Comprehensive Overview on Computational Intelligence Techniques for Machinery Condition Monitoring and Fault Diagnosis

ANN is a special case of neural computation, which is inspired by the human brain. This neural network is a mathematical model that can achieve distributed parallel information processing. ANN can adjust the interconnections among internal nodes to achieve information processing of a complex system.

Diagnostic inference can be interpreted as a solution of a problem based on the specific mapping relationship between fault symptoms and fault causes. For complex mechanical systems, the mapping relationship is generally nonlinear. Therefore, ANN has been widely used in fault diagnosis because it can effectively approximate various mapping relations. At present, most of fault classification methods utilize time-frequency analysis methods as the early feature extraction, and ANN or its optimized forms are then employed for fault classification. Fig. 2 shows the flowchart of fault diagnosis based on ANN. In Ref. [26], wavelet packet transform (WPT) and ANN were integrated to diagnose fault in internal combustion engine, in which WPT was used to extract the fault characteristics, and generalized recurrent neural network (RNN) was proposed to classify various fault conditions. Lei, et al [27], proposed an intelligent diagnosis method based on ensemble empirical mode decomposition (EEMD) and wavelet neural network. EEMD was used to extract the characteristics of time and frequency domains from the sensitive intrinsic mode functions (IMFs). Wavelet neural network was adopted to complete the pattern recognition. WPT and empirical mode decomposition (EMD) were utilized to preprocess and extract features, and ANN was used to diagnose early fault in rotating machinery [28]. Cui, et al [29], proposed a new backpropagation neural network (BPNN) based on the coefficient entropy of wavelet packet decomposition to realize quantitative diagnosis of fault severity trend of rolling bearings. Saravanan, et al [30], presented a new hybrid method based on discrete wavelet transform (DWT) and ANN to diagnose various faults of spur bevel gearbox. Zhao, et al [31], utilized BPNN and improved shuffled frog-leaping algorithm (SFLA) to perform fault classification. The accurate selection of suitable features that reflect the running status of equipment in practical application of fault diagnosis is the key point of research. Therefore, fault feature selection based on ANN is an important research direction.

Fault diagnosis of mechanical system based on ANN has some limitations. First, extraction and selection of features depend largely on the prior knowledge of signal-processing technique and diagnosis experience, and generalization is weak. Second, ANN adopts a shallow structure, which also limits ANN to learn complex nonlinear structures in fault diagnosis [32]. Deep neural network (DNN) is developed based on deep learning theory, which can enhance the accuracy of big data classification [33] and effectively overcome the preceding shortcomings. Deep learning was first introduced into the field of fault diagnosis by Tran, et al [34], who applied deep belief network (DBN) based on Teager energy operator to achieve fault diagnosis of reciprocating compressor valves. A multisensor health diagnosis method based on the DBN was presented in Ref. [35], which classified the sensor signals collected from a damaged structure. Guo, et al [36], developed a hierarchical adaptive deep convolutional neural network for bearing fault diagnosis. Jia, et al [32], used DNN for intelligence fault diagnosis in rotating machinery, especially in the case when the vibration data were massive. ELM has been extensively applied and popularized in the fault diagnosis of mechanical system in recent years. Yang, et al [37], proposed a multilayer ELM based on representational learning for fault diagnosis. The effectiveness of this method was successfully verified by applying a wind turbine system. Wei, et al [38], proposed a method based on local mean decomposition to identify the different fault types of gearbox, combining permutation entropy and ELM. More references on the applications of ELM in machine fault diagnostics were provided in Refs. [39‐42].

