Electric vehicles are increasingly recognized as one of the most effective ways to reduce carbon emissions and achieve energy transition. Yet the safety and durability of an electric vehicle strongly depend on its rechargeable energy storage devices. In my work, I focus on the early detection and prediction of failures in these battery systems using real operational data and advanced machine-learning architectures. The central challenge is that battery failure is often slow and hidden; it emerges under complex interactions among thermal stress, chemical aging, and dynamic load conditions. By modeling the temporal evolution of sensor streams collected inside an electric vehicle, I can estimate the remaining useful life of the onboard energy storage system before a severe fault occurs.
I was motivated by the limitations of conventional fault alarms, which usually trigger only after a threshold is exceeded. Such alarms are often too late for preventive action. Instead, a data-driven, continuous remaining-useful-life estimate can give maintenance teams more time to respond. In this thesis, I present a hybrid architecture combining a Kolmogorov-Arnold network with a long short-term memory network, called KAN-LSTM, to achieve accurate electric vehicle battery failure prediction. The model was tested on real electric vehicle trip records captured during normal operation and eventual failure. My experiments show that the KAN-LSTM model provides a substantial improvement over traditional machine-learning baselines and single deep-learning models.
Background and Fault-Prognosis Principles
Rechargeable energy storage devices in electric vehicles are complex electrochemical systems. Their degradation can be roughly divided into two categories. The first category is gradual degradation, which manifests as a sustained change in sensor measurements such as voltage, current, capacity, and internal resistance. The second category is sudden failure, which is difficult to anticipate because the monitored values may look normal until the failure moment. The prediction method studied in this thesis is mainly designed for gradually increasing risk, but it also produces a health indicator that can warn users before an unexpected breakdown.
In the engineering literature, time-series prediction and remaining-useful-life prediction are the two main ways to forecast device failures. Time-series prediction extrapolates future values of selected signals. If the extrapolated signal crosses a pre-defined safety threshold, an alarm can be issued. However, an electric vehicle can have many correlated sensors, and the relationship between those sensors and failure is not simple. Remaining-useful-life prediction is more direct because the model is trained to map sensor sequences to a scalar that is the time left before failure. This formulation naturally supports supervised learning and can integrate multiple sensor inputs.
Sliding-average filtering is also useful before model construction. A sliding average smooths the original sequence and removes high-frequency noise while preserving long-term trends. In my work, I used this technique to help determine the length of the input window and to reduce spurious fluctuations in the battery signals. The moving-average value at time \(t\) for window size \(n\) is defined as:
$$
\bar{x}_t = \frac{1}{n} \sum_{i=t-n+1}^{t} x_i
$$
I estimated the appropriate window size by minimizing the total distance between the moving-average series and its global mean. The resulting window length was then used to construct training samples for the prediction model.
Machine-Learning Methods Used in the Study
Support Vector Regression and Ensemble Methods
Support vector regression, often abbreviated as SVR, is a classical machine-learning model for regression. SVR looks for a hyperplane that has the largest feasible margin while keeping most prediction errors inside an insensitive tube of width \(\varepsilon\). The optimization can be written as:
$$
\min_{w,b,\xi_i,\xi_i^*}\;\frac{1}{2}\|w\|^2 + C\sum_{i=1}^{n}(\xi_i+\xi_i^*)
$$
$$
\text{subject to } y_i – w^\top x_i – b \le \varepsilon+\xi_i,\quad w^\top x_i + b – y_i \le \varepsilon+\xi_i^*,\quad \xi_i,\xi_i^*\ge 0
$$
In my electric vehicle battery experiments, I used a polynomial kernel because the capacity-degradation trend is strongly nonlinear and the historical trajectories were not strictly linear. The hyperparameters of SVR were tuned with grid search.
I also evaluated XGBoost, a powerful gradient-boosted decision-tree algorithm. It treats a prediction as a sum of decision trees and minimizes a regularized objective. At the \(t\)-th boosting iteration, the objective can be expressed as:
$$
L^{(t)} = \sum_{i=1}^{n}\left[g_i f_t(x_i) + \frac{1}{2}h_i f_t^2(x_i)\right] + \Omega(f_t)
$$
$$
\Omega(f_t) = \gamma T + \frac{1}{2}\lambda\sum_{j=1}^{T} w_j^2
$$
Here \(g_i\) and \(h_i\) are the first and second derivatives of the loss function, \(T\) is the number of leaves, and \(w_j\) are leaf weights. The regularization terms help prevent overfitting and improve generalization.
