified online to improve the predictive precision in the process of real-time control. Based on the obtained model, discrete optimization of the control action is carried out according to the principle of Branch and Bound (B&B) method. The test results demonstrate the effectiveness and advantage of this approach

This paper presents an adaptive control system using wavelet neural networks (WNN) for an induction servo drive motor with complex of non-linear, multivariable, strong coupling, slow time-varying properties, etc., and many uncertainties such as mechanical parametric variation and external disturbance. The motivation of developing a new method is to overcome the limitation of conventional control methods which depends on the accurate model and cannot guarantee satisfactory control performance. The proposed scheme with on-line learning has the good tracking and dynamic performance, the ability of adaptive learning from the process and good robustness to uncertainties. Simulation results demonstrate the effectiveness of the proposed method.

In this paper, an adaptive single neuron-based PID controller for the Triaxial and Torsional Shear Apparatus is proposed. Considering variations of the control precision, the single neuron adaptive PID controller is used to construct the control system to achieve the adaptive control of triaxial and torsion testing. The single neuron adaptive PID controller using hybrid Supervised-Hebb rule is proposed to tune control parameters. The testing results of actual application show that the single neuron adaptive PID controller makes better improvement in the control precision and robustness of control system.

In plastic injection molding, the ram velocity plays an important role in production quality. This paper introduces a new method, which is a combination of the current cycle feedback control and neural network (NN) learning, to control the ram velocity in injection process. It consists of two parts: a PD controller (current cycle feedback control) is used to stabilize the system, and the feedforward NN learning is used to compensate for nonlinear/unknown dynamics and disturbances, thereby enhancing the performance achievable with feedback control alone. The simulation results indicate that the proposed NN learning control scheme outperforms the conventional PD controller and can greatly reduce tracking errors as the iteration number increase.

For a discrete-time nonlinear MIMO system, a multiple models neural network decoupling controller is designed in this paper. At each equilibrium point, the system is expanded into a linear and nonlinear term. These two terms are identified using two neural networkss, which compose one system model. Then, all models, which are got at all equilibrium points, compose the multiple models set. At each instant, the best model is chosen as the system model according to the switching index. To design the controller accordingly, the nonlinear term and the interactions of the best model is viewed as measurable disturbance and eliminated by the use of the feedforward strategy. The simulation example shows that the better system response can be got even when the system is changed around these equilibrium points.

An algorithm of the feedback-assisted iterative learning control (FBAILC) was proposed for a batch repeatable operation process. Control law of FBAILC was based on the inverse of process model, added the filter polynomial in iterative learning and analyzed the convergence of FBAILC algorithm. On-line estimator method of the process parameters was introduced in application, which achieved the parameter self-tuning of controller. The effectiveness of the proposed method was demonstrated by simulation results.

Artificial neural networks (ANNs) have found many potential applications in power systems operation and control recently. This paper presents a categorization of the main significant applications of neural networks, which includes power system controller design, power system security assessment and load forecasting. It is desired that they are helpful to the construction of more efficient, robust ANNs to solve a broader range of problems in power systems.

This paper proposes a novel fermentation control method. Two stages are involved. First, propose the fermentation time model and the optimal fermentation temperature model based on RBF Neural networks. Second, on the base of the two models, propose the novel fermentation control method by which different fermentation batch can adopt different optimal fermentation temperature trajectory which fits itself. Using this method, each fermentation batch can be fermented at optimal fermentation temperature trajectory and will improve average product proportion. The practical application showed that this method can improve average product proportion 3% effectively.

A study is performed to investigate the state evolution of a kind of recurrent neural network. The state variable in the neural system summarize the information of external excitation and initial state, and determine its future response. The recurrent neural network is trained by the data from a dynamic system so that it can behave like the dynamic system. The dynamic systems include both input-output black-box system and autonomous chaotic system. It is found that the state variables in neural system differ from the state variable in the black-box system identified, this case often appears when the network is trained with input-output data of the system. The recurrent neural system learning from chaotic system exhibits an expected chaotic character, its state variable is the same as the system identified at the first period of evolution and its state evolution is sensitive to its initial state.

A new adaptive robust friction compensation for servo system based on luGre model is proposed. Considered the uncertainty of steady state parameters in the friction model, a RBF neural network is adopted to learn the nonlinear friction-velocity relationship in steady state. The bristle displacement is observed using the output of the network. Nonlinear adaptive robust control laws are designed based on backstepping theory to compensate the unknown system parameters. System robustness and asymptotic results is proved and shown in simulation results.

In general, RBF neural network cannot match nonlinear systems exactly. Unmodeled dynamic leads parameters drift and even instability problem. According to system identification theory, robust modification terms must be included in order to guarantee Lyapunov stability. This paper suggests new learning laws for normal and adjustable RBF neural networks based on Input-to-State Stability (ISS) approach. The new learning schemes employ a time-varying learning rate that is determined from input-output data and model structure. The calculation of the learning rate does not need any prior information such as estimation of the modeling error bounds.

Artificial Neural Networks (ANNs) have provided an interesting and labor-saving approach for system identification. However, ANNs fall short when an explicit model is needed. In this paper, a method of getting the explicit model by extracting it from a trained ANN is proposed. To identify a system, a Multi-Layer Perceptron (MLP) is constructed, trained and a polynomial model is extracted from the trained network. This method is tested in the experiments and shows its capability for system identification, compared with the Least Squared method.

Based on the principle of Radial Basis Function (RBF) Neural Network, a learning method is presented for the identification of a complex system model. The RBF algorithm is employed on the learning and identifying process of the nonlinear model. The simulation results show that the presented method has good effect on speeding up the learning and approaching process of the nonlinear complex model, and has an excellent performance on learning convergence.

Network topological optimization in communication network is to find the topological layout of network links with the minimal cost under the constraint that all-terminal reliability of network is not less than a given level of system reliability. The all-terminal reliability is defined as the probability that every pair of nodes in the network can communicate with each other. The topological optimization problem is an NP-hard combinatorial problem. In this paper, a noisy chaotic neural network model is adopted to solve the all-terminal network design problem when considering cost and reliability. Two sets of problems are tested and the results show better performance compared to previous methods, especially when the network size is large.

The capacity and performance of code division multiple access (CDMA) systems are limited by multiple access interference (MAI) and ”near-far” problem. Space-time multiuer detection combined with adaptive wavelet networks over frequency selective fading multipath channels is proposed in this paper. The structure of the multiuser detector is simple and its computational complexity mostly lies on that of wavelets networks. With numerical simulations and performance analysis, it is shown that the proposed detector can converge at the steady state rapidly and offer significant performance improvement over the space-time matched filtering detector, the conventional RAKE receiver and matched filter detector only at the time domain. Therefore, it can suppress the MAI and solve the ”near-far” problem effectively.

Wireless sensor networks have recently emerged as a premier research topic. They have great longterm economic potential, ability to transform our lives, and pose many new system-building challenges. One of the fundamental problems in sensor networks is the calculation of the coverage. In this paper, an optimal distribution based on Genetic Algorithm is proposed in the initial planning of sensor network. Moreover, a new optimizing algorithm of sensor node distribution is designed by utilizing node topology in sensor network, which provides a sound effective means for the topology management of sensor network. Simulation shows that this efficient algorithm optimally solved the best-coverage problem raised in [1].

A Fast de-hopping and FHP estimation model for Direct Sequence/Frequency Hopping Spread Spectrum (DS/FHSS) system is proposed. The Neural Networks (NNs) were used to mimic the Parallel Matched Filtering (PMF). The signal samples and its Fast Fourier Transform (FFT) were used for Back propagation Neural Network (BNN) training. The FH patterns designated as concatenated prime codes [8] were used for the Radial Basis Function (RBF) training. Computer simulations show that the proposed method can effectively identify the frequency and estimate its pattern. Small hardware resources compared with PMF hardware.

In Time-Division-Duplex (TDD) wireless communications, downlink beamforming performance of a smart antenna system can be degraded due to variation of spatial signature vectors in vehicular scenarios. To mitigate this, downlink beams must be adjusted according to changing propagation dynamics. This can be achieved by modeling spatial signature vectors in the uplink period and then predicting them for new mobile position in the downlink period. This paper examines time delay feedforward neural network (TDFN), adaptive linear neuron (ADALINE) network and autoregressive (AR) filter to predict spatial signature vectors. We show that predictions of spatial signatures using these models provide certain level of performance improvement compared to conventional beamforming method under varying mobile speed and filter (delay) order conditions. We observe that TDFN outperforms ADALINE and AR modeling for downlink SNR improvement and relative error improvement with high mobile speed and higher filter order/delay conditions in fixed Doppler case in multipaths.

It is difficult for numerical method to forecast and control the anode shape in Electrochemical Machining (ECM) with an uneven interelectrode gap, so this paper introduces Artificial Neural Network (NN) to solve this problem. The experiments with different cathode shapes and minimal interelectrode gaps are carried out and the corresponding anode shapes are obtained. Those cathode and anode shapes are discretized and taken as the input samples of a B-P network. Quasi-Newton algorithm is used to train this network. To verify the validity of the trained network, results obtained by NN are compared with that obtained by the experiments, and the results show that the former is close to the later, which indicates it is feasible to apply NN to solve this problem.

