Research On Fine-Grained Fault Identification Algorithm Of Rotating Machinery Vibration Signal Based On Time-frequency Image And Multi-level Sparse Coding

Intelligent diagnosis of fine-grained faults in rotating machinery is an indispensable task in modern industry.Considering that the one-dimensional vibration signals of rotating machinery are complex and non-stationary,a fine-grained fault recognition algorithm based on time-frequency images and multi-level sparse coding is proposed by combining the theory of signal time-frequency analysis,deep semi-negative matrix factorization(Deep Semi-NMF),and multi-layer cascading fine-grained fault detection model.The overall framework can be divided into three parts,including time-frequency image construction,multi-level sparse coding,and a multi-level cascade classifier model,to implement fine-grained fault classification for rotating machinery faults.Considering the traditional time-frequency analysis method with low focus and severe cross-terms,a multisynchrosqueezing S-Transform(MSSST)is proposed based on the theories of S-transform(ST)and multisynchrosqueezing,combined with a denoising algorithm based on non-local means(NLM).It is used to construct time-frequency images of one-dimensional vibration signals.The algorithm squeezes ST through several iterations,so that the time-frequency result of ST is redistributed many times,which greatly improves the focusing of the time-frequency coefficient.In mathematical simulation experiments,comparing with the short-time Fourier transform(STFT),ST,synchrosqueezing transform(SST),and second-order synchronousqueezing transform(SST2)and multisynchrosqueezing transform(MSST),the corresponding entropy value of the time-frequency image generated by MSSST is 0.4938 from the corresponding entropy value,which is less than other time-frequency images,and MSSST only needs 3 times of iterative squeezing to achieve the desired effect.It can be concluded that the focus of MSSST is better than other algorithms?Considering the dimensions of time-frequency image is too high,the theories non-negative matrix factorization(NMF)and sparse expression are combined to propose a multi-level sparse coding model based on Deep Semi-NMF and orthogonal matching pursuit(OMP)algorithms.First,the model calculates the multi-level basic matrix through a deep semi-NMF theory and a small batch iterative algorithm.Then,an OMP sparse expression algorithm is introduced to extract multi-level feature vectors of time-frequency images based on multi-layer dictionary.Finally,the obtained feature vector set is used to train a linear SVM,and the test set is used to test the accuracy of the algorithm.Experiments on the MFPT data set show the fault recognition accuracy on the MFPT dataset can reach 97.82%,which is higher than other algorithm models.Considering the complexity of fine-grained fault diagnosis for rotating machinery,a two-level cascade detection framework based on time-frequency image and multi-level sparse coding is proposed by using time-frequency analysis theory and multi-level sparse coding algorithm.First,the framework uses multisynchrosqueezing S-Transform(MSSST)algorithms to construct time-frequency images of one-dimensional vibration signals;then,it uses Deep Semi-NMF and OMP algorithms to extract the two-level sparse features of time-frequency images;finally,it uses Two levels of feature sets are used to train two levels of cascade classifier frameworks.The trained framework performs fault recognition on the test set in two steps.The first step identifies the basic fault category of the test sample,and the second step identifies the damage level of the basic fault.Experiments on the CWRU database show that the accuracy of fine-grained fault recognition corresponding to the detection framework in this paper can reach 98.92%,which is higher than other algorithms.