With the widespread promotion of new energy vehicles, safety issues have attracted considerable attention. The normal aging of high-voltage battery packs or abusive usage during operation can elevate the likelihood of safety problems. Consequently, monitoring and evaluating the safety of power batteries throughout their entire lifecycle holds significant importance. To address this challenge, my research focuses on state of health estimation algorithms and thermal runaway warning algorithms for lithium-ion batteries, alongside the construction of a cloud-based battery safety monitoring platform that enables the engineering deployment of these algorithms.
This dissertation primarily focuses on the research of power battery full-lifecycle safety monitoring algorithms and the development of a corresponding cloud-based platform. The core work involves four major components: real-world data preprocessing and construction of experimental datasets, development of a battery state of health estimation method based on inconsistency evaluation, proposal of a multi-modal thermal runaway warning framework, and the architectural design of a cloud-based safety monitoring platform.
1. Introduction and Research Background
The transformation of the global energy structure, driven by environmental protection requirements, has accelerated the adoption of clean energy sources. In the transportation sector, which accounts for nearly 10% of global energy consumption, the electrification of vehicles plays a crucial role in alleviating energy supply-demand contradictions and reducing environmental pollution. As of 2024, annual sales of new energy vehicles in China reached 12.866 million units, a 35.5% year-on-year increase. However, with this rapid proliferation, the number of safety incidents has also increased, raising concerns that hinder further industrial development.
As the core system of electric vehicles, the high-voltage battery’s safety and lifespan directly influence vehicle reliability and consumer acceptance. Lithium-ion batteries, favored for their superior energy density, good charge-discharge performance, and long cycle life, have become the unanimous choice for electric vehicles. However, due to uncertainties in production and usage—such as impurities introduced during manufacturing, physical impacts during use, and lithium dendrite growth caused by fast charging—internal insulation failure can occur, leading to internal short circuits. These events release significant heat within a short period, potentially triggering thermal runaway. National new energy vehicle big data alliance statistics indicate that lithium-ion battery faults account for over 50% of fire accidents in electric vehicles.
Battery lifespan issues similarly hinder the further adoption of electric vehicles. Although the cycle life of lithium iron phosphate batteries can now exceed 2,000 cycles, inconsistencies among cells within the battery pack may cause individual cells to fail prematurely, reducing the overall lifespan of the pack. This is particularly challenging because such defects are difficult to detect at the manufacturing stage.
To promote wider adoption of electric vehicles, improvements in battery safety and lifespan are indispensable. While physical and chemical system enhancements are necessary, indirect approaches—such as enhanced monitoring, fault diagnosis, and life prediction algorithms—can also significantly improve safety. The evolution of vehicle connectivity and cloud computing infrastructure now supports more sophisticated off-board algorithms, creating opportunities for continuous improvement of battery safety through advanced data analytics.
1.1 Review of State of Health Estimation Algorithms
Battery state of health (SOH) estimation is fundamental to battery management systems. Table 1 summarizes the main methodological categories. Aging of lithium-ion batteries manifests through capacity fade, internal resistance increase, and power capability degradation. SOH can be characterized as the ratio of current performance parameters to initial values. Changes in SOH not only affect vehicle range but also influence safety, making accurate SOH assessment crucial.
| Category | Typical approaches | Advantages | Limitations |
|---|---|---|---|
| Model-based | Equivalent circuit, electrochemical, mathematical models | Physical interpretability | Complex, computationally intensive |
| Data-driven | Neural networks, Gaussian process regression, support vector machines | High accuracy, nonlinear adaptability | Require extensive training data |
| Hybrid | Combinations of model and data-driven | Enhanced accuracy | Increased complexity |
| Cloud-based | Real-world data mining, feature engineering | Applicability to actual vehicles | Lower data quality, missing labels |
Model-based methods employ mathematical formulations, equivalent circuit models, or electrochemical models to track degradation. Equivalent circuit models identify circuit parameters that correlate with aging. Electrochemical models simulate internal reactions through partial differential equations, providing detailed mechanistic insight. However, their computational burden limits practical implementation. Data-driven methods leverage machine learning to estimate SOH from large volumes of data. Techniques include recurrent neural networks, long short-term memory networks, convolutional autoencoders, and hybrid structures. While achieving impressive accuracy, these methods require substantial training data.
Real-world conditions are considerably more complex, with operational and sampling randomness, making laboratory methods less transferable. Cloud-based approaches utilizing vehicle data have emerged, addressing the gap between laboratory validation and field application. These methods use low-rate, long-duration data, although lacking direct capacity labels necessitates estimation strategies.
