Remaining Useful Life Transfer Prediction and Cycle Life Test Optimization for Different Formula Li-ion Power Batteries Using a Robust Deep Learning Method

Abstract

Aiming at providing life information of different formula batteries for designers iteratively selecting an appropriate formula, cycle life tests are demanded to long-term perform until battery capacity reaching a pre-set failure threshold. However, the time-consuming test brings a high and unbearable cost to battery enterprise specifically focusing on cost and efficiency. For this practical problem, a prediction-based test optimization method is proposed to estimate the battery remaining useful life to replace its test life, and to shorten the test cycles for saving the test-cost. The prediction accuracy and robustness to the variation on battery formula and test temperature are guaranteed by an instance-based transfer learning method combined with a highly robust deep learning method named stacked denoising autoencoder. An average Euclidean distance-based transferability measurement method selects the most similar historical test data of batteries with other different formulas. It helps to compensate for the lost trend information of the predicted battery caused by cycles reduction and to augment the data for effectively training the prediction model. The actual test data from a battery company verify the accurate prediction and significant cost saving. Nearly more than 30% of the test cycles are optimized for different formula batteries on average.