Battery cell thermal runaway detection method and device based on CCS data collection

Abstract

The application discloses a cell thermal runaway detection method and device based on CCS collected data driving, relates to the field of battery management, and comprises the following steps: acquiring voltage time sequence data and temperature time sequence data of multiple cells synchronously collected by a cell connection assembly CCS, and acquiring cell spatial arrangement information, working condition parameters and battery health state; determining voltage spatial gradients and temperature spatial gradients of each cell relative to adjacent cells according to the spatial arrangement information; constructing a multi-dimensional time sequence feature matrix in combination with time sequence change characteristics and spatial gradients of voltage and temperature; extracting a feature vector and an attention weight through a time sequence transformation encoder model; and determining an abnormal score of each cell by using an isolated forest sub-model corresponding to a current composite working condition category; determining a reference abnormal threshold value according to the battery health state to obtain an adaptive abnormal threshold value; determining a thermal runaway detection result according to a comparison result of the abnormal score and the adaptive abnormal threshold value, and executing continuous window verification, hierarchical early warning and safety protection strategies.

Description

Technical Field

[0001] This application relates to the field of battery safety monitoring and battery management technology, and in particular to a method and device for detecting thermal runaway of battery cells driven by data collected from the cell connection component CCS. Background Technology

[0002] The Cell Connector System (CCS) is installed within the battery module of a power battery or energy storage battery to enable electrical connections between multiple cells and to collect voltage and temperature data for each cell. The Battery Management System (BMS) can monitor the operating status of the battery module based on the data collected by the CCS and issue alarms or implement safety protection measures when a cell experiences an internal short circuit, abnormal heating, or thermal runaway risk.

[0003] Existing testing methods typically set fixed thresholds for the absolute voltage, absolute temperature, or rate of temperature rise of individual battery cells. When a cell is subjected to high-rate charge / discharge, at the state of charge boundary, or in high or low temperature environments, voltage and temperature may fluctuate normally. As the battery's health deteriorates, the cell's internal resistance, temperature rise, and voltage response characteristics will also change. Therefore, consistently using the same fixed threshold can easily misinterpret normal fluctuations as abnormalities, or fail to adapt to changes in abnormal behavior after cell aging.

[0004] Furthermore, in the early stages of thermal runaway, a single cell may not have reached the preset absolute threshold, exhibiting only continuous changes relative to its own historical state or local deviations relative to physically adjacent cells. If only the absolute data of a single cell is analyzed independently without utilizing the spatial differences between multiple adjacent cells, it is difficult to distinguish between the common fluctuations caused by changes in the overall operating conditions of the battery module and the local anomalies generated by a single cell.

[0005] Therefore, a cell thermal runaway detection scheme is needed, which, based on the continuous voltage and temperature data of multiple cells, further utilizes the spatial differences between physically adjacent cells, and enables the anomaly judgment criteria to adapt to changes in battery operating conditions and battery health status. Summary of the Invention

[0006] This application provides a cell thermal runaway detection method and device based on CCS data acquisition, which improves the ability to distinguish between overall fluctuations of the battery module and local anomalies of a single cell under different operating conditions and different battery health states.

[0007] Firstly, this application provides a cell thermal runaway detection method driven by CCS acquisition data, which can be executed by a battery management system. The battery management system acquires voltage and temperature time-series data of multiple cells within a battery module synchronously acquired by the cell connection component CCS, and obtains spatial arrangement information, operating parameters, and battery health status of the multiple cells; determines the voltage and temperature spatial gradients of each cell relative to adjacent cells based on the spatial arrangement information; determines anomaly scores for each cell based on the voltage and temperature time-series data, voltage and temperature spatial gradients; determines a baseline anomaly threshold based on the battery health status, and corrects the baseline anomaly threshold based on the operating parameters to obtain an adaptive anomaly threshold; and determines the thermal runaway detection result of each cell based on the comparison between the anomaly score of each cell and the adaptive anomaly threshold.

[0008] In the above method, by synchronously acquiring the voltage and temperature of multiple cells using CCS, continuous state data of multiple cells under the same sampling time reference can be obtained. By determining the voltage spatial gradient and temperature spatial gradient of the current cell relative to its adjacent cells, the state deviation of the current cell relative to its local physical neighborhood can be characterized. When the overall operating conditions of the battery module change, multiple cells usually have a common trend of change, while a local anomaly in a single cell is prone to deviation relative to its adjacent cells. Therefore, combining time-series data and spatial gradients to determine anomaly scores is beneficial for distinguishing between overall common fluctuations and local anomalies.

