Research on Safety Monitoring of Traction Battery Pack Throughout Life Cycle and Cloud Platform Development

As the electrification of transportation accelerates, the traction battery pack, as the primary energy storage device in electric vehicles, has become a central focus of both academic research and industrial application. The operational safety of the traction battery pack over its entire life cycle, ranging from initial manufacturing to eventual retirement, directly determines the reliability, durability and public acceptance of electric vehicles. The motivation of this dissertation originates from the urgent need to address two coupled challenges: the accurate estimation of the state of health (SOH) of the traction battery pack under unpredictable real-world operating conditions, and the early detection of thermal runaway risks before catastrophic failures occur. Traditional battery management systems embedded in vehicles are often constrained by limited on-board computational resources and storage capabilities. Cloud-based platforms, however, offer an unprecedented opportunity to implement computationally intensive algorithms, store long-term historical data, and perform fleet-level analysis. This work therefore develops a holistic framework that integrates data preprocessing, laboratory experiment design, algorithmic innovation for SOH assessment, a multimodal early-warning method for thermal runaway, and the construction of a complete cloud-native platform to deploy these algorithms in real-world applications.

1. Real-World Data Preprocessing and Experimental Dataset Construction

High-quality data are the foundation of any data-driven algorithm. The real-world driving data used in this dissertation were mainly collected from the National New Energy Vehicle Big Data Platform and enterprise cloud databases. These datasets conform to the Chinese national standard GB/T 32960-2016 and are reported at an approximate sampling frequency of 0.1 Hz. The records contain essential operation variables such as bus voltage, total current, state of charge, cell voltages, and module temperatures derived from diverse vehicle types, including private passenger cars, ride-hailing electric vehicles, and city transit buses. The analysis covers both lithium iron phosphate batteries and ternary lithium battery chemistries, as well as varied driving patterns. A typical example of the raw collected variables from a traction battery pack is summarized in Table 1 with its statistical characteristics during normal operation.

Variable Unit Range Sampling frequency Usage purpose
Total voltage V 250–750 0.1 Hz Misfire check, state segmentation
Total current A −200 to 300 0.1 Hz Charging/discharging identification
Cell voltage V 2.5–4.35 0.1 Hz SOH estimation and fault diagnosis
Cell temperature °C −20 to 60 0.1 Hz Thermal safety evaluation
State of charge % 0–100 0.1 Hz Operating condition segmentation
Vehicle speed km/h 0–120 0.1 Hz Driving/charging/idle discrimination

Despite the richness of the cloud-collected real-world data, numerous practical obstacles emerge before they can be fed into machine-learning models. The data transmission process is susceptible to interruptions, signal noise, and measurement anomalies. Therefore, a rigorous preprocessing pipeline was developed. First, state segmentation is performed based on the charging flag, vehicle speed, and accumulated mileage. A data point is classified as charging when the charging flag equals 1, vehicle speed remains zero, and the mileage increase is negligible. Conversely, a discharge condition is identified by the charging flag equaling 0 and either a significant speed change or mileage accumulation. The idle state is recognized when the vehicle speed and mileage remain continuously unchanged. In order to preserve meaningful battery operational windows, the charging and discharging data segments are sliced into individual charge/discharge episodes. After segmenting the data, abnormal data are identified and treated using various strategies tailored to the measurement type. Cell voltage abnormal readings are detected when a sudden change in voltage is observed, accompanied by a distribution standard deviation that is substantially smaller than the voltage jump itself. The total voltage is cross-checked against the cell voltage data, and any inconsistencies are marked as abnormal. The current data are routinely checked against a threshold, and temperature data are processed in a similar manner to the voltages. The final step of data cleaning employs three main operations: linear interpolation for isolated abnormal points when the cumulative anomalous ratio is less than five sampling points, resampling and deletion for more severe corruption, and segment rejection whenever the anomalous data exceed one percent of the segment length. The flow diagram of the preprocessing operations is presented in the left portion of Figure 1.

The preprocessing outputs at each stage involve many thousands of records, making a quantitative validation possible. Specifically, after cleaning, the dataset includes a total of more than twelve million valid sample points from more than twenty electric buses and passenger vehicles. The anomalous proportion is reduced from an initial average value of 3.2% to below 0.05%, guaranteeing data integrity for subsequent analysis.

