As global environmental pollution and energy crises intensify, the development of electric vehicles has become a strategic imperative. Among numerous technologies, the lithium-ion power battery system is the core energy storage unit of an EV battery pack. Its safety, reliability, and lifespan directly determine the overall performance and driving safety of the vehicle. In recent years, fire incidents involving electric vehicles have frequently occurred, and many of these accidents are closely linked to latent faults within the battery modules, such as loose connections (virtual connection) and micro-short circuits (MSC).
In my research, I focus on the fault diagnosis of an EV battery pack, aiming at mitigating the influence of critical faults by employing a data-driven scheme together with model-based quantitative analysis. A complete diagnostic framework is proposed, which consists of signal denoising, voltage feature enhancement, window-based sample reconstruction, convolutional neural network classification, and incremental capacity analysis. The rest of this article is organized as follows. First, the working principle and failure mechanisms of lithium-ion cells are elaborated. Second, the experimental setups for both virtual connection and MSC faults are described. Third, data pre-processing methods, including an improved voltage normalization and sliding-window augmentation, are introduced. Fourth, a one-dimensional and a two-dimensional CNN are designed and compared for virtual connection fault grading. Fifth, an incremental capacity method is adopted to quantitatively estimate the MSC resistance. Finally, conclusions are drawn.
1. Basic Characteristics of Lithium-Ion Cells in EV Battery Packs
A lithium-ion cell typically consists of a positive electrode, a negative electrode, an electrolyte, a separator, current collectors, and an outer casing. The positive electrode materials used in an EV battery pack determine the cell’s energy density, thermal stability, and cost. Table 1 compares the major positive electrode materials for lithium-ion batteries, highlighting their influence on EV battery pack attributes.
| Material | Specific Energy (Wh/kg) | Max Charge Rate (C) | Max Discharge Rate (C) | Cycle Life | Safety | Cost |
|---|---|---|---|---|---|---|
| LiCoO2 | 150-200 | 0.7-1 | 1 | 500-1000 | Medium | High |
| LiMn2O4 | 100-150 | 0.7-1 | 1 | 300-700 | High | High |
| LiFePO4 | 90-120 | 1 | 2 | 1000-2000 | Medium | High |
| NMC | 150-220 | 0.7-1 | 2 | 1500-2000 | High | Medium |
| NCA | 200-260 | 0.7 | 1 | 500-1000 | Medium | Medium |
In my work, the tested cells used in the EV battery pack experiments are commercial pouch cells from Tianjin Lishen Battery Co., with a nominal capacity of 50 Ah. The detailed specification of these cells is listed in Table 2.
| Capacity (Ah) | 50 |
| Charge cutoff voltage (V) | 4.2 |
| Discharge cutoff voltage (V) | 3.0 |
| Dimensions (mm) | 170 × 43 × 100 |
2. Failure Mechanisms and Fault Principles
Lithium-ion cells undergo several ageing and degradation mechanisms during operation. These include dissolution of the transition metal from the positive electrode, decomposition of the electrolyte, phase transition of active materials, growth of the solid–electrolyte interphase film, and loss of conductive contact. Such degradation not only reduces the capacity and increases the internal resistance, but also influences the voltage characteristics that are used for fault detection in an EV battery pack.
For series-connected cells, the terminal voltage of each cell can be described by the Thevenin equivalent circuit model. The cell circuit model containing the connection resistance is shown in Figure 1. The terminal voltage equation is expressed as
$$U_n(t) = U_{ocv,n}(t) – I(t)(R_{n} + R_{c,n}) – U_{p,n}(t),$$
where \(I(t)\) is the operating current of the EV battery pack, \(R_{c,n}\) is the contact resistance between the connection strip and the cell terminal, \(R_n\) is the ohmic internal resistance, \(U_{p,n}\) is the polarization voltage, and \(U_{ocv,n}\) is the open-circuit voltage. When a virtual connection occurs, the contact area fluctuates due to vibration and mechanical shock, leading to a stochastic change of \(R_{c,n}\). As a result, the terminal voltage shows abnormal high-frequency spikes and drops.
For micro-short-circuit faults, the direct cause is the formation of small-scale internal short paths between the positive and negative electrodes. A common model is obtained by inserting a leakage resistor \(R_{ISC}\) in parallel with the battery equivalent circuit. The relevant current balance is
$$I_{leak} = I – I_{effective},$$
where \(I\) is the applied current, \(I_{leak}\) is the leakage current through the MSC resistor, and \(I_{effective}\) is the actual current that participates in the electrochemical reaction. Because the leakage current consumes part of the charging current, the MSC cell in an EV battery pack displays a gradual voltage deviation from a healthy cell.
