Configuration Selection for Degradation Trajectory Prediction of Power Modules Based LSTM Model

This paper investigates a data-driven approach for degradation trajectory prediction aimed at reducing reliability testing time, specifically employing iterative sequence-to-sequence prediction based on the long short-term memory (LSTM) model. It provides a comprehensive understanding of the application of the data-driven method to the scenarios analyzed and the related details, which involves data processing and the hyperparameter selection process. The degradation data support is from 18 samples under three test conditions in the power cycling test. The study considers the impact of different configurations (i.e., hyperparameters) of deep learning models on the prediction results, namely input/output features, the data down-sampling coefficient, the number of network layers, the number of hidden layer units, and the lengths of input and output sequences. Moreover, two indicators, the prediction accuracy and the degree of testing time reduction, are defined to quantify the prediction analysis performance. Finally, the sensitivity analysis quantifies the contribution of each of the six factors to both predicted performance metrics.