A power battery monomer thermal runaway early warning method and system based on deep learning
By constructing a single-cell voltage time-series matrix based on deep learning, performing variational mode decomposition and feature extraction, and combining it with an autoencoder for unsupervised anomaly detection, the problem of lag and accuracy in early warning of thermal runaway of power battery cells is solved, and efficient early warning is achieved in real vehicle environment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-25
- Publication Date
- 2026-06-26
Smart Images
Figure CN122283473A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power battery technology, specifically to a method and system for early warning of thermal runaway in a single power battery cell based on deep learning. Background Technology
[0002] As the core power source of new energy vehicles, the health status of individual cells of the power battery directly determines the safety of the vehicle operation. If early performance degradation, internal short circuits and other abnormalities of the cell are not warned in time, they can easily lead to serious accidents such as thermal runaway and fire. Therefore, individual cell-level safety warning is a core requirement of the power battery management system.
[0003] Existing power battery safety early warning technologies mostly rely on global macroscopic features such as total voltage, total current, and temperature to construct early warning models. However, the essence of power battery failures originates from individual cells. Different cells within the same battery pack exhibit significant differences in degradation rates and health. Global features can easily mask subtle anomalies in individual cells, leading to delayed early warnings and difficulty in identifying risks in their early stages. Furthermore, some technologies only use simple time-domain features such as voltage mean, variance, and peak value. However, power battery voltage signals are nonlinear and non-stationary, containing deeper information such as electrochemical reactions and electrode aging. A single time-domain feature cannot comprehensively characterize the cell's state, easily resulting in low early warning accuracy and high false alarm and missed alarm rates.
[0004] Existing technologies also face challenges in data adaptability. Most solutions rely on laboratory simulation data for model training and validation. However, in real-world vehicle scenarios, power batteries face complex operating conditions such as fluctuating road conditions, changing ambient temperatures, and frequent charging / discharging switching. Voltage data suffers from noise interference, overall offset, and timing jitter, making it difficult for laboratory data to replicate these complex characteristics, thus limiting the technology's adaptability in real-world applications. Furthermore, traditional machine learning methods and simple neural network models are insufficient in capturing the long-term dependencies and high-dimensional nonlinear characteristics of voltage time-series data, making it difficult to adapt to the performance degradation patterns throughout the cell's lifespan. As the number of cell cycles increases, its voltage response characteristics dynamically change, and existing models struggle to adapt to these dynamic changes, resulting in significant fluctuations in warning accuracy throughout the entire lifespan.
[0005] In terms of sample acquisition, most existing early warning methods adopt a supervised learning model, which relies on a large number of labeled fault samples to train the model. However, there are very few serious fault samples in real vehicle scenarios, and it is even more difficult to obtain early weak anomaly samples. The problem of sample imbalance is prominent, which limits the training of supervised models and the ability to identify unlabeled new anomalies or early weak anomalies needs to be improved.
[0006] The aforementioned problems stem from multiple factors. On the one hand, existing technologies primarily focus on monitoring the overall performance of the battery pack, with insufficient research on the correlation between individual cell consistency differences, microscopic performance degradation, and macroscopic signals. Furthermore, the adaptability of signal processing methods to the complex and non-stationary characteristics of power battery voltage signals needs improvement. On the other hand, real-world vehicle power battery operation data is massive in volume and complex in dimensions, making data cleaning and standardization challenging. Moreover, traditional model architectures are not well-matched to the characteristics of time-series data, and there is a lack of reasonable feature reduction and extraction structures designed specifically for the high-dimensional characteristics of voltage data. Summary of the Invention
[0007] The purpose of this invention is to propose a method and system for early warning of thermal runaway of a single power battery cell based on deep learning, which can improve the accuracy and timeliness of early warning of thermal runaway of a single power battery cell.
[0008] To achieve the above objectives, in a first aspect, the present invention proposes a deep learning-based early warning method for thermal runaway of a single power battery cell, comprising: Collect battery monitoring data during actual vehicle operation and divide discharge segments based on the time-series changes in battery charging status; For each discharge segment, construct the individual voltage timing matrix and calculate the relative voltage timing matrix; Variational mode decomposition is performed on the relative voltage time-series signal of each individual cell to obtain the intrinsic mode components corresponding to each individual cell. The noise-dominant modes are removed, and the effective modes are retained to form the mode feature sequence of each individual cell. The modal feature sequence of each individual cell is input into a bidirectional LSTM-multi-head attention feature extraction network, which outputs a high-dimensional robust feature vector. An unsupervised autoencoder model is constructed using high-dimensional robust feature vectors. The high-dimensional robust feature vectors of each individual cell are input into the autoencoder, and the reconstruction error is calculated as the anomaly score of that individual cell. Statistical analysis is performed on the abnormal scores of all individual cells in the current discharge segment to determine the adaptive abnormal judgment threshold. Based on the comparison between the abnormal scores and the adaptive abnormal judgment threshold, an early warning is issued for thermal runaway of individual power battery cells.
