Fault Detection of Li–Ion Batteries in Electric Vehicles: A Comprehensive Review

Abstract

Lithium–ion (Li–ion) batteries are fundamental for advancing intelligent and sustainable transportation, particularly in electric vehicles, due to their long lifespan, high energy density, and strong power efficiency. Ensuring the safety and reliability of EV batteries remains a critical challenge, as undetected faults can lead to hazardous failures or gradual performance degradation. While numerous studies have addressed battery fault detection, most existing reviews adopt isolated perspectives, often overlooking interdisciplinary and intelligent approaches. This paper presents a comprehensive review of advanced battery fault detection using modern machine learning, deep learning, and hybrid methods. It also discusses the pressing challenges in the field, including limited fault data, real-time processing constraints, model adaptability across battery types, and the need for explainable AI. Furthermore, emerging AI approaches such as transformers, graph neural networks, physics-informed models, edge computing, and large language models present new opportunities for intelligent and scalable battery fault detection. Looking ahead, these frameworks, combined with AI-driven strategies, can enhance diagnostic precision, extend battery life, and strengthen safety while enabling proactive fault prevention and building trust in EV systems.

1. Introduction

Electric vehicles (EVs) are revolutionizing the transportation sector by offering a sustainable and intelligent alternative to conventional fuel-powered vehicles. Central to this advancement are Li–ion batteries, now widely used as the main energy storage system due to their high energy density, lightweight, and extended battery life [1]. These batteries enable EVs to cover longer distances while maintaining stable performance and long-term reliability. Compared with older battery technologies, Li–ion batteries offer a longer lifespan, making them well-suited for the ongoing use needs of EVs [2]. As advancements in battery technology continue to evolve alongside smart infrastructure, renewable integration, and intelligent transportation systems, the rapid adoption of EVs exemplifies the multidisciplinary innovation required to achieve sustainable, resilient transportation networks precisely the kind of innovation this special issue aims to spotlight.

Battery Management Systems (BMSs) are crucial for maintaining battery safety, performance, and useful life. Important functions of the BMS of an EV include modeling the battery, estimating internal states, and managing the charging process [3]. A strong battery model is necessary to understand how the battery behaves, monitor its performance, control its temperature, apply real-time management, and detect possible faults [4]. State estimation is a central task of the BMS, acting as a monitoring system for the entire power setup. In addition, BMSs help protect battery packs from dangers such as short circuits and overvoltage by using smart control systems, sensors, and actuators [5]. Another important function is cell balancing, which ensures that all cells within the battery pack maintain uniform voltage and state-of-charge (SOC) levels. Over time, individual cells may become unbalanced due to aging and environmental conditions, leading to pack-level inefficiencies and safety concerns. Balancing can be performed through passive or active balancing techniques. While often underemphasized in system schematics, cell balancing plays a key role in maximizing battery life, improving pack utilization, and preventing localized overcharge or overdischarge scenarios in EV applications [6]. Figure 1 presents a summary of the key components of a BMS and how it is used in electric vehicle technology.

As transportation moves toward electrification, Li-ion batteries have become increasingly important, making their safety and reliability a top priority. Incidents like unexpected fires and rapid loss of battery capacity in EVs highlight the urgent need for reliable fault detection and detection systems [7]. A failure in just one cell can trigger thermal runaway, where the cell’s temperature rises rapidly and uncontrollably, possibly causing nearby cells to overheat and leading to a fire across the entire battery pack [8]. These events present serious dangers to passenger safety and can damage or destroy vehicles in a short time. Therefore, real-time fault detection and early intervention are crucial for protecting users and extending battery lifespan [9]. Despite significant research efforts, maintaining the safety of billions of EV battery cells throughout their life cycle is still a major challenge.

Accurate and timely fault detection is crucial to prevent battery failures. Detecting issues early enables the BMS to respond with actions such as adjusting cell voltages, lowering the load, or separating a faulty cell to avoid accidents [10]. This becomes even more important under dynamic operating conditions, as EV batteries face fluctuating loads, temperatures, and environments that can hide or imitate fault symptoms. In recent years, several battery failures have resulted in recalls, raising both safety concerns and financial burdens [11]. There is a growing demand for technologies that can provide early warnings and enable timely intervention [12]. Various methods have been studied for detecting battery faults, generally grouped into model-based techniques, signal analysis methods, and data-driven approaches. Model-based techniques use first principles or equivalent circuit models to predict battery behavior, and a fault is identified when the actual measurements differ significantly from these predictions [13].

However, these techniques depend on precise modeling of battery behavior and may not perform well as the battery ages or encounters unexpected conditions [14]. Signal processing and statistical methods identify anomalies by analyzing data for unusual patterns, sudden changes, or outliers. While these methods are quick and easy to interpret, they often miss early-stage faults hidden within noisy data [15]. Machine learning (ML) and deep learning (DL) techniques detect faults by learning patterns directly from battery data, making them more effective at identifying early signs of failure. However, these methods can be difficult to interpret and typically need large amounts of labeled data to train effectively [16]. Each fault detection approach offers distinct advantages and drawbacks, but none fully satisfies the combined demands of high accuracy, fast response, robustness to varying conditions, and broad applicability required for real-world battery diagnostics [17].

1.1. Research Growth in Battery Fault Detection

In recent years, the growing dependence on Li–ion batteries in EVs and energy storage systems has brought battery fault detection to the forefront of research. The demand for more accurate, efficient, and predictive diagnostic methods has driven significant advancements, especially in AI-based techniques. This surge in innovation is evident from the increasing volume of research publications, as shown in Figure 2, reflecting the field’s fast-paced progress and growing importance to the industry.

As shown in Figure 2, research on battery fault detection has seen a remarkable surge, showing steady year-over-year growth fueled by progress in AI-driven diagnostics and the expanding use of electric vehicles. Research activity has nearly tripled since 2015, signaling a strong commitment from both academia and industry to boost battery safety, efficiency, and reliability. This increase parallels the rapid development of EV markets and energy storage technologies, where machine learning and deep learning methods have greatly improved real-time fault detection and predictive accuracy. The growing research momentum emphasizes the importance of intelligent BMSs and predictive maintenance in extending battery lifespan, minimizing failure risks, and advancing sustainable energy solutions.

1.2. Existing Reviews and Their Limitations

Previous review papers have explored battery fault detection, but most have approached the topic from a broad or non-specialized perspective. For instance, Zhao et al. [18] extend the discussion from lab-scale diagnostics to real-world applications, identifying practical challenges in transferring fault detection models from controlled environments to actual EV deployments. Despite highlighting the behavior of failures in real use, the review falls short of analyzing how modeling frameworks can be scaled or adapted across vehicle types and use cases. Shang et al. [19] present a comprehensive overview of battery fault detection mechanisms, including fault types and detection principles, particularly focusing on parameters such as voltage inconsistencies and distributional analysis across cells. However, it lacks an in-depth discussion on advanced learning techniques like graph neural networks (GNNs) or entropy methods, which are increasingly relevant in modern battery diagnostics.

