In recent years, the rapid growth of conventional internal combustion engine vehicles has intensified both energy security concerns and environmental degradation. The transition toward new energy vehicles has become an inevitable trend in the global automotive industry. As the core energy storage component of electric vehicles, the EV battery pack plays a pivotal role in determining vehicle safety, reliability, and overall performance. However, the inherent electrochemical complexity of lithium-ion cells, combined with harsh operating conditions such as high charge-discharge rates, wide temperature variations, and mechanical stress, often leads to various fault conditions including overcharging, over-discharging, internal resistance growth, capacity degradation, and thermal runaway. These faults not only degrade the performance of the EV battery pack but also pose serious threats to human life and property. Consequently, accurate and timely fault diagnosis of the EV battery pack has become an urgent research topic.
In this dissertation, I propose a data-driven fault diagnosis methodology for the EV battery pack based on a hybrid Long Short-Term Memory (LSTM) neural network and Back-Propagation (BP) neural network, referred to as the LSTM-BP model. My research systematically addresses the entire diagnostic pipeline: experimental platform construction, data acquisition, data preprocessing, equivalent circuit modeling, fault simulation, neural network model development, and diagnostic threshold design. The proposed method leverages the temporal sequence learning capability of LSTM networks and the nonlinear mapping capability of BP networks, thereby overcoming the limitations of each individual model and achieving superior diagnostic accuracy for the EV battery pack.
1. Introduction and Research Background
The rapid depletion of fossil fuels and the escalating severity of environmental pollution have motivated governments worldwide to prioritize the development of clean and efficient transportation solutions. Electric vehicles, particularly battery electric vehicles, have emerged as the most promising alternative to conventional automobiles. The EV battery pack is the most critical subsystem in an electric vehicle, directly influencing driving range, acceleration performance, and operational safety. Despite significant advances in battery technology, lithium-ion batteries still suffer from performance degradation and potential safety hazards under abusive or abnormal operating conditions.
Statistical data indicate that more than 65% of fire and combustion accidents involving electric vehicles are caused by failures of the EV battery pack. For instance, in 2024 alone, numerous fire incidents involving new energy vehicles were reported across China, with the majority attributed to thermal runaway or electrical faults within the battery system. These incidents underscore the importance of developing robust fault diagnosis techniques that can promptly detect abnormal states and provide early warnings before catastrophic failures occur.
Existing fault diagnosis methods for the EV battery pack can be broadly classified into three categories: knowledge-based methods, analytical model-based methods, and data-driven methods.
Knowledge-based approaches rely on expert experience and established fault knowledge bases. These methods utilize rule-based reasoning, fuzzy logic, and expert systems to infer fault conditions. Although these methods do not require complex mathematical models, they depend heavily on the completeness and accuracy of the knowledge base. Moreover, as battery technology evolves rapidly, maintaining an up-to-date fault knowledge base becomes increasingly challenging.
Analytical model-based methods employ equivalent circuit models or electrochemical models to generate residuals between predicted and measured battery states. Methods such as parity space, Kalman filtering, and parameter identification have been widely adopted. While these approaches provide physical insights into battery behavior, they require precise model construction and careful threshold selection. Additionally, most model-based methods are tailored for specific fault types and may not generalize well to multiple simultaneous faults.
Data-driven methods have attracted considerable attention in recent years due to their ability to learn complex nonlinear relationships directly from measured data. Techniques including artificial neural networks, support vector machines, deep learning, and ensemble learning have been successfully applied to EV battery pack fault diagnosis. Among these, LSTM neural networks have demonstrated exceptional performance in handling time-series battery data, as they can capture long-term dependencies and temporal dynamics inherent in charge-discharge cycles. However, standalone LSTM models may suffer from overfitting and insufficient feature extraction for high-dimensional battery data. To address these limitations, I propose the LSTM-BP hybrid model, which integrates the strengths of both LSTM and BP neural networks.