[0009] The beneficial effects of the basic scheme: This calculation scheme provides characteristics of battery charging state time-series changes to divide discharge segments, accurately distinguishes the actual discharge operation stage of the battery, eliminates invalid data interference from non-discharge conditions, closely matches the actual operating state of real vehicles, maximizes the value of big data from real vehicle on-board monitoring, and lays a high-quality data foundation for subsequent thermal runaway early warning. By constructing a time-series matrix of relative voltage of individual cells, the interference of common voltage fluctuations of the battery pack on the individual cell differences is eliminated; and by combining variational mode decomposition to remove noise-dominant modes and retain effective feature modes, the dual processing effectively suppresses various noises such as environmental interference and acquisition errors under real vehicle operating conditions, solving the technical problems of high noise and low feature recognition in real vehicle operating data.
[0010] A bidirectional LSTM-multi-head attention fusion network is introduced. The bidirectional LSTM fully captures the forward and backward temporal correlation patterns of voltage mode features, and the multi-head attention mechanism adaptively weights and strengthens key temporal anomaly features. This overcomes the limitations of traditional single temporal feature extraction, which is insufficient and key details are easily lost. It outputs a high-dimensional and highly discriminative battery health feature vector.
[0011] The algorithm employs an autoencoder-based unsupervised anomaly detection model, eliminating the need for extensive fault sample labeling and perfectly adapting to the scarcity of fault samples and the variability of operating conditions in real-world vehicle scenarios. Simultaneously, it uses statistical analysis to adaptively generate anomaly judgment thresholds, eliminating reliance on manually fixed thresholds and avoiding the problems of poor adaptability and repeated parameter tuning required by manual thresholds, thereby reducing the manual cost and maintenance difficulty of algorithm implementation.
[0012] This technical solution has strong adaptability to the changing driving conditions of new energy vehicles. It can directly realize online early warning based on real-time on-board monitoring data, providing reliable technical support for early prevention and control of thermal runaway of power battery cells. It has high industrialization and engineering application value.
[0013] As a feasible and preferred option, battery monitoring data includes individual cell voltages, state of charge, operating sequence, driving range, total battery pack voltage and total battery pack current, and battery pack temperature. Battery monitoring data is obtained in real time by reading the monitoring indicators reported by the battery management system through the vehicle-mounted remote service terminal via the CAN bus.
[0014] As a feasible preferred solution, the discharge segment is divided as follows: when the charging state switches from charging or idling to discharging, and the discharging state is continuously maintained for at least a preset number of sampling points, it is determined as the start point of the discharge stage; when the charging state switches back from discharging to charging or idling, or when data monitoring is continuously interrupted for more than a preset number of sampling points, it is determined as the end point of the discharge stage; the data interval from the start point to the end point of the discharge stage is extracted to form a discharge segment.
[0015] As a feasible and preferred approach, the individual voltage timing matrix is constructed, specifically including: Extract the individual cell voltage string at each time step in each discharge segment, and divide it into a string list using a delimiter; convert each string element in the string list into numerical data. If the conversion fails, use the voltage value of the individual cell at the adjacent time steps to perform linear interpolation to fill the gap; stack the numerical lists of all time steps in chronological order to construct the individual cell voltage time series matrix.
[0016] As a feasible and preferred approach, the relative voltage timing matrix is calculated, specifically including: Using the median voltage value of all individual cells at each time step as a benchmark, the relative voltage value of each individual cell is calculated, and a relative voltage time series matrix is constructed. The calculation formula is as follows:
[0017] in, Let be the relative voltage value of the i-th individual cell at the t-th time step. This is the original voltage value. V median(t) This represents the median of the voltage values of all individual cells at the t-th time step.
[0018] As a feasible preferred option, the parameters of variational mode decomposition include a penalty factor α, the number of mode decompositions K, the DC component suppression coefficient DC, and the convergence error tol; wherein, the number of mode decompositions K is adaptively adjusted according to the complexity of the actual vehicle signal using an adaptive strategy; VMD decomposition is performed on the relative voltage time series signal of each individual cell in the relative voltage time series matrix; The mathematical principle of VMD decomposition is: The input signal f(t) is decomposed into K eigenmode components, each of which is amplitude-modulated and frequency-modulated around a center frequency. The decomposition process is achieved by solving the following constrained optimization problem: ; The constraints are: ; In the formula, {u k} is obtained from decomposition K Each intrinsic mode component, {w k} represents the center frequency corresponding to each mode. δ ( t ) is the Dirac function, This represents the convolution operation. j The imaginary unit; Solving the above optimization problem yields the corresponding values for each individual cell. K Each intrinsic mode component Calculate the energy percentage of each intrinsic mode component, identify modes with energy percentages below a preset threshold as noise-dominant modes and remove them, retain the remaining effective modes, and form the mode feature sequence of each cell. To eliminate noise-dominant modes, the energy percentage of each intrinsic mode component is calculated, and modes with an energy percentage lower than a preset threshold are identified as noise-dominant modes and eliminated.
