Real-time Diagnosis Of Motor Bearing Faults Based On Deep Learning

The development of modern industry has led to the widespread use of electric motors in different aspects of national construction.If they fail unexpectedly,they will result in economic losses,equipment downtime,and even serious personal injury.The rapid development of embedded systems and artificial intelligence technology has provided new ideas for fault diagnosis and real-time monitoring of motors.This article proposes a real-time diagnosis method for motor bearing faults based on deep learning.Using convolutional neural networks to train the model of motor bearing fault signals,and then deploying the model file to an embedded platform for real-time diagnosis.The main research content and methods are as follows:1.Design an embedded hardware system for real-time diagnosis of motor bearing faults,including modules such as signal acquisition,signal transmission,signal storage,signal processing,and result display.Based on the types of bearing faults that exist in the actual production process,the bearing is selected and processed for different faults.Then,the vibration signals corresponding to the faults are collected through the built embedded platform,and the establishment of the fault database is ultimately completed.2.A rolling bearing fault diagnosis method based on residual connection and onedimensional separable convolution(1D-RSCNN)is proposed to address the problems of traditional fault diagnosis models with multiple parameters,long training and diagnosis time,poor noise resistance,and unsuitable for real-time diagnosis.Utilizing one-dimensional separable convolution and global average pooling to compress the model size and improve the computational efficiency of traditional convolutions;By using a wide convolutional kernel and introducing Dropout in the convolutional layer,the tolerance to noise is improved.The experimental results show that compared with other models,the proposed 1D-RSCNN model has the advantages of high diagnostic accuracy,good real-time performance,and strong antiinterference ability.3.Aiming at the problem of mixed compound fault signals and unclear fault features,which rely on complex signal processing technology for feature extraction,LSEUnet and 1DRSCNN methods is proposed for intelligent separation and diagnosis of compound faults.The feature extraction module LSE of the lightweight LSEUnet network is constructed.To reduce the number of parameters,a convolutional layer with a convolutional kernel size of 1×1 is used as the first layer and next is 3×3 deep separable convolutions.After convolution operation,SE blocks are added to adaptively assign different weights to different channels.Using 1×1 point by point convolutional layer to output the final features.LSEUnet is used to train fault signals after time-frequency conversion.Binary masks with excellent performance are trained to achieve intelligent separation of compound faults;Using the lightweight network model 1DRSCNN as a feature learning model,effectively learning and identifying features from the original vibration signal for single fault classification.The fault test bench conducted real-time diagnosis and achieved good results.