A knowledge transfer method for water faults diagnosis of proton exchange membrane fuel cell based on sample re-weighting

Diagnosing water faults in proton exchange membrane fuel cell (PEMFC) often suffers from a shortage of fault samples. To address this problem, this paper proposes an innovative knowledge transfer method for water faults diagnosis that combines prior knowledge with sample re-weighting (PK-SR). Firstly, artificial prior features extraction is performed, mapping raw samples to fault feature space. Then, the fault feature similarity between source and target domain is calculated based on the extracted fault features vectors. Subsequently, initial weights for samples are calculated and applied to modified TrAdaBoost algorithm, which updates sample weights based on both fault feature similarity and classifier prediction results. Finally, the high-precision water faults diagnosis task was achieved with insufficient faults data, and overfitting was essentially avoided. Through comparative analysis with the latest methods, the proposed PK-SR method has been verified to have significant performance advantages. To our knowledge, this is the first successful attempt to combine prior knowledge of PEMFC water faults with transfer learning method for knowledge transfer and taking into account support for edge computing devices.