Driven by the pressing challenges of environmental degradation and the accelerating depletion of conventional energy reserves, electric vehicles (EVs) have emerged as a pivotal substitute for traditional internal combustion engine vehicles. Central to the EV’s propulsion architecture is the battery system, which serves as the exclusive energy reservoir. Among the various battery chemistries, lithium-ion batteries have ascended to prominence in the EV battery pack domain, owing to their exceptional energy density, minimal self-discharge propensity, and extended operational longevity. The Nobel laureate Akira Yoshino has astutely observed that lithium-ion batteries are destined to play an integral role in the forthcoming energy revolution. Nevertheless, the rapid proliferation of EVs is shadowed by persistent safety concerns, particularly the catastrophic incidents arising from battery thermal runaway. The evolution from a minor electrical abuse to a full-blown thermal event invariably transits through a short circuit phase. If this phase can be detected at its nascent stage, commonly referred to as a soft short circuit (SSC), and isolated before its progression, the occurrence of heat-induced calamities can be effectively thwarted. Consequently, the development of an efficacious SSC fault diagnostic strategy for the EV battery pack is critical for ensuring vehicular safety and reliability.
My research is dedicated to the diagnosis of SSC within the EV battery pack, specifically targeting the lithium iron phosphate (LiFePO₄) chemistry. The inherent challenge lies in the fact that SSC faults often manifest with subtle signatures, which are readily obscured by the intrinsic parameter inconsistencies present among cells. To surmount this barrier, I propose a holistic diagnostic framework that systematically addresses the measurement, detection, and localization facets of fault management. This methodology is anchored in an interpretable full-dimensional statistical analysis, which aims to empower the battery management system (BMS) with the ability to discern millivolt-level anomalies expeditiously. In this paper, I expound upon the theoretical underpinnings, architectural design, and empirical validation of this diagnostic paradigm. My experimental apparatus comprises a compact array of fourteen series-connected LiFePO₄ cells, subjected to the rigorous World Light-duty Vehicle Test Cycle (WLTC). The outcomes affirm that my approach can robustly detect and isolate both transient and cumulative SSC faults across a broad state-of-charge (SOC) window, under operational stresses peaking at a 2C discharge rate.
1. Introduction and Background
The automotive industry is undergoing a transformative shift towards electrification, propelled by stringent emission regulations and a collective global consciousness about climate change. As the demand for EVs surges, the imperative for an unyielding commitment to battery safety becomes paramount. The EV battery pack, being a high-energy component, inherently harbors risks. In recent years, numerous fire incidents have been attributed to battery thermal runaway, a phenomenon often initiated by internal or external short circuits. While severe short circuits are typically conspicuous and may be readily detected, it is the insidious soft short circuit that presents a more complex dilemma.
An SSC can be conceptualized as a high-resistance internal short. This resistance acts as a parasitic load, slowly siphoning energy from the affected cell. This process generates localized heat and accelerates self-discharge, inducing a subtle deviation in voltage relative to neighboring healthy cells. The diagnostic problem is further compounded when considering the intrinsic variability of cells within a pack. Variations in manufacturing tolerances, temperature gradients, and aging trajectories lead to discrepancies in capacity, internal resistance (IR), and polarization characteristics. During highly dynamic load conditions, these discrepancies can produce voltage disparities that mimic the early signatures of an SSC, leading to a high rate of false alarms or, conversely, masking the genuine fault signal.
The limitations of existing diagnostic techniques are twofold. First, model-based methods, such as those utilizing Extended Kalman Filters or electrochemical models, often necessitate substantial computational overhead and prior knowledge of accurate cell models. Their performance can be contingent upon the fidelity of the model to the actual cell under non-ideal operational states. Second, while data-driven approaches do not require rigorous physical models, they frequently suffer from a lack of interpretability. In many instances, the fault indicators are computed via black-box transformations, obscuring the relationship between the statistical metric and the physical fault magnitude. This opacity makes the task of threshold determination a matter of empirical guessing rather than principled calculation. Furthermore, in conventional structures, fault localization is usually treated as a subsequent, dependent step following detection, which prevents the localization results from informing and refining the detection process in real time.
