An Anti-Noise Convolutional Neural Network for Bearing Fault Diagnosis Based on Multi-Channel Data

In real world industrial applications, the working environment of a bearing varies with time, and some unexpected vibration noises from other equipment are inevitable. In order to improve the anti-noise performance of neural networks, a new prediction model and a multi-channel sample generation method are proposed to address the above problem. First, we proposed a multi-channel sample representation method based on the envelope time–frequency spectrum of a different channel and subsequent three-dimensional filtering to extract the fault features of samples. Second, we proposed a multi-channel data fusion neural network (MCFNN) for bearing fault discrimination, where the dropout technique is used in the training process based on a dataset with a wide rotation speed and various loads. In a noise-free environment, our experimental results demonstrated that the proposed method can reach a higher fault classification of 99.00%. In a noisy environment, the experimental results show that for the signal-to-noise ratio (SNR) of 0 dB, the fault classification averaged 11.80% higher than other methods and 32.89% higher under a SNR of −4 dB.