Review of Fault Detection Approaches for Large-Scale Lithium-Ion Battery Systems: A Spatio-Temporal Perspective

Abstract

Battery fault detection is crucial for maintaining the safety and reliability of large-scale lithium-ion battery systems, especially in demanding applications like electric vehicles and energy storage power stations. However, existing research primarily addresses either temporal patterns or spatial variations in isolation. This paper presents a comprehensive review of fault detection from a spatio-temporal perspective, with a specific focus on AI-driven methods that integrate temporal dynamics with spatial sensor data. The contributions of this review include an in-depth analysis of advanced techniques such as transfer learning, foundation models, and physics-informed neural networks, emphasizing their potential for modeling complex spatio-temporal dependencies. On the engineering side, this review surveys the practical application of these methods for early fault detection and diagnostics in large-scale battery systems, supported by case studies and real-world deployment examples. The findings of this review provide a unified perspective to guide the development of robust and scalable spatio-temporal fault detection methods for EV batteries, highlighting key challenges, promising solutions, and future research directions.

1. Introduction

Electric vehicles (EVs) are revolutionizing transportation, offering a cleaner and more energy-efficient form of transport compared to conventional fuel-based vehicles. The driving force behind this new wave is lithium-ion batteries, which have emerged as the go-to energy storage solution based on their high energy capacity, lightweight design, and long lifespan of operation [1]. EVs enabled by these batteries can cover long distances with consistent performance and reliability. Relative to traditional battery technology, Li-ion enables the longer life, which makes it best suited to the constant usage required by EVs. With such advantages, EV penetration is spreading fast, making them a central mover of future sustainable mobility [2].

Battery management systems (BMSs) form the core of battery safety, efficiency, and lifespan. Figure 1 illustrates the operational architecture of a grid-connected large-scale energy storage system (ESS). Central to the system are battery racks, which store energy and provide real-time voltage, current, and temperature data to sensors. These sensors relay information to both the energy management system and safety management system. The EMS directs energy flow between the battery racks, grid, load, and charging equipment and shares data with the spatio-temporal monitoring layer, which uses AI models for advanced fault detection. The thermal management system maintains safe temperature conditions by interacting with both the grid and the sensors. The supervisory control and data acquisition (SCADA) and data communication system provides centralized monitoring and control of the ESS [3].

Electrification of transport has brought Li-ion batteries into the spotlight, and their safety and reliability are now the highest priority [4]. Prominent EV battery failures, from spontaneous fires to sudden capacity loss, underscore the need for effective fault detection in a timely manner [5]. Failure of a single cell can cascade to thermal runaway in which the cell temperature increases exponentially, potentially causing adjacent cells to combust and consume the entire pack in flames [6,7]. These incidents can be life-threatening to passengers and result in the burning of vehicles within minutes. Thus, the ability to detect faults in real time and mitigate them before they induce more severe effects is crucial for consumer safety and battery system lifespan. Despite numerous studies on the subject, protecting millions of automotive battery cells across their operational life cycle still poses an open challenge [8].

Traditional BMSs monitor parameters like voltage, current, and temperature of each cell. They rely on models and threshold-based checks like overvoltage or over-temperature limits to flag abnormalities. While useful, these conventional methods are often reactive and may only detect faults at advanced stages [9]. Moreover, many existing diagnostic techniques analyze signals in a single domain either over time or across the pack but not both. Some methods monitor each cell’s voltage trajectory independently over time to identify deviations through temporal analysis, while others compare differences between cells at a given time instance to identify abnormal behavior through spatial analysis [10]. Such one-dimensional approaches can overlook complex fault patterns that involve small changes over time and across multiple cells simultaneously.

Real-time and accurate fault detection is essential to avoid battery malfunctions. Key materials and manufacturing breakthroughs have accelerated Li-ion adoption for both transport and grid applications [11]. Early detection allows the BMS to perform corrective measures, such as balancing cell voltages, reducing the power demand, or disconnecting defective cells to prevent hazards [12]. This becomes particularly important during dynamic operational scenarios. Large-scale battery systems encounter fluctuating loads, temperatures, and ambient conditions, which can mask or imitate fault signatures. Recent years have seen frequent battery failures leading to recalls, raising safety concerns and economic costs [13]. Technologies for early warning and intervention are therefore in high demand. Numerous approaches have been explored for battery fault detection, broadly categorized into model-based methods, signal processing methods, and algorithms guided by data. Model-based methods use fundamental principles or circuit-based models of the battery, and a fault is identified when actual behavior differs from the model’s prediction beyond a set limit [14]. However, these require accurate modeling of battery dynamics and often struggle as batteries age or face unexpected conditions. Signal processing and statistical methods detect anomalies by examining measured data for out-of-range values, quick changes, or abnormal correlations. They are fast and interpretable but may fail to detect incipient faults buried in noise [15,16]. Each category has limitations; no single technique currently meets all the requirements of accuracy, speed, robustness, and generality needed for battery fault detection in the field.

In recent years, battery fault detection has gained significant attention due to the increasing reliance for Li-ion batteries used in EVs and energy storage systems, supported by the increasing number of related studies found in the Web of Science database identified through search queries such as “battery fault detection,” “battery fault diagnosis,” and “battery health” in article titles and abstracts. The need for more reliable, efficient, and predictive fault detection methods has led to a surge in research contributions, particularly in AI-driven diagnostics and spatio-temporal modeling approaches. This growing research interest is reflected in the rising number of research papers, as shown in Figure 2, which demonstrates the field’s quick development and rising importance in the industry [17].

In 2025, the field of battery fault detection research has seen a major rise in publications, showing steady annual growth fueled by improvements in AI-based diagnostics and the growing use of EVs. Compared to previous years, research output has nearly tripled since 2015, highlighting the growing industry and academic commitment to improving battery reliability, efficiency, and safety. This trend aligns with fast growth in EV markets and energy storage solutions, where innovations in ML, DL, and spatio-temporal modeling have significantly enhanced fault prediction accuracy and real-time monitoring capabilities. The increasing research interest underscores the critical role of predictive maintenance and intelligent BMSs in ensuring longer battery life cycles, reduced operational risks, and greater sustainability in energy storage applications.

Prior survey papers and reviews have addressed fault detection but typically from a general perspective. For instance, Zhang et al. (2023) [18] proposed a cloud energy storage model to improve energy storage utilization through shared and aggregated ESS platforms. However, the study does not consider fault detection or any diagnostic strategies within the cloud framework. Our work addresses this gap by integrating spatio-temporal fault modeling into the CES to improve system safety and reliability. Hu et al. (2020) [19] reviewed various fault types in Li-ion battery systems, including sensor and actuator faults, and discussed model-based diagnostic approaches. The study lacks a focus on spatial system structure and does not utilize digital twins for fault visualization. In our research, we model faults across all levels, cell, module, and pack, using digital twin technology for enhanced spatial insight. Adasah et al. (2024) [20] examined fault diagnosis using protection mechanisms and sensor data, focusing mainly on feature extraction and traditional models. However, the study does not include predictive analytics or cloud integration for large-scale systems. We enhance this by applying time-series prediction techniques within a cloud-based monitoring setup for proactive fault detection. Li et al. (2024) [21] reviewed station-level battery energy storage systems, discussing system topologies and fault diagnosis methods. While comprehensive at the system level, the study does not address internal fault progression or apply AI techniques. Our work models internal fault evolution using deep learning and links it to external system-level symptoms.

Hong et al. (2023) [22] introduced digital twin and big data technologies to improve energy storage system control and visualization. Despite its innovation, the study lacks a focus on fault types and their spatial propagation. We expand this by using digital twins specifically to track, classify, and simulate faults across time and battery system layers. Song et al. (2024) [23] presented a fault evolution mechanism using Failure Modes, Mechanisms, and Effects Analysis (FMMEA) to understand fault behavior across different battery levels. However, the study does not integrate real-time monitoring or cloud-based diagnostics. Our research integrates predictive algorithms for early fault warning. Nazaralizadeh et al. (2024) [24] reviewed battery health metrics such as SOH, SOC, and RUL using machine learning for battery life prediction. The study does not address spatial modeling or real-time multi-level system diagnostics. We fill this gap by linking health indicators to spatio-temporal fault detection in large-scale Li-ion battery systems. Kumar et al. (2025) [25] reviewed AI-based PHM methods for lithium-ion batteries, focusing on data acquisition, feature extraction, and SOH/RUL prediction using ML and DL models. However, it overlooked real-time fault detection and spatial–temporal fault behavior. We extend this by addressing fault detection across battery modules through a spatio-temporal framework for improved reliability.