In another review, Li et al. [20] provide a comprehensive review of fault prognosis techniques for lithium–ion batteries in electric vehicles, covering model-based, data-driven, and hybrid approaches. It highlights the importance of battery chemistry, sensor fusion, and electrochemical modeling in improving fault prediction accuracy. Xu et al. [21] provide a comprehensive overview of model-based fault detection methods for Li–ion batteries, focusing on state-space modeling, observer design, and fault mechanism analysis. It discusses the challenges associated with modeling uncertainties and the implementation of diagnostic algorithms. Wang et al. [22] provide an exhaustive review of battery faults and diagnostic techniques, categorizing faults into mechanical, electrical, thermal, inconsistency, and aging faults. While comprehensive, the review does not delve into the fusion of data for enhanced fault detection. Kumar et al. [23] present a comprehensive review of machine learning-based, data-driven fault detection techniques for lithium–ion batteries. It discusses various algorithms, including artificial neural networks, support vector machines, and random forests, highlighting their applications in battery management systems.

Although prior studies have investigated either modeling independently, a unified framework for battery fault detection remains relatively underdeveloped. This review addresses that gap by presenting modeling as a comprehensive and effective approach for improving early and accurate fault identification. It critically reviews the latest methodologies, including autoencoders, transformer-based attention mechanisms, and hybrid models. Recent innovations, such as graph neural networks (GNNs), large language models (LLMs), and self-attention-based architectures tailored to battery diagnostics are systematically consolidated and discussed, which previous reviews have yet to cover holistically. Furthermore, the review offers a comparative evaluation of these approaches in terms of dataset types, evaluation criteria, and their applicability to real-world battery systems.

Table 1. Comparison of recent review papers on battery fault detection: contributions, gaps, and our advancements.

Study	Contribution	Missing Aspects	Our Contributions
Zhao et al. (2024) [18]	Bridged lab diagnostics with real-world application challenges in fault propagation.	Limited exploration of modeling frameworks in diverse EV architectures.	Evaluates model adaptability across chemistries, topologies, and field data.
Shang et al. (2024) [19]	Comprehensive review of parameters in battery fault detection, including fault types and voltage distribution analysis.	Lacks detailed treatment of learning techniques like GNNs and entropy models.	Introduces a deep analysis of ML/DL models, including GNNs and entropy-based methods.
Li et al. (2025) [20]	Reviewed fault prognosis techniques (model-based, data-driven, hybrid) and their relationship with battery chemistry and sensor fusion.	Does not explore recent developments in deep learning architectures, attention mechanisms, or model interpretability.	Incorporates modern AI frameworks, interpretable learning techniques, and cross-domain fusion for fault prediction.
Xu et al. (2024) [21]	Provided an overview of model-based diagnostic methods emphasizing observer design, state estimation, and fault modeling.	Lacks coverage of learning-based methods that adapt to nonlinear degradation behaviors and uncertain real-world conditions.	Proposes adaptive AI-driven models that account for nonlinear degradation, uncertainty, and real-time implementation challenges.
Wang et al. (2024) [22]	Provides an exhaustive review of battery faults and diagnostic techniques.	Does not delve into the fusion of data.	Highlight the benefits of combining data for enhanced fault detection.
Kumar et al. (2021) [23]	Summarized ML-based fault detection methods (SVM, ANN, Random Forest) for BMS applications.	Omits advanced hybrid modeling, multi-dimensional data analysis, and integration of physics-based reasoning.	Bridges AI and domain knowledge through hybrid models that combine physical insights with robust learning algorithms.

summarizes this comparison, outlining each study’s contributions and the research gaps they leave open.

This review advances the field of intelligent and sustainable transportation by addressing critical challenges in Li–ion battery fault detection through multidisciplinary AI-driven approaches. The key contributions are as follows:

• This review emphasizes the integration of model dependencies in analyzing electric vehicle battery faults, highlighting the importance of modeling cell-level interactions over multiple charge-discharge cycles for early fault detection and predictive maintenance.
• Comprehensive review of ML and DL methods: A systematic evaluation of ML and DL techniques, focusing on their effectiveness in handling fault patterns. This includes CNN-LSTM hybrids and physics-informed neural networks (PINNs), offering a unified discussion on their role in battery fault detection.
• Critical discussion of challenges: This paper identifies challenges, including limited labeled fault datasets and computational constraints in real-time BMS, and addresses the lack of interpretability in deep learning models, proposing strategies to enhance model robustness and explainability.
• Emerging trends and future research directions: This highlights emerging trends and future research directions such as GNNs, LLMs, self-supervised learning, and edge computing for real-time BMS diagnostics. This review provides a strategic roadmap for advancing AI-driven battery fault detection.

The remainder of this paper is organized as follows. Section 2 discusses the fundamentals of EV battery fault detection, focusing on major fault types and their effects. Section 3 provides an in-depth review of ML- and DL-based fault detection techniques, showcasing their effectiveness in analyzing battery degradation patterns. Section 4 discusses critical challenges, including data scarcity, computational constraints, and the interpretability of AI-driven models. Section 5 presents emerging future research directions, such as GNNs, LLMs, self-supervised learning, and edge computing, offering a structured pathway for future developments. Lastly, Section 6 concludes by emphasizing the importance of hybrid AI-based methods for enhancing fault detection accuracy and reliability. This review bridges conventional diagnostic methods with advanced AI-driven fault detection, serving as a valuable resource for researchers and industry professionals.

2. Fundamentals of EV Battery Fault Detection

Before exploring advanced modeling methods, it is essential to first understand the various types of faults that can develop in EV batteries, their underlying causes and consequences, and the spatial and temporal features of battery data that support the need for integrated analysis [24]. Typically, an EV battery pack is composed of multiple individual cells arranged into modules, which are then assembled to form the complete battery system responsible for powering the vehicle [25].

Figure 3 presents a functional overview of a BMS, integrating multiple key operations to ensure battery safety and performance. It highlights core functions such as data acquisition, fault detection, thermal management, charge or discharge control, and battery state estimation and also emphasizes the interaction between sensors, signal processors, control algorithms, and actuators to monitor, manage, and protect the battery system in real-time. The protection function in the BMS refers to safety operations such as high-voltage contactor control, overvoltage response, and short-circuit isolation. These mechanisms help prevent battery failure propagation by promptly disconnecting affected modules. The comprehensive architecture and critical responsibilities of a modern BMS in electric vehicle applications.

