The transition to net-zero emissions requires transportation electrification and decarbonization, with electric vehicles (EVs) using lithium-ion batteries (LiBs) at the forefront. A major concern is battery safety; EV fires are longer-lasting and less predictable than gasoline car fires, leading to costly inspections. Early prediction of battery failures is vital to reduce social costs and enhance EV adoption. However, EV batteries are complex nonlinear systems, making it challenging to design algorithms that accurately understand their failure mechanisms (short circuit, physical damage, overcharge/overdischarge, thermal abuse, etc.). Existing research on battery safety has explored both physics-based and data-driven approaches. However, these methods face limitations. Firstly, validation often occurs in small-scale experimental settings, unlike successful data science applications in other fields that rely on large-scale, real-world datasets. Secondly, many algorithms rely on information unavailable in real-world settings. A practical EV battery fault detection algorithm must consider factors like data availability, economic trade-offs, sensor noise, and model privacy. This study addresses these shortcomings by releasing three EV charging datasets (over 690,000 charging snippets from 347 EVs) to benchmark existing deep learning models and conventional data-driven approaches. While these models show some detection power, they incur high economic costs (around 10³ CNY per vehicle). To overcome these limitations, a new deep learning model, the dynamical autoencoder for anomaly detection (DyAD), is developed. DyAD is designed for large-scale real-world EV LiB data, exploiting the hidden Markov model of battery data and balancing accident and inspection costs. It differs from existing models by using a dynamical system formulation (partitioning data features into system inputs and responses) and exploiting the structure in EV fault labels.

The paper reviews existing literature on battery safety, categorizing approaches into physics-based and data-driven methods. Physics-based methods leverage the underlying physical and chemical processes within the battery to model and predict failures. Data-driven approaches, on the other hand, utilize machine learning techniques to analyze historical data and identify patterns associated with battery failures. The authors highlight the limitations of both approaches when applied to real-world scenarios. Physics-based models often lack sufficient accuracy in capturing the complex interactions within a battery, while data-driven models often require extensive data and may not generalize well to new scenarios. The lack of large-scale, real-world datasets for validating these models is also pointed out as a major impediment to progress in the field. The authors emphasize the need for a data-driven model that accounts for practical limitations of real-world scenarios such as the availability of data, cost constraints, and data privacy.

The researchers created three large-scale datasets from an EV data platform, containing over 690,000 charging snippets from 347 EVs. The datasets include time series data (current, voltage, temperature) for each charging event and vehicle-level fault labels from driver reports and engineer confirmations. The challenge of accurately detecting faults was demonstrated by showing that traditional methods based on simple features such as voltage variance yield poor results. This motivated the development of a novel deep learning framework. The proposed DyAD model utilizes a dynamical autoencoder architecture. This architecture differs from traditional autoencoders in that it explicitly models the input-output relationship of the battery system. The encoder part of the model takes as input the system's input variables (state of charge (SOC) and current) and outputs (voltage and temperature) and encodes them into a latent representation. The decoder then reconstructs the system's output variables from the latent representation and the system's input variables. The model learns to reconstruct normal battery behavior, and anomalies are detected as significant deviations between the reconstructed and observed outputs. The model incorporates a robust scoring procedure to aggregate predictions from individual charging snippets to generate vehicle-level predictions, addressing the sparsity of vehicle-level labels. This procedure involves thresholding reconstruction errors and averaging the top percentile of errors to determine if a vehicle is abnormal. The training process minimizes a combined loss function comprising reconstruction loss, KL divergence for regularization, and mileage supervision. The DyAD model's deployment is designed for privacy preservation: encoders are deployed at charging stations, while the fault detection module is cloud-based. This design safeguards sensitive data while reducing data communication. The model was benchmarked against various state-of-the-art algorithms (GDN, AE, SVDD, GP, VE) using the three datasets. The performance was evaluated using the area under the receiver operating characteristic curve (AUROC) and the expected direct cost of battery faults and inspections.

The DyAD model significantly outperforms state-of-the-art algorithms in detecting LiB anomalies. It achieved a 16-33% AUROC boost compared to baselines. The auxiliary loss terms further improved performance. The improved detection translates to substantial economic benefits, reducing expected direct costs of LiB faults and inspections by 33-50%. This cost reduction is based on analysis considering the range of LiB fault rates, fault costs, and inspection costs from real-world data. Visualization of the model's internal representations using t-distributed stochastic neighbor embedding (t-SNE) shows that the DyAD effectively separates normal and abnormal charging snippets in the output layer, indicating that the model learns meaningful representations of battery behavior. The model's ability to effectively cluster abnormal snippets is also highlighted through analysis of specific examples, showing that larger prediction errors accurately identify abnormal charging snippets. Table 1 summarizes the average AUROC and average direct costs for each algorithm.

The study demonstrates the effectiveness of the DyAD model for realistic EV LiB fault detection, showcasing the potential of dynamical deep learning for improving battery safety and reducing economic costs. The model's ability to leverage limited anomaly samples is a significant advantage in real-world scenarios. The integration of social and financial factors into the model's design is a key contribution, highlighting the importance of considering practical constraints in developing AI solutions for real-world problems. The privacy-preserving design of the model, utilizing separate encoders and decoders, allows for wider adoption while ensuring data security. The findings emphasize the need for collaborations between AI and energy communities to address the safety challenges posed by EV batteries.

This research presents a novel dynamical deep learning approach, DyAD, for LiB fault detection in EVs, coupled with the release of a large-scale real-world dataset. DyAD surpasses existing methods in accuracy and cost-effectiveness. Future work could explore the incorporation of additional battery parameters, improve model interpretability, and refine cost models to further enhance the practical applicability of such methods. The framework presented is promising for broader applications in other dynamical systems.

While the DyAD model shows significant improvement, several limitations exist. The study relies on data from a specific set of EV manufacturers; generalizability to other battery chemistries and vehicle types needs further investigation. Indirect costs associated with battery failures (e.g., reputational damage) were not explicitly quantified in the cost analysis. The interpretation of the model's internal representations is still limited, requiring further research to gain deeper insights into the learned patterns.