基于 GRU 和改进注意力机制的多信息融合的EMA 故障诊断方法

Addressing the issues of insufficient time-series features and incomplete fault information in fault diagnosis methods for electromechanical actuators (EMAs) based on traditional machine learning and deep learning, a fault diagnosis method for EMAs based on multi-source signal fusion with gated recurrent unit (GRU) and an improved attention mechanism is proposed. First, the collected signals from different sensors are divided into separate channels, and the time-series features of each channel’s signal are extracted using GRU. The self-attention mechanism is then introduced to further distinguish the important relationships between different time points of the signal. A multi-channel attention mechanism is employed to adaptively fuse the features from different channels. Finally, fault diagnosis is achieved through the classifier. Experimental results based on the test rig dataset show that the diagnostic accuracy improves by 10% compared to the single-sensor model and by 5.2% compared to the model without the attention mechanism. Compared to classical machine learning, deep learning and recent improvements in deep learning-based algorithms from the past two years, the diagnostic accuracy of the proposed model exceeds 98.5%, demonstrating optimal diagnostic performance.