[0009] Determining a baseline anomaly threshold based on battery health status allows the anomaly assessment benchmark to correspond to the battery's lifecycle state. Adjusting the baseline anomaly threshold based on operating parameters adapts the anomaly assessment standard to the current charge / discharge rate, state of charge, and ambient temperature. In one optional implementation, upper and lower threshold limits can be set for the adjusted result to ensure that the adaptive anomaly threshold matches the range of anomaly scores.

[0010] In one possible design, the set of adjacent cells corresponding to each cell is determined based on the spatial arrangement information. The average value of the voltage data of each adjacent cell in the set of adjacent cells is determined as the neighborhood voltage reference value, and the difference between the current cell voltage data and the neighborhood voltage reference value is determined as the voltage spatial gradient. Correspondingly, the average value of the temperature data of each adjacent cell in the set of adjacent cells is determined as the neighborhood temperature reference value, and the difference between the current cell temperature data and the neighborhood temperature reference value is determined as the temperature spatial gradient.

[0011] In one possible design, voltage timing difference and temperature timing difference are determined based on the voltage and temperature data of each cell at adjacent sampling times. Then, voltage timing data, temperature timing data, voltage timing difference, temperature timing difference, voltage spatial gradient, and temperature spatial gradient are combined according to the sampling time order to construct a multi-dimensional timing feature matrix.

[0012] In one possible design, the multi-dimensional temporal feature matrix is ​​mapped to feature dimensions and injected with positional encoding; the correlation between different sampling times and different feature dimensions is determined through a multi-head self-attention network; temporal encoded features are obtained through a feedforward network, residual connections, and layer normalization; the temporal encoded features are pooled to obtain feature vectors, and attention weight vectors corresponding one-to-one with each feature dimension of the feature vector are determined based on the attention response of each feature dimension in the temporal encoding process.

[0013] In one possible design, the feature dimensions in the feature vector are weighted according to the attention weight vector to obtain a weighted feature vector; the current working condition category is determined according to the working condition parameters, and the corresponding target isolated forest sub-model is selected from multiple isolated forest sub-models; the anomaly score is determined according to the path length of the weighted feature vector in the multiple isolated trees of the target isolated forest sub-model.

[0014] In one possible design, the current life cycle state of the battery is determined as a healthy stage, a decay stage, or an aging stage based on the battery's health status, and a corresponding baseline abnormal threshold is determined. Correction parameters are determined based on the charge / discharge rate, state of charge, and ambient temperature, and then the operating condition correction coefficient is determined. The baseline abnormal threshold is corrected using the operating condition correction coefficient, and the correction result is limited to between the lower limit and the upper limit of a preset threshold.

[0015] In one possible design, multiple warning levels are set according to an adaptive anomaly threshold, corresponding to anomaly scoring intervals; the initial anomaly level of each cell within the current sliding time window is determined; when a target cell has the same initial anomaly level within at least a preset number of sliding time windows in multiple consecutive sliding time windows, the final anomaly level of the target cell is determined, and the cell identifier of the target cell is output. In a more stringent verification method, the preset number of sliding time windows may be further required to be adjacent to each other.

[0016] In one possible design, sampling frequency adjustment and temperature pre-regulation, charging and discharging power limitation and forced cooling, high voltage power-off and emergency cooling are performed according to the first warning level, the second warning level and the third warning level, respectively.

[0017] In one possible design, the battery management system acquires voltage time-series data, temperature time-series data, operating parameters, battery health status, thermal runaway detection results, and safety protection strategy execution results throughout the battery's lifecycle to construct a model update sample set. Abnormal samples are labeled, and the labeled model update sample set is used to incrementally train the time-series encoder model, calibrate the isolated forest sub-model and the baseline abnormal threshold, and verify the detection accuracy of the updated model and threshold. When the preset verification conditions are met, the updated model and threshold are used for subsequent cell thermal runaway detection.

[0018] Secondly, this application provides a cell thermal runaway detection device driven by CCS acquisition data. The device is applied to a battery management system and includes a data acquisition module, a spatial gradient determination module, an anomaly scoring determination module, a threshold determination module, and a detection result determination module. The modules cooperate with each other to achieve the method described in the first aspect.