2. Working Principles and Fault Analysis of Power Batteries
2.1 Fundamental Working Principles
The lithium-ion battery, which serves as the core of the EV battery pack, operates based on the principle of lithium-ion intercalation and deintercalation between the positive and negative electrodes. During charging, lithium ions are extracted from the cathode material (e.g., LiMn₂O₄, LiFePO₄, or Li(NiCoMn)O₂), migrate through the electrolyte, and intercalate into the anode material (typically graphite). During discharging, the reverse process occurs, with lithium ions moving from the anode back to the cathode. The overall electrochemical reaction can be represented as:
$$ \text{LiMeO}_2 + C_6 \xrightarrow[\text{discharge}]{\text{charge}} \text{Li}_{1-x}\text{MeO}_2 + \text{Li}_x C_6 $$
where Me represents the transition metal element. The voltage difference between the two electrodes, known as the open-circuit voltage (OCV), is determined by the difference in electrochemical potentials of the two electrodes, which in turn depends on the state of charge (SOC) and temperature.
2.2 Key Performance Parameters
To effectively diagnose faults in the EV battery pack, it is essential to understand the key performance parameters that characterize battery behavior. These parameters include capacity, voltage, current, internal resistance, specific energy, energy density, specific power, power density, and SOC. The relationships among these parameters provide critical clues for fault detection and diagnosis. Below I summarize the primary parameters in the following table.
| Parameter | Symbol | Unit | Description |
|---|---|---|---|
| Nominal capacity | Cₙ | Ah | Total charge stored under specified discharge conditions |
| Open-circuit voltage | U_oc | V | Terminal voltage at zero current and equilibrium |
| Terminal voltage | U | V | Actual voltage across battery terminals under load |
| Internal resistance | R₀ | Ω | Ohmic resistance arising from electrodes, electrolyte, and contacts |
| Polarization resistance | R_p | Ω | Resistance caused by electrochemical polarization effects |
| State of charge | SOC | % | Ratio of remaining capacity to nominal capacity |
| Depth of discharge | DOD | % | Ratio of discharged capacity to nominal capacity |
| State of health | SOH | % | Ratio of current maximum capacity to initial rated capacity |
2.3 Common Fault Types and Characterization
Based on extensive failure analysis of the EV battery pack, I focus on five representative fault categories in this dissertation: overcharge fault, over-discharge fault, internal resistance increase fault, capacity degradation fault, and thermal runaway fault. Each fault exhibits distinct observable signatures in voltage, current, temperature, and derived quantities such as voltage change rate and temperature change rate. The following table summarizes the fault types and their diagnostic criteria.
| Fault type | Observable symptoms | Diagnostic criterion |
|---|---|---|
| Overcharge | Charging voltage exceeds rated cutoff; longer charging duration; faster discharge rate | Charging voltage exceeds 4.2 V threshold after reaching rated value |
| Over-discharge | Discharge voltage drops below cutoff; extended discharge duration | Discharge voltage falls below 2.9996 V threshold |
| Internal resistance increase | Rapid voltage drop under load; accelerated discharge; higher temperature | Internal resistance value significantly exceeds normal range |
| Capacity degradation | Shorter charge and discharge durations; reduced usable energy | Actual capacity falls below rated capacity with significant deviation |
| Thermal runaway | Rapid temperature rise; voltage surge; eventual voltage collapse | Temperature exceeds 50°C with abnormal voltage/current derivative |
Understanding the fundamental principles and failure mechanisms of lithium-ion batteries provides the theoretical basis for developing effective diagnostic algorithms. In the following sections, I describe the experimental platform, data acquisition process, equivalent circuit modeling, and the proposed LSTM-BP neural network diagnostic method.
3. Experimental Platform and Data Acquisition
3.1 Battery Selection and Experimental Setup
To acquire realistic charge-discharge cycling data for the EV battery pack, I constructed a dedicated experimental platform at the battery testing laboratory. A ternary lithium-ion battery with a nominal capacity of 21.673 Ah and a nominal voltage of 4.2 V was selected as the test specimen. The discharge cutoff voltage was set at 3.0 V, and the operating temperature range was -20°C to 50°C. The experimental setup comprised the following components:
- Battery performance tester: JFBTS model, with discharge voltage range of 2-100% full scale and resolution of 0.02%, providing ±(0.05%FS + 0.05%RD) accuracy.
