TFARNet: A novel dynamic adaptive time-frequency attention residual network for rotating machinery intelligent health prediction

Abstract

Rotating machinery is critical functional part of industrial mechanical equipment, and health status of rotating machinery is closely related to equipment stability, reliability and safety. Vibration signals for health prediction are often collected under operating conditions with variable loads and speeds, which makes health prediction more challenging. STFT-based time-frequency representation methods are widely used for the health prediction of rotating machinery. However, these methods lack specific learning mechanisms to effectively distinguish the time-frequency representations at different time points and frequency bands and highlight important feature information. To vanquish the weakness, this paper develops a novel dynamic adaptive time-frequency attention residual network (TFARNet) for rotating machinery intelligent health prediction. Firstly, a new adaptive STFT time-frequency attention (TFA) unit is developed to recalibrate time-frequency features, thereby enhancing important information and suppressing redundant information. Secondly, the TFA unit is inserted into the residual network, by stacking multiple residual blocks and TFA units to establish TFARNet and efficiently learn more discriminative features. Thirdly, label smoothing regularization and dynamic learning rate are employed to accelerate model convergence and optimize the model training process. Finally, three cases are carried out to evaluate the developed method. Compared with the other seven health prediction methods, the developed method can achieve higher prediction accuracy.