LightGBM is another tree-based ensemble algorithm. It is particularly efficient because it builds histograms of feature values and uses leaf-wise growth. For every decision tree \(h_m\), a new model is obtained as:
$$
F_m(x)=F_{m-1}(x)+\eta\,h_m(x)
$$
Each new tree is fitted to the negative gradient of the loss function so that the ensemble gradually reduces the residual errors from previous iterations. I included LightGBM because it often achieves lower error than XGBoost on high-dimensional data while being faster to train.
Long Short-Term Memory Networks
Long short-term memory networks are a special type of recurrent neural network introduced to overcome the vanishing-gradient problem in standard recurrent networks. Each LSTM unit keeps an internal cell state and uses three gates to update it. The input gate, forget gate, and output gate determine which information should be kept or discarded. The LSTM equations can be written as:
$$
f_t = \sigma(W_f [h_{t-1}, x_t] + b_f)
$$
$$
i_t = \sigma(W_i [h_{t-1}, x_t] + b_i)
$$
$$
\tilde{C}_t = \tanh(W_C [h_{t-1}, x_t] + b_C)
$$
$$
C_t = f_t \odot C_{t-1} + i_t \odot \tilde{C}_t
$$
$$
o_t = \sigma(W_o [h_{t-1}, x_t] + b_o)
$$
$$
h_t = o_t \odot \tanh(C_t)
$$
Because battery aging is a sequential process, LSTM can model the historical dependence of voltage, current, and temperature in an electric vehicle. My LSTM baseline was used as a strong reference to judge whether the hybrid KAN-LSTM architecture could lead to further improvements.
Kolmogorov-Arnold Network
A Kolmogorov-Arnold network is a recently developed neural-network architecture inspired by the Kolmogorov-Arnold representation theorem. The theorem states that a multivariate continuous function can be expressed using sums of univariate functions. KAN replaces the usual linear weights on network edges with learnable univariate functions. A KAN layer can be expressed as:
$$
\mathrm{KAN}(x) = \sum_{q=1}^{2n+1} \Psi_q\left(\sum_{i=1}^{n} \phi_{q,i}(x_i)\right)
$$
The learnable functions \(\phi_{q,i}\) are parametrized as B-spline curves. In practice, I used a residual representation:
$$
\phi(x) = \omega_b\,b(x) + \omega_s\,\mathrm{spline}(x)
$$
where \(b(x)\) is a smooth base function and \(\mathrm{spline}(x)\) is a linear combination of B-spline basis functions. Because the spline knots are trainable, the network is able to approximate very sharp or very smooth changes in the data. For an electric vehicle battery, such sharp changes can occur near the end of life when the internal resistance suddenly increases or when temperature rises quickly during fast charging. KAN is especially useful in this area because it provides more flexible nonlinear feature extraction than fixed activation functions.
KAN-LSTM Hybrid Model
I designed a hybrid model that exploits the complementary strengths of both KAN and LSTM. The KAN module is placed at the front end to transform raw sensor measurements into a richer, high-dimensional representation. The B-spline functions in KAN can separate local nonlinear effects from the global sensor trends. Then the LSTM module receives this representation and models its temporal dynamics through the gate mechanism. The result is a full KAN-LSTM architecture that can reproduce both abrupt abnormal transitions and slow degradation patterns.
The model can be summarized in two steps. First, the extracted feature vector at time \(t\) is produced by KAN:
$$
z_t = \Phi_{\mathrm{KAN}}(x_t)
$$
Then, an LSTM layer processes the sequence \(z_t\) over the sliding window and maps it to the remaining useful life:
$$
\hat{y}_t = W_o\, h_t + b_o
$$
I trained the model by using a multi-stage procedure. In the first stage, I pre-trained the KAN module while keeping the LSTM parameters fixed. In the second stage, I fixed the KAN weights and fine-tuned the LSTM parameters. Finally, the whole KAN-LSTM network was trained end to end. This procedure allows each part of the electric vehicle fault prediction model to specialize: KAN extracts local nonlinear features, while LSTM learns how those features evolve over time.