2.1. Common Fault Types in Lithium–Ion Batteries

In general, battery faults are classified into two main types: cell-level faults and system-level faults [26]. Cell-level faults originate within individual battery cells and are often the most severe in terms of safety risks. These faults can be classified into two types: gradual progressive faults and abrupt sudden faults [27].

2.1.1. Overcharge Fault

Charging a cell beyond its specified voltage threshold, known as overcharging, is a type of progressive fault that triggers harmful internal chemical reactions. One common consequence is lithium plating, where excess lithium accumulates on the anode surface, leading to increased internal resistance and excessive heat generation [28]. Gradually, overcharging deteriorates the cell’s solid-electrolyte interphase (SEI), potentially resulting in gas formation or the development of microscopic dendrites. If not addressed, this condition can escalate into a sudden fault, as lithium dendrites may extend far enough to puncture the separator, leading to an internal short circuit [29]. An internal short circuit can rapidly generate excessive heat, which may escalate into a thermal runaway event [30]. Therefore, an overcharge fault involves changes over time and can also spread spatially, especially if several cells are overcharged or if the failure of one cell leads to increased temperatures in adjacent cells.

2.1.2. Overdischarge Fault

Overdischarge refers to depleting a cell beyond its safe lower voltage threshold. This condition can cause copper from the anode current collector to dissolve and trigger other permanent chemical changes. Although less instantly hazardous than overcharging, overdischarge results in capacity degradation and contributes to imbalances among cells within a battery pack [31]. Cells subjected to overdischarge can undergo faster aging and may develop internal short circuits when recharged. This fault type usually progresses gradually and often stems from poor BMS calibration or inappropriate usage patterns. Overdischarge not only diminishes long-term performance but can also lead to abrupt drops in capacity over time [32].

2.1.3. Overheating and Thermal Runaway

High temperatures can serve as both an indicator and a trigger of battery faults. One of the most critical and abrupt failures is thermal runaway (TR), where internal heat-generating reactions, often caused by mechanical abuse or internal malfunctions, lead to a rapid temperature increase. This temperature spike then accelerates the chemical reactions even further, creating a self-reinforcing feedback loop that can result in catastrophic failure [33]. Thermal runaway (TR) typically begins in a single cell, often triggered by an internal short circuit or exposure to excessive external heat. As the affected cell’s temperature surges, sometimes exceeding 500 °C, it can ignite the flammable electrolyte and emit hazardous gases. The intense heat generated may then transfer to adjacent cells, raising their temperatures and potentially initiating TR in those cells as well. This chain reaction can rapidly escalate, spreading throughout the battery module or entire pack [34]. TR propagation is a process influenced by the arrangement of cells, the presence of thermal insulation, and the rate at which heat disperses across the pack.

2.1.4. Internal Short Circuit

An internal short occurs when the positive and negative electrodes within a battery cell become directly connected, bypassing the intended electrical pathway. This fault may result from a damaged separator, often due to production flaws, dendrite intrusion, or physical impact. Internal shorts vary in severity; some are mild with high resistance, gradually discharging the cell, while others are severe with low resistance, causing sudden high current flow and rapid heat buildup [35]. A mild internal short might initially result in a slow self-discharge and minimal heating, making it difficult to identify using basic threshold-based detection. Over time, the situation can deteriorate, as the heat produced may further degrade the separator, increasing the size of the short circuit. In contrast, severe internal short presents as a sudden failure, rapidly producing intense heat that can lead to thermal runaway within the affected cell [36]. Identifying internal shorts at an early stage is difficult and remains a critical objective in advanced battery diagnostics. It involves recognizing subtle changes in a cell’s voltage or temperature, often requiring analysis. If a single cell’s voltage declines more rapidly than others while the battery is at rest, it may signal a self-discharge caused by an internal short [37].

2.1.5. Cell Imbalance

Although not a catastrophic failure itself, imbalance occurs when cells within a battery pack diverge in their state of charge or health [38]. This condition can lead to future issues, as an imbalanced cell may reach overcharge or overdischarge limits sooner than others, making it more vulnerable. Such imbalances, whether in temperature or capacity, reflect spatial differences across the pack and typically evolve gradually over multiple charge-discharge cycles, highlighting their temporal nature as well [39]. When a specific module consistently operates at higher temperatures, its cells tend to age more rapidly over time, resulting in capacity degradation. This deterioration eventually creates a performance gap compared with other modules, revealing an imbalance [40]. BMS usually addresses such issues through cell balancing, redistributing charge among cells to equalize their states [41]. However, if the imbalance persists or becomes excessive, it may signal a deeper issue, such as a cell with a partial internal short that repeatedly underperforms in voltage. Beyond individual cell issues, batteries can also experience system-level faults that impact overall performance and safety [42].

2.1.6. Causes and Effects of Faults

Battery faults may arise due to multiple factors, including flaws during manufacturing, gradual wear over time, external damage, or stressful operating conditions [43]. Additionally, Figure 4 presents a classification of typical lithium–ion battery faults, clearly separating those that occur at the cell level from broader system-level issues.

On the other hand, aging-induced degradation, such as the loss of active lithium, breakdown of electrode materials, or the thickening of the SEI, progresses gradually over months or years, eventually resulting in reduced capacity and increased internal resistance [44,45]. Aging alone doesn’t constitute a fault, but it increases a cell’s vulnerability to faults. An older cell with diminished ion diffusion capability is more likely to experience lithium plating, especially during fast charging at low temperatures [46]. Therefore, progressive faults are frequently linked to aging patterns. Tracking how each cell’s capacity or internal resistance changes over time, commonly referred to as state-of-health (SOH) estimation, can help identify cells that are degrading faster than expected, signaling a higher risk of sudden failure [47]. The model presented here is initially pre-trained on a large-scale dataset compiled from previous research, covering a wide range of fault categories and failure scenarios, as summarized in

Table 2. Multiple battery fault types and failure scenarios to support robust and generalized model learning.

Battery Fault	Description
Internal electrical fault	A short circuit forms between the terminals, causing a sudden energy release and excessive heat buildup.
Lithium deposition	Lithium ions accumulate on the anode while charging, lowering battery capacity and increasing safety hazards.
Voltage Overload/Voltage Depletion	Exceeding the recommended voltage or current limits during charging or discharging can cause serious damage to the battery.
Irregular Energy Loss	Slow, unintentional reduction in battery capacity over time caused by ongoing internal chemical processes.
Irregular Capacity Loss	Consistent reduction in the battery’s capacity to hold and supply electrical energy.
Irregular Voltage Loss	Irregular voltage loss indicating possible internal faults or cell imbalance.
Abnormal temperature behavior	Irregular rise in temperature during use, signaling possible underlying faults.
Chemical Leakage	Chemical leakage is typically caused by physical damage or structural failure of the battery.
Charge Imbalance	Uneven charge distribution among battery cells, resulting in reduced efficiency and potential long-term degradation.
Fire Risk	Rapid and excessive heat buildup that may trigger fire hazards or explosive failure.

