Research On Fault Diagnosis Method Based On Feature Mining And Convolutional Neural Network

In order to deeply explore the hidden feature information in the original fault signal,explore the construction of different dimensional samples and the improvement strategy of matching deep learning model,the research of intelligent fault diagnosis method based on feature mining and convolutional neural network is carried out.The paper proposes a sample construction method based on signal down-sampling,while improving the traditional one-dimensional convolutional neural network;investigates the sample construction method based on wavelet packet decomposition,and designs a two-dimensional convolutional neural network structure matched with the sample;by improving the classical network model with light weight,the computational resource consumption is reduced while completing the fault mode recognition task.The main components of the paper are shown below.(1)An intelligent diagnosis method based on signal down-sampling and one-dimensional convolutional neural network is proposed.For the fault signal containing noisy components,the original signal feature vector information is first extracted by fast principal component analysis,and then a one-dimensional sample sequence is constructed by random sampling method after normalization,which preserves the main feature information while down-sampling the signal.The first convolutional layer at the front end of the network and the batch normalization layer realize the extraction of feature components in the initial sample and the data distribution arrangement;the RReLU activation function with random parameters replaces the traditional ReLU activation function to realize the nonlinear transformation between nodes while maintaining the data integrity during the iterative process;The global average pooling with Softmax classifier at the end of the network achieves the mapping of global features to each class.The accuracy and stability of the proposed method are confirmed by the confusion matrix of the final experimental results.(2)An intelligent diagnosis method based on wavelet packet decomposition sample construction method and improved two-dimensional convolutional neural network is presented.The wavelet packet decomposition method is used to mine the different frequency feature components of the samples,and the wavelet packet coefficients corresponding to the frequency band with the highest energy share are captured and converted into gray-scale sample maps.In order to learn the feature information contained in multiple possible frequency bands,a feature fusion module is designed to achieve the fusion of a specified number of frequency bands and complete the shaping of the original feature samples.The learning rate adaptive adjustment algorithm improves network learning efficiency and effectively ensures the stability of the iterative process.The superiority of the proposed method in dealing with various types of complex faults is demonstrated by comparing it with the proposed method on data sets containing different types of faults,reflecting the superiority of the proposed method in dealing with various types of complex faults.(3)An intelligent diagnosis model based on the lightweight Ghost-Net is constructed.For the blind spot of the frequency domain analysis method in feature information mining,the phase spectrum and amplitude spectrum of the target fault signal are first extracted,and the original signal is stitched with its amplitude and phase information and converted into gray-scale sample images.In order to save computational resources as much as possible,the Ghost-Net model,which can generate redundant sample feature maps,is used,and the existing model is lightened and improved:only the main module for feature map size transformation is retained;the modified bottleneck module plays the role of down-sampling while feature extraction;the compression-excitation module focuses more on mining channel heavy intrinsic features;Dropout strategy avoids possible overfitting on the test dataset.The experimental validation results show that the intelligent diagnosis model based on lightweight neural network has a more stable and efficient performance in the field of fault diagnosis.