1.2 Review of Fault Diagnosis Algorithms
Battery faults arise from electrical, thermal, or mechanical abuse, or from production defects, leading to capacity loss, power limitation, or even catastrophic thermal runaway. Fault diagnosis methods generally fall into two categories: laboratory data-based and real-world vehicle data-based approaches. Laboratory-based diagnosis includes knowledge-driven, model-driven, signal processing, and data-driven methods. Table 2 provides a comparative analysis.
| Category | Subtype | Core principle | Advantages | Disadvantages |
|---|---|---|---|---|
| Knowledge-based | Expert systems, fault trees, rules | Derive diagnosis from expert knowledge | No model required | Heavily dependent on expert input |
| Model-based | State estimation, parameter identification | Generate residuals via models | Objective analysis | Complex, computationally demanding |
| Signal-based | Thresholds, signal processing | Analyze signal attributes (amplitude, frequency) | Stable and robust | High computational effort, weaker discrimination |
| Data-driven | Machine learning, multivariate statistics | Use historical data patterns | No physical model required, high accuracy | Data quality dependent |
Knowledge-driven methods were developed early and applied widely, converting expert knowledge into computable forms. Model-based diagnostics create battery models to estimate key parameters and compare against measurements. While physically interpretable, these frequently require significant computation. Signal-based techniques process sensor data using thresholds and analysis methods. Data-driven approaches, particularly machine learning, have become a research hotspot due to their adaptability to nonlinear battery behavior without explicit physical models.
Real-world vehicle environments exhibit more complex conditions, making direct application of laboratory methods unsuitable. Cloud data, collected at low sampling rates, poses additional challenges including data loss and interference. To address these, various algorithms using deep learning, ensemble methods, and multi-feature fusion have been developed. A significant challenge is obtaining labeled data. Real-world data for thermal runaway is rare, and diagnosis under such conditions often requires unsupervised or semi-supervised approaches.
1.3 Current Deployment of Safety Monitoring Algorithms
While substantial research progress has been achieved in both SOH and thermal runaway warning algorithms, their practical deployment lags significantly. Onboard battery management systems are limited by computational resources, commonly relying on traditional threshold-based methods. Vehicle connectivity through cloud platforms allows for deployment of more sophisticated algorithms. Several industrial systems and patents have emerged, offering some degree of battery health evaluation and lifecycle management. However, several issues remain: many algorithms require complex pre-training and large labeled datasets. Algorithm deployment on generic cloud infrastructure with robustness and scalability remains challenging.
From the analysis of existing research, several unresolved issues remain. For SOH estimation: the discrepancy between real-world and laboratory data; the inherent difficulty in obtaining accurate capacity data from real-world vehicles; and the high adaptability required for varying data quality. For thermal runaway warning: obtaining rare-event training data; mapping single features to actual failures; and the limited scalability of some algorithms. For monitoring platforms: either overly simplistic functional scope, or complexity that limits practical deployment.
2. Real-world Data Preprocessing and Experimental Dataset Construction
The research on battery safety monitoring algorithms requires a comprehensive dataset. My data consisted of two main sources: real-world vehicle cloud-based big data and laboratory-simulated real-vehicle experiments. Real-world data complying with GB/T 32960-2016 serves as the primary validation source.
2.1 Preprocessing of Real-world Vehicle Data
Real-world vehicle data collected from cloud platforms presents several challenges. In this study, data from the National New Energy Vehicle Big Data Alliance and partner-provided fleet data, sampled at ~0.1 Hz, were used. The data contains cell voltage, total voltage, total current, temperature, SOC, and vehicle status information. I developed a comprehensive preprocessing pipeline to handle various anomalies. The data was first segmented into charging and discharging states based on the charging flag, vehicle speed, and mileage. Charging state is identified when the charging flag is active, speed is zero, and no mileage accumulation occurs. Discharging or driving states are characterized by the discharge flag with continuous mileage increase. Idle states are defined by the discharge flag with zero continuous mileage change.
After segmentation, each segment underwent data cleaning. For cell voltage data, sudden jumps—where the voltage distribution standard deviation is far smaller than the average level—were considered anomalous. I applied linear interpolation when anomalous values account for less than 5 samples, and the fragment was deleted if anomalies exceeded 1% of the segment length. For the total voltage data, I cross-validated with the sum of cell voltages to detect jumps. Current data received threshold-based validation. Individual temperature and SOC data underwent similar anomaly assessment. After processing, the real vehicle database comprised clean charging and discharging fragments.