[0019] As a feasible preferred embodiment, the bidirectional LSTM-multi-head attention feature extraction network includes an input layer, a bidirectional LSTM layer, a multi-head attention layer, a Dropout layer, and a fully connected layer; the bidirectional LSTM layer has 2 to 4 network layers, a hidden layer dimension of 32 to 128, and uses ReLU as the activation function; the multi-head attention layer has 2 to 8 attention heads.
[0020] As a feasible preferred embodiment, the autoencoder includes an input layer, an encoding layer, and a decoding layer; the encoding layer has 1 to 2 hidden layers with a dimension of 4 to 16, and the activation function is ReLU; the output dimension of the decoding layer is the same as the input layer dimension, and the activation function is linear activation. The autoencoder is trained using the Adam optimizer and the mean squared error loss function. The feature matrix is used as both the input and label of the model for unsupervised training. The formula for calculating the outlier score is:
[0021] in, ( j ) is the eigenvector of the i-th individual entity. j Each latitude value, Reconstruct its value, P The dimension of the feature vector.
[0022] As a feasible and preferred approach, the anomaly determination threshold is adaptively determined using the interquartile range method, specifically by calculating the 20th quartile (Q1) and 80th quartile (Q3) of the anomaly score to obtain the interquartile range. ; The adaptive anomaly detection threshold is: ; If the abnormal score of any single cell exceeds the threshold, it is determined to be an abnormal cell, and an early warning signal for thermal runaway is triggered.
[0023] Secondly, this invention also proposes a deep learning-based early warning system for thermal runaway of a single power battery cell, which utilizes the aforementioned deep learning-based early warning method for thermal runaway of a single power battery cell. Attached Figure Description
[0024] Figure 1 This is a logical diagram of a deep learning-based early warning method for thermal runaway of a single power battery cell.
[0025] Figure 2 This is a schematic diagram of the RV curve for each discharge segment of the battery pack in the thermal runaway vehicle.
[0026] Figure 3 This is a schematic diagram of the diagnostic results for thermal runaway of vehicle A.
[0027] Figure 4 This is a schematic diagram of the diagnostic results for thermal runaway of vehicle B.
[0028] Figure 5 This is a schematic diagram showing the normal diagnostic results for vehicle C. Detailed Implementation
[0029] To make the technical solution and advantages of this application clearer, the technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. It is understood that the specific embodiments described herein are only some embodiments of the present invention, and are only used to explain this application, not to limit it. It should be noted that the technical features or combinations of technical features described in the following embodiments should not be considered isolated; they can be combined with each other to achieve better technical effects. The same reference numerals appearing in the accompanying drawings of the following embodiments represent the same features or components, and can be applied to different embodiments.
[0030] Furthermore, unless otherwise defined, the technical or scientific terms used in this invention description shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains.
[0031] The present invention will now be described in further detail with reference to the accompanying drawings.
[0032] Reference Figure 1 This disclosure provides a deep learning-based method for early warning of thermal runaway in a single power battery cell, comprising: Step S100: Real-vehicle big data collection and standardized discharge segment division, including: Step S101: Real-vehicle monitoring data acquisition. This involves collecting core monitoring data during the actual operation of the new energy vehicle. The data acquisition terminal is an onboard remote service terminal (T-BOX), which reads various monitoring indicators reported by the battery management system (BMS) in real time via the CAN bus. Data fields cover key indicators such as the voltage of each individual cell in the battery pack (CELL_VOLTAGE), charging status (CHARGE_STATUS), runtime sequence (TIMESTAMP), driving range (MILEAGE), total battery pack voltage (PACK_VOLTAGE), total battery pack current (PACK_CURRENT), and battery pack temperature (PACK_TEMP). The individual cell voltage data is recorded as comma-separated strings, such as "3.215,3.218,3.212,3.219,...,3.198", with each string containing the voltage values of all individual cells in the battery pack at that time step. The charging status is represented by time-series labels, with values including discrete labels such as "Charging" (corresponding to CHARGE_STATUS=1), "Discharging" (corresponding to CHARGE_STATUS=0), and "Still" (corresponding to CHARGE_STATUS=2). The runtime sequence is recorded in Unix timestamp or standard date and time format with precision at the second or millisecond level.
