Enabling unsupervised fault diagnosis of proton exchange membrane fuel cell stack: Knowledge transfer from single-cell to stack

Fault diagnosis has been considered as the most promising technique to strengthen reliability and durability of proton exchange membrane fuel cell (PEMFC) stack. However, the contradictory between sufficient labeled stack data requirement from existing methods and unlabeled stack data from real-world applications brings great challenges to unsupervised PEMFC stack fault diagnosis. For breaking through the bottleneck, this paper proposes an innovative deep transfer learning-based unsupervised PEMFC stack fault diagnosis method through knowledge transfer from single-cell to stack (DTL-PEM). Specifically, on the one hand, the proposed DTL-PEM method combines adversarial learning and conditional distribution adaptation to reduce both marginal and conditional distribution bias between single-cell and stack data, which greatly encourages capturing rich domain-invariant features to promote knowledge transferability from single-cell to stack. On the other hand, a weighting module is introduced in DTL-PEM network to eliminate the negative effect stemming from asymmetric label space. The effectiveness of the proposed DTL-PEM network is verified using labeled single-cell and unlabeled stack voltage data at various PEMFC states. Compared with the existing state-of-the-art methods, the proposed DTL-PEM network can not only achieve accurate unsupervised PEMFC stack fault diagnosis by knowledge transfer from single-cell to stack, but also have superior adaptability to different data openness, which make it promising in real-world PEMFC stack fault diagnosis. To the best of our knowledge, this is the first successful attempt to solve the unsupervised PEMFC stack fault diagnosis problem based on knowledge transfer from single-cell to stack.