2.2 Simulated Battery Pack Experiment Based on Real-world Operating Conditions
To address the absence of capacity information in real-world data and validate SOH estimation with known labels, I designed a battery pack simulation experiment. Prismatic lithium iron phosphate cells were selected with parameters listed in Table 3.
| Parameter | Value |
|---|---|
| Cell chemistry | Lithium iron phosphate |
| Charge cutoff voltage | 3.65 V |
| Discharge cutoff voltage | 2.00 V |
| Nominal capacity | 20 Ah |
| Maximum continuous discharge current | 3C |
| Nominal voltage | 3.2 V |
| Cell dimensions | 27 × 70 × 145 mm |
I selected 16 cells with close capacity and internal resistance for assembly into an 8-series-2-parallel configuration (Figure 2-5), giving 16 individual cells monitored within the experiment. Figure 2-8 shows capacity and internal resistance distribution before assembly to minimize initial inconsistency.
The experimental protocol comprised: (1) cell capacity testing (0.5C constant current to 3.65V, constant voltage to C/20, then 0.3C discharge); (2) pack capacity testing; (3) HPPC testing; (4) real-world driving cycle procedures; and (5) accelerated aging cycles. Each aging point consisted of these stages. During pack formation, a pre-selection process ensured minimal initial differences between branches. The battery pack is arranged with impedance-matched modules, and each branch was validated to have similar total capacity and internal resistance.
Cell capacity test procedure:
The cell (at 25°C within a thermal chamber) was first rested for one hour, charged at 0.5C to the voltage limit, maintained at constant voltage until the current dropped to C/20, rested for one hour, and finally discharged at 0.3C to the lower cutoff. This procedure was repeated twice.
Pack capacity test procedure:
All cells were discharged to the lower cutoff voltage on individual channels before assembling into the pack. Pack capacity tests used 0.3C constant current charging to first cutoff, followed by 0.1C charging to final cutoff, rest, and 0.3C constant-current discharge to first discharge cutoff, repeated twice.
HPPC test: The pack was charged at 1C to 3.6V on any cell, then using 0.1C to 3.65V, rested for 1 hour. A discharge pulse of 1C for 10 s was applied, followed by 40 s rest, then a 1C charge for 10 s and 40 s rest. SOC was decreased in increments of 10%, with one-hour rests between increments.
Real-world driving cycle: After completing the characterization, the pack underwent multi-stage charging (1C to first cell at 3.44V, 0.85C to 3.468V, 0.5C to 3.497V, 0.2C to 3.55V, 0.1C to 3.65V). Real-world discharge profiles extracted from the cloud data were applied to the pack until any cell reached 2V, followed by a one-hour rest. Each aging point included six such cycles using randomized real-world profiles. Figure 2-15 displays examples of the current excitation profiles and Figure 2-17 illustrates the results obtained from the pack under real-world conditions.
Accelerated aging: After the real-world cycle, multi-stage charging was repeated, followed by 2C discharges to cutoff. The number of cycles per aging point was adjusted according to observed degradation state. Several aging points were completed with full characterization.
This comprehensive dataset—comprising 15 aging points—provided both validation for the SOH estimation and necessary context for the thermal runaway warning studies.
3. Battery Health State Estimation via Inconsistency Assessment
As a battery pack ages, the aging of constituent cells progresses unevenly due to initial manufacturing differences and varying operational stresses. The pack capacity is affected both by overall cell aging and the worsening inconsistency among cells. This worsening inconsistency frequently results in capacity loss and diminished performance, contributing to an elevated risk of accelerated aging. My method leverages this relationship to estimate SOH by evaluating cell-to-cell inconsistency.
3.1 Feature Extraction
Determination of reference curve: After data segmentation and cleaning, to establish a representative baseline voltage during charge segments, I employ a 3σ multi-level filtering process. Because individual cells within a module may have large deviations, simply averaging voltage curves is sensitive to the presence of outliers. The 3σ filter iteratively excludes samples lying beyond ±3σ from the average, recalculates the mean and standard deviation, and stops when the change in standard deviation between iterations falls below a tolerance threshold. This dynamic threshold helps preserve the core distribution of typical cells. The final filtered mean across healthy samples at each sampling point constitutes the reference curve $V_m=[V_{m0}, V_{m1}, …, V_{mn}]$.
Similarity assessment with DTW: Given that data collection timing may vary across cells, I employ Dynamic Time Warping to align the curves before measuring similarity. Let a measured single cell voltage curve be $V_c=[V_{c0}, V_{c1}, …, V_{cn}]$ and reference curve $V_m$. I initialize a distance matrix $D$ having rows roughly corresponding to $V_c$ and columns to $V_m$ (the matrix may be non-square due to time-warp, though curves have the same length after alignment).
