Realistic fault detection of li-ion battery via dynamical deep learning

The study addresses the challenge of accurately and proactively detecting faults in lithium-ion batteries used in electric vehicles. EV batteries are complex nonlinear systems subject to varied failure mechanisms (short circuits, physical damage, over/over-discharge, thermal abuse, etc.), making algorithmic detection difficult. Prior approaches have often been validated only in small-scale lab settings and rely on information unavailable in real-world deployments. The authors aim to build and validate a practical, scalable deep-learning framework using large-scale, real-world EV charging data while accounting for practical social and financial considerations (data availability, inspection costs, fault rates, privacy, and sensor noise). The purpose is to demonstrate that deep learning, when tailored to the dynamical nature of LiB systems and configured using economic statistics, can improve safety by enabling early, cost-effective fault detection.

Existing research spans physics-based and data-driven approaches. Physics-based methods involve parameter identification and modeling (e.g., open-circuit voltage, internal resistance, impedance-based diagnosis, thermal modeling, and safety improvements through materials and separators). Data-driven studies include LSTM-based voltage anomaly prognosis, statistical distribution methods, mutual information for micro-short-circuit identification, and broader big-data fault diagnostics. However, translation to real-world EV settings is hindered by small-scale validations, reliance on unavailable parameters, and lack of large public datasets. The authors note the success of AI in other domains driven by large, real-world datasets and emphasize the need for similar datasets and deployment-aware methods in EV battery safety. They underline gaps: sparse vehicle-level labels, absence of snippet-level ground truth, practical constraints (privacy, communication efficiency), and the need to incorporate social and financial factors into model configuration.

Datasets and labels: The authors release three EV charging datasets (vehicles from three anonymized manufacturers) with over 690,000 charging snippets from 347 EVs. Each snippet includes time series of current, voltage, and temperature. Vehicle-level fault labels are derived from driver reports and engineer confirmations (e.g., lithium plating, low range, over-high temperature, abnormal voltage). Data at or near failure events are removed to ensure models must predict issues days in advance. Model: Dynamical autoencoder for anomaly detection (DyAD). The approach treats the battery as a dynamical system. Inputs (system controls) are state of charge (SOC) and current; responses are voltage and temperature (e.g., min/max/avg voltage, min/max temperature). The encoder maps input-output pairs to a latent variable representing system parameters; the decoder predicts system response from input using these parameters. Anomalies are detected via reconstruction error between predicted and observed responses. The formulation is motivated by viewing battery behavior with a hidden Markov model perspective and partitions features into inputs and responses, unlike generic anomaly models that treat dimensions equally. Losses: Total training loss combines (1) reconstruction loss: MSE between decoder output and observed response; (2) latent regularization via KL-style regularization to avoid overfitting by structuring the latent distribution; and (3) mileage supervision loss that encourages the representation to preserve mileage information via an MLP head, providing weak supervision and improving anomaly detection performance. Robust scoring from snippets to vehicles: Because labels are sparse and at vehicle level, a robust aggregation converts snippet-level errors to vehicle-level scores. First, threshold snippet reconstruction errors at t to mark abnormal snippets; then compute the average over the top p percentile errors of a vehicle; if this averaged top-p error exceeds threshold t (or a tuned vehicle-level threshold), the vehicle is flagged abnormal. Hyperparameters t and p are tuned on the training set. Training details: Encoder and decoder are implemented with GRU-based recurrent networks (3 layers, 32 hidden units per layer). Latent space dimension is 32. Optimization uses Adam with learning rate 0.001 and minibatch size 128. Deployment and privacy: The encoder runs at charging stations to produce privacy-friendly encoded representations; the cloud-based module performs reconstruction and fault scoring. This split preserves customer privacy (e.g., mileage, time, location), reduces communication costs by transmitting encoded features, and protects model IP. Economic configuration: Social and financial statistics parameterize decision thresholds to minimize expected direct costs. The expected cost function depends on battery fault rate p, fault cost c_f, inspection cost c_r, true positive rate, and false positive rate. Empirical ranges used: p from 0.038% to 0.075% (based on 1.2 million EVs), c_f from 1–5 million CNY per vehicle, c_r from 8–55 thousand CNY per vehicle. Model operating points along the ROC are selected to minimize total expected direct costs. Baselines: The study compares DyAD against graph deviation network (GDN), vanilla autoencoder (AE), support vector data description (SVDD), Gaussian process (GP), and a variation evaluation (VE) method. Five-fold cross-validation is used; AUROC is the primary metric. Interpretability is explored by visualizing embedding evolution (input, latent, output error spaces) via t-SNE, showing separation of abnormal versus normal snippets in the output-error space.

