Advances in Neural Network Research and Applications

It is an important issue to allocate appropriate resources to mobile calls for wireless cellular networks owe to scarce wireless spectrums. The call admission control (CAC) will maintain better performance metrics of mobile call such as call dropping probability (CDP) and call blocking probability (CBP) if the future utilization of wireless spectrums can be predicted and provided to the decision of CAC. Therefore, a prediction mechanism which can predict most information such as system utilization is proposed in this paper. The techniques of data mining and pattern matching which adopts gradient to fuzz time series data for representations of chain code are applied to mining a possible repetitive pattern. Our proposed prediction mechanism can provide prediction information in advance whether the repetitive time series pattern of information exists or not. Furthermore, an update of confident level will be conducted continuously for performing each prediction in the proposed scheme. Our proposed mechanism is developed and tested with four cases which can be regarded as using scenarios of wireless cellular network. The experimental results show that the proposed scheme can capture repetitive time series patterns and perform following predictions with these repetitive time series patterns. Besides, the required storage is less than traditional schemes and lower computation power is required for the proposed scheme.

Forecasting model of the rate of rock burst acoustic emission time series has been established by BP neural network, and it is combined with the catastrophe theory to determine whether the rock burst. And then the experimental data recorded are used for examining the model. The results show that the degree of prediction accuracy is high, and it proves that the prediction model of rock burst is feasible.

In the course of oilfield development, accurate measurement of water content in crude oil has always been playing important role in practicing development adjustment and enhancing the effects of stimulation operation, and moreover it determines the development perspectives of oilfield. After research of the method of measuring the rate of water content, and the non-linear mapping relation between the rate of water content and impact factors of crude oil, a model of predicting the rate of water content in crude oil about vertical well based on wavelet neural network is proposed. Results of the simulation in MATLAB indicated that the way of WNN (wavelet neural network) has better convergent rate, prediction precision, learning ability and generalization ability than the traditional BP neural network. The method of WNN can predict the water content in crude oil with high precision and it owns much more powerful theoretical guide and much better application effects. It will have a broad application in the future.

In order to improve prediction accuracy of urban heat island intensity, we chose 9 main influencing factors from 1981 to 2006 and predicted urban heat island intensity in Chuxiong city in 2006 with backpropagation neural network. The predicted value was 2.1815°C. Compared with its measured value, residual error was 0.1042, relative error was 4.5588model and GM(1,1) model. The result shows that backpropagation neural network is effective to predict urban heat island intensity.

Based on the relevance vector machine and wavelet theory, a new machine learning method–wavelet relevance vector is introduced. The method performs wavelet transform to decompose the original series into some filtered series, and a relevance vector machine whose kernel functions is wavelet function models each of them. Then, this method is used to predict fouling thermal resistance of Heat exchanger. Construction of wavelet relevance vector machine prediction model is presented. Simulations show that wavelet relevance vector machine requires dramatically fewer kernel functions and it can get high prediction precision, and work in the paper offers a new method for the research of heat exchanger fouling.

Accurate electricity price forecasting is critical to market participants in wholesale electricity markets. The problem becomes more complex because the acquired data series are non-linear and non-Gaussian. In this paper, Multi Layer Perceptrons (MLP) trained with minimizing error entropy (MEE) algorithm is utilized to forecast electricity price. Compared with the conventional MLP with mean square error (MSE) criterion, the proposed approach can achieve better performance in simulated examples.

Accurate forecasting of rainfall has been one of the most important issues in hydrological research. In this paper, a novel nonlinear regression ensemble model is proposed for rainfall forecasting. The model employs Least Square Support Vector Machine (LS-SVM) based on linear regression and nonlinear regression. Firstly, Projection Pursuit (PP) technology and Particle Swarm Optimization (PSO) algorithm are used to obtain the main factors of the rainfall, which optimize projection index from high dimensionality to a lower dimensional subspace. Secondly, using different linear regressions extract linear characteristics of the rainfall system, and using different Neural Network (NN) algorithms and different network architectures extract nonlinear characteristics of the rainfall system. Finally, LS-SVM regression is used for nonlinear ensemble model. This technique is implemented to forecast daily rainfall in Guangxi, China. Empirical results show that the prediction by using the LS-SVM ensemble model is generally better than those obtained using other models presented in this study in terms of the same evaluation measurements. The results suggest that our nonlinear ensemble model can be extended to meteorological applications in achieving greater forecasting accuracy and improving prediction quality.

This paper introduces GARCH-EVT-Copula model and applies it to study the portfolio risk of exchange rates. Multivariate Copulas including Gaussian Copula, t Copula and Clayton Copula were used to describe the structure and extend the analysis from bivariate to any n-dimension. We apply this methodology to study the returns of a portfolio of four major foreign currencies in China. Our results suggest that the optimal investment allocations are similar across different Copula and confidence levels and the optimal investment concentrates in the USD investment. Generally speaking, t Copula and Clayton Copula can better portray the correlation structure of multiple assets than Normal Copula.

Cerebellar model articulation controller neural network (CMAC NN) has many advantages, such as very fast learning, reasonable generalization capability and robust noise resistance. Thus, CMAC NNs are conventionally used in robot control. To solve financial time series forecasting, this paper presents an efficient CMAC NN scheme. The proposed CMAC NN transforms continuous values of input variables to discrete indexes by using a quantization operator. To enhance generalization ability, the CMAC NN employs high quantization resolution and a large generalization size. To perform many-to-few mappings, the CMAC NN uses an efficient and fast hashing code based on bitwise XOR operator. The proposed CMAC NN was used to Nikkei 225 closing cash indexes collected from Japanese stock market. The forecasting results of the proposed CMAC NN were compared with those of support vector regression (SVR), which is statistical/ machine learning algorithm. Experimental results indicate that the performance of the CMAC NN was better than SVR in the tested case. Therefore, the CMAC NN may be considered as an efficient tool for forecasting financial time series.

This paper present one forecasting method with

Exponential Weighted Moving Average(EWMA)

and

Support Vector Regression (SVR)

. The daily cash turnover of the banks is time-serial data, banks need to forecasting daily cash turnover for banking reserve. First, the time series is preprocessed with EWMA method. The EWMAs with different coefficients are selected for forecasting features. And then SVR is used in the transformed dataset with EWMA for forecasting. The experimental result shows that the EWMA can improve the forecasting accuracy, and the SVR is more effective than other method such as 1-NN and MLP. Statistical correlation of SVR between the forecasted and actual values is much higher than other method such as 1-NN and MLP.

This study attempts to use GreyART network and grey model to construct a financial distress prediction model. The inputs used to train the network are the historical data containing 17 different financial ratios of 22 healthy and 5 distressed Taiwan’s listed banks. With the help of the developed performance index, this study also proposes a growing extraction method for financial variables not only to further improve the classification ability in the training and testing phases, but also to use fewer extracted variables to build the financial distress prediction model. Simulation results show that the optimal condition is the one using four extracted variables as inputs and the vigilance threshold of 0.80. Under this condition, the proposed method generates only two clusters with corresponding classification hit rates of 96.30% and 95.24% for the training and testing results, respectively.

As new field of computer intelligence research, Artificial Immune System inspired by biological immune system, provides a strong paradigm for information processing and problem solving. Artificial immune network theory is an important theory of AIS and has already wildly applied in the fields of data clustering, data analysis and robot control. This research proposes firstly to apply artificial immune network theory into risk assessment of managerial field. According to the comparability of the e-commerce risk problem and the biological immune system, presents specific model construction process relating to case study and testifies the result with data from real test.

In the many influencing factors of load forecasting, some are related to each other, and some are independent, so it is not very necessary of all factors to the conclusions. The paper adaptively selects input features making use of Bayes method and Rough Set, and chose a group of input characteristic set which report most out the reason of output change to train the BP Neural Network model for forecasting, which has been testified to be valid.