First, the local distance matrix is computed as:
$$
D_{\text{dist}}(i, j) = |V_{ci} – V_{mj}|
$$
Then, a cumulative distance matrix is built recursively. For boundaries:
$$
D[0, j] = \sum_{k=0}^{j} D_{\text{dist}}(0, k)
$$
$$
D[i, 0] = \sum_{k=0}^{i} D_{\text{dist}}(k, 0)
$$
and for interior points:
$$
D[i, j] = D_{\text{dist}}(i, j) + \min\{D[i-1, j],\, D[i, j-1],\, D[i-1, j-1]\}
$$
By tracing back from the top-right corner to the bottom-left corner, the optimal alignment path is obtained. This path re-indexes $V_c$ and $V_m$ to produce aligned sequences $V_{c\_\text{new}}$ and $V_{m\_\text{new}}$. Subsequently, similarity between a cell and the pack reference is measured via Euclidean distance:
$$
S_{\text{line}}(i) = \sqrt{\sum_{j=1}^{n}\left(V_{c\_\text{new}}(j) – V_{m\_\text{new}}(j)\right)^2}
$$
3.2 Probability Density Estimation for Inconsistency Metrics
After computing the similarity for each charging fragment across all cells, I evaluate their statistical distribution. The discrete set $S_{\text{line}}$ is interpreted as samples from an underlying distribution. Kernel density estimation yields a smooth probability density function $\hat{f}_h(x)$.
First the standard deviation of the similarity samples is calculated:
$$
\sigma = \sqrt{\frac{1}{N-1}\sum_{i=1}^{N}\left(S_{\text{line}}(i) – \bar{S}_{\text{line}}\right)^2}
$$
where $N$ denotes the number of samples and $\bar{S}_{\text{line}}$ is the sample mean. The bandwidth is determined by Scott’s rule:
$$
h = \sigma \left(\frac{4}{5N}\right)^{1/5}
$$
A Gaussian kernel is chosen:
$$
K_h(x) = \frac{1}{h\sqrt{2\pi}} \exp\left(-\frac{x^2}{2h^2}\right)
$$
The kernel density estimate is:
$$
\hat{f}_h(x) = \frac{1}{N h} \sum_{i=1}^{N} K_h\left(\frac{x – S_{\text{line}}(i)}{h}\right)
$$
The distribution of similarities reveals the degree of inconsistency. In a healthy, highly consistent pack, most cells will display similarly low distances from the reference, resulting in a narrow, high-peaked density. Conversely, inconsistent packs show a more dispersed and lower-peaked distribution.
Metric extraction: From the PDF, I extract two features: accumulated density $P$ within 10% intervals flanking the peak, and the span $S$ of the distribution. The inconsistency indicator is obtained as the quotient $S/P$. After this, DBSCAN clustering removes outliers from the indicator set. Final trends are obtained via spline fitting, enabling correlation analyses with capacity degradation trends.
3.3 Algorithm Validation Results
The proposed method was validated using data from vehicles #1 to #5, all city buses. Vehicles #1–#3 used LFP chemistry, while #4–#5 used NCM chemistry. Each vehicle contributed more than 1 million data rows, spanning a full year of operations. Figure 3-3 shows the inconsistency estimation results before and after outlier removal for each vehicle.
Correlation with battery capacity: Because capacity values are not measured directly in real-world vehicles, I applied a Monte Carlo-based capacity estimation technique. The Pearson correlation coefficient (PCC) was employed to evaluate correlation between the inconsistency trajectory and the capacity curve:
$$
\rho_{A,B} = \frac{ \sum_{i=1}^{N} \left( A_i – \mu_A \right)\left( B_i – \mu_B \right)}{ (N-1)\sigma_A \sigma_B }
$$
Across the five vehicles, PCC values ranged from 0.864 (vehicle #2) to 0.99 (vehicle #5), with a mean of 0.921. These results indicate a robust linear relationship, implying that the inconsistency metrics effectively track capacity fade.
Laboratory verification: The experimental dataset from Chapter 2 was used to quantitatively assess SOH estimation accuracy. At each aging point, the real-world operating cycles before capacity measurement provided the voltage inputs. The resulting SOH estimates were aligned with measured pack capacities, as illustrated in Figure. The fitted SOH curve yielded errors of approximately 2% (MAE of 0.22). The algorithm performed robustly across different aging stages, confirming its high practical utility.