Computing the Multicast QoS routing is an NP-complete problem. Generally, it was solved by heuristic algorithms, which include tabu search, simulated annealing, genetic algorithms (GA), neural networks (NN), etc. In this paper, a hybrid neural network and genetic algorithm approach is described to compute the multicast QoS routing tree. The integration of neural network and genetic algorithm can overcome the premature and increase the convergence speed. The simulation results show that the proposed approach outperforms the traditional GA and NN algorithm in terms of both solution accuracy and convergence speed.

With dynamics property and highly parallel mechanism, recurrent neural networks (RNN) can effectively implement blind adaptive multiuser detection at the circuit time constant level. In this paper, the RNN based blind adaptive multiuser detection is extended to ubiquitous asynchronous DS-CDMA systems, and the performance of the output signal to interference plus noise ratio, asymptotic multiuser efficiency, computational complexity, operating time, and mismatch of the detector are quantitatively analyzed. With performance analysis and numerical simulations, it is shown that RNN based blind adaptive multiuser detection can converge at the steady quickly and offer significant performance improvement over some existing popular detectors in eliminating multiple access interference and ”near-far” resistance.

A neural network model is described which simulate the Parallel Matching Filtering (PMF) for Direct Sequence Spread Spectrum (DSSS) signal acquisition. This system is based on training the Counter Propagation Network (CPN) in all half chip phase shifts of the Pseudo Noise (PN) code. The trained network can be used at the receiver for the signal acquisition. The CPN performance in Additive Wight Gaussian Noise (AWGN) channel is evaluated. Computer simulations carried on maximal length sequences of length N=256, show that the proposed system can effectively decide the half chip phase shift of the received code even at much lower Signal to Noise ration (S/N) until S/N = -27.74dB. This model has a simple architecture, so can be realized in a simple hardware. This makes the neural network based acquisition technique faster and more robust than the other conventional acquisition techniques.

In this paper, we consider the channel assignment problem in cellular mobile communication systems, which assigns a channel to every requested call with the minimum span of channels subject to interference constraints. In this paper, a multi-stage heuristic algorithm using neural networks for the channel assignment problem is proposed. The proposed algorithm is devised to find first a good initial feasible assignment, then a locally optimal assignment, and finally a overall best quality assignment. The proposed algorithm has been applied to the Philadelphia benchmark instances. Experimental results show that the proposed method is competitive with the other existing algorithms.

A new spread spectrum communication system based on CNN is proposed in this paper. Chaos is generated with three cell CNN, then it’s transferred to a digital sequence. The chaotic sequence is better than gold sequence when they are utilized in direct sequence spread spectrum system. Compared with the traditional gold sequence system, there is 2dB improvement in CNN chaotic sequence system when the channel is additive white Gaussian noise channel, and there is more improvement in CNN chaotic sequence system when the channel is multi-path channel. The structure of hardware CNN spread spectrum system is also shown at last.

In ATM networks, congestion control is a distributed algorithm to share network resources among competing users.It is important in situation where the availability of resources and the set of competing users vary over time unpredictably, round trip delay is uncertain and constraints on queue, rate and bandwidth are saturated, which results in wasted bandwidth and performance degradation. A neural congestion control algorithm is proposed by real-time scheduling between the self-tuning neural controller and the modified EFCI algorithm, which makes the closed-loop systems more stable and robust with respect to uncertainties and more fairness in resources allocation. Simulation results demonstrated the effectiveness of the proposed controller.

In this paper, we proposed a novel and simple linear distortion compensation scheme utilizing Kohonen’s neural network for digital communications. The scheme compensates the distortions that exist in modulator and demodulator, at the decision stage of the receiver at the same time of data transmission and decision process. The scheme can compensate not only static but also changing distortions. Computer simulations using QPSK signal have confirmed the effectiveness of the proposed scheme.

For conventional space-time block coding (STBC), the decoding usually requires accurate channel state estimation. However, the accuracy of the channel estimation strongly determines the system performance. Independent component analysis (ICA) techniques can be applied to perform blind detection so as to detect the transmitted symbols without any channel information. In this paper, we establish the special ICA model for the STBC system and study the performance of STBC schemes based on ICA neural networks; what is more, several different ICA algorithms of blind separation are used for performance evaluation. By using the ICA based schemes, the good robustness against channel estimation errors and time variation of the fading channels can be acquired. The computer simulation analyzes the bit error rate (BER) performance of these methods and indicates the optimal separation algorithm suitable for STBC scheme.

Space-time block coding (STBC) can provide a fairly good diversity advantage over the Rayleigh fading channel, and the beam space-time block coding (BSTBC) scheme combining STBC with the transmit beamforming (TBF) can achieve both diversity gains and beamforming gains. In this paper, we establish a BSTBC model firstly, and then present a method which considers transmit antenna array selection (TAAS) used for BSTBC at the transmitter side and applies independent component analysis (ICA) techniques to assist the channel estimation and achieve blind detection at the receiver side. The utility of TAAS can obtain higher diversity gains as well as keep the original number of the radio-frequency chains. The ICA-based blind scheme can enhance the flexibility of the communication system and adapt to more rapidly varying channels. Simulation results for Rayleigh channel demonstrate the validity and significant performance improvement of the proposed scheme.

On condition that the noise, such as the reverberation, could be modeled as a dynamical model with lower dimensions, it would be picked out from its mixture with a useful signal by using the nonlinear dynamic method proposed in this paper. In other words, the useful signal could be separated from the noise with this method, which constructs the nonlinear inverse to linear filter, based on the spectrum difference between the noise and the useful signal, in virtue of successive approximation with Radial Basis Function Neural Networks. Two examples, with the sine pulse as the useful signal, are displayed. The artificial chaotic signal plays the role of the noise in one example, and the actual reverberation in another. These examples confirm the feasibility of this method.

Mixed pixels exist in almost all multispectral and hyperspectral remote sensing images. Their existence impedes the quantification analysis of remote sensing images. This paper proposes a new scheme for detection and classification of subpixel spectral signatures in multispectral remote sensing images. By minimizing the energy function with two special constraints, the mixed pixels can be decomposed more precisely. Further, we point out that our scheme can also be used for the decomposition of mixed pixels in hyperspectral remote sensing imagery. Finally, the performances of the proposed scheme are demonstrated experimentally and the comparisons of the performances with conventional methods such as back-propagation (BP) neural network are made.

A voting-mechanism-based fuzzy neural network model for identifying 11 kinds of mineral waters by its taste signals is proposed. In the model, A classification rule extracting algorithm based on discretization methods in rough sets is developed to extract fewer but robust classification rules, which are ease to be translated to fuzzy if-then rules to construct a fuzzy neural network system. Finally, the particle swarm optimization is adopted to refine network parameters. Experimental results show that the system is feasible and effective.

According to the recently neurophysiology research results, a novel signal detection subsystem of radar based on HA-CNN is proposed in this paper. With a kind of improved chaotic neuron that is based on discrete chaotic map and Aihara model, Hierarchical-Associative Chaotic Neural Network (HA-CNN) exhibits promising chaotic characteristics. The function of HA-CNN in the signal detection subsystem of radar is to reduce the influence of the environmental strong noisy chaotic clutter and distill the useful signal. The systematic scheme of signal detection with HA-CNN and the detailed chaotic parameter region of HA-CNN applied in signal detection are given and the results of analysis and simulation both show this kind of signal detection subsystem has good detecting ability and fine noise immunity.

The problem of real time signal detection in the noise and its applications to the denoising single-trial evoked potentials (EP) was investigated. The main objective is to estimate the amplitude and the latency of the single trail EP response without losing the individual properties of each epoch, which is important for practical clinical applications. Based on the radial basis function neural network (RBFNN), a method in terms of normalised RBFNN was proposed to obtain preferable results against other nonlinear methods such as ANC with RBFNN prefilter and RBFNN. The performance of the proposed methods was also evaluated with MSE and the ability of tracking peaks. The experimental results provide convergent evidence that the NRBFNN can significantly attenuate the noise and successfully identify the variance between trials. Both simulations and real signal analysis show the applicability and the effectiveness of the proposed algorithm.

Investigation of the states of human brain through the elec-troencephalograph (EEG) is an important application of EEG signals. This paper describes the application of an artificial neural network technique together with a feature extraction technique, the wavelet packet transformation, in classification of EEG signals. Feature vector is extracted by wavelet packet transform. Artificial neural network is used to recognize the brain statues. After training, the BP and RBF neural network are able to correctly classify the brain states, respectively. This method is potentially powerful for brain states classification.

