Fault Prediction of Bearing Based on Dual Dimensional Perception and Composite Gated Recurrent Network

Bearing failures seriously affect the operational reliability of rotary equipment. The early degradation characteristics of bearing faults are not obvious, and it is crucial to effectively extract fault features. It is more difficult to achieve predictive research on bearing faults based on the identification of early and unclear fault characteristics of bearings. This paper proposes a research method for early degradation fault perception and prediction of bearings based on dual dimensional feature perception and composite gate controlled recurrent network (composite GRU). Firstly, analyze the early fault characteristics of bearing degradation and propose feature extraction of bearing vibration waveform data from both trend and detail dimensions at the data layer. In the trend dimension, 7 preprocessed vibration waveform feature quantities were proposed from 33 feature data as the feature data for the trend dimension. In the detail dimension, the original vibration waveform is subjected to spectral transformation and discrete wavelet packet decomposition, and a neural network based on attention mechanism is used to extract features from the data obtained by wavelet packet decomposition, in order to obtain 16 dimensional feature data in the detail dimension. Secondly, a composite gated recurrent network main algorithm model with enhanced post attention algorithm layer was designed. After fusion and feature principal component extraction of the aforementioned two-dimensional data, the proposed composite gated recurrent network model with algorithm level attention enhancement is used for degradation state fitting identification research. Finally, by using actual bearing degradation data, the proposed algorithm’s ability to perceive and identify early degradation states of bearings was verified, demonstrating the effectiveness and superiority of the proposed method for bearing fault prediction research.