In this review, I focus on diagnostic algorithms for safety issues of lithium-ion batteries in electric cars. The growing ownership of electric cars has been accompanied by an increasing number of fire accidents caused by battery failures. In my analysis, battery thermal runaway is considered the leading root cause of such catastrophic events, and the need for robust fault diagnosis is therefore critical. I first summarize the current safety challenges of lithium-ion batteries in electric cars, then explain the intrinsic electrochemical mechanisms that lead to failure under elevated temperature conditions. I further classify and review both model-based and non-model-based diagnostic algorithms under three common abuse conditions: mechanical, electrical, and thermal. Technical challenges and promising research directions are also addressed, including big-data-driven health monitoring, real-time online diagnosis, cloud computing, and machine learning.
Throughout this work, I emphasize that a reliable advanced battery management system is essential for the safe deployment of lithium-ion batteries in electric cars. Since the first commercial lithium-ion cells entered the market, their high specific energy, long cycle life, low self-discharge rate, and negligible memory effect have propelled the transition of road transportation from fuel-based vehicles to pure electric cars. However, lithium-ion batteries in electric cars may encounter various abusive conditions during daily use, such as overcharging due to capacity fade, overheating caused by cooling system failure, and severe mechanical deformation during collisions. These abnormal events can drive the cell temperature to excessive levels and trigger a series of exothermic side reactions, eventually resulting in thermal runaway and destructive fires. According to my review of recent field data, an electric-car battery fire can last for hours after the initial thermal event, with multiple jet flames and complex re-ignition behaviors, thus presenting substantial challenges for firefighting rescue operations.
1. Current Safety Issues of Lithium-Ion Batteries in Electric Cars
Lithium-ion cells used in electric cars are commonly manufactured in three form factors: cylindrical, prismatic, and pouch. From a safety perspective, the larger the cell size, the harder it is to ensure uniform internal reliability. Therefore, the risk ranking generally increases in the order of cylindrical, pouch, and prismatic cells under particular conditions. Cylindrical cells have small individual volumes and moderate gaps between cells, so their heat dissipation is often better. Prismatic cells offer higher capacity and better system stability, but they suffer from poor heat rejection and significant cooling challenges. Pouch cells are enclosed by aluminium-laminated films, which are mechanically weaker, and thus they tend to swell and release gas rather than explode when internal faults arise. In my experience as a battery safety engineer, the choice of cell form factor is a trade-off between integration technology and battery performance. Some electric cars integrate thousands of small cylindrical cells, while others adopt a much smaller number of large-format pouch or prismatic cells. The larger the number of series-parallel connected cells, the greater the difficulty in managing cell-to-cell inconsistency and safety boundaries. Even under moderate operating conditions, the generated heat cannot be fully dissipated, especially in a densely packed battery pack, and this localized temperature rise easily induces undesirable chemical reactions that become uncontrollable in a short period.
The chain of battery safety events is complex. Figure 1, shown above, illustrates the typical progression from an electric-car battery fault to a fire disaster. During the initial stage, a trigger factor — such as a mechanical crash, electrical abuse, or local thermal spike — may cause a single cell to enter thermal runaway. If the thermal runaway is not suppressed, it propagates to adjacent cells and modules, eventually leading to a pack-level fire that can spread to nearby electric cars and even the surrounding environment. The entire process can last from minutes to hours, and it often involves toxic smoke and violent flame jets. Therefore, early detection of internal short circuits and abnormal temperature rise is one of the most important tasks that an electric-car battery management system (BMS) must perform.
1.1 Overview of Safety Issues in Electric-Car Lithium-Ion Battery Systems
In my review, I identify several common failure mechanisms. Figure 2 shows four typical abuse scenarios: mechanical abuse during collisions, electrical abuse during charging, thermal abuse caused by insufficient heat dissipation, and local heat impingement from an external fire. Each scenario has unique characteristics. In a crash, the cell is compressed or punctured, creating localized high-stress regions through the separator. If the separator fractures, active electrode materials come into direct contact and initiate an internal short circuit. The consequent Joule heat and exothermic electrochemical reaction rapidly overheat the cell. During charging, abnormal overcurrent or overvoltage conditions can accumulate heat inside the cell faster than the cooling system can remove it; because the heat is trapped, the electrode–electrolyte interfaces degrade and release flammable gases. During battery operation, if the waste heat cannot be distributed to the cooling system, the cell core becomes significantly hotter than the surface, and this non-uniform temperature profile accelerates the aging and decomposition of the electrolyte. Finally, when a battery is subjected to local heating from an external flame or a nearby burning object, the temperature gradient across the cell can cause severe internal damage and lead to immediate ignition.