.

3. Methods for Fault Detection in EV Batteries

Battery fault detection methods have progressed from basic threshold-based rules to sophisticated algorithms powered by machine learning and deep learning. This section outlines the current range of techniques, grouped into three main categories: (1) ML approaches, which rely on manually selected features and traditional algorithms; (2) DL techniques, which automatically extract patterns and are particularly effective for time-series and spatial data. EV battery systems comprise numerous interconnected Li–ion cells grouped into modules and packs. These system-level structures introduce unique diagnostic challenges compared with isolated cell studies. Unlike single-cell diagnosis, which assumes uniform conditions, EV battery packs experience spatial temperature gradients, cell imbalances, and inter-module propagation of faults. Thus, system-level fault detection must account for hierarchical data dependencies, thermal coupling effects, and real-time monitoring requirements across multiple levels. Recent studies have attempted to scale single-cell methods to pack-level diagnostics, but additional attention is required to model inter-cell dynamics and cross-sensor anomaly detection.

In system-level battery applications, the interaction between thermal, electrical, and mechanical factors becomes more complex. A fault that begins in a single cell, such as overheating or overcharging, can affect nearby cells and spread across modules if not detected in time. Additionally, large EV battery systems operate under changing loads and varying environmental conditions, which can create inconsistencies in sensor signals and make it more difficult to locate the source of a problem. To manage these challenges, system-level diagnostics must combine data from multiple sensors and apply techniques that can understand both time-based and spatial relationships. Recent research has explored methods such as graph-based learning, shared data monitoring, and combining different types of sensor data to improve accuracy. However, applying these methods in real time is still difficult because of the large amount of data and high processing requirements. Addressing this issue will require both better algorithms and smarter ways of organizing battery data that reflect how the pack is built and how its parts work together. A comparative overview of these methods, including their data needs and detection performance, is provided in Figure 5 to enable direct comparison and deeper insight.

3.1. Machine Learning Methods

Conventional machine learning approaches for detecting battery faults typically follow a two-step process: initially, relevant features are extracted from raw sensor data; subsequently, an ML algorithm is trained to differentiate between healthy and faulty conditions or to categorize fault types. Feature engineering plays a vital role in this process. Frequently used features include statistical values such as voltage means and variances, temperature rise rates, inter-cell voltage differences indicating imbalance, and parameters derived from electrochemical models through system identification [48]. The machine learning model can function either as a classifier to determine the specific type of fault or as a regression or anomaly detection tool to produce a health index or detect unusual behavior [49]. Commonly used ML techniques in this area include Support Vector Machines (SVMs), Random Forests (RFs), Gradient Boosting algorithms, k-Nearest Neighbors, and unsupervised or one-class classifiers such as one-class SVM and Isolation Forests for identifying anomalies [50].

RF is a widely used ensemble learning method well-suited for structured, tabular battery datasets. In battery fault detection, RF models typically utilize features such as cell voltage variance, temperature differentials across modules, and SOC spread. These features help detect gradual faults like cell imbalance, capacity degradation, or charge heterogeneity. RFs are interpretable and computationally efficient, making them practical for onboard diagnostics [51]. However, they are better suited for static or short-time window analysis and are limited in modeling temporal fault evolution without feature engineering. SVM classifiers perform well in scenarios involving well-defined binary classifications, such as fault or no-fault status. SVMs rely on features like voltage slope, a derivative of SOC, and current inflection points, making them particularly effective in detecting internal short circuits, sudden cell disconnections, or discrete fault events. Due to their margin-maximizing nature, SVMs work best with balanced, high-quality labeled data and are less effective in environments with noise, class imbalance, or drifting fault behavior over time.

Isolation Forest is an unsupervised anomaly detection algorithm that isolates anomalies based on feature partitioning [52]. It is highly effective for identifying unknown or rare faults, particularly in real-time BMS data where labels may be unavailable. Isolated Forest typically uses statistical features such as interquartile voltage range, temperature skewness, or signal dropout patterns. It is best applied to scenarios involving sensor noise, novel battery failures, or fault types not previously seen during training, though it does not provide fault-type classification. Daniels et al. [53] utilized a random forest algorithm to detect indicators of electrolyte leakage in battery cells subjected to external short-circuit testing. Their model was trained on experimental datasets where leakage events were clearly labeled, allowing it to accurately classify new instances as either leakage-related or normal. Random forests are widely used in such cases because they can manage complex, nonlinear relationships among features and offer insights into which features contribute most to a decision, improving interpretability [54]. A segmented regression technique in conjunction with an optimized GRU network to improve detection accuracy during charging and discharging phases. While this work incorporates variations in voltage dynamics, its focus remains on a narrow methodological scope, excluding comparative insights across other modeling paradigms.

Similarly, SVMs have been employed to differentiate between healthy and faulty battery states by constructing an optimal boundary in the feature space. One study used SVMs to detect battery pack faults, highlighting that a large and diverse training dataset is essential to effectively capture various fault scenarios [55]. Some machine learning methods focus on detecting unusual behavior by identifying deviations from normal battery performance rather than labeling specific fault types [56]. For instance, an isolation forest, an ensemble-based anomaly detection algorithm, is used in their fault detection approach. They first processed the voltage data to isolate residual signals and then applied the Isolation Forest to detect anomalies based on these residual patterns [57]. This approach operates without needing labeled fault data, as it learns the typical behavior of voltage signals and identifies unusual patterns, such as sudden drops or unexpected fluctuations. Its main strength lies in reducing reliance on rare fault data, though it can be challenging to fine-tune the model’s sensitivity to prevent false alarms triggered by harmless anomalies.

One major drawback of traditional machine learning is its reliance on manually selected features. If these features fail to accurately represent the characteristics of a fault, the model’s performance can suffer [58]. A metric such as the “voltage range in module X” might successfully detect significant imbalances but could overlook gradual voltage changes in a single cell [59]. To overcome these challenges, researchers have begun merging physical modeling techniques with data-driven approaches. For instance, Ahsan et al. [60] introduced a fault detection strategy that involved estimating key battery parameters, such as internal resistance, using a microcontroller-based setup. These estimated parameters, along with real-time sensor readings, were then used as input features for a data-driven classification model. This combined approach enabled the identification of specific faults like internal short circuits or degradation-related performance drops by associating unusual parameter behavior with known fault signatures [61]. An alternative machine learning technique employs ensemble strategies or layered frameworks to enhance fault identification. For instance, Zhao et al. [62] developed a fault detection method utilizing a three-step screening mechanism that progressively isolates and detects battery abnormalities. Their method involved a probabilistic analysis of irregular cell voltage patterns within an EV battery pack, systematically narrowing down potential faulty cells and fault conditions through a staged evaluation process [63]. This progression sets the stage for exploring deep learning approaches, which are discussed in the next section.

