Electric vehicles (EVs) have emerged as a cornerstone of the global shift towards sustainable transportation and energy systems. The heart of an EV’s powertrain is its battery pack, predominantly composed of lithium-ion (Li-ion) cells due to their high energy density and long cycle life. Ensuring the safety and reliability of these EV battery packs is paramount, as failures can lead to severe economic losses and, more critically, pose significant risks to human life. Among various failure modes, the internal short circuit (ISC) is particularly insidious. It can initiate from mechanical, thermal, or electrical abuse, leading to the growth of lithium dendrites that pierce the separator, creating an internal conductive path. This fault evolves through early, intermediate, and late stages. In the critical early stage, the short-circuit resistance is relatively high, resulting in subtle changes in cell voltage, temperature, and state of charge (SOC). These faint anomalies are notoriously difficult to detect with conventional monitoring techniques, yet timely identification at this juncture is essential to prevent progression to thermal runaway—a catastrophic event characterized by rapid temperature rise, gas venting, and potential fire or explosion.
Traditional fault diagnosis methods for EV battery packs have heavily relied on voltage signal analysis. Techniques such as correlation-based methods, principal component analysis (PCA), and kernel PCA have been applied to series-connected packs with considerable success. However, these voltage-based approaches face a fundamental limitation in modern EV battery pack architectures, which often employ complex series-parallel configurations to meet voltage and capacity requirements. In a parallel module, the inherent cell balancing effect masks voltage discrepancies, making it nearly impossible to detect and, more importantly, to pinpoint the exact faulty cell based solely on voltage inconsistency. This creates a significant gap in safety management for contemporary EV battery pack designs. Furthermore, the early-stage ISC signals are weak and easily buried within normal operational noise, challenging the sensitivity of many data-driven algorithms. While advanced techniques like electrochemical impedance spectroscopy (EIS) or ultrasonic testing offer high precision, their integration into cost-sensitive and space-constrained EV battery packs remains challenging for widespread deployment.
To address these challenges, this paper proposes a novel early fault detection and localization method for EV battery packs based on an improved local gravitation analysis of the temperature field. Unlike voltage, temperature is a more direct indicator of internal energy dissipation and is not subject to the masking effects present in parallel circuits. The core of our method lies in quantifying the inconsistency of the thermal field within the pack. We first calculate a Local Resultant Force (LRF) for each cell, a metric derived from the concept of gravitational attraction between data points, applied here to normalized temperature readings. The LRF for cell \(i\) at time \(t\) is defined as:
$$F_i(t) = \sum_{j=1}^{n} \left( T’_j(t) – T’_i(t) \right)$$
where \(T’_i(t)\) and \(T’_j(t)\) are the standardized (z-score) temperature values of cell \(i\) and its neighboring cell \(j\), respectively, and \(n\) is the total number of other cells in the pack. Standardization is performed over a moving time window to ensure comparability:
$$T’_i(t) = \frac{T_i(t) – \mu_i(t)}{\sigma_i(t)}$$
Here, \(T_i(t)\) is the raw temperature, and \(\mu_i(t)\) and \(\sigma_i(t)\) are the mean and standard deviation of cell \(i\)’s temperature within the predefined time window ending at \(t\). In a healthy EV battery pack, thermal forces from surrounding cells on a given cell tend to cancel out, resulting in a small LRF magnitude. Conversely, a cell undergoing an ISC begins to exhibit thermal behavior divergent from its peers. The forces acting upon it become more unidirectional, leading to a significant increase in the magnitude of its LRF.
To transform this physical insight into a system-level detection statistic, we define the Temperature Abnormality Index (TAI), \(h(t)\), as the maximum absolute LRF value across all cells at each time instant:
$$h(t) = \max\{ |F_1(t)|, |F_2(t)|, …, |F_N(t)| \}$$
where \(N\) is the total number of cells. The TAI serves as a real-time gauge of the overall thermal anomaly within the EV battery pack. A key innovation of our work is the adaptive, data-driven determination of the detection threshold for the TAI. We employ non-parametric Kernel Density Estimation (KDE) on TAI values calculated from a fault-free training dataset to model its underlying probability distribution. The threshold \(h_r\) at a confidence level \(\Omega\) (e.g., 0.99) is found by solving:
$$\int_{0}^{h_r} f(s) ds = \Omega$$
The probability density function \(f(s)\) is estimated as:
$$f(s) = \frac{1}{C a} \sum_{i=1}^{C} K\left(\frac{s – h(i)}{a}\right)$$
where \(\{h(i) | i=1,2,…,C\}\) are the TAI samples from the training set, \(a\) is the bandwidth parameter (calculated via Silverman’s rule), and \(K(\cdot)\) is the Gaussian kernel function. Fault detection is declared at the first time \(t_f\) when \(h(t) \geq h_r\).
Following fault detection, precise localization of the faulty cell is achieved through a contribution analysis. We argue that the faulty cell’s LRF remains elevated and contributes dominantly to the system’s abnormal state after the fault onset. Therefore, we define a Local Gravitation Contribution Function \(C(z)\) for each cell \(z\) over a time window \(p\) following the detected fault time \(t_f\):
$$C(z) = \sum_{t=t_f}^{t_f + p} |F_z(t)|$$
The cell with the maximum contribution value is identified as the faulty cell \(z_f\):
$$z_f = \arg\max_z C(z)$$
This localization strategy effectively pinpoints the source of thermal anomaly, even within parallel modules, overcoming a major shortcoming of voltage-based methods.