Normal fusion procedure of neural networks and fuzzy systems is to use neural learning techniques to train membership functions of fuzzy system. If no mechanistic prior knowledge can be used, fuzzy systems should be obtained from a set of data. The same data are usually used to train these fuzzy systems in the framework of fuzzy neural networks. But modeling accuracy cannot be improved extraordinary, because neural training and fuzzy modeling use the same data set. In this paper, we propose a new modeling idea, the fuzzy system will not be changed, and modeling error between real plant and the fuzzy system is compensated by a neural network. Another contribution of this paper is fuzzy model is generated automatically by kernel smoothing technique. The third contribution of this paper is a new learning approach for neural compensator is proposed, which assures stable and faster learning.

Based on deeply analyzing the characteristic of the battlefield ferromagnetic targets, according to the problems of magnetic detection system, for instance single detection pattern, low detection resolution and poor anti-interference performance, the Giant Magneto-Impedance(GMI) micro-magnetic sensor in combination with the technology of fuzzy neural networks(FNN) were carried as the core of the magnetic detection system. Take advantage of GMI sensor and FNN to realize accurate recognition of the target in the range of nan-otesla magnetic field. In this paper, equable magnetization rotation ellipsoid is used to simulate the tank and military truck, taking the triaxial magnetic moments and semi-focal length, that is

M

x

,

M

y

,

M

z

,

c

, as recognition characte-ristic quantity , and the FNN is used to recognize the tank and military truck including the categories and motion directions. The method reaches good recognition effect through experimental verification, and it has significance to improve detection range and recognition accuracy.

Inspired by the idea of combining multiple models to improve prediction accuracy and robustness, a soft sensing modeling of flotation process based on multiple T-S fuzzy neural networks and fuzzy c-means clustering algorithm (FCM) is proposed. Firstly, the model adopts principal component analysis (PCA) to reduce dimensions of the input variables data composed of texture characteristics of floatation froth image and process variables. FCM algorithm is used for separating a whole training data set into several clusters with different centers and each subset is trained by T-S FNN. The degrees of membership are used for combining several models to obtain the finial soft sensing result. Simulation results show that the proposed modeling is effective in the prediction of indexes and meets the requirement for optimization computation for the flotation process.

Using reliability evaluation for diesel engine is a complex, dynamic and uncertain process. In order to make an objective and right evaluation, and offer a stronger decision-making tool for designer and user of diesel engine, the neural-network-driven fuzzy reasoning mechanism of using reliability evaluation was developed based on the detail analysis of engine using reliability in the case that there is no sufficient quantitative information or the information is fuzzy and imprecise, where a feedforward neural network was used to replace fuzzy evaluation in the fuzzy system. Applications show that the evaluation result can be used as references for the improvement of reliability and maintainability of engines, and for the establishment of maintenance strategy.

This paper is concentrated on control system design for an XY parallel micromanipulator with piezoelectric actuation. The decoupled property of the manipulator enables the employment of single-input-single-output controllers for the two working axes. To compensate for nonlinear hysteresis effect stemming from piezoelectric actuator, the dynamics with Bouc-Wen hysteresis model is derived and identified for the system. Afterwards, a real-time fuzzy sliding mode control with perturbation estimation (FSMCPE) along with fuzzy switching control is proposed to attenuate the chattering phenomenon. Experimental studies reveal that the designed FSMCPE is superior to conventional SMCPE in terms of both positioning and tracking performances of the micromanipulator system, which validates the effectiveness of the presented control system design as well.

This study integrates the characteristics of credit risk rating and artificial intelligence technology into a credit risk rating model based on fuzzy neural network. The combination of fuzzy theory and neural network provides a good foundation for credit risk rating, making this model with fewer parameters, faster learning and less training samples. This study confirms that fuzzy neural network is an effective method for credit risk rating. The results of this study can solve the shortcomings in existing credit risk rating model and provide more information for decision-making reference.

In this paper, we introduce the concept of interval-valued fuzzy control. Based on the idea of the interpolation mechanism of fuzzy control, we propose the inference algorithm of interval-valued fuzzy inference and mathematical model of interval-valued fuzzy control, investigate the interpolation mechanism of interval-valued fuzzy control. Finally, we use a simulation experiment of interval-valued fuzzy control to illustrate our proposed algorithm reasonable.

The use of multiple intelligent agents in the decision support systems has become common. However, there are many questions in this field, such as the consensus of the agents and the method on choosing the best decision. In order to solve these problems, this paper provides a method, which uses the fuzzy aggregation operator to aggregate the opinions of the agents, ranks the alterative with the fuzzy preference relations, and then chooses the best alternative. This approach can be widely used in many fields.

Compared with related studies of stochastic demand inventory routing problem (IRP), this paper takes the constraints of soft time windows into consideration. Specially, penalties for violating time window limits are included in value function; upper limit of time window is dependent on inventory level. These assumptions of IRP are more practical. Firstly, the paper represents stochastic demand inventory routing problem with soft time windows (IRPSTW) as a discrete time Markov decision process model. Then the paper proposes a solution with considering constraints of soft time windows. Finally the application of this solution is illustrated with a numerical example.

The research aims at profit maximization, and firstly, according to the purchase methods for one-to-many purchase situation, the research installs purchase agent system model of virtual organization to solve the problems of purchase possibly faced by the virtual organization on supply chains. The research uses the results of value allocation model of virtual organization researched by AHP to install the module of value distribution agent system. In the last step, the research uses the joint procurement cycle and quantity estimation model made by back-propagation neural network algorism to install the model of enterprise demand estimation agent system and the model joint procurement agent. The goal aims at construction of the optimal solutions and methods to the negotiation between buyers, the joint procurement between the raw material suppliers, the item purchase assignment of jointly the goods with various shoppers, distribution of the purchase goods and so on.

An interactive hybrid evolutionary computation (IHC) process for MEMS design synthesis is described, which uses both human expertise and local performance improvement to augment the performance of an evolutionary process. The human expertise identifies good design patterns, and local optimization fine-tunes these designs so that they reach their potential at early stages of the evolutionary process. At the same time, the feedback on local optimal designs confirms and refines the human assessment. The advantages of the IHC process are demonstrated with micromachined resonator test cases.

To deal with the changing traffic in SONET/WDM tree networks, reconfigurable grooming of non-deterministic dynamic traffic (REGNT) is introduced. The full-fit case of REGNT is studied according to different network demand. Then two heuristic algorithms are developed to tackle the case by combining genetic algorithms with traffic-splitting heuristics. To evaluate the algorithms, the theoretical bounds are also derived. Computer simulations show that the proposed algorithms can achieve good results in reducing the number of additional ADMs and wavelengths to fit new traffic.

In multi-sensor multi-target tracking systems, the arrivals of “out-of-sequence” measurement (OOSM) can occur even in the absence of communication time delays. A optimal OOSM update algorithm are derived to solve one-lag as well as multi-lag OOSM update problems. In order to extend the OOSM update algorithms to multi-sensor multi-target tracking in clutter, the probabilistic data association (PDA) have been incorporated into the OOSM update algorithms with economic storage and efficient computation based on the nonsingularity assumption of some special matrices. The simulation results shows that PDA with the OOSM update algorithms have compatible RMS errors to the in-sequence PDA filter.

This paper proposes a Hopfield neural network guided evolutionary algorithm for aircraft penetration path planning. The combination of the algorithms benefits from the advantages of each and is intended to achieve a fast and adaptive path searching mechanism. The HNN works as guidance for the EA path planner by restricting the searching area around the gradient of the states of the network. Meanwhile the complex penetration factors are well integrated into the algorithm by EA so making the combined algorithm applicable to penetration path planning. Extensive simulations are conducted to demonstrate the effectiveness of the propose algorithm.

Based on credibility theory, this paper presents a class of two-stage fuzzy programming with value-at-risk (VaR) to deal with material procurement planning (MPP) problem. Since the MPP problem usually includes continuous fuzzy variable parameters with infinite supports, it is inherently an infinite-dimensional optimization problem that can rarely be solved directly. To overcome this difficultly, this paper introduces an approximation approach (AA) which can turn the infinite-dimensional optimization problem into a finite-dimensional optimization one. Furthermore, in order to solve the proposed MPP model, a hybrid algorithm is designed which combines AA, neural network (NN) and particle swarm optimization (PSO). Additionally, one numerical example is also presented to illustrate the effectiveness of the designed algorithm.