[0033] The data acquisition frequency is set to 1Hz, meaning a complete battery pack monitoring data is collected once per second. For a passenger vehicle with an average daily mileage of 50km, approximately 86,400 data records can be collected per day (calculated over 24 hours); for a city bus with an average daily mileage of 200km, the number of data records that can be collected per day is even greater. The data is uploaded to the cloud server via a 4G / 5G communication module and stored in CSV or JSON format. The raw data contains fields that are stored in a mix of character and numeric types, where the individual cell voltage field is a character string, the charging status field is a character label, and the remaining fields are numeric.
[0034] Step S102 involves preliminary cleaning of the collected raw vehicle data, including: Invalid flag removal involves iterating through all data records to identify and remove invalid flags from the individual voltage field, including but not limited to null values (NULL), missing values (NaN), abnormal characters (such as "#N / A", "ERROR"), and out-of-range flags (such as obvious error values like -1 or 0). For the CHARGE_STATUS field, records with values outside the preset valid set {0,1,2} are removed.
[0035] Special character filtering is performed on individual voltage strings using regular expressions to remove invisible characters such as spaces, tabs, and newlines. For example, "3.215, 3.218" is cleaned into "3.215,3.218".
[0036] Timestamp alignment is performed by sorting all data records in ascending order by the TIMESTAMP field and checking the continuity of timestamps. If the time difference between two adjacent records exceeds a preset threshold (e.g., 5 seconds), it is determined that the data transmission is interrupted, and a breakpoint is marked at that position. This breakpoint will be used as the boundary when dividing subsequent discharge segments.
[0037] The format is standardized by converting all cleaned numeric fields to 32-bit floating-point format (float32) and timestamps to Unix timestamp format (unit: seconds) to ensure data type consistency in subsequent calculations.
[0038] After cleaning, the remaining valid data records constitute the cleaned dataset, which retains pure numerical values and time format information, avoiding interference from invalid characters in subsequent calculations.
[0039] Step S103: Based on the time-series changes in battery charging state and the continuous monitoring interval of individual cell voltage, the discharge operation stage is standardized and defined. The specific division rules are as follows: The discharge phase start point is determined by iterating through the CHARGE_STATUS field in the cleaned dataset. When CHARGE_STATUS changes from "Charging" (CHARGE_STATUS=1) or "Still" (CHARGE_STATUS=2) to "Discharging" (CHARGE_STATUS=0), and this state is maintained continuously for at least 3 sampling points (i.e., at least 3 seconds), that moment is determined as the start point of the discharge phase. Setting a minimum duration of 3 seconds is to avoid misjudgments caused by instantaneous state fluctuations.
[0040] The discharge phase termination point is determined when CHARGE_STATUS switches back from "discharging" (CHARGE_STATUS=0) to "charging" (CHARGE_STATUS=1) or "stationary" (CHARGE_STATUS=2), or when data monitoring is continuously interrupted (more than 5 consecutive sampling points are missing).
[0041] Discharge segments are generated by extracting corresponding data intervals from the cleaned dataset, using the start point of the discharge phase as the starting index and the end point of the discharge phase as the ending index. Each discharge segment contains information such as the voltage strings of all individual cells, charging status, and runtime sequence within that time period.
[0042] Validity screening involves filtering the generated discharge segments and retaining only those that meet the following condition: the discharge duration is no less than 60 seconds (i.e., the time step). T ≥60); the integrity rate of single-cell voltage data within the segment is not less than 95%; the number of single-cell batteries included in the segment N No fewer than 12 (covering mainstream battery pack configurations).
[0043] After the above division, several continuous discharge segments are obtained. Each discharge segment is a complete set of real vehicle discharge condition monitoring data. Its time step T ranges from 200 to 800. The number of individual batteries N is determined according to the vehicle configuration (usually 12 to 100+). All discharge segments together constitute the basic data set for subsequent diagnosis.
[0044] Step S200, Discharge segment data preprocessing and relative voltage feature extraction, includes: Step S201, Individual voltage timing matrix construction: For each discharge segment obtained in step S103, perform the following data format normalization operation: Extract the individual cell voltage string at each time step t in each discharge segment. This string is in the form of "3.215,3.218,3.212,...,3.198". Separate the strings using commas to obtain a list of strings.
[0045] For each string element in the list of strings, a floating-point conversion function (such as the float() function in Python) is used to convert it into numeric data, resulting in a list of numeric values. During the conversion process, if the conversion of a numeric data fails (e.g., the string is empty or not a number), linear interpolation is performed using the voltage values of the individual battery at adjacent time steps to fill the gap.
