Frequency Domain Feature Extraction and Long Short-Term Memory for Rolling Bearing Fault Diagnosis

Abstract

With the rapid development of the high-speed rail-ways, the speed of trains is getting faster and faster, and the dynamic load between the wheels and rails of the vehicle increases accordingly. The rolling bearing is a key part of the high-speed train transmission system. The train is subjected to high-frequency vibration for a long time during operation, and the bearing is prone to fatigue damage, which affects the safe operation of the train. Nowadays, many methods have been applied in fault diagnosis like reinforcement learning, convolutional neural networks and autoencoders. One of the typical methods is the reinforcement neural architecture research method. It makes neural network design automatic and eliminates the bottleneck associated with choosing network architectural parameters. However, this method focuses on the time domain signal, and a time domain signal cannot capture the particular properties of a frequency domain signal. In order to solve these problems, we propose a new method containing two Steps: Use FFT to convert the time domain signal to the frequency domain and use Bi-LSTM neural network model to recognize different faults. For each fault, the time series signal has some correlation with some specific frequencies. The frequency domain is more intuitive than the time domain and describes different states of faulty types. For recognition, LSTM is better at classifying sequence data than other methods, and Bi-LSTM can predict the sequence from both directions, achieving higher accuracy. Experiments on public data sets demonstrate the efficiency of the proposed method.