[0019] Compared to detection methods that only use absolute data of a single cell and a fixed threshold, this application utilizes the continuous temporal information of the cell itself and the spatial deviation information between adjacent cells, and makes judgments based on abnormal thresholds adapted to the battery health status and current operating conditions. This improves the operating condition adaptability and result reliability of cell thermal runaway detection, and provides a detection basis for abnormal cell location and subsequent graded safety protection. Attached Figure Description

[0020] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0021] Figure 1 This is a schematic diagram of the architecture of a battery cell thermal runaway detection system provided in an embodiment of this application; Figure 2 A flowchart of a cell thermal runaway detection method provided in this application embodiment; Figure 3 A schematic diagram of a cell spatial arrangement and an assembly of adjacent cells provided for an embodiment of this application; Figure 4 A flowchart illustrating the determination of voltage and temperature spatial gradients provided in this application embodiment; Figure 5 A schematic diagram illustrating the construction of a multidimensional temporal feature matrix provided in an embodiment of this application; Figure 6 A flowchart illustrating the processing of a timing-transform encoder model provided in an embodiment of this application; Figure 7 A schematic diagram illustrating the correspondence between working condition categories and isolated forest sub-models provided in this application embodiment; Figure 8 A flowchart illustrating the determination of an adaptive anomaly threshold is provided in an embodiment of this application. Figure 9 This application provides a schematic diagram of an anomaly level verification for a continuous sliding time window. Figure 10 A flowchart illustrating a graded security protection strategy provided in this application embodiment; Figure 11 A model update and rollback flowchart is provided for embodiments of this application; Figure 12 This is a schematic diagram of a battery cell thermal runaway detection device provided in an embodiment of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0023] In this application, "multiple" means two or more; "adjacent cells" refers to one or more cells that are marked as having an adjacency relationship with the current cell by spatial arrangement information according to the physical installation position of the cell in the battery module; "spatial arrangement information" can be represented by row and column numbers, spatial coordinates, adjacency matrix, or a mapping relationship between cell number and adjacent cell number.

[0024] The cell connection assembly (CCS) may include a flexible circuit board, copper or aluminum busbars, voltage acquisition contacts, a temperature sensor, signal conditioning circuitry, and analog-to-digital conversion circuitry. CCS synchronously acquires data from multiple cells, meaning that each acquisition channel uses the same sampling clock or is in the same sampling period after timestamp alignment, allowing for channel time deviations caused by hardware sampling order that are less than a preset synchronization error threshold.

[0025] like Figure 1 and Figure 2 As shown, the cell connection assembly (CCS) is physically integrated with the battery module and connected to the battery management system via the vehicle bus. Figure 1 In this diagram, N represents the total number of battery cells in the battery module, and cells 1, 2, to N are used to illustrate the multiple battery cells. The battery management system may include a processor, a memory, a communication interface, and a model inference unit. The memory stores spatial arrangement information, model parameters, threshold parameters, and detection programs.

[0026] The system comprises several data acquisition components: a voltage acquisition branch is formed by voltage signals from multiple battery cells via voltage acquisition contacts corresponding to each cell; a temperature acquisition branch is formed by temperature sensors from multiple battery cells; both voltage and temperature acquisition branches are processed by signal conditioning circuits and then input to an analog-to-digital converter (ADC); the ADC sends the converted digital acquisition data to the battery management system (BMS) via the vehicle bus; the BMS receives the digital acquisition data through a communication interface; its processor retrieves spatial arrangement information, model parameters, and threshold parameters stored in memory; and its model inference unit performs cell thermal runaway detection; the BMS also interacts with the thermal management system, the high-voltage main relay control unit, the vehicle controller, and the emergency call module through the communication interface.

[0027] Step S101: Acquire detection data. The battery management system acquires voltage and temperature time-series data of multiple cells synchronously collected by the CCS. Taking the i-th cell as an example, Vi(t) represents the voltage data of the i-th cell at sampling time t, and Ti(t) represents the temperature data of the i-th cell at sampling time t. The battery management system also acquires spatial arrangement information, operating parameters, and battery health status. The operating parameters are used to characterize the current operating conditions that affect the normal fluctuations of cell voltage or temperature, and may include at least one of charge / discharge rate, state of charge, and ambient temperature. In some embodiments, they may also include battery pack cooling status or battery pack current change status. The battery health status can be characterized by at least one of SOH, capacity retention rate, internal resistance change, or cumulative cycle count. This embodiment takes SOH as an example.