Fuzzy theory is the process of imitating the way that people logically think in dealing with fuzzy information, which is suitable for the qualitative analysis of complex large-scale systems. The relationship between fault and symptom is difficult to describe using an accurate mathematical model due to the complexity of engineering practice. Therefore, the application based on fuzzy logic theory in fault diagnostics is closer to human thinking habits and language expression. Fuzzy logic is an effective pattern recognition method, which has been successfully used in power [43], transmission line [44], transportation [45], and industrial production [46]. Fuzzy logic mainly imitates human’s logical thinking and thus has strong capability of expressing knowledge. ANN imitates the function of human brain neuron, which has the strong capability of self-learning and direct processing of data. Adaptive neuro-fuzzy inference system (ANFIS) comprises both merits of neural network and fuzzy logic. Zheng, et al [47], proposed a fault diagnosis method based on local characteristic-scale decomposition (LCD) and fuzzy entropy (FuzzyEn) for fault diagnosis of rolling bearings. A series of intrinsic scale components (ISCs) was first obtained using LCD, and the FuzzyEns of the first few ISCs that contained the main failure information were then extracted. Finally, the FuzzyEns were used as the input of the ANFIS, which achieved the accurate classification of bearing fault types. Zhang, et al [48], proposed an early fault diagnosis method based on multiscale entropy and ANFIS. Tran, et al [49], combined ANFIS with decision tree to achieve fault diagnosis. Result indicated that the proposed method could effectively diagnose the fault of induction motors. Wu, et al [50], combined DWT with ANFIS to identify gear faults. DWT was used to extract the energy spectrum feature vector, which was regarded as ANFIS input.

Neural computation is the process of imitating the physiological structure and information processing of the human brain, whereas EA is the process of imitating the biological evolution and group intelligence. EAs provide a new way to address complex optimization problems, which have the advantages of simple principle and convenient implementation, especially in the case of solving large-scale dynamic optimization problems.

The present review of EAs used for fault diagnosis can be broadly divided into two sections, as shown in Fig. 3.

Broadly, neural computation includes kernel methods, such as SVM. However, SVM does not belong to ANN [20], and it is a new machine-learning method based on statistical theory.

SVM solves the optimal classification hyperplane by using the structural risk minimization principle, overcomes the dimensionality disaster and local minimum problem, and has a small demand for the samples. Therefore, SVM is particularly suitable for establishing a fault diagnosis model. SVM, which was first introduced into the field of fault diagnosis by Jack and Nandi [71], was used to achieve rolling bearing fault classification. SVM has been widely used in mechanical system fault diagnosis in recent years. Konar, et al [72], used wavelet transform and SVM to detect bearing faults in induction motors. Li, et al [73], proposed a fault diagnosis method based on redundant second-generation wavelet transform to achieve fault diagnosis, which was combined with neighborhood rough set and SVM. Cheng, et al [74], proposed singular value decomposition and SVM based on EMD to perform fault diagnosis of rolling bearings and gear. In this method, feature vector matrix was the singular value of the sensitive IMF component decomposed by EMD, which was regarded as SVM input for intelligent fault diagnosis. Zhang, et al [75], proposed a hybrid model based on permutation entropy, EEMD, and SVM for motor bearing fault diagnosis. The vibration signal was initially decomposed into a set of IMF components by EEMD. Then, the permutation entropy feature vector of the first few IMFs was obtained, which was regarded as the input of the optimized SVM for achieving fault-type classification. Some improved SVM has been proposed, such as ensemble SVM (ESVM) and fuzzy SVM (FSVM), to solve the problem of multiple fault classification. The ensemble classifier not only solves the multi-fault classification problem but also significantly improves the classification performance compared with the single SVM [76, 77]. Zheng, et al [78], proposed composite multiscale FuzzyEn and ESVM for rolling bearing fault diagnosis. FSVM was also used to solve multi-classification problems [79]. Hang, et al [80], employed EEMD to extract fault feature vectors, and FSVM was adopted to solve multi-classification problems in fan fault diagnosis. Comparison of FSVM classification results with backpropagation and standard SVM indicated that FSVM had higher classification accuracy. More recent references of SVM in fault diagnosis can be found in Refs. [81‐88].