For a more systematic comparison, I summarize the causes, side reactions, and consequences of battery failures in the following table.
| Abuse category | Typical causes | Key side reactions | Possible failures |
|---|---|---|---|
| Mechanical | Collision, puncture, vibration, deformation | Separator rupture, internal short circuit, lithium plating | Heating, fire, explosion |
| Electrical | Overcharging, over-discharging, external short circuit | SEI decomposition, electrolyte decomposition, gas generation, lithium dendrite growth | Capacity loss, thermal runaway |
| Thermal | Local heating, cooling failure, extreme ambient temperature | Cathode oxygen release, polymer binder decomposition, separator melting | Cell rupture, fire propagation |
| Internal defects | Manufacturing impurities, electrode misalignment, contamination | Micro-short circuit, accelerated aging | Gradual capacity fade, sudden failure |
| External system faults | Sensor offset, connection loosening, cooling valve failure | Poor state estimation, uneven temperature distribution | Reduced safety margin, functional degradation |
Faults can also be categorized by their temporal evolution. Sudden or abrupt faults, such as external short circuits, appear in a very short time and require the BMS to react within milliseconds or seconds. Progressive faults, including internal lithium plating, SEI thickening, and gradual capacity loss, evolve over many charge–discharge cycles. Both categories can be effectively identified only if the monitoring algorithms are sensitive enough to capture subtle deviations of voltage, current, temperature, and incremental capacity from their expected behavior.
1.2 Electrochemical Characteristics and Intrinsic Generation Mechanisms
To understand why safety failures happen, I need to examine the internal electrochemical reactions of the cell. A lithium-ion cell consists of a negative electrode (usually graphite or silicon-based), a positive electrode (e.g., LiNiCoAlO₂, LiFePO₄, or LiMn₂O₄), a separator that electrically isolates the electrodes, and an electrolyte that facilitates lithium-ion transport. The separator is designed to block electrons but allow ions to pass through. Under normal conditions, the solid electrolyte interphase (SEI) formed on the anode surface is essential because it prevents continuous electrolyte decomposition and provides a protective barrier against further side reactions. However, as the cell temperature rises, the SEI begins to decompose and re-form repeatedly. When the temperature reaches approximately 80–100 °C, SEI decomposition accelerates, leading to a direct reaction between the lithiated graphite and the electrolyte. This parasitic reaction produces flammable gases such as hydrogen and ethylene (derived from carbonate solvents), which increases the internal pressure and can rupture the cell casing. The decomposition of conventional carbonate electrolytes also forms additional gas species including CO, CO₂, and hydrocarbons.
From the perspective of thermal stability, the failure sequence can be described by the chain of exothermic reactions that occur in distinct temperature ranges. A simplified mathematical representation based on an Arrhenius-type kinetic model is often used:
$$ \frac{d\alpha}{dt} = A \exp\left(-\frac{E_a}{RT}\right) f(\alpha), $$
where $\alpha$ is the conversion degree of the decomposition process, $A$ is the pre-exponential factor, $E_a$ is the activation energy, $R$ is the universal gas constant, $T$ is the absolute temperature, and $f(\alpha)$ is a kinetic function that depends on the reaction mechanism. When the self-generated heat exceeds the heat dissipating capability of the cell, a positive feedback loop occurs:
$$ \rho C_p \frac{\partial T}{\partial t} = \nabla \cdot \left(k \nabla T\right) + \dot{q}_{\text{exo}} + \dot{q}_{\text{joule}} – \dot{q}_{\text{cool}}, $$
where $\rho$ is the density, $C_p$ is the specific heat capacity, $k$ is the thermal conductivity, $\dot{q}_{\text{exo}}$ denotes the exothermic side reaction power, $\dot{q}_{\text{joule}}$ is the Joule heating power, and $\dot{q}_{\text{cool}}$ represents the cooling power. If the local temperature continues to increase and the trigger thermal events spread to adjacent cells, thermal runaway propagation occurs. The critical condition for an accelerating temperature rise can be expressed through the Semenov number:
$$ \mathrm{Se} = \frac{V A \Delta H E_a \exp\left(-E_a / RT_0\right)}{h A_s R T_0^2}, $$
where $V$ is cell volume, $A$ is the pre-exponential factor, $\Delta H$ is heat of reaction, $h$ is heat transfer coefficient, $A_s$ is cooling surface area, and $T_0$ is the reference temperature. When $\mathrm{Se}$ exceeds a critical threshold, the system enters thermal runaway. In my review, I note that cathode oxygen release from layered metal oxides becomes active typically above 150–200 °C. Oxygen generated from the cathode then reacts with the combustible electrolyte, causing a dramatic temperature jump. At the same time, the polyvinylidene fluoride binder decomposes and releases hydrogen fluoride, while the separator (especially polyolefin membranes) begins to melt. The melting behavior of the separator increases the cell internal resistance and thereby reduces the current temporarily; however, the collapse of the separator also creates localized regions where the anode and cathode may contact, inducing internal short circuits.