One major limitation of the real-world datasets is the absence of true capacity labels. Unlike laboratory cycling tests, vehicle batteries cannot be easily discharged at a constant low rate for controlled capacity measurements. The actual available capacity is an internal unobservable state that degrades over time. To overcome this challenge, and to generate rigorous ground-truth labels for validating SOH estimation algorithms, a battery-pack aging experiment was designed based on real-world operating conditions. The experimental platform consisted of eight series-connected blocks, each block comprising two parallel-connected commercial 20 Ah lithium iron phosphate prismatic cells. The resulting configuration corresponds to an 8-series-2-parallel structure, which is a representative architecture for a small-scale traction battery pack. Table 2 lists the main parameters of the chosen cell for this study.

Parameter Value Parameter Value
Cathode chemistry LiFePO₄ Charging upper cutoff voltage 3.65 V
Rated capacity 20 Ah Discharging lower cutoff voltage 2.0 V
Nominal voltage 3.2 V Maximum continuous discharge current 3.0 C
Cell dimensions 27 mm × 70 mm × 145 mm Maximum peak discharge current 5.0 C

Before assembling the cells into a pack, a strict screening procedure was applied to reduce the initial inconsistency. Each cell underwent two capacity measurements using a 0.5 C charge-and-discharge protocol, and internal resistance measurements were repeated three times at 25 degrees Celsius. After obtaining the capacity and resistance data, sixteen cells with closely matched performance were selected and configured into the 8S2P structure. The statistical distribution of cell capacity and resistance before the screening process is shown in Table 3, highlighting the necessity of screening for pack-level consistency.

Statistical indicator Cell capacity (Ah) Internal resistance (mΩ)
Mean value 21.6 0.94
Standard deviation 0.28 0.04
Maximum value 22.16 1.03
Minimum value 21.40 0.90

The experimental protocol was carefully designed to simulate both normal and stressful operational cycles of a traction battery pack. Each aging point includes the following sequence of tests: (i) single-cell capacity tests at the beginning; (ii) pack-level capacity measurement at a moderate 0.3 C discharge rate; (iii) hybrid pulse power characterization (HPPC) to track resistance changes; (iv) alternating real-world big-data driving cycles and accelerated aging cycles; and (v) repeated pack capacity checks every few cycles. For the driving-cycle portion, six segments out of twelve available real-world discharge-current profiles are randomly selected and normalized according to the pack capacity, then loaded onto the pack through a programmable charger. The charging phase mimics modern fast charging by adopting a five-step constant-current staircase. The acceleration phase applies a 2 C discharge current until any single cell reaches the lower cutoff voltage. Between each test, the pack is rested at 25 degrees Celsius for one hour to ensure thermal equilibrium. This protocol was repeated continuously for two months and generated fifteen distinct aging points, spanning a capacity degradation from 100% to approximately 80%. The obtained data include the complete evolution of cell voltages, branch currents, pack voltage and current at every stage, thereby providing rich validation datasets for subsequent algorithms.

The data collected from the entire experimental campaign enable both short-term dynamic assessment and long-term degradation analysis. The laboratory datasets include detailed capacity measurements that act as ground truth, supplementing the cloud-based real-world datasets. The combination of real-world archival data and well-controlled laboratory experiments thus constitutes a complete foundation for developing and testing the algorithms described in the following chapters.

2. Health State Estimation Method Based on Inconsistency Assessment

The overall performance of a traction battery pack depends on the performance of the weakest cell and is greatly affected by cell-to-cell inconsistency. Over the course of aging, this inconsistency often increases monotonically, since cells within the same pack may experience slightly different thermal environments, current distributions, and internal degradation pathways. The useful capacity of the pack is therefore not merely the sum of individual cell capacities, but also depends on the degree of consistency among cells. This observation motivates the development of a health state assessment method that infers the degradation of the pack from the evolution of inconsistency. In this approach, the departure of individual voltage curves from a representative standard voltage curve is measured during identical operating conditions. When all cells are healthy and operating in a coherent manner, the ensemble of cell voltage curves is tightly clustered. However, when one or several cells exhibit accelerated capacity fade, increased resistance, or micro-short-circuit faults, the corresponding voltage traces will deviate, thereby increasing the dispersion of the family of voltage curves and consequently augmenting the cell-inconsistency feature. This chapter builds upon such a premise to estimate the SOH of the entire pack.

