Early prediction of battery lifetime based on graphical features and convolutional neural networks

Accurate lifetime prediction of lithium-ion batteries in the early cycles is critical for timely failure warning and effective quality grading. Convolutional neural network (CNN), with excellent performance in feature extraction, has gained increasingly attentions in battery prognostics. However, since degradation test normally takes years to complete, employing end-to-end CNNs directly for battery lifetime prediction is impractical due to the limited number of available training samples and the scarcity of features in the early cycles. Instead of directly feeding the raw data, in this work, we propose to use graphical features for early lifetime prediction. Three feature curves, including capacity-voltage curve, incremental capacity curve, and capacity difference curve are used to construct graphical features. Specifically, the incremental capacity curve and capacity difference curve are derived from capacity-voltage curve, aiming to extract more information from both intra-cycle and inter-cycle perspectives. The evolution patterns of these feature curves over the initial 100 cycles show evident correla-tions with battery lifetime, and are termed as the graphical features. The three graphical features, after some proper transformation, are stacked into a three-channel image before feeding to the CNN model. Five classical CNNs, with different structures and key parameters, are investigated for battery lifetime prediction. Comparative experiments are conducted to study the influence of different feature combinations, voltage segments, and discharge cycles on the prediction performance. Experimental results demonstrate that simple CNNs with only a few convolutional layers can achieve satisfying prediction results. Additionally, networks with rectified linear unit are shown to outperform those with other activation functions.