Although SVM has made some achievements in the research of machinery fault diagnosis, some issues need to be further studied, such as the following: (1) selection of the appropriate kernel function and its parameters; (2) selection of the appropriate multi-classification algorithm to meet the needs of a multi-fault classifier; (3) improvement of training speed to satisfy the real-time requirement of fault diagnosis; and (4) combination of other knowledge-based fault diagnosis methods, such as fuzzy logic and neural network, with SVM for fault diagnosis.

ANN implements RUL estimation and prediction of future health conditions by including direct or indirect observation data, independent from the failure process of a physical model. Various types of data can be used as the input of ANN, such as some process variables, monitoring data (vibration signals), evaluation characteristics (age and stop time), and some historical features. The output of ANN is RUL prediction or performance degradation assessment, which is used for conducting effective maintenance strategies. ANNs widely used in fault prediction include BPNN [91‐95], radial basis function network (RBFN), and RNN [96]. Ahmadzadeh, et al [94], proposed a three-layer feedforward BPNN for RUL estimation of grinding mill liners, which considered degeneration and condition monitoring data as the inputs of ANN, and used RUL as the output of ANN. Rodriguez, et al [95], presented ANN (six input layers, three hidden layers, and one output layer) to predict and simulate the behavior of life-cycle assessment in blades of steam turbines. In view of the shortcomings of traditional incremental training methods in long-term prediction, Malhi, et al [96], proposed an RNN based on competitive learning method to improve the accuracy in long-term prediction of rolling bearings. Mahamad, et al [97], used feedforward neural network and the Levenberg-Marquardt training algorithm to predict the RUL of rolling bearings. Considering the complexity and nonlinearity of the pitch system, and the difficulty of describing with precise mathematical model, Chen, et al [98], proposed ANFIS based on a prior knowledge, which was used to predict wind turbine pitch faults. Existing ANN methods predict the RUL by using failure history data, but suspended historical data are rarely utilized. Hong, et al [99], used a self-organizing map, which combined wavelet packet and EMD for feature extraction, to estimate bearing performance degradation. Javed, et al [100], used ELM and fuzzy clustering to predict the degradation state and the RUL of complex nonlinear systems. Compared with ANN, ELM improved the algorithm efficiency by randomly selecting hidden layer parameters. Zhang, et al [101], proposed a multi-objective DBN ensemble method for RUL estimation.

The main purpose of introducing fuzzy logic is to overcome uncertainty and inaccuracy, and fuzzy logic has obvious superiority in dealing with large time delay, time variation, and nonlinear processing. This method is based on fuzzy mathematics theory, including the appropriate membership and fuzzy rules; then, fuzzy inference is continued to implement fuzzy prediction. Baban, et al [102], acquired the vibration and temperature signals of the textile machine, evaluated the machine status-based fuzzy logic, and implemented preventive maintenance. Stetter, et al [103], used fuzzy logic to monitor the health status of the pump system in an engine. Ishibashi, et al [104], combined decision tree, fuzzy logic, and GA for RUL prediction of aeroengine, and acquired the historical data from the sensor on the aeroengine. Tian, et al [105], proposed a fuzzy-adaptive unscented Kalman filter to improve the prediction accuracy of nonlinear processes. Zio, et al [106], proposed a method based on fuzzy similarity analysis to estimate system RUL. This method is subjective to manually select membership function of fuzzy logic. Therefore, this method is generally used in conjunction with fault tree, expert system, and neural network.

