Cross-Scenario Multi-Type Fault Diagnosis for Lithium-Ion Batteries via Deep Transfer Learning

Faults in lithium-ion batteries, such as low capacity (LowC) and micro short-circuits (MSC), can progressively deteriorate system reliability and may eventually trigger catastrophic safety incidents. Although existing diagnostic methods show promising performance in controlled laboratory settings, their reliance on artificially simulated data and their focus on single-scenario or single-fault conditions significantly limit their applicability in real-world deployments. To address these challenges, this paper proposes a lightweight neural network framework for multi-fault diagnosis with cross-scenario adaptability, enabling effective knowledge transfer from electric vehicles (EVs) to industrial and commercial energy storage systems (ICES). First, an effective feature extraction method, termed the 10-bin probability distribution matrix (10-BPDM), is introduced to compactly encode fault-relevant temporal–statistical characteristics. Then, a depthwise separable convolutional network integrated with a hybrid attention mechanism is designed to accurately detect and classify LowC and MSC faults. Extensive experiments conducted on real-world battery datasets demonstrate that the proposed method achieves accurate multi-fault diagnosis while maintaining strong cross-scenario transferability. In particular, by fine-tuning a model pre-trained on EV data using only a small number of target samples, a fault detection rate of 98.25% is achieved on the ICES dataset, confirming both the effectiveness and deployment feasibility of the proposed lightweight framework.