[0046] Matrix construction involves stacking the lists of values for all time steps in chronological order to construct the time-series matrix of the individual cell voltage corresponding to the discharge segment. V∈RT×N The row index of the matrix represents the discharge time step. t ( t =1,2,...,T), the column index is the individual battery number i (i=1,2,...,N). This represents the initial voltage value of the i-th individual cell at the t-th time step, in volts (V).
[0047] For example, for a discharge segment with a time step T=500 and a number of individual cells N=96, the constructed individual cell voltage time series matrix is... V The dimension is 500×96, which means it contains 500 time steps and 96 individual cell voltage values for each time step.
[0048] Step S202, for each discharge segment, the individual voltage timing matrix V The specific steps for calculating the relative voltage characteristics are as follows: For each row of the matrix cell voltage time series matrix V (i.e., each time step t), calculate the median of all N cell voltage values in that row, denoted as . V median(t) This median can characterize the normal trend of the overall voltage change of the battery pack at the current time step t, effectively avoiding the interference of the overall charge and discharge state switching of the battery pack on the characteristics of individual cells.
[0049] Using this median as a benchmark, the relative voltage value of each individual cell at the corresponding time step is calculated, and the relative voltage time series matrix RV is constructed. The calculation formula is as follows:
[0050] in, This represents the relative voltage value of the i-th individual cell at the t-th time step, in volts (V).
[0051] The physical significance of this relative voltage characteristic lies in its ability to eliminate the interference of overall battery pack voltage fluctuations on individual cell characteristics. For example, when the overall battery pack voltage drops from 3.5V to 3.2V, the voltages of all individual cells will drop synchronously. However, if a single cell experiences a voltage drop greater than the overall trend due to an internal micro-short circuit, its relative voltage value will show a significant negative deviation, thus being highlighted by the relative voltage characteristic. Similarly, if a single cell experiences increased internal resistance and lower voltage due to aging, its relative voltage value will also continuously show a negative deviation. Therefore, the relative voltage characteristic can intuitively reflect the deviation between the voltage change trend of individual cells and the overall battery pack, highlighting the voltage differences and consistency between individual cells, providing a core foundational feature for subsequent anomaly identification.
[0052] Step S300, Variational Mode Decomposition (VMD) denoising and modal feature extraction: The relative voltage signal of a real vehicle inevitably contains sensor noise, data transmission noise, and operating condition fluctuation interference. VMD decomposition is needed to achieve signal denoising and effective feature extraction, including: Step S301, VMD parameter settings. The core parameters of VMD are set as follows: The penalty factor α ranges from [1000, 5000]. This parameter is used to balance the fidelity and sparsity of each modal component after signal decomposition. The larger α is, the sparser the modal components are, but effective signal details may be lost; the smaller α is, the smoother the modal components are, but too much noise may be retained.
[0053] The modal decomposition number K has a value range of [2, 15]. In this embodiment, an adaptive strategy is used to determine the value of K, which is adaptively adjusted according to the complexity of the actual vehicle signal.
[0054] The DC component suppression coefficient (DC) is set to 0 in this embodiment, meaning that the DC component is not retained. Since the relative voltage characteristics have eliminated the overall voltage offset, the signal should not contain a DC component; retaining the DC component would introduce redundant information.
[0055] The convergence error tol has a range of
[10] . -7 , 10 -6 ], that is, when the L2 norm of the changes in each modal component in two adjacent iterations is in [10 -7 , 10 -6 When the condition is met, VMD decomposition is considered converged.
[0056] Step S302, VMD decomposition is performed on the relative voltage timing matrix. RV The relative voltage timing signal RV(:,i) of each individual cell (i.e., the i-th column of the matrix, i=1,2,...,N) is decomposed into VMD.
[0057] The mathematical principle of VMD decomposition is as follows: the input signal f(t) is decomposed into K eigenmode components, and each eigenmode component is amplitude-modulated and frequency-modulated around the center frequency. The decomposition process is achieved by solving the following constrained optimization problem: ; The constraints are: ; In the formula, {u k} is obtained from decomposition K Each intrinsic mode component, {w k} represents the center frequency corresponding to each mode. δ ( t ) is the Dirac function, This represents the convolution operation. j It is the imaginary unit.
[0058] The physical meaning of this optimization problem is: by demodulating each modal component to the baseband (multiplied by e) -jw_k t) After that, the L2 norm of the gradient of the analytic signal is calculated so that each modal component is as compact as possible in the time domain (i.e., with finite bandwidth), while the sum of all modal components accurately reconstructs the original signal.