4. Multi-modal Thermal Runaway Warning Method
Fault diagnosis of high-voltage battery packs in electric vehicles faces significant challenges including limited labeled data, variability of real-world operating conditions, and stringent computational constraints. To address these issues, I developed a multi-modal fault diagnosis framework for thermal runaway early warning based on unsupervised learning. This framework uses fused data from multiple domains to yield robust and timely warnings without requiring model pre-training or massive labeled datasets.
4.1 Feature Extraction Methodology
Charging resistance: The battery internal resistance is a key parameter that reflects the battery’s internal state. Multi-stage charging strategies common in electric vehicles induce voltage jumps at current switch points. By measuring the voltage changes and current differences, an effective charging resistance can be computed:
$$
R_C = \frac{U_B – U_A}{I_A – I_B}
$$
where $U_A$ and $U_B$ are the average cell voltages immediately before and after the current switch, and $I_A$ and $I_B$ are the corresponding currents. This estimation, while not identical to an ohmic or polarization resistance measured under controlled conditions, is well-suited for cloud-based analytics.
Relative voltage and Trend Skewing Measurement (TSM): To decouple voltage changes due to SOC from underlying anomalies while capturing subtle fault evolution, I derived a standard voltage curve $V_N(i)$ computed from cell voltages after excluding the top and bottom 20%:
$$
V_N(i) = \frac{\sum_{n}^{m} V_x(i)}{m-n}
$$
The relative voltage for each cell is:
$$
V_{R_i}(t) = V_i(t) – V_N(t)
$$
In a healthy pack, $V_R$ values cluster tightly near zero. The trend skewing measurement (TSM) cumulative metric integrates the relative voltage over a driving segment:
$$
TSM_i = \frac{\displaystyle\sum_{t} V_{R_i}(t)}{t_{\text{sum}}}
$$
Because thermal-runaway-prone cells typically exhibit internal short circuits with slightly lower voltage, I further emphasize such failures through an optimized weighting parameter $\alpha$:
$$
TSM_i^* = \begin{cases} TSM_i \cdot \alpha & \text{if } TSM_i < 0 \\ TSM_i & \text{if } TSM_i \ge 0 \end{cases}
$$
with a scaling factor $\alpha = 0.8$.
Time-frequency spectrogram extraction: To augment time-domain information, STFT was applied to the relative voltage sequences. The STFT calculation is:
$$
P(t, \omega) = \left| \int_{-\infty}^{\infty} V(\tau) h(\tau – t) e^{-j\omega \tau} d\tau \right|^2
$$
where $h(\tau-t)$ is the window function. The resulting spectrogram is transformed into decibel units:
$$
S_{\text{edit}} = 10 \cdot \log_{10}\left(\text{abs}(S)\right)
$$
After comparison, I selected a window size of 4 and an overlap of 3 to achieve a balanced time-frequency resolution. These spectrograms offer rich dynamic fault signatures that time-domain indicators may miss.
4.2 Multi-modal Fault Diagnosis Framework Architecture
To effectively fuse the diverse features derived above, my diagnostic framework incorporates three modules: graphical feature extraction, feature fusion processing, and multi-dimensional recurrent fault diagnosis.
CNN-based Graphical Feature Extraction. A convolutional neural network extracts high-dimensional features from the spectrograms. Table 4 lists the CNN architecture configuration used for the images.
| Layer | Type | Kernel size | Activation |
|---|---|---|---|
| 1 | Convolutional | 4 × 4 | — |
| 2 | Activation function | — | ReLU6 |
| 3 | Convolutional | 3 × 3 | — |
| 4 | Activation function | — | ReLU |
| 5 | Convolutional | 3 × 3 | — |
| 6 | Activation function | — | ReLU |
| 7 | Convolutional | 3 × 3 | — |
| 8 | Max Pooling | 3 | — |
| 9 | Average Pooling | — | — |
| 10 | Flattening | — | — |
| 11 | Fully Connected | — | — |
The ReLU6 function, which bounds activations between 0 and 6, improves robustness under low precision computation, while the combination of max and average pooling captures salient regional features while maintaining globally consistent feature statistics. The output is a 32-dimensional feature vector.
KPCA-based nonlinear dimensionality reduction. The CNN output was fed into KPCA for dimensionality reduction. Given the high-dimensional CNN output represented as $X = \{x_1, x_2, …, x_n\}$, after mapping to feature space via nonlinear function $\Phi(X)$, the covariance matrix is:
$$
C = \frac{1}{n} \sum_{i=1}^{n} \Phi(x_i) \Phi(x_i)^T
$$
A Gaussian kernel is selected:
$$
K_{ij} = \exp\left( -\frac{\left\| x_i – x_j \right\|^2}{N} \right)
$$
Eigen-decomposition of the kernel matrix produces lower-dimensional feature representations. The final dimension was selected as one to achieve effective fusion with the numerical features.
