Fault pattern recognition of rolling bearing based on smoothness prior approach and dual-input depth spatial–temporal fusion

Abstract

Accurate fault diagnosis is critical for the safe and stable operation of mechanical equipment. Current deep learning (DL)-based fault diagnosis can extract various and deep discriminative features efficiently. Different DL methods are applicable to different data types, and the extracted features are also different. In addition, the equipment working conditions are complex, and noise from the working environment is inevitable. If only a single network or single input is used, it is difficult to extract the information that can comprehensively describe the fault features, which affects the diagnosis accuracy, especially when multiple faults (more than ten types) are considered. Hence, this study proposes a method for fault diagnosis based on the smoothness prior approach (SPA) and a dual-input depth spatial–temporal fusion network. First, the original signal is decomposed via SPA and two different types of input are constructed. Second, a dual-input depth spatial–temporal fusion network is proposed to extract deeper information by simultaneously learning temporal and spatial features. These two types of features are fused using a two-dimensional convolutional neural network to complete classification tasks. The average recognition accuracy of the proposed network can be maintained above 99% on two commonly used benchmark vibration datasets. Compared with several state-of-the-art diagnosis methods under different noise level experiments, the proposed method can achieve higher diagnosis accuracy for each dataset under different working conditions.