The approach of the Dynamic Spectrum evaluates the pulsatile part of the entire optical signal at different wavelength. In the course of collecting the pulsatile spectrum signal in vivo, it is inevitable to be interfused with yawp signals as high frequency interference, baseline drift and so on. Using the traditional adaptive filter, it is very difficult to collect the reference signal from the in vivo experiment. In this paper, Daubechies wavelet adaptive filter based on Adaptive Linear Neuron networks is used to extract the signal of the pulse wave. Wavelet transform is a powerful tool to disclose transient information in signals. The wavelet used is adaptive because the parameters are variable, and the neural network based adaptive matched filtering has the capability to ”learn” and to become time-varying. So this filter estimates the deterministic signal and removes the uncorrelated noises with the deterministic signal. This method can get better result than nonparametric results. This filter is found to be very effective in detection of symptoms from pulsatile part of the entire optical signal.

This paper presents a novel spectral analysis and classification technique, which is based on multi-scale feature extraction and neural networks. We propose two feature extraction methods in wavelet domain to implement de-noising process and construct feature spectra. Then a radial basis function network is employed for classifying spectral lines. The input of the neural network is the feature spectra, which is produced by the proposed methods. Real world data experimental results show that our technique is robust and efficient. The classification results are much better than the best results obtained by principle component analysis feature extraction method.

In this paper, we propose a novel Neural Fuzzy Filter (NFF) to remove impulse noise from highly corrupted images. The proposed filter consists of a fuzzy number construction process, a neural fuzzy filtering process and an image knowledge base. First, the fuzzy number construction process will receive sample images or the noise-free image, then construct an image knowledge base for the neural fuzzy filtering process. Second, the neural fuzzy filtering process contains of a neural fuzzy mechanism, a fuzzy mean process, and a fuzzy decision process to perform the task of impulse noise removing. By the experimental results, NFF achieves better performance than the state-of-the-art filters based on the criteria of Mean-Square-Error (MSE). On the subjective evaluation of those filtered images, NFF also results in a higher quality of global restoration.

The normal Kalman filter (KF) is deficient in adaptive capability, at the same time, the estimation accuracy of the neural network (NN) filter is not very well and the performance depends on the artificial experience excessively. It is proposed to incorporate a back-propagation (BP) neural network into the adaptive federal KF configuration for the SINS/GPS/TAN (Terrain Auxiliary Navigation)/SAR (Synthetic Aperture Radar) integrated navigation system. The proposed scheme combines the estimation capability of adaptive KF and the learning capability of BP NN thus resulting in improved adaptive and estimation performance. This paper addresses operation principle, algorithm and key techniques. The simulation results show that the performance of the BP NN aided filter is better than the stand-alone adaptive Kalman filter’s.

A solely excitatory oscillator network (SEON) is proposed for color image segmentation. SEON utilizes its parallel nature to reliably segment images in parallel. The segmentation speed does not decrease in a very large network. Using NBS distance, SEON effectively segments color images in term of human perceptual similarity. Our model obtains an average segmentation rate of over 98.5%. It detects vague boundaries very efficiently. Experiments show that it segments faster and more accurately than other contemporary segmentation methods. The improvement in speed is more significant for large images.

This paper presents a new image de-noising method, which based on the image representation model of radial basis function neural network. In this model, the number and distribution of the centers (which are set to the pixels of the observed image) are fixed, and the model parameters of the image representation are chosen by cross-validation method. Experimental results show that the model can represent the image well, and proposed method can reduce the noises in images without need any noise knowledge in priori.

A method for restoration of ultrasonic C-scan images is presented by using a radial basis function network. The method attempts to reproduce the mapping between the degraded C-scan image and the high quality one by training a RBF network. The inputs for training are the sub-images divided from C-scan image of flat-bottom hole of size 3mm and the output is the corresponding center in high quality image. After the network was trained, the other C-scan images were used to verify the network. The results show that the network produces good restored results, in which the noise is removed and the edges are deblurred. Comparing the restored results by the networks trained by the different sub-images, the sub-images with size 7x7, scanning step of 3 are determined as the optimal inputs for training.

Recent researches indicate that pulse coupled neural network (PCNN) can be effectively utilized in image segmentation. However, the near optimal parameter set should always be predetermined to achieve desired segmentation result for different images, which impedes its application for segmentation of various images. So far as that is concerned, there is no method of adaptive parameter determination for automatic real-time image segmentation. To solve the problem, this paper brings forward a new automatic segmentation method based on a simplified PCNN with the parameters determined by images’ spatial and grey characteristics adaptively. The proposed algorithm is applied to different images and the experimental results demonstrate its validity.

In this paper, a new method is proposed to estimate the accurate pose of face images. A face image collection device is designed to collect face image with accurate pose. Principle Component Analysis (PCA) is used to set up the eigenspace of face pose, then a fuzzy C-means clustering method is applied to divide the train samples into several classes, and to decide the centers of the classes. When a face image is presented to our system, we first project the image into the eigenspace, then the subordinate degree of the input face image to each class is calculated, finally the pose of the input face image is calculated combining the subordinate degrees. Experiments show that the proposed method can effectively estimate the accurate pose of face image.

Large computational quantity and cumulative error are main shortcomings of add- weighted one-rank local-region single-step method for multi-steps prediction of chaotic time series. A local-region multi-steps forecasting model based on phase-space reconstruction is presented for chaotic time series prediction, including add-weighted one-rank local-region multi-steps forecasting model and RBF neural network multi-steps forecasting model. Simulation results from several typical chaotic time series demonstrate that both of these models are effective for multi-steps prediction of chaotic time series.

Although, in theory, the neural network is able to fit, model and predict any continuous determinant system, there is still an obstacle to prevent the neural network from wider and more effective applications due to the lack of complete theory of model identification. This paper addresses this issue by introducing a universal method to achieve nonlinear model identification. The proposed method is based on the theory of information entropy and its development, which is called as nonlinear irreducible autocorrelation. The latter is originally defined in the paper and could determine the optimal autoregressive order of nonlinear autoregression models by investigating the irreducible auto-dependency of the investigated time series. Following the above proposal, robust prediction of chaotic time series became realizable. Our idea is perfectly supported by computer simulations.

A simple model based on the combination of neural network and wavelet techniques named wavelet neural network (WNN) is proposed. Thanks to the time-frequency analysis feature of wavelet, a selection method that takes into account the domain of input space where the wavelets are not zero is used to initialize the translation and dilation parameters. A proper choice of initialize parameters is found to be crucial in achieving adequate training. Training algorithms for feedback WNN is discussed too. Results obtained for a nonlinear processes is presented to test the effectiveness of the proposed method. The simulation result shows that the model is capable of producing a reasonable accuracy within several steps.

Japan has so may solitary islands and wide coastal areas. In the air of these areas, there are so many grains of sea-salt. The salt grains adhere to distribution lines and damage the lines. In particular, the damage of covered electric wires is a serious problem. However, any method has not been proposed that judges if distribution lines are injured by salt water. This paper proposes a neural network method to determine injury of distribution lines by salt.

Determining the location of cortical activity from electroencephalographic (EEG) data is important theoretically and clinically. Estimating the location of electric current source from EEG recordings is not well-posed mathematically because different internal source configurations can produce an identical external electromagnetic field. In this paper we propose a new method for EEG source localization using Independent Residual Analysis (IRA). First, we apply Independent Residual Analysis on EEG data to divide the raw signals into the independent components. Then for each component, we employ the least square method to locate the dipole. By localizing multiple dipoles independently, we greatly reduce our search complexity and improve the localization accuracy. Computer simulation is also presented to show the effectiveness of the proposed method.

G-protein coupled receptors (GPCRs) are a class of pharmacologically relevant transmembrane proteins with specific characteristics. They play a key role in different biological process and are very important for understanding human diseases. However, ligand specificity of many receptors remains unknown and only one crystal structure solved to date. It is highly desirable to predict receptor’s type using only sequence information. In this paper, Support Vector Machine is introduced to predict receptor’s type based on its amino acid composition. The prediction is performed to the amine-binding classes of the rhodopsin-like family. The overall predictive accuracy about 94% has been achieved in a ten-fold cross-validation.

An individual blood glucose control model (IBGCM) based on the Mixture of Experts (MOE) neural networks algorithm was designed to improve the diabetic care. MOE was first time used to integrate multiple individual factors to give suitable decision advice for diabetic therapy. The principle of MOE, design and implementation of IBGCM were described in details. The blood glucose value (BGV) from IBGCM extremely approximated to training data (r=0.97± 0.05, n=14) and blood glucose control aim (r=0.95± 0.06, n=7).

Evoked potential (EP) is non-stationary during the recording of electroencephalograph (EEG). This paper promotes a method to track the variation of EP’s amplitude by the application of independent component analysis (ICA) and wavelet transform (WT). The utilization of the spatial information and multi-trial recording improves the signal-to-noise ratio (SNR) greatly. The variation trend of EP’s amplitude across trials can be evaluated quantitatively. Our result on real auditory evoked potential shows a drop of about 40% on the amplitude of EP during 10 minutes recording. The present work is helpful to study the uncertainty and singularity of EP. Furthermore, it will put forward the reasonable experiment design of EP extraction.