Feature fusion and normalization. Since different features exist on disparate scales, I normalize all quantities. Taking TSM as an example, reference value $TSM_{\text{normal}}$ is calculated from the middle 60% of values after sorting:
$$
TSM_{\text{normal}} = \frac{\displaystyle\sum_{i=m}^{n} TSM(i)}{n-m}
$$
The uniform feature is derived as:
$$
TSM_{\text{uniform}}(i) = \frac{TSM(i)}{TSM_{\text{normal}}}
$$
This process normalizes each feature so that healthy cells exhibit values near unity. The resulting three-dimensional feature vector $F_i$ is:
$$
F_i = \left[ TSM_{\text{uniform}}(i),\; R_{C,\text{uniform}}(i),\; G_{\text{uniform}}(i) \right]
$$
where $R_C$ is charging resistance and $G$ is the CNN-KPCA spectral feature.
Multidimensional recursive fault diagnosis module. Fault detection relies on the principle that faulty cells are a small minority and deviate from the majority. I first compute the centroid position of all feature vectors in feature space:
$$
F_{\text{xc}} = \frac{1}{n} \sum_{i=1}^{n} F_{xi}, \quad F_{\text{yc}} = \frac{1}{n} \sum_{i=1}^{n} F_{yi}, \quad F_{\text{zc}} = \frac{1}{n} \sum_{i=1}^{n} F_{zi}
$$
Euclidean distances to the centroid are:
$$
D_i = \sqrt{\left(F_{xi} – F_{\text{xc}}\right)^{2} + \left(F_{yi} – F_{\text{yc}}\right)^{2} + \left(F_{zi} – F_{\text{zc}}\right)^{2}}
$$
I compute the average distance $D_{\text{pre}}$, remove the maximum distance, and recalculate $D_{\text{aft}}$. The ratio is:
$$
\beta = \frac{D_{\text{aft}}}{D_{\text{pre}}}
$$
This iterative process repeats, progressively removing the most extreme cells, until $\beta \ge 0.95$. The final centroid is then used to compute each cell’s distance $D_{fi}$, which serves as the diagnostic score. A threshold $D_{\text{fault}} = 7.5$ was established empirically as an effective balance between false alarms and missed detections.
4.3 Algorithm Validation Results
The proposed framework was evaluated using multiple real-world vehicles. I present detailed results for representative cases, focusing on three categories—thermal runaway vehicles, suspected faults, and normal vehicles. One of the more prominent cases studied involved a vehicle that experienced a thermal runaway event. Table 5 shows the battery cell specifications for the case-study vehicle.
| Parameter | Value |
|---|---|
| Cell type | NCM ternary |
| Nominal capacity | 50 Ah |
| Nominal voltage | 3.65 V |
| Charge cutoff voltage | 4.2 V |
| Discharge cutoff voltage | 3.0 V |
| Dimensions | 148 × 27 × 92 mm |
In the thermal runaway vehicle, the module configuration was 4 series × 3 parallel per module, leading to a total of 8 modules across the pack. The thermal runaway originated from cell #76. Based on data gathered from April 1 to April 14, 2023, vehicle diagnostics identified cell #76’s anomaly score exceeding the threshold 11 days before the thermal runaway, with #50 and #66 also surpassing the threshold. The pack’s normal cells maintained scores well below the threshold of 7.5.
Validation across multiple vehicles. Results for the thermal runaway LFP battery vehicle align with observed data: when tested on a bus with a 2p189s configuration (378 total cells), the algorithm assigned an abnormal score to cell #372, which was confirmed as the source of thermal runaway, while other cells scored low. In the suspected fault case, the algorithm had identified cell #41 and later #9 and #93 as anomalous, well before the eventual fire. For normal vehicles, throughout all examined cycles, the fault scores never approached critical threshold levels. A statistical summary of validation results is presented in Table 6.
| Vehicle type | Battery chemistry | Case outcome | Warning timing | False alarm rate |
|---|---|---|---|---|
| Passenger vehicle (fault) | Ternary NCM | Thermal runaway | 11 days before | None |
| Bus (fault — LFP) | LFP | Thermal runaway | Early diagnostics | None |
| Passenger vehicle (suspected) | Ternary NCM | Fire after cutoff | Multi-week detection | None |
| Passenger vehicle (normal) | Ternary NCM (large cells) | No fault | Not applicable | Very low |
| Buses (normal) | LFP | No fault | Not applicable | None observed |
Feature effectiveness evaluation. Figures and analyses of the extracted features confirmed that the TSM feature accumulates subtle time-domain anomalies effectively. The digital CNN-KPCA features extracted from spectrograms provided uncorrelated but complementary fault information. Normalization enabled consistent comparison across different data scales. The multi-stage recursive centroid calculation method improved the diagnostic discrimination capability by an average of 7.34% across various faulted cells.
