Abstract: DC-DC converters work under various complex stresses such as thermal and electrical stresses for a long time, resulting in problems like fluctuating interference in the circuit degradation trend and low prediction accuracy. To address these issues, a fault prediction method combining time series decomposition and deep learning is proposed. Firstly, the output voltage is taken as the sensitive signal of the circuit fault and its fuzzy entropy feature extraction is performed, in order to address noise interference with circuit signal acquisition. Then, ICEEMDAN decomposition of fuzzy entropy is conducted, in order to reduce the effect of local fluctuations in the feature data and to extract monotonic trend components. Finally, multi-step prediction of IMF components and trend terms using BiLSTM with Attention mechanism, and reconfigure the results of each prediction component to enable circuit-level fault prediction. By building a physics experiment platform with boost converters as the research object, compared with the algorithms such as LSTM and AM-LSTM, the prediction error of the proposed method AM-BiLSTM is reduced by 61.44%, which verifies its feasibility and accuracy.