Hierarchical Stochastic Spatial–Temporal Transformer for Trustworthy State-of-Health Estimation of Batteries in Industrial Applications

With lithium-ion batteries prevalent in safety-critical industries, accurate and reliable estimation of battery state-of-health, termed trustworthy prognosis, has become crucial. However, neglecting model uncertainty during representative correlation learning may lead to inappropriate aggregation of ambiguous segments. Moreover, shifts in dominant spatial channels and sparsity in useful temporal features further hinder the trustworthy estimation. Furthermore, the misalignment between the predicted and actual distributions leads to a confidence biases issue and degrades performance in distinguishing out-of-distribution samples. To address these challenges, we propose a hierarchical stochastic spatial–temporal Transformer (HSSTT). First, HSSTT implements stochastic self-attention utilizing Gumbel–Softmax reparameterization for uncertainty quantification. Then, a hierarchical spatial–temporal Transformer is designed to leverage uncertainty-aware timestep-wise dilated convolution and clustered stochastic self-attention. Finally, we theoretically analyse the confidence bias issue through bias-variance decomposition and develop a principled calibration strategy. Experimental results on four datasets demonstrate the superiority of HSSTT in trustworthy prognosis against seven State-of-the-Art models.