5. Development of a Cloud-based Safety Monitoring Platform
The preceding chapters introduced robust SOH estimation and thermal runaway warning methods. However, the full realization of enhanced safety through cloud monitoring demands an application-level platform that integrates these algorithms, data management, and real-time interaction. To accomplish this, I developed a cloud-based battery safety assessment platform for new energy vehicles, promising continuous, near-real-time monitoring of both health condition and safety risk.
5.1 Overall Assessment Process
The platform architecture follows the full-chain structure illustrated in Figure 5-1, connecting vehicle-side data acquisition to cloud-based analytics. On the vehicle side, the BMS acquires high-voltage battery information including voltages, currents, temperatures, and SOC, which are forwarded by the vehicle control unit to a telematics unit. Data transfer occurs via the on-board cellular module to the cloud. The cloud-based platform receives and processes the datastream, applies the SOH estimation and thermal runaway warning algorithms, and delivers results back to fleet managers and drivers. This process leverages cloud computation resources, addressing the BMS’s inherent limitations in storing large-scale historical data and running complex diagnostic algorithms.
5.2 Technical Implementation Strategy
To satisfy the project requirements of high scalability, storage robustness, and visualization effectiveness, the platform uses the following technology architecture. At the data storage layer, MySQL is used, structured query language (SQL) supporting massive data management with transactional security and data consistency. MySQL’s InnoDB storage engine supports efficient indexing. The platform partitions historical data to manage the volume effectively, achieving fast responses and secure access controls.
At the frontend layer, HTML determines page structure, CSS controls the visual styling, and JavaScript implements real-time user interactions and dynamic data rendering. The asynchronous fetch and event-driven nature of JavaScript facilitates live updates. In this manner, the interface supports responsive, real-time data visualization across the different modules of the dashboard.
The backend is developed using Python due to its strength in algorithm deployment and rapid prototyping. The platform uses the Django web framework for URL routing, request/response handling, and API provisioning. For database interactions, the integrated Object Relational Mapping model simplifies data handling. Django also provides administrative interfaces and layered security measures, including SQL injection protection, session security, and cross-site request forgery (CSRF) defense.
The system architecture is modular. This modularity eases integration of new algorithms. Table 7 summarizes the principal technology stack components and their responsibilities.
| Layer | Technology | Function |
|---|---|---|
| Data Storage | MySQL | Store raw and processed battery data, maintain high concurrent access support |
| Backend computation | Python (scikit-learn, PyTorch) | Execute SOH, hot-runaway algorithms, and API handling |
| Web framework | Django | Control URL routing, handle HTTP requests/responses, ORM model mapping, CSRF defense |
| Frontend | HTML, CSS, JavaScript | Interface structure, styling, dynamic visualizations, and interactive operation |
| Visualization | ECharts library | Display plots, geographical distributions, historical trends |
5.3 Platform Functional Layout and Design
The platform structure is organized into five primary pages: a Home Dashboard, a Vehicle List, Overall SOH Monitoring, a Safety Risk Monitoring, and individual Battery Detail Views. This route layout enables intuitive navigation from an overview down to specific battery modules.
Home dashboard. The initial page communicates at a glance the platform’s overall operational statistics including active vehicle count, online status, vehicle models, and geographical distributions. It pushes real-time battery fault notifications and alerts via the algorithms, enabling fleet managers to spot issues before they escalate. Real-time statistical updates combined with graphical maps provide a global monitoring capability.
Battery list page. This is the look-up menu of vehicles connected to the platform. It lists battery pack basic information such as vehicle model, battery type, rated capacity, and current state. Users can select or filter vehicles. This page also contains lists of batteries grouped by the risk levels derived from algorithm outcomes, accessible from the overall Risk Monitoring Page for fast response protocols.
SOH overall status monitoring. The page displays the summarized SOH estimation results across vehicles. In addition to key statistics on pack health, it presents histograms and distribution maps describing the state-of-health distribution of the entire fleet. This view allows fleet managers to schedule maintenance replenishment, observe trends across diverse vehicle topologies, and understand which factors contribute to battery aging.
Safety risk status page. This page integrates historical trend statistics, real-time risk counts, and proportions of vehicles with various safety risk levels. As soon as the diagnosis algorithm flags a vehicle, this page highlights it for immediate action. A click-through yields a list of risky battery records sorted by severity. This page also explains the thermal runaway methodology, so that end users are aware of the warning’s origin.