3. Experimental Design of Fault Simulation
3.1 Virtual connection fault simulation
The virtual connection test bench consists of a vibration platform, voltage acquisition sensors, a battery charge/discharge tester (BST-600), and a data acquisition system. Several cells were connected in series to form a simplified EV battery pack. The test module was placed on the shaking table to reproduce the vibration environment that occurs while an electric vehicle is running on a rough road. I applied a simplified but realistic loading current profile to the module, as shown in Table 3.
| Current (A) | 0 | -30 | -45 | 10 | -10 | 45 | 30 |
|---|---|---|---|---|---|---|---|
| Time (s) | 30 | 180 | 60 | 90 | 90 | 60 | 180 |
The experiment was initially healthy. Then the bolt that connects one cell was loosened gradually, reflecting the dynamic growth of a virtual connection fault. According to the voltage fluctuation intensity, four health states were defined for the EV battery pack: healthy, slight fault, moderate fault, and severe fault. The collected voltage curves show that in the severe stage, sparks were observed at the loosened terminal, which indicates a dangerous state.
3.2 Micro-short-circuit fault simulation
For the MSC fault study, a simulation model was built in MATLAB/Simulink using several lithium-ion battery blocks connected in series. A resistance of 5 Ω, 50 Ω, or 100 Ω was placed in parallel with one cell to mimic severe, moderate, and light MSC faults, respectively. The cell model had a rated capacity of 2.3 Ah, a nominal voltage of 3.3 V, an initial SOC of 100%, an internal resistance of 0.014 Ω, and an initial temperature of 25 °C. The charging current was set to 0.5 C. The voltage curves from the simulation clearly show an increasing deviation for smaller parallel resistance.
To further validate the algorithm, a hardware-in-the-loop MSC experiment was performed on a Neware CT-4008-DB battery tester. A healthy Li-ion cell was first activated by three charge/discharge cycles with 1 C current in the voltage range 3.0–4.2 V. Then a controllable external resistor of 20 Ω was connected across the cell terminals during a constant-current charge to simulate an MSC. This experimental dataset was used to verify the quantitative diagnosis result.
4. Data Pre-Processing and Feature Enhancement
4.1 Noise filtering using EMD and correlation
The raw voltage signals collected from the EV battery pack always contain unavoidable noise caused by sensor error and electromagnetic interference. To extract the fault information hidden in the signal, I employed an EMD-based denoising strategy. Empirical mode decomposition adaptively decomposes the measured voltage signal \(x(t)\) into a finite number of intrinsic mode functions {\(c_i(t)\)} and a residue \(r_n(t)\):
$$x(t) = \sum_{i=1}^{n} c_i(t) + r_n(t).$$
After decomposing the voltage signal, I computed the Pearson correlation coefficient between each IMF and the original voltage signal to distinguish informative IMFs from noise:
$$r = \frac{\sum_{i=1}^{n}(X_i – \bar{X})(Y_i – \bar{Y})}{\sqrt{\sum_{i=1}^{n}(X_i – \bar{X})^2}\sqrt{\sum_{i=1}^{n}(Y_i – \bar{Y})^2}}.$$
IMFs with very low correlation values are treated as noise and are discarded. Then, the remaining IMFs and the residue are summed to reconstruct the denoised voltage. In Table 4, the correlation values of different IMFs are presented. For the first channel, IMF15 and IMF16 have the highest correlation and are preserved, while the first fourteen IMFs are regarded as noise due to their low correlation coefficients.