3.2. Deep Learning Methods

Deep learning techniques are increasingly used in battery fault detection because they can automatically extract layered features from raw data without the need for manual feature design. These models are capable of capturing intricate nonlinear behaviors and identifying subtle fault-related patterns that traditional methods may overlook. Various deep neural network architectures have been applied, each tailored to specific data structures and detection needs.

• Convolutional Neural Networks (CNNs) are employed to extract spatial features from battery sensor data by detecting localized anomalies, such as uneven voltage distribution or concentrated areas of excessive heat across individual cells.
• Recurrent Neural Networks (RNNs) and their advanced forms, like Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU), are utilized to capture time-based patterns in battery data, making them effective for identifying anomalies that develop or evolve over time.
• Transformer models apply self-attention techniques to effectively capture long-term dependencies within battery time-series data, enabling more accurate fault identification and advanced predictive analysis across extensive and high-dimensional datasets.
• Hybrid architectures that integrate convolutional and recurrent layers are well-suited for battery data, where CNNs extract characteristics such as cell-level patterns and RNNs maintain relationships to track changes over time.

LSTM networks are a class of RNNs tailored for sequence prediction tasks. LSTM models are effective for detecting gradual faults by modeling long-range dependencies in time-series data. In battery diagnostics, they ingest multivariate time-series inputs such as voltage, current, and SOC trajectories to forecast and detect trends leading to overcharge, capacity fade, or low-rate degradation [64]. Their memory cells allow them to retain historical information, making them highly suitable for long-duration faults. However, LSTMs require large labeled datasets and are computationally intensive. CNNs are traditionally used in image recognition but have been adapted to detect spatial fault patterns in battery packs. By converting sensor array temperature or voltage maps across multiple cells into 2D feature representations, CNNs can effectively detect faults like thermal runaway, cell overheating, or hotspot development. CNN-based models are best suited for structured spatial data and can be extended using 3D-CNNs to incorporate both temporal and spatial correlations [65]. However, they are not ideal for sequential modeling unless coupled with RNNs.

CNNs, traditionally used in image processing, have been creatively adapted for battery systems. In one approach, the sensor layout of a battery pack is visualized as an image, with each cell’s readings representing a pixel or small area [66]. More frequently, CNNs are applied to time-series data by organizing it into a two-dimensional matrix, placing time on one axis and sensor indices on the other, allowing convolutional filters to detect localized features and trends within this structure [67]. A 2D convolutional layer can move both across time steps and different battery cells, making it effective in identifying localized patterns. Expanding on this concept, 3D convolutional neural networks (3D-CNNs) introduce an additional dimension, considering time as a depth axis, enabling the model to learn how spatial features change over short temporal intervals, thus capturing the dynamic evolution of faults more effectively [68]. CNNs are effective at detecting battery faults by recognizing visual patterns, especially when paired with attention mechanisms that help focus on key sensors or time steps [69].

Qiu et al. [70] introduced a battery fault detection model using a stacked sparse autoencoder integrated with a Convolutional Block Attention Module (CBAM) inside a Capsule Network to enhance feature learning and fault detection accuracy. The autoencoder compresses and cleans the input data, while the CBAM-CapsNet focuses on key spatial sensors and temporal steps for fault classification. This targeted attention enhances diagnostic accuracy beyond that of traditional CNNs or fully connected networks [71]. RNNs are well-suited for analyzing sequential battery sensor data [72]. However, due to limitations like vanishing gradients in long sequences, advanced variants such as LSTM and GRU are preferred. These models effectively learn long-term patterns and are widely applied in battery fault detection for forecasting future behavior and identifying anomalies from historical data sequences [73].

Zhang et al. [74] developed an LSTM-based model for battery fault prognosis, enabling early anomaly detection through multi-step voltage prediction. The LSTM framework was designed to forecast battery string voltage over multiple future steps with high accuracy [75]. Studies have shown that RNN-based fault detectors can identify issues like thermal runaway earlier than traditional threshold methods by detecting rapid rise patterns and their nonlinear progression [76]. Additionally, Figure 6 presents an overview of the mechanism behind DL approaches, specifically highlighting the integration of CNN and RNN architectures in a hybrid model for battery fault detection.

DL methods, when paired with sufficient data and optimal architectures, significantly improve fault detection and provide earlier identification compared with traditional methods [77]. A large-scale EV study showed a 46% increase in true positive detection rates, with low false-positive rates [78]. These figures highlight their capability to detect subtle faults. However, careful validation is necessary to avoid overfitting and ensure robustness across different battery types and operating conditions [79]. Despite their advantages, DL models remain black-box systems, making interpretability a concern. To address this, incorporating domain knowledge or using spatio-temporal and hybrid approaches is gaining traction, enhancing both model transparency and performance [80].

Conventional machine learning methods, such as SVM, RF, gradient boosting algorithms, k-NN, and isolation forests, rely extensively on manual feature engineering to enable fault detection in lithium–ion batteries. These techniques require the extraction of domain-specific features from raw sensor data, which are then used as inputs for classification or anomaly detection tasks. Typical features include voltage variance across cells, SOC slope, inter-cell voltage imbalance, and thermal gradients. These metrics are carefully designed to capture battery behavior and degradation signatures. The primary advantage of these ML models lies in their interpretability and computational efficiency, making them suitable for deployment in resource-constrained environments such as onboard BMS. RF models are particularly effective at identifying gradual degradation or imbalance by analyzing structured and tabular datasets. Similarly, SVMs are commonly used for binary fault classification tasks, especially in scenarios where distinct fault signatures can be separated by a decision boundary. In contrast, unsupervised models such as isolation forests are well-suited for detecting anomalies in datasets where labeled fault data are unavailable. However, these traditional ML approaches often struggle to model temporal dynamics unless specific time-domain features are crafted and integrated into the training process.

