State‐of‐Health Estimation of Lithium‐Ion Battery Based on Partial Charge Data and Multimodal Image Fusion

Accurate estimation of lithium‐ion battery state of health (SOH) is essential for battery safety, performance, and reliability. However, existing SOH estimation methods depend on full charging data, while single‐modal image approaches capture limited battery information. To overcome these problems, this article proposes a novel SOH estimation method using partial charging data and multimodal image fusion. First, voltage and temperature data are collected within specific state of charge intervals after wavelet denoising, eliminating dependence on complete charge curves. Second, the partial charge data are transformed into multimodal image data using Gramian angular field, recurrence plot, and adaptive average pooling. These transformations individually capture global, local, and statistical battery features. Subsequently, the adaptive graph channel attention mechanism is integrated into the residual network architecture, forming the improved residual network structure. This enhances the model's ability to synergistically fuse multimodal image information and extract deep features. Next, the fused features are input into a bidirectional long short‐term memory network to capture dynamic relationships between features. The self‐attention mechanism is then used to assign feature weights, and a fully connected layer outputs the SOH. Finally, testing the proposed method on National Aeronautics and Space Administration and Oxford datasets demonstrates its accuracy and generalization capability, with mean absolute error and root mean square error both below 1.3%.