[0028] Before performing thermal runaway risk detection on the acquired data, channel validity checks can be performed. If a voltage or temperature acquisition channel experiences an open circuit, short circuit, continuous exceedance of the sensor's effective range, timestamp interruption, or communication verification failure, the corresponding data will be marked as invalid. Data anomalies caused by channel failures are not directly identified as cell thermal runaway risks; missing data can be handled using remaining valid channels, adjacent cell data, or data from the previous valid window.

[0029] Step S102: Determine the spatial gradient. The battery management system determines the adjacent cells corresponding to the i-th cell based on the spatial arrangement information, and determines the voltage spatial gradient and temperature spatial gradient respectively based on the voltage deviation and temperature deviation between the i-th cell and its adjacent cells at the same sampling time. When the spatial gradient retains its positive or negative sign, it can simultaneously represent the direction and degree of deviation; in implementations that only focus on the degree of deviation, the absolute value of the spatial gradient can also be used.

[0030] Step S103: Determine the anomaly score. The battery management system organizes voltage time-series data, temperature time-series data, voltage spatial gradient, and temperature spatial gradient according to the sampling time relationship into data objects that can reflect the current state, time changes, and spatial deviations. Anomaly scores are determined based on the degree of deviation of these data objects from the normal cell sample distribution. The anomaly score can be normalized according to a preset score range; a higher anomaly score indicates a greater deviation of the current cell from the normal sample distribution under the corresponding operating condition. Examples 3 to 5 provide a specific method for determining anomaly scores using a multi-dimensional time-series feature matrix, a time-series transform encoder model, and an isolated forest sub-model.

[0031] Step S104: Determine the adaptive anomaly threshold. The battery management system determines a baseline anomaly threshold based on the battery health status, and corrects the baseline anomaly threshold based on the current charge / discharge rate, state of charge, and ambient temperature to obtain an adaptive anomaly threshold. In an optional implementation, to prevent the threshold from exceeding the anomaly score range, a limiting process can be performed on the correction result.

[0032] Step S105: Determine the thermal runaway detection result. When the anomaly score is less than the adaptive anomaly threshold, it is determined that no anomaly reaching the preset risk level has been detected in the current window; when the anomaly score reaches or exceeds the adaptive anomaly threshold, it is determined that the corresponding cell has a thermal runaway risk. The thermal runaway detection result is used to characterize whether the cell has reached the preset thermal runaway risk level, and is not equivalent to confirming that the cell has experienced thermal runaway. Furthermore, the risk level can be determined according to the range in which the anomaly score falls.

[0033] Anomaly scores characterize the degree of deviation of the current feature vector from the normal sample distribution. To establish a correlation between anomaly scores and thermal runaway risk, confirmed anomaly samples used to calibrate the baseline anomaly threshold and verify test results must include at least thermal runaway precursor samples confirmed by internal short-circuit tests, local heating tests, overcharge tests, fault reproduction results, or maintenance and testing results. Thresholds are jointly calibrated using normal samples and thermal runaway precursor samples, so that when anomaly scores reach the threshold, it corresponds to a preset thermal runaway risk level.

[0034] Through the above steps, when multiple battery cells undergo similar changes synchronously due to charging / discharging or environmental variations, the spatial gradient remains relatively small; however, when a single battery cell experiences a local anomaly, its spatial deviation relative to adjacent cells increases. Therefore, overall common fluctuations and local anomaly deviations can be distinguished by combining timing characteristics.

[0035] As a verification implementation method, normal operation segment, overall ambient temperature rise segment, and single-cell local heating segment are selected within the same battery module. During overall ambient temperature rise, the temperature time series data of multiple cells rise synchronously, and the temperature spatial gradient of each cell relative to its adjacent cells remains within the normal fluctuation range of the neighborhood under the corresponding operating conditions. During single-cell local heating, the temperature spatial gradient of the target cell increases with the local heating process, and the anomaly score obtained from the multi-dimensional time series feature matrix reaches the corresponding adaptive anomaly threshold. Through the above comparison process, it can be verified that the combination of spatial gradient and time series features can be used to distinguish between the overall temperature rise of the battery module and the local temperature rise of a single cell.