| IMF index | Ch1 | Ch2 | Ch3 | Ch4 |
|---|---|---|---|---|
| IMF1 | 0.046 | 0.025 | 0.022 | 0.035 |
| IMF2 | 0.044 | 0.018 | 0.032 | 0.033 |
| IMF3 | 0.024 | 0.020 | 0.027 | 0.042 |
| IMF4 | 0.031 | 0.027 | 0.045 | 0.044 |
| IMF5 | 0.043 | 0.034 | 0.051 | 0.041 |
| IMF6 | 0.028 | 0.027 | 0.075 | 0.065 |
| IMF7 | 0.047 | 0.064 | 0.082 | 0.061 |
| IMF8 | 0.062 | 0.058 | 0.089 | 0.137 |
| IMF9 | 0.052 | -0.198 | 0.179 | 0.084 |
| IMF10 | 0.339 | 0.371 | 0.152 | 0.784 |
| IMF11 | 0.893 | 0.883 | 0.743 | 0.149 |
| IMF12 | 0.139 | 0.103 | 0.526 | — |
| IMF13 | -0.094 | — | — | — |
I further compared EMD with two improved variants, EEMD and CEEMDAN. EEMD adds random white noise to the target signal and averages the ensemble results, which effectively mitigates mode-mixing. CEEMDAN adds adaptive white noise at each stage and obtains a complete decomposition with negligible reconstruction error. However, both methods consume more computation time than EMD. Considering the requirement of accurate and rapid fault diagnosis for the EV battery pack, EMD was selected as the default filter. This choice is validated by the diagnostic experiments presented in Section 5.
4.2 Improved voltage data transformation
Even when the cells in an EV battery pack are selected with similar initial capacity and internal resistance, manufacturing tolerances and ageing imbalance produce unavoidable inconsistencies. This inconsistency often masks weak fault features. In addition, the absolute voltage differences between the cells are relatively small, which hinders the feature extraction ability of a convolutional neural network. I therefore propose a simple but effective voltage transformation:
$$\tilde{v}_{i,j} = \left( v_{i,j} – \bar{v}_j \right) \times \alpha,$$
$$\bar{v}_j = \frac{1}{n} \sum_{i=1}^{n} v_{i,j},$$
where \(v_{i,j}\) is the denoised voltage of the \(i\)-th cell at the \(j\)-th sample time, \(\bar{v}_j\) is the average voltage of all cells at the same sample time, \(n\) is the total number of cells, and \(\alpha\) is an amplification coefficient. In this study, \(\alpha\) is set to 1000.
Subtracting \(\bar{v}_j\) reduces the influence of the common-mode voltage component, which increases with the cell state of charge and with dynamic current variations. Multiplying by \(\alpha\) enlarges the tiny differential voltage so that the CNN can rapidly identify the abnormal pattern. The transformed voltage may be positive or negative, such that the polarity itself carries useful information.
4.3 Sliding-window sample augmentation
Deep learning models usually require a large number of training samples. My raw fault experiment produced only 7200 sampling instants for each cell. If I simply divide the entire sequence into non-overlapping segments, the number of samples is limited to \(7200/m\). To provide enough data for training the CNN, I apply a sliding-window approach with window length \(m\). For every time step, a window of size \(m\) is moved forward by one step, generating many overlapping samples from the same sequence. The augmented sample set greatly improves the stability of the CNN classifier.
5. Convolutional Neural Network Based Online Fault Diagnosis
CNN is a powerful deep learning architecture widely used for image and signal classification. A typical CNN structure is composed of an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. During the forward phase, each convolutional kernel slides over the input and computes the dot product with the local receptive field. The pooling layer reduces the feature dimension and helps to extract invariant features. Activation functions such as Sigmoid, Tanh, and ReLU introduce nonlinearity. For multi-class classification, the softmax layer is placed at the end of the network.
In this work, I treat the transformed voltage data of the EV battery pack as an input sequence or an input pseudo-image. For the one-dimensional CNN, each sample is a one-dimensional vector containing the differential voltages of four cells in consecutive time points. For the two-dimensional CNN, a sample is organized as a matrix with dimensions (number of time steps, number of cells), which is similar to an image. The two CNN structures are compared for virtual connection fault classification.
5.1 One-dimensional CNN for virtual connection classification
The architecture of the proposed 1D CNN is given in Table 5. It includes two convolutional layers with 32 and 64 filters, two max-pooling layers, a flatten layer, a dense layer with 128 neurons, and a final dense layer with 4 neurons. Tanh is used as the activation function in convolutional and dense layers.
| Layer | Output size | Details |
|---|---|---|
| Input | 200 × 1 | 200 time steps, 1 channel |
| Conv1D_1 | 198 × 32 | kernel size 3, stride 1, tanh |
| MaxPooling1D_1 | 99 × 32 | pool size 2 |
| Conv1D_2 | 97 × 64 | kernel size 3, stride 1, tanh |
| MaxPooling1D_2 | 48 × 64 | pool size 2 |
| Flatten | 3072 | — |
| Dense_1 | 128 | tanh |
| Dropout | 128 | rate 0.5 |
| Dense_2 | 4 | softmax |
During training, the categorical cross-entropy loss is minimized using the Adam optimizer. The loss function is
$$\mathcal{L} = -\frac{1}{N}\sum_{i=1}^{N}\sum_{c=1}^{M} y_{ic} \log p_{ic},$$
where \(N\) is the batch size, \(M\) is the number of classes, \(y_{ic}\) is a binary indicator of the true label and \(p_{ic}\) is the predicted probability for class \(c\).
