An Intelligence Fault Diagnosis Approach for Rolling Bearings Based on Non-local Convolutional Neural Network

Due to its superior feature extraction ability, convolutional neural network (CNN) has been widely researched and applied in the domain of intelligence fault diagnosis and gained outperformance compared with other traditional machine learning methods. However, CNN could only process the signal within a local region for each operation. Thus, it lacks long-range correlation learning ability which is extremely important for time-series signal classification. Therefore, we introduce non-local means into CNN and propose non-local block (NLB) which could enhance the ability of long-range correlation capturing ability of CNN. Based on this block, we also present non-local convolutional neural network (NLCNN) for rolling bearings fault diagnosis. In the experiment section, we firstly verify the effectiveness of the NLB on the wheelset bearing dataset of high-speed trains (HST) under the noise environment of -6dB. Then, we also compare NLCNN with other four outstanding intelligence fault diagnosis methods and our proposed method gains the best performance.