[0036] like Figure 3 and Figure 4 As shown, cell (r, c) represents the cell located in row r and column c of the battery module, where r represents the row number, c represents the column number, and i represents the index of the cell to be detected. The set of adjacent cells of the i-th cell is denoted as Ω_i, and the number of adjacent cells in the set Ω_i that participate in the current spatial gradient calculation is denoted as M_i. Figure 3 On the left, cell (2,2) is taken as the current cell located inside the battery module, and its adjacent cells include cell (1,2), cell (2,1), cell (2,3) and cell (3,2). Figure 3 On the right, cell (1,2) is used as the boundary cell, and its adjacent cells include cells (1,1), (1,3), and (2,2). For cells located inside the battery module, cells in front, behind, to the left, and to the right can be selected as adjacent cells; for cells located at the boundary, the set of adjacent cells is determined based on the actual physical adjacent cells. Therefore, the number of adjacent cells for boundary cells and internal cells can be different.

[0037] At each sampling time, the battery management system acquires the voltage data of all effective neighboring cells corresponding to the current cell, calculates the arithmetic mean of the voltage data of all effective neighboring cells, and determines the average value as the neighborhood voltage reference value of the current cell.

[0038] The voltage spatial gradient of the current cell is obtained by subtracting the neighborhood voltage reference value from the voltage data of the current cell at the sampling time. The larger the absolute value of the voltage spatial gradient, the greater the deviation of the current cell's voltage from the average voltage of its neighboring cells.

[0039] Correspondingly, the arithmetic mean of the temperature data of all effective adjacent cells corresponding to the current cell is calculated, and the obtained average value is determined as the neighborhood temperature reference value; the temperature data of the current cell at the sampling time is subtracted from the neighborhood temperature reference value to obtain the temperature spatial gradient of the current cell.

[0040] A positive voltage or temperature spatial gradient indicates that the current cell's corresponding data is higher than the average level of adjacent cells; a negative voltage or temperature spatial gradient indicates that the current cell's corresponding data is lower than the average level of adjacent cells. The spatial gradient can retain its sign to simultaneously represent the direction and degree of deviation, or it can use its absolute value when only the degree of deviation is of concern.

[0041] When a data acquisition channel in an adjacent cell is missing, malfunctions, or continuously exceeds the sensor's effective range, that adjacent cell can be temporarily removed from the set Ω. i Remove from the middle and recalculate M based on the remaining valid adjacent cells. i Neighborhood reference value and spatial gradient; when the number of effective adjacent cells is less than the preset minimum number, the output of the spatial gradient of the cell can be paused, or the neighborhood reference value of the previous effective window can be used.

[0042] For example, if cell 3's neighboring cells are cells 2 and 4, and at the same sampling time, the voltages of cells 3, 2, and 4 are 3.695V, 3.702V, and 3.699V respectively, then the neighborhood voltage reference value for cell 3 is 3.7005V, and the voltage space gradient is -0.0055V. This negative gradient indicates that the voltage of cell 3 is lower than the average voltage level of its neighboring cells.

[0043] like Figure 5 As shown, ΔVi(t) represents the voltage timing difference of the i-th cell at sampling time t relative to the previous adjacent sampling time, and ΔTi(t) represents the temperature timing difference of the i-th cell at sampling time t relative to the previous adjacent sampling time. The L sampling times within the sliding time window are denoted as t1, t2, ..., tL in chronological order. Within the sliding time window containing L sampling points, for each sampling time of the i-th cell, the voltage data at the current sampling time is subtracted from the voltage data at the previous adjacent sampling time to obtain the voltage timing difference; the temperature data at the current sampling time is subtracted from the temperature data at the previous adjacent sampling time to obtain the temperature timing difference.

[0044] At each sampling moment, the current voltage, current temperature, voltage time-series difference, temperature time-series difference, voltage spatial gradient, and temperature spatial gradient are combined in a preset channel order to form a six-dimensional single-moment feature corresponding to that sampling moment. The voltage time-series difference and temperature time-series difference are used to represent the voltage change and temperature change of the same cell between adjacent sampling moments, respectively.

[0045] According to the sampling time from early to late, the L six-dimensional single-time features within the sliding time window are arranged row by row to obtain a multi-dimensional time series feature matrix containing L rows and six feature channels; each row of the matrix corresponds to a sampling time, and the six feature channels correspond to voltage, temperature, voltage time series difference, temperature time series difference, voltage spatial gradient and temperature spatial gradient in sequence.