- DyAD outperforms baselines with an AUROC of 88.6 ± 2.9% across datasets. Baseline AUROCs: GDN 70.3 ± 5.5%, AE 72.8 ± 13.4%, SVDD 51.5 ± 8.26%, GP 66.6%, VE 55.6%. - DyAD provides a dominant average ROC curve and 16–33% AUROC boost over state-of-the-art deep learning baselines; auxiliary losses (KL regularization and mileage supervision) further enhance performance. - Economic impact: Using empirical ranges (fault rate 0.038%–0.075%, fault cost 1–5 million CNY/vehicle, inspection cost 8–55 thousand CNY/vehicle), DyAD reduces the expected direct costs by about 33% versus deep learning baselines and 50% versus VE. Average direct costs (million CNY/vehicle/year): DyAD 0.085, GDN 0.126, AE 0.133, SVDD 0.152, GP 0.162, VE 0.169. - Data scale and realism: Evaluation on three real-world EV datasets totaling over 690,000 charging snippets from 347 EVs (55 abnormal, 292 normal). The method detects anomalies without relying on unavailable parameters (e.g., OCV, internal resistance) and is robust to diverse charging patterns, mitigating false alarms on rare but normal inputs. - Interpretability: t-SNE visualizations show that while abnormal and normal snippets overlap in input and latent spaces, they separate in the output error space, supporting reconstruction error as a discriminative feature. - Deployment benefits: The encoder-at-edge, cloud-decoder design is privacy-friendly, communication-efficient, and suitable for multi-party collaboration.

The findings demonstrate that tailoring deep learning to the dynamical nature of EV LiB systems and incorporating social and financial statistics for operating-point selection yields superior fault detection and lower costs in realistic settings. By modeling input-to-response mappings, DyAD avoids misclassifying rare but normal charging behaviors, addressing a key limitation of distribution-based anomaly detectors. The robust scoring bridges snippet-level predictions to vehicle-level decisions under sparse labels. Economically optimal operating points, selected along the ROC using fault rates and cost statistics, substantially lower expected direct costs relative to baselines. The deployment design preserves privacy and reduces communication, facilitating real-world adoption. Interpretability analyses indicate that error-based features align with abnormal behavior, offering actionable insights for maintenance and potential manufacturing guidance.

This work introduces DyAD, a dynamical autoencoder tailored to EV LiB anomaly detection, and releases large-scale, real-world datasets enabling rigorous benchmarking. DyAD achieves state-of-the-art accuracy, delivers a dominant ROC, and reduces expected direct fault and inspection costs by 33–50% compared with baselines. The framework integrates social and financial statistics for economically optimal configuration and supports privacy-preserving deployment. Future directions include: (1) enhancing interpretability by aligning learned representations with physics-informed battery parameters (e.g., capacity, internal resistance); (2) quantifying forecast horizons to determine how far in advance faults can be predicted; (3) adapting models to regional, chemistry, and manufacturer heterogeneity; and (4) extending the framework to other dynamical-system anomaly detection tasks such as photovoltaics, robotic navigation, water treatment testbeds, and spacecraft.

- Labels are sparse and at vehicle level; snippet-level ground truth is unavailable. Robust scoring mitigates but does not eliminate this limitation. - Data near or at failure were removed, so the model must infer issues from earlier signals; this complicates precise forecast horizon estimation. - Some physics parameters (e.g., open-circuit voltage, internal resistance) are often unavailable in real-world data; while the model does not require them, incorporating estimates could improve performance and interpretability. - Economic configuration relies on regional statistics (fault rates, costs) that may vary; indirect costs (e.g., reputational, sales impacts) are not fully quantified. - Raw EV data cannot be shared due to privacy laws; only processed datasets are available, which may limit external replication of certain analyses.