In response to these challenges, I have crafted a diagnostic strategy that operates on three synergistic fronts: an advanced measurement topology, a statistically rigorous detection indicator with a self-adjusting threshold, and a concurrently running localization mechanism. The novelty of my approach lies not only in its components but also in their synergistic integration.
2. Methodology for Voltage Measurement
2.1. The Challenge of Parameter Inconsistency
A fundamental obstacle in the quest for reliable SSC detection within an EV battery pack is the inherent parameter inconsistency among the series-connected cells. When a current excitation \( I_{p,k} \) flows through the pack, the terminal voltage of a cell is governed by its open-circuit voltage (OCV) characteristic and its effective internal impedance. This relationship can be approximated by:
$$V_{i,k}^{c} \approx V_{oc,i,k}(SOC_{i,k}) – I_{p,k} \cdot R_{s,i,k}(SOC_{i,k})$$
where \( V_{oc,i,k} \) is the OCV, representing an electrochemical potential which is a function of the state-of-charge (SOC), and \( R_{s,i,k} \) is the total dynamic internal resistance. Given that capacity and impedance cannot be manufacture-identical for every cell, the SOCs and effective resistances will differ. This disparity results in a baseline differential voltage even under healthy conditions, which is exacerbated when the current magnitude is high. This baseline is destructive to the sensitivity of fault detection. For lithium iron phosphate cells, the OCV curve is predominantly flat, which is an advantageous feature for SOC estimation but also means that even moderate IR differences can lead to significant deviations in voltage response when current is applied.
2.2. Virtual Voltage Measurement for Signal Alignment
To circumvent the confounding effects of cell inconsistencies, I have introduced a “virtual voltage measurement” (VVM) scheme. Rather than attempting to physically alter the pack’s architecture, I manipulate the way voltage samples are processed.
Conventional differential calculations, which are instrumental in highlighting deviations between contiguous cells, suffer from the pollution of the IR-related noise described above. I have implemented a two-pronged strategy to refine this differential signal:
1. Spatial Reconfiguration (Virtual Mapping): During the pack’s assembly, cells are often adjacent to one another. However, in terms of their dynamic impedance, cells that are physically adjacent may not be the most similar. Therefore, I develop a mapping matrix \( \mathbf{A}_v \) that rewires the logical topology of the voltage measurements in the digital domain to pair cells with analogous resistance characteristics. The transformed voltage vector is expressed as \( \mathbf{V}^{sv}_{k} = \mathbf{A}_v \mathbf{V}^s_k \). This ensures that the differential operations between logical channels minimize noise contributions from impedance mismatches.
2. Temporal Alignment (Near-Cut-off Voltage Aligned, NCVA): Exploiting the inherent flat OCV plateau of LiFePO₄ chemistry, I perform a one-off voltage alignment operation. When the EV battery pack‘s terminal voltage reaches a predetermined near-cut-off voltage (approximately 2.94V/cell; greater than the lower knee point), the active balancing system is engaged to ensure all cells possess the same starting potential. Subsequently, balancing is disabled. This practice ensures that cells begin their charge/discharge cycle from a common reference point, preventing the natural drift of OCV due to capacity mismatches from introducing extraneous voltage components.
The efficacy of this method is summarized in Table I, which compares the standard deviation (SD) of the differential channel voltages across three scenarios: conventional charging without alignment, alignment only, and alignment with virtual mapping. My proposed VVM scheme effectively narrows the spread of data, yielding a robust baseline for fault analysis.
| Cell Number | Scenario 1: Original | Scenario 2: NCVA only | Scenario 3: NCVA + Virtual |
|---|---|---|---|
| 1 | 3.380 | 0.742 | 0.923 |
| 2 | 4.073 | 1.030 | 0.742 |
| 3 | 4.284 | 1.152 | 0.837 |
| 4 | 3.269 | 0.990 | 0.816 |
| 5 | 2.318 | 0.997 | 0.811 |
| 6 | 2.181 | 1.435 | 0.999 |
| 7 | 0.932 | 1.078 | 0.828 |
| 8 | 1.088 | 0.732 | 1.014 |
| 9 | 2.911 | 1.208 | 0.995 |
| 10 | 2.314 | 0.927 | 0.893 |
| 11 | 4.761 | 1.029 | 1.122 |
| 12 | 3.367 | 0.634 | 0.814 |
| 13 | 2.975 | 0.749 | 0.997 |
| Mean Dev. (MD) | 0.880 | 0.170 | 0.094 |
The application of the virtual measurement scheme, as demonstrated in Table I, leads to a dramatic reduction in differential voltage volatility. The mean deviation (MD) is compressed from 0.880 mV in the conventional scenario to a mere 0.094 mV with my complete VVM protocol, which is primordial for maximizing fault detectability.