[0046] The number of sampling points L within the sliding time window can be determined by the product of the sampling frequency and the duration of the sliding time window. The sliding time window moves forward according to a preset step size, so that adjacent windows retain continuous temporal evolution information.

[0047] In one alternative implementation, before constructing the multidimensional time-series feature matrix, missing value completion, acquisition pulse identification, low-pass filtering, and normalization are performed on the raw data. The low-pass filter cutoff frequency is lower than half of the sampling frequency to preserve low-frequency variation trends and suppress high-frequency interference.

[0048] When normalizing voltage or temperature data, first subtract the minimum value of that type of data within the corresponding sliding time window from the filtered current data. Then divide the resulting difference by the difference between the maximum and minimum values ​​within the same sliding time window. To avoid division by zero caused by the maximum and minimum values ​​being equal, a preset positive number is added to the divisor. The normalized data is used to reduce the dimensional differences between voltage and temperature.

[0049] To avoid the sliding window normalization masking absolute voltage or absolute temperature levels, both the original physical quantity and the window baseline value can be retained as additional inputs, or the online data can be normalized using fixed statistics from the initial training sample set. The same normalization method and feature channel order are used in both the training and online detection phases.

[0050] like Figure 6 As shown, the temporal transform encoder model described in this application is a Transformer encoder model for processing temporal features. First, a linear mapping layer maps the multi-dimensional temporal feature matrix from a six-dimensional feature space to a preset model dimension. Then, positional encoding is injected into the mapped features corresponding to each sampling position to preserve the chronological order of each sampling point in the time series. Positional encoding can employ fixed sine-cosine positional encoding or learnable positional encoding.

[0051] When using fixed sine-cosine positional encoding, for even-numbered dimensions in the model's feature dimensions, a sine function is used to generate the corresponding positional code; for odd-numbered dimensions, a cosine function is used. As the feature dimension number increases, the variation period corresponding to the sine and cosine functions increases according to a preset exponential relationship. Each sampling position in the sliding time window corresponds to a set of positional codes, and this set of positional codes is added element-wise to the mapping features of the corresponding sampling position.

[0052] The multi-head self-attention module maps the input features into a query matrix, a key matrix, and a value matrix, respectively. For each attention head, the product between the query matrix and the transpose of the key matrix is ​​first calculated, and the product result is scaled using the square root of the key vector dimension of the single attention head; then, the scaled result is normalized along the direction of the attention sampling position to obtain the attention matrix of that attention head; finally, the value matrix is ​​weighted and summed using the attention matrix to obtain the output of that attention head.

[0053] The outputs of multiple attention heads are concatenated according to their feature dimensions and processed through an output mapping layer to obtain multi-head attention features, while retaining the attention matrix corresponding to each attention head. The model dimension is divisible by the number of attention heads, and the query matrix, key matrix, and value matrix of each attention head all have a time dimension corresponding to the number of sampling points in the sliding time window.

[0054] The multi-head attention features are sequentially transformed through the first linear transformation layer, the nonlinear activation layer, and the second linear transformation layer to complete the feedforward feature transformation. The feedforward processing result is merged with the previous input through residual connection and layer normalization is performed to obtain the temporal coding features.

[0055] When performing feature pooling on time-series coded features, the average and maximum responses of each feature dimension can be calculated along the sampling time direction, and then concatenated according to a preset feature dimension order to obtain the feature vector of the corresponding battery cell. Feature pooling can also employ average pooling only, max pooling only, or other processing methods that can convert time-series coded features from multiple sampling times into fixed-dimensional feature vectors.

[0056] To determine the attention weight vector corresponding to the feature dimension of the feature vector, the attention matrices output by multiple attention heads are first aggregated to obtain the comprehensive attention response at each sampling time. Then, using the comprehensive attention response at each sampling time, the response values ​​of each feature dimension in the temporal coding feature are weighted and summarized at all sampling times to obtain the feature importance corresponding to each feature dimension.

[0057] The feature importance corresponding to each feature dimension is normalized to obtain the attention weight vector. The elements in the attention weight vector correspond one-to-one with the feature dimensions in the feature vector in the same order. When the feature vector is formed by concatenating the results of average pooling and max pooling, the feature importance of the average pooling part and the max pooling part are determined respectively, and the attention weight vector is formed in the same concatenation order as the feature vector, so that the attention weight vector has the same dimension as the feature vector.

