Cross-condition fault diagnosis method for hydraulic systems based on domain adaptation and ensemble learning

In response to the challenge of low fault diagnosis accuracy for hydraulic components in cross-condition fault diagnosis of multi-sensor fused hydraulic system using deep domain adaptation method, a method based on domain adaptation and ensemble learning is proposed. This method employs convolutional neural network as the primary architecture to recognize hydraulic signals collected by each sensor. It aligns the features of samples from source domain and target domain using kernel maximum mean discrepancy and central moment discrepancy. Additionally, a dual-attention mechanism is applied to select features, and the domain adaptation models trained for each sensor are integrated. This ensemble learning approach aims to diagnose faults in target domain hydraulic components across varying working conditions. Experimental validation is conducted using data from a hydraulic cooling system testbed and simulated data from an underwater robot hydraulic system. The method demonstrates higher accuracy in cross-condition fault diagnosis of hydraulic systems compared to other domain adaptation methods.