While previous studies have explored spatial or temporal modeling in isolation, a unified spatio-temporal perspective for battery fault detection remains largely underexplored. This review uniquely positions spatio-temporal modeling as a comprehensive framework, emphasizing its potential to enhance early and accurate fault detection. This critically examines cutting-edge approaches including spatio-temporal autoencoders, transformer-based attention models, and hybrid architectures that jointly learn spatial and temporal dependencies. Crucially, this review combines recent advancements such as GNNs, physics-informed neural networks (PINNs), large language models (LLMs), and self-attention mechanisms specifically designed for battery systems, which have not been collectively analyzed in the prior literature. Moreover, this study offers a comparative assessment of these methods based on dataset sources, evaluation metrics, and their practical applicability, offering a valuable resource for advancing intelligent BMSs.

Table 1. A comparative analysis of recent review studies on battery fault detection: contributions, limitations, and our advancements.

Study	Contribution	Missing Aspects	Our Contributions
Zhang et al. [18]	Proposed cloud energy storage model for shared ESS operation	Lacks fault detection or safety mechanism integration	We incorporate spatio-temporal fault modeling into cloud BMS frameworks
Hu et al. [19]	Reviewed battery, sensor, and actuator faults with modeling techniques	No focus on system-level spatial structure or digital twin use	We address faults at all levels using digital twins and spatial mapping
Adasah et al. [20]	Analyzed protection-based fault diagnosis with sensor data	Limited attention to cloud-based or temporal prediction systems	We integrate prediction-based models using time-series fault patterns
Li et al. [21]	Reviewed topologies and diagnosis methods for storage stations	Does not detail internal fault progression or AI methods	We model internal fault evolution using AI and spatio-temporal layers
Hong et al. [22]	Introduced big data and digital twins for energy storage management	Lacks specific focus on fault types and fault propagation	We apply digital twins to trace multi-level fault evolution
Song et al. [23]	Modeled multi-level fault evolution with failure analysis	No integration of cloud or data-driven systems	We unify FMMEA with cloud-based monitoring and prediction
Nazaralizadeh et al. [24]	Reviewed battery health metrics and machine learning for SOH	Does not address spatial system structure or real-time detection	We combine SOH tracking with real-time spatial fault diagnosis
Kumar et al. [25]	Reviewed AI-based PHM for Li-ion batteries using ML and DL models (GPR, CNN, LSTM) for SOH/RUL prediction	Did not address real-time fault detection in large systems with a spatio-temporal perspective	Focus on fault detection framework to enhance safety and scalability in large-scale battery systems

provides a comparative analysis of existing studies, highlighting their contributions and identifying gaps.

The main contributions of this study are summarized as follows:

• It categorizes common EV battery faults, such as capacity fade and thermal imbalance, based on their causes and manifestations, distinguishing between temporal patterns (e.g., degradation trends) and spatial patterns (e.g., cell-level inconsistencies) to guide spatio-temporal modeling.
• It systematically analyzes state-of-the-art AI techniques, including transfer learning, foundation models, PINNs, and hybrid DL architectures, highlighting their effectiveness in capturing complex spatio-temporal dependencies.
• It identifies and discusses key challenges, such as scarcity of labeled fault data, computing limitations in real-time BMS, poor model explainability, and issues with generalization across battery chemistries.

The structure of this paper is as follows: Section 2 explains the fault mechanisms in Li-ion batteries for EVs, emphasizing the main fault types and their impacts. Section 3 delivers a detailed overview of spatio-temporal fault detection methods, describing their ability to link spatial and temporal factors and analyzing modeling techniques that effectively study battery degradation. Section 4 covers key challenges such as data shortages, computational limits, and the interpretability of AI-based models. Section 5 outlines upcoming research directions like GNNs, LLMs, self-supervised learning, and edge computing, providing a clear roadmap for future progress. Section 6 gives this paper’s conclusion.

2. Spatio-Temporal Characterization of Fault Mechanisms in ESSs

Li-ion batteries are the core energy source in modern energy storage systems, yet their performance and safety are highly exposed to various fault mechanisms. These faults, ranging from overvoltage and thermal runaway to internal short circuits, can evolve silently over time or appear suddenly due to operational stress, aging, or manufacturing errors. Understanding these failure modes is essential not only for preventing safety hazards but also for improving battery reliability, efficiency, and lifespan [26]. This section provides a foundational overview of common fault types in EV battery systems, highlighting their physical causes, their functional impact, and the need for early, intelligent fault detection and the spatio-temporal characteristics of battery data that motivate integrated analysis [27].

Figure 3 shows the functional layout of a BMS, which combines several essential operations to maintain battery safety and performance. It highlights key functions, such as data collection, fault detection, thermal control, charge and discharge regulation, and state estimation. The figure also illustrates how sensors, signal processors, control algorithms, and actuators work together to monitor and protect the battery system in real time, reflecting the overall structure and main responsibilities of a modern BMS used in EVs.

2.1. Major Failure Modes in Li-Ion Batteries: Categories, Causes, and Their Consequences

Broadly, battery faults are categorized as cell faults or system faults. Cell faults occur within the battery cells themselves and are usually the most critical for safety. They can be further divided into spatial faults and temporal faults [28].

2.1.1. Spatial Fault Characteristics in Battery Systems

Spatial faults are those that take place at specific locations in the battery system, typically at the level of modules or individual cells. These faults are typically identified by making comparisons of voltage, temperature, or current in multiple cells at a single point in time [29]. One of the most common spatial faults is cell imbalance, which occurs when certain cells possess a higher or lower state of charge compared to the others. This imbalance is most frequently the result of unequal aging, manufacturing variability, or non-uniform thermal conditions and can lead to overcharging or over-discharging of weak cells [30]. A thermal hotspot is another significant spatial fault and can be due to internal resistance, poor heat release, or external factors like exposure to high surrounding temperatures. If left unaddressed, it can lead to thermal stress or reduced capacity, and it can even induce other safety risks in that particular section of the battery pack [31].

Furthermore, internal short circuits, at least in their early stages, are spatial in nature, occurring within a given cell as a result of physical damage, separator breakdown, or dendrite growth [32]. Figure 4 illustrates the typical factors that trigger short circuits in Li-ion batteries. Manufacturing defects are also spatial faults, since they will only impact specific cells as a result of such defects as misalignment, material anomalies, or contamination during assembly. These faults are not identifiable during initial operation but can result in performance irregularities or premature failure [33]. Lastly, sensor failure that results in inaccurate reading for a particular cell also qualifies as a spatial fault. It will cause misleading control action by the BMS if not efficiently detected and isolated. Precise cell-level measurement and comparative analysis are required to identify spatial faults and calculate deviations throughout the pack [34].

2.1.2. Temporal Fault Characteristics in Battery Systems

Temporal faults develop over time and are typically associated with performance degradation, cyclic stress, or continuous exposure to environmental and operating conditions. Temporal faults do not reveal themselves in the form of spot measurements but as trends when battery parameters are monitored continuously over numerous charge and discharge cycles [35]. One of the most hazardous temporal faults is overcharge, in which a cell is charged continuously beyond its voltage limit. It leads to side reactions such as lithium plating or SEI (solid electrolyte interphase)-layer breakdown, which induces long-term degradation [10]. Similarly, over-discharge occurs when a cell is discharged below its minimum recommended level, which facilitates copper dissolution, structural damage, and eventual failure. These changes are not usually visible in a single cycle but become noticeable as they accumulate with each passing day and affect the electrochemical stability of the cell [36].

The remaining temporal faults are capacity fade and resistance growth, both of which characterize the normal aging of the battery. Capacity fade refers to the loss of charge-holding ability of the battery over time, while resistance growth characterizes the increasing internal resistance that reduces the battery’s power output [37]. Voltage drift is a temporal fault where the voltage response of the cell begins to deviate slowly from normal operation, although the system is initially discovered to be normal. Abnormal self-discharge, where the battery is losing energy at a greater than usual rate when it is in its resting state, is also a temporal fault [38,39]. All of these issues require long-term monitoring and analysis of trends in history. Temporal fault detection involves tracking gradual deviations of key parameters and employing predictive mechanisms capable of tracking aging, degradation, and repetitive patterns of stress over time [40]. Figure 5 illustrates the TR mechanism, emphasizing major factors such as electrical abuse like overcharging or over-discharging, mechanical abuse such as puncturing or crushing, and exposure to elevated temperatures. This visual representation underscores the interconnected nature of these factors and their collective role in initiating TR events [41]. All-solid-state batteries (ASSBs) and conventional liquid-electrolyte Li-ion cells both exhibit rapid temperature rise under severe abuse, yet the underlying physical pathways differ. ASSBs remove flammable liquid electrolytes, reducing vapor-mediated ignition, but can still undergo exothermic interface-level or cathode decomposition reactions and mechanical failure that produce rapid localized heating. Consequently, while qualitative detection goals and early localization and propagation prediction remain similar, sensor placement, thermal models, and the signature features used for prognostics must be adapted to ASSB-specific modes like boundary-layer heating and solid-state reaction fronts.

