Intelligent fault diagnosis of rolling bearings with low-quality data: A feature significance and diversity learning method

Abstract

In real engineering applications, low-quality and insufficient vibration signals of rolling bearings, which are usually multiple faults with low signal-to-noise ratios (SNR), restricts the effective application of intelligent diagnosis methods based on deep learning. In this study, a new rolling bearing fault diagnostic method is proposed based on an improved deep fused convolutional neural network (DFCNN) combined with complementary ensemble empirical mode decomposition (CEEMD) and a short-time Fourier transform (STFT). First, the DFCNN is developed by introducing batch normalisation (BN) to overcome the gradient disappearance in convolutional neural networks (CNNs) and fuse multi-scale CNNs to improve feature diversity. Meanwhile, to enhance the quality of information, the original vibration signals are processed by CEEMD and STFT to obtain reconstructed denoising signals and time-frequency spectrograms, respectively. Then, these two types of data are input into a concatenation layer of DFCNN to extract the deep fault features, which are transformed by SoftMax to achieve bearing fault recognition. Finally, two bearing cases with data collected from the accelerated life degradation and different distributions are studied. The results reveal that the proposed method can fully mine fault features with low-quality data and realise efficient fault diagnosis.