Battery detail page. This page is designed for in-depth analysis of a selected high-voltage battery pack. Its user interface includes modules for:
| Module | Content |
|---|---|
| Topology & basic info | Graphic cell arrangement, connection method, rated parameters |
| Live status | Current total voltage, total current, state of charge, instantaneous power |
| Algorithm ratings | SOH score and safety anomaly index for each battery cell, real-time status |
| Real-time curves | Continuously updated curves showing cell voltage and current trends |
| Detailed results | SOH history trajectory, thermal runaway diagnostics results for each individual cell over time |
Using topology, operators can locate failing cells in the physical layout pattern. Live curves help to visually inspect severe changes associated with incipient faults. The detailed algorithm page offers specific diagnostic output history and comparisons against fleet baselines.
5.4 Algorithm Deployment and Integration
To ensure the algorithms can operate within the cloud environment, I modified their design for scalability and efficiency. State of health estimation modules were encapsulated into independently callable computation services. All nonvolatile model parameters are stored in the database and loaded into the runtime environment during initialization. For thermal runaway warning, which utilizes spectrogram computations and the feature-fusion module with heavy matrix operations, I implemented asynchronous processing. When a user request is submitted, Django immediately returns a job ID while the CPU- or GPU-bound workload is scheduled into the compute queue. Once finished, the result is written back to the database and announced by pushing a notification to the frontend.
During deployment, I focused on high-concurrency robustness. Optimizations included cached hot data and compressed payload formats for data exchange. The data uploaded from vehicles uses effective management, being transformed into MySQL tables on the fly. In order to avoid heavy loads from simultaneous access, I leveraged the threading support from the Python task queue. The access control also was enhanced, bounding access to the battery data by credentials and roles, thereby preventing unauthorized visibility of sensitive information.
Conclusions and Future Perspectives
My dissertation investigated safety monitoring methods for electric vehicle high-voltage battery systems and their deployment. To achieve this, the study integrated real-world big data analytics, laboratory simulation experiments, advanced machine learning methodologies, and modern cloud platform technologies.
The main contributions are summarized as follows:
1. A physical-data-linked dataset: I processed large-scale field data, showcasing key cleaning techniques adapted to cloud-based battery datasets. I further constructed a real-world operating condition-based aging experiment, creating a controlled analog of vehicle battery pack aging processes. The experiment provided essential validation data where true capacity was measured.
2. SOH estimation method based on inconsistency evaluation: My proposed technique derives inconsistency features from the individual cell voltage curves. Dynamic time warping ensures robust alignment and comparison of low-frequency sampled curves despite asynchronous sampling; Gaussian kernel density estimation yields a stable distribution; and indicators extracted from that distribution strongly correlate (PCC mean of 0.921) with capacity changes in operating vehicles. Laboratory data showed a consistent SOH estimation error of about 2%, better meeting real-world requirements with simpler implementation than existing methods.
3. Multi-modal thermal runaway early warning: Building from a unique combination of trends over time, spectrograms from the frequency domain, and internal battery state estimation, the framework autonomously detects multivariate anomalies in an unsupervised manner. I introduced TSM for fault signal accumulation and applied CNN-KPCA methods for advanced image feature extraction. Across multi-vehicle data sets, an average of 11 days early warning was observed, with very low false-negative rates.
4. Deployment in a functional cloud-based platform: To move algorithms from analysis to action, I developed and validated an end-to-end cloud service using MySQL, Python/Django, and HTML/CSS/JavaScript. The module structure provides versatile, responsive user interfaces for monitoring, inspecting, diagnosing, and real-time alerting on the battery systems. This significantly helps to bridge the gap between battery laboratory research and operational implementation.
Limitations and future directions: the proposed SOH estimation method requires data from a sufficiently large number of series-connected cells to obtain stable statistics. If the battery pack uses few cells, performance may weaken. Future work will investigate the fusion of additional capacity indicators, permitting single-cell or small-string applicability.
Additionally, while the data processing pipeline is semi-automated, some steps depend on human intervention. Future algorithms should aim to eliminate these issues for end-to-end automation. Although thermal runaway warnings have demonstrated excellent performance across different chemistries and vehicle types, open questions remain for the early diagnosis of even more subtle gradual faults under harsh conditions and on a wider variety of chemistries.
Due to the rapid progress and constantly expanding scale of cloud data, greater ability to analyze all available vehicles will make it possible to further optimize the methodologies. The proposed platform’s modular architecture makes it an appropriate base for these future extensions.