The sequence of thermal-runaway reactions can be summarized in the following table for three representative temperature intervals.
| Region | Approximate temperature range | Dominant reactions | Observable signatures |
|---|---|---|---|
| SEI growth/decomposition | 70–130 °C | SEI dissolution and re-formation, anode–electrolyte reactions | Mild gas generation, cell swelling |
| Electrolyte decomposition and separator collapse | 130–180 °C | Solvent decomposition, HF release, separator shutdown/melt-down | Rapid voltage drop, internal pressure rise |
| Cathode oxidation and full thermal runaway | 180–300 °C | Cathode lattice oxygen release, electrolyte combustion, binder decomposition | Abrupt temperature spike, fire, explosion |
Because lithium-ion batteries contain high energy in a small volume, any uncontrolled chemical reaction can lead to a catastrophic energy release. Therefore, I emphasize that the key to battery safety lies in predicting and detecting these early stages while the cell is still in a reversible pre-abuse condition. Diagnostic algorithms that can estimate the internal temperature, the degree of SEI degradation, and the severity of micro-internal-short-circuits are essential for enhancing electric-car safety.
2. Safety Fault Diagnosis Methods
In this section, I systematically review fault diagnosis algorithms for lithium-ion batteries in electric cars. A complete diagnostic process typically consists of three stages: fault detection, fault isolation, and fault mitigation. The goal is not only to detect that a fault has occurred but also to identify which type of fault is present and to estimate its severity. From a technical standpoint, the diagnostic algorithms can be classified into two major groups: model-based and non-model-based methods. Model-based methods rely on an electrochemical, equivalent-circuit, thermal, or coupled model to predict the expected voltage and temperature; the residual between the measured and predicted outputs is then monitored. Non-model-based methods, by contrast, analyze signals and data directly through statistical features, information theories, or machine learning classifiers.
2.1 Classification of Safety Faults
Faults can be divided from different perspectives. In my work, I adopt a practical classification that combines fault location and triggering conditions. Internal battery faults include overcharging, over-discharging, internal short circuits, external short circuits, overheating, accelerated degradation, and thermal runaway. External faults include sensor failures (voltage, current, temperature), cooling system faults, and cell-connection faults. The cause–effect relationships of these faults are illustrated in the following conceptual table.
| Type | Fault | Potential cause | Detection feature |
|---|---|---|---|
| Internal faults | Overcharge | BMS algorithm failure, incorrect cell state estimation, charger malfunction | Voltage exceeding upper cut-off limit, abnormal temperature rise |
| Over-discharge | Long-term storage with low SOC, cell-to-cell inconsistency | Copper dissolution, voltage reversal | |
| Internal short circuit | Separator puncture, lithium dendrite penetration, mechanical abuse | Self-discharge increase, unequal cell terminal voltages | |
| External short circuit | Crush of pack, insulation failure, contamination | High current spike, voltage collapse, fast temperature rise | |
| External faults | Sensor faults | Harness disconnection, component aging, electromagnetic interference | Out-of-range measurement, frozen signal, unrealistic signal jumps |
| Cooling system fault | Pump failure, coolant leakage, channel blockage | Cell temperature drifting above expected set-point | |
| Connector fault | Loose busbar, corrosion, vibration fatigue | Abnormal voltage drop across connector, arc signal, local hot spot |
Among these faults, internal short circuits and thermal runaway have the most severe consequences. Internal short circuits can be induced by microscopic lithium dendrites, which grow during fast charging in low-temperature environment or when the anode is over-lithiated. A dendrite can slowly pierce the separator and create a small local short path. The local leakage current gradually drains the affected cell and produces local heat. Because the initial heat generation is small, the BMS may not trigger an alarm; however, the local temperature accelerates the surrounding separator shrinkage, which in turn expands the shorted area. This positive feedback eventually causes catastrophic thermal runaway.
2.2 Diagnostic Algorithms for Safety Faults
In the following, I present a comprehensive summary of diagnostic algorithms for battery safety issues in electric cars. According to the literature, these algorithms can be organized into seven categories: state estimation, parameter estimation, equivalent-space methods, structural analysis, signal processing, knowledge-based methods, and purely data-driven methods. Table 3 lists these categories, indicates the core diagnostic rules, and gives representative methods that I have found to be effective.