The first step in the proposed SOH estimation framework is the isolation of a robust reference voltage curve from the recorded segment of data. Since the pack contains many cells, each having slightly different characteristics, a computationally efficient yet robust approach using multi-stage 3-sigma rejection is adopted. The algorithm works recursively: at each sampling instant, the average voltage of all cells is computed; then the standard deviation is derived. Any cell voltage outside the interval of the mean plus or minus three standard deviations is temporarily removed. The remaining values are utilized to recalculate the mean and standard deviation. This procedure is repeated until the relative change in the standard deviation between successive iterations is smaller than an acceptable tolerance. A dynamic threshold, rather than a fixed one, is employed in the final stages to better adapt to varying operating conditions. The final standard voltage value at each time index is then preserved, giving the batch-normalized reference curve V_m. The computation is summarized as:

$$ V_m(i) = \frac{1}{N_{\text{sel}}(i)} \sum_{j=1}^{N_{\text{sel}}(i)} V_j(i), \quad \text{for cells passing the } 3\sigma \text{ gate at time } i.$$

Since the raw voltage signals are occasionally sampled at irregular intervals, an exact one-to-one correspondence between the cell voltage vector and the reference voltage vector cannot always be assumed. A dynamic time warping (DTW) algorithm is therefore applied to align the two curves in the temporal domain. Dynamic time warping computes the minimal cumulative distance between two sequences under the constraint of local continuity, enabling the optimal matching of time indices. The local distance matrix element between a single-cell voltage and the reference curve is defined as the squared Euclidean distance. A dynamic programming recurrence is then:

$$ D(i,j) = |V_c(i)-V_m(j)| + \min\big\{D(i-1,j), D(i,j-1), D(i-1,j-1)\big\},$$

where D(i,j) is the accumulated distance between the first i samples of the cell voltage and the first j samples of the reference voltage. The optimal path is reconstructed by backtracking from the final element to the initial element, which yields a synchronized pair of voltage arrays. Once the synchronization is achieved, the similarity E_sline between an individual cell voltage vector and the reference curve is computed by the mean squared difference along the warped path. The similarity between a particular cell and the average behavior reflects the degree to which this cell deviates from the majority of the battery population. It is noteworthy that this ensures a fair comparison even if the sampling times differ between channels or between different vehicle reports.

Given the similarity values of all cells in the pack during one particular charging or discharging event, a probability density function (PDF) of the similarities is constructed by kernel density estimation. The Gaussian kernel is selected because it yields a smooth and continuous estimate of the distribution. The bandwidth of the kernel is adapted according to Scott’s rule:

$$ h = \sigma \left(\frac{5}{4N}\right)^{1/5},$$

with σ being the standard deviation of the similarity values and N the number of similarity observations. The kernel density estimate at a query point x is then obtained by summing all the individual Gaussian kernels:

$$ \hat{f}(x)=\frac{1}{N h}\sum_{i=1}^{N}\frac{1}{\sqrt{2\pi}}\exp\left[-\frac{(x-S_i)^2}{2h^2}\right].$$

From the resulting probability density curve, two fundamental metrics are extracted: the cumulative probability enclosed in the central region around the mode, and the spread of the distribution. The former reflects how concentrated the cells are near the normal behavior, while the latter describes how widely the cell similarities are dispersed. A robust Inconsistency Index is then defined as the inverse of the ratio between the probability mass around the mode and the total bandwidth of the distribution. In mathematical form:

$$ \Phi = \frac{P_{10\%}}{S_{10\%}},$$

where the subscripts refer to the central 10% probability region. Larger values of the index correspond to a more consistent pack, whereas smaller values signal growing inconsistency. Since the raw Inconsistency Index can be noisy, particularly for short segments or when a vehicle is operating under rapidly varying load conditions, a density-based spatial clustering of applications with noise (DBSCAN) algorithm is adopted to remove outliers. The remaining clean index points are fitted with a smoothing spline to produce a continuous inconsistency trajectory over the life of the vehicle.

To assess the correlation between the derived Inconsistency Index and the true SOH, five real-world vehicles were selected from the database. Among them, vehicles #1 through #3 are equipped with lithium iron phosphate batteries, while vehicles #4 and #5 use ternary lithium batteries. The data length covers more than one year of continued operation, providing up to one million data rows per vehicle. The correlation between the fitted inconsistency curve and the pack capacity is quantified by the Pearson correlation coefficient:

$$ \rho = \frac{\sum_{i=1}^{N_{p}} (A_i – \mu_A)(B_i – \mu_B)}{\sqrt{\sum_{i=1}^{N_{p}} (A_i – \mu_A)^2 \sum_{i=1}^{N_{p}} (B_i – \mu_B)^2}},$$

where A_i and B_i are the capacity and inconsistency variables, respectively. The correlation values for vehicles #1 to #5 range from 0.864 to 0.990, with an average value of 0.921. Such a high correlation strongly confirms the existence of a meaningful physical relationship between the inconsistency evolution and the capacity degradation of a traction battery pack under real-world usage. It also implies that a purely voltage-based inconsistency indicator can serve as an effective proxy to track the SOH of the pack.