2.1.3. Integrated Spatio-Temporal Fault Propagation

Spatio-temporal faults in Li-ion batteries are failure modes that begin locally but increasingly evolve or spread over time and impact more than one cell or more than one module [42]. Such types of faults require spatial as well as temporal examination to detect and predict correctly because they do not remain localized in one area or at a single instance of time. One of the priority issues is thermal runaway propagation, which starts as localized overheating of a single cell and then spreads quickly to neighboring cells through heat transfer and exothermic reaction [23]. This propagation leads to a chain reaction that compromises the entire battery pack and can be an extremely dangerous safety risk if detected late. Since the fault begins in a location but evolves through time, it cannot be meaningfully understood or managed without considering both spatial place and temporal evolution [43]. Moreover, we introduce the classification of typical Li-ion battery faults, differentiating between cell-level faults and system-level faults in Figure 6.

Yet another is lithium plating, which begins as a gradual chemical process due to conditions like low-temperature charging or long-term overcharging. Initially, it will not show any symptoms, but plated lithium can actually end up growing dendrites and hence have localized internal short circuits [44]. Progressive imbalance is another error where small disparities among the cells keep growing bigger with every cycle until the point of performance degradation and safety concerns across the pack [45]. Propagation caused by mechanical abuse is another form that is defined by physical stress causing damage in a single cell that spreads step by step to surrounding areas. These spatio-temporal faults require a holistic monitoring approach, where real-time sensor data and historical patterns are analyzed together to detect how localized faults evolve and spread through both space and time [46].

Table 2. Various types of battery faults and their corresponding failure situations.

Battery Fault	Description
Internal short circuit (hard)	A conductive path develops between the terminals, leading to rapid discharge and overheating.
Internal short circuit (soft)	A milder short circuit condition causing partial discharge and potential issues.
Lithium plating	Accumulation of lithium ions on the anode while charging, causing a decrease in capacity and possible safety hazards.
Over-discharging	Charging or discharging beyond recommended thresholds, leading to battery damage.
Abnormal self-discharge	Gradual and unintended loss of capacity over time due to internal chemical reactions.
Abnormal capacity degradation	Steady decline in the battery’s ability to store and deliver charge.
Abnormal voltage fluctuations	Unusual variations in voltage suggesting internal issues or imbalance.
Abnormal temperature behavior	Unusual increase in temperature during operation, indicating potential faults.
Electrolyte leakage	Leakage of the electrolyte, often due to physical damage to the battery.
Thermal runaway	Uncontrollable overheating leading to a risk of fire or explosion.
Cell balancing issues	Imbalance in the charge levels of individual cells in a battery pack, resulting in less efficient performance.

outlines various fault types and failure scenarios commonly observed in Li-ion batteries, serving as the foundation for understanding their spatial, temporal, and spatio-temporal characteristics.

2.2. Spatio-Temporal Characteristics of Battery Data

EV battery packs are heavily instrumented systems. A typical pack has voltage sensors for every cell or every few cells, temperature sensors at various locations, and sometimes pressure or gas sensors [47]. The BMS records these at a certain frequency from seconds up to minutes, depending on the system. Thus, the battery generates multivariate time-series data, and at each time step an array of measurements (voltages, temperatures, etc.) is collected [48]. The spatial dimension refers to the arrangement of these cells in the pack. The temporal dimension is the sequence of such vectors over time as the battery operates charging, discharging, and idling. The battery generates multivariate time-series data at each time step, with spatial and temporal relationships crucial for detecting faults [49]. Figure 7 illustrates a spatio-temporal fault detection framework that combines feature extraction, a multi-head attention autoencoder, and a two-stage thresholding method to process spatial inconsistencies and temporal patterns for accurate battery pack and cell fault detection.

2.2.1. Spatial Domain

Faults often have spatial signatures. In a well-balanced healthy pack, all cell voltages and temperatures should be within a narrow range of each other at any given time, especially if the cells are identical and equally used over time [50]. A fault can break this uniformity. For instance, an internal short in one cell will cause its voltage to sag compared to others under load or its temperature to rise relative to neighboring cells [51]. An overheating cell might stand out in the thermal profile of the pack, for example, when one module corner gets more heated, possibly indicating a cooling failure or a failing cell generating heat [52]. Cell-to-cell variation is thus a key indicator; techniques look at the spatial residual difference of each cell’s reading from the pack average to detect outliers. Spatial relationships are also important; cells physically near each other may influence each other’s temperature. If one cell goes into TR, the cells next to it will heat up in a spatial cluster. Graph-based representations of nodes as cells and edges connecting adjacent cells or cells in the same module can capture such relationships for analysis [53].

2.2.2. Temporal Domain

Battery faults also have temporal dynamics. A progressive fault by definition involves changes over time, with a cell’s capacity dropping faster than others or a slow increase in self-discharge current [54]. Even sudden faults have early signs over time. For example, an internal short might show a tiny voltage drop hours or days before a major failure; as another example, before a thermal runaway event, there might be a period of accelerating temperature rise minutes before a warning [55]. Therefore, time-series analysis of each signal and of derived features like voltage differences is critical. Fault propagation is a temporal process once a fault starts; how it evolves in terms of quick escalation or slow burn is a time-domain behavior. Temporal modeling techniques like filtering, change-point detection, or sequence modeling with neural networks are employed to spot the beginning of an abnormal trend [56].

Crucially, space and time are interlinked in battery fault scenarios. Consider a thermal runaway propagation: it is a spatial sequence of failures occurring over time. Cell A fails and a few seconds later, adjacent cell B fails, and so on [57]. Pure spatial analysis at a single time could identify cell A as hot and others as cool, so A is faulty, but without temporal tracking, one might not predict that B will soon fail. Conversely, pure temporal analysis of one cell might catch its temperature rising, but without spatial context, one might not realize the risk of it spreading to neighbors or detect if multiple cells show a correlated rise that could indicate an external cause like ambient heating [58]. Spatio-temporal modeling aims to capture patterns like “Cell 5’s voltage has been dropping faster than others for the last 10 cycles” or “temperatures in module 3 are rising together faster than normal, possibly indicating a developing issue localized to that module.” By combining both dimensions, the hope is to achieve earlier and more robust fault detection [59].

Research has shown that integrating these perspectives can reveal early-warning signals across multiple spatio-temporal scales that isolated analysis would miss [60]. To summarize the fundamentals, EV batteries can fail in various ways, each with distinct signatures in the data, like overcharge, over-discharge, internal shorts, thermal runaway, and cell imbalance, each showing unique signs in cell voltages, temperatures, and other parameters. These indicators appear as deviations among cells, such as spatial anomalies and changes over time trends or abrupt events [61]. The importance of spatio-temporal modeling lies in its ability to detect faults that have multivariate footprints, leveraging the full richness of the battery’s sensor network and its historical data [62].

3. Approaches and Techniques for Spatio-Temporal Fault Detection in ESSs

Spatio-temporal techniques explicitly consider both spatial and temporal dependencies in battery systems. While some methods in Section 3.1 and Section 3.2 implicitly use spatio-temporal information (CNN-LSTM hybrids), here we focus on approaches that are fundamentally designed for spatio-temporal data or that combine separate spatial and temporal modeling stages. We break this discussion into spatial analysis techniques, temporal dynamics modeling, and integrated spatio-temporal approaches. The advanced category includes 3D-CNNs, GNNs, and transformer-based models tailored for spatio-temporal sequences, as well as special case studies like spatio-temporal sequence-to-sequence learning. Figure 8 shows the mechanism of spatial–temporal and temporal–spatial feature extraction.

3.1. Spatial Analysis Techniques

Spatial modeling in battery packs involves understanding and leveraging the relationships between cells or modules in space, which could be physical space or logical arrangement. Two prominent techniques are graph-based models and thermal imaging approaches, along with methods focusing on sensor placement and spatial interpolation [64]. A battery pack has each cell as a node, and edges connect nodes that influence each other as neighbors in the pack’s physical layout or in the electrical circuit. Graph-based fault detection builds a model of the pack as a network. A graph-based voltage model to predict each cell’s voltage should be given the other cells’ voltage; a fault is detected when the actual voltage deviates from this prediction beyond tolerance [65]. Popp et al. [66] did exactly this by training a GNN to learn the complex interactions of cell voltages in a pack under various conditions.