[0058] Before the time-transform encoder model is put into testing, the model training equipment acquires multi-dimensional temporal feature matrices of multiple normal battery cells under different composite operating conditions to construct an initial training sample set. The initial training samples are normal battery cell samples, enabling the time-transform encoder model to learn the temporal changes and spatial deviation patterns of normal battery cells under different composite operating conditions; confirmed thermal runaway precursor samples are used for benchmark anomaly threshold calibration and model verification, and are not used as normal reconstruction training samples.

[0059] Randomly mask some sampling locations or feature channels in the initial training samples. The masked multidimensional temporal feature matrix is ​​then input into the temporal transform encoder model, and the masked data is predicted by reconstructing the output layer. For each masked element, the square of the difference between the model's predicted value and the true value before masking is calculated. The arithmetic mean of the squared differences corresponding to all masked elements is then calculated to obtain the reconstruction loss.

[0060] The temporal transform encoder model parameters are updated with the goal of reducing the reconstruction loss. After training, the reconstruction output layer is removed, and the temporal transform encoder model is retained as the feature extraction model. As one implementation, the model dimension is set to 64, the number of attention heads is set to 8, the number of encoder layers is set to 2, and the intermediate dimension of the feedforward network is set to 4 times the model dimension. Sampling positions and feature channels are randomly occluded according to a preset occlusion ratio of 10% to 30%. The model parameters are updated using an adaptive moment estimation optimization algorithm, and the learning rate is adjusted based on the validation set reconstruction loss. Training stops when the validation set reconstruction loss does not decrease within a preset number of consecutive training rounds. The above parameters are merely an example and do not constitute a limitation on the model parameters.

[0061] The same version of the time-transform encoder model, after training, is used to process normal battery cell samples under different composite operating conditions to obtain normal battery cell feature vectors. These feature vectors are then used to train the corresponding isolated forest sub-model. During online detection, the same data preprocessing method, feature matrix channel order, and model parameter version are used as in the model training phase to maintain consistency between training and detection inputs.

[0062] This embodiment uses a composite operating condition label to represent composite operating condition categories. The composite operating condition label consists of a basic operating state, a rate state, an ambient temperature state, and a state of charge. The basic operating state can be idle, charging, or discharging; the rate state can be a normal rate or a high rate; the ambient temperature state can be low temperature, normal temperature, or high temperature; and the state of charge can be low state of charge, intermediate state of charge, or high state of charge.

[0063] The basic operating state can be determined based on the charging / discharging direction indicator and the charging / discharging rate. When the absolute value of the charging / discharging rate is not greater than the resting rate threshold, it is determined to be in a resting state; when the charging / discharging direction indicator is charging and the absolute value of the charging / discharging rate is greater than the resting rate threshold, it is determined to be in a charging state; when the charging / discharging direction indicator is discharging and the absolute value of the charging / discharging rate is greater than the resting rate threshold, it is determined to be in a discharging state.

[0064] When the absolute value of the charge / discharge rate reaches the high-rate threshold, it is determined to be in a high-rate state; otherwise, it is determined to be in a normal-rate state. The ambient temperature state is determined based on whether the ambient temperature falls within a preset low-temperature range, normal-temperature range, or high-temperature range. The state of charge (SOC) is determined based on whether it falls within a preset low-charge range, intermediate-charge range, or high-charge range. These thresholds and range boundaries can be set based on the battery model, cycle life test data, and operating condition calibration data.

[0065] For example, if at a certain moment the battery is charging, the charging rate has reached a high rate threshold, the ambient temperature is in a high-temperature range, and the state of charge (SOC) is in a high-charge range, its operating condition category can be determined as "charging-high rate-high temperature-high charge". Different composite operating condition categories correspond to an isolated forest sub-model; when the number of training samples for a certain composite operating condition category is insufficient, it can fall back to the previous level operating condition sub-model with the same basic operating state and ambient temperature state, or use a general isolated forest model.

[0066] Each isolated forest sub-model is trained based on the feature vector of a normal battery cell under the corresponding operating condition category. The feature vector of the corresponding battery cell is weighted dimension by dimension according to the attention weight vector, that is, the value of each feature dimension in the feature vector is multiplied by the weight of the corresponding position in the attention weight vector to obtain a weighted feature vector; then the weighted feature vector is input into the target isolated forest sub-model to obtain the path length of the weighted feature vector in multiple isolated trees.