3. Robust Fault Detection Indicator and Self-Tuning Control Limit
3.1. Formulating the Fault Detection Index
In a meticulously monitored EV battery pack, the health of the cells can be discerned by analyzing the correlation matrix of voltage measurements. Let \( \mathbf{x} \) denote the differential voltage vector in the virtual space. For a sliding window of width \( w \), I construct a matrix \( \mathbf{X}_k \in \mathbb{R}^{w \times (n-1)} \). To ensure data comparability across different SOCs and temperatures, this data is standardized using historical reference data, \( \boldsymbol{\mu}_0 \) and covariance scaling \( \boldsymbol{\Sigma}_0 \). The standardized matrix \( \overline{\mathbf{X}}_k \) is then used to calculate the covariance matrix \( \mathbf{C}_k \) as follows:
$$ \mathbf{C}_k = \frac{1}{w}\overline{\mathbf{X}}_k^T \overline{\mathbf{X}}_k \approx \mathbf{V}_k \boldsymbol{\Lambda}_k \mathbf{V}_k^{T}$$
where \( \boldsymbol{\Lambda}_k = \mathrm{diag}(\lambda_{1,k}, \lambda_{2,k}, \ldots, \lambda_{n-1,k}) \) is the diagonal matrix of eigenvalues and \( \mathbf{V}_k \) is the matrix of orthogonal eigenvectors. This is a full-fledged eigen-decomposition. In this new eigenspace, the primary axis captures the direction of maximum variance in the data.
When an SSC event commences, the associated abnormal voltage fluctuations introduce additional variance into the data. This perturbation predominantly aligns along a specific eigenvector direction, causing a correlating spike in the corresponding eigenvalue. To quantify this, I devise the Detection Indicator (DI) based on the normalized deviation of the eigenvalues from their healthy statistical distribution:
$$ DI_k = \max_{1 \leq j \leq n-1} \left\{ \frac{\lambda_{j,k} – \mu_{\lambda_j^*}}{\sigma_{\lambda_j^*}} \right\} $$
In this equation, \( \mu_{\lambda_j^*} \) and \( \sigma_{\lambda_j^*} \) represent the mean and standard deviation of the \( j^{th} \) eigenvalue, estimated from a historical dataset of healthy operation.
3.2. Interpretable Threshold Derivation via Detectability Analysis
A pivotal aspect of my contribution is adding interpretability to the data-driven detection. For a fault occurring in a channel represented by vector \( \mathbf{A} \) (the linear transformation coefficients), I model the faulty measurement as \( \mathbf{x} = \mathbf{A}\mathbf{s} + \boldsymbol{\xi} f_{sc} \). where \( \boldsymbol{\xi} \) locates the fault and \( f_{sc} \) is the scalar fault magnitude. By propagating this fault through the covariance matrix estimation, its expected trace can be decomposed into three components:
$$ \mathbb{E}\{ \mathrm{Tr}(\mathbf{C}_k) \} = \mathbf{C}_{1,k} + \mathbf{C}_{2,k} + \mathbf{C}_{3,k} $$
Given that \( \mathbb{E}\{ \mathbf{C}_{1,k} \} \) corresponds to the healthy eigenvalue sum, \( \mathbb{E}\{\mathbf{C}_{2,k}\} \) averages to zero, and \( \mathbf{C}_{3,k} \) relates to the fault signature, I derive the crucial relationship connecting the sensor fault amplitude to the detection threshold. This leads to the quantitative condition that ensures fault detectability:
$$ \delta_{sc} \cdot \sum_{i=1}^{n-1} \sigma_{\lambda_i^*} \geq \frac{\bar{f}_{sc}^2}{w} \cdot \bar{t}_f \cdot \max_{l \in [1,n-1]} \left( 2\sigma_{x_l}^* \right) $$
Here \( \delta_{sc} \) is the control limit, \( \sigma_{\lambda_i^*} \) is the eigenvalue standard deviation, \( \bar{f}_{sc} \) is the magnitude of the fault signature (in volts), \( w \) is the window width, \( \bar{t}_f \) is the fault duration within the window, and \( \sigma_{x_l}^* \) is the standard deviation of the \( l^{th} \) differential voltage channel during normal operation. This closed-form solution explicitly provides a mapping between the threshold limit and the physical parameters of the fault and the system. Thus, given a tolerable SSC fault amplitude that the BMS should detect, I can directly compute the necessary threshold, enabling systematic control limit self-commissioning.