To rigorously quantify the accuracy of the proposed algorithm, a validation experiment using the laboratory battery pack was carried out. Since the experimental pack was disassembled and re-tested at each aging point, the actual capacity is precisely known from the reference performance test. The algorithm is then applied to the voltage data collected from the large real-world driving cycles. When the SOH values obtained by the algorithm are compared with the reference capacities at the corresponding aging points, the average estimation error is about 2%. The mean absolute error of the fitted SOH curve with respect to the reference capacity values is only 0.22 Ah, corresponding to a relative error of 1.13% with respect to the nominal capacity. Table 4 summarizes the error statistics across all aging points.

Aging point Reference capacity (Ah) Estimated SOH (%) Absolute error
1 37.42 99.8 0.8
3 36.15 96.5 0.9
5 34.87 92.9 1.6
7 33.79 90.2 0.6
9 33.0 88.4 1.4
11 32.17 86.1 1.8
13 31.50 84.3 0.7
15 30.68 82.2 1.2

The most significant advantage of the proposed method is its high adaptability to real-world applications. It does not require a trained deep-learning model, does not require any historical capacity label for calibration, and is robust against missing data or sampling misalignment. The final outcome of this chapter is thus an efficient, simple, yet accurate algorithm that derives SOH of a traction battery pack directly from the cloud-recorded cell voltages, thereby providing a dimensionless and interpretable health indicator that is suitable both for individual vehicle tracking and for fleet-level statistical analysis.

3. Multimodal Thermal Runaway Warning Method for Traction Battery Pack

Thermal runaway remains the most dangerous and detrimental failure mode of the traction battery pack. It may originate from manufacturing defects, mechanical damage, prolonged overcharging, or internal short-circuit evolution. In the early stage of a fault, the change in voltage is often weak and is buried under the effects of current fluctuation, temperature drift, and normal cell-to-cell inconsistencies. In contrast, later-stage signals such as smoke, fire, or massive voltage collapse provide too little time for intervention. Therefore, a robust early-warning method capable of capturing subtle precursors days or even weeks before the thermal runaway is of critical importance. This chapter develops a multimodal fault-diagnosis framework that makes full use of the rich information contained in cloud-collected data.

To ensure that the features contain complementary information, three features are extracted from two scales and three analytical dimensions. First, a time-domain feature, named the trend skewing measurement, is derived from the relative voltage curve of each cell. Given a charge or discharge segment, the standard voltage of the traction battery pack is once again obtained by taking the robust mean of the cell voltages after trimming the top and bottom 20% of sorted voltages. The relative voltage of cell k at time i is:

$$ V_{R,k}(i) = V_k(i) – V_{m}(i).$$

Having removed the effect of the SOC-dependent nominal voltage, the relative voltage is small in absolute value but still captures any anomalous behavior. To amplify this behavior, the relative voltage is integrated over the entire segment and then divided by the total segment duration, which yields the trend skewing measurement:

$$ TSM_k = \frac{1}{t_{\text{sum}}} \sum_{i=1}^{N_s} V_{R,k}(i) \Delta t_i,$$

where Δt_i is the discrete time interval. Since a cell prone to internal short circuit usually has a voltage that tends to be lower than the pack average, a modified TSM score is further computed to reduce the influence of cells that are consistently above the average. The mathematical expression reads:

$$ TSM_k^* = \left\{ \begin{array}{ll} \alpha \cdot TSM_k & \text{if } TSM_k > 0, \\ -TSM_k & \text{if } TSM_k < 0, \end{array} \right. $$

with α being a scaling factor assigned a value of 0.8. In this way, abnormally underperforming cells become more distinguishable than cells with a high voltage profile.