The GNN effectively acted as a “fault-free model” of the pack. It output the expected voltage of each cell based on its neighbors and historical data. During operation, if a cell’s measured voltage fell outside a predicted confidence interval, the system flagged a potential fault on that cell [67]. This approach inherently accounts for spatial correlations. For instance, if one cell drops, normally its neighbors might also show slight effects due to load redistribution or thermal coupling, so the GNN learns those patterns. A fault like an isolated drop stands out because it breaks the learned spatial correlation. This GNN-based method achieved accurate fault detection and localization in packs, illustrating the power of spatial modeling for anomaly detection [68]. It was noted, however, that such data-driven spatial models require a variety of training data and may be sensitive to sensor noise. Ensuring robustness via regularization or incorporating physics in the graph is a key consideration [69]. Beyond GNNs, simpler graph-based methods include using graph inference or spatial interpolation. One could model the battery pack as a resistive network and use techniques like Kirchhoff’s laws to infer a missing sensor value. A significant mismatch would indicate a fault in a node cell [70].

Figure 7 illustrates a spatio-temporal framework designed for fault detection in Li-ion batteries within EVs [71,72]. This framework consists of three main components that work together to identify faulty cells accurately and efficiently. The first part focuses on feature extraction. Here, voltage data is collected from the entire battery pack, capturing both spatial differences between cells and their changes over time. These spatio-temporal features help highlight inconsistencies within the pack, which are crucial indicators of potential faults.

The second part of the figure presents the spatio-temporal autoencoder model. An autoencoder is a neural network that learns to reconstruct its input data while minimizing reconstruction errors. In this case, the autoencoder is trained on voltage data from healthy battery cells, establishing a baseline for normal operation. When voltage data from a faulty cell is fed into the model, the presence of anomalies such as unusual voltage patterns leads to higher reconstruction errors. This enables the system to detect faults effectively. A key enhancement here is the inclusion of a multi-head attention mechanism, which dynamically focuses on important temporal and spatial features. This mechanism improves the model’s ability to understand complex patterns in battery behavior over time and across different cells.

The third part introduces a two-stage fault threshold method. First, the framework determines whether the entire battery pack is faulty by analyzing its reconstruction error. If a fault is detected at the pack level, the system then evaluates individual cells within that pack. Both stages apply the interquartile range (IQR) method to calculate fault thresholds, which makes the detection process robust against outliers and adaptable to batteries with different aging levels or material properties. This two-stage approach ensures precise identification of faulty cells, even when battery packs have varied conditions. This figure represents a comprehensive, data-driven method that integrates DL with domain knowledge of battery systems. By combining spatio-temporal feature extraction, attention-enhanced autoencoders, and statistical fault thresholding, the approach provides a reliable solution for detecting faults in EV batteries.

Another study by Wei et al. [73] introduced a spatio-temporal inference system using graph theory. It created a probabilistic model of heat diffusion in a battery pack graph to detect and locate thermal abnormalities. By observing temperature readings at various nodes, the system inferred the likely origin of an abnormal heat source fault and its propagation path. This not only detected the fault but also localized it, an important aspect for mitigation. Spatial thermal monitoring is crucial for detecting issues like cooling system failures or cell overheating [74]. Some fault detection approaches use thermal cameras to capture infrared images of the battery pack. These images provide a 2D temperature field. Computer vision techniques, often CNN-based, can then identify hotspots or unusual thermal patterns, as mentioned earlier. U-Net, a type of CNN, has been tested on thermal images to isolate defects in cells by segmentation approaches [75]. While IR imaging can be powerful since thermal runaway can be directly observed as a hotspot, it is generally used in lab settings or as an add-on for research; it is not yet common in vehicles due to cost and complexity [76]. A more industry-friendly approach is using distributed temperature sensors, like fiber optic sensors that run through the pack, to provide a continuous temperature profile [77]. These yield spatially dense data, such as a temperature reading every few centimeters.

Techniques like spatial entropy have been applied to such data; for instance, Wei [78] proposed a spatio-temporal entropy method to detect and localize thermal abnormalities in large-format battery packs. They used the idea that a normal temperature field has a certain spatial coherence, whereas an abnormal heating event increases the disorder (entropy) in the spatial temperature distribution. By performing Karhunen–Loève decomposition (KLD) or PCA (Principal Component Analysis) in space and calculating entropy, they could flag anomalies and even triangulate the position of the fault since the method indicated which spatial modes were excited. Many BMSs perform routine checks like “maximum cell voltage minus minimum cell voltage” as a consistency metric. Advanced methods expand on this by looking at distributions of cell parameters. Some research uses clustering of cells based on their performance; an outlier cell cluster might indicate that those cells are drifting towards a fault state [79]. Spatial statistics across the pack like variance and skewness of cell voltages can be monitored over time; a rising variance may warn of an imbalance fault developing. These methods bridge both spatial and temporal trends; the metric is spatial at a given time but it is tracked across time [80].

CNNs, initially designed for computer vision tasks, have been effectively applied to identify spatial fault patterns in battery systems [81]. On transforming spatially distributed sensor data such as temperature distributions or voltage maps across cell arrays into structured two-dimensional feature representations, CNNs can identify localized anomalies including thermal runaway, uneven heating, and the emergence of hotspots. This spatial awareness makes CNNs particularly advantageous in module- or pack-level diagnostics [82]. Furthermore, extensions such as 3D-CNNs allow for the simultaneous modeling of spatial and temporal correlations, enhancing their capability to capture dynamic fault propagation. However, CNNs are inherently limited in sequential modeling tasks unless integrated with temporal architectures such as RNNs or LSTM networks [83].

To capture dynamic fault evolution more comprehensively, 3D-CNNs extend this architecture by treating time as an additional depth axis [84]. This enhancement enables the model to learn how spatial fault features develop across short temporal intervals, improving sensitivity to evolving anomalies such as thermal propagation or gradual voltage divergence. CNN performance can be further augmented through attention mechanisms, which selectively emphasize critical time steps or sensor channels. Such mechanisms help the model focus on diagnostically relevant patterns while filtering out noise [85]. To properly capture and interpret both the spatial and temporal features that constitute battery behavior, most diagnostic systems first address the extraction of features individually from each of these groups. They extract these features and then integrate them into DL models to enable accurate, holistic fault detection, as seen in Figure 9.

Qiu et al. [86] introduced a CNN-based framework for battery fault detection that combines a stacked sparse autoencoder and a convolutional block attention module within a capsule network structure. In their design, the autoencoder performs denoising and dimensionality reduction, while the convolutional block attention module—CapsNet—emphasizes crucial spatial and temporal features to improve fault classification accuracy. This attention-driven structure outperformed traditional CNNs and fully connected networks by prioritizing high-impact data regions during learning.

3.2. Temporal Dynamics Modeling

Temporal modeling focuses on how battery signals evolve. Many classical methods fall here, including forecasting models, filtering techniques, and change detection methods [87]. Autoregressive Integrated Moving Average (ARIMA) models, along with other statistical time-series approaches, can forecast future battery parameters based on previous data values. If the future measurement deviates beyond a prediction interval, it signals an anomaly [88]. ARIMA, being linear, might not capture battery nonlinearities except in a narrow operating range, but it can model trends and seasonal effects. ARIMA could model the gradual drift of capacity or resistance with cycles, helping to distinguish normal aging from an unexpected drop fault. More complex variants, like ARIMAX, which include external inputs like current or temperature to predict voltage could be employed [89]. However, ARIMA requires stationary data, and often, battery data is non-stationary due to changing loads. Thus, often, a preprocessing step like focusing on residuals after subtracting a baseline model is needed [90].

RNNs, including their advanced forms such as the LSTM and GRU, are utilized for modeling temporal dependencies in battery data, enabling effective anomaly detection over time [91]. These models are designed to analyze sequential sensor data such as voltage, current, and SOC, making them highly suitable for capturing the time-evolving behavior of battery systems. RNN-based architectures are particularly adept at identifying long-term dependencies and slow-developing anomalies, which are common in Li-ion batteries used in EVs [92].

LSTM networks, in particular, are widely used for sequence prediction and temporal modeling tasks. In battery fault diagnostics, LSTM networks have demonstrated effectiveness in detecting slow-evolving faults by learning from multivariate time-series data [10]. These models retain relevant historical context through internal memory cells, allowing them to forecast potential trends such as overcharging, capacity fade, or progressive degradation [93]. However, LSTM networks need extensive labeled datasets to achieve peak accuracy and typically demand high computational resources, which may challenge their deployment in real-time embedded systems. Despite this, LSTM networks continue to be a popular choice due to their ability to model long-range behavior in battery performance [94,95].

A particle filter can maintain multiple hypotheses of battery state, and some researchers have augmented PFs to include fault modes as part of the state space [96]. If the measurements align more with the internal short scenario, the PF’s probability weight for that scenario increases, thereby detecting the fault. Ye et al. [97] used a dual-scale adaptive particle filter for joint parameter and SOC estimation. While their goal was SOC, the same framework could detect faults by noticing a sudden change in estimated parameters, like a sudden drop in the capacity parameter, which might indicate a cell fault. These methods treat the time series as piecewise segments of stable behavior [98]. A fault is a change point where the statistical properties of the signal change mean, variance etc. Algorithms like Cumulative Sum and Sequential Probability Ratio Test have been applied to battery data [99].

