Research On Bearing Fault Diagnosis Based On Deep Learning With High-dimensional Hierarchical Features

Rolling bearings are an important part of rotating machinery.Due to the high degree of integration of mechanical components,the sampling points and sampling periods of bearing fault monitoring systems are increasing.The ability to process large amounts of data has become a necessary requirement for modern fault diagnosis methods.The traditional machine learning fault diagnosis model is a shallow network structure.If the input feature representation of the model is insufficient,it may lead to misdiagnosis.The difficulty of feature selection in rolling bearing high-dimensional data leads to inaccurate classification results.Although deep learning can automatically learn basic features from original data,standard deep learning methods only consider a single deep feature and ignore shallow features.And cause the problem of feature loss.This paper is mainly based on the deep learning method,focusing on the high-dimensional data features,multi-level feature extraction and fault diagnosis and recognition of bearing vibration signals.The research work and results are summarized as follows:(1)Aiming at the problem of inaccurate classification results caused by the difficulty of feature extraction under high-dimensional data,an improved regularized extreme learning machine is proposed to apply to the fault classification method of noise reduction autoencoder.The stacked noisereducing autoencoder is used to perform feature learning on high-dimensional data,extract more robust features,and use this feature as the input of the regularized extreme learning machine to obtain the classification result.Aiming at the difficult problem of selecting regularization parameters in regularized extreme learning machines,quantum particle swarm optimization algorithm is used to optimize the parameters.Through the use of Case Western Reserve University bearing data set for verification,the experimental results show that the classification result of the regularized extreme learning machine is about 1%higher than that of the extreme learning machine,and the stability is better,and the test result of the quantum group optimization algorithm is better than the particle swarm optimization algorithm is 1%～2% higher.Therefore,the model can obtain a more accurate classification effect under high-dimensional data sets.(2)The standard deep learning method usually only considers the deep features of the input data after convolution,ignoring the shallow features,and the problem of partial feature loss.A convolutional neural network combining multi-level features is established.Fault identification model.Use a convolutional neural network with two convolutional layers,and input the features extracted after the two layers of convolution into the long and shortterm memory network for further extraction of time series features,and then convolve one layer of convolution and two layers of convolution The feature segments are connected to realize multi-level feature extraction.Through the use of case Western Reserve University bearing data and the University of Ottawa bearing data for verification,the experimental results show that the average accuracy of the model is 1% to 7% higher than the standard deep learning method,and the standard deviation is smaller and the stability is better.