<p>Data and code.</p>

To address these issues, this paper presents an unsupervised deep learning framework for anomaly detection. In the framework, a hybrid LSTM-Autoencoder architecture is used to integrate the temporal feature extraction capability of long short-term memory networks with the reconstruction mechanism of Autoencoders, thus markedly improving the temporal modeling of battery data.

Since the detection performance is highly dependent on the selection of reconstruction error metrics, multiple such metrics are employed in the model, and are systematically evaluated and compared to determine the most suitable metric for battery anomaly detection. The model is trained separately for static and driving conditions based on corresponding subsets of the battery datasets. With a fixed model structure and experimental conditions, model performance is evaluated using various reconstruction error metrics, and Bayesian optimization is used to identify the optimal detection threshold for each metric.