| Algorithm category | Core diagnostic rule | Representative techniques |
|---|---|---|
| State estimation | Estimate system states from the model; detect faults from residuals between estimated and measured states | Particle filter, extended/unscented Kalman filter, Luenberger observer, nonlinear observer, sliding-mode observer |
| Parameter estimation | Estimate model parameters from measurements; identify faults through anomalous parameter changes | Recursive least squares, Lyapunov-adaptive law, joint state-parameter estimation |
| Equivalent space | Generate residuals based on input/output algebraic relations | Parity equations, nonlinear parity relations |
| Structural analysis | Analyze the over-determined parts of a model graph to isolate faults | Grey-box structural modelling, bipartite graph matching |
| Signal processing | Transform measured signals into fault-sensitive features (entropy, correlation, frequency) and monitor their deviations | Wavelet transform, Shannon entropy, Pearson correlation, Z-score analysis |
| Knowledge-based | Use human experience or expert rules to make decisions | Threshold rules, fuzzy logic, expert systems |
| Data-driven | Train a classifier or anomaly detector using labeled/unlabeled historical data | Random forest, support vector machine, neural network, deep learning, sparse data observer |
For electric-car battery systems, model-based algorithms often start with an equivalent-circuit model to track the terminal voltage under different current load conditions. A generic equivalent-circuit model with one RC pair has the following equations:
$$ \dot{U}_{1} = -\frac{U_1}{R_1 C_1} + \frac{I}{C_1}, $$
$$ U_{t} = U_{\mathrm{ocv}}(z) – U_1 – R_0 I, $$
where $U_1$ is the polarization voltage across the RC pair, $R_1$ is the polarization resistor, $C_1$ is the polarization capacitor, $R_0$ is the ohmic resistance, $I$ is the current, $z$ is the state of charge (SOC), and $U_{\mathrm{ocv}}(z)$ is the open-circuit voltage as a function of SOC. State estimation filters such as the extended Kalman filter can recursively estimate the SOC and $U_1$ while also detecting sensor faults:
$$ \hat{x}_{k}^{-} = f\left(\hat{x}_{k-1}, u_{k-1}\right), $$
$$ P_{k}^{-} = A_{k} P_{k-1} A_{k}^{\mathsf{T}} + Q, $$
$$ K_{k} = P_{k}^{-} C_{k}^{\mathsf{T}} \left( C_{k} P_{k}^{-} C_{k}^{\mathsf{T}} + R \right)^{-1}, $$
$$ \hat{x}_{k} = \hat{x}_{k}^{-} + K_{k}\left( y_{k} – h\left( \hat{x}_{k}^{-} \right) \right). $$
A residual $r_k = y_k – h(\hat{x}_k^{-})$ with a magnitude greater than a threshold indicates a possible sensor fault or a system fault. In general, model-based techniques allow fast and physically interpretable diagnosis, but they suffer from modeling inaccuracies when the battery ages or when cell-to-cell variations occur. Electrochemical parameters such as the diffusion coefficient and exchange current density are not known exactly, so the observer gains need to be carefully tuned to avoid false alarms.
Signal-processing methods have become popular for battery fault diagnosis in electric cars due to their low computational burden and model-free nature. For example, the Shannon entropy of the voltage or temperature sequence can quantify the uncertainty and randomness caused by early faults. Assume a window of $N$ normalized voltage increments $\Delta U_i$; these increments are partitioned into $M$ bins and the probability of bin $j$ is $p_j$. The Shannon entropy is then defined as:
$$ H = -\sum_{j=1}^{M} p_j \log_2 p_j . $$
In healthy cells, voltage differences between adjacent cells are stable and therefore the probability distribution is concentrated, resulting in a low entropy value. When a micro-short circuit or abnormal internal resistance develops in one cell, the voltage increments become irregular and the entropy increases rapidly. The moving-window entropy can be computed online from a rolling buffer of measurements. In one of my recent investigations, I have noticed that a Z-score based on the correlation coefficient between the voltages of neighboring cells is particularly effective for locating the faulty cell:
$$ z_{i} = \frac{x_{i} – \mu_i^{\text{ref}}}{\sigma_i^{\text{ref}}}, $$
where $x_i$ is the real-time voltage of cell $i$, $\mu_i^{\text{ref}}$ and $\sigma_i^{\text{ref}}$ are the mean and standard deviation obtained from a historical healthy reference. A cell whose Z-score exceeds a high threshold for several consecutive sampling intervals is classified as an outlier. The advantage of such non-model-based approaches is that they have no dependency on precise battery parameters. The disadvantage is that they may produce many false alarms under dynamic driving conditions with rapid current changes.
For multi-fault diagnoses in large battery packs, I have found that the voltage correlation matrix is an excellent tool. Define the Pearson correlation coefficient between cells $i$ and $j$ during a sliding time window:
$$ \rho_{ij} = \frac{\sum_{k=1}^{N} (v_{i,k} – \bar{v}_i)(v_{j,k} – \bar{v}_j)}{\sqrt{\sum_{k=1}^{N} (v_{i,k} – \bar{v}_i)^2} \sqrt{\sum_{k=1}^{N} (v_{j,k} – \bar{v}_j)^2}}. $$
In a normally operating pack, all cells exhibit a high positive correlation because they are exposed to the same load current. If one cell suffers from an internal short circuit, its terminal voltage tends to deviate from the pack average and the correlation coefficient with all other cells decreases significantly. By comparing the correlation coefficients matrix, the faulty cell can be isolated even in the absence of a full battery model. This method has been validated using field data from electric cars, where an internal short circuit often develops gradually over hundreds of cycles before the thermal runaway event.
