Research On Mechanical Fault Recognition And State Assessment Based On Deep Learning

With the introduction and promotion of the strategies of "Industry 4.0" and "China made 2025",the research of high end manufacturing system and intelligent mechanical equipment is being paid more and more attention.However,if the mechanical equipment fails to be detected in time in actual production,it will cause serious economic loss and even safety accident.Therefore,it is of great significance to study fault identification and state evaluation of mechanical equipment.Based on the advanced technologies and theories of modern signal analysis,large data processing,manifold learning and deep learning,this paper studied the two main lines of fault identification and state assessment by taking the key components of mechanical equipment,bearings,as the research objects,which mainly completed the following three aspects:(1)After analyzing the characteristics of the vibration signal of the mechanical equipment in the engineering practice,the method of extracting the 11 time domain features,13 frequency domain features and 14 wavelet packet energy features from the signal to construct the high dimensional mixed domain feature data set are proposed.In view of the troubles of information redundancy,easily causing "over-fitting" and low computational efficiency of the high-dimensional feature set,it is proposed to convert the high-dimensional features to low-dimensional space using the manifold learning algorithms.The experimental results showed that the manifold learning transformation can map high-dimensional data to low-dimensional space without distortion,and is more conducive to subsequent calculation and analysis.(2)After analyzing the excellent performance of deep learning theory in various pattern recognition problems and its shortcomings,deep belief network(DBN)was introduced into the field of fault identification.Aiming at the instability and uncertainty of the DBN model and the difficulty of establishing an effective model for the actual problem,the method of optimizing the parameters of DBN model by using the simulated annealing algorithm(SAA)was proposed,and the theoretical basis and implementation processed of the optimization method was also introduced.Six different types of bearing failure data were used for experiments to compare the fault recognition effects based on different feature sets and different machine learning methods.The results showed that the principal component analysis(PCA)of the linear feature transformation method is more suitable for the transformation of the high-dimensional features of the data,and the DBN model optimized by SAA has more accurate fault recognition capabilities than the un-optimized DBN model,support vector machine(SVM)and probabilistic neural network(PNN)and other algorithms.(3)Most of the existing equipment state assessment technologies need to do a lot of feature extraction,feature conversion,feature screening and other processing on the original signal data,aiming at the issues of increasing the computation complexity,reducing reliability,introducing human factors and so on,the method based on original signal and DBN Self-learning for evaluating the operational status of mechanical equipment was proposed.The advantages of research based on the original signal are analyzed,and the process and implementation procedures based on the DBN evaluation algorithm are formulated.Employing the bearing whole-life degradation data for verification and comparison,through comparing the evaluation results based on PCA and Isomap conversed characteristics and the evaluation results based on Hidden Markov model(HMM)and back propagation neural network(BPNN)methods as well as the quantitative analysis,the proposed method was proved to be more stable and accurate in identifying machinery operating status during different time periods.Additionally,the research also discovered that the method of extracting features is more suitable for mechanical fault recognition(classification),while the method using the original signal without extracting features is more suitable for machine state assessment.