[0059] The above optimization problem is solved by the alternating direction multiplier method to obtain the corresponding values for each individual cell. KEach intrinsic mode component is analyzed, and noise-dominant modes with an energy percentage below 15% are removed, leaving only the remaining effective modes. This results in the mode feature sequence U for each individual cell. i (i=1,2,...,N), to achieve denoising and frequency domain feature extraction of real vehicle signals.
[0060] Step S303, noise mode removal and effective mode feature sequence construction, includes: For each cell i, the energy percentage of each of the K modal components is calculated. Modes with an energy percentage below 15% are identified as noise-dominant modes and are removed. For example, if K=6, the calculated energy percentages of each mode are [0.35, 0.28, 0.18, 0.12, 0.05, 0.02]. Modes 4, 5, and 6 (with energy percentages of 12%, 5%, and 2%, respectively, all below the 15% threshold) are removed, and the first three effective modes are retained.
[0061] The retained valid modes are arranged in the original index order to form the modal feature sequence for each individual cell i. To facilitate subsequent network input, the modal feature sequence is transposed, i.e., the time step comes first and the modal dimension comes last.
[0062] Step S400, Bidirectional LSTM-Multi-head Attention Deep Feature Extraction, includes: Step S401, Network architecture design and parameter initialization, including: A bidirectional LSTM-multi-head attention feature extraction network is constructed to capture deep temporal correlations and key features from modal feature sequences. The network structure consists of an input layer, a bidirectional LSTM layer, a multi-head attention layer, a Dropout layer, and a fully connected layer. The parameters of each layer are set as follows: Input layer: The input dimension is the modality decomposition number, and the input data format is the modal feature sequence of each individual cell; Bidirectional LSTM layer: The number of network layers N∈ [2, 4]; Hidden layer: Dimension M∈ [32, 128]; The activation function is ReLU; The dropout probability p∈[0.2, 0.5] of the dropout layer is used to prevent model overfitting; The number of attention heads H∈[2, 8] of the multi-head attention layer, and the key vector dimension (Ensure dimensionality matching for multi-head parallel computing).
[0063] Step S402: Feature extraction is performed, including: For each individual cell i (i=1,2,...,N) in the current discharge segment, perform the following operations: The modal characteristic sequence U of the single cell i The temporal dimensions are rearranged and converted into a T×K input format, which is then fed into a bidirectional LSTM-multi-head attention network.
[0064] The rearranged sequence is then input into a bidirectional LSTM layer. Through parallel computation of forward LSTM units (capturing future temporal correlations) and backward LSTM units (capturing historical temporal correlations), the bidirectional temporal fusion feature F is output. LSTM Key feature enhancement is achieved through parallel weight allocation of multiple attention points. The calculation process is as follows:
[0065]
[0066] This network fully captures the forward and backward temporal correlations of battery voltage mode features through bidirectional LSTM, and adaptively strengthens the weights of key temporal features through a multi-head attention mechanism, enabling it to extract more discriminative high-dimensional robust features from real vehicle time-series data.
[0067] Step S500: Construct an autoencoder model and calculate anomaly scores, including: Step S501: Autoencoder model construction. An unsupervised autoencoder model is constructed based on the feature matrix obtained in step S402. The autoencoder structure includes an input layer, an encoding layer, and a decoding layer. The dimension of the input layer is consistent with the dimension P of the high-dimensional robust feature vector.
[0068] The encoding layer has 1-2 hidden layers with dimensions D∈ [4, 16] and ReLU is used as the activation function.
[0069] The output dimension of the decoding layer is the same as the input dimension P, and the activation function is linear activation (i.e., no activation function).
[0070] Step S502, unsupervised training of the autoencoder, includes: The Adam optimizer was chosen as the model optimizer, with an initial learning rate set to 1×10⁻⁶. -3 The loss function used is mean squared error (MSE).
[0071] During training, using the feature matrix X It also serves as both input and label for the model (i.e., unsupervised training, no fault labels required).
[0072] Unsupervised learning enables the decoding layer to accurately reconstruct input features.
[0073] The training objective is to minimize the global reconstruction loss.
[0074] In the formula, x i (j) represents the j-th dimension value of the i-th individual feature vector. Reconstruct its value; After training, the autoencoder learns to accurately reconstruct the feature vectors of normal individual cells, but for abnormal individual cells (whose feature vectors deviate from the normal pattern), the reconstruction error will increase significantly.
[0075] Step S503: After training is complete, convert the high-dimensional feature vectors of each individual unit... The input is fed into the trained autoencoder to obtain the corresponding reconstructed feature vector. The mean squared error of both is calculated as the outlier score for that individual. The calculation formula is as follows:
[0076] The physical meaning of the anomaly score is the degree of deviation between the characteristics of a single cell and the normal characteristic pattern of the battery pack as a whole. The higher the score, the more abnormal the state of the single cell and the higher the risk of thermal runaway.