Cloud-based and big-data-driven algorithms have recently received much attention. In electric cars equipped with telematics units, the central cloud platform can collect battery voltage, current, temperature, SOC, and charging records in real time. The cloud has almost unlimited storage and strong computing capability. A representative approach is the sparse-data observer, which uses only a small amount of data during charging and discharging to detect multiple faults. The method first computes the average normalized voltage score for each cell. Suppose we have $N_s$ cells in series; the normalized voltage of cell $j$ at time step $k$ is:
$$ \Delta v_j(k) = \frac{v_j(k) – \bar{v}(k)}{\bar{v}(k)}, $$
where $\bar{v}(k)$ is the average voltage of all series-connected cells at time $k$. The fault signature is then obtained by adaptive Kalman filtering of $\Delta v_j(k)$ to remove measurement noise. A spatial clustering algorithm, such as the $k$-means or density-based spatial clustering of applications with noise, is then applied to the filtered values. Cells that form small isolated clusters are flagged as potentially faulty. Since this technique does not need a complete battery model, it is robust to cell inconsistency and measurement noise. I believe that this cloud-based healthcare system is the only rational solution for a future large-scale electric-car fleet.
In the following table, I present a side-by-side comparison of the performance characteristics of several commonly used diagnostic algorithms for fault detection in electric-car lithium-ion battery systems.
| Method | Computational cost | Accuracy | Robustness to noise | Physical interpretability | Adaptation to aging | Typical detection time |
|---|---|---|---|---|---|---|
| Kalman filter observer | Medium | High (if model accurate) | Medium | High | Low | Seconds |
| Recursive least squares | Low | High for parameter drift | Low | High | Medium | Seconds to minutes |
| Wavelet transform | Medium | Good for abrupt changes | Good | Medium | Low | Milliseconds |
| Shannon entropy | Very low | Good for gradual anomaly | Good | Medium | Medium | Minutes |
| Correlation coefficient | Very low | Good for cell isolation | Good | Medium | High | Minutes |
| Fuzzy logic | Low | Depends on rules | High | High | Low | Real-time |
| Random forest | Medium training, low inference | High (with data) | High | Low | High if retrained | Milliseconds to seconds |
| Deep neural network | High training, medium inference | Very high | Very high | Very low | High if retrained | Milliseconds |
| Cloud sparse-data observer | High cloud side, low vehicle side | High for short-circuit | Very high | Medium | High | Minutes to hours |
One important trend that I have observed in the literature is the combination of model-based state estimation with data-driven classifiers. For example, the residual of a Kalman filter is not simply compared to a fixed threshold; instead, the time-series residuals are fed into a classifier such as a support vector machine or a recurrent neural network. This hybrid approach enables simultaneous detection, isolation, and estimation of multi-fault conditions. In electric-car battery applications, common combinations include:
$$ \hat{\theta} = \arg \min_{\theta} \sum_{k} \left\| y_k – \hat{y}_k(\theta) \right\|^2 + \lambda \|\theta\|_1, $$
where the first term models the system output and the second term encourages sparse fault parameters, that is, only a few components of the battery system are abnormal at the same time. The sparse structure is beneficial for isolating a single faulty cell among hundreds of cells in a large electric-car battery pack.
Another important family of diagnostic algorithms is based on the incremental capacity (IC) curve analysis, which can reveal internal degradation mechanisms such as loss of active anode material or lithium inventory. The incremental capacity is defined as:
$$ IC = \frac{dQ}{dV} \approx \frac{\Delta Q}{\Delta V} , $$
where $Q$ is the charge or discharge capacity and $V$ is the terminal voltage. During charging, the IC curve of a healthy cell has characteristic peaks. A shift in the peak position or a decrease in peak amplitude signifies an aging mechanism. Accelerated aging may act as a precursor to safety hazards, so the IC-based features can be fed into a data-driven diagnostics model. The following table gives an example of mapping IC feature changes to possible failure modes:
| IC feature change | Possible internal mechanism | Safety implication |
|---|---|---|
| Peak amplitude decrease at high SOC region | Loss of lithium inventory | Accelerated capacity fade but low immediate risk |
| Peak shift to lower voltage | Increase of cell impedance | Higher heat generation under fast charging |
| Appearance of additional peaks | Structural change of cathode or phase transition | May induce oxygen release in extreme conditions |
| Irregular wiggles during current pulses | Micro-short circuit or electrode inhomogeneity | Early warning of internal short circuit |
In practice, no single diagnostic method can guarantee robust detection and isolation of every possible fault in a lithium-ion battery pack of an electric car. I have therefore learned that a hierarchical fusion framework is usually necessary. The diagnostics hierarchy can be represented as:
$$ S(t) = w_1\,\mathrm{KL}(t) + w_2\,\mathrm{AN}(t) + w_3\,\mathrm{IC}(t) + w_4\,\mathrm{Res}(t), $$
where $\mathrm{KL}$ is the Kullback–Leibler divergence of the voltage distribution, $\mathrm{AN}$ is the anomaly score from temperature sensors, $\mathrm{IC}$ is a feature derived from incremental capacity, and $\mathrm{Res}$ is the normalized model residual. The weights $w_1,\ldots,w_4$ are optimized using a set of validation cells that cover different abuse modes. I have found that this group-level fusion can simultaneously reduce false alarms and miss rates even under severe electromagnetic interference and sensor noise.