Zhang et al. (2022) [100] designed an LSTM-based prognosis model capable of forecasting multi-step battery string voltage with high accuracy, enabling early detection of anomalous behaviors. Studies have shown that these frameworks can identify subtle nonlinear patterns that precede critical events, such as thermal runaway, earlier than traditional threshold-based detection methods [101]. The ability of RNNs and LSTM networks to recognize early warning signs of failure events makes them valuable for predictive maintenance and fault prevention strategies in BMSs [102].

Transformers have transformed sequence modeling within natural language processing and are increasingly utilized in time-series applications such as battery diagnostics [103]. At the heart of transformer models lies the self-attention mechanism, allowing them to assign importance to different time steps within a sequence. This is particularly beneficial for long battery usage sequences, where traditional RNNs may struggle with gradient vanishing or limited context [104]. Unlike RNNs, transformers can more efficiently capture long-range dependencies without being limited by sequential propagation steps [105].

One such application is the BERTtery architecture introduced by Zhao et al. [10], which adapts the BERT model for early detection of battery faults, utilizing a dual-tower transformer architecture. One tower encodes temporal measurement sequences, while the other analyzes cross-channel data such as multiple cells or sensor types. This architecture enables the model to detect relationships between a shape in the voltage curve and a temperature pattern across cells, often an early indicator of failure. The self-attention mechanism in the temporal tower highlights past time points that are relevant to the current prediction, while the channel-wise attention learns relationships among different sensors [106].

This spatio-temporal transformer structure demonstrated an ability to detect early warning signs across multiple scales, identifying subtle deviations in early battery cycles that foreshadow failure many cycles later [107]. The BERTtery model showed robust performance in real-world EV data, providing weeks or even months of fault warning using only standard sensor data like voltage and temperature. In a related study, Jaguemont et al. [29] introduced another two-tower transformer for state-of-charge (SOC) estimation. Their model applied multi-head attention to automatically learn temporal features, eliminating manual feature extraction and enhancing interpretability.

Attention-based models also offer a degree of explainability. By visualizing attention weights, researchers can interpret which time points or sensors influenced a prediction [108]. For example, the model might identify that the voltage and the temperature are highly correlated with a future fault, supporting engineers in validating the model’s predictions or even discovering new early indicators. These developments highlight the growing importance of temporal modeling, particularly transformer-based approaches, in enhancing early fault prediction and interpretability in EV battery systems [109].

3.3. Integrated Spatio-Temporal Architectures

Integrated architectures that combine convolutional and recurrent layers are being widely adopted to model the spatio-temporal dynamics of batteries [110]. These models simultaneously learn spatial distributions and temporal dependencies, allowing for more accurate and early fault detection. In a CNN-LSTM hybrid model, convolutional layers generally extract spatial information like temperature distributions or voltage correlations, which are then input into an LSTM network to model their temporal variations [111]. For instance, a PCC-CNN-LSTM model was utilized in a study to identify faults in EV batteries. The Pearson Correlation Coefficient (PCC) matrix was first computed from battery sensor data, serving as an input to the CNN, which extracted spatial features. The LSTM then modeled the sequential evolution of these features to predict faults [112,113]. Such hybrid models have shown superior accuracy compared to individual CNNs or LSTM networks, as they can simultaneously capture spatial dependencies across battery cells and long-term temporal dynamics [114].

To enhance the model’s ability to detect abnormal behavior over time, temporal convolutional autoencoders have been applied to learn latent representations of normal operation [115]. When a fault occurs, the autoencoder’s reconstruction error increases, providing a signal of anomaly. This latent space can also be modeled using LSTM layers instead of raw input, improving the detection of gradual or evolving fault patterns. However, DL models often require large datasets, which is challenging in battery applications due to the scarcity of real-world fault data [103,116]. To mitigate this, researchers use a combination of lab-induced faults, transfer learning, and synthetic data generation through simulations [117].

Transformers are another class of DL models that have been adopted for integrated spatio-temporal learning in battery diagnostics [118]. In the BERTtery model, a dual-tower transformer architecture is employed, where one tower handles the temporal sequences of voltage and current data and the other examines cross-channel characteristics, such as sensor measurements from multiple cells or modules [119]. This architecture enables the model to simultaneously learn temporal dependencies and spatial interactions among sensors. The self-attention mechanism enables the model to concentrate on diagnostically relevant time steps, such as anomalies at the start of a charging cycle, while also learning relationships across sensor channels [120]. The model demonstrated the ability to detect subtle early indicators of faults, providing warnings weeks before failure occurred using only standard onboard data like voltage and temperature [19,121].

More recently, attention has also turned to the 3D-CNN, which uses a three-dimensional convolution kernel that spans spatial dimensions, such as cell index or sensor index, and the temporal dimension [122]. A 3D filter sliding through this cube can pick out a feature like “a localized hot spot in space that grows over a short time interval.” This is very useful for phenomena like thermal propagation. A case study from another domain, that of traffic prediction, introduced a model called ST-ConvS2S (Spatio-Temporal Convolutional Sequence-to-Sequence), which used layers of temporal and spatial convolutions to forecast future traffic states [123]. Analogously, an ST-ConvS2S model for a battery could input a sequence of thermal images of a pack and predict the future temperature distribution, effectively predicting thermal runaway propagation [124].

Zhang et al. [125] developed a spatio-temporal CNN to predict temperature evolution in a battery pack for thermal management. While their goal was prediction for cooling control, the same model could be used to predict normal thermal behavior and detect anomalies when reality deviates from a form of fault detection. Additionally, graph-based temporal neural networks, such as diffusion convolutional recurrent neural networks or temporal GNNs, have emerged as powerful spatio-temporal frameworks [126]. These models process battery pack data by applying graph convolution across the cell layout, capturing spatial interaction at each time step, followed by a recurrent layer that propagates the state forward in time. This allows the model to learn how localized faults like a voltage drop in one cell influence surrounding cells in later cycles [127]. Early experiments using GNNs on degradation data have demonstrated the ability to track fault propagation and cell interactions in complex battery architectures. These models treat the battery pack as a dynamic network where each node or cell interacts with others both in space and across time [128,129].

A dynamic GNN could identify that an internal short in one cell will affect its neighbors in subsequent time steps through thermal or electrical coupling edges. Early experiments with GNNs on battery degradation data show promise in capturing spatio-temporal degradation dynamics [130]. These models effectively treat the battery pack like a small “social network” of cells, learning how a change in one might influence others over time. Regarding transformer-based models for multivariate time series, beyond BERTtery, another concept is a hierarchical attention network, where one level of attention operates across sensors at a single time through spatial attention and another across time for each sensor through temporal attention [131]. In batteries, spatial attention could learn to weight each cell’s importance, maybe focusing on the worst cell, and temporal attention could focus on critical moments like the start of charge and end of charge, where faults often manifest [132].

A recent study by Peng et al. [133] proposed a Spatial–Temporal Self-Attention Network, further highlighting the value of layered attention structures in capturing complex interdependencies. In such models, spatial attention is used to prioritize information from critical cells (e.g., the hottest or weakest), while temporal attention identifies crucial time segments like charging phases or rapid voltage drops [134]. These transformer-based methods not only enhance detection accuracy but also improve interpretability, helping engineers trace back the model’s predictions to physical events in the battery system [135].

Hybrid fault detection frameworks for Li-ion batteries offer a promising approach by combining the accuracy of physics-based modeling with the adaptability of data-driven DL techniques [136]. This combination enables systems to benefit from the interpretability and theoretical rigor of physical modeling while leveraging the adaptive pattern recognition and generalization capabilities of neural networks [137]. A commonly used technique is residual-based learning, where a deterministic battery model, typically an equivalent circuit model, simulates expected voltage or temperature behavior under defined operating conditions. The deviation between simulated results and real sensor readings, known as the residual, signals possible abnormalities [138]. The residuals are subsequently input into classifiers like SVMs, gradient boosting models, or neural networks, which are trained to recognize and classify distinct fault patterns. Decoupling of fault modeling from the full system behavior reduces learning complexity and improves diagnostic precision [139,140].

PINNs represent another notable category within hybrid modeling approaches. PINNs embed physical laws such as Ohm’s law, electrochemical kinetics, and heat transfer equations directly into the training loss function [141]. Such constraints guarantee that the model adheres to known physical principles, even in situations where experimental data is limited. These models have demonstrated success in scenarios such as thermal anomaly detection and internal short circuit detection, where labeled data are hard to obtain but governing equations are well understood [142]. By enforcing physics-constrained outputs, PINNs offer improved generalization across battery chemistries and conditions, addressing a major limitation of purely black-box AI models [143].