[0077] Step S600, achieving early warning of thermal runaway through interquartile range threshold determination, includes: Step S601: Statistical analysis of outlier scores. Statistical analysis was performed on the abnormal scores of all N individual cells in the current discharge segment: Calculate the 20th percentile Q1 and 80th percentile Q3 of the outlier scores, where Q1 is the value at the 20th percentile after the outlier scores are sorted in ascending order (Q1=percentile(AnomalyScore,20)) and Q3 is the value at the 80th percentile (Q3=percentile(AnomalyScore,80)).
[0078] Solve for the interquartile range: ; The interquartile range reflects the dispersion of the middle 50% of outlier scores. It is a robust statistic for measuring the spread of data distribution and is not affected by extreme values.
[0079] Step S602: Adaptive threshold calculation and anomaly detection, including: The anomaly detection threshold is set as follows: ; The threshold is adaptively determined through statistical methods, which can effectively adapt to the abnormal score distribution characteristics under different real vehicle operating conditions and avoid the problem of poor adaptability caused by manually setting the threshold.
[0080] The anomaly score for each individual is compared with a threshold. If AnomalyScore(i) > Threshold, then the i-th monomer is determined to be an anomalous monomer.
[0081] Step S603: If one or more abnormal cells appear in the current discharge segment, the early warning signal for thermal runaway is triggered.
[0082] Simultaneously, abnormal information is recorded, including: discharge segment index (identifying the sequence number of the discharge segment in the entire day's data), abnormal individual number ( i The data includes the anomaly score (AnomalyScore(i),) the corresponding runtime sequence (TIMESTAMP), the relative voltage curve of the cell, and the modal characteristic curve, providing accurate information for subsequent battery maintenance and troubleshooting.
[0083] To verify the effectiveness of the method in this embodiment, cloud data from three new energy vehicles were used as an example for testing. The data collection fields were based on the requirements of the national standard GB / T 32960, including one ternary lithium battery thermal runaway vehicle (vehicle A, passenger car, battery pack with 96 cells), one lithium iron phosphate battery thermal runaway vehicle (vehicle B, passenger car, battery pack with 96 cells), and one normal vehicle (vehicle C, city bus, battery pack with 192 cells), basically covering various operating conditions.
[0084] Figure 2 The RV curves of each discharge segment of the battery pack in the thermal runaway vehicle are shown. Figure 3 The diagnostic results for thermal runaway of vehicle A were presented. Figure 4 The diagnostic results for the vehicle's thermal runaway were presented. Figure 5 The diagnostic results for vehicle C are shown to be normal. Combined with vehicle disassembly results, this method achieves early warning of thermal runaway (≥7 days) under a given threshold, achieving high accuracy and low false alarm and false negative rates.
[0085] This disclosure also provides a deep learning-based early warning system for thermal runaway of a single power battery cell, which utilizes the aforementioned deep learning-based early warning method for thermal runaway of a single power battery cell.
[0086] The above content is merely an embodiment of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can improve and implement this solution based on the guidance provided in this application and their own capabilities. Typical well-known structures or operating methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.
Claims
1. A method for early warning of thermal runaway in a single power battery cell based on deep learning, characterized in that, include: Collect battery monitoring data during actual vehicle operation and divide discharge segments based on the time-series changes in battery charging status; For each discharge segment, construct the individual voltage timing matrix and calculate the relative voltage timing matrix; Variational mode decomposition is performed on the relative voltage time-series signal of each individual cell to obtain the intrinsic mode components corresponding to each individual cell. The noise-dominant modes are removed, and the effective modes are retained to form the mode feature sequence of each individual cell. The modal feature sequence of each individual cell is input into a bidirectional LSTM-multi-head attention feature extraction network, which outputs a high-dimensional robust feature vector. An unsupervised autoencoder model is constructed using high-dimensional robust feature vectors. The high-dimensional robust feature vectors of each individual cell are input into the autoencoder, and the reconstruction error is calculated as the anomaly score of that individual cell. Statistical analysis is performed on the abnormal scores of all individual cells in the current discharge segment to determine the adaptive abnormal judgment threshold. Based on the comparison between the abnormal scores and the adaptive abnormal judgment threshold, an early warning is issued for thermal runaway of individual power battery cells.
2. The method for early warning of thermal runaway of a single power battery cell based on deep learning according to claim 1, characterized in that, Battery monitoring data includes individual cell voltage, charging status, operating sequence, driving range, total battery pack voltage, total battery pack current, and battery pack temperature; Battery monitoring data is obtained in real time by reading the monitoring indicators reported by the battery management system through the vehicle-mounted remote service terminal via the CAN bus.
