Fault Diagnosis Of Rolling Element Bearings Based On Deep Belief Network And Multiple Sensors Information Fusion

Rolling element bearings are one of the most critical but vulnerable components in rotating machines.The healthy condition of rolling element bearings directly affects the performance and reliability of mechanical equipment.The fault diagnosis researches about rolling element bearings at present are mainly around two aspects: fault location detection and fault severity estimation.However,few researches combined the two aspects together.An integrated fault diagnosis scheme should not only distinguish the fault locations,but also recognize the different severities in each fault locations.This study is operated according to the following three aspects to construct the intergrated fault diagnosis scheme.1.The fault features that can effectively reflect the fault locations and corresponding severities are extracted.Lempel-Ziv complexity(LZC)and permutation entropy(PE)are the two common used nonlinear complexity analysis indexes,which can effectively distinguish the fault samples away from healthy samples.However,fault samples of each fault locations are composed by different fault severities,which can not be distinguished only by LZC and PE.To reflect the fault information comprehensively,LZC combined with wavelet packet analysis and PE combined with ensemble empirical mode decomposition(EEMD)are served as two fault feature extraction methods in this study.Meanwhile,a feature dimension evalution method based on the within-class scatter matrix and between-class scatter matirx is proposed to demonstrate the reasonable of the fault feature dimension selected in this study.2.The deep belief networks(DBNs)are severed as classifiers to achieve fault recognition with excellent result.Though DBN classifier possess a great capability in discriminating different fault classes,a number of parameters in DBN,i.e.,learning rate,momentum and number of neurons in each hidden layers,must be given before recognition process.However,the parameter setting of DBN is a challenging task and can be very time consuming.A hybrid optimization algorithm is proposed to improve the calculation efficiency of DBN parameters optimization process,in which the social thinking capacity in particle swarm optimization(PSO)is combined with the local search capability in genetic algorithm(GA).The expriment result demonstrate that the proposed optimization method can obtain the optimal parameters and complete the parameter setting task quickly.3.D-S evidence theory is used as a fault information fusion method to improve the final recognition accuracy.Two DBNs classifiers are used to evaluate both bearing fault conditions and its severity ranking,respectively.The basic probability assignment(BPA)functions in D-S evidence theory are constructed based on the global reliability and local reliability of classification results.The final identification result is given by fusing the two BPA functions.4.A hierarchical identification framework is proposed to evalute both fault locations and corresponding severities,in which the first layer is used to detect the fault locations and the second layer is used to estimate the fault severities.In addition,the formula of final accuracy rate is also provided.Two contrast cases are set to verify the effectiveness of the proposed fault diagnosis scheme in this study.In the first contrast case,only one classifier in single direction that without fault information fusion is used to achieve fault conditions and its severity ranking recognition.In the second contrast case,SVM is substituded for DBN as classifiers to achieve fault conditions and its severity ranking hierarchical identification.The experiment result demonstrates the propoesd fault diagnosis scheme is significantly superior to the two contrast cases,which can precisely distinguish the different faulty classes.