SOH Estimation Using Contrastive Learning to Extract Cross-Battery Transferable Degradation Representations

Due to the complexity of real-world conditions for lithium-ion batteries, the scarce labeled data for state of health (SOH), making SOH prediction task challenging. This article proposed a SOH prediction method using contrastive learning (CL) to extract battery aging characteristics from a large amount of unlabeled data. CL learn from positive sample pairs, enhancing the model’s robustness to noise and environmental factors, and learn from negative sample pairs, effectively capturing battery degradation information. We detailed the strategy for constructing the positive and negative sample pairs, meanwhile analyzed the data from constant current (CC) charging stage in detail explained why CL is effective. The experimental results show that after extracting aging features with strong transferability through CL, in SOH prediction task, training with only one battery that has an SOH label can significantly reduce the mean squared error (MSE) to 0.006%, demonstrates the effectiveness and high generalization capability of the proposed method. This article offers a new approach for SOH prediction under conditions with limited labeled data.