3. Challenges, Outlook, and Recommendations
3.1 Current Challenges in Battery Fault Diagnosis
Although the research achievements in the diagnosis of electric-car lithium-ion battery safety faults are extensive, there remain several limitations that prevent the practical implementation in a commercial BMS. The first challenge is the ambiguities caused by similar fault features. For example, an internal short circuit, a cooling system failure, and a sensor drift may all lead to a voltage or temperature deviation of very similar magnitudes. Therefore, isolating the true root cause is difficult. Current methods often assume that all other components are working nominally; this assumption is not valid in real systems. Second, a model-based algorithm requires an accurate cell model that accounts for nonlinear behavior, hysteresis, temperature dependence, and aging. Such a model is hard to build and even harder to embed in a low-cost microcontroller. Third, signal-processing methods are sensitive to measurement noise; although they have good dynamic response, their ability to detect slow-developing faults is not sufficient. Fourth, knowledge-based methods require explicit rules extracted from expert knowledge, but some battery failure mechanisms are still not fully understood. Finally, many data-driven algorithms, especially deep neural networks, consume a large amount of training data and computational resources, and their “black-box” nature makes it difficult for engineers to interpret the decision logic.
From the point of view of the entire system, the diagnostic algorithm must operate under unpredictable driving conditions in an electric car. The battery load profile is determined by the driver, the road type, the environmental temperature, and traffic conditions. A diagnostic threshold that works well in one city might produce malfunction warnings in another. Therefore, I propose a robust design of health indicators that are independent of driving conditions. For example, features based on accumulated heat and charge transfer rather than instantaneous electric variables are more reliable.
Model-based approaches and data-driven methods both face a common challenge: the need for reliable ground-truth data from fault experiments. Constructing a thermal runaway experiment in a laboratory is expensive and potentially hazardous. Consequently, many fault samples are generated by artificial simulations and are not fully representative of real electric-car failure processes. To cope with this shortage, transfer learning and generative adversarial networks are emerging as tools to expand the fault data set while maintaining the subtle characteristics of battery behavior.
3.2 Future Directions and Recommendations
I foresee several clear trends that will shape the next generation of safety fault diagnostic algorithms for lithium-ion batteries in electric cars. The first and most important trend is the multi-source sparse-data based precise state monitoring. Instead of relying on a single voltage channel, the future BMS will fuse data from current, temperature, and additional sensors such as cell surface strain gauges, optical fiber sensors, gas sensors, and impedance spectrometers. The strain and gas-sensing signals may provide early warnings of internal gas generation that occur before voltage collapse. The second trend is big-data fault feature extraction and global fault evaluation. With cloud-connected electric cars, fleet-level data can be used to discover rare but dangerous fault patterns. For example, an unusual correlation between the charging station type and battery anomalies can reveal regional charging infrastructure problems. The third trend is the use of model-based fusion for closed-loop fault estimation. If a fault is detected and isolated, the BMS should handle the fault through fault-tolerant control: e.g., reducing power limits, rebalancing the pack, turning on a more aggressive cooling regime, or activating a pyrotechnic disconnector. The fourth trend is the implementation of intelligent algorithms in embedded hardware with minimal power consumption. Lightweight neural networks, tiny machine-learning frameworks, and model compression techniques have already been deployed in some electric-car battery management units.
One of the most promising directions is the combination of a physics-informed neural network (PINN) with battery models. The governing differential equations of battery dynamics are used as regularization terms during neural network training, so even with a sparse data set, the PINN can predict internal temperature and lithium distribution within acceptable errors. The governing equation for lithium diffusion inside solid particles is Fick’s second law:
$$ \frac{\partial c_s}{\partial t} = \frac{1}{r^2} \frac{\partial}{\partial r} \left( D_s r^2 \frac{\partial c_s}{\partial r} \right), $$
where $c_s$ is the lithium concentration in the electrode particles, $r$ is the radial coordinate, and $D_s$ is the solid-phase diffusion coefficient. The degradation of parameters can be inferred through an inverse PINN framework. The trained network can reconstruct the concentration profile without explicitly solving the partial differential equation, which greatly improves the speed of model-based diagnosis.
I also recommend that future research focuses on “digital twin” technology for battery packs in electric cars. In a digital twin, a virtual battery model is continuously updated with real-time sensor measurements. The digital twin runs ahead of the physical battery to predict the near-future temperature and state-of-health. The difference between the twin prediction and physical sensor output is a powerful fault indicator. It can be expressed as:
$$ e(t) = \left| x_{\text{physical}}(t + \Delta t) – x_{\text{digital}}(t + \Delta t) \right|, $$
where $x$ represents state variables such as cell temperature or terminal voltage. A sudden divergence in $e(t)$ is a much earlier fault precursor than a fixed-threshold voltage alarm.