In the semiconductor manufacturing process, the endpoint of plasma etch process can be determined by the graphics based detection in order to avoid the loss of over-etching and under-etching. Our approach in current study can be conducted as one way to real-time monitor and judge the endpoint instead of observing it manually. When the endpoint occurs, this system can improve the etch processes and provide instant shutdown recommendations. This method makes use of Radial Basis Function (RBF) network’s functional approximation in time-series modeling and in pattern classification. By training with enough samples, the judge will be more accurate. All the samples are probed with optical emission spectroscopy (OES) sensor in real plasma etch process and for both network training and test.

A Novel Cellular Neural Network (CNN) entitled the shortest path CNN (SP-CNN) is proposed in this paper. Compared with general CNN, it is distinguished in the network structure and neural dynamics. As a result of these distinctions, SP-CNN has a good performance in motion planning for mobile robots. By mapping environment information to parameters in this neural network, motion planning can be transformed to the state evolvement of SP-CNN and the generated state represents information of the optimal path. The proposed method generates the best solution in static environments in real time. Extensive simulations about the above mentioned aspects demonstrate the effectiveness of the proposed approach.

Based on the basic theory of ERP (Enterprise Resource Planning), this paper summarized a series of indicators which reflected the effect of the ERP system. Then we used BP neural network and ant colony algorithm together to evaluate the effect of the ERP system. After we get the scores of the evaluation of the subsystems, we can find which subsystem is the weakest, so we can improve that subsystem and make the whole system better. The empirical results show that the ant colony network not only has the extensive mapping capability as the neural network, but also has the high-speed and global convergence features as the ant colony algorithm. When we use this model, we can get better accuracy and efficiency.

Satellite-based synthetic aperture radar (SAR) is a powerful tool for ship detection because it can work in all weather conditions, day and night. However, speckles and heterogeneous regions in SAR images pose great challenges on automatic detection of ships. This paper introduces a bottom-up visual attention model and proposes a visual attention based method for ship detection in SAR images. The proposed method is very simple and fast in computation, and has powerful capability of targets detection. The analysis of the detection performance over both simulated and real images confirms the robustness of the proposed algorithm.

Recognizing multi-warship in infrared image is very important to missile imaging guidance, an algorithm is suggested in this paper to recognize one ship among the other ones in infrared images. Four silhouette features are defined to describe infrared ships. A rule based on BP neural network with these four features is set up to distinguish one ship from the other ones. A complete algorithm is presented. This algorithm was simulated in computer with twenty-five images. The results showed this algorithm is valid when the sides of the ships face the camera.

An image feature extraction method, generalized 2-dimensional clustering-based discriminant analysis (G2DCDA), is proposed. This method can compress images along both rows and columns, and thus overcoming the drawbacks of too large feature matrices of 2DCDA. In addition, similar to 2DCDA, G2DCDA not only has the computational advantages of the 2D subspace methods available, but can deal with the multimodal distribution problems. Experiments conducted on the moving and stationary target acquisition and recognition (MSTAR) public database demonstrate that G2DCDA is more efficient than some 2D existing subspace methods, such as 2D principal component analysis (2DPCA), 2D linear discriminant analysis (2DLDA) and 2DCDA. Moreover, G2DCDA achieves higher recognition rates and lesser memory requirements.

A novel double-level parallelized firing pulse coupled neural networks (DLPFPCNN) model is presented in this paper, which is used for the segmentation of remote sensing image with water area as low contrast, low signal-to-noise ratio(SNR), and uniform slowly varying grayscale values of object or background. Its theory and work process is detailedly introduced as well, base on which the novel DLPFPCNN model is used to segment remote sensing image containing bridges above water. By a series of sequential processing combining with the priori knowledge of the bridge itself, such as linear feature et al., the target is finally recognized. Experimental results show that the proposed method has a good application effect.

Intelligent robot human-machine interactive technology will be incorporated into our daily lives and industrial production. This paper presents an autonomous mobile robot control system for human-machine interaction, The use of computer vision, 3D reconstruction of the two-dimensional image information of a specific coordinate point to point system. The use of a known point to the appearance characteristics of objects, by vision recognition algorithm, the original color image data for target screening and recognition. Allows users to easily through simple body movements, issuing commands to the robot. And by support vector machine(SVG) to classify non-linear non-separable type of data, accept user input and recognition actions to improve the robot vision system for target identification accuracy, and thus to achieve the goal of human-computer interaction.

With the development of the remote-sensing imaging technology, there are more and more applications of hyperspectral image classification tasks, in which to select a minimal and effective subset from a mass of bands is the key issue. This paper put forward a novel band selection strategy based on conditional mutual information between adjacent bands and branch and bound algorithm for the high correlation between the bands. In addition, genetic algorithm and support vector machine are employed to search for the best band combination. Experimental results on two benchmark data set have shown that this approach is competitive and robust.

The problem of automatic detecting impulsive sounds such as human sound (screams, shout), gun shots, machine gun, thunder, fire alarm, and car horn are useful for hearing impairment person. In this paper, instead of filtering the frequency of each sound for identifying types of sound, the frequency of sound is transformed into a recognizable image. The transformation is based on audio spectrogram (Power Spectrum) and Mel-frequency cepstral coefficients (MFCC). The images of both power spectrum and Mel-frequency spectrum are used as the inputs for an artificial neural network to recognize the corresponding sound. The proposed technique is tested with six different types of sound, i.e. machine gun, human scream, gun shot, thunder, fire alarm, and car horn from a sound database containing more than one hour of six different impulsive sounds. The experimental results on impulsive sounds detection using a spectrogram with feed-forward neuron network can effectively detect the segments of impulsive sound region in audio signal with more than 94% accuracy.

This paper proposes a method to protect the communication band through machine learning in cognitive networks. A machine learning cognitive radio (MLCR) extracts features from the signal waveforms received from various radios. A machine learning radio user (MLRU) assigns the states, i.e., unauthorized/authorized, and the associated actions, i.e., interfering/no interfering, to each waveform. The MLCR learns through a proposed hierarchical neural network to classify the signal states based on their features. The {signal, action} pairs are stored in the knowledge base and can be retrieved by MLCR automatically based on its prediction of the signal state related to the presented signal waveform. A case study of protecting the band of a legacy radio using our proposed method is provided to validate the effectiveness of this work.

There are many complicated and non-linear spatial patterns on land use and cover change at different scales. It is very difficult to express these geographic phenomena at a known scale. The aim of this paper is to propose self-organizing map to identify spatial pattern of land use and cover change at different scales. The procedure and steps of identifying spatial pattern of LUCC based on SOM are discussed in detail. An example of application research is experimented in Echeng District through remote sensing imagines in 1992 and 2002. The results suggest that SOM is very distinct and visual, which can avoid too many clusters to misty spatial pattern or be short of clusters so as not to express its change diversity.

The efficient feature subset selection for predictive and accurate classification is highly desirable in bioinformatic datasets. This paper proposes a method to apply our previously proposed neural network to microarray classification problem. The adjustable linguistic features are embedded in the network structure. After the training process, the informative features are selected. The network performs classification task either by the direct calculation or by rule-based approach. The structure of the three-layer feedforward neural network is designed with the consideration of useful information during the training process. The hidden layer is embedded with the linguistic feature tuning and mechanism for rule extraction. The colon tumor microarray dataset is used in the experiments. Good results from both direct calculation and from logical rules are achieved using the 10-fold cross validation. The results demonstrate the importance of the linguistic features selected by the network. The results show that the proposed method achieves better classification performance than the other previously proposed methods.

Speech denoising for two microphones is discussed using signal’s sparsity in this paper. A novel speech denoising algorithm is proposed. The algorithm firstly divides the noisy speech into a disjoint part and a joint part, and then removals the joint part which is mainly noise. Second, it deletes the noise samples on the “cross” in the disjoint part. Third, the denoisy speech is smoothed by a moving filter. Finally, several speech experiments demonstrate it performance and practices.