Data Description and Preprocessing
For this research, I used real operating data from two different electric vehicle models. They were monitored through a large-scale municipal platform for new-energy intelligent vehicles. The data covered a period from normal driving to the moment of failure. The original sensor sampling frequency was one sample every 20 seconds. I collected 7,023 time steps before the failure and obtained 27 sensor variables. The sensor variables include battery voltage, current, state of charge, temperature, odometer, vehicle state, alarm level, and several other status indicators. Table below summarizes the main sensor characteristics used in the initial data matrix.
| Feature ID | Short Name | Description |
|---|---|---|
| 1 | HAL | Highest alarm level |
| 2 | SPD | Vehicle speed |
| 3 | VST | Vehicle state |
| 4 | CHS | Charging state |
| 5 | OPM | Operation mode |
| 6 | ODM | Odometer reading |
| 7 | TOT | Total voltage |
| 8 | TCI | Total current |
| 9 | SOC | State of charge |
| 10 | DCS | DC-DC converter state |
| 11 | GEA | Current gear |
| 12 | GDF | Gear driving force |
| 13 | GBF | Gear braking force |
| 14 | INS | Insulation resistance |
| 15 | HVB | Subsystem number of highest cell voltage |
| 16 | HVN | Cell number of highest cell voltage |
| 17 | HVM | Maximum cell voltage |
| 18 | LVB | Subsystem number of lowest cell voltage |
| 19 | LVN | Cell number of lowest cell voltage |
| 20 | LVM | Minimum cell voltage |
| 21 | HTS | Subsystem number of highest temperature |
| 22 | HTP | Probe number of highest temperature |
| 23 | HTV | Maximum temperature |
| 24 | LTS | Subsystem number of lowest temperature |
| 25 | LTP | Probe number of lowest temperature |
| 26 | LTC | Minimum temperature |
| 27 | UAF | Universal alarm flag |
Before modeling, I inspected the data for missing values and format problems. The original dataset was complete enough that no imputation was necessary. Because many categorical variables had already been encoded as integers according to communication standards, no additional label encoding was needed.

Normalization
Feature magnitudes vary greatly. For example, total voltage can be several hundred volts, while insulation resistance might be thousands of kilohms. If I had used raw values, the model would have been dominated by features with larger ranges. I therefore applied Min-Max normalization to scale every variable into the interval \([-1,1]\). The normalized value \(x’\) is defined as:
$$
x’ = \frac{x – x_{\mathrm{mean}}}{x_{\max} – x_{\min}}
$$
In this formula, \(x\) is the original sensor value, \(x_{\max}\) and \(x_{\min}\) are the maximum and minimum values, and \(x_{\mathrm{mean}}\) is the mean. This operation preserves the relative ordering of each sensor and removes the effect of different measurement units.
Feature Selection
Not all 27 variables are useful for remaining-useful-life prediction. Some features contain too little variance, while others are strongly correlated with a primary feature and therefore redundant. I used a two-stage feature-selection strategy to identify the most informative subset.
In the first stage, I computed Spearman correlation between each feature and a temporary target variable. Spearman correlation measures monotonic associations without assuming linearity. For two variables \(X\) and \(Y\), after replacing each value with its rank, Spearman’s coefficient is:
$$
\rho_s = \frac{\sum_{i=1}^{n}(x_i-\bar{x})(y_i-\bar{y})}{\sqrt{\sum_{i=1}^{n}(x_i-\bar{x})^2 \sum_{i=1}^{n}(y_i-\bar{y})^2}}
$$
In the second stage, I ran a random forest regressor to compute feature-importance scores. Random forest is an ensemble of decision trees trained on bootstrap samples. Feature importance is measured by how often a feature is selected for splitting and by how much the model error increases if that feature is randomly shuffled. I combined the Spearman ranking and the random-forest ranking to remove variables with low importance and low correlation. After this evaluation, I discarded the following nine variables: total current, highest-temperature subsystem number, highest-temperature probe number, highest-voltage battery subsystem number, highest-voltage battery cell number, lowest-temperature subsystem number, lowest-temperature probe subsystem number, lowest-voltage battery cell number, and lowest-voltage battery subsystem number.