- Temperature and humidity chamber: HTH80-40A model, capable of regulating temperature from -40°C to 150°C and humidity from 20% to 98% RH.
- Host computer: equipped with data acquisition software to record and export battery data in Excel format.
The battery cycling experiments were conducted over a period from April 2022 to May 2023, encompassing 3,565 charge-discharge cycles. The battery capacity gradually decayed from 21.673 Ah to 17.057 Ah over this period. A total of 596,703 data points were recorded during the experiments. Each charge-discharge cycle consisted of five stages: constant current charging (CCC), constant voltage charging (CVC), rest (R), constant current discharge (CCD), and final rest (R). The charging current was set to 21.003 A, and the charging voltage was limited to 4.2 V.
3.2 Data Preprocessing
To ensure data quality and consistency, I performed comprehensive preprocessing on the raw experimental data. Since battery capacity degrades progressively with cycling, I selected data from the first 1,775 cycles, corresponding to 303,349 data points, during which the battery capacity remained above 86.7% of its rated value. The following preprocessing steps were implemented:
First, the depth of discharge (DOD) was calculated from SOC measurements using:
$$ \text{DOD} = 1 – \text{SOC} $$
Second, voltage change rate and temperature change rate were computed using:
$$ \dot{V}_t = \frac{V_{T+1} – V_T}{\Delta T}, \quad \dot{T}_t = \frac{T_{T+1} – T_T}{\Delta T} $$
Third, the relationship between open-circuit voltage (OCV) and SOC was established through polynomial fitting. A ninth-degree polynomial was selected due to its superior fitting accuracy:
$$ U_{oc}(SOC) = k_0 + k_1 SOC + k_2 SOC^2 + k_3 SOC^3 + k_4 SOC^4 + k_5 SOC^5 + k_6 SOC^6 + k_7 SOC^7 + k_8 SOC^8 + k_9 SOC^9 $$
The fitted coefficients for both charging and discharging states are presented in the following table.
| Coefficient | Charging state | Discharging state |
|---|---|---|
| k₀ | 3.376215 | 2.797240 |
| k₁ | 4.438361 | 12.932791 |
| k₂ | -51.256807 | -147.982115 |
| k₃ | 400.245979 | 949.392946 |
| k₄ | -1891.500436 | -3602.984507 |
| k₅ | 5410.129756 | 8434.967743 |
| k₆ | -9245.370676 | -12300.356520 |
| k₇ | 9186.216501 | 10884.024795 |
| k₈ | -4891.018942 | 5348.501875 |
| k₉ | 1078.947329 | 1119.765249 |
The fitting results demonstrated that the maximum OCV error was less than 0.01 V when SOC ranged from 0.1 to 1.0, corresponding to a fitting error rate below 0.238%. This level of accuracy satisfied the requirements for subsequent fault diagnosis analysis. After preprocessing, the final lithium battery dataset contained the following features: current, voltage, temperature, SOC, capacity, internal resistance, voltage change rate, temperature change rate, and OCV.
4. Equivalent Circuit Modeling and Fault Simulation
4.1 Comparison of Equivalent Circuit Models
Equivalent circuit models are mathematical representations that simulate the electrical behavior of the EV battery pack using resistors, capacitors, and voltage sources. I evaluated several commonly used models: the Rint model, Thevenin model, PNGV model, and second-order RC model. The following table compares their structural complexity and accuracy characteristics.
| Model | Components | Advantages | Disadvantages |
|---|---|---|---|
| Rint | Voltage source + series resistor | Simple, low computational cost | Poor accuracy, ignores polarization |
| Thevenin | Voltage source + resistor + RC network | Moderate accuracy, captures polarization | Limited to single time-constant dynamics |
| PNGV | Voltage source + capacitor + RC network | Captures OCV variation with load | Requires additional parameters |
| Second-order RC | Voltage source + resistor + two RC networks | High accuracy, captures both short-term and long-term dynamics | More complex, higher computational cost |
Considering the trade-off between accuracy and computational complexity, I selected the second-order RC equivalent circuit model for fault simulation of overcharge, over-discharge, internal resistance increase, and capacity degradation faults.