Blind separation of convolutive speech mixtures in frequency domain has obvious advantages in term of convergence and computation, but suffers from permutation ambiguity. Motivated by the fact that speech signals have strong correlations across frequency, the paper presents an energy correlation method for solving permutation ambiguity after separation of instantaneous speech mixtures at each frequency bin. Extensive experiments with synthetic and recorded speech signals are carried out to compare the energy correlation method to amplitude correlation method, three different complex-valued independent component analysis (ICA) algorithms are compared as well. The results show that the proposed method achieves better performance than the amplitude correlation method, and the complex ICA algorithm based on negentropy maximization yields the best separation.

A blind broadband beamforming method is presented in this paper for speech enhancement in the reverberation environment. The broadband beamforming is carried out using frequency blind source separation to generate the signals needed by the multichannel adaptive noise cancellation, which lies in the formulation of the mean frame skewness approximation. Simulation results demonstrate its effectiveness.

Aiming at solving the blind discrete nonlinear system identification problem, a cyclostationarity-based blind identification method of nonlinear system is proposed. In order to turn the identification process with input into the process without input, the first-order statistical characteristics of cyclostationary input and the inverse nonlinear mapping of the Hammerstein-Wiener model are introduced. The paper describes the statistical characteristics of the input and the structure of Hammerstein-Wiener model, and then discusses the mechanism of blind identification algorithm. Simulation results demonstrate the effectiveness of this approach in solving a class of discrete nonlinear system blind identification.

Digital image correlation method using artificial neural networks for subpixel displacement measurement is described in this paper. The integer pixel accuracy displacement is calculated based on cross correlation between subimages from undeformed and deformed images by two-dimensional discrete Fourier transform. Subpixel accuracy is obtained by training ANNs. Computer-simulated images are then used to verify this method. Results indicate it can obtain similar accuracies compared with other subpixel algorithms, but the ANN approach has the advantage that it can obtain subpixel displacement faster without knowledge of the analytical form of correlation coefficient of the interested point and its neighbours. Then the effects of speckle size, noise of images are studied. An optimal speckle size for optimal accuracy and the performance of the noise robustness are obtained.

Trees are important objects in our living environment. Modeling of living tree in our environment is hard work in computer vision and pattern recognition, since trees are related to large shape diversity and geometry complexity. In this paper, we present a range image analysis based approach to model a 3D tree from a single range image data. Range image pixels are thought of as 3D discrete points. Points from leaves and points from branches are segmented based on a new metrics on the convergence of local directions. A region growing method is then adopted to classify points from different branches. Skeletons of main branches are then computed by clustering each branch segment into small bins. The shape patterns of visible branches are used to predict those of obscured branches. Experiments show that this approach is applicable to modeling living trees.

In the computer vision research field, the “Bag of Words” model is known as a popular method for image representation. The Information Bottleneck principle derived from the rate-distortion theory in basic information theory has been applied to many applications in machine learning. In this paper, we introduce a method which combines the two state-of-the-art techniques for image clustering. Images are firstly represented using the “Bag of Words” model, and in the process of clustering based on Information Bottleneck principle, we utilize the Bregman divergence algorithm which works like k-means to get the optimal clustering result. Through the experimental results, we present several points of improvement obtained by the proposed method.

Fuzzy c-means clustering (FCM) with spatial constraints (FCM-S) is an effective algorithm suitable for image segmentation. Its effectiveness contributes not only to the introduction of fuzziness for belongingness of each pixel but also to exploitation of spatial contextual information. The contextual information can raise its insensitivity to noise to some extent. Although the robustness of the FCM-S algorithm is better, the convergence speed of it is lower. In this paper, to overcome the problem that FCM-S algorithm is time consuming, a fast fuzzy c-means clustering algorithm with spatial constraints (FFCM-S) is proposed. To speed up FCM-S calculations, FFCM-S algorithm modified the degree of memberships. Experiments on the artificial and real-world datasets show that our proposed algorithm is more effective.

In this paper, an airborne hyper-spectral image, which has 218 bands within a range of spectral resolution from 427.2nm to 945.7nm, is used to classify the vegetation of Mountain Jou-Jou. However, redundant bands could not significantly increase the accuracy of vegetation classification, but increase the computation cost of pattern recognition. Thus, the dimension of the hyper-spectral image is reduced using Principle Component Analysis (PCA) to extract the useful information for vegetation classification. Finally, Support Vector Machines (SVM) is employed to classify the vegetation based on the extracted useful information. In order to illustrate the classification accuracy of the aforementioned procedure, we tested hyper-spectral images of Purdue’s Indian Pines test site with its ground truth data. SVM gives the classification accuracy reach up to 80%.

By analyzing the extraction capacity of the image edge feature based on the wavelet transform, and the abilities of the nonlinear processing, the self-adaptive learning and the pattern recognition based on the identification method of the image definition with the neural networks, the identification method of the image definition based on a W-N model is put forward using the human eyes’ focusing mechanism based on the neural networks. The wavelet component statistics obtained by the wavelet transform are taken as the inputs of the 5 layer BP neural network model. The model identifies the image definition applying the steepest descent method of the additional momentum in a variable step size to adjust the network weights. The W-N model is first trained by 75 images from a training set, and then is tested by 102 images from a testing set. The results show that it is a very effective identification method which can obtain a higher recognition rate.

Based on olfactory neural system and KIII model proposed by Walter J. Freeman, a simplified olfactory system model was constructed to generated texture images. The output of different nodes (mitral cells, M) in the model can generate texture images. But the difference among these texture images is not analyzed. In this paper, the features of generated texture images are extracted based on gray level co-occurrence matrices (GLCM). According to different features, the texture images are compared. Experimental results show that the texture images generated by different M nodes have different characteristics and these texture images are different.

With learning-based natural language processing (NLP) becoming the main-stream of NLP research, neural networks (NNs), which are powerful parallel distributed learning/processing machines, should attract more attention from both NN and NLP researchers and can play more important roles in many areas of NLP. This paper tries to reveal the true power of NNs for NLP applications as supervised or unsupervised learning devices by concretely introducing two practical applications: part of speech (POS) tagging and self-organizing documentary maps for high-precision, visual information retrieval.

Blind channel identification can be cast into a single-input-multi-output (SIMO) identification problem by oversampling and then solved easily by SIMO identification methods. Due to this, SIMO identification is of great interest and attracts a lot of attention in the past twenty years. Many efficient methods have been developed for this problem. However, most of them are sensitive to overestimation of channel order. Based on sparse representation, an efficient SIMO identification method is proposed in this paper. Differing from the Prediction Error Method, the new algorithm does not require the input signal to be independent and identical distribution, and even the input signal can be non-stationary. In addition, the new algorithm is more robust to the overestimation of channel order.

Human head modeling is the foremost research topic of the multi-modal fusion for medical information in brain function research based on EEG/fMRI. Finding an effective way to construct the geometric model of human head is one of the focal point of research. Here a new method based on NURBS surface reconstruction is proposed. In the experiment, surface interpolation is employed, and a fairly good result is got with comparatively less points. Thus NURBS is proved a rapid and effective method for head modeling.

Robust filter based on risk sensitive estimator is derived to estimate the state of the uncertain models, while the estimation error involves two terms, the first term coincides with the minimum value of the risk sensitive cost function, the second one is the distance between the true and design probability models. The proposed algorithm, which introduces risk sensitive estimator into the unscented particle filter, could automatically change the state noise covariance according to the magnitude of the risk function. As a result, sample impoverishment could be mitigated. In the simulation of submarine bearing and frequency tracking, the performance of the new algorithm is compared with the unscented kalman filter and the unscented particle filter. Simulation results show that the new algorithm performs better than the two others.

In this paper a novel fully complex multiplicative neural network (MNN) algorithm is proposed to extract Quadrature Amplitude Modulation (QAM) signals when passed through a non linear channel in the presence of noise. The inputs, weights, activation functions and the output of the proposed MNN are complex valued. The training algorithm for the multilayer feed forward fully complex MNN is derived. The equalizer is tested on 4, 16 and 32 QAM signals and compared with split complex feed forward MNN equalizer. The proposed equalizer is implemented on nonlinear and nonminimum phase stationary channel. The fast converging algorithm gives lower bit error rate performance even in the presence of substantial noise.