The remaining 18 variables were used to train the final electric vehicle fault prediction models. These selected features include odometer, state of charge, gear position, minimum cell voltage, operation mode, minimum temperature, vehicle speed, charging state, vehicle state, alarm level, DC-DC state, braking force, insulation resistance, drive force, universal alarm flag, maximum cell voltage, total voltage, and maximum temperature.
Data Split and RUL Labels
I divided the data into four subsets to prevent data leakage and to make sure that the model is evaluated on a vehicle not used during training. The first two subsets are anomaly-detection modules. The third subset is the training set of the prediction model. The fourth subset is the testing set. The table below shows the chronological division.
| Subset Name | EV Model Source | Time Step Range | Purpose |
|---|---|---|---|
| EDSM-A | Model A | 1-5000 | Anomaly detection calibration |
| EDSM-B | Model B | 1-5000 | Anomaly detection calibration |
| PSM-train | Model A | 5001-7023 | Supervised model training |
| PSM-test | Model B | 5001-7023 | Model evaluation |
Because the goal is to predict remaining useful life before failure, I need a target value for every input sample. I created a piecewise-linear RUL label. For the normal phase before the first abnormal event, the RUL value is set to a constant time-to-detection value. After the anomaly is detected, the RUL decreases linearly until the final failure. Mathematically, the label can be written as:
$$
y_{\mathrm{RUL}}(t) =
\begin{cases}
T_f – t_{\mathrm{det}}, & t \le t_{\mathrm{det}} \\
T_f – t, & t > t_{\mathrm{det}}
\end{cases}
$$
where \(t_{\mathrm{det}}\) denotes the first anomalous time step, and \(T_f\) is the time of final failure. This design stabilizes the beginning of the training signal while retaining a sharp downward trend in the aging phase.
To determine \(t_{\mathrm{det}}\), I used the Mahalanobis distance. The Mahalanobis distance of a sample \(x\) from a reference population with mean vector \(\mu\) and covariance matrix \(\Sigma\) is:
$$
\mathrm{MD}(x) = \sqrt{(x-\mu)^\top \Sigma^{-1}(x-\mu)}
$$
I estimated \(\mu\) and \(\Sigma\) from the first normal-operation block of each vehicle. I defined the anomaly threshold as the mean distance plus three standard deviations:
$$
T_{\mathrm{th}} = \bar{D} + 3\sigma_D
$$
Once the first sample with \(\mathrm{MD}(x)>T_{\mathrm{th}}\) was found, I used that point as the beginning of battery degradation. This approach uses the joint correlation of many sensors rather than a single voltage or temperature limit, making the detection more robust under different electric vehicle usage patterns.
Experimental Environment and Model Tuning
The model experiments were conducted using an AutoDL cloud GPU server equipped with an NVIDIA RTX 3090 GPU with 24 GB memory, 14 virtual CPU cores, and an Intel Xeon processor. The software environment included Linux, PyTorch, and Python 3.8. I selected this setup because deep-learning architectures such as LSTM and KAN-LSTM require intensive matrix operations and are trained much more effectively on a GPU server.
I constructed supervised samples by applying a sliding window over the time series. Each sample in the input matrix is a sequence \(X_1, X_2, \dots, X_L\), where each \(X_i\) is an 18-dimensional feature vector. The corresponding label is the RUL at the final time step of the sequence. I varied the window size from 10 to 160 and searched for the window length that minimized the total distance between the moving average and its mean. The first local minimum was found at a window length of 53 time steps, so I used \(L=53\) as the input sequence length.
For hyperparameter tuning, I used different strategies depending on model complexity. The SVR model has only a few hyperparameters, so grid search was sufficient. For XGBoost, LightGBM, LSTM, and KAN-LSTM, I used Bayesian optimization. Bayesian optimization treats the validation error as an unknown function and builds a probabilistic surrogate model. After each trial, it updates the posterior distribution and chooses the next hyperparameter combination that is most likely to improve the result. This approach is more efficient than grid search because it explores promising regions of the parameter space instead of evaluating all possible combinations.