4.2 Second-Order RC Equivalent Circuit Model
The second-order RC equivalent circuit model employs two parallel RC networks to represent the short-term and long-term polarization effects within the battery. The circuit topology consists of an ideal voltage source representing the OCV, a series ohmic resistance R₀, and two RC networks (R₁C₁ and R₂C₂) connected in series, as illustrated in figure. The terminal voltage and state-space equations for this model are expressed as:
$$ U(t) = U_{oc}(SOC(t)) – I(t)R_0 – U_1(t) – U_2(t) $$
$$ \frac{dU_1(t)}{dt} = -\frac{U_1(t)}{R_1 C_1} + \frac{I(t)}{C_1} $$
$$ \frac{dU_2(t)}{dt} = -\frac{U_2(t)}{R_2 C_2} + \frac{I(t)}{C_2} $$
The discrete-time representation of these equations, used for digital implementation, is given by:
$$ U(k+1) = U_{oc}(k) – I(k)R_0 – U_1(k) – U_2(k) $$
$$ U_1(k+1) = U_1(k)e^{-\Delta T/(R_1 C_1)} + I(k)R_1(1 – e^{-\Delta T/(R_1 C_1)}) $$
$$ U_2(k+1) = U_2(k)e^{-\Delta T/(R_2 C_2)} + I(k)R_2(1 – e^{-\Delta T/(R_2 C_2)}) $$
4.3 Electro-Thermal Coupling Model
To accurately simulate thermal runaway faults, I employed an electro-thermal coupling model that combines the equivalent circuit model with a thermal dynamics model. The thermal behavior of the battery is governed by:
$$ \rho C_p \frac{\partial T}{\partial t} = \nabla \cdot (\lambda \nabla T) + q_{gen} $$
where the heat generation rate q_gen is composed of irreversible Joule heating and reversible entropy change:
$$ q_{gen} = I(E – U) + IT \frac{\partial E}{\partial T} $$
This coupled model enables the prediction of temperature evolution under various operating conditions, which is essential for detecting thermal runaway in the EV battery pack.
4.4 Fault Simulation Results
Using the second-order RC equivalent circuit model and the electro-thermal coupling model, I simulated five fault scenarios based on the experimental dataset. The simulation results revealed distinct voltage and temperature signatures for each fault type, enabling the establishment of fault sample databases. The following table summarizes the fault simulation observations.
| Fault type | Charging observations | Discharging observations |
|---|---|---|
| Overcharge | Voltage continues rising above 4.2 V; extended charging time | Faster discharge rate; reaches normal cutoff |
| Over-discharge | No significant difference from normal | Voltage drops below 2.9996 V; extended discharge time |
| Internal resistance increase | Higher charging cutoff voltage; faster voltage rise | Rapid voltage decline; lower discharge cutoff |
| Capacity degradation | Shorter charging time; faster charge rate | Shorter discharge time; faster discharge |
| Thermal runaway | Temperature exceeds 50°C; voltage surges | Voltage collapses to 0 V; rapid temperature rise |
Based on the fault simulation results, I integrated the voltage and temperature data to construct a comprehensive fault sample database for the EV battery pack, which served as the training and validation basis for the neural network diagnostic model.
5. LSTM-BP Neural Network Fault Diagnosis Method
5.1 Neural Network Fundamentals
Artificial neural networks are computational models inspired by biological neural systems, capable of learning complex input-output relationships through iterative training. The fundamental processing unit of a neural network is the neuron, which computes a weighted sum of its inputs and applies an activation function to produce the output:
$$ y = f\left(\sum_{i=1}^{n} w_i x_i – \theta\right) $$
where x_i represents the input values, w_i denotes the corresponding weights, θ is the threshold, and f indicates the activation function. Common activation functions include the sigmoid function, hyperbolic tangent (tanh), and rectified linear unit (ReLU).