In order to assist different kinds of disabled people and senior citizens, and to improve the effectiveness and convenience of human-machine-environment interaction, a novel economical embedded multi-mode control system for intelligent wheelchair is developed which is based on a high-powered 16-bit single chip system. In this paper the real-time image processing and control strategies for the visual navigation mode of the embedded system are studied. A less data image array, an adaptive recognition algorithm, a fuzzy control strategy and fusion control strategies with other modes are addressed. Some verifying experiments are carried out and results are given.

A new

L

2

-gain disturbance rejection controller and adaptive adjustment are combined into a hybrid robust control scheme, which is proposed for robot tracking control systems. The proposed controller deals mainly with external disturbances and nonlinear uncertainty in motion control. A neural network (NN) is used to approximate the uncertainties in a robotic system. Meanwhile, the approximating error of the NN is attenuated to a prescribed level by the adaptive robust controller. The adaptive techniques of NN will improve robustness with respect to uncertainty of system, as a result, improving the dynamic performance of robot system. A simulation example demonstrates the effectiveness of the proposed control strategy.

We address the problem of tracking relative translation in a leader-follower spacecraft formation architecture using Robust Integral of the Sign Error (RISE) based Neural Network (NN) technique. Based on the relative translational dynamic model of the spacecraft formation, RISE is introduced to approximate the dynamics of the follower as well as various practical disturbances. It is shown that the errors of the entire formation closed-loop are asymptotical stability (AS), which takes significant advantage over the typical Uniformly Upper Bounded (UUB) property of most NN controllers in high-precision formation tasks. Finally, numerical simulation is provided to verify the effectiveness of the proposed algorithm.

The walking assistant robot (WAR) provides powder assist and navigation operation for the elderly and injured. The WAR need integrate different kinds of sensors to recognize people’s intents and execute different functions to help people, such as assist seating/standing up, assist walking, human tracking, location, obstacle detection, avoidance and so on. What’s more, WAR need to be flexible enough to integrate new equipments for different requests of the elderly and injured. In this paper, a Simplified Modular Petri Net (SMPN) is proposed to describe and control the distributed module behaviors of the walking assistant robot system. The simplified MPN has been applied to control SJTU’s WAR-Walkmate. The experiment results show that SMPN can control the Walkmate to integrate different modules and perform various tasks successfully.

This paper provides a novel WAR(Walking Assistant Robot) which only uses ODVS (Omni-Directional Vision System) as its main sensor to implement the task of Tracking and Guiding. We use Blob Detection and Camshift Algorithm to track, and use APF(Artificial Potential Fields) approach to guide. The experiments show that those methods have high precision and good robustness, and also show the advantages of the wide view field and low cost of the ODVS.

This paper presents a new adaptive controller for a robot manipulator by using the visual feedback from an eye-in-hand camera. The controller is designed to cope with the case when the target 3-D positions are unknown. The controller employs the depth-independent interaction matrix to map the image errors onto the joint inputs of the manipulator. A new algorithm is developed to estimate the unknown parameters on-line. The Lyapunov theory is used to prove asymptotic stability of the proposed controller based on the nonlinear dynamics of the robot manipulator. Experiments have been conducted to demonstrate the performance of the proposed controller.

Stacking speed is an important parameter to estimate the capacities of robotic system. In this paper, an innovative method based on ILC (Iterative Learning Control) could be used to optimize the speed of the industrial robots. By learning following error, the robotic system could reach the optimum speed within the permissible error range and achieve the optimization of the stacking robotic system. Furthermore, requiring little to model the robotic system, ILC is easy to be realized in practice. The results of the system simulation indicate, under the circumstances of computing robotic system parameter and comparing the optimization with ILC to the original results, that ILC could increase the stacking speed a lot through improving local speed.

To track passive target efficiently and accurately, an improved particle filter algorithm based on genetic algorithm (SGAPF)is proposed.By incorporating the newest observation into sampling process and using genetic algorithm, the degeneracy problem is overcome and the predication performance of particle filter is improved. The improved algorithm guarantees the diversity of the particles and particles are moved to the regions where they have larger values of posterior density function. Simulation experiments show the validity of the proposed algorithm.

This paper investigates how to design the limit of failure rate and the adjust number of robots in the distribution control of the multiple robots which may be broken through the simulation of space solar power satellite assembly. For this purpose, we conduct simulations with changing the failure rate of robots that employ our proposed deadlock avoidance method. Intensive simulations have revealed the following implications: (1) from the viewpoint of the completion rate, our deadlock avoidance method enables the robots to complete the assembly in 80completion rate even if the 1/3 robots are broken; (2) from the viewpoint of the recovery rate (

i.e.

, the rate of completing a task when some of robots are broken), the maximum failure rate which enables robots to complete the assembly in 80% is 0.2%,

i.e.

, the 1/3 robots can be broken; and (3) when the failure rate is 0.2%, it is possible to maximize the completion rate from 80% to 90% by adjust the number of the robots.

This paper presents a discrete-time decentralized control strategy for trajectory tracking of a five degrees of freedom (DOF) redundant robot. A recurrent high order neural network (RHONN) modified structure is used to identify the plant model and based on this model, a discrete-time control law is derived, which combines block control and sliding mode techniques. The neural network learning is performed on-line by Kalman filtering. The local controller for each joint use only local angular position and velocity measurements. The proposed control scheme is implemented in real-time.

Multiple Radial Basis Function (RBF) neural network models are used to approximate a kind of discrete time nonlinear system described by a combination of nonlinear and linear part. Based on these models, Multiple model adaptive control strategy is given to improve the transient response of closed-loop system. Simulations according to different uncertain changing parameters in the model of the system are studied to show the efficiency of the proposed method.

This paper presents a control system design strategy for multi-input and multi-output (MIMO) networked control systems with random delays. The performance index of the control systems is constructed by entropies of tracking error which can be estimated by Modha’s density estimator via multi-layer perceptron (MLP). Finally, a simulation example is given to illustrate the efficiency and feasibility of the proposed approach.

In this paper, the traditional linearized model of the flight dynamics of a rigid airplane, in which aerodynamic forces, moments and thrust coefficient tables play an important role, are transformed into an analytical affine form using a newly proposed neural networks structure. This turns feasible the on-line inversion of the nonlinear flight dynamics. A neural decoupling control law based on nonlinear inverse control techniques (Input-Output Linearization) can then be used to perform flight trajectory tracking, Very good tracking performance is shown compared to a theoretical (tables based) autopilot.

From theory based on adaptive observers, this paper presents a structure for black-box identification based on state-space recurrent neural networks for a class of dynamic nonlinear systems in discrete-time. The network catches the dynamics of the unknown plant and jointly identifies its parameters using only output measurements. The stability and the convergence of the training algorithm and the ultimate bound on the identification error as well as the parameter error are established in the Lyapunov sense. Numerical examples using simulated and experimental systems are included to demonstrate the effectiveness of the proposed method.

In this paper, the passivity problem is investigated for a class of stochastic neural networks with discrete time-varying delay and distributed time-varying delay as well as generalized activation functions. By constructing appropriate Lyapunov-Krasovskii functionals, and employing the free-weighting matrix method and stochastic analysis technique, a delay-dependent criterion for checking the passivity of the addressed neural networks is established in terms of linear matrix inequalities (LMIs), which can be checked numerically using the effective LMI toolbox in MATLAB. An example is given to show the effectiveness and less conservatism of the proposed criterion. It is noteworthy that the traditional assumptions on the differentiability of the time-varying delays and the boundedness of its derivative are removed.

In this contribution, the inverse problem of spatiotemporal complex dynamical system is investigated on the basis of a typical diffusively coupled map lattices (CML) model. A novel method of recovering initial conditions for CML system is proposed by utilizing symbolic dynamics. It is shown that convergent property goes hand in hand with coupling strength. Both theoretical and experimental results show that the proposed algorithm performs well in noiseless case when the coupling strength is not very strong. The presented results also provide a theoretical and factual basis for better analysis and description of the spatiotemporal complex dynamical behaviors of the actual system.