The second feature is the charging resistance, which is computed for cells during the step changes of the charging current. Many electric vehicles employ staged battery charging strategies that alter the charging current at predefined thresholds. These slope discontinuities induce sharp changes in both the cell terminal voltage and the total current. By taking the voltage difference before and after the current transition and dividing it by the difference of the corresponding currents, an effective battery resistance can be determined:

$$ R_C = – \frac{U_A – U_B}{I_A – I_B},$$

in which U_A and U_B are the measured cell voltages immediately before and after the current step, and I_A and I_B are the corresponding total currents. Although this estimated resistance is not exactly equal to the ohmic or polarization resistance, it reflects the internal conducting state of the cell and its evolution with aging or fault development. Cells with internal anomalies, such as separator degradation or dendrite growth, often exhibit reduced internal resistance under a charging current.

The third feature is a time-frequency spectrogram. The cell voltage signal is first transformed into a relative voltage signal as described above, thus removing the dominant dependence on the state of charge. The relative voltage then becomes approximately stationary for healthy cells, making it suitable for joint time-frequency analysis. A short-time Fourier transform is applied to the relative voltage to obtain a spectrogram:

$$ P(t,\omega) = \left| \int_{-\infty}^{\infty} V_R(\tau) h(\tau – t) e^{-j\omega \tau} d\tau \right|^2,$$

where h(τ) is a moving analysis window. In this work, a sliding window length of four seconds with an overlap of three seconds is chosen to balance time and frequency resolution. The resulting spectrogram is normalized in a logarithmic scale before being fed into the feature-extraction network:

$$ S_{\text{edit}} = 10 \log_{10} \left(\text{abs}(S)\right).$$

These three families of features contain information from different views: TSM captures slowly evolving voltage trends; charging resistance relates to the internal resistive path; and the spectrogram exposes transient anomalies in the frequency content. Crucially, the spectral information cannot be easily deduced from the TSM alone, and vice versa. Therefore a multimodal strategy is necessary.

Once the features have been extracted, they are subsequently processed through the proposed multimodal fault diagnostic framework. The block diagram of the framework consists of a CNN feature extraction layer, a KPCA dimension reduction layer, a normalization and feature fusion module, and a multi-dimensional recursive fault diagnosis module. The entire architecture is shown in Figure 1.

For spectrogram inputs, a convolutional neural network (CNN) is first utilized to extract latent high-dimensional features. The CNN architecture contains three convolution layers with ReLU6 and ReLU activations, followed by both max pooling and average pooling layers, a flatten layer, and a fully connected layer. The convolution operation performed at layer l + 1 is expressed as:

$$ H_j^{l+1} = f\left(\sum_{i \in M_j} H_i^l * W_{ij}^{l+1} + b_j^{l+1}\right),$$

where H_j^{l+1} is the j-th feature map at the (l+1)-th layer, W is the convolution kernel, b is the bias, f(·) is the activation function, and M_j is the set of input feature maps connected to the j-th output feature map. The adopted activation function is expressed as:

$$ f(x) = \left\{ \begin{array}{ll} 0, & x\le 0, \\ x, & 0< x \le 6, \\ 6, & x>6, \end{array} \right. $$

for the initial layer and the standard rectified linear unit for the subsequent layers. The pooling operations reduce the spatial size of the feature maps while preserving the important information. The output of the last pooling layer is flattened into a one-dimensional vector and passed through a fully connected layer to form a 32-dimensional feature vector. This vector is considered the deep representation of the original spectrogram.

To reduce this high-dimensional feature vector into a compact and effective representation, a kernel principal component analysis (KPCA) layer follows the CNN extraction. In KPCA, the original feature vectors are mapped into a higher-dimensional feature space through a nonlinear feature map φ. The covariance matrix in this feature space is:

$$ C = \frac{1}{N} \sum_{i=1}^{N} \phi(\mathbf{x}_i)\phi(\mathbf{x}_i)^T.$$

Its eigenvalue problem is then solved by means of the kernel matrix K, which is constructed from the Gaussian radial basis kernel:

$$ K_{ij} = \exp\left(-\frac{\|\mathbf{x}_i – \mathbf{x}_j\|^2}{2\sigma_N^2}\right).$$

The feature dimension is finally reduced from 32 to 1 after selecting the direction that contains the largest variance in the kernel space. The combination of linear and nonlinear operations ensures robust information compression while preserving meaningful structure that is hidden in the original spectrogram.