Beyond residual-based and PINN architectures, next-generation hybrid frameworks integrate DL with high-fidelity electrochemical or multi-physics simulations [144]. These systems use neural networks to learn unmodeled nonlinearities such as degradation effects or hysteresis while relying on simulators to enforce physical behavior under various loading profiles [145]. The synergy between simulation accuracy and data-driven flexibility enhances robustness, especially under uncertain or noisy real-world conditions. These models support both real-time diagnostics and long-term health forecasting, making them well-suited for use in EV BMSs [146].

A major challenge in developing such hybrid models is the scarcity of real-world fault datasets, particularly those involving severe or catastrophic failures [147]. To overcome this, researchers increasingly use a mix of lab experiments, where faults are induced in controlled environments, along with transfer learning and data augmentation techniques [148]. In other cases, physics-based simulations are used to generate synthetic training samples, such as temperature maps from thermal propagation models. Although there is always a domain gap between simulated and real data, careful model tuning and hybrid training strategies help mitigate this issue [149].

Table 3. Comparative summary of battery fault detection methods, illustrating their use of spatial and temporal data, targeted fault types, dataset scale, and notable performance metrics or outcomes.

Method and Reference	Spatio-Temporal Features	Target Fault (s)	Data Source	Performance Metrics	Remarks
Zhao et al. 2017 [43]	Cell voltage anomalies over time (ST)	Internal short, degraded cells	EV field data (fleet)	Sensitivity: 92%; false positives: 7%	Three-level screening improved detection reliability.
Yang et al. 2018 [151]	Engineered features (S and T): voltage drop, short indicators	External short, leakage	Lab experiments (short circuit tests)	Accuracy: 94%; precision: 91%	RF classifier required careful feature selection.
Hong et al. 2019 [152]	Sequential voltage per cell (T); per-cell model	Predicted voltage anomaly	Lab cycling data	RMSE: 0.012 V; fault lead time: 5 min earlier	LSTM achieved early anomaly detection.
Zhou et al. 2024 [153]	Time window as image (S across sensors, T in window)	Various battery faults (classification)	Lab fault dataset	Accuracy: 98.3%; F1: 97.8%	CNN attention improved interpretability.
Zhao et al. 2023 [10]	Two-tower transformer (temporal seq. and cross-cell features)	Early failure prediction	EV field data (1000s of cells)	Accuracy: 96.3%; fault prediction horizon: months ahead	Demonstrated long-term fault predictability.
Wei & Li 2022 [154]	Distributed temperature field over time (S+T)	Thermal anomalies (TR localization)	Lab thermal abuse tests	Detection lead time: 30 s earlier	Entropy-based method improved thermal awareness.
Ali et al. 2024 [155]	Physically constrained DL (S implicit, T)	Safety faults (TR, leakage, short)	18.2 million data points (515 EVs)	Detection improvement: +46.5%; F1: 95%	Demonstrated scalability for large fleets.
Hong et al. 2024 [156]	CNN-LSTM fusion of voltage, current, and temperature sequences (spatio-temporal)	Short circuit, imbalance, thermal faults	Real-world EV and ESS datasets	F1-score: 0.96; accuracy: 97.3%	Employs physics-informed deep learning; detects faults up to 48 h early with high interpretability.

summarizes representative methods from each category, noting the data they use, whether they capture spatial, temporal, or both aspects, the faults targeted, and key performance metrics or findings. It is evident that methods integrating both space and time generally achieve better detection performance, especially for incipient faults, than those using only one domain. Moreover, the ability to localize faults, not just detect them, is greatly enhanced by spatial analysis, while the ability to predict or prognosticate faults before they fully appear relies on temporal analysis of trends. Hence, the fusion of both in spatio-temporal models offers a powerful framework to advance EV battery fault detection [150].

PINNs integrate physical laws into the learning process and help reduce reliance on labeled fault data. However, they also have notable limitations [157]. Their performance depends heavily on the accuracy of the boundary and initial conditions, as well as on how well the underlying physical models are defined. When the physical constraints conflict with noisy real-world measurements, the training process can become slow or unstable. In addition, solving physics-based (PDE-constrained) loss functions is computationally expensive, making it difficult to scale PINNs to large spatial domains such as full ESS thermal fields [158]. To address these challenges, hybrid methods that couple simplified physics models for residual estimation with efficient data-driven classifiers can provide a better balance among interpretability, accuracy, and computational efficiency.

This comparative view highlights that hybrid and spatio-temporal methods generally outperform single-domain methods in terms of early detection and accuracy. Notably, the transformer-based and physics-informed DL approaches, which leverage both spatial and temporal data in complex models, achieve some of the best results with high accuracy on real-world data and early warnings. The field is clearly moving toward models that can learn the full battery system behavior across all sensors and over time and thereby detect when the system strays into an unsafe or abnormal regime.

4. Current Challenges and Open Issues

Although substantial advancements have been achieved in battery fault detection methods, several challenges and open issues still persist. These must be tackled to ensure dependable, real-time fault detection in commercial EV battery systems and to promote industry adoption of advanced spatio-temporal techniques. As illustrated in Figure 10, six key barriers currently hinder the implementation of sophisticated fault detection within practical BMSs: the lack of labeled fault data, difficulties in real-time deployment, limited integration with control frameworks, vulnerability to sensor noise and anomalies, poor interpretability of complex models, and the difficulty of generalizing methods across different chemistries and applications. Overcoming these obstacles is crucial to achieving robust, scalable, and intelligent battery health monitoring in EVs.

4.1. Lack of Labeled Data for Battery Faults

Reliable and well-documented data on battery faults, especially severe ones like thermal runaway, are extremely limited. In actual electric vehicle fleets, failures seldom occur, a positive outcome for safety but a challenge for data-driven model development [160]. Most EV batteries are designed to avoid failure throughout their warranty period; when failures do occur, companies may not disclose the data. Consequently, researchers depend on laboratory tests, intentionally triggered failures, or simulated datasets that might not fully represent the range of real-world operating scenarios [161]. There is also a lack of standardized datasets; many studies rely on private data that cannot be shared, limiting reproducibility. This scarcity of labeled fault data makes training supervised models difficult. Models like DNNs risk overfitting to the few fault examples they see. Moreover, failures in the field may occur due to complex combinations of factors such as manufacturing variance, specific duty cycle, climate, etc., that are hard to replicate in the lab [162].

Unsupervised and self-supervised approaches which do not require labeled faults are promising, but even they need abundant normal-operation data and might struggle to define what constitutes an anomaly when the boundary between safe and unsafe is unclear [163]. The research community should promote data-sharing initiatives, potentially via collaborations or government-backed programs, to build extensive datasets covering both normal and faulty battery behaviors. Active learning and reinforcement learning methods can be used in a safety-conscious way to guide experiments toward informative failure data [164].

4.2. Real-Time Deployment Challenges

Many advanced algorithms, such as DNNs with millions of parameters or graph models on large packs, are computationally intensive, requiring substantial memory and processing power. In contrast, a BMS in an EV is typically built around a low-power microcontroller or a small microprocessor with strict real-time requirements [165]. Fault detection algorithms often need to run at 10–100 Hz sampling rates for hundreds of channels. Running a large CNN or transformer on an automotive-grade electronic control unit in real time is nontrivial. While some modern EVs may have more powerful domain controllers that could handle heavier algorithms, the power and cost constraints push designers to simpler algorithms [142]. There is thus a challenge in distilling complex spatio-temporal models into lightweight forms suitable for the edge on board the vehicle [166].

Researchers are investigating “TinyML” approaches for BMSs, for instance, taking an LSTM and reducing its size while keeping performance acceptable. Another aspect is latency. Some fault detection decisions might be needed within milliseconds, and complex algorithms that introduce even a one-second delay might be unacceptable for safety [167]. Balancing algorithm complexity with the need for fast, deterministic execution is an open engineering problem. Notably, some studies have managed real-time implementations: Kim et al. [168] use microcontroller-based fault detection and tailor the solution to embedded hardware. It shows that careful optimization can allow advanced analysis like data mining or parameter estimation within a BMS. Another approach is to utilize dedicated hardware, for instance, an FPGA or a neural network accelerator chip in the BMS for heavy computations, though this increases cost [169].

Many advanced models, such as deep transformers, large GNNs, and PINNs, achieve excellent fault detection accuracy but require considerable computational resources for both training and real-time operation. For onboard or edge applications in ESSs and EV BMSs, factors like latency, memory usage, and power consumption become critical. Key metrics to consider include the number of model parameters, inference time on target hardware like embedded SoCs or edge GPUs, and peak memory demand. To improve feasibility, researchers have applied model compression (pruning and quantization), knowledge distillation to smaller models, event-driven inference, and hybrid approaches that use lightweight filters for preliminary screening before activating more complex models. Reporting these deployment metrics consistently across studies would help evaluate the practical viability and real-time performance of emerging fault detection frameworks.