5.2 Back-Propagation Neural Network
The BP neural network is a multi-layer feedforward network that learns through error back-propagation. It consists of an input layer, one or more hidden layers, and an output layer. During forward propagation, the hidden layer output is computed as:
$$ z_j = \sum_{i=1}^{n} w_{ij} x_i + b_j $$
where w_ij is the weight connecting input neuron i to hidden neuron j, and b_j is the bias term. The final output is obtained through the output layer:
$$ y_k = \sum_{j} w_{jk} a_j + b_k $$
where a_j is the activation output of hidden neuron j, and w_jk is the weight connecting hidden neuron j to output neuron k. The BP algorithm adjusts the weights and biases by minimizing the loss function through gradient descent.
While BP networks possess strong nonlinear mapping capabilities, they struggle to capture temporal dependencies in time-series data and may converge to local minima.
5.3 Long Short-Term Memory Network
The LSTM network, a special variant of recurrent neural networks (RNN), addresses the vanishing gradient problem through a sophisticated gating mechanism. Each LSTM cell contains three gates: the forget gate, input gate, and output gate, which control the flow of information. The gating operations are mathematically expressed as follows:
$$ f_t = \sigma(W_f \cdot [h_{t-1}, x_t] + b_f) $$
$$ i_t = \sigma(W_i \cdot [h_{t-1}, x_t] + b_i) $$
$$ \tilde{C}_t = \tanh(W_C \cdot [h_{t-1}, x_t] + b_C) $$
$$ C_t = f_t \odot C_{t-1} + i_t \odot \tilde{C}_t $$
$$ o_t = \sigma(W_o \cdot [h_{t-1}, x_t] + b_o) $$
$$ h_t = o_t \odot \tanh(C_t) $$
where f_t, i_t, o_t denote the forget gate, input gate, and output gate activation values, respectively; C_t represents the cell state; h_t is the hidden state; and σ indicates the sigmoid activation function.
Although LSTM networks excel at sequence learning, they may overfit when applied to relatively small datasets and may not fully exploit spatial or global features present in non-sequential inputs.
5.4 Architecture of the Proposed LSTM-BP Model
To leverage the complementary strengths of LSTM and BP networks while mitigating their individual weaknesses, I proposed the LSTM-BP hybrid model. The overall architecture integrates an LSTM network as the frontend feature extractor for temporal dependencies, followed by a BP network as the backend nonlinear regressor for final state prediction and classification. The model architecture is structured as follows:
In the LSTM stage, the preprocessed battery data (current, voltage, temperature, and SOC) are normalized and segmented into training and validation sets. The normalized sequence data are fed into the LSTM hidden layers, where the gating mechanisms capture temporal patterns across charge-discharge cycles. The output sequence from the LSTM layers is then passed to the BP stage.
In the BP stage, the input layer dimension is matched to the LSTM hidden layer output dimension. The BP network processes the extracted features through its input layer, hidden layers, and output layer via forward propagation, ultimately generating the predicted battery state parameters (capacity, OCV, internal resistance, voltage change rate, and temperature change rate).
The training process employs back-propagation to compute losses and updates the weights and biases of both the LSTM and BP components through an optimization algorithm, thereby minimizing prediction errors and improving model accuracy.