A control scheme was designed for nonlinear time-varying systems with time delay, based on time-delay compensating theory and fuzzy logic. It consists of a Smith neural-network predictor and a modified parameter-self-tuning fuzzy controller, given ACR (Acrylate Copolymer Resin) polymerizing-kettle as the control plant. In the experiment, we verified the performance of the control system in two scenarios: one with invariant parameters and the other with time-varying parameters. Experimental results illustrate the effectiveness of the proposed scheme. Moreover, the comparison to other three typical control methods is also presented, which demonstrates that the proposed control scheme has satisfying dynamic performances, and when the system parameters varied with time it can still control stably with good robustness.

In this paper, an adaptive wavelet neural network controller (AWNNC) is proposed to synchronize two nonlinear identical chaotic gyros. The proposed AWNNC system is composed of a neural controller and a compensation controller. The neural controller uses a wavelet neural network to online approximate an ideal controller, and the compensation controller is used to guarantee system stable based on Lyapunov function candidate. Some simulation results verify the chaotic behavior of two nonlinear identical chaotic gyros can be synchronized by the proposed AWNNC scheme.

In this paper, an intelligent transportation control system (ITCS) using wavelet neural network (WNN) is developed to increase the safety and efficiency in transportation process. The proposed control system is composed of a neural controller and compensation controller. The neural controller acts as the main tracking controller, which is designed via a WNN to mimic the merits of an ideal total sliding-mode control (TSMC) law. The learning algorithms are derived from the Lyapunov stability theorem, which are utilized to adjust the parameters of WNN on-line for further assuring system stability. Moreover, based on robust control technique, the compensation controller is developed to attenuate the effect of the approximation error, so that the desired attenuation level can be achieved. Finally, it is applied to control a marine transportation system. The simulation results demonstrate that the proposed control system can achieve favorable control performance than other control methods.

A video-based incident detection system for monitoring the urban road is investigated in this paper. The developed surveillance system can monitor not only vehicles and motorcycles on the road surface but also pedestrians on the walkway and the prohibitive zones in the image. Several different kinds of incident can be detected in the presented system, such as the congestion, the illegally parking, the lane-changing vehicle, the falling object, the pedestrian across the road, and the pedestrian appearing in the prohibitive zone. The proposed method is based on the background subtraction. Therefore, the background image and lane markings are estimated in the beginning. Then the foreground image is obtained from the difference between the current image and the background image. In order to have isolated objects, the objects in the foreground image are separated by lane marking. The occlusion caused by the vehicles and the pedestrians can also be handled by the temprol-spacial analysis for the exact detection. After that, the tracking is applied to track the targets. Finally the proposed system can detect the incident by the integrated information from the pre-defined information and the tracking. Several challenge video are performed to verify and the experimental results demonstrate that our system is satisfying and effective.

With the convergence of the Internet and the Global Positioning System (GPS), a common approach called GPS tracking system has been developed nowadays. A GPS tracking system is a system that uses GPS to determine the precise location of a vehicle, person, or other assets and the information can also be shared with remote clients through the Internet. For the device used in the outdoor, it usually connects to the Internet through mobile network like GPRS or 3G. However, the weaknesses of mobile network are low bandwidth and high charging strategies compared to Ethernet or even Wi-Fi. In this paper, an efficient web-based tracking system through reduction of redundant connections is proposed. By integrating several technologies nowadays we design a new connection scheme to improve the efficiency for the location-aware application. The proposed architecture is focus on two problems. One is the redundant connection, causing the waste of bandwidth and fees during the period when connecting to the mobile network. The other is the data asynchronization during the period of transmission. This paper not only provides a new connection scheme to solve above issues, but also has verified this architecture on an embedded platform with a prototype implementation for a web-based tracking system.

Blind spot warning (BSW) systems play an important role in advanced driving assistance systems (ADAS). Crashes are frequently happened when drivers change their host lanes without taking notice of the vehicles in the blind spot area. In this paper, a BSW system is developed and has been real road verified with CCD/CMOS cameras. The proposed algorithm is robust to adapt different weather conditions in day and night, including sunny, cloudy and rainy. It has been implemented on our self-designed DSP system with a 600M Hz core processor. It deserves to be mentioned that the bilateral BSW function is completed on one DSP system and it processes more than 20 frames per second in CIF image format. The average detection ratio achieves 95.09%. In the future, this system will be integrated with lane departure warning systems, previous vehicle warning systems and parking assistance systems to be the omni-directional ADAS.

A fuzzy logic controller (FLC) with feed-forward Compensation is proposed for autonomous parallel parking in this paper. Two ultrasonic radars are used to detect the parking space. The law of cosines is used to increase the accuracy of the corner detection of the parking space. Triangular function is applied to obtain the desired trajectory, orientation angle, and feed-forward steering angle. Trajectory estimation is used to obtain the lateral position error and orientation error with respect to the reference trajectory. These two errors are inputs to the FLC and the control output is the steering angle correction. Preliminary experimental results show that the proposed FLC with feed-forward compensation can achieve better trajectory following and reduce the length requirement of parking space.

OSGi is considered as one of the fundamental technologies underpinning telematics services, and has attracted both academic and commercial interest. Because of a desire and mandates to deliver telematics services on OSGi environments, professionals have identified a key issue that need to be addressed: the service delivery would suffer from the situation of low bandwidth and unstable links. In this work, we benefit from the inherent characteristics of mobile agents and develop OSGi-Based mobile agents to address this challenge. Furthermore, the mobile agents are enhanced with a Java bytecode extractor which makes a mobile agent partially reside in a contactless smart card for service invocations, and with a risk-enabled reputation model which supports a mobile agent for information filtering to reduce the size of data that need to be carried.

Intelligent control and guiding of urban traffic syste-m become so important nowadays especially under the current traffic status, which is more and more congested and complicated. In this paper, we proposed a method of modeling the urban traffic flow system combining the global and local model information for the whole city net. We consider that the traffic digraph is composed from several nodes and those nodes are linked with routes lines. The proposed system is inspired by the random walk theory: for each traffic flow in the urban network, we simulated it with a random walkprocessing, with vehicle flow density and driver strategy independent. These flows only shared traffic lights and affected each other in the congestion situation. Finally we get simulator solution only by seeking the stable solutions of random walk. This intelligent system is very powerful andonly the topology structure of city, the start and destination and numbers of each vehicles flow are known, it can return all of the optimized control strategy for each traffic light and driver in the traffic net. For evaluation, different road situations with various system parameters are simulated on the proposed system. The experiments results are satisfied and show the feasibility and robustness our system.

This paper presents a noval computational framework, BN+CoBE, Bayesian enhaced Cascades of Boosted Ensemble, for on-road vehicle detection. The objective of this research is to reduce false alarms while keeping the detection rate high. In the proposed system, BN+CoBE, the CoBE is trained on image texture features and Bayesian conditional probability function is trained on vehicle features of location, size and confidence values generated by all the stages in CoBE. Experiment results on real world data show that the proposed BN+CoBE system is effective in reducing false alarms significantly while keeping the detection rate high.

Vibration and noise reduction and control have obvious significance for submarines. A novel vibration analysis method based on an improved ICA algorithm is proposed in this paper. By using the clustering evaluation method, the stability and separating performance of the algorithm are significantly enhanced. The improved ICA algorithm is applied to feature extraction of the vibration signals and quantitative calculation of the source contributions of a scaled submarine model. The result shows that the proposed method is effective, and this research provides a primary basis for the vibration and noise control of submarines.