After all three feature modalities have been converted into numerical scalars, a normalization procedure is required to make them comparable and to prepare for fusion. The normal value of each feature is calculated from the population of cells in the pack. Since most cells inside a traction battery pack are normal, the mode of the cell feature distribution serves as the baseline. Each cell’s feature value is then divided by this baseline to produce a normalized dimensionless indicator, the mean of which is approximately unity for normal cells. In a mathematical formulation:

$$ F_{\text{norm},k} = \frac{F_k}{F_{\text{baseline}}},$$

where F_baseline is computed by trimming the top and bottom 20% of the sorted feature values. This normalization method prevents a few highly abnormal cells from corrupting the reference baseline, thereby enabling consistent comparison across vehicles, pack types, and operating conditions.

Once the normalized feature values for each cell are obtained, they are combined into a three-dimensional feature vector:

$$ \mathbf{F}_k = \left[ F_{\text{TSM},k},\; F_{\text{C},k},\; F_{\text{image},k} \right].$$

The multi-dimensional recursive fault diagnostic module then begins by determining the centroid of the feature vectors. Initially, the arithmetic mean of all cell feature vectors is computed as the centroid position. However, since abnormal cells may skew the centroid, an iterative refinement is applied. The Euclidean distance from each cell vector to the centroid is calculated.

$$ D_k = \sqrt{\left(F_{\text{TSM},k} – F_{\text{TSM},c}\right)^2 + \left(F_{\text{C},k} – F_{\text{C},c}\right)^2 + \left(F_{\text{image},k} – F_{\text{image},c}\right)^2}.$$

The average distance D_pre is obtained. Then the cell with the largest distance is temporarily excluded from the centroid calculation, after which a new average distance D_aft is computed from the remaining cells. The ratio β between D_aft and D_pre is monitored. As long as β is below 0.95, the iteration repeats. This iterative process effectively excludes potentially anomalous cells from influencing the centroid location. When the stopping criterion is met, the final centroid is considered as the most representative center of normal cells.

In the final stage, the normalized Euclidean distance from each cell vector to the refined centroid is determined, and a uniform threshold is applied to classify the cell into a healthy or anomalous state:

$$ D_k^* \le D_{\text{fault}}, \Rightarrow \text{normal}; \quad D_k^* > D_{\text{fault}}, \Rightarrow \text{abnormal},$$

with D_fault set to 7.5 based on the empirical analysis of many datasets.

To demonstrate the performance of the proposed framework, results from nine different vehicles were analyzed. Three representative cases are discussed in detail. The first one is a vehicle that suffered a thermal runaway in April 2023. Its battery system consisted of 8 modules, each containing 12 cells in a 4-series-3-parallel configuration. The cell parameters are those reported in Table 2. The manufacturer retrieved the cloud data from the fourteen days before the thermal runaway. The diagnosis demonstrates that the fault score for the origin cell #76 exceeded the threshold as early as eleven days before the thermal runaway, and then continued to stay above the threshold until the event. Neighboring cells #50 and #66 also showed anomalous scores, which corroborates the post-event teardown findings. The early warning was therefore provided more than seven days in advance, which is essential for a driver to stop the vehicle and leave sufficient time before the thermal runaway.

The second case is a suspected fault vehicle. Its data were recorded for about three months prior to a subsequent fire. During the recorded time period the fault score of cell #41 gradually increased, first exceeding the threshold on April 17, followed by cells #9 and #93 around May 10. This progressive deterioration reflects the inherent nature of internal cell defects that develop over time. The algorithm successfully detected the latent fault while other cells still exhibited normal behavior.

The third case is an entirely normal vehicle. None of the cells exceeded the fault threshold at any time during the collection period. The highest fault scores did not approach the threshold, and no sustained upward tendency was observed. These results confirm the low false-positive rate of the proposed warning method.

In addition to the passenger-car cases, two city buses equipped with lithium iron phosphate batteries of a completely different cell format (302 Ah) were also tested. One of them suffered a thermal runaway that originated from cell #372. The proposed algorithm once again successfully isolated cell #372 from the remaining cells of the pack, even though no visual voltage abnormality could be identified by simple inspection. For the normal bus, none of the 162 cells crossed the threshold. The algorithm thus demonstrates good transferability across battery chemistry, cell capacities, pack topology, and operating conditions. A summary of the experimental alarms is presented in Table 5.

