Abstract: Abstract: To address the challenges of high complexity in multi-source heterogeneous time-series data, scarcity of anomaly samples, and strong inter-variable dependencies in lithium-ion battery fault detection for electric vehicles, this paper proposes a prediction-evaluation framework based on dynamic autoencoders and attention mechanisms. First, a joint feature encoder is designed by integrating a dynamic autoencoder with gated recurrent units (GRUs) to perform feature fusion and dimensionality reduction on multi-source time-series data, thereby extracting deep cross-modal representations. In parallel, a pre-response encoder based on self-attention is constructed to capture long-term temporal dependencies, enhancing the efficiency and accuracy of feature extraction. Furthermore, a memory parsing module is introduced, which leverages residual contrastive learning to align the prediction path with the actual response path, thereby improving the model's capability to detect anomalous patterns. On this basis, an evaluation model is developed using a synergistic anomaly detection algorithm, which analyzes reconstruction error distributions to generate anomaly scores and performs threshold traversal and ROC-based assessment to robustly distinguish between normal and faulty states. Operating under an unsupervised learning paradigm, the proposed framework effectively extracts key response patterns and identifies latent anomalies from complex multi-source data. Experimental results on multiple lithium-ion battery datasets demonstrate that the proposed framework significantly outperforms baseline methods such as AE, DeepSVDD, and GDN in terms of detection accuracy and model robustness. Notably, compared to the DyAD model, the proposed DTMAD achieves an AUROC improvement to 0.901 (an increase of 0.032), while reducing performance variance from 0.029 to 0.026, indicating superior detection stability and generalization capability.

Key words: Keywords: lithium-ion battery, fault detection, federated learning, feature fusion, variational autoencoder

摘要： 摘 要: 针对电动汽车锂离子电池故障检测中多源异构时序数据复杂性高、异常样本稀缺及多变量关联性强的挑战，提出了一种基于动态自编码器与注意力机制的预测-评估框架，以提升故障识别准确性与模型泛化能力。首先，设计了融合动态自编码器与门控循环单元的联合特征编码器，对多源时序数据进行特征融合与降维处理，提取跨模态深层特征表示；同时，构建基于自注意力机制的预响应编码器，捕捉时序数据中的长期依赖关系，提升特征提取效率与精度；进一步引入记忆解析模块，通过残差对比学习机制融合预测路径与实际响应路径，增强异常模式检测能力。其次，基于重构误差的分布特性，通过协调异常检测算法，设计了评估模型。最终，通过综合的预测-评估框架，在无监督学习条件下从多源数据中提取关键响应模式并识别潜在异常。实验结果表明，所提框架在多组多源锂离子电池数据集上相较于AE、DeepSVDD、GDN在故障检测准确率和模型稳定性方面具有明显优势。其中，相较于基准模型DyAD，DTMAD模型的AUROC提升至0.901，提高了0.032，且结果波动幅度由0.029降至0.026，展现出更高的检测稳定性与泛化能力。

关键词: 锂离子电池, 故障检测, 联合学习, 特征融合, 变分自编码器