For thermal management of electric-car batteries, I believe that advanced control architectures will incorporate diagnostic information into the thermal management loop. The block diagram of such a system is schematically represented in terms of target temperature and corrective actions. Because temperature has a dominant effect on the electrochemical degradation rate, maintaining the cell temperature within a uniform band is essential. The actual temperature of the cell is a 3D field; a simplified lumped thermal model is often used for online state observation:
$$ m C_p \frac{dT_c}{dt} = Q_{\text{gen}} – h_{\text{eff}} A_{\text{cool}} \left( T_c – T_{\text{coolant}} \right), $$
where $m$ is the cell mass, $T_c$ is the representative core temperature, $h_{\text{eff}}$ is an effective heat transfer coefficient, and $Q_{\text{gen}}$ includes both reversible and irreversible heat. In my outlook, the next-generation BMS will estimate $Q_{\text{gen}}$ online using an electrochemical-thermal model and then decide whether to increase the coolant flow rate or limit the charge current. A diagnostic algorithm should be able to identify a gradually blocked cooling channel by evaluating the increase in the effective thermal resistance $R_{\mathrm{th}}$:
$$ R_{\mathrm{th}} = \frac{\bar{T}_{\text{cell}} – T_{\text{coolant}}}{\dot{q}_{\text{total}}}. $$
If $R_{\mathrm{th}}$ rises above a baseline calibrated for the healthy pack, the BMS can both alert the driver and request a service inspection.
Table below lists the potential future technologies and their expected contributions to the safety of electric-car battery systems.
| Technology | Expected contribution | Implementation challenge |
|---|---|---|
| Cloud-based battery management | Fleet-level diagnostics, remote health and safety monitoring, big-data mining for rare fault modes | Data security, communication latency, standard formats |
| Machine learning with interpretable models | High detection accuracy plus physical understanding | Need for explainable AI techniques |
| Sparse data observers | Robust detection of short-circuit and open-circuit without high-fidelity model | Selection of features and cluster thresholds |
| Hardware-embedded tiny neural networks | Real-time edge inference in the vehicle | Memory restrictions, quantization loss |
| Hybrid algorithms (model + data) | Best utilization of physical constraints and statistical learning | Complexity of software integration |
| Electrolyte modification and solid-state cells | Fundamental reduction of flammability and thermal-runaway propagation | Manufacturing cost, scalability, temperature performance |
Another essential point that I have noticed from accident investigations is the disconnect between pre-failure fault diagnostics and post-failure firefighting. The electric-car firefighting procedures often require manual validation of battery condition after a severe impact. If the BMS is capable of storing a safety event data recorder (similar to a flight recorder), containing voltage–current–temperature profiles at a high sampling rate during the event, the cause of the accident can be accurately reconstructed. This information will greatly enhance the safety design of next-generation electric cars.
In the area of machine learning, I recommend using recurrent neural networks or long short-term memory networks for time-series fault prediction. A long short-term memory model can learn from a sequence of battery cells measurements and predict the risk score of thermal runaway. The hidden state update equation is:
$$ f_t = \sigma\left( W_f \cdot [h_{t-1}, x_t] + b_f \right), $$
$$ i_t = \sigma\left( W_i \cdot [h_{t-1}, x_t] + b_i \right), $$
$$ \tilde{C}_t = \tanh\left( W_C \cdot [h_{t-1}, x_t] + b_C \right), $$
$$ C_t = f_t * C_{t-1} + i_t * \tilde{C}_t, $$
$$ o_t = \sigma\left( W_o \cdot [h_{t-1}, x_t] + b_o \right), $$
$$ h_t = o_t * \tanh(C_t). $$
These equations allow the network to retain memory of slow degradation while being sensitive to abrupt events. For training an effective LSTM classifier, one must provide enough examples of clear thermal runaway events, which are rarely available in the public domain. Transfer learning across different cell chemistries and form factors is therefore a key future focus.
I also emphasize the importance of uncertainty quantification in diagnostic algorithms. The health state of a battery must not be reported as a point estimate alone; rather, the confidence interval should be included. For example, if the internal-resistance estimate has a large uncertainty because the current sensor is noisy, then a fault alarm raised based on that estimate may be unreliable. Bayesian neural networks and ensemble methods can provide such uncertainty bounds. In model-based state estimation, the covariance matrix $P_k$ is inherently an uncertainty representation; a fault can be declared if a predicted optimum exceeds a precomputed confidence boundary.