Accurate diagnosis among a group of histologically similar cancers is a challenging issue in clinical medicine. Microarray technology brings new inspiration in solving this problem on genes level. In this paper, a novel gene selection method is proposed and a BP classifier is constructed for gene expression-based cancer classification. By testing on the open leukemia data set, it shows excellent classification performance. 100% classification accuracy is achieved when 46 informative genes are selected. Reducing genes number to 6, only a sample is misclassified. This study provides a reliable method for molecular cancer classification, and may offer insights into biological and clinical researches.

Nonnegative matrix factorization (NMF) is a powerful feature extraction method for nonnegative data. This paper applies NMF to feature extraction for Electroencephalogram (EEG) signal classification. The basic idea is to decompose the magnitude spectra of EEG signals from six channels via NMF. Primary experiments on signals from one subject performing two tasks show high classification accuracy rate based on linear discriminant analysis. Our best results are close to 98% when training data and testing data from the same day, and 82% when training data and testing data from different days.

This paper proposes a novel clustering analysis algorithm based on principal component analysis (PCA) and self-organizing maps (SOMs) for clustering the gene expression patterns. This algorithm uses the PCA technique to direct the determination of the clusters such that the SOMs clustering analysis is not blind any longer. The integration of the PCA and the SOMs makes it possible to powerfully mine the underlying gene expression patterns with the practical meanings. In particular, our proposed algorithm can provide the informative clustering results like a hierarchical tree. Finally, the application on the leukemia data indicates that our proposed algorithm is efficient and effective, and it can expose the gene groups associated with the class distinction between the acute lymphoblastic leukemia (ALL) samples and the acute myeloid leukemia (AML) samples.

In this paper the visual processing architecture is assumed to be hierarchical in structure with units within this network receiving both feed-forward and feedback connections. We propose a neural computational model of visual system, which is based on the hierarchical structure of feedback selectiveness of visual attention information and feature integration theory. The proposed model consists of three stages. Visual image input is first decomposed into a set of topographic feature maps in a massively parallel method at the saliency stage. The feature integration stage is based on the feature integration theory, which is a representative theory for explaining all phenomena occurring in visual system as a consistent process. At last stage through feedback selection, the saliency stimulus is localized in each feature map. We carried out computer simulation and conformed that the proposed model is feasible and effective.

This paper proposes a new method to discriminate the vascular conditions from biological signals by using a probabilistic neural network, and develops the diagnosis support system to judge the patient’s conditions on-line. For extracting vascular features including biological signals, we model the dynamic characteristics of an arterial wall by using mechanical impedance and estimate the impedance parameters ”beat-to-beat”. As a result, this system can be utilized for the actual surgical operation, and the vascular conditions can be discriminated with high accuracy using the proposed method.

Bidirectional recurrent neural network (BRNN) is a noncausal system that captures both upstream and downstream information for protein secondary structure prediction. Due to the problem of vanishing gradients, the BRNN can not learn remote information efficiently. To limit this problem, we propose segmented memory recurrent neural network (SMRNN) and obtain a bidirectional segmented-memory recurrent neural network (BSMRNN) by replacing the standard RNNs in BRNN with SMRNNs. Our experiment with BSMRNN for protein secondary structure prediction on the RS126 set indicates improvement in the prediction accuracy.

Independent component analysis (ICA) is a new powerful tool for blind source separation. This paper proposes a new algorithm that combines two existent algorithms, the improved infomax algorithm and the fastICA algorithm. Utilizing the initial weights obtained by the improved infomax algorithm, we can not only reduce the length of data which fastICA algorithm needs, but also enhance the convergence stability of fastICA algorithm. The effectiveness of the algorithm is verified by computer simulations.

Building a precise respiration system model is very helpful for setting appropriate ventilation conditions to fit each patient when artificial respiration is performed on the patient. In this paper, a new respiration system model is proposed, which is a second order nonlinear differential equation including volume dependent elastic term described by RBF network. The model is able to describe the nonlinear dynamics of respiration. By using Sagara’s numerical integration technique, a discrete-time identification model is derived. Then, off-line and on-line parameter estimation algorithms are presented. It is easy to obtain pulmonary elastance from identified model. The proposed model and the parameter estimation method are validated by clinical examples.

Inter-residue contacts map prediction is one of the most important intermediate steps to the protein folding problem. In this paper, we focus on protein inter-residue contacts map prediction based on radial basis function neural network (RBFNN), and propose a novel binary encoding scheme for the purpose of learning the inter-residue contact patterns, and get a better prediction results.

In this paper, we focus on modeling and exploring the genetic regulatory systems (GRS) with an artificial recurrent neural network (ARNN). Based on the approximation capability of the ARNN, the proposed model can be used to express and analyze the genetic regulatory systems of genetic components, even the large-scale genetic systems. Unlike the conventional ARNN, the ARNN used in the paper is the echo state network (ESN), in which the connection weights of internal neurons are fixed and only the output weights are adjustable. Thus, there are no cyclic dependencies between the trained readout connections and, training the genetic regulatory system becomes a simple linear regression task. The experiment studies shows the new genetic regulatory system modeled by ESN and trained from the fluorescent density of reporter protein has a satisfactory performance in modeling the synthetic oscillatory network of transcriptional regulatory of Escherichia coli cells.

This paper designs a new adaptive disparity filter which implements the binocular matching of disparity features in stereo images, and constructs a computational model as well as a stereopsis system based on the disparity filter along with mechanism of biological vision by integrating it with Grossberg’s FACADE theory, to process real-world images of 3-D scenes. By using this stereopsis system, depth perception of surfaces in real-world stereo images is simulated and realized.

A new reconstruction approach to computerized tomography(CT) with Cauchy Radial Basis Functions network is presented. The distribution of material parameters is represented by the weighting sum of Cauchy functions. The analytical formula of the line integral of Cauchy functions along any straight-line path is deduced, and the theoretical projection data along a bent ray is computed by optimization algorithm. The parameters in RBFs network are found by the learning rule based on the gradient decent method. The new reconstruction approach is suitable for the CT with a relatively small number of bent ray paths or projection dada, such as seismic tomography. Computer simulations show its good effects.

With increasing demands on higher performance, more safety and reliability of dynamic systems, especially on safety-critical systems, fault diagnosis became a research interests in recent years. In this paper, a systematic approach to design fault diagnosis and accommodation with compound approach such as applying improved robust observer to fault diagnosis on linearized system aided by dynamic neural network, which is trained to bridge the gap between simulated system and real system on nonlinear attributes and modeling errors. Using an instance of fault diagnosis on attitude control system of satellite attitude, advantages of new scheme are tested to be effective.

Parity space is a well-known model-based scheme for fault detection and diagnosis. However, the construction of parity vector is strictly based on the formulation of linear systems, and can hardly be extended to nonlinear cases. From the view of analytical redundancy and the nature of parity space, we propose a new parity check scheme for nonlinear deterministic systems, which can be realized by neural networks and the condition that the parity relation exists are also given theoretically. Simulation studies on the model of the three tank system DTS200 demonstrate the effectiveness of the new strategy, which can detect faults fast and accurately.

In view of non-stationary and non-linearity of vibration signal from machine surface, a new method, which is called Local-Wave Method (LWM), is presented to decompose it into number of Intrinsic Mode Weighs (IMW). Then improved RBF network model is constructed and trained using IMW as inputs. Taking the diesel fault diagnosis as an example, the method, which is checked through theory and practice, provides a power means for condition monitoring and fault diagnosis for the diesel engine.

The fault diagnosis of hybrid systems is a challenging research topic at present. Model based fault diagnosis methods have been paid much attention in recent years, however, because of the complexity of hybrid systems, it is usually difficult to achieve a first principle model. To address this problem, this paper proposes a novel hybrid neural network, and based on it, a gray-box approach to fault detection of hybrid systems is presented, which combines some priori knowledge with neural networks instead of the common first principle model. Simulation results illustrate the effectiveness of the proposed approach.

In this paper, a new Eigencurves method to detect leaks in oil pipelines is presented based on enhanced independent component analysis (EICA) and wavelet transform. EICA is used to derive Eigencurves from appropriately reduced principal component analysis (PCA) space of the training pressure images set. Wavelet transform de-noising (WTDN) is employed to preprocess measured pressure signals before getting the training and test images. In order to detect leaks, a classifier is designed to recognize negative pressure wave curve images by training set. The test results based on real data indicate that the method can detect many leak faults from a pressure curve,and reduce the ratio of false and missing alarm than conventional methods.

This paper presents a novel monitoring-oriented and diagnosis-oriented software architecture model of Internet-based remote monitoring and fault diagnosis system based on Unified Modeling Language. This system combines rule-based reasoning and data mining techniques with case-based reasoning for fault diagnosis. A novel knowledge-based fuzzy neural network has been proposed and implemented to mine data stored in monitoring database. The applied technologies can be used for other research domains.