The final hyperparameter choices are listed in the table below.
| Model | Hyperparameters | Search Method |
|---|---|---|
| SVR | C=1, degree=3, gamma=1, coef0=0.1 | Grid search |
| XGBoost | learning_rate=0.086, max_depth=7, subsample=0.5, min_child_samples=9 | Bayesian optimization |
| LightGBM | max_depth=6, num_leaves=36, subsample=0.8386, min_child_samples=16 | Bayesian optimization |
| LSTM | hidden_dim=64, num_layers=1, learning_rate=0.1, batch_size=16, dropout_rate=0.3 | Bayesian optimization |
| KAN-LSTM | KAN hidden_dim=30, B-spline size=5, spline_order=3; LSTM same as LSTM row | Bayesian optimization |
Experimental Results and Comparison
I evaluated all models on the PSM-test data, which was collected from the second electric vehicle model and never used during training. The primary metrics were root mean square error, relative root mean square error, and coefficient of determination.
The root mean square error is defined as:
$$
\mathrm{RMSE} = \sqrt{\frac{1}{n}\sum_{i=1}^{n}\left(y_i-\hat{y}_i\right)^2}
$$
The relative root mean square error is:
$$
\mathrm{RRMSE} = \frac{\mathrm{RMSE}}{\bar{y}}\times 100\%
$$
where \(\bar{y}\) is the mean observed RUL. The coefficient of determination is:
$$
R^2 = 1-\frac{\sum_{i=1}^{n}\left(y_i-\hat{y}_i\right)^2}{\sum_{i=1}^{n}\left(y_i-\bar{y}\right)^2}
$$
For electric vehicle battery life prediction, the experimental-level requirement is typically \(\mathrm{RRMSE}\le 10\%\). In high-safety applications, however, an even stricter requirement is needed: \(\mathrm{RRMSE}\le 5\%\).
The results on the test set are summarized in the table below.
| Model | RMSE (seconds) | RRMSE | R² | Meets 10% criterion | Meets 5% criterion |
|---|---|---|---|---|---|
| SVR | 1331.75 | 10.97% | 0.8985 | No | No |
| XGBoost | 1323.78 | 10.90% | 0.9145 | No | No |
| LightGBM | 1208.57 | 9.95% | 0.9282 | Yes | No |
| LSTM | 632.87 | 5.21% | 0.9565 | Yes | No |
| KAN-LSTM | 320.95 | 2.64% | 0.9775 | Yes | Yes |
SVR performed better than a simple linear model, but it could not fully capture the time-dependent degradation behavior of an electric vehicle battery. The SVR model produced large errors in the phase where the RUL begins to decline. The reason is that SVR is essentially a static regressor. It maps input features to the output without an explicit memory structure. Consequently, the model could not remember how the battery state evolved over the previous minutes.
XGBoost improved slightly on SVR. The tree-based model captured some nonlinear interactions among voltage, current, and temperature variables. However, its RUL curve still had noticeable deviations from the true values. XGBoost has no built-in mechanism to treat the data as a time-ordered sequence. The order of samples in a decision tree does not matter, so important temporal dependencies may be ignored.
LightGBM achieved an RRMSE of 9.95%, which met the experimental-level threshold. The histogram-based algorithm and leaf-wise growth allowed it to fit sharp changes more effectively than XGBoost. Still, LightGBM could not meet the high-safety 5% threshold. The prediction error became larger near the end of battery life, where the battery state often changes quickly and irregularly.
LSTM substantially outperformed all traditional models. Its RMSE was about half of LightGBM’s error. The recurrent architecture uses memory cells and gates, which enable it to learn the gradual decline of battery health over long time windows. The LSTM model produced smoother and more accurate predictions than SVR, XGBoost, and LightGBM. The relative error, 5.21%, was close to the high-safety standard but still not strictly below 5%.
The proposed KAN-LSTM model gave the best results. Its RMSE on the test set was only 320.95 seconds, and its RRMSE was 2.64%. This value is below both criteria, including the strict 5% high-safety threshold. The coefficient of determination reached 0.9775, meaning that 97.75% of the variation in RUL was explained by the model. The improvement is not small. Compared to LSTM, the RMSE was reduced from 632.87 seconds to 320.95 seconds, almost halving the prediction error.
