A bearing fault diagnosis and monitoring software system based on lightweight neural networks to resist coloured noise

In actual industrial sites, the application of bearings is becoming increasingly widespread. In order to better monitor the faults of bearings, this article combines the concept of deep learning and designs a bearing fault diagnosis and monitoring software system based on lightweight neural networks to resist coloured noise. This system is developed based on MATLAB App Designer. When testing the system, five different bearing datasets, namely MFPT, Paderborn, IMS, Ottawa, and CWRU, are applied. Considering that the data in actual scenarios contains complex noise, coloured noise signals are added. Compared to traditional fault diagnosis software that requires pre writing data into the program, this software can perform real-time processing on any single column vibration data file. By using lightweight neural network methods to preprocess the data collected by sensors, the SqueezeNet network has a faster speed to extract significant features of vibration. This software system can achieve time-frequency domain image output of signals, with multiple noise reduction methods. It can also calculate the frequency of faults based on bearing model data. Through envelope spectrum images, the location of faults can be monitored and email reminders can be sent to engineers.