In this paper, we present a hierarchical salient-region based algorithm and apply it for automatic ship detection in remote sensing images. The novel framework breaks down the complex problem of scene analysis by hierarchical attention, in a computationally efficient manner, such that only the salient-regions which contain potential targets can be analyzed in detail. Firstly, a parallel method is adopted for crudely selecting saliency tiles from entire scene by using low-level feature extraction mechanisms, and then the Region-of-Interest (ROI) around each saliency object is taken out from the saliency tiles to pass to the further processing. Shape and texture features are extracted from the multiresource ROIs to describe more details for candidate targets respectively. Finally, Support Vector Machine (SVM) is applied for target validation. Experiments show the proposed algorithm achieves high probabilities of recall and correct detection, as well as the false alarms can be greatly diminished, with a reasonable time-consumption.

There are several stages of tool wear in turning process. We collect of the force signals and vibration signals at each stage. Using wavelet filtering and power spectrum methods, typical parameters changes are detected. We extract the signal feature for fuzzy clustering. Experimental results show that the tool wear monitoring is achieved in turning by using this pattern recognition method.

The aim of part-machine clustering (PMC) in cellular manufacturing systems is to cluster parts that have similar processing requirements into part-families; and machines that meet these requirements into machine-groups. Although PMC problems are known as NP-complete in the literature, extensive research is still conducted in this field because of the considerable practical value of PMC for industries. In this paper, conventional adaptive resonance theory (ART1) neural network method and a novel meta-heuristic approach called ant colony system (ACS) are proposed for solving PMC problems. The experimental results show that ACS performs better than ART1 neural network on the same selected benchmark test problems. A PMC performance measure called grouping efficiency (GE) is also employed to evaluate the clustering result.

As one of the most common parts of various rolling mechanical equipments, rolling element bearing is vulnerable. Therefore, great attentions have been attributed to the theories, Afailure diagnosis methods and their applications for rolling bearings. Vibration analysis is also a very important means for condition monitoring and fault diagnosis. This paper aims at the research on the methods of signal processing and pattern recognition. Vibration signals collected were analyzed by using methods of Time-delayed correlation demodulation and effective signal features were extracted. B-spline neurofuzzy networks were established to carry out the recognition of faults of bearings. Experimental results have proved that the developed error diagnostic architecture is reliable and effective.

Visible and near infrared (Vis/NIR) spectroscopy was investigated for the fast and nondestructive determination of viscosity of lubricating oil. A total of 150 oil samples were scanned, and different calibration models were developed with the pretreatment of smoothing and standard normal variate. The input variables of calibration were the principal component selected by principal component analysis (PCA) and characteristic wavelengths selected by successive projections algorithm (SPA). The calibration model were developed using partial least squares (PLS), multiple linear regression (MLR) and back propagation neural networks (BPNN). The results indicated that PCA-BPNN and SPA-BPNN models were better than the linear models (PCA-PLS, PCA-MLR, SPA-PLS and SPA-MLR). The correlation coefficients were 0.971 for PCA-BPNN and 0.964 for SPA-BPNN. This demonstrated that BPNN could make good use of the nonlinear information in spectral data, and SPA was a powerful way for the selection of characteristic wavelengths. The selected wavelengths were helpful for the development of portable lubricating oil viscosity detection instrument.

In this paper, a chattering-free adaptive wavelet neural network controller (CAWNNC) is proposed using the dynamic sliding-mode approach. The proposed CAWNNC system is composed of a neural controller and a switching compensator. The neural controller uses a wavelet neural network to online approximate an ideal controller, and the switching compensator is designed to eliminate the approximation error introduced by neural controller. Finally, the proposed CAWNNC system is implemented based on a field programmable gate array (FPGA) chip for low-cost and high-performance industrial applications and it is applied to control a brushless DC (BLDC) motor to show its effectiveness. The experimental results demonstrate the proposed CAWNNC scheme can achieve favorable control performance without occurring chattering phenomena.

After thorough analyses of the standard PSO, BPSO and the relative varied algorithms, this paper proposes a new natural learning mode, named MNLBPSO, which is based on BPSO. Using a relaxing restriction method, we establish an unconstrained multi-objective optimization model for the Interruptible load management problem, and apply the new algorithm, MNLBPSO, to solve this model. Two experiments are made with different load interruption demands and MNLBPSO can obtain better schemes with much less iteration time than BPSO. And the results of the experiments indicate that MNLBPSO algorithm is more powerful and effective than BPSO and has excellent properties, such as naturalness, simplicity and speediness.

A novel method is presented to determine the external dynamic forces applied on structures from measured structural responses in this paper. The method utilizes a new SVM-DE model that hybridized the differential evolution (DE) technique and support vector machines (SVM) to resolve the problem of force identification. Both numerical simulations and experimental study are conducted to demonstrate the effectiveness, robustness and applicability of the proposed method. It is promising that the proposed method is practical to the real-life application.

In wet ball mill, measurement accuracy of mill load (ML) is very important. It affects production capacity and energy efficiency. A soft sensor method is proposed to estimate the mill load in this paper. Vibration signal of mill shell in time domain is first transformed into power spectral density (PSD) using fast Fourier transform (FFT), such that the relative amplitudes of different frequencies contain mill load information directly. Feature variables at low, medium and high frequency bands are extracted through principal component analysis (PCA), which selects input as a preprocessing procedure to improve the modeling performance. Three support vector machine (SVM) models are built to predict the mill operating parameters. A case study shows that proposed soft sensor method has higher accuracy and better predictive performance than the other normal approaches.

This work discuss the fault diagnosis of tool wear based on HHT (Hilbert-Huang Transform) and SVM (Support Vector Machine). Firstly, we introduce a novel approach to extract features of sampling data based on HHT. Our emphasis is to analyze the Hilbert amplitude spectrum and the Hilbert marginal spectrum of each IMF (Intrinsic Mode Function), which contains the physically meaningful characteristics of the tool states. Secondly, the Hilbert spectrum and marginal spectrum are used as the features, which would be as the inputs of LS-SVM (Least Square-SVM) model for pattern recognition to monitor cutting processing. Based on SVM theory, this dissertation develops a research on machine fault pattern classification by using the extracted features for tool wear applications. The simulation results show that the HHT has good performance in extracting the features, and the proposed method can increase the correct rate of classifier.

Due to the uncertainty of the factors that influence the income and other characters of floating women in Jiangsu province, we propose using Bayesian Networks to model this kind of system. We use different algorithms for learning Bayesian Networks in order to compare several models. This study of a real problem includes preliminary data processing, the comparison of different algorithms, and the role of using Bayesian Networks for social problems. We suggest that researchers can use Bayesian Networks to explore the potential relationship between variables of complex social problems.

The groundwater level has continuously decreased due to the rapid increase of paddy field acreage in area of well irrigation paddy in Sanjiang Plain in recent years, some time happen such things as more and more “hanging pump” and local overdraft. The authors took 853 Farm as an example and established the dynamic prediction model of groundwater depth by using the multi-resolution function of wavelet analysis and nonlinear approximation ability of artificial neural network in order to solve above problems. The results of dynamic variation regularities analysis and precision inspection and comparison showed that the model had high accuracy in fitting and prediction. The prediction results also showed the groundwater level will descend continually in the future years and has an average annual downrange of about 0.66m. Therefore, the local government should reinforce the scientific groundwater management. This model revealed the dynamic variation regularities of regional groundwater and provided the scientific basis for sustainable utilization of groundwater resource in area of well irrigation paddy in 853 Farm and even entire Sanjiang Plain.

A Petri-Net modeling method of agent’s belief-desire-intention called BDIPN is proposed to describe managers’ belief. It is combined into Batch Deterministic and Stochastic Petri Nets to describe inventory systems, where the model is called BDI-BDSPN. Two types of logistics, Complex Logistics and Agile Logistics, are both modeled using BDI-BDSPN.

This paper compares two vastly different methods of analysis – multiple regression and neural networks, in supply chain flexibility assessment. Data of manufacturing firms evaluating their prominent suppliers were analysed by multiple regression and simulated using three-layer multilayer perceptron (MLP) neural networks. Our study shows that NN can accurately determine a supplier’s flexibility capability within an error of 1% The incorporation of these two methods can lead to better understanding and dynamic prediction of supply chain flexibility for buyers.

