Degradation prognosis for proton exchange membrane fuel cell based on hybrid transfer learning and intercell differences

Proton exchange membrane fuel cell (PEMFC) has been widely used in diverse applications. However, degradation and durability problem is one of the biggest barriers to take PEMFCs into extensive commercial use. Prognostics and health management is an effective solution to this problem. In this study, we focus on its core technology prognostics and propose an individual difference conscious prediction method for PEMFC using a hybrid transfer learning approach to get higher accuracy. Firstly, a time-scale self-optimization local weighted regression method is designed to adaptively smooth the raw data to prominent the performance degradation trend. Then, to obtain a more similar curve to the predicted fuel cell as the training data of the prediction model, a transferability measurement method using cosine-distance selects the most similar historical test data. Furtherly, it is utilized to generate a more similar curve by a data transfer method combining a deep learning model named stacked autoencoder and a hybrid transfer learning strategy. Two types of transfer learning approaches are fused to maximally mine available information from historical data and previous models to help improve the similarity of the generated curve. In this process, the common degradation information of all cells and individual information of the predicted cells are considered to improve generation quality. Finally, a prediction model using stacked Long-short Term Memory(LSTM) having a significant advantage in modeling series relation is trained by the generated samples cut with variable width sliding windows and estimates remaining useful life(RUL) the target fuel cell. Experimental validation data are employed to verify the effectiveness of the proposed algorithm. Satisfying results are also obtained by accuracy comparison under different smoothing scales, numbers of transferable samples, and prediction methods.