Neural network and fuzzy logic are important methods in computational intelligence. The advantages of neural network are its parallel processing capability, strong fault-tolerant capability, self-learning, and self-adaptability. However, the neural network is similar to a black box, which lacks transparency. Expressive knowledge of weights in the network is not easy to understand, and ANN cannot utilize the language knowledge of experts. Fuzzy logic make the inference process understand easily, which can use expert knowledge directly, with lower requirements on the samples. Nevertheless, the disadvantages of fuzzy logic are slow inference speed and low precision. Adaptive learning is difficult to implement. Thus, the two above mentioned methods are integrated to form NFSs, which have the merits of both methods. NFSs have been widely used in machinery performance prediction in recent years. Zhang, et al [107], used neuro-fuzzy network to predict tool wear and RUL, considering that tool condition monitoring is critical to the manufacturing industry. Gokulachandran, et al [108], used neuro-fuzzy and support vector regression to evaluate RUL of cutting tools. The experimental results showed that the neuro-fuzzy method could obtain a more accurate prediction. Ali, et al [109], combined fuzzy neural network and Weibull distribution to predict RUL of rolling bearings. Chen, et al [110], employed ANFIS using a prior-knowledge-based method to predict wind turbine pitch fault. Zhao, et al [111], used neuro-fuzzy method to predict the health condition of bearings. Compared with the RBFN, the proposed method is superior with respect to reliability and robustness. Chen, et al [112], proposed a method based on NFSs and Bayesian algorithms to predict the health status of helicopter gearboxes and bearings. The results showed that the proposed algorithm was superior to RNN and NFSs in prediction accuracy. Ramasso, et al [113], proposed a method combining NFSs and belief function theory to evaluate the RUL of a turbofan engine. Wang, et al [114], proposed an evolutionary neuro-fuzzy (eNF) predictor on time-varying dynamic systems. A newly enhanced least squares estimator was used to train the linear parameters of the eNF predictor. Experiments showed that the proposed method could be successfully applied to mechanical condition monitoring. The existing literature shows that the NFSs are superior to other nonlinear prediction methods (that is, RNN and RBFN). Recent development of NFSs based on fault diagnosis can be found in Refs. [17, 115‐122]. Nevertheless, several issues still need to be considered. For example, the number of nodes in the hidden layer is still difficult to be determined, and the selection of fuzzy parameters requires human intervention, all of which affect the prediction results of NFSs.

The establishment of a suitable model under the limited monitoring data is the key problem to estimate RUL and an urgent need for industrial production. SVM is a machine-learning algorithm based on Vapnik-Chervonenkis theory, which is used to solve the problems of classification and prediction when the sample size is small [123]. At present, SVM research mainly focuses on algorithm solution and model establishment. The purpose of algorithm solution is to address a constrained optimization problem by adopting the appropriate algorithm. The problem of model establishment includes optimization of model parameters, kernel function selection, feature vector extraction, and unbalanced sample. The aforementioned factors directly affect the model accuracy. Lu, et al [124], considered the difficulty of obtaining ideal prediction results under a small sample size. Thus, they proposed a least squares SVM (LSSVM) to estimate the degradation trend of slewing bearings and used PSO to optimize LSSVM parameters. Compared with the RBFN, the LSSVM model was demonstrated to be more accurate and effective. Dong, et al [125], used principal component analysis (PCA) to fuse the original features and reduce the dimension. Then, they used the LSSVM model to predict the bearing degradation process. Chen, et al [126], proposed an RUL prediction method based on relative feature and multivariate SVM (MSVM). In contrast to univariate SVM, MSVM overcame the shortcomings of insufficient condition monitoring information and mined potential, and useful information from small sample size. Caesarendra, et al [127], combined Cox proportional hazard model and SVM for failure degradation prediction of bearings. Widodo, et al [128], proposed a prediction method based on survival analysis and SVM. Kaplan-Meier and probability density function estimators were used to generate survival probability, and the kurtosis of measured data and survival probability were used as input and output of the SVM, respectively. The trained SVM successfully predicted machine failure time. Tran, et al [129], combined auto-regressive and moving average (ARMA) model, Cox proportional hazard model, and SVM for RUL prediction. Loutas, et al [130], used wavelet packet nodal energies and Wiener entropy as the feature vector, and proposed ε-support vector regression method in predicting the RUL of rolling bearings. He, et al [131], proposed a hidden Markov-SVM to predict surface roughness in hard turning. More published literature of applying SVM in RUL estimation of mechanical systems can be found in Refs. [132‐137]. SVM has been successfully developed rapidly in recent years, with an increasing number of studies focusing on SVM. However, SVM still has some problems to be solved, such as feature selection and large-scale training sample problem.