Structural faults, such as unbalance, misalignment and looseness etc., are often occurring in a shaft of rotating machinery. These faults may cause serious machine accidents and bring great production losses. In order to detect faults and distinguish fault type at an early stage, this paper proposes a diagnosis method by using ”Partially-linearized Neural Network (PNN)” by which the type of structural faults can be automatically distinguished on the basis of the probability distributions of symptom parameters. The symptom parameters are non-dimensional parameters which reflect the characteristics of time signal measured for diagnosis of rotating machinery. The knowledge for the PNN learning can be acquired by using the Rough Sets of the symptom parameters. The practical examples of diagnosis for rotating machinery are shown to verify the efficiency of the method.

A practical approach is proposed to power system transient stability assessment by monitoring and controlling active power flow on critical transmission lines. A radial basis function network is trained to estimate transient stability limit of active power flow on a critical line. Once the transient stability limit is violated, a preventive control decision is made in terms of generation rescheduling. Control amount is determined using first-order sensitivities of transient stability margin with respect to generator outputs, which can be derived directly from partial derivatives of the trained radial basis function network’s output to inputs. Simulation results of a real-world power system demonstrate the effectiveness of the proposed approach.

This paper presents an optimized shunt hybrid power quality conditioner (SHPQC) for the compensation of harmonics and reactive power in power distribution network. A novel signal processing technique based on the adaptive neural network algorithm is applied to determine harmonics for generating the reference signals for the current controller of the SHPQC. The conventional hysteresis current controller for the three-phase insulated gate bipolar transistor (IGBT) voltage source inverter (VSI) is replaced by a current regulated space vector pulse width modulated controller. Simulations and experimental results are presented verifying the excellent power quality improvement performance of the proposed topology while keeping the apparent power of the SHPQC minimal.

This paper proposes a novel cyclic statistics based artificial neural network for early fault diagnosis of rolling element bearing, via which the real time domain signals obtained from a test rig are preprocessed by cyclic statistics to perform monitoring fault diagnosis. Three kinds of familiar faults are intentionally introduced in order to investigate typical rolling element bearing faults. The testing results are presented and discussed with examples of real time data collected from the test rig.

The objective of this paper is to model the abnormal mould friction in continuous casting, using back propagation (BP) neural network. It is also discussed two capturing algorithms of abnormal characteristics, ample-change judgement and modified pass-zero-ratio judgement algorithms, which are designed to capture the abnormal modes, sharp pulse and big ramp. A set of software for mould friction abnormity has been developed, and the results of simulating prediction accord elementarily with the abnormal records in steel plant. Compared to the traditional technique, the BP neural network in combination with ramp and pulse judgements is much more convenient and direct, and can achieve a much better prediction effect.

It is well known that the time-varying parametric model based on wavelet neural network has excellent performance on modeling a signal. The spark discharge signal is a typical nonstationary one. The spark discharge has two types in a static dust catcher. one is normal discharge, another is anti-electric-corona discharge. A few simulations indicate that the performance of the TVAR model on modeling a discharge signal is fineness, especially, on distinguishing between normal discharge and anti-electric-corona discharge.

Fuzzy neural networks fault diagnosis technology and diagnosis mode are used to diagnose cracks. It is trained with promoted BP arithmetic. The faults of cracked cantilever plate are diagnosed. Firstly the mode and frequency of numerical simulation intact plate and different cracked plates are calculated. Then five crack diagnosis indexes are calculated. Divide five indexes into three groups and create three fuzzy neural networks. The fuzzy neural networks are trained using these indexes, and diagnosis is taken to the crack in the end.

A blind diagnosis method of photovoltaic radar digital and analog integrated circuit based on the CPN neural networks is presented. By measuring the temperature and voltage of circuit component, the membership functional assignment of two sensors to circuit component is calculated, and the fusion membership functional assignment is gained by using the CPN neural networks, then according to the fusion data, the fault component is found. Comparing the diagnosis results based on separate original data with the ones based on CPN fused data, it is shown that the blind diagnosis method is more accurate.

In this paper, a cipher based-on chaotic neural network is proposed, which is used to encrypt JPEG2000 encoded images. During the image encoding process, some sensitive bitstreams are selected from different subbands, bit-planes or encoding-passes, and then are completely encrypted. The algorithm has high security with low cost; it can keep the original file format and compression ratio unchanged, and can support direct operations such as image browsing and bit-rate control. These properties make the cipher very suitable for such real-time encryption applications as image transmission, web imaging, mobile and wireless multimedia communication.

A combined hash and encryption scheme by chaotic neural network is proposed. With random chaotic sequences, the weights of neural network are distributed and the permutation matrix P is generated. The nonlinear and parallel computing properties of neural network are utilized to process hash and encryption in a combined mode. In-depth analyses of the performance indicate that the scheme is efficient, practicable and reliable, with high potential to be generalized.

Recently, a novel encryption algorithm that integrates Haar wavelets transformation into chaotic signal generator, was proposed by R.Luo et al. In this paper, we first analyzed the merits and demerits of this algorithm. Then an improved scheme which uses a clipped neural network is proposed. Both theoretical analysis and computer simulations show our proposed scheme succeeds in overcoming the defects of Luo’s algorithm while retaining all its merits. Moreover, the way that the clipped neural network evolves may present a new idea to the cryptography.

A new blind watermarking scheme by combining neural network with chaotic map is proposed. Using a chaotic sequence, the binary highpass watermark is generated. An encrypted watermark is to map to a selected pixel is modified according to the bit will be embedded. Embedding and extraction of watermark are based on the relationship between the pixel and its neighborhood in each block of image. A neural network and an adaptive embedding algorithm are adopted to enhance the characters of the watermarking system. Experimental results show that this watermarking scheme is very robust to common image processing and the extracted watermarks are readily recognizable.

This paper presents a color image watermarking scheme based on neural networks. A binary image watermark is embedded into the spatial domain of the host image. A fixed binary sequence is added to the head of the payload image watermark as the samples to train the neural networks. Each bit of the watermark is embedded in multiple positions. Because of the good adaptive and learning abilities, the neural networks can nearly exactly extract the payload watermark. Experimental results show good performance of the proposed scheme resisting common signal processing and geometric attacks.

Intrusion detection is an important technique in the defense-in-depth network security framework and a hot topic in computer security in recent years. In this paper, a new intrusion detection method based on Principle Component Analysis (PCA) with low overhead and high efficiency is presented. System call data and command sequences data are used as information sources to validate the proposed method. The frequencies of individual system calls in a trace and individual commands in a data block are computed and then data column vectors which represent the traces and blocks of the data are formed as data input. PCA is applied to reduce the high dimensional data vectors and distance between a vector and its projection onto the subspace reduced is used for anomaly detection. Experimental results show that the proposed method is promising in terms of detection accuracy, computational expense and implementation for real-time intrusion detection.

Digital watermarks have been proposed in recent literature as a means for copyright protection of multimedia data. In the absence of standardization and specific requirements imposed on watermarking procedures, anyone can claim ownership of any watermarked images. In order to protect against these counterfeiting techniques, we examine the properties that are necessary for resolving ownership via invisible watermarking. A method for watermarking of chaos and neural network is proposed in this paper. The watermark is embedded in the wavelet descriptors. Watermarks generated by this nonlinear technique can be successfully detected even after rotation, translation, scaling. And watermarks of our scheme are good at protecting from many kind watermark attacks. The experimental results demonstrate that the watermark is useful and practical.

In this paper, we investigate the application of Quantum neural network (QNN) to the document image watermarking problem. The method first divides the document into some weight-invariant partitions in the spatial domain. For better performance, we then reduce the watermarking problem to a classification problem, and use the Quantum neural network to solve it. QNN, characterized by the principles of quantum computing including concepts of qubits, superposition and entanglement of states, is a relatively new type of neural networks. Owning to the power of Quantum search, QNN is considered to have at least the same computational power as classical networks. We test the performance of QNN and the experimental results indicate the soundness of our method.

Detection of unknown malicious executables is one of most important tasks of Computer Immune System (CIS) studies. By using non-self detection, anomaly detection based on thickness, diversity of anti-body (Ab) and artificial neural networks, this paper proposes an NN-based malicious executables detection algorithm. This algorithm includes three parts, i.e., detector generation, anomaly information extraction and classification. At last, a number of experiments illustrate that this algorithm has high detection rate with very low the false positive rate.

Steganographic messages can be embedded into digital image while most methods in use today are invisible to an observer’s senses. Mathematical analysis may reveal statistical anomalies in the stego images. Since the difference of adjacent DCT coefficients can better reflect the statistical difference between the cover image and the stego image, we select six differences of adjacent DCT coefficients as the feature vectors used by the classifier for training. Experimental results show that our approach based on SVM can determine the existence of hidden messages in JPEG images reliably and get excellent computational efficiency.