3.3 Recommendations for BMS Implementation
Based on my review of the current status and future trends, I propose the following concrete recommendations for battery safety diagnostic systems in electric cars:
First, always design the diagnostic algorithm with the constraint of the BMS hardware. The algorithm should be coded in a compact linear algebra library and benchmarked on the actual ECU. For every diagnostic feature, evaluate the false-alarm rate under standard driving cycles and under a no-fault field dataset. Second, adopt a hierarchical diagnostic architecture. The first layer consists of signal-based quick screening checks: overvoltage, undervoltage, over-temperature, over-current, insulation leakage, and voltage inconsistency. The second layer uses state observers or model-based residuals to detect soft faults that do not trigger immediate thresholds. The third layer applies data-driven classifiers only after a potential fault is localized to a subset of cells, reducing computational costs. Third, harness the cloud. Since modern electric cars are always connected when charging, every charging event can be regarded as a complete battery scanning test. By comparing the charging voltage curves with historical data of the same car, a slow-evolving internal short circuit can be detected days before the cell enters thermal runaway. Fourth, design diagnostic experiments with realistic cells. I advise the industry to publish more valuable cells test data on controlled mechanical abuse, electrical abuse, and overcharging tests in order to standardize the comparison of diagnostic performance.
The following table provides a concise guideline for selecting diagnostic methods according to the fault type in electric-car lithium-ion battery systems.
| Fault type | Recommended primary method | Secondary method | Remark |
|---|---|---|---|
| External short circuit | Current and voltage double-check | Temperature rate-of-rise detector | Reaction in less than 1 s |
| Internal short circuit | Voltage correlation-based isolation | Cloud-based sparse data observer | May develop over weeks |
| Overcharge | Model-based voltage threshold with SOC correction | Coulomb counting check | Charge current limiting response |
| Over-discharge | Minimum cell voltage under load | Copper-free anode simulation | Requires high-quality current sensor |
| Cooling fault | Lumped thermal model residual | Thermistor redundancy | Check via effective thermal resistance |
| Sensor drift | Kalman filter observer innovation | Signal correlation between sensors | Sensor fusion with virtual reference |
| Thermal runaway precursor | Gas/strain detection | Local voltage and temperature abrupt change | Trigger battery disconnect |
In all cases, it is crucial to implement the diagnostic system with a safe default action. When a severe fault is immediately detected, the BMS should open the high-voltage contactors, discharge any pyrotechnic fuse connected to the pack, and warn the user through the car display and smartphone app. When a less severe fault is detected, the system may derate the maximum power, balance the cells, adjust the cooling set-point, or request a service event. Fault-tolerant control can be formulated as an optimization problem:
$$ \min_{u_c, u_t, P_{\max}} J = \sum_{t=0}^{N} \left\{ \alpha\left(T_{t} – T_{\mathrm{target}}\right)^2 + \beta \cdot \mathrm{Risk}\left( \hat{x}_t \right) + \gamma \cdot \left| u_t – u_t^{\mathrm{nom}} \right|^2 \right\}, $$
where $u_c$ is the coolant flow command, $u_t$ is the heater or disconnector command, $P_{\max}$ is the maximum allowable power, and $\mathrm{Risk}(\hat{x}_t)$ is an estimated risk score from the diagnostic module. This type of optimal fault-tolerant control prevents unnecessary panic disconnections while keeping the electric car safe.
4. Conclusions
In this paper, I have reviewed the current state and future possibilities of safety fault diagnostic algorithms for lithium-ion batteries used in electric cars. I first described the main safety issues and their electrochemical origin, highlighting the role of increasing cell temperature on the exothermic chain reactions. I explained that mechanical, electrical, and thermal abuse can all induce internal short circuits and eventually thermal runaway. Then, I discussed in detail the classification of battery faults and summarized the mainstream diagnostic algorithms, including model-based state estimation, parameter estimation, signal-processing, knowledge-based, and data-driven approaches. I provided equations for residual generation, entropy computation, correlation-based cell isolation, and cloud data processing, as these formulas form the basis of many advanced algorithms. I also introduced a hierarchical fusion framework to reduce false alarms and improve fault isolation.
I clearly concluded that no single diagnostic algorithm is omnipotent. The future of electric-car lithium-ion battery safety will rely on a synergy of multi-sensor measurement, cloud computing, big data, digital twin, real-time model updating, and embedded machine learning. In the future, I believe that artificial intelligence will dramatically improve the performance of fault diagnosis, while physics-based models will provide the necessary interpretability and guarantee that decisions never violate the fundamental laws of electrochemistry and thermodynamics. Through these intelligent algorithms, the electric-car industry will be able to develop a highly safe advanced battery management system, which will not only improve battery performance and lifespan but also ensure the reliable and secure operation of electric cars in real-world scenarios and contribute to the sustainable development goals of clean energy and carbon neutrality.
I hope this review has provided a comprehensive and practical perspective for both researchers and engineers who are developing novel diagnostic methods for the safety of lithium-ion batteries in electric cars. Every electric car on the road deserves a robust safety monitoring solution, and every driver deserves confidence in the battery system that drives their car. This is the ultimate motivation for continuing my research on diagnostic algorithms for the safety of lithium-ion batteries in electric cars.