5.5 Model Parameter Selection
Selecting optimal hyperparameters is critical to achieving high diagnostic accuracy while avoiding underfitting or overfitting. I employed a cross-validation strategy to systematically explore 16 different parameter configurations for the LSTM-BP model. The mean absolute error (MAE) was used as the evaluation metric for parameter selection:
$$ \text{MAE} = \frac{1}{n} \sum_{i=1}^{n} |y_i – \hat{y}_i| $$
The results of the parameter combinations are summarized in the following table.
| No. | LSTM layers | LSTM neurons | BP layers | BP neurons | MAE |
|---|---|---|---|---|---|
| 1 | 2 | (256,128) | 2 | (256,128) | 0.0044 |
| 2 | 2 | (256,128) | 2 | (256,64) | 0.0059 |
| 3 | 2 | (256,128) | 2 | (128,64) | 0.0071 |
| 4 | 2 | (256,64) | 2 | (256,128) | 0.0064 |
| 5 | 2 | (256,64) | 2 | (256,64) | 0.0088 |
| 6 | 2 | (256,64) | 2 | (128,64) | 0.0103 |
| 7 | 2 | (128,64) | 2 | (256,128) | 0.0079 |
| 8 | 2 | (128,64) | 2 | (256,64) | 0.0096 |
| 9 | 2 | (128,64) | 2 | (128,64) | 0.0124 |
| 10 | 2 | (256,128) | 3 | (256,128,64) | 0.0041 |
| 11 | 2 | (256,64) | 3 | (256,128,64) | 0.0059 |
| 12 | 2 | (128,64) | 3 | (256,128,64) | 0.0077 |
| 13 | 3 | (256,128,64) | 2 | (256,128) | 0.0019 |
| 14 | 3 | (256,128,64) | 2 | (256,64) | 0.0031 |
| 15 | 3 | (256,128,64) | 2 | (128,64) | 0.0047 |
| 16 | 3 | (256,128,64) | 3 | (256,128,64) | 0.0055 |
Based on these results, the optimal configuration was identified as three LSTM layers with (256,128,64) neurons and two BP layers with (256,128) neurons, yielding the lowest MAE of 0.0019. I adopted this configuration as the final model architecture.
Additionally, I employed multiple optimization strategies to enhance the model’s generalization capability. The dataset was divided into training, test, and validation sets in a 7:2:1 ratio. The Adam optimizer was selected for its adaptive learning rate and momentum characteristics. Dropout with a rate of 0.2 and batch normalization techniques were incorporated to prevent overfitting and improve training efficiency. The final parameter settings are summarized in the following table.
| Parameter | Value |
|---|---|
| Input sequence length | 1 |
| Batch size | 32 |
| Learning rate | 0.001 |
| Dropout rate | 0.2 |
| Epochs | 200 |
| LSTM layers | 3 |
| LSTM neurons | (256,128,64) |
| BP layers | 2 |
| BP neurons | (256,128) |
| Optimizer | Adam |
| Activation function | ReLU |
5.6 Overall Fault Diagnosis Workflow
The proposed LSTM-BP fault diagnosis framework consists of two main stages: a model training stage and a diagnostic inference stage. In the training stage, the experimental data are processed and split into training, validation, and test sets. The LSTM-BP model is trained using the training set, hyperparameters are tuned based on the validation set, and model performance is evaluated using the test set.
In the inference stage, a given battery sample is preprocessed, and the current, voltage, temperature, and SOC data are fed as inputs to the trained LSTM-BP model. The model predicts the battery’s output characteristics, including capacity, internal resistance, voltage change rate, and temperature change rate. These predicted values are then compared against diagnostic threshold criteria to determine whether the battery is in a fault state, and if so, to identify the specific fault type. Batteries exceeding the threshold are flagged as faulty, while those within the threshold are classified as normal.
6. Fault Diagnosis Data Analysis and Result Verification
6.1 Single-Cycle Fault Data Analysis
To establish reliable diagnostic thresholds, I conducted detailed single-cycle charge-discharge data comparisons between normal and faulty battery samples for each fault type using the constructed datasets and the LSTM-BP model predictions.
For the overcharge fault analysis, the charging voltage of the faulty battery continued to increase beyond the rated 4.2 V cutoff voltage, eventually reaching approximately 4.4 V. The charging time was extended by 710 seconds (13% increase) compared to the normal battery. The discharge phase showed a faster voltage decay rate for the overcharged battery. The predicted characteristics for the overcharge fault sample were: capacity of 21.664 Ah, internal resistance of 0.1547176 Ω, charging cutoff time of 6,170 seconds, charging cutoff voltage of 4.4 V, and discharge cutoff voltage of 2.9996 V.