The communication via electronic mail is one of the most popular services of Internet. The volume of emails that we get is constantly growing. In particular, unsolicited messages or spam, flood our email boxes, and result in causing frustration, and wasting bandwidth and time. The paper presents a novel schema to automatically filter spam emails by using the principal component analysis(PCA) and the Self Organized Feature Map (SOFM). In our schema, each email is represented by a series of textual and non-textual features. To reduce the number of textual features, PCA is used to select the most relevant features. Finally the output of the PCA and the non-textual features should be inputted into a well-trained SOFM to classify (spam or normal). In comparison with some traditional classification methods, the experimental result denotes that the scheme will increase the accuracy of filtering emails.

Self-organizing maps (SOMs) have proven extremely useful because of their ability to provide a condensed representation of data. This is accomplished by creating a mapping from an often continuous data space to a discrete map grid, relationships on which ideally preserve much of the structure of the original data. When well-trained, Kohonen’s original SOM successfully preserves continuity in the mapping from the SOM grid to the data space. When the dimension of a map approximately matches the dimension of the data, or when the folding of a map into higher-dimensional data space is controlled, mapping algorithms can preserve the organization of the data more comprehensively. In these cases, the structure of the map grid may reflect the data structure at several levels of granularity, allowing the search for the best-matching unit (BMU) of the trained map to speed up significantly.

As the activation function of neural networks is nonlinear, an adaptive model and an algorithm are suggested for real-time calculation of the network weights using an Unscented Kalman Filter (UFK). Because the method can simplify the calculation of nonlinear systems, a global optimization of data fusion information can be achieved by real-time adjusting the weights of the networks according to the variations of the system inputs. An experiment was made on a vessel board equipped with DGPS/GPS/ROLANDC/COMPASS to testify the availability of the algorithm. Three algorithms are applied to the system: the 1st with fixed weights of the network, the 2nd adjusting the weights of the system with linear activation function by KF, and the 3rd adjusting the weights of the system with nonlinear activation function by UKF. Experimental results show that the proposed algorithm can improve the tracking accuracy of the system. This method can be applied to integrated navigation and other data fusion systems.

In order for the map to visualize the structure of input data more naturally, the positions of neurons of the SOM should be adjusted based on their similarities. We study these position-adjustable SOM algorithms using Himberg’s contraction model, and then improve them into a uniform PASOM algorithm, which is easier to implement and control. In addition, by making use of the SOM’s topological ordering, the side effect of randomly selected initial weight vectors and the excess contraction to one point can also be avoided mostly. Finally, the PASOM algorithm is verified by experimental results very well.

This paper discusses an important problem emerged in applying GPS to wetland study–transformation method between coordinates system of local electrical maps and global coordinates system of GPS data. Traditional method is complex and not precise enough for large area. In this paper, a feed-forward neural network is build and the back propagation algorithm is used. The Levenberg-Marquardt algorithm and a modified error function are adopted to improve convergence rate and generalization ability. Neural network model, learning principle, simulation curves and results comparation are discussed in detail. The study is undertaken on Songhua River Basin. Practice results prove that compared with traditional method, neural network method succeeds in solveing complex problem and is effective in dynamic data analysis of wetland.

Disk arrays are widely used in Storage Area Networks (SANs) to achieve mass storage capacity and high level I/O parallelism. Data partitioning and distribution among the disks is a promising approach to minimize the file access time and balance the I/O workload. But disk I/O parallelism by itself does not guarantee the optimal performance of an application. The disk access rates fluctuate with time because of access pattern variations, which leads to a workload imbalance. The user access pattern prediction is of great importance to dynamic data reorganization between hot and cool disks. Data migration occurs according to current and future disk allocation states and access frequencies. The objective of this paper is to develop a neural network based disk allocation trend prediction method and optimize the disks’ file capacity to their balanced level. A Levenberg-Marquardt neural network was adopted to predict the disk access frequencies with the I/O track. History. Data reorganization on disk arrays was optimized to provide a good workload balance. The simulation results proved that the proposed method performs well.

We integrate probability distribution into pure competitive analysis to improve the performance measure of competitive analysis, since input sequences of the leasing problem have simple structure and favorably statistical property. Let input structures be the characteristic of geometric distribution, and we obtain optimal on-line algorithms and their competitive ratios. Moreover, the introducing of interest rate would diminish the uncertainty involved in the process of decision making and put off the optimal purchasing date.

Regression analysis is an essential tools in most research fields such as signal processing, economic forecasting etc. In this paper, an regression algorithm using probabilistic wavelet network is proposed. As in most neural network (NN) regression methods, the proposed method can model nonlinear functions. Unlike other NN approaches, the proposed method is much robust to noisy data and thus over-fitting may not occur easily. This is because the use of wavelet representation in the hidden nodes and the probabilistic inference on the value of weights such that the assumption of smooth curve can be encoded implicitly. Experimental results show that the proposed network have higher modeling and prediction power than other common NN regression methods.

For the reason that the conventional illumination models are empirical and non-linear, the traditional shape from shading (SFS) methods with conventional illumination models are always divergent in the process of iteration and are difficult to initialize the parameters of the illumination model. To overcome these disadvantages, a new approach based on the neural network for indirect SFS is proposed in this paper. The new proposed approach applies a series of standard sphere pictures, in which the gradients of the sphere can be calculated, to train a neural network model. Then, the gradients of the reconstructed object pictures, which are taken in the similar circumstances as that of the standard sphere pictures, can be obtained from the network model. Finally, the height of the surface points can be calculated. The results show that the new proposed method is effective, accurate and convergent.

This paper presents an approach to predict dimensional errors in end milling by using a hybrid radial basis function (RBF) neural network. First, the results of end milling experiments are discussed and the effects of the cutting parameters on dimensional errors of machined surfaces are analyzed. The results showed the dimensional errors are affected by the spindle speed, the feed rate, the radial and axial depths of cut. Then, a hybrid RBF neural network is applied. This neural network combines regression tree and an RBF neural network to rapidly determine the center values and its number, and the radial values of the radial basis function. Finally, the prediction models of dimensional errors are established by using the RBF neural network, the ANFIS (adaptive-network-based fuzzy inference system), and the hybrid RBF neural network for end milling. Compared with the predicted results of the above three models, the performance of the hybrid RBF neural network-based method is shown to be the best.

In this paper we apply Bayesian neural networks to life modeling and prediction with real data from a dynamically tuned gyroscopes (DTG). The Bayesian approach provides consistent way to inference by integrating the evidence from data with prior knowledge from the problem. Bayesian neural networks can overcome the main difficulty in controlling the model’s complexity of modeling building of standard neural network. And the Bayesian approach offers efficient tools to avoid overfitting even with very complex models, and facilitates estimation of the confidence intervals of the results. In this paper, we review the Bayesian methods for neural networks and present results of case study in life modeling and prediction of DTG.

Dynamic modeling for the quality of large-scale process is studied in this paper combined with continuous annealing process. This kind of process is constituted with several sub-processes. There is complex nonlinear mapping between the sub-process set points and the final quality. The quality model should be constructed and updated based on the new data from the real process. To meet this demand, a novel generalized growing and pruning RBF (GGAP-RBF) network is used to establish the quality model. GGAP-RBF is a sequential learning algorithm so that we can establish the model dynamically. Last, we do some on-line application study on the continuous annealing furnace in a steel factory. The quality model between the furnace temperature of each zone in the furnace and the exit strip temperature is constructed.

It is crucial to calculate the Signorini contact problem at high speed in real-time simulation of the belt grinding process. Finite Element Method (FEM) is the traditional way to solve such a contact problem. However, FEM is too time-consuming to fulfill the real-time simulation requirement. This paper demonstrates a new approach to model the Signorini contact problem based on learning. This new model averts doing the expensive optimization problem for each contact by FEM; hence dramatically reduces the calculating time. Multi-Layers Neurons (MLN) and Support Vector Regression (SVR) are adopted as a learning machine. The testing errors of these two learning machines are also compared. SVR showed superior performance than MLN in this application.

The General Regression Neural Network (GRNN) is briefly introduced. The BIC method for determining the order of Auto Regression (AR) model is employed to select the number of input neurons, and the Genetic Algorithm is applied to calculate the optimal smoothing parameter. The GRNN is used to predict the vibration time series of a large turbo-compressor, and its performance is compared with that of Radial Basis Function Neural Network (RBFNN), Back Propagation Neural Network (BPNN), and AR. It is indicated that the GRNN is more appropriate for the prediction of time series than the others, and is qualified even with sparse sample data.

Starting from an equivalent presentation of projection to latent structures (PLS), a novel nonlinear PLS approach is presented where both nonlinear latent structures and nonlinear reconstruction are obtained straightforwardly through two consecutive steps. First, an radial basis functions (RBF) network is utilized to extract the latent structures through linear algebra methods without the need of nonlinear optimization. This is followed by two feed-forward networks (FFN) to reconstruct both the original predictor variables and response variables. The proposed algorithm exhibits fast convergence speed and its efficiency is assessed through both mathematical example and modelling of a pH neutralization process.