To rigorously validate the proposed method, we constructed a simulation model of a 6s4p EV battery pack, comprising 24 cylindrical 21700 cells. The model, whose parameters are summarized in Table 1, captures the coupled electrochemical-thermal dynamics and has been validated against experimental data under various current rates. We designed nine distinct ISC fault scenarios with varying parameters, as detailed in Table 2. These scenarios test the method’s robustness against different fault locations (including cells within parallel modules), discharge rates (constant 2C and dynamic UDDS profile), internal short resistances, and fault initiation times.
| Parameter | Value | Source/Note |
|---|---|---|
| Cell Diameter | 0.021 m | Cell Specification |
| Cell Height | 0.070 m | Cell Specification |
| Number of Cells | 24 | 6s4p Configuration |
| Inter-cell Gap | 0.002 m | Forced Air Cooling |
| Number of Temperature Sensors | 24 | One per Cell |
| Ambient Temperature | 20 °C | Simulation Setting |
| Cell Nominal Capacity | 4.8 Ah | Cell Specification |
| Cell Nominal Voltage | 3.7 V | Cell Specification |
| Fault ID | Faulty Cell Location | Discharge Profile | ISC Resistance (Ω) | Fault Trigger Time (s) |
|---|---|---|---|---|
| 1 | #4 | 2C | 10 | 1000 |
| 2 | #5 | 2C | 10 | 1000 |
| 3 | #11 | 2C | 10 | 1000 |
| 4 | #16 | 2C | 10 | 1000 |
| 5 | #18 | 2C | 10 | 1000 |
| 6 | #23 | 2C | 10 | 1000 |
| 7 | #23 | 2C | 5 | 1000 |
| 8 | #23 | 2C | 10 | 1500 |
| 9 | #23 | UDDS | 10 | 1000 |
The performance of the proposed method was evaluated using three key metrics: Fault Recall Rate (FRR), False Alarm Rate (FAR), and Fault Detection Delay (FDD). The comprehensive results for all nine fault scenarios are presented in Table 3. The method demonstrated consistently high FRR (exceeding 94% for most constant-current cases), very low FAR (≤0.5%), and short FDD (≤90 seconds). Most significantly, the localization accuracy was 100% across all scenarios, correctly identifying the faulty cell even when it was part of a parallel module. For instance, in Fault 1, the TAI rapidly exceeded its threshold shortly after the ISC was triggered at 1000s, leading to detection within 8 seconds and accurate identification of cell #4 as the fault source.
| Fault ID | Fault Recall Rate (FRR) % | False Alarm Rate (FAR) % | Fault Detection Delay (FDD) s | Localization Result (Cell ID) |
|---|---|---|---|---|
| 1 | 99.2 | 0.5 | 8 | 4 |
| 2 | 98.1 | 0.5 | 17 | 5 |
| 3 | 98.4 | 0.5 | 16 | 11 |
| 4 | 96.8 | 0.5 | 32 | 16 |
| 5 | 96.1 | 0.5 | 35 | 18 |
| 6 | 94.4 | 0.5 | 56 | 23 |
| 7 | 97.8 | 0.5 | 22 | 23 |
| 8 | 99.2 | 0.4 | 4 | 23 |
| 9 | 55.9 | 1.5 | 26 | 23 |
A comparative analysis against established methods highlights the advantages of our approach. As shown in Table 4, the proposed method significantly outperforms the original voltage-based local gravitation method, correlation coefficient method, and PCA in terms of FRR and FDD for a representative fault scenario (Fault 1). Crucially, while the other methods either cannot provide localization or are ineffective for parallel configurations, our temperature-based method achieves precise fault cell identification.
| Method | FRR (%) | FAR (%) | FDD (s) | Localization Capability |
|---|---|---|---|---|
| Proposed Method | 99.2 | 0.5 | 8 | Accurate (Cell #4) |
| Original Local Gravitation | 16.5 | 0.5 | 639 | Not for Parallel Packs |
| Correlation Coefficient | 80.3 | 1.0 | 197 | Not for Parallel Packs |
| Principal Component Analysis | 7.5 | 0.5 | 619 | Not for Parallel Packs |
In conclusion, this paper presents an effective early fault detection and localization framework for EV battery packs based on improved local gravitation analysis of the thermal field. The method introduces the Temperature Abnormality Index (TAI) for system-level anomaly quantification and employs adaptive KDE-based thresholding for reliable detection. The designed contribution function enables precise identification of the faulty cell, a capability that is notably absent in voltage-based methods when applied to series-parallel EV battery pack topologies. Experimental validation on a 6s4p pack model under various ISC conditions confirms the method’s high accuracy, low false alarm rate, rapid response, and perfect localization performance. This work provides a viable and promising solution for enhancing the safety management of complex EV battery packs by leveraging their inherent temperature field characteristics.
Future research will focus on several extensions to improve practicality and robustness. Firstly, while the fixed threshold derived from normal operation data showed good performance, exploring adaptive thresholding mechanisms that account for dynamic operating conditions and aging effects could further enhance robustness. Secondly, the current method is particularly sensitive to faults exhibiting thermal signatures like ISC. Its efficacy for faults with minimal thermal manifestation (e.g., connection loosening, sensor drift) may be limited. Integrating multi-modal information from voltage and current sensors could create a more comprehensive diagnostic framework. Finally, for EV battery packs with highly non-uniform cooling structures, incorporating spatial weighting factors into the LRF calculation or the TAI could improve generalization by accounting for inherent, non-fault-related temperature gradients.