For the over-discharge fault analysis, the faulty battery exhibited no significant differences during charging but continued discharging below the rated cutoff voltage of 2.9996 V, reaching a final discharge voltage of 2.8001 V—a 6.65% reduction. The predicted characteristics were: capacity of 21.667 Ah, internal resistance of 0.1558343 Ω, charging cutoff time of 5,460 seconds, charging cutoff voltage of 4.2 V, and discharge cutoff voltage of 2.8001 V.
For the internal resistance increase fault, the faulty battery showed a slower charging rate initially but reached the rated cutoff voltage, followed by a faster discharge rate and a lower discharge cutoff voltage of 2.902 V. The internal resistance increased by 13.82% to 0.1765217 Ω, and the charging time increased by 18.74% to 6,483 seconds.
For the capacity degradation fault, the faulty battery exhibited slower charging and faster discharging compared to normal, with a predicted capacity reduction of 8.84% to 19.757 Ah. The charging time increased by 11.53% to 6,090 seconds, while voltage thresholds remained unchanged.
For the thermal runaway fault, the analysis revealed distinct temperature stages: normal phase (20-50°C), self-heating phase (50-120°C), thermal accumulation phase (120-200°C), and thermal runaway phase (above 200°C). During the self-heating phase, voltage and temperature began to rise; during the thermal accumulation phase, voltage and voltage change rate increased sharply; and during the thermal runaway phase, the voltage instantaneously dropped to 0 V with rapid temperature escalation.
6.2 Diagnostic Threshold Determination
Based on the fault data analyses and the labeled threshold nodes in the voltage comparison charts, I established six diagnostic thresholds for fault classification. The charging cutoff voltage threshold was set at 4.2 V, the discharging cutoff voltage threshold at 2.9996 V, the voltage change rate threshold at ±0.003 V/s, the temperature change rate threshold at 0.009°C/s, the maximum temperature threshold at 50°C, and the charging cutoff time threshold at 5,460 seconds. These thresholds form the quantitative basis for the diagnostic decision logic.
| Diagnostic feature | Threshold value |
|---|---|
| Charging cutoff voltage | 4.2 V |
| Discharge cutoff voltage | 2.9996 V |
| Voltage change rate | ±0.003 V/s |
| Temperature change rate | 0.009°C/s |
| Maximum temperature | 50°C |
| Charging cutoff time | 5,460 s |
6.3 Diagnostic Flowchart
The complete fault diagnosis result flowchart integrating the trained LSTM-BP model with the diagnostic threshold logic is designed as follows. The battery sample data are first preprocessed, and the extracted current, voltage, temperature, and SOC features are fed into the LSTM-BP model. The model predicts the capacity, open-circuit voltage, internal resistance, voltage change rate, and temperature change rate. Subsequently, the charging cutoff voltage, discharge cutoff voltage, maximum temperature, and charging cutoff time are recorded as diagnostic characteristics. These characteristics are then evaluated against the predetermined thresholds using the decision logic to classify the battery state as normal or faulty.
6.4 Model Comparison and Performance Evaluation
To validate the effectiveness of the proposed LSTM-BP model, I conducted both vertical and horizontal comparisons. In the vertical comparison, the LSTM-BP model was compared with the standalone BP and LSTM models. In the horizontal comparison, the LSTM-BP model was benchmarked against advanced architectures including RNN, GRU, and CNN-LSTM models. I used three evaluation metrics to assess prediction performance: mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE). These metrics are defined as:
$$ \text{MSE} = \frac{1}{n} \sum_{i=1}^{n} (y_i – \hat{y}_i)^2 $$
$$ \text{RMSE} = \sqrt{\frac{1}{n} \sum_{i=1}^{n} (y_i – \hat{y}_i)^2} $$
$$ \text{MAE} = \frac{1}{n} \sum_{i=1}^{n} |y_i – \hat{y}_i| $$
The prediction results for battery capacity, OCV, internal resistance, temperature change rate, and voltage change rate were evaluated. The comparison results are summarized in the following table.