I believe the improvement is due to the architecture of KAN-LSTM. KAN extracts flexible local nonlinear features with B-spline basis functions. For example, when a sudden temperature spike occurs, the B-spline interpolation can encode that spike without blending it into a global trend. Then LSTM uses its gate mechanism to decide how much of that local feature should be kept over time. This decomposition is useful for electric vehicle battery diagnosis because the cell voltage response and internal resistance do not follow a fixed mathematical formula. The interactions among state of charge, temperature, and current are strongly nonlinear, and KAN can adapt to them more freely than fixed activation functions.
Another important observation is that KAN-LSTM is especially effective near the end of the battery’s life. In the last stage of the test sequence, the RUL drops rapidly. Traditional models tended to produce lagged predictions, meaning that they were too optimistic at the moment of imminent failure. The KAN-LSTM model was better at detecting the accelerated degradation pattern because its spline basis could fit the changing curvature of the degradation trajectory. For real electric vehicle monitoring, this property is very valuable because the most dangerous period is exactly the final phase before the energy storage device fails.
Discussion
My experiments show that the choice of algorithm matters more than simply adding more data or more features. The conventional tree-based models are powerful on tabular data, but they are not designed to model time-dependent processes. LSTM is an appropriate choice for time-series regression, but it may not fully represent short-lived nonlinear anomalies. KAN-LSTM is not a simple combination; it is a conscious division of labor. The KAN module acts as a local feature transformer, and the LSTM module is a temporal analyzer. Together they create a model that is both nonlinear and stateful.
I also found that anomaly detection is very important before training a supervised remaining-useful-life model. Using Mahalanobis distance to locate the first abnormal point makes the RUL label physically meaningful. The covariance matrix captures correlations among voltage, current, and temperature variables. If I had instead assumed that all samples before the failure were equally risky, the model could not learn the process of degradation. The abnormal-point detection step contributes to the final success of the electric vehicle fault prediction system.
The feature-selection step also had a visible effect. Removing variables with low information content reduced model complexity and prevented the model from fitting noise. The selected 18 features are closely related to the health state of the rechargeable energy storage system. For example, total voltage, maximum cell voltage, minimum cell voltage, maximum temperature, and state of charge are direct indicators of battery electrical behavior and thermal behavior.
Despite the strong results, I should mention the limitations of this study. First, the data were collected from only two electric vehicle models. A model trained on these vehicles may need additional calibration before it can be applied to a completely different battery chemistry or battery management system. Second, the RUL label was generated with a piecewise-linear approximation. In reality, the degradation curve is not exactly linear, and this approximation may introduce a certain amount of label noise. Third, the current study focuses on predicting the remaining time until failure, but it does not directly identify which exact failure mechanism is dominant, such as lithium plating or thermal runaway. In future work, I hope to extend the label structure and add diagnostic outputs.
| Comparison Dimension | LightGBM | LSTM | KAN-LSTM |
|---|---|---|---|
| Model type | Tree ensemble | Recurrent neural network | Hybrid KAN and recurrent network |
| Ability to model time order | Weak | Strong | Strong |
| Nonlinear local feature extraction | Moderate | Moderate | Strong |
| RRMSE on test set | 9.95% | 5.21% | 2.64% |
| Suitability for high-safety EV applications | Not sufficient | Marginal | Suitable |
From the practical perspective, a model with a relative RUL error below 5% can significantly improve the maintenance planning of electric vehicle battery packs. Imagine a fleet of electric vehicles used for public transportation. If the KAN-LSTM model says that a rechargeable energy storage device will fail in 10,000 seconds, the actual failure time is likely to fall between about 9,744 and 10,256 seconds. This narrow interval is much more actionable than a prediction with 10% error. The maintenance team can schedule an inspection before the vehicle enters a dangerous condition.
In addition, KAN-LSTM reduces the occurrence of “false late” alarms. Traditional threshold alarms are usually triggered only when voltage or temperature exceeds a hard limit. However, in many fault cases, no single feature reaches an alarming level until the very last minute. By learning from the joint evolution of all selected features, the model can issue an early warning based on subtle cross-feature patterns. This is exactly what I observed in the real data: the Mahalanobis-based anomaly detector found the first abnormal signal many time steps before the final failure, and KAN-LSTM was then able to track the RUL decrease over that interval.