4.3. Generalization and Transferability Issues

A model trained on one type of battery or operating regime might not work on another. This is a major concern for ML models in BMSs. Batteries vary in chemistry between NMC, LFP, and other form factors, and between cylindrical and pouch, among cooling systems, and among usage patterns, all of which have different profiles [170]. A fault detection model needs either to be retrained for each scenario or to be flexible enough to generalize. Currently, many data-driven models are developed and tested on relatively narrow datasets of the same type of cells under similar cycling. If we deploy such a model fleet-wide, there is a risk that it might misclassify unfamiliar patterns as faults (false alarms) or miss faults that look different. Domain adaptation plays an important role in adjusting a model trained in one context so that it performs effectively in another field using data without extensive relabeling [171].

Shi et al. [172] emphasize that as battery technologies evolve with new chemistries and designs, predictive models must adapt rapidly and that “variabilities in manufacturing and operating conditions compound the challenge.” They suggest that models capable of learning from diverse data sources and hybrid modeling combining physics with data will generalize better. For instance, a transformer could be pretrained on one battery dataset and fine-tuned on another. Another is to incorporate physical parameters like capacity and internal resistance as part of the input space so the model can condition on those [173]. Nevertheless, ensuring reliability across different conditions remains open. Regulatory bodies will likely demand evidence that any AI-based detection system works for all expected conditions requiring extensive testing or formal verification, both of which are difficult for black-box models. Gaining trust in AI-driven fault detection is paramount for adoption in industry, especially in safety-critical automotive applications [174].

A BMS engineer or a safety officer would want to know why the algorithm says a fault is present [175]. Traditional model-based methods have clear reasoning, for example, “voltage exceeded limit hence fault.” But a complex neural network might trigger an alarm due to a combination of sensor readings in a way that is not immediately transparent. This lack of clarity causes some problems: (1) difficult in checking and fixing the model when it makes a mistake; (2) lack of operator insight—ideally, if a fault is flagged, the system should also indicate the likely cause or location, but pure data-driven models might just give a yes or no without context; and (3) regulatory approval—in some cases, explainability might be needed to certify the system under functional safety standards in automotive applications, encouraging predictable, explainable behavior. To tackle this, researchers are exploring XAI techniques for battery management [176].

As mentioned, attention weights can highlight which inputs are important; similarly, methods like SHAP (Shapley Additive Explanations) can rank sensor contributions to a decision. Some recent models explicitly output interpretable indicators. A model might output an estimated internal resistance for each cell that is human-interpretable, for example, that a high resistance cell is bad, as an intermediate step to predicting a fault [177]. The review by Song et al. pointed out that purely data-driven methods have poor interpretability, whereas more classical methods are easier to explain [178]. A promising direction is hybrid modeling, to be discussed later, where part of the model is physical and thus understandable and part is ML for flexibility. Finding an optimal trade-off between model accuracy and interpretability remains an unsolved challenge. There is a research opportunity in developing explainable fault reasoning on top of black-box detectors: essentially, using the outputs of a neural net and analyzing them with an expert system or logical rules to generate a human-readable detection [179].

However, a fault detection system can itself be victim to faults, specifically sensor faults [180]. If a voltage sensor gives a spurious reading, an algorithm might wrongly perceive a cell fault. Some advanced methods might detect inconsistency, for example, when one sensor disagrees with others, indicating the sensor itself is bad, but not all methods allow this. Ensuring the detection algorithm is robust to noise, outliers, and missing data is crucial [181]. In practice, BMSs have redundant measurements or perform reasonableness checks. For ML models, augmenting training data with noisy samples can teach the model to be tolerant. In graph-based approaches, techniques exist to identify a node whose data does not fit with neighbors, which might mean that the sensor has failed [182]. The challenge remains to avoid false positives due to benign anomalies; RF interference causing one erroneous temperature spike reading should not trigger a panic. Filtering inputs and requiring persistence of an anomaly for a certain time before alarm are practical measures, but these delay detection [183]. This trade-off between sensitivity and noise robustness is a classic signal detection problem, now magnified by the complexity of ML models.

The generalizability issue is mainly due to the lack of standardization, a challenge highlighted by Ward et al. [184] in their proposal for a Battery Data Genome to overcome the fragmented data ecosystem. In a step towards this goal, a study by Lininger et al. (2022) [185] introduces the Voltaiq Data Format (VDF), a standardized framework for organizing and sharing battery data across the entire life cycle. It defines common structures for time-series data, metadata, and file organization to enable consistent data collection and interoperability. By promoting standardized labeling and measurement conventions, the VDF supports reproducibility, large-scale benchmarking, and AI-driven analysis, directly contributing to improved comparability and transparency in battery fault detection and health management research.

4.4. Incorporation into BMS and Control Mechanisms

Identifying a fault is merely the initial stage; the system must then do something about it. If an algorithm is too sensitive and frequently triggers, the BMS might end up unnecessarily throttling power or taking the pack offline, impacting vehicle performance or availability. On the other hand, if it is too conservative, it may act too late [186]. Defining the correct actions upon detection of load limiting, cooling activation, cell bypass, warning the user, etc., is part of the broader fault management system design. This requires not just detection accuracy but also calibration of confidence levels. Many data-driven models output a continuous “health index” or probability of fault. Open problems remain in creating user-friendly and safe responses: for instance, how to inform the driver. Current EVs occasionally show messages like “battery fault service required,” which are very generic [187]. A more advanced system might say “battery fault detected cell overheating, please stop the car” using the richer output of a spatio-temporal model that can even pinpoint the issue. But deciding these thresholds and messages lies in that gray area between technical algorithm and practical usage. While spatio-temporal fault detection techniques show great promise, addressing these challenges is essential for their real-world deployment [188]. We further discuss these challenges in

Table 4. Overview of key challenges and open problems in spatio-temporal battery fault detection.

Challenge Area	Limitation/Gaps	Prospective Solutions
Scarcity of labeled fault data	Real-world battery faults are rare and underreported. Synthetic or lab-generated data may lack diversity.	Self-supervised learning, federated learning, synthetic augmentation, public datasets from consortiums.
Real-time deployment constraints	Models trained on specific chemistries or systems may not work on others, lack of cross-domain reliability.	Domain adaptation, transfer learning, hybrid physics-informed models.
Generalization and transferability	Models trained on specific chemistries or systems may not work on others; lack of cross-domain reliability.	Domain adaptation, transfer learning, hybrid physics-informed models.
Robustness to noise and sensor faults	Sensor anomalies or hardware noise can cause false alarms or missed detections.	Noise-aware training, redundancy checks, sensor fusion, smoothing filters.
Integration with BMS and control	Detection alone is insufficient. Systems must take reliable action and communicate effectively.	Fault classification with confidence scores, adaptive thresholds, actionable health indices.

and below:

• Data issues: Require innovative training schemes and perhaps industry collaboration on data sharing.
• Computational hurdles: Push us toward algorithm optimization and possibly new hardware.
• Generalization: Demands more adaptive and physics-aware models.
• Interpretability: Will likely involve hybrid approaches or additional layers to explain AI decisions.
• Robustness and integration: A fault detection system must work reliably within a complex BMS environment, not just in a research setting.

The next section will look ahead at emerging research directions that aim to tackle some of these challenges and push the field forward.

5. Future Research Directions

Looking ahead, numerous emerging research directions and technologies are expected to significantly advance battery fault detection, particularly through spatio-temporal approaches. These efforts aim to overcome existing challenges while utilizing recent progress in DL and battery modeling. Figure 11 outlines a circular hierarchical roadmap that summarizes potential future research paths in this field. The sections below highlight key areas needing deeper investigation and development.

5.1. Self-Supervised Learning for Battery Systems

Motivated by the achievements of large-scale foundation models in natural language processing and computer vision, researchers are now extending these models to time-series domains such as battery data analysis. The idea is to train a very large model on an extremely diverse set of time series from many domains or many fleets of batteries in a self-supervised manner so that it learns general patterns between normal and abnormal behavior [189]. Lag-Llama is a decoder-only transformer trained on a wide range of univariate time series; it demonstrated strong zero-shot generalization to unseen datasets and improved accuracy when fine-tuned, outperforming many specialized models. Translating this to batteries, we envision training a large model, perhaps a transformer, with tens of millions of parameters on sequences from hundreds of thousands of battery cells from different sources including normal operation, various aging trajectories, and any fault incidents available [190]. The training could be self-supervised, for instance, using objectives like forecasting the next data point, reconstructing masked portions of the sequence as in BERT’s masking, or contrastive objectives distinguishing between scrambled and real sequences. By doing so, the model would build a robust representation of battery dynamics [191].

