Research On Fault Diagnosis Algorithms Of Rotating Machinery Based On Vibration Signals

Rotating machinery is widely used in industrial field.Irregular health monitoring of its parts can effectively avoid accidents and economic losses.This paper proposes three kinds of rotating machinery fault diagnosis algorithms based on vibration signals.The main research contents are as follows:In the fault diagnosis models based on machine learning,the time-frequency energy generated by the traditional time-frequency analysis methods is very fuzzy,and the dimension of time-frequency representation is too high.For these problems,fine-grained fault diagnosis method of rolling bearing combining multisynchrosqueezing transform(MSST)and sparse feature coding based on dictionary learning(SFC-DL)was proposed.Firstly,the high-resolution time-frequency images of raw vibration signals were constructed by MSST.Then,SFC-DL was designed to reduce the dimension of time-frequency representation while preserve the subtle differences between different fault states.Moreover,the effectiveness and completeness of the sparse feature coding set obtained by SFC-DL were verified by max-relevance and min-redundancy(MRMR).Finally,the proposed method was applied on the bearing dataset from Case Western Reserve University(CWRU),achieved the fine-grained classification of 10 mixed fault states,and the classification accuracy was up to 98.03%.The end-to-end fault diagnosis models have the weak ability of feature extraction for raw vibration signals,and increasing the depth of network will increase the number of the model’s parameters.Therefore,the fault diagnosis model based on the shallow Inception-Mobilenet network was proposed.The model spliced the raw vibration signals into two-dimensional images,then used the multi-scale convolutional kernels to extract different resolution feature maps,and combined deepwise separable convolution to realize feature learning and classification.The shallow Inception-Mobilenet network was tested on the CWRU bearing dataset and the Machinery Failure Prevention Technology(MFPT)Society dataset,respectively,and achieved ten kinds of fault classification with 99.5% classification accuracy and three kinds of fault classification with 95.78% classification accuracy.Since the time-frequency analysis methods can effectively extract the implicit features from raw vibration signals,and convolutional neural network has the function of feature extraction and classification,the fault diagnosis models based on time-frequency images and convolutional neural network was proposed.Firstly,to describe the fault features of impulsive-like signals clearly and accurately,second-order time-reassigned multisynchrosqueezing transform(STMSST)was proposed.Then,to improve the training performance of the model,convolutional neural network based on evenly mini-batch training was designed.The experimental results showed that the training method was stable in the training process.At the same time,the method obtained the test accuracy with 99.83%,98.67% and 98.25% in CWRU dataset,MFPT dataset and loudspeaker pure-tone detection dataset,respectively.By comparison,it can be found that the time-frequency analysis method can make up for the lack of poor feature extraction ability of convolutional neural network,and convolutional neural network can solve the problem of low recognition accuracy of traditional feature extraction engineering and classifier.Therefore,the fault diagnosis model based on time-frequency analysis and convolutional neural network solves the problems of the existing fault diagnosis models,at the same time,achieves the best diagnosis effect,and is more universal in industrial application.