5.2 Two-dimensional CNN for virtual connection classification
The 2D CNN uses two-dimensional convolution filters both in the time and cell dimensions. The architecture is listed in Table 6. Several convolutional blocks with 2×2 filters are used, followed by max-pooling layers and an adaptive average pooling layer. A dense layer maps the high-level features to the four fault classes.
| Layer | Output shape | Details |
|---|---|---|
| Input | (36, 4, 1) | 36 time steps, 4 cells, 1 channel |
| Conv2D_1 | (35, 3, 64) | 2×2 kernels, tanh |
| Conv2D_2 | (34, 2, 64) | 2×2 kernels, tanh |
| MaxPooling2D_1 | (17, 1, 64) | pool size 2×2 |
| Conv2D_3 | (16, 1, 128) | 2×1 kernels, tanh |
| Conv2D_4 | (15, 1, 128) | 2×1 kernels, tanh |
| GlobalAveragePooling2D | (128) | — |
| Dense_1 | 128 | tanh |
| Dense_2 | 4 | softmax |
5.3 Evaluation of the sliding-window size
For the 1D CNN, I first set the sliding window length to 20, and the classification accuracy was 99.77% after 75 epochs. The improvement was stable and converged quickly. When the raw un-augmented samples were used, the training accuracy increased significantly but the testing accuracy remained around 25%, because the size of the training dataset was too small. This comparison substantiates the indispensable role of sliding-window augmentation.
For the 2D CNN, I varied the window length from 20 to 40, and the results are summarized in Table 7.
| Window length | Accuracy (%) | Training time (s) |
|---|---|---|
| 20 | 100 | 78.59 |
| 30 | 99.95 | 68.40 |
| 36 | 100 | 63.71 |
| 40 | 100 | 80.34 |
From Table 7, all window sizes lead to high accuracy, but the window length of 36 gives the shortest training time. Therefore, I choose a window size of 36 as the default for the 2D CNN, which reduces the time by 23.36% compared with the window size of 20 and by 26.10% compared with the window size of 40.
5.4 Effect of denoising algorithms on classification
I also compared three denoising methods with the 2D CNN. Table 8 shows the accuracy and training time when EMD, EEMD, and CEEMDAN are used to filter the voltage signals.
| Window | EMD accuracy | EMD time (s) | EEMD accuracy | EEMD time (s) | CEEMDAN accuracy | CEEMDAN time (s) |
|---|---|---|---|---|---|---|
| 20 | 100% | 78.59 | 99.63% | 104.29 | 99.30% | 82.77 |
| 30 | 99.95% | 68.40 | 99.72% | 89.47 | 99.25% | 89.92 |
| 36 | 100% | 63.71 | 99.76% | 86.39 | 99.91% | 105.27 |
| 40 | 100% | 80.34 | 100% | 107.14 | 99.53% | 77.78 |
EMD offers the highest accuracy and generally the shortest training time. Although EEMD and CEEMDAN improve the decomposition quality, their extra computational burden may delay the fault warning, which is unacceptable for EVs because a fire can develop within seconds. Therefore, EMD is preferred in my diagnostic framework.
5.5 Comparison with conventional classifiers
To demonstrate the superiority of the proposed CNN methods, I compared them with a genetic-algorithm-optimized extreme learning machine (GA-ELM) and a grid-search-optimized support vector machine (GS-SVM). The diagnostic results are listed in Table 9.
| Method | Accuracy (%) | Computation time (s) |
|---|---|---|
| GA-ELM | 97.97 | — |
| GS-SVM | 97.50 | 167.38 |
| 1D CNN | 99.77 | 100.48 |
| 2D CNN | 100 | 63.71 |
Both CNN structures outperform the traditional machine-learning classifiers in accuracy. The 2D CNN achieves the highest accuracy and the shortest training time when the window size is 36. Compared with GS-SVM, the 2D CNN reduces the diagnostic time by more than 160%. This is because CNN learns hierarchical features automatically and is less sensitive to manually designed features.