3.3. Fault Severity and Limit Determination
I define the fault voltage amplitude term \( f_{sc} \) that encompasses both instantaneous and cumulative failures. It reflects contributions from both the equivalent short-circuit resistance and the OCV drift over time:
$$ f_{sc}(t_2) = f_{ocv}^{sv}(t_2) + \frac{V_{oc}^{sv}}{R_{sc}(t_2)} \cdot \Delta R $$
In the case of gradual deterioration of the cell’s internal separator, Rsc may decrease slowly. The total detection threshold \( \delta_d \) is determined as the maximum of the boundaries \( \delta_{sc,1} \) and \( \delta_{sc,2} \) derived from the two pivotal differential voltage subspaces, guaranteeing coverage over all possible fault channels. \( \delta_d = \max\{\delta_{sc,1}, \delta_{sc,2}\} \). Given the resilience of my virtual voltage measurement against inconsistencies, the resulting detection threshold exhibits a high sensitivity, allowing the BMS to respond to truly minuscule fault signatures.
4. Concurrent Fault Localization and Detection Refinement
4.1. Localization through Eigenvector Contribution
Traditional fault diagnosis in an EV battery pack typically runs detection and localization sequentially. If the detection threshold is breached, a separate algorithm is then implemented to identify the faulty cell. In contrast, my methodology capitalizes on the inherently orthogonal computations of eigen-decomposition to execute these tasks in parallel. When a fault exists, the covariance matrix eigenvalues deviate from healthy norms. The spatial location of the anomaly is encoded in the weighting of the corresponding eigenvector. By analyzing the \( m^{th} \) eigenvector \( \mathbf{v}_{m,k} \) associated with the abnormal eigenvalue \( \lambda_{m,k} \), I define a normalized localization index \( \eta_p \) for each cell position \( p \):
$$ \eta_{p} = \frac{ (v_{m,k}^{(p)})^2 }{ \sum_{p=1}^{n-1} (v_{m,k}^{(p)})^2 } $$
This index extracts the contribution of the fault to each physical differential channel. After remapping through the virtual topology matrix, the algorithm pinpoints the cell exhibiting the most profound anomaly by comparing against a threshold \( \eta_{th} \). The process is illustrated as an assessment of the abnormal eigenvector’s contribution. This parallel implementation does not add latency to the fault management process, providing the BMS with an instant address for the affected cell.
4.2. Synergistic Correction of Detector Alerts
A primary caveat of a highly sensitive threshold \( \delta_d \) is that transient, high-current driving conditions—which induce temporary parameter mismatches—might generate false-positive alerts. However, these current-driven noise signals typically have low spatial coherence. The eigen-decomposition localizes the variance across several principal components. Thus, the corresponding localization indices \( \eta_p \) will be diffuse, with no single cell having a dominant contribution. In contrast, a genuine SSC event is spatially concentrated.
By requiring both conditions—the Detection Indicator exceeding the control limit \( DI_k > \delta_d \) and a specific localization index exceeding the contribution threshold \( \eta_p > \eta_{th} \)—the BMS can effectively filter out current-pulse-induced ‘pseudo-faults’. This built-in redundancy acts as a logic AND gate. Moreover, it offers the strategic advantage of potentially reducing the control limit \( \delta_d \), thereby increasing the detection speed for mild SSCs, while simultaneously preserving a low false positive rate. This is crucial in real-world dynamic driving cycles where current spikes are frequent.