By integrating the PBL model with book resources in libraries, this study identifies the advantages of libraries in supporting e-Learning. This study supports problem-based learning (PBL) in library using a novel intelligent mobile location-aware book recommendation system (IMLBRS) with map-based guidance. Experimental results reveal the learning performance during PBL supported by the proposed IMLBRS for book searches is superior to Online Public Access Catalogue (OPAC) system. Furthermore, this study confirm that the proposed system facilitates better learning performance for learners with the field-dependent learning style than for learners with the field-independent learning style.

This paper presents least squares support vector regression with genetic algorithms (LS-SVRGA) models for the prediction of radio-wave path-loss in suburban environment. The least squares support vector regression (LS-SVR) model is a novel forecasting approach and has been successfully used to solve time series problems. However, the application of LS-SVR models in a radio-wave path-loss forecasting has not been widely investigated. This study aims at developing a LS-SVRGA model to forecast radio-wave path-loss data. Furthermore, in the LS-SVRGA model genetic algorithms is applied in order to select two parameters of LS-SVR models. In this study, four forecasting models, Egli, Walfisch and Bertoni (W&B), generalized regression neural networks (GRNN), and support vector regression with genetic algorithms (SVRGA) models are employed for forecasting the same data sets. Empirical results indicate that the LS-SVRGA outperforms others models in terms of forecasting accuracy. Thus, the LS-SVRGA model is an effective method for radio-wave path-loss forecasting in suburban environment.

This paper evaluates the impact of tobacco control policies on female and male smokers. The data used in this study are from the 2002-2006 International Tobacco Control Four Country Survey. Based on eleven smokers’ motivational attributes used in the survey, principle component analysis is used to identify tobacco control policy drivers which are labeled as personal concerns, cigarette price, environmental restrictions and social encouragement. To examine the relative impact degrees of these four policy drivers on the groups of female and male smokers for their quit attempts, probabilistic neural network models are developed using hypothetical policy impacted populations. The experimental result shows that the most significant motivator for female smokers to make a quit attempt is their personal concerns. For male smokers, social encouragement plays a dominant role for them to make a quit attempt. The result indicates that smoking restrictions in public places or at workplace can somewhat encourage them to make a quit attempt. However, increasing the cigarette price is less likely to affect the MQA rate of both female and male smokers.

This study is an attempt to provide a methodology for predicting the stability of building foundation pits basing on the employment of intelligent tools: neural network and fuzzy logic. For this purpose, a back-propagation network (BP network) model was established by training the prophase horizontal displacement data of a building foundation pit, which was situated in center of Shanghai city. The obtained BP network model was then adopted to predict the later stage displacements of the above-mentioned foundation pit. Following it, a fuzzy logic model was set up by employing Mamdani fuzzy algorithm and applied to the pre-control of foundation pit excavation. In the fuzzy logic model, the displacement values, which were predicted by BP neural network, were employed as input data. The obtained output result, namely building foundation pit pre-control measures, is proven to be very ideal through practical verification. The developed theory and methodology offers a new way for the pre-control of building foundation pit excavation and other underground construction.

The accurate inversion of model parameters is a major difficulty for marine water quality model. In this paper, a data-driven model (DDM) method based on back-propagation neural network (BPNN) is developed to inverse the value of model parameters and to find out the relationship between model parameters and the pollution concentration values of interior stations. All training data are calculated by numerical water quality model from results of multi-parameter matching design cases, so the physical properties are not disturbed. Filed data is imported into the relationship for inversing optimal parameter. Finally optimal prediction method is applied to validate the long term stability of inversion results. Case tests are carried out in Bohai Sea, China. Dissolved inorganic nitrogen (DIN) and its sensitive parameters are considered for validating the present method. Case studies show that the present method can make a more satisfactory inversion for a practical problem.

Visible and near infrared (Vis/NIR) spectroscopy combined with least squares-support vector machine (LS-SVM) was investigated to determine the sugar content of instant milk-teas. The Savitzky-Golay smoothing, standard normal variate and 1

st

derivative were applied as preprocessing methods. The PLS model was developed and the optimal latent variables (LVs) and effective wavelengths (EWs) were also selected. The LV-LS-SVM model with LVs outperformed PLS models. Wavelengths at 484, 515 and 957 nm were confirmed to be EWs and the EW-SL-SVM model achieved the best performance in all developed models. The correlation coefficient (

r

), root mean square error of prediction (RMSEP) and bias for validation set were 0.964, 0.087 and 0.005, respectively. The results indicated that the EWs combined with LS-SVM method was successfully implemented for the prediction of sugar content of instant milk-teas, and the confirmed EWs were helpful to develop commercial instrument to progress the quality evaluation of instant milk-teas.

In this study, we presented a novel summarization method for generating sports video abstracts, which utilized motion entropy analysis and mutual information. Both of them are based on an attentive model. In order to capture and detect significant segments among a video, we exploited saliency maps by calculating color contrast, intensity contrast, and orientation contrast of frames. In the next step, motion vectors between maps were computed and converted into salient motion entropy. Meanwhile, a new algorithm based on mutual information was proposed to improve the smoothness problem when we selected boundaries of segments. The experiments showed that our proposed algorithm could not only detect highlights effectively but also generate smooth playable clips. Compared with the traditional approaches, our system improved the precision by 7.6% and enhanced smoothness by 1.2, which also verified feasibility of our system.

A neural network (NN) based algorithm for retrieval of precipitable water vapor (PWV) from the Moderate Resolution Imaging Spectroradiometer (MODIS) radiance is proposed. A multilayer feedforward neural network (MFNN) is selected, in which the at-sensor brightness temperature, the surface emissivity of MODIS chs. 31 and 32, and the land surface temperature (LST) are input variables, and PWV is the output variable. The input parameters for the MFNN are mainly based on the radiative transfer simulation with MODTRAN 4.0 code and the latest global assimilation data. The algorithm is applied to retrieval of the PWV over northeast area in china using MODIS data. Compared with the MODIS PWV products, the RMSE of the PWV retrieved by our algorithm is 0.45g/cm

2

. Furthermore, a comparison of our retrieval PWVs with radiosonde data is carried out. The results show that the MFNN-based retrieval algorithm for PWV is robust and efficient.

It is commonly acknowledged that wind energy is the leading renewable energy generation method; currently producing a power yield equivalent to 35 GW, with an estimated projection of 40-60 GW by 2012. In order to successfully integrate wind energy with traditional generation supplies it is necessary to have the ability to accurately forecast the available yield of a wind installation during a period of time. In this paper we present a neural network based estimation tool which produces wind speed estimates for a given wind installation. These predications are subsequently used in industry standard calculations to produce an energy yield estimate for the wind installation over a given time period. The proposed approach produces forecasts that can be used for two main purposes; firstly, delivery of wind (energy) yield estimations and secondly to assess the suitability of a given location for development into a wind park site. The tool makes use of a Multi-layered Perceptron which has been trained with historical data to produce a set of predicted wind speed data for a given period. This data is then processed in conjunction with independent variables, including Turbine Generator type and altitude to give an estimated power yield and expected uncertainty of the forecast (in terms of percentage capacity factor). Our results indicate that by using such a neural network approach the accuracy of the tool is sufficiently accurate to for short to medium estimations and could prove a valuable tool for wind energy producers and utility operators.

Distribution system optimal planning has vital significance, but there isn’t efficient and practical algorithm at Traditional genetic algorithm has a poor expressive power for complicated problem because of the restriction of its norm mode, which limits the application fields of genetic algorithm. This paper adapts the idea of “Ethogenetics” reference, and presents a new type of genetic algorithm based on Agent behavior and paradigm learning. Unlike the based creating mode of feasible solution in traditional genetic algorithm, a feasible solution is created by ~ series o! ye behaviors of Agent based on knowledge in the new genetic algorithm. To adapt the new creating mode of feasible the traditional mechanism of evolution optimization based on Darwinism is abandoned and the mechanism of learning’ is adopted to realize the evolution optimization. At last, an example distribution network is optimized by Ilene tic algorithm and traditional genetic algorithm respectively. The comparative result proves the new genetic I has higher expressive power, computing efficiency, convergent stability and extendable capability.