3. The method for early warning of thermal runaway of a single power battery cell based on deep learning according to claim 1, characterized in that, The discharge segment is divided as follows: when the charging state switches from charging or resting to discharging, and the discharging state is continuously maintained for at least a preset number of sampling points, it is determined as the start point of the discharge stage; when the charging state switches back from discharging to charging or resting, or when data monitoring is continuously interrupted for more than a preset number of sampling points, it is determined as the end point of the discharge stage; the data interval from the start point to the end point of the discharge stage is extracted to form a discharge segment.
4. The method for early warning of thermal runaway of a single power battery cell based on deep learning according to claim 1, characterized in that, Constructing the individual voltage timing matrix specifically includes: Extract the individual cell voltage string at each time step in each discharge segment, and divide it into a string list using a delimiter; convert each string element in the string list into numerical data. If the conversion fails, use the voltage value of the individual cell at the adjacent time steps to perform linear interpolation to fill the gap; stack the numerical lists of all time steps in chronological order to construct the individual cell voltage time series matrix.
5. The method for early warning of thermal runaway of a single power battery cell based on deep learning according to claim 1, characterized in that, Calculate the relative voltage time series matrix, specifically including: Using the median voltage value of all individual cells at each time step as a benchmark, the relative voltage value of each individual cell is calculated, and a relative voltage time series matrix is constructed. The calculation formula is as follows: in, Let be the relative voltage value of the i-th individual cell at the t-th time step. This is the original voltage value. V median(t) This represents the median of the voltage values of all individual cells at the t-th time step.
6. The method for early warning of thermal runaway of a single power battery cell based on deep learning according to claim 1, characterized in that, The parameters of variational mode decomposition include the penalty factor α, the number of mode decompositions K, the DC component suppression coefficient DC, and the convergence error tol; wherein, the number of mode decompositions K adopts an adaptive strategy to be adaptively adjusted according to the complexity of the actual vehicle signal; VMD decomposition is performed on the relative voltage time series signal of each individual cell in the relative voltage time series matrix; The mathematical principle of VMD decomposition is: The input signal f(t) is decomposed into K eigenmode components, each of which is amplitude-modulated and frequency-modulated around a center frequency. The decomposition process is achieved by solving the following constrained optimization problem: ； The constraints are: ； In the formula, {u k } is obtained from decomposition K Each intrinsic mode component, {w k } represents the center frequency corresponding to each mode. δ ( t ) is the Dirac function, This represents the convolution operation. j The imaginary unit; Solving the above optimization problem yields the corresponding values for each individual cell. K Each intrinsic mode component Calculate the energy percentage of each intrinsic mode component, identify modes with energy percentages below a preset threshold as noise-dominant modes and remove them, retain the remaining effective modes, and form the mode feature sequence of each cell. To eliminate noise-dominant modes, the energy percentage of each intrinsic mode component is calculated, and modes with an energy percentage lower than a preset threshold are identified as noise-dominant modes and eliminated.
7. The method for early warning of thermal runaway of a single power battery cell based on deep learning according to claim 1, characterized in that, The bidirectional LSTM-multi-head attention feature extraction network includes an input layer, a bidirectional LSTM layer, a multi-head attention layer, a Dropout layer, and a fully connected layer; the bidirectional LSTM layer has 2 to 4 layers, a hidden layer dimension of 32 to 128, and uses ReLU as the activation function; the multi-head attention layer has 2 to 8 attention heads.
8. The method for early warning of thermal runaway of a single power battery cell based on deep learning according to claim 1, characterized in that, The autoencoder includes an input layer, an encoding layer, and a decoding layer; the encoding layer has 1 to 2 hidden layers with a dimension of 4 to 16, and the activation function is ReLU; the output dimension of the decoding layer is the same as the input layer dimension, and the activation function is linear activation. The autoencoder is trained using the Adam optimizer, with mean squared error as the loss function, and unsupervised training is performed using the feature matrix as both the input and label of the model. The formula for calculating the outlier score is: in, ( j ) is the eigenvector of the i-th individual entity. j Each latitude value, Reconstruct its value, P The dimension of the feature vector.
9. The method for early warning of thermal runaway of a single power battery cell based on deep learning according to claim 1, characterized in that, The anomaly threshold is adaptively determined using the interquartile range method by calculating the 20th quartile (Q1) and 80th quartile (Q3) of the anomaly score to obtain the interquartile range. ; The adaptive anomaly detection threshold is: ; If the abnormal score of any single cell exceeds the threshold, it is determined to be an abnormal cell, and an early warning signal for thermal runaway is triggered.
10. A deep learning-based early warning system for thermal runaway of a single power battery cell, characterized in that: The method for early warning of thermal runaway of a single power battery cell based on deep learning, as described in any one of claims 1-9, is employed.