| Model | MSE | RMSE | MAE |
|---|---|---|---|
| BP | 0.02154 | 0.1468 | 0.1643 |
| LSTM | 0.00976 | 0.0988 | 0.0872 |
| RNN | 0.02779 | 0.1667 | 0.1561 |
| GRU | 0.02419 | 0.1555 | 0.1179 |
| CNN-LSTM | 0.01032 | 0.1016 | 0.0835 |
| LSTM-BP | 0.00231 | 0.0481 | 0.0313 |
From the table, it is evident that the LSTM-BP model achieved the lowest values across all three evaluation metrics, confirming its superior predictive accuracy and reliability for EV battery pack fault diagnosis.
6.5 Diagnostic Verification Using Confusion Matrix
To further verify the diagnostic accuracy of the proposed method, I randomly selected a total of 600 battery samples from the test set and fault sample database, with 100 samples from each category: normal, overcharge, over-discharge, internal resistance increase, capacity degradation, and thermal runaway. These samples were evaluated using the LSTM model, CNN-LSTM model, and the proposed LSTM-BP model. Confusion matrices were constructed to visualize classification results, and the overall accuracy was calculated as:
$$ \text{Accuracy} = \frac{TP + TN}{TP + TN + FP + FN} $$
The diagnostic results are summarized in the following table.
| Model | Misclassified samples | Diagnostic accuracy |
|---|---|---|
| LSTM | 43 | 92.83% |
| CNN-LSTM | 34 | 94.33% |
| LSTM-BP | 14 | 97.67% |
The LSTM-BP model achieved the highest diagnostic accuracy of 97.67% with only 14 misclassified samples out of 600, outperforming both the LSTM model (92.83%) and the CNN-LSTM model (94.33%). These results conclusively demonstrate the effectiveness and superiority of the proposed LSTM-BP fault diagnosis method for EV battery pack applications.
7. Conclusion
In this dissertation, I proposed a data-driven fault diagnosis method for the EV battery pack based on a hybrid LSTM-BP neural network. The main contributions of this research are summarized as follows:
First, I established a comprehensive experimental platform for lithium battery charge-discharge cycling and collected extensive battery data. Through rigorous preprocessing, including SOC-OCV polynomial fitting, voltage and temperature change rate computation, and dataset construction, I created a high-quality lithium battery dataset suitable for machine learning applications.
Second, I performed systematic fault simulation for five critical fault types—overcharge, over-discharge, internal resistance increase, capacity degradation, and thermal runaway—using the second-order RC equivalent circuit model and the electro-thermal coupling model. The simulation results were analyzed and integrated into a fault sample database that provides the foundation for data-driven diagnostic model development.
Third, I proposed the LSTM-BP hybrid neural network architecture, which combines the temporal sequence learning capability of LSTM networks with the powerful nonlinear mapping ability of BP networks. Through cross-validation and multiple optimization strategies, I determined the optimal model hyperparameters. Comprehensive comparisons with BP, LSTM, RNN, GRU, and CNN-LSTM models demonstrated that the LSTM-BP model achieved superior prediction accuracy and reliability.
Fourth, I established quantitative diagnostic threshold criteria based on fault data analysis and labeled threshold nodes. The overall fault diagnosis workflow, integrating the LSTM-BP model with threshold-based decision logic, was designed and validated using 600 randomly selected battery samples. The proposed method achieved a diagnostic accuracy of 97.67%, significantly outperforming other neural network models.
While the proposed method demonstrates promising results for single-fault diagnosis, future research directions include addressing multiple simultaneous faults in the EV battery pack and improving the temporal resolution of thermal runaway detection to the millisecond level for faster and more precise fault identification. Overall, the LSTM-BP-based fault diagnosis framework presented in this work provides a reliable and effective solution for enhancing the safety and reliability of the EV battery pack in electric vehicles.