Vehicle Battery system Number of cells Actual thermal runaway Predicted faulty cell Advance warning time (days)
Vehicle A Ternary lithium 96 Yes (cell #76) #76, #50, #66 11
Vehicle B Ternary lithium 96 Yes (cell #41) #41, #9, #93 10
Vehicle C Ternary lithium 96 Yes origin cell 8
Vehicle D LFP 378 Yes (cell #372) #372 8
Vehicle E Ternary lithium 96 No None —
Vehicle F LFP 162 No None —

The framework demonstrates several significant advantages: (i) the average alarm time is 11 days, giving enough time to react; (ii) it is a purely unsupervised method that does not require labeled training data; (iii) it avoids threshold drift due to battery aging because all features are normalized relative to the whole pack; (iv) it is computationally efficient enough for cloud deployment; and (v) it is easily extensible to add other features. Overall, the multimodal warning method provides a significant layer of safety for the traction battery pack and enables fleet-level proactive supervision.

4. Development of the Cloud-Based Safety Monitoring Platform

The ultimate goal of the previous algorithmic research is to transfer the academic advances into practical deployment for large-scale monitoring of the traction battery pack. Realizing this ambition requires the creation of a robust cloud-based platform that can store large volumes of time-series data, execute the SOH estimation and thermal runaway warning algorithms automatically, and provide intuitive visualizations and timely alerts to fleet operators or vehicle owners.

To satisfy these requirements, a comprehensive cloud monitoring platform has been designed and implemented. The architecture follows a three-tier model that separates the data storage layer, the application logic layer, and the presentation layer. The database tier uses the MySQL relational database management system. MySQL is open-source, mature, and optimized for concurrent read and write operations, and its performance can be scaled horizontally through clustering and partitioning. In the context of monitoring large fleets of electric vehicles, the accumulated quantity of data is enormous; each vehicle can upload more than one data report per second. A dedicated database schema was designed for managing vehicle information, cell voltage arrays, module temperatures, algorithm results, and warning events. Table 6 illustrates the core tables and their main fields.

Table name Main fields Key constraints
vehicle_info vehicle_id, model, battery_type, pack_capacity Primary key vehicle_id
cell_data vehicle_id, timestamp, cell_no, voltage Index (vehicle_id,timestamp)
pack_data vehicle_id, timestamp, total_voltage, current, soc Index vehicle_id
soh_result vehicle_id, timestamp, soh_value, inconsistency_index Foreign key vehicle_id
fault_score vehicle_id, timestamp, cell_no, fault_score Foreign key vehicle_id
alarm_event vehicle_id, alarm_time, alarm_type, severity Foreign key vehicle_id

The frontend of the platform is built using the three core technologies of the Web: HTML, CSS, and JavaScript. HTML provides the semantic structure of the pages, CSS governs the visual layout and stylistic elements, while JavaScript enables dynamic interactions. In addition, the ECharts library is utilized for rich real-time data visualization. Through the Web frontend, platform administrators can inspect the real-time status of all registered electric vehicles, view field statistics, analyze the geographical distribution of vehicles, visualize health evaluation results, and access detailed battery information for each individual car. The interface is composed of several major pages, each designed for a specific monitoring function. The homepage, the vehicle list, the overall SOH monitoring page, the overall safety overview page, and the detail page of each battery form the major navigation hierarchy.

The backend application layer is developed in Python. Python was selected because it enables seamless integration with the machine-learning libraries used in the previous chapters, including Scikit-learn for kernel density estimation and kernel principal component analysis, and PyTorch for the CNN feature extraction module. The Django Web framework is utilized as the main backend framework. Django’s ORM (Object-Relational Mapping) capability makes it easy to interact with the MySQL database using Python classes instead of direct SQL queries. Django also provides built-in security features, such as protection against SQL injection, cross-site scripting, and cross-site request forgery. In addition, Django’s RESTful API support allows seamless communication with the frontend, enabling dynamic updates without page reloads.

The platform workflow starts with the reception of real-time data uploaded by the vehicle telemetry terminal. Upon data reception, the data are first validated and then stored in the corresponding tables. Whenever new data are received, the backend automatically triggers the appropriate algorithm modules. For SOH evaluation, the algorithm runs after a sufficient amount of new charging/discharging data has been accumulated. For thermal runaway diagnosis, the new data segment is analyzed through the multimodal diagnostic framework, and each cell’s fault score is calculated. The results are then stored back into the database and simultaneously pushed to the monitoring frontend. Figure 1 provides a conceptual illustration of the platform’s data flow and the algorithms embedded in the monitoring center.