Deep learning techniques, on the other hand, are increasingly adopted in battery fault detection due to their ability to automatically learn hierarchical and nonlinear feature representations directly from raw sensor inputs. Unlike ML methods, DL models such as LSTM networks, CNNs, and autoencoder-based architectures do not require manual feature engineering. LSTM networks, a class of RNNs, are highly effective at learning long-term dependencies within multivariate time-series data such as voltage, current, and temperature. These models are particularly suited for detecting gradual and evolving faults, including overcharging, capacity fade, and thermal imbalance, by identifying patterns and trends over extended time intervals. CNNs, although originally developed for image classification, have been adapted to battery diagnostics by transforming spatial sensor layouts or sequential data into structured matrix formats. These matrices allow the CNN to detect localized fault patterns, such as thermal hotspots or cell-level anomalies, within a pack. Additionally, hybrid deep learning architectures that combine CNN and LSTM models have demonstrated improved performance by simultaneously capturing spatial and temporal correlations. While DL models offer superior accuracy and robustness in handling complex datasets, they typically require significant computational resources and large volumes of labeled training data, and their internal decision-making processes remain less interpretable than those of traditional ML methods.

To illustrate the practical applications and performance of these approaches, several comparative studies have been included in the revised manuscript. For instance, Daniels et al. [81] applied a random forest algorithm to identify electrolyte leakage in battery cells subjected to external short-circuit testing. Their model leveraged manually extracted features from experimental datasets, demonstrating the effectiveness of RF in structured fault classification tasks. In contrast, Zhang et al. [82] implemented a deep LSTM-based model for early fault detection by predicting voltage trajectories over multiple future time steps. This approach enabled accurate forecasting of cell behavior and timely identification of abnormalities, outperforming conventional models in scenarios involving gradual degradation. Furthermore, advanced DL models such as stacked sparse autoencoders integrated with CBAM and CapsNet have been employed to improve feature learning and fault classification performance. These models enhance spatial and temporal focus through attention mechanisms, leading to greater diagnostic accuracy. Such developments highlight the advantages of DL in extracting complex patterns and making informed predictions, particularly when dealing with high-dimensional, noisy, or incomplete sensor data prevalent in real-world electric vehicle battery systems.

The distinction between ML and DL methods in battery fault detection lies in their handling of data, model complexity, computational demands, and application suitability. ML models are characterized by their reliance on manual feature engineering, faster training times, and interpretability, making them ideal for real-time or embedded diagnostic systems. However, their performance is limited when applied to raw or unstructured data without carefully crafted features. DL models, in contrast, offer a powerful alternative by learning deep and abstract representations from raw input, thereby eliminating the need for hand-engineered features. They are particularly effective for modeling complex spatial-temporal interactions in battery data, especially in large-scale or high-resolution diagnostic environments. Nevertheless, the higher accuracy and flexibility of DL come with increased computational costs and reduced transparency. Both paradigms serve important and complementary roles depending on the specific use case, available data, and deployment environment.

3.3. Transformer Architectures (Self-Attention Networks)

Transformer models, including architectures adapted from NLP, such as BERTtery, have recently emerged as powerful tools for battery fault detection. Transformers leverage self-attention mechanisms to model temporal sequences without recurrent structures, enabling them to learn both short- and long-term dependencies effectively [83]. They are particularly suitable for detecting subtle, slowly evolving faults such as gradual SOC imbalance, aging-related internal resistance increase, or delayed sensor drift. Feature inputs include time-position encoded battery signals (voltage, temperature, current), which allow the model to focus attention on critical sequence regions. Transformers are highly scalable and interpretable but require large datasets and significant computational resources for training.

Transformers have revolutionized sequence modeling in NLP and are now entering time-series analysis, including battery applications. The core idea is the self-attention mechanism, which allows the model to weigh the relevance of different time steps when making a prediction or classification. For long battery sequences, which can be thousands of time points long if considering hours or days of data, transformers are appealing because they can capture long-range dependencies without the time-step-by-time-step propagation of RNNs [84]. A specialized Transformer-based architecture called BERTtery (a play on BERT for batteries) was introduced by Zhang et al. [85] for early fault prediction. Their model employs a two-tower Transformer design: one tower processes the temporal sequence of measurements, and the other processes the channel information, essentially modeling across cells or across different measurement types [86]. By doing so, the model can capture patterns like “a certain shape between the voltage and time curve, coupled with a certain temperature rise pattern across sensors, is an early indicator of failure.”

Another Transformer approach by Guirguis et al. [87] targeted SOC estimation, but it introduced architectural ideas applicable to fault detection. They used a two-tower transformer to handle two different sequences and leveraged multi-head attention to automatically extract temporal features, eliminating the need for manual feature engineering [88]. In the context of fault detection, a similar attention mechanism can highlight that “the cell voltage at time t1 and the temperature at time t2 are strongly linked to the eventual outcome,” which could correspond to a fault signature. Transformers, due to attention, also lend themselves to interpretability to some extent; one can visualize the attention weights to see which time periods or sensors the model focuses on when predicting a fault [89]. This can sometimes reveal known precursors, validating the model, or even new ones that engineers had not noticed. Hybrid models CNN + RNN: there is a trend of combining CNNs and RNNs to exploit structures [90].

A CNN-LSTM hybrid might use convolutional layers to extract features from the pack and then feed these features into an LSTM that captures overall evolution [91,92]. An application of this is found in a study where a PCC-CNN-LSTM model was used for EV battery fault detection [93]. The “PCC” likely denotes some preprocessing like the Pearson Correlation Coefficient matrix as input; the CNN extracts features from that matrix, and the LSTM models the sequence of those features to predict faults [94]. Such hybrid models have shown improved accuracy over pure CNN or pure LSTM because they can simultaneously handle multi-sensor correlations and time dependencies. Another emerging architecture is the use of autoencoders for unsupervised anomaly detection, followed by sequence models [95]. In some cases, simulation or physics-based models might generate synthetic fault data, like a thermal propagation simulation, to augment training, though the gap between simulation and reality must be managed [96].

3.4. Hybrid Approaches for Battery Fault Detection

Hybrid fault detection frameworks in lithium–ion battery systems have emerged as a powerful paradigm that integrates the rigor of physics-based modeling with the adaptability of machine learning techniques. This dual approach capitalizes on the strengths of both methodologies: the interpretability and structure of physical models and the pattern recognition capabilities of data-driven models. One of the most widely adopted hybrid strategies is residual-based learning. IIn this approach, a deterministic battery model, such as an equivalent circuit model, is used to simulate the expected electrical or thermal response under defined operating conditions. The discrepancy between the simulated output and actual sensor data, known as the residual, often reveals anomalous behavior not captured by the model [97]. These residual signals are then fed into machine learning classifiers, such as SVM, gradient boosting trees, or neural networks, which are trained to detect and classify specific fault patterns. This architecture reduces the complexity of the learning task, as the machine learning algorithm focuses solely on capturing deviations that reflect real-world anomalies rather than modeling the entire battery system from scratch.