6. Virtual Connection Fault Grading Results
Based on the voltage fluctuation strength of the virtual connection fault, I categorized the states into four grades:
– Healthy (H): no abnormal voltage fluctuation,
– Slight fault (SF): very small but detectable voltage perturbations,
– Moderate fault (MF): visible voltage deviations in a few pulses,
– Severe fault (VF): large voltage drops and sparks at the loose terminal.
Using the proposed voltage transformation and sliding-window augmentation, I trained both the one-dimensional and two-dimensional CNN models. The models achieved almost perfect classification on the test set. In particular, the 2D CNN reached 100% classification accuracy for all classes after about 12 epochs, while the 1D CNN achieved 99.77% accuracy after 75 epochs. This verifies that the proposed data-processing steps greatly assist CNN in extracting discriminative fault features.
7. Incremental Capacity Method for MSC Fault Diagnosis
7.1 Qualitative diagnosis using CNN
Before quantitative analysis, I applied the same CNN framework to detect whether a micro-short-circuit exists in a battery pack. The simulation produced four classes: healthy, MSC with 100 Ω, MSC with 50 Ω, and MSC with 5 Ω. The voltage data were denoised, transformed, and augmented using the procedure described above. The CNN model was trained on those samples. The accuracy of the classifier exceeded 95% after 13 epochs, indicating that the proposed model has good generalization ability. Because the same CNN architecture is successfully used for virtual connection faults and MSC faults, the framework can be extended to other fault types without major structural changes.
7.2 Incremental capacity curve analysis
The incremental capacity (IC) curve is defined by the derivative of charging capacity with respect to voltage. For constant-current charging, the IC value is calculated as
$$IC = \frac{dQ}{dV} \approx \frac{\Delta Q}{\Delta V} = \frac{I_c \Delta t}{\Delta V},$$
or alternatively,
$$IC = \frac{\Delta Q_c}{V_{c,2} – V_{c,1}},$$
where \(I_c\) is the charging current, \(\Delta t\) is the time interval, \(Q_c\) is the charged capacity, and \(V_{c,1}, V_{c,2}\) are two adjacent voltage points.
If a cell has an MSC, part of the charging current is shunted through the internal short path. Consequently, under the same terminal voltage rise, the MSC cell needs a larger apparent charging capacity than the healthy cell. The IC curve of the MSC cell is therefore shifted upward compared with the healthy cell. By integrating the difference in the IC curve over the same voltage window, I can obtain the amount of leaked charge \(\Delta Q_{leak}\):
$$\Delta Q_{leak}^{(AB)} = \int_{A}^{B} \left( IC_{ISC}(V) – IC_{healthy}(V) \right) dV.$$
In practice, I approximate the integral by summing the area over the voltage interval between two well-defined IC peaks. Then the average leakage current in this interval is
$$I_{leak}^{(AB)} = \frac{\Delta Q_{leak}^{(AB)}}{\Delta t_{AB}},$$
where \(\Delta t_{AB}\) is the time elapsed while the terminal voltage increases from voltage A to voltage B.
7.3 Quantitative estimation of MSC resistance
The average terminal voltage in the window from A to B is approximated by the arithmetic mean of all sampled voltages between those points:
$$\bar{U}_{AB} = \frac{1}{n}\sum_{i=1}^{n} U_i,$$
where \(n\) is the number of sample points and \(U_i\) is the measured voltage. Using Ohm’s law, the MSC resistance is obtained by
$$R_{ISC} = \frac{\bar{U}_{AB}}{I_{leak}^{(AB)}}.$$
In the simulation study, I selected the voltage range from 3.3675 V to 3.3924 V. The healthy battery and MSC batteries with 100 Ω and 50 Ω resistances were analyzed. The integrated charges in this window are shown in Table 10.
| Battery condition | Charge variation (Ah) | Leaked charge (Ah) |
|---|---|---|
| Healthy | 1.96718 | — |
| MSC (50 Ω) | 2.08830 | 0.12112 |
| MSC (100 Ω) | 2.02630 | 0.05912 |
The corresponding charging time, computed leakage current, average voltage in the window, and estimated MSC resistance are listed in Table 11.
| Battery condition | Time interval (s) | Leakage current (A) | Average voltage (V) | Actual resistance (Ω) | Estimated resistance (Ω) | Error (%) |
|---|---|---|---|---|---|---|
| MSC (50 Ω) | 6537 | 0.0667 | 3.38179 | 50 | 50.72 | 1.44 |
| MSC (100 Ω) | 6343 | 0.03355 | 3.38225 | 100 | 100.81 | 0.81 |
The good agreement between the estimated and actual resistance values demonstrates that the proposed IC-based method is capable of quantitatively diagnosing MSC faults with high accuracy.