This paper presents a new method of using ANN (Artificial Neural Networks) to solve the inverse problem: predicting reinforced concrete (RC) beams flexural load rating (ratio of flexural load to ultimate bearing capacity) given apparent damage parameter (crack width, crack height, deflection). A hybrid algorithm (G-Prop) that combines a genetic algorithm (GA) and BP (back-propagation) is used to train ANN with a single hidden layer. The GA selects the initial weights and changes the number of neurons in the hidden layer through the application of specific genetic operators. The case study shows that ANN based on GA and BP is a powerful instrument for predicting flexural load rating and further for evaluating RC beams safety.

This paper presents an two weighted neural network approach to determine the delay time for a heating, ventilating and air-conditioning (HVAC) plan to respond to control actions. The two weighted neural network is a fully connected four-layer network. An acceleration technique was used to improve the General Delta Rule for the learning process. Experimental data for heating and cooling modes were used with both the two weighted neural network and a traditional mathematical method to determine the delay time. The results show that two weighted neural networks can be used effectively determining the delay time for AVAC systems.

Evolutionary artificial neural network is applied for the prediction of the slope stability from the geotechnical material properties and the slope geometries. Coupling the genetic algorithm with artificial neural network, an effective forecast procedure is presented to analyze slope stability. In order to deal with the local minimal problem of artificial neural network with Back-Propagation rule, the connection weights of the artificial neural network are changed by using the genetic algorithm during the iteration process. The practical application demonstrates that the forecast of slope stability using artificial neural network is feasible and a well trained artificial neural network reveals an extremely fast convergence, a better generalization and a high degree of accuracy in the intelligent forecast for the slope stability.

When the performance function cannot be expressed exactly, response surface method is often adopted for its clear thought and simple programming. The traditional method fits response surface with quadratic polynomials, and the accuracy can not be kept well, which only the area near checking point coincides well with the real limit state surface. In this paper, a new method based on global response surface of BP neural network is presented. In the present method, all the sample points for training network come from global area, and the real limit state surface can be fitted well in global area. Moreover, the examples and comparison are provided to show that the present method is much better than the traditional one, the amount of calculation of finite element analysis is reduced quite a lot, and the accuracy is increased.

The large temperature drift caused by variation of environmental temperature is the main factor affecting the performance of fiber optical gyroscope (FOG). Based the advantages of artificial neural network and the fact that the temperature drift of FOG is a group of multi-variable non-line time series related with temperature, this paper presents modeling temperature drift of fiber optical gyro rate by improved back propagation (BP) training algorithm and by Gauss-Newton training algorithm, comparison between the modeling results of by improved BP algorithm and by gauss-newton algorithm is presented. Modeling results from measured temperature drift data of FOG shows that Gauss-Newton algorithm has higher training precision and shorter convergence time than improved BP algorithm on the same training conditions for application of modeling temperature drift of FOG.

Aiming at improving automatic level of traditional material ratio control system, we propose a novel chaotic neural network (NCNN) for automatic material ratio control system. The NCNN controller has following properties: (1) separation of superimposed ratio patterns, (2) learning unknown ratio patterns successively. As for the first feature, we utilize the feature of chaotic neural network proposed by K.Aihara [1]. As for the first property, when a stored pattern is given to the network, the network searches around the input pattern by chaos, and it recalls the ratio patterns from superimposed patterns. As for the second one, when an unknown input pattern is given, a different response will be received. So it can distinguish unknown patterns from the known patterns and learn the unknown patterns successively. A series of computer simulations demonstrate the effectiveness and stability of the proposed method.

According to the property that global stiff matrix is positive definite after being adjusted and specific formation of elastomer potential energy function, linear saturated system model (LSSM) is introduced into finite element neurocomputing. Based on the neural network, a circuit implementation of an example is given and the time, error characteristic and simulation of the circuit are analysed.

The crack density and crack growth rate are important parameters, which are used to describe the fatigue damage and predict the fatigue life of a material. There are many researches on the quantitative description of the fatigue cracks density and the crack growth rate, and several models are proposed, but these models cannot be widely used. In this paper, the BP network is used to describe the evolution of the fatigue crack density and the crack growth rate. It can be seen that the proposed method is feasible. The proposed method does not need to determine the interface between the long and short crack, and overcome the shortcoming of traditional models in which physical background of the parameters are uncertain, so it is difficult to determine in engineering.

Selective laser sintering (SLS), one of rapid prototyping technologies, employs laser beam to selectively fuse fully powder into a solid object layer by layer. However, density prediction of SLS parts using finite elements analysis (FEA) having been reported, heavily depends on the precision of the FEA model. An Artificial neural network (ANN) approach presented in this paper has been developed for density prediction of SLS parts. Two-layer supervised neural networks are used, and the inputs to the neural network are known SLS process parameters such as laser power, scan speed, scan spacing and layer thickness. Orthogonal experimental method is employed for collection of experimental training and test sets. The construction of network is also investigated. Comparison of predicted and experimental data has confirmed the accuracy of the ANN approach.

An application of neural networks to tire optimization designs is presented to alleviate the stress concentration of toe opening. As well known, it is either uncertain or time-consuming to obtain the global optimum solution by using classical local search methods when objective function of optimization is both nonconvex and implicit. In addition, it is infeasible to use local search method based on iteration to optimize tire mechanical property because analysis of tire mechanical responses is involved with material nonlinearity, geometry nonlinearity and boundary nonlinearity. In this paper, a GRNN is constructed to optimize the stress of toe opening by looking at an optimum Young’s modulus and cord direction of tire body rubber-cord composite material layer.

In order to estimate the bath endpoint phosphorus [P] content and manganese [Mn] content for Basic Oxygen Furnace (BOF) steelmaking process, three BOF endpoint estimation models are given according to the fast speed case, the slow speed case and the middle speed case of analyzing the sublance in-blowing sample. The first one is modeled from metallurgic mechanism, the second is with RBF NN and the last one is modeled using least square method. With theses models, steelmaker can get the estimation of BOF endpoint [P] and [Mn] based on the process information and the sublance measurement. The industrial experiment shows that these models are helpful and powerful.

This paper investigates the regular wave impact force on open-piled wharf deck by Artificial Neural Network. A three-layered neural network is employed and the units of input layer are wave period, T, incident wave height, H, and relative clearance, s/H. The unit of output layer is the maximum wave impact force, F_max. It is shown that the neural network with parameters determined through self-learning can predict wave impact force reasonably.

For aluminum powder nitrogen atomizing process, it is important but difficult to establish the aluminum powder granularity distribution using real-time measurements of exiting online sensors. In this paper, a novel software sensor model based on RBF Neural Networks is presented to estimate the granularity distribution of aluminum powder by means of measurements of melted aluminum level and temperature, atomizing nitrogen temperature and pressure, and recycle nitrogen temperature and pressure combined with granularity statistics distribution of the aluminum powder. The software sensor model can be obtained by training the RBFNN offline or online iteratively. An error analysis is carried out to illustrate effectiveness of the proposed software sensor model of aluminum powder granularity distribution.

Applying data analysis-based intellectual optimization approach to the industrial process optimization may improve the weakness of the mechanical model-based optimization method which is mostly based on the incomplete knowledge and understanding of a system. Therefore, it shows great application prospects. In this paper, an operating parameter optimization model of crude oil distillation process is developed using BP neural network. Then genetic algorithm is applied to search for the optimal technological parameters. Simulation results indicate that this approach is effective and practical.

For the production of high quality chemicals batch manufacture plays an important role. Since some parameters of a batch chemical reaction are difficult to measure with conventional instruments, Artificial Neural Network (ANN) model has become a useful tool for the estimation of chemical reaction indexes. Data restructuring is an effective method for improving properties and accuracy of the ANN model. This paper constructed an ANN model for estimating the number-average molecular weight and polydispersity in batch polymerization process and presented the method of training data restructuring. The results were satisfactory.

Industrial production of lycopene and β-carotene from tomatoes appears to be in highly demand in recent years by food industry and pharmaceutical companies for the development of functional foods. In this paper, a novel neuro-fuzzy model is first proposed to model the supercritical CO₂ extraction of lycopene and β-carotene from tomato paste waste by considering several process parameters. It can help to achieve the optimal control, and to develop a simulation and auto-control system for quality assurance in the industrial production. The effectiveness of the proposed model is demonstrated by simulation study.

A fundamental task of radar is the detection of targets in noise and clutter. It is very difficult to detect weak targets in heavy sea clutter. In this paper we apply the discrete wavelet transform (WT), along with multilayer feedforward neural networks, to the detection of radar weak targets. The wavelet transform employs the Haar scaling function, which is well matched to the signals from targets. The hard threshold filter is adopted to remove sea clutter. The neural networks are trained with the back-propagation rule, which are used to detect weak targets. The simulated results show that the proposed method is very effective for radar detection of weak targets.

An artificial olfactory system is constructed to determine the individual gas concentrations of gas mixture (CO and H2) with high accuracy. And a Back-Propagation (BP) neural network algorism has been designed using MATLAB neural network toolbox. And an effective study to enhance the parameters of the neural network, including pre-processing techniques and early stopping method is presented in this paper. It is showed that the method of BP artificial neural improves the selectivity and sensitivity of semiconductor gas sensor, and is valuable to engineering application.