The platform also provides a user-friendly dashboard with several distinct features. The homepage section displays the total number of registered vehicles, online status statistics, operating mode distribution (charging, driving, idle), and provincial distribution on a map. The overall SOH monitoring page demonstrates the statistical distribution of the estimated SOH of all vehicles in the fleet, allowing an operator to quickly identify vehicles with premature degradation. The safety overview page summarizes the results of the thermal runaway warning algorithms by category of risk and provides a link to the list of suspicious vehicles. Finally, the battery detail page offers a graphical view of a single battery pack, including the connection topology, current status values, individual cell fault scores, as well as a historical plot of the SOH trajectory. A combination of tables, bar charts, line graphs, and heat maps is used to convey the needed information intuitively.

To guarantee high availability and responsiveness, certain optimization measures are added to the platform. First, a task queue mechanism manages the execution of heavy computation, preventing the backend from being blocked by simultaneous algorithm requests. Second, a caching layer stores frequently accessed data such as vehicle basic information and recent warning statuses, shortening the response time. Third, the database tuning incorporates proper indexing to speed up query execution. Finally, a role-based access control mechanism restricts the operations allowed to different users, thereby enhancing data security.

5. Conclusion and Outlook

This dissertation has systematically addressed the full life-cycle safety monitoring of the traction battery pack, ranging from data collection and preprocessing, algorithmic development for health state estimation and thermal runaway warning, to the final construction of a cloud-based monitoring platform. The major conclusions are summarized as follows:

(1) A comprehensive data preprocessing pipeline was developed, including state segmentation, anomaly detection, resampling and interpolation, thereby enabling the use of noisy and asynchronous real-world vehicle data for algorithm training and validation. In parallel, a dedicated battery pack aging experiment mimicking real driving profiles, fast charging and accelerated aging was designed and executed. The experiment provided accurate capacity labels and high-quality voltage data from a real 8S2P battery configuration over 15 aging points for validating the proposed SOH algorithm.

(2) An SOH estimation approach based on inconsistency evaluation was proposed. The method extracts a representative voltage baseline using a dynamic multi-stage 3-sigma rejection, aligns each cell voltage curve using dynamic time warping, constructs a similarity probability density function with a Gaussian kernel, and derives a robust inconsistency index. The average Pearson correlation coefficient between the inconsistency curve and the capacity curve of five real-world vehicles is 0.921, demonstrating a strong relationship. Experimental validation shows that the algorithm estimates the SOH of a traction battery pack with an average error of about 2%, while requiring only low-frequency voltage data and no model pretraining, making it highly suitable for cloud deployment.

(3) A multimodal early-warning method for thermal runaway was developed based on trend skewing measurements, charging resistance, and time-frequency spectrograms. The multimodal inputs are processed by a CNN-KPCA feature-extraction module, normalized and fused into a three-dimensional feature vector, and classified by a recursive fault diagnosis module with iterative centroid refinement. The algorithm was validated on multiple thermal runaway vehicles and several normal vehicles, achieving an average advance warning time of 11 days while maintaining a very low false-positive rate. Moreover, the algorithm performs well on different battery chemistries and vehicle types, proving its feasibility and universality for fleet-level monitoring.

(4) A power-battery cloud safety monitoring platform was developed using MySQL, Django, HTML, CSS, JavaScript and ECharts. The platform integrates real-time data storage, SOH estimation, thermal runaway alarm, user management, and data visualization. It solves the limitation of on-board BMS computational resources and enables transparent and scalable monitoring of any number of traction battery packs. The modular design provides a robust foundation for future expansion such as adding other safety indicators, predictive maintenance recommendations, or advanced big-data analytics.

Despite the developments accomplished in this dissertation, several limitations are acknowledged. For instance, the data cleaning stage is not yet fully automatic. Although the proposed SOH estimation algorithm works well for typical vehicles, its performance may degrade for battery packs with a very small number of parallel cells. The thermal runaway framework relies on a threshold that is currently tuned using a limited set of real-world data. In future work, the adaptive selection of thresholds across the entire fleet will be further studied. Additional directions include the inclusion of synthetic fault data from electrochemical and thermal models, the fusion of additional information such as battery temperature gradients and charging behavior with charging stations, and the application of advanced transfer-learning approaches to reduce the amount of labeled data needed. Overall, the presented framework constitutes a significant step toward an intelligent, proactive, and cloud-assisted safety management system for the traction battery pack.

Scroll to Top