Discriminant Analysis-Guided Alignment Network for Multimachine Fault Collaborative Learning and Diagnosis

In general, the safety and efficiency of thermal power plants require the collaboration of multiple coal mills. However, running data of different coal mills will introduce significant inconsistent distribution, resulting in suboptimal performance or even the unavailability of conventional diagnosis methods. For this end, this paper presents an advantageous discriminant analysis aided collaborative alignment network (DA-CAN) for cross-device fault diagnosis. Firstly, the contribution of each feature in distinguishing source and target domains is determined by assigning the adaptive updated weight, which is helpful to keep the gradient direction more stable in domain transferring. To avoid destroying the inherent data structure of different domains, we design multiple complementary class-wise discrepancy metrics to enhance the domain consistency during the domain adaption process. After that, a joint training loss term with adjustment factor is introduced to transform the private data of individual coal mill into collective representations, and smooth the conditional and marginal distributions discrepancy collaboratively. Finally, the experiment results of real-world coal mill group indicate that the DA-CAN is more effective and practical than the state-of-the-art transfer learning methods regarding multi-machine fault diagnosis.