5. Experimental Validation and Discussion
5.1. Experimental Platform and Fault Emulation
To validate the performance of my diagnostic scheme, I constructed a representative EV battery pack experimental platform. This platform integrates a real-time BMS prototype, capable of monitoring voltages (via the MC33771 analog front end) at a high sampling rate of 10 Hz, which is critical for capturing transient voltage dynamics. The cells in the test rig are 14 series-connected A123 ANR26650M1B LiFePO₄ cylindrical cells, which are renowned for their high-power density and safety. Key parameters are shown in Table II.
| Parameter | Value | Parameter | Value |
|---|---|---|---|
| Nominal Voltage | 3.3 V | Nominal Capacity | 2.5 Ah |
| Internal Resistance | 6 mΩ | Standard Charge Rate | 1C |
| Voltage Range | 2.0 – 3.6 V | Max. Discharge Rate | 28C |
The test sequence adhered to the WLTC protocol, whose current profile has a maximum magnitude of 2C (approx. 5A) to emulate standard urban driving. Two classes of faults were induced on cell #7:
Transient SSC: This simulates a sudden mechanical breach. A 10Ω resistor is abruptly connected in parallel with the healthy cell via a relay, creating a “hard” but mild short. The resistance is selected to emulate the early stage of internal short circuit (IS). The fault is applied and then shorted briefly.
Cumulative SSC: This simulates the growth of lithium dendrites puncturing the separator. Rather than a fixed resistance, a time-varying impedance \( R_{sc}(t) \) is emulated through a programmable load, which slowly decreases its resistance over minutes according to the smooth profile. This leads to a gradual voltage divergence. Table III contrasts the two fault classes.
| Type | Etiology | Equivalent Feature | Simulation Mode |
|---|---|---|---|
| Transient SSC | Mechanical crush / sudden rupture | Sudden voltage drop | Constant parallel resistor |
| Cumulative SSC | Dendrite growth / metallic dissolution | Gradual voltage drift | Time-varying parallel resistance |
5.2. Comparative Diagnostic Performance
The performance of my proposed full-dimensional statistical analytics method, termed the ‘Eigen Contribution Fault Diagnosis’ (ECFD) was compared against two recently documented benchmarks in the field: the Voltage Cosine Similarity (VCS) method and the Independent Component Analysis based on Voltage Correlation (ICA-VC). I assessed the detection time relative to fault initiation, the continuity of the alert, and the fault detection rate (FDR) across a comprehensive SOC scope, spanning from 93.01% down to 9.66% (covering 83% of the rated capacity), with the initial NCVA alignment presetting the cell voltage. The quantitative results are detailed in Table IV.
| Fault Case | Method | Detection Time (s) | Continuous Alert | Alert after Removal | Fault Cell Detected | Detection Rate (%) |
|---|---|---|---|---|---|---|
| Transient SSC | ICA-VC | 9.9 | No | Yes | #7 | 0.03 |
| VCS | 80.1 | No | No | #7 | 0.0002 | |
| My ECFD Method | 109.6 | Yes | Yes | #7 | 98.4 % | |
| Cumulative SSC | ICA-VC | 649.8 | No | No | None | 0.0 |
| VCS | — | No | No | None | 0.0 | |
| My ECFD Method | 171.3 | Yes | Yes | #7 | 97.6 % |
The comparison clearly illuminates the superior robustness of my ECFD method. While the ICA-VC method exhibited a faster alert in the transient scenario (9.9 seconds), it failed to sustain the alert, giving only a momentary spike that decays as the polarization dynamics settle. It also failed to catalogue the cumulative fault. VCS significantly lagged in the transient case and entirely missed the cumulative-type fault. My diagnostic framework, in contrast, detects the soft instability in minutes and, due to the inherent OCV offset brought by the fault, correctly triggers a persistent alarm even after the physical short circuit is removed. This post-removal alarm is a crucial safety feature that alerts the driver to a latent degradation in the pack, which persists even if the primary short circuit is thought to have been cleared. The fault detection rates of 98.4% and 97.6% for transient and cumulative scenarios dwarf those of the comparison models (under 1%), demonstrating that the eigen-space transformation amplifies the weak signal-to-noise ratio of the mild SSC faults to a readily measurable quantity.
