Data-Driven Battery Health Prognostics Using Time-Frequency Feature Maps and Spatial-Temporal Neural Network

Lithium-ion batteries have been widely used in various application scenarios, acting as the heart of power storage systems. Reliable prognostics and health management (PHM) play essential roles in the safe operation and reliable maintenance of battery systems. Within this context, a data-driven method based on time-frequency feature maps and spatial-temporal neural networks is proposed for state-of-health estimation and cycle-to-knee prediction. First, raw data is acquired from partial charging curves of different fast-charging protocols to better align with real-world application scenarios. Second, to make full use of frequency domain information, the time-frequency feature maps are generated through continuous wavelet transformation. Then, spatial-temporal information is mapped to battery state-of-health and cycle-to-knee through a convolutional neural network and bidirectional long short-term memory network sequentially. The fusion of spatial-temporal features and the organization done by the attention mechanism contribute to improving battery PHM accuracy. Finally, experiments conducted on LFP/graphite A123 batteries under different fast-charging protocols indicate the effectiveness and superiority of the proposed method. In addition, the ablation experiments are carried out to demonstrate the necessity of each model component. Experimental results show that using time-frequency feature images significantly enhances accuracy, and each component plays a pivotal role in enhancing the overall performance.