Another prominent class of hybrid models involves PINNs. Unlike traditional black-box networks, PINNs incorporate the fundamental laws governing battery behavior, such as conservation of charge, energy, and electrochemical kinetics, directly into the training process by embedding partial differential equations within the model’s loss function. This constraint ensures that the neural network learns solutions that are not only data-consistent but also physically valid [98]. PINNs are particularly valuable in scenarios where labeled fault data is sparse but robust theoretical models exist. For instance, they can detect internal short circuits or thermal instabilities by ensuring the learned predictions adhere to Ohm’s law, thermal diffusion, and mass transport equations.

In addition, emerging hybrid models extend beyond residual analysis and PINNs by integrating neural networks with electrochemical or multi-physics simulation platforms. These systems employ deep learning to approximate unmodeled dynamics, such as degradation behavior or non-linear hysteresis, enhancing the predictive accuracy of long-term battery performance and health forecasts [99]. The synergy between simulation accuracy and machine learning adaptability allows these hybrid frameworks to operate effectively in uncertain, noisy environments typical of electric vehicle applications. More importantly, they enhance system safety and real-time diagnostic capability, providing a viable path toward scalable, interpretable, and robust battery management in the next generation of electric mobility systems.

To bridge the gap between fault classification (as discussed in Section 2) and corresponding detection methodologies (explored in Section 3),

Table 3. Correlation between fault types and suitable detection methods.

Fault Type	Suitable Methods	Key Features Used	Best Use Case Scenario	Data Requirement
Overcharge	LSTM, Transformer	Voltage increase over time, SOC drift	Gradual faults during prolonged charging	Time-series, labeled
Thermal Runaway	CNN, 3D-CNN	Localized temperature spike patterns	Rapidly evolving spatial-temperature faults	sensor data
Internal Short Circuit	Residual and SVM, GRU	Unexpected voltage or current deviation	Sudden failures under normal operation	Real and model residual data
Cell Imbalance	Random Forest, Isolation Forest	Voltage variance across cells	Asynchronous cell behavior during discharge	Structured tabular data
Overdischarge	GRU, Isolation Forest	Declining voltage slope, low terminal voltage	Degradation under prolonged use	Time-series
Sensor Fault	Autoencoders, Isolation Forest	Signal dropout, constant value anomaly	Random or infrequent sensor failure	Unlabeled, anomaly detection

below provides a fault method correlation matrix. This matrix aligns specific types of battery faults commonly found in electric vehicle applications with the most suitable detection techniques based on their underlying features, performance characteristics, and typical data requirements. This mapping enables researchers and practitioners to understand which diagnostic tools are best suited for particular fault scenarios, thereby supporting targeted deployment and system-level optimization.

4. Challenges and Open Problems

Despite notable advancements in battery fault detection technologies, there are still several unresolved challenges. Addressing these challenges in

Table 4. Summary of major challenges and unresolved issues in battery fault detection.

Challenge Area	Description
Data Availability	The need for new strategies to acquire and share data across industries.
Computational Complexity	Requires advancements in algorithm efficiency and possibly the introduction of specialized hardware.
Model Flexibility	Calls for the creation of more adaptable models that are aware of physical constraints.
Model Transparency	Emphasizes the necessity of making models more interpretable for better decision-making.
Fault detection Accuracy	Strives for more precise and reliable identification of battery faults over time.
Real-Time Detection	Highlights the challenge of implementing fast, real-time fault detection systems.
Generalization Across Systems	The need for models that generalize well across different battery types and usage conditions.
Hybrid Approaches	Encourages combining physical models with data-driven techniques for more accurate predictions.

is essential to ensure reliable, real-time fault detection in commercial EV battery packs and to foster the widespread adoption of advanced methods in the industry.

Figure 7 outlines six key obstacles hindering the implementation of advanced fault detection in real-world BMS. These include the lack of labeled fault data, limitations in real-time deployment, difficulties in integrating with control systems, vulnerability to sensor noise and anomalies, challenges in interpreting complex models, and the issue of generalizing across different chemistries and applications. Overcoming these barriers is crucial for achieving reliable, scalable, and intelligent battery health monitoring in electric vehicles.

4.1. Scarcity of Labeled Fault Data

High-quality battery fault detection is limited by the limited availability of labeled data, especially for severe failures like thermal runaway. Real-world EV failures are rare and often undocumented, making it hard to train supervised models. Researchers rely on lab-induced or synthetic data, which may not reflect real-world conditions [100]. The lack of standardized, shareable datasets also hinders reproducibility. As a result, models risk overfitting to limited examples, and replicating the complex causes of real faults like manufacturing differences or environmental stress is difficult in controlled settings [101].

As highlighted by Zhao et al. [102], uncertainties in battery materials, manufacturing differences, and the lack of high-quality datasets make it difficult to accurately model battery failures. While unsupervised and self-supervised learning methods offer potential since they don’t rely on labeled fault data, they still require large volumes of normal operational data and often struggle to clearly define what counts as a fault, especially when the line between normal and abnormal is not obvious [103]. To overcome these issues, collaborative efforts such as data-sharing initiatives through industry consortia or government-backed programs are needed to build large, standardized datasets that include both healthy and faulty battery behavior. Additionally, techniques like active learning and reinforcement learning could be used to intelligently design experiments that generate more useful fault data in a safe and controlled manner [104].

4.2. Real-Time Deployment Challenges

Many advanced fault detection algorithms, such as deep neural networks and graph-based models, are highly complex and require significant memory and computational resources. However, most BMS in electric vehicles use low-power microcontrollers with strict real-time processing limits [105]. These systems must analyze data at high sampling rates across hundreds of channels, making it difficult to run large models like CNNs or Transformers in real time. Although some modern vehicles use more capable processors, cost and energy constraints often limit their use. Therefore, a major challenge is simplifying these heavy models into lightweight versions that can operate efficiently on embedded systems [106]. Solutions from machine learning, such as model pruning, quantization, and knowledge distillation, where a smaller model learns from a larger, one offer promising ways to reduce model size while preserving performance [107].

Kim et al. [108] developed a microcontroller-based fault detection method optimized for embedded systems. This proves that, with smart design, advanced techniques like parameter estimation can work within hardware limits. Additionally, dedicated chips like FPGAs or neural accelerators can handle heavier models in real time, though they may raise system costs.

4.3. Generalization and Transferability Issues

Machine learning models for battery fault detection often face challenges in generalizing across different battery chemistries, configurations, and usage patterns. A model trained on one setup may misclassify new patterns in others, highlighting the need for domain adaptation techniques to bridge this gap. Additionally, the lack of interpretability in deep learning models raises concerns in real-world applications, especially for safety validation and regulatory approval. To address this, explainable AI (XAI) methods are being explored to provide clearer reasoning behind fault predictions, improving trust and transparency in battery health management systems [109].

