Lightweight fault diagnosis for EV battery packs via SpikingFormer and frequency slice wavelet transform

Abstract

Accurate fault diagnosis of power batteries is crucial for ensuring the safe and reliable operation of electric vehicles (EVs). Existing fault diagnosis methods have increasingly adopted deep neural networks due to their powerful learning and feature extraction capabilities. However, two significant limitations remain. Firstly, these methods fail to exploit the time–frequency coupling characteristics from multiple battery operational signals, leading to suboptimal feature representation. Secondly, the employed deep network models, such as Transformers, often require substantial computational resources, making them unsuitable for real-time deployment. To address these challenges, this paper proposes a novel fault diagnosis framework that integrates frequency slice wavelet transform (FSWT) with a lightweight SpikingFormer architecture. FSWT is employed to decompose and analyze multiple raw battery signals, capturing comprehensive time–frequency domain features that enhance fault representation. SpikingFormer, inspired by spiking neural networks, serves as an efficient alternative to the Transformer model, reducing computational complexity through event-driven processing while maintaining its capability to capture long-term dependencies. The proposed method, validated using real-world EV battery datasets collected from 100 EVs over a period of 6 to 12 months, demonstrates superior performance compared to state-of-the-art (SOTA) techniques. Specifically, it achieves a 4%–6.8% increase in mean fault-diagnosis accuracy and reduces the time-to-fault error by 1.2–3.2 min. Moreover, its inference time accounts for only 2.8%–28.4% of that required by SOTA methods, while its energy consumption is limited to 13.3%–14.4% of their levels.