Later, it can be fine-tuned on specific fault detection tasks with relatively few labeled. This approach could dramatically reduce the data requirement for each new scenario, as the foundation model already encodes a wealth of knowledge about typical battery behavior [192]. Additionally, such models might capture subtle precursors to failure that are consistent across different batteries, for example, a certain small oscillation pattern that tends to appear before an internal short. IBM’s Time Series Transformer and models like TimeGPT are also exploring this space [193], indicating a broad interest. A challenge for batteries is the multimodal nature of data on voltage, current, and temperature streams together; foundation models might need to handle multiple channels. Self-supervised techniques like forecasting both voltage and temperature given past data or aligning sequences from different cells to teach spatial awareness could be used. As these models tend to be large, deploying them in vehicles might not be directly feasible, but they could live in the cloud. We may see a cloud–edge hybrid where a foundation model in the cloud monitors fleet data and sends distilled insights or smaller models to edge BMS units [194].

5.2. GNNs for Enhanced Spatial Learning

Representing battery packs as graphs is intuitive, and GNNs continue to evolve rapidly as a research area. Future research can exploit GNN variants to improve spatial learning beyond what was achieved in early attempts. GATs could allow a model to learn which cell connections are most important for fault propagation, maybe giving more weight to immediate neighbors and less to far cells unless a thermal runaway is detected, in which case it could suddenly connect the whole module [195]. Hierarchical GNNs could treat modules as super-nodes, enabling pack-level fault reasoning like isolating which module has a fault and then zooming in to the cell. GNNs also can incorporate edge features. The distance or thermal resistance between cells can be an edge attribute, bringing physical context into the model [196]. One can imagine a GNN that simultaneously predicts each cell’s future state of health and detects anomalies, effectively performing distributed prognostics [197].

Another aspect is using spatio-temporal graphs where time itself is built into a graph. This yields a very large graph, but spatio-temporal GNN frameworks like structural RNNs or temporal message passing could handle it. For instance, Fan et al. [198] in another field created a DCRNN for traffic, which could be adapted to batteries, treating the voltage of each cell as a time series on a node. We expect GNN approaches to be particularly useful for localization of faults by looking at the activation of nodes in the network, which nodes’ states are most anomalous, etc. Additionally, GNNs could help with fault propagation modeling, essentially learning cause–effect relationships in the spatial domain over time.

5.3. LLMs for Battery Fault Detection

LLMs are quickly becoming powerful tools for intelligent battery fault detection due to their strong abilities in pattern recognition, multi-domain reasoning, and interpretability [199]. Although originally created for natural language processing, recent research has demonstrated their potential in sensor-driven and engineering applications. LLMs excel at modeling complex relationships within high-dimensional datasets and can be fine-tuned to reason across heterogeneous inputs, such as time-series sensor data, maintenance logs, and environmental conditions.

In the context of battery diagnostics, LLMs offer several advantages over traditional DL models. Their ability to ingest multimodal data such as voltage, temperature, current signals, and textual metadata enables a more holistic understanding of system behavior [200]. For instance, the recently introduced SensorLLM framework has demonstrated the effective alignment of inertial sensor data with pretrained language models for human activity recognition, a concept that is directly adaptable for anomaly detection in EV battery systems. By embedding temporal sensor sequences into language-like tokens, SensorLLM enables the model to learn semantic representations of operational patterns, which can be leveraged for early fault detection and system health monitoring [201].

Further research has explored the use of LLMs for battery SoH estimation. In a recent study, Xu et al. [202] proposed an LLM-based framework that combines historical cycling data with contextual knowledge to predict degradation trends. This approach has shown promising results in forecasting capacity fade and identifying precursors to failure, demonstrating the model’s potential in predictive diagnostics.

5.4. Physics-Informed AI and Hybrid Models

Combining physical battery models with data-driven methods can offer the advantages of both improved interpretability and reduced data requirements, as well as enhanced flexibility and accuracy. One such approach is seen in PINNs, where the neural network’s loss function incorporates differential equation terms, ensuring that outputs remain consistent with physical laws. In battery fault detection, PINNs can help estimate internal states that are otherwise unobservable [202]. For instance, a PINN could incorporate the diffusion equation for lithium concentration in the electrode and the heat equation for the cell. The network would take in measured surface temperatures and voltages and try to infer internal temperature distribution or lithium plating amount as a function of time, constrained by these equations. If the inferred internal state goes beyond safe limits, that effectively detects a fault [203].

Lixin et al. [204] described how physics-informed ML can integrate data and PDEs, and applying that to batteries is already in progress in some groups. There are PINNs for battery charging optimization, for instance. Another hybrid approach is electrochemical model augmentation, using simplified battery models like the single particle model or equivalent circuit and having neural nets learn the discrepancy between the model and real data [205]. Such model-constrained networks were demonstrated to diagnose faults under varying conditions, as in the 2023 study combining neural nets with model constraints that achieved great results on 515 vehicles. This idea can be further expanded by employing an equivalent circuit model to produce baseline voltage predictions, while a neural network processes the residual differences between actual and predicted values to identify faults [206]. The neural network only needs to learn the unmodeled aspects, such as internal short circuits or sensor drift, making the task simpler than training entirely from raw data [207].

Additionally, event-driven processing could be used. Instead of running the fault detection model at a fixed high frequency, one could run simpler monitors that trigger the heavy model only when something looks off. This way, 99% of the time the system uses negligible resources, and only when needed does it spawn the deep analysis. This is akin to some security systems working with multi-tier intrusion detection [208]. Another aspect is leveraging edge computing hardware. Modern cars increasingly include GPUs or NPUs (neural processing units) in their infotainment or driving systems. Perhaps the BMS could offload heavy computation to such a processor when the car is on or even to the cloud when the car is charging, uploading data and obtaining a cloud-based analysis. However, safety-critical decisions likely have to occur on board, so at least a pared-down model must run locally [209]. Research into BMS-specific hardware accelerators could emerge, for example, an FPGA in the BMS performing parallel computations for a GNN. Future research might also focus on making diagnostics actionable and understandable to users and technicians. This involves translating the output of spatio-temporal models into natural language explanations or simple indicators. Techniques from NLP could be applied to auto-generate reports from model outputs. Additionally, integrating fault detection with maintenance scheduling systems will be valuable. A prediction of failure within X months could trigger a preemptive battery module replacement under warranty [210].

The future of battery fault detection is likely to be characterized by smarter, more adaptive models that leverage both data and physics, are deployed in a hierarchical manner from cell level to cloud level, and that provide transparent insights into battery health. Foundation models could provide the broad knowledge, GNNs and transformers the fine-grained spatio-temporal analysis, and physics-informed frameworks the trust and generality. Combined with improved sensing (more sensors, fiber optics, etc.), the next generation of BMSs could possibly predict and prevent failures with unprecedented accuracy. One could imagine an EV that never suffers an advanced battery failure because its BMS would catch even the smallest discrepancy and gracefully retire the affected component or alert the user long before a dangerous situation arises.

6. Conclusions

This review comprehensively explored the evolving landscape of fault detection in ESSs through a spatio-temporal lens. We demonstrated that integrating spatial dependencies such as cell-to-cell interactions with temporal signal evolution significantly enhances early fault identification and diagnostic accuracy. Traditional methods, which treat battery data in isolation, often miss the complex interdependencies that precede critical failures. In contrast, advanced DL architectures like transformers, CNN-LSTM hybrids, and GNNs effectively capture these multivariate patterns. Our comparative analysis consistently showed that spatio-temporal models outperform single-domain approaches, particularly in identifying subtle, incipient faults. The evidence underscores the growing necessity of treating the battery pack as a dynamic, interconnected system rather than as isolated components.

Despite remarkable progress, several challenges remain unresolved. These include limited access to real-world fault data, the computational demands of deploying complex models on embedded systems, and the generalizability of models across battery chemistries and operating conditions. Future studies should emphasize hybrid modeling strategies that integrate physics-based understanding with data-driven learning to enhance both robustness and interpretability. Real-time deployment strategies such as lightweight neural networks, edge computing, and model compression techniques must be explored to enable onboard execution within BMS constraints. Moreover, XAI methods are essential to enhance transparency and regulatory acceptance. Emerging techniques like foundation models for time-series data, self-supervised learning, and PINNs represent promising pathways for developing adaptable and intelligent diagnostic systems.

The path forward demands a coordinated effort across academia, industry, and government. Open collaboration is crucial to developing standardized datasets, accelerating model validation, and ensuring that battery fault detection tools are both scalable and trustworthy. Bridging battery science with AI innovation will lead to the next generation of intelligent BMSs capable of real-time monitoring, early intervention, and continuous learning. As EV adoption accelerates worldwide, embedding such intelligent diagnostics into battery infrastructure will be key to maximizing safety, performance, and long-term sustainability in electrified transportation.