A BMS engineer or safety expert needs to understand the reason behind a fault detection decision. Traditional model-based approaches offer straightforward logic; for instance, a voltage reading exceeding a predefined limit clearly indicates a fault [110]. In contrast, complex neural networks may generate fault alerts based on intricate combinations of sensor inputs, often lacking immediate interpretability. This lack of transparency presents several challenges: (1) it becomes difficult to validate or troubleshoot the model when it makes an incorrect prediction; (2) users or operators are left without insights into the underlying cause or affected location when a fault is detected; and (3) meeting regulatory standards, especially in automotive safety, often requires systems to demonstrate clear, explainable behavior [111].

4.4. Integration with BMS and Control Actions

Detecting a fault is just the first step in ensuring the proper functioning of a BMS. Once a fault is identified, the system must take appropriate action. However, this comes with a challenge: if the fault detection algorithm is too sensitive, it may trigger unnecessarily, leading the system to throttle power or take the battery pack offline too often. This could negatively impact the vehicle’s performance and availability [112]. On the other hand, if the algorithm is too conservative, it might delay the necessary actions, allowing faults to worsen before they are addressed. Striking the right balance between sensitivity and conservatism is critical to ensure that the BMS responds to faults in a timely and efficient manner without causing unnecessary disruptions.

• Data challenges: necessitate the development of novel training approaches and potential industry partnerships for data sharing.
• Computational obstacles: drive the need for algorithm optimization and possibly the introduction of new hardware solutions.
• Generalization: requires the creation of more adaptable and physics-informed models.
• Interpretability: calls for improved model transparency to enhance understanding and trust in the outcomes.

The following section explores emerging research directions designed to address these challenges and drive further advancements in the field.

5. Future Research Directions

Looking ahead, there are several promising research areas and technologies that could significantly improve battery fault detection, particularly from a spatio-temporal perspective. These advancements aim to tackle the challenges outlined earlier and take advantage of emerging developments in machine learning and battery modeling. Figure 8 provides a brief introduction to the Circular Hierarchical Roadmap for Future Research in Battery Fault Detection.

5.1. Self-Supervised Learning for Battery Systems

Drawing inspiration from the success of foundation models in fields like natural language processing and computer vision, researchers are now turning their attention to time-series data, with a focus on battery data [113]. The approach involves developing large-scale, pre-trained models capable of handling vast amounts of time-series information from a wide array of domains. By training these models on diverse datasets, including data from various fleets of batteries, the goal is to create a highly versatile model [114]. Such models would be able to understand and predict battery performance and faults more effectively by leveraging the patterns found in time-series data. This method promises to enhance the accuracy and efficiency of battery monitoring systems, offering valuable insights for managing and optimizing battery health across different applications [115].

5.2. GNNs for Improved Spatial Learning

Graph-based representations of battery packs are inherently suited for analysis, and GNNs are an emerging and rapidly developing area of research [116]. Future studies could explore different variants of GNNs to enhance spatial learning, improving upon the methods used in earlier attempts. One promising approach involves GATs, which could enable models to identify the most crucial cell connections for fault propagation, providing a more targeted and effective way of understanding battery behavior and detecting issues [117].

Another approach involves incorporating time into graph structures, where each temporal slice of the battery pack at a given time (t) is connected to the next slice at time (t + 1). This creates an extensive graph, but GNN frameworks, such as structural RNNs or temporal message passing, are well-equipped to manage such complex structures [118].

5.3. LLMs for Battery Fault Detection

LLMs are emerging as powerful tools for battery fault detection because of their ability to process complex patterns and reason over multimodal data. Initially designed for text-based tasks, LLMs are now showing promise in sensor-driven domains [119]. Their capacity to generate explainable outputs adds to their value in this context. Recent research, such as SensorLLM, has demonstrated the effectiveness of LLMs in aligning sensor data for improved fault detection. This marks a significant shift in how sensor data can be analyzed. LLMs could enhance the accuracy and interpretability of battery fault detection. Their potential in this field continues to grow [120].

Other research has focused on combining LLMs with ontologies and knowledge graphs to enhance interpretability in maintenance diagnostics [121]. LLMs have also been applied in industrial settings for querying sensor data and performing semantic reasoning. They are used to predict faults by analyzing multimodal data sources and can transform unstructured logs into valuable, actionable insights for maintenance [122].

5.4. Physics-Informed AI and Hybrid Models

Merging physical models of batteries with data-driven models offers a balance between interpretability and the need for less data while also providing flexibility and accuracy [123]. One such approach is PINNs, where the loss function of a neural network incorporates terms from differential equations. This ensures that the network’s outputs align with physical laws to some degree. For battery fault detection, PINNs can help improve the model’s adherence to real-world behavior, enhancing both performance and reliability in predicting battery health [124].

Karniadakis et al. [125] discussed how physics-informed machine learning can integrate data with partial differential equations (PDEs), and this approach is already being applied to batteries in several research groups. One example is the use of PINNs for optimizing battery charging. Another hybrid approach involves augmenting electrochemical models, where simplified models are enhanced with data-driven methods to improve accuracy and efficiency in battery-related tasks [126]. This concept can be extended by using an equivalent circuit model to generate baseline voltage predictions. A neural network is then used to process the residuals, which represent the difference between the actual and model-predicted voltage, to detect faults. In this setup, the neural network only needs to learn the unmodeled phenomena, such as internal shorts or sensor drifts, making the task easier compared with learning all the complex dynamics from scratch.

6. Conclusions

Li–ion batteries are fundamental to the continued growth and performance of EVs, yet their ability to develop internal faults poses ongoing challenges to safety, reliability, and long-term efficiency. Fault detection, particularly predictive and early-stage fault prognosis, has become essential for identifying and mitigating risks before they escalate into catastrophic failures. This review has provided a comprehensive overview of fault detection approaches, moving beyond conventional threshold-based techniques to explore advanced machine learning and deep learning frameworks.

Among the key methods surveyed, deep learning architectures such as CNNs, LSTMs, and Transformer-based models show strong potential in large-scale battery systems. Hybrid models that combine physics-based insights with data-driven learning are particularly promising, offering a balance of accuracy, interpretability, and adaptability. Multi-sensor data fusion, transfer learning, and health-index prediction further strengthen diagnostic performance in real-world settings.

To address these, recent advances such as XAI and edge computing show potential for safe and scalable deployment in BMS. Looking ahead, LLMs open up exciting new possibilities for battery diagnostics, from adaptive health modeling to automated reasoning and maintenance support. Ultimately, achieving robust and intelligent fault detection will require close collaboration across academia, industry, and regulatory bodies. Such collective efforts will accelerate the development of next-generation diagnostic tools that not only enhance battery safety and lifespan but also support the global shift toward sustainable electric mobility.