This paper presents a new approach for designing form elements of a product based on consumers’ perception of images of the product. Neural network (NN) models are used to suggest the best combination of product forms for matching a given set of product images. An experimental study on the form design of mobile phones is conducted to evaluate the performance of four NN models, which are developed by linking 9 form elements with 3 product images of mobile phones. The evaluation result shows that NN models can be used to help product designers determine the best combination of form elements for matching a set of desirable product images.

This paper presents an intelligent simulation method for regulation of container yard operation on container terminal. This method includes the functions of system status evaluation, operation rule and stack height regulation, and operation scheduling. In order to realize optimal operation regulation, a control architecture based on fuzzy artificial neural network is established. The regulation process includes two phases: prediction phase forecasts coming container quantity; inference phase makes decision on operation rule and stack height. The operation scheduling is a fuzzy multi-objective programming problem with operation criteria such as minimum ship waiting time and operation time. The algorithm combining genetic algorithm with simulation is developed. A case study is presented to verify the validity and usefulness of the method in simulation environment.

This paper proposes a novel freeway traffic incident detection algorithm. Two stages are involved. First, get the freeway traffic flow model based on BP neural networks and use the model to obtain the output prediction. The residual signals will be gotten from the comparison between the actual and prediction states. Second, a SOM neural networks is trained to classify characteristics contained in the residuals. Hence, based on the classification given by the SOM neural networks, traffic incidents can be detected. Both theory analysis and simulation research show that this algorithm is effective.

A method is described for using Radial Basis Function (RBF) network to predict the chaos time series. The structure of RBF network is introduced first, then the two-step training procedure for the network is proposed, that is the unsupervised learning in first layer using K-means clustering algorithm and the supervised learning in second layer using gradient-based methods. Results obtained for a chaos logistic map time series analysis is presented to test the effectiveness of the proposed method. This approach is shown to improve the overall reliability of chaos time prediction.

Identification of underground nuclear explosion events and natural earthquake events is always worth to study because of the great military importance. The earthquake signal is non-stationary and FMmlet atom can delineate this signal more subtly, so we apply FMmlet Transform to extract the features of the nuclear explosion and natural earthquake signals, and then these features are recognized by the BP neural network. Experimental results indicate that the feature extraction and classification methods are effective and get good recognition results.

Internet traffic prediction plays a fundamental role in network design, management, control, and optimization. The self-similar and non-linear nature of network traffic makes highly accurate prediction difficult. In this paper, a boosting-based framework is proposed for self-similar and non-linear traffic prediction by considering it as a classical regression problem. The framework is based on Ada-Boost on the whole. It adopts Principle Component Analysis as an optional step to take advantage of self-similar nature of traffic while avoiding the disadvantage of self-similarity. Feed-forward neural network is used as the basic regressor to capture the non-linear relationship within the traffic. Experimental results on real network traffic validate the effectiveness of the proposed framework.

In Intelligent Transportation Systems (ITS), traffic flow forecasting is important to Traffic Flow Guidance System (TFGS). However most of the traffic flow forecasting models cannot meet the requirement of TFGS. This paper presents a traffic flow forecasting model based on BP neural network according to the correlation theory. This model greatly reduces the size of input patterns. Meanwhile, a new parallel training algorithm based on training set decomposition is presented. This algorithm greatly reduces the communication cost. Experiment results show that the new algorithm converges faster than traditional one, and can meet practical requirement.

This paper presents a combination of ELMAN and SOM neural networks in order to enhance the prediction precision. A new method of training and predicting of samples is developed. In this method, the training and predicting are divided into two steps, clustering at first and then training and predicting the samples at the clustered areas. As examples, this method is applied to weather broadcasting and disaster prediction. Simulation results show that this method can enhance the ability of local generalization of the network. The prediction precision of combined network presented in this paper is higher than that with normal BP network or just one of ELMAN or SOM network.

In this paper expanded Bayesian network method for short-term traffic flow forecasting in case of incomplete data is proposed. Expanded Bayesian network model is constructed to describe the causal relationship among traffic flows, and then the joint probability distribution between the cause and effect nodes with dimension reduced by Principal Component Analysis (PCA) is approximated through Gaussian Mixture Model (GMM). The parameters of the GMM are learned through Competitive EM algorithm. Experiments show that the expanded Bayesian network method is appropriate and effective for short-term traffic flow forecasting with incomplete data.

A highway traffic flow model using a distributed radial basis function (RBF) neural network based on fuzzy c-means (FCM) clustering is presented. FCM clustering is used to classify training data into a couple of clusters, each cluster is trained by a sub-RBF neural network, and membership values are used for combining several RBF outputs to obtain the final results. A highway traffic flow model with four segments is studied. The training data for traffic flow modeling were generated using a well-known macroscopic traffic flow model at different densities and average velocities. The simulation result proves the effectiveness of this method.

Earthquake Prediction is one of the most difficult subjects in the world. It is difficult to simulate the non-linear relationship between the magnitude of earthquake and many complicated attributes arising the earthquake. In this paper, RBF neural network ensemble was employed to predict the magnitude of earthquake. Firstly, the earthquake examples were divided to several training sets based on Bagging algorithm. Then a component RBF neural network, which was optimized by Adaptive Genetic Algorithm, was trained from each of those training sets. The result was obtained by majority voting method, which combined the predictions of component neural networks. Experiments demonstrated that the prediction accuracy was increased through using RBF neural network ensemble.

A reliable correlation between rainfall-runoff enables the local authority to gain more amble time for formulation of appropriate decision making, issuance of an advanced flood forewarning, and execution of earlier evacuation measures. Since a variety of existing methods such as rainfall-runoff modeling or statistical techniques involve exogenous input and different assumptions, artificial neural networks have the potential to be a cost-effective solution, provided that their drawbacks can be overcome. Usual problems in the training with gradient algorithms are the slow convergence and easy entrapment in a local minimum. This paper presents a particle swarm optimization model for training perceptrons. It is applied to forecasting real-time runoffs in Siu Lek Yuen of Hong Kong with different lead times on the basis of the upstream gauging stations or at the specific station. It is demonstrated that the results are both more accurate and faster to attain, when compared with the benchmark backward propagation algorithm.

In the paper, a multi-objective programming of portfolio is proposed according to the assumption that total risk loss can be measured by the maximum of risk loss in all securities. After analyzing the risk preference of the investor and taking transaction cost function’s linear approximation, the multi-objective programming model is transformed into simple-objective linear programming model. Based on neural network, a differential dynamical system for solving linear programming is constructed, and optimal portfolio decision is obtained.

Portfolio optimization has been researched for several years in financial engineering. The problem can be described as maximization of the expected growth rate of a portfolio with relatively small variance of the portfolio growth rate. In this paper, novel method of multi-stage portfolio optimization with partial information is introduced in detail. Investors can redistribute wealth in multiple periods in order to maximize the returns and minimize the risks in financial market. The experiments with random data can prove the method efficient and proper.

Nonparametric approaches of option pricing have recently emerged as alternative approaches that complement traditional parametric approaches. In this paper, we propose a novel neural network learning algorithm for option-pricing, which is a nonparametric approach. The proposed method is devised to improve generalization and computing time. Experimental results are conducted for the KOSPI200 index daily call options and demonstrate a significant performance improvement to reduce test error compared to other existing techniques.

The study examines the effectiveness of two types of neural networks in predicting corporate financial distress in China. Back-propagation and LVQ neural networks are considered. The neural networks are compared against Logistic Regression, which is the most popular one among the traditional methods. The results show that the level of Type I and Type II errors varies greatly across techniques. The neural networks have low level of Type I error and high level of Type II error, while Logistic Regression has the reverse relationship. Since the cost of Type I is more expensive than that of Type II error in this field. We demonstrate that the performance of the neural networks tested is superior to Logistic Regression.

Conditional Value-at-Risk (CVaR) is a new approach for credit risk optimization in the field of finance engineering. This paper introduces the concept of α-CVaR for the case of multiple losses under the confidence level vector α. The problem of solving the minimal α-CVaR results in a multiobjective problem (MCVaR). In order to get Pareto efficient solutions of the (MCVaR), we introduce a single objective problem (SCVaR) and show that the optimal solutions of the (SCVaR) are Pareto efficient solutions of (MCVaR). We construct a nonlinear neural networks model with an approximate problem (SCVaR)′ of (SCVaR). We may get an approximate solution (SCVaR) by solving this nonlinear neural networks model.

A systematic approach for Digital Signal Processing (DSP) structure optimizations is introduced. The method is based on the Multirate Signal Flow Graph (MSFG) representation and transformations. It reduces the optimization of dsp algorithms/structures to a series of MSFG transformations that makes automation of dsp structure optimization possible. As an example, an efficient Digital Down Converter structure is derived through MSFG transformations. It is applicable to linear time-invariant or periodically time-varying systems.