Further analyzing the robustness of my method, I studied the interplay between the detection threshold and the minimum detectable fault voltage. Figure 1 in my experimental notes illustrates that when the fault localization step is active, the required detection voltage plateaus even as the detection threshold is reduced significantly. When the detection threshold \( \delta_d \) falls below 8.122, the theoretical and experimental detectable fault levels begin to show a marked variance. This demonstrates that without the localization step’s filter, the detector would suffer from an exorbitant false alarm rate, especially if the threshold was reduced without this guard. The localization filter effectively decouples detection limit from false alarm rate, allowing an aggressive reduction of the threshold. By reducing the confidence interval from \( \alpha = 1-10^{-6} \) to \( \alpha = 0.998 \), the immediate fault detection time was expedited to 101.8 seconds earlier for the transient fault and 106 seconds earlier for the cumulative fault, a minimization of at least 61.9% in alarm response time without triggering spurious ‘pseudo-fault’ alerts. This indicates that the threshold can be lowered aggressively in my method since the localization step can validate the true spatial nature of the fault.
6. Conclusion and Future Directions
In my investigation, I have systematically engineered and validated an advanced diagnostic architecture tailored for the safety monitoring of the EV battery pack. My research findings, both in theoretical derivation and empirical validation, yield several key conclusions:
1. Measurement enhancement is foundational. The introduction of a virtual voltage measurement strategy effectively suppresses the noise imprint caused by cell-to-cell parameter dispersion. By simultaneously aligning the operating point on the OCV plateau and re-mapping the spatial sequence via eigenvectors, the differential voltage variability metric (mean deviation) was curtailed by 89.3%, from 0.880 mV to 0.094 mV, critically sharpening the contrast for weak fault signals.
2. Interpretability in detection. The Fault Detection Indicator (DI) based on the convergent eigenvalues of the covariance matrix provides a statistically robust metric. Crucially, I derive an analytic, quantitative expression relating this DI statistical threshold to the fault signature magnitude. This interpretability empowers the BMS to adapt the control limit proactively. A user-defined requirement for fault amplitude (e.g., 10.7 mV) can be directly converted to the corresponding detection threshold, thereby liberating the system from relied ad-hoc tuning.
3. Synergistic detection and localization. Capitalizing on the mathematical properties of eigen-decomposition, localization does not lag but occurs simultaneously with detection. An isolated anomaly manifests a high energetic contribution in one specific eigenvector. This not only accurately identifies the offending cell but also acts as a spatial filter against noise originating from high-frequency current ripple. I demonstrate that by using a contribution threshold \( \eta_{th} = 0.8 \) and an advanced eigenvalue decomposition along with recursive time updates, the diagnostic process maintains high accuracy while being computationally lean for embedded processors.
4. Comprehensive experimental validation. Under the diverse WLTC current profile, my method was able to identify a 10Ω SSC, replicating a reasonable worst-case internal fault condition. It boasted a fault detection rate exceeding 97.6%, which is an exponential improvement over the performance of benchmark VCS and ICA-VC methods under identical settings. Furthermore, its ability to identify degradation and emit sustained alerts is an invaluable asset for proactive battery health management, rather than a mere binary “fault/no-fault” system.
Looking ahead, an avenue for further refinement is the quantification and decoupling of the “pseudo-fault” characteristics inherent in dynamic IR shifts under high C-rate pulses in large-scale EV battery packs. As electric vehicle fast-charging capabilities push operational limits to 4C or higher, standardizing this fault-filtration process becomes paramount. My future work will focus on storing and applying this “virtual alignment” in a dynamic, cycle-life-aware manner to ensure these advanced safety diagnostics remain robust over a 10-year, 300,000-mile lifespan. By maintaining such rigorous fault detection capabilities, the transition to sustainable e-mobility can be accelerated with confidence in its inherent safety.