1DCNN Based Mechanical Fault Diagnosis Algorithm Analysis And Software Platform Design

In a modern society,machinery and equipment are becoming larger,more complex,more sophisticated,and more intelligent.However,as its complexity continues to increase,more and more types of failures occur.The traditional maintenance method takes a long time and has a high cost,which seriously affects production efficiency.Machine learning methods are required for automatic fault diagnosis.In recent years,mechanical fault diagnosis technology has been widely used in battery circuits,electronic equipment,gearbox gearboxes,wind turbines,railway switches and so on.As the number of sampling points,sampling frequency,and continuous sampling time continue to increase,the amount of data in the monitoring system has greatly increased,making the fault features more and more complicated,the connection between mechanical parts has also greatly increased,and the difficulty of feature extraction has greatly increased.The end-to-end one-dimensional convolutional neural network does not require manual feature extraction,which is significantly better than the traditional machine learning method of feature extraction and feature classification.This subject has designed a 1DCNN(1 Dimension Convolutional Neural Networks)network structure that can directly act on time-domain signals.Based on 1DCNN,a compact network structure is designed.Features are automatically extracted using a large and shallow network design.Then,fully connected neural networks and softmax activation functions are used for multi-classification.A cross-entropy loss function is adopted.Adam is used to optimize the learning rate and decline.direction.The network adopts an end-to-end design,which does not require manual feature extraction;it adopts a compact structure,so it has a fast prediction speed.In the prediction of I7 9750 H CPU,each prediction can be completed within 5s,which has a very high Real-time;At the same time,due to fewer network parameters,compared with the deep network,the demand for data is smaller.In the end,the recognition rate of more than 99% was achieved on the CWRU(Case Western Reserve University)rolling bearing data set.Since CWRU is data measured in the laboratory,its noise content is small,and its characteristics are very clear.It is easy to achieve a high accuracy.In the actual production environment,there may be large noise,so this topic adds Gaussian distributed noise on the basis of the CWRU data set,and explores to improve the anti-noise performance of the model.The size of the convolution kernel,Batch size,regularization and other aspects have been optimized to improve the anti-noise performance of the model.Finally,the accuracy rate can reach 98.31% when the signal-to-noise ratio is-8d B and 87.27% when the signal-to-noise ratio is-10 d B,which significantly improves the model’s anti-noise performance.In order to carry out real-time detection of the health status of mechanical equipment,this subject has designed a software platform for the fault diagnosis PC.It realizes the functions of WIFI receiving data,node switching,waveform display,local data storage,real-time diagnosis and so on.After real-time diagnosis is turned on,the host computer will automatically store the vibration data received from WIFI every 10 s,and diagnose the data,showing the mechanical operation status and failure probability distribution.