7.4 Hardware experiment validation
To further confirm the reliability of the IC-based diagnostic method, I performed an offline experiment in which an MSC resistance of 20 Ω was deliberately connected in parallel with a real lithium-ion cell. I used a constant-current charging protocol after activation. The measured voltage of the MSC cell increasingly deviated from that of the healthy cell during charging. I then constructed the IC curves for both the healthy and the MSC cell. The second IC peak was chosen as the target voltage window, because the first peak was distorted by the transition from the constant-current phase, and the fourth peak was truncated by the cutoff voltage.
The charge variations in the selected voltage window are given in Table 12.
| Cell status | Charge variation (Ah) | Leaked charge (Ah) |
|---|---|---|
| Healthy | 0.761989 | — |
| MSC cell | 0.824923 | 0.062934 |
Using the charging time of this voltage window, the leakage current was calculated as 0.19087 A. The average terminal voltage in the window was 3.7556 V. Therefore, the estimated MSC resistance was
$$R_{ISC,est} = \frac{3.7556}{0.19087} \approx 19.68 \, \Omega.$$
The actual resistance of the parallel resistor is 20 Ω, resulting in an estimation error of only 1.6%. This experiment validates that the proposed methodology is sufficiently accurate for practical applications.
8. Discussion
Through extensive experimental and numerical investigations, I have verified the efficacy of the whole diagnosis chain for an EV battery pack. The main findings can be summarized as follows:
First, EMD-based denoising can remove high-frequency noise effectively. Compared with EEMD and CEEMDAN, EMD has lower computational cost and is therefore more suitable for real-time fault warning. Second, subtracting the average voltage of all series-connected cells enlarges the subtle deviation caused by faults and suppresses the common mode caused by SOC changes. The amplification factor α further accelerates the CNN feature extraction process. Third, sliding-window sample generation overcomes the small-sample problem of deep learning. Fourth, the 2D CNN with the optimal window length of 36 reaches the highest diagnosis accuracy while using the shortest training time. Finally, the IC curve difference between an MSC cell and a healthy cell can be used to estimate the leakage current and the MSC resistance with an error of less than 2%.
It should be mentioned that this study assumes that a single cell experiences only one type of fault at a time. In real driving environments, multiple failures may coexist, such as virtual connection plus micro-short-circuit, or thermal degradation plus sensor drift. Future research should address multi-fault coupling mechanisms and develop fault-isolation logic for an EV battery pack under realistic dynamic profiles. In addition, aging-induced parameter variations should be separated from the fault features by introducing an aging factor into the diagnostic model.
9. Conclusion
In this paper, I have proposed a comprehensive fault-diagnosis framework for lithium-ion power batteries used in an EV battery pack. The framework integrates a noise-filtering method based on empirical mode decomposition, a voltage difference transformation, a sliding-window data augmentation algorithm, convolutional neural network classifiers for fault grading, and an incremental-capacity method for micro-short-circuit quantification.
Experimental results show that both the one-dimensional and the two-dimensional CNN can identify healthy, slight, moderate, and severe virtual connection states in an EV battery pack. The 2D CNN with a window size of 36 offers a classification accuracy of 100% and requires only 63.71 s for training. The same CNN architecture is successfully applied to the qualitative detection of MSC faults, exhibiting good generalization ability. For quantitative MSC diagnosis, the incremental capacity method estimates the MSC resistance with errors of 1.44% and 0.81% for simulated short-circuit resistances of 50 Ω and 100 Ω, respectively. In a hardware experiment with a 20 Ω parallel resistor, the equivalent MSC resistance was predicted as 19.68 Ω, leading to an error of merely 1.6%.
In the future, I plan to extend the method to diagnose more complex faults such as simultaneous virtual connection and MSC conditions, as well as to incorporate thermal and mechanical data from the EV battery pack to improve early warning capability. The proposed framework is expected to become a useful component of next-generation battery management systems, improving the safety and reliability of electric vehicles.