The high safety requirements are particularly relevant for electric vehicle rechargeable energy storage systems. Thermal runaway is a rare but catastrophic failure mode. A robust remaining-useful-life prediction model can assist the battery management system in making decisions such as limiting charging current, activating the cooling system, or alerting the driver to move the electric vehicle to a safe location. KAN-LSTM is not a replacement for low-level battery management protection, but it can provide a higher-level forecasting layer that is complementary to voltage protection and current limits.
During the tuning process, I found that Bayesian optimization was especially beneficial for deep models. The number of hyperparameters in KAN-LSTM is larger than in LSTM because the KAN part has its own spline order, grid size, and weight initialization choices. A grid-search approach would require thousands of trials and would be impractical with a deep model. Bayesian optimization required only a limited number of evaluations and still found a combination that produced the best performance. This result suggests that the combination of Bayesian optimization with the KAN-LSTM architecture is a reliable method for developing electric vehicle battery health-management systems.
Conclusion and Future Directions
In my master’s research, I set out to answer a practical question: how accurately can an electric vehicle’s remaining useful life be predicted using only onboard sensor data before a battery fault? I have shown that the answer depends on choosing a model that can capture both nonlinear feature interactions and temporal dependencies. The KAN-LSTM hybrid model fulfills this need better than SVR, XGBoost, LightGBM, or a standalone LSTM.
The main contributions of my work are as follows. First, I used real operating data from two different electric vehicle models to test the fault prediction framework under conditions close to actual use. Second, I proposed a feature-selection pipeline based on Spearman correlation and random forest importance that reduced the initial 27 features to 18 key variables. Third, I proposed a Mahalanobis-distance-based anomaly detection method to locate the starting point of battery degradation and construct piecewise-linear remaining-useful-life labels. Fourth, I designed and implemented the KAN-LSTM model with Bayesian optimization for remaining-useful-life prediction. Fifth, I compared multiple models with standardized evaluation metrics and demonstrated that KAN-LSTM meets the strict 5% high-safety error criterion.
The KAN-LSTM model achieved a root mean square error of 320.95 seconds on the test set, a relative root mean square error of 2.64%, and a coefficient of determination of 0.9775. These numbers show that the model is not only accurate on average, but also highly consistent across different degradation stages. The architecture is especially successful at reducing the large terminal-stage error that appears in conventional models, making it suitable for early warning and proactive maintenance in electric vehicle applications.
In the future, I would like to extend this work in several ways. First, I plan to collect data from a larger fleet of electric vehicles to improve the generalization and robustness of the model. Different operating climates, road conditions, charging habits, and battery types could all influence the degradation path. A multi-vehicle dataset would allow the model to learn a more universal fault pattern. Second, I will investigate attention mechanisms and Transformer-based replacements for the LSTM component in the hybrid architecture. Attention-based models offer the ability to directly focus on the most informative time steps in a sequence, which could improve prediction when the failure pattern is irregular.
Third, I would like to increase the interpretability of the model. KAN has the advantage that its activation functions can be visualized and compared with physical expectations. In future electric vehicle fault prediction systems, I want to extract the learned spline activation curves for voltage and temperature features and analyze whether they match known electrochemical aging mechanisms. If the model can explain why a certain sensor pattern leads to failure, battery designers can use that information to improve the next generation of rechargeable energy storage devices.
Fourth, I plan to integrate the proposed KAN-LSTM model into an online battery health-management system. The current study is an off-line experiment. In a real system, the model must read streaming data, update selected statistics, calculate RUL repeatedly, and issue alerts in real time. Online parameter adaptation and concept-drift detection will be necessary to keep the model accurate as the battery ages and as the operating environment changes. I believe the strong performance of KAN-LSTM in this thesis makes it a promising foundation for such an online electric vehicle safety monitoring platform.
In conclusion, the fault prediction of electric vehicle rechargeable energy storage devices remains a challenging and important problem. Traditional fault-detection methods often fail to capture slow degradation and sudden abnormal behavior simultaneously. My proposed KAN-LSTM model, supported by careful preprocessing and feature engineering, provides a powerful data-driven solution. It demonstrates that deep learning models can meet the strict accuracy demands of high-safety electric vehicle applications if they are designed with the right inductive biases. I expect that further developments in hybrid neural architectures and automated machine learning will continue to improve the reliability of electric vehicles and help accelerate the transition toward cleaner transportation.
