Fault Transfer Diagnosis of Hydraulic Pump via Dynamic Simulation and an Improved Domain Adversarial Neural Network

Unsupervised domain adaptation (UDA) technology has been proven to effectively tackle the fault diagnosis (FD) issue of hydraulic pump (HP) across different working conditions, which utilizes a large amount of labeled data under one operating condition to achieve FD of unlabeled data under another new operating condition. However, a current challenge of this method is the difficulty in acquiring sufficient labeled data of the source domain in practical engineering and overcoming the significant distribution differences to extract transferable features. In this article, a novel fault transfer diagnosis approach for HP, named semi-supervised domain adversarial neural (DS-IDANN), is proposed by integrating dynamic simulation and an improved domain adversarial neural network (DANN). The proposed DS-IDANN method mainly includes two steps. To begin with, the simulated system of HP is built to obtain the simulation data under different health statuses by deriving the relationship between impact force and fault deformation. Then, a semisupervised domain adversarial graph convolutional network (GCN) is constructed to realize the fault transfer diagnosis from the simulation data to measured data by introducing a maximum entropy and improving the structure of the traditional DANN (TDANN). Finally, one case study conducted on a plunger pump is implemented to demonstrate the effectiveness and practicability of our developed method, and the result indicates that the accuracy of FD is higher than other existing approaches.