[0067] The average path length of the weighted feature vector across multiple isolated trees is calculated, and a benchmark for the average path length of normal samples is determined based on the number of subsamples used in a single isolated tree. This benchmark is then used to normalize the average path length of the current weighted feature vector, and the normalization result is converted into anomaly scores within a preset scoring range using an exponential mapping. Shorter path lengths indicate that the weighted feature vector is more easily isolated from normal samples, resulting in higher anomaly scores; longer path lengths correspond to lower anomaly scores.

[0068] During the model training phase, training samples are divided into training, validation, and test sets according to time to avoid the same consecutive execution segment appearing in both the training and test sets simultaneously. The number of isolated trees, the number of subsamples, and the maximum tree depth can be determined based on the embedded inference latency and validation set performance, and are not limited to a fixed value.

[0069] like Figure 8 As shown, the current life cycle state of a battery is divided into three stages: healthy, degraded, and aging, based on the State of Harmony (SOH). For example, a SOH ≥ 80% is defined as the healthy stage, 60% ≤ SOH < 80% as the degraded stage, and SOH < 60% as the aging stage. The stage boundaries can be recalibrated for different battery models based on cycle life tests.

[0070] The health stage, degradation stage, and aging stage each correspond to different baseline anomaly thresholds. As the battery health decreases, the baseline anomaly threshold can be gradually reduced to improve sensitivity to abnormal changes in aging cells. The baseline anomaly threshold is calibrated using a validation set of normal samples and confirmed thermal runaway precursor samples. Confirmed thermal runaway precursor samples can be identified through internal short-circuit tests, local heating tests, overcharge tests, fault reproduction results, or maintenance and inspection results.

[0071] The rate correction parameter is determined based on the current charge / discharge rate. One implementation method involves multiplying the absolute value of the charge / discharge rate by a non-negative rate correction factor and then adding one to obtain the rate correction parameter, ensuring that the larger the absolute value of the charge / discharge rate, the larger the rate correction parameter.

[0072] The state of charge (SOC) correction parameter is determined based on whether the SOC falls within a preset SOC boundary range. When the SOC is within a preset low SOC range or a preset high SOC range, the SOC boundary indicator is set to one; otherwise, it is set to zero. The SOC boundary indicator is multiplied by a non-negative SOC correction factor and then incremented by one to obtain the SOC correction parameter. The temperature correction parameter is determined in the same way: when the ambient temperature is within a preset low temperature range or a preset high temperature range, the temperature boundary indicator is set to one; otherwise, it is set to zero. The temperature boundary indicator is multiplied by a non-negative temperature correction factor and then incremented by one to obtain the temperature correction parameter.

[0073] Multiply the rate correction parameter, state of charge correction parameter, and temperature correction parameter to obtain the operating condition correction coefficient; then multiply the baseline anomaly threshold by the operating condition correction coefficient to obtain the threshold correction result, and determine the threshold correction result as the adaptive anomaly threshold.

[0074] In one optional implementation, a lower threshold and an upper threshold can be preset. When the threshold correction result is lower than the lower threshold, the adaptive anomaly threshold is determined as the lower threshold; when the threshold correction result is higher than the upper threshold, the adaptive anomaly threshold is determined as the upper threshold; otherwise, the threshold correction result is determined as the adaptive anomaly threshold. The lower threshold is lower than the upper threshold, and both are within the range of anomaly scores.

[0075] For example, when the battery health status corresponds to the degradation stage, the baseline anomaly threshold can be set to 0.55. When charging at twice the rated rate, with the state of charge in the high state of charge range and the ambient temperature in the high temperature range, the rate correction parameter can be set to 1.1, and both the state of charge correction parameter and the temperature correction parameter can be set to 1.05. Multiplying the three correction parameters together yields an operating condition correction coefficient of approximately 1.213. Multiplying the baseline anomaly threshold by this operating condition correction coefficient yields an adaptive anomaly threshold of approximately 0.667. The above values ​​are merely calculation examples and do not constitute a limitation on the parameter range.

[0076] like Figure 9 As shown, the adaptive anomaly threshold is determined as the first-level threshold. A remaining scoring interval is determined between the first-level threshold and the upper limit of the anomaly score value. This remaining scoring interval is then truncated according to a first preset ratio, and the sum of the first-level threshold and the truncated interval value is determined as the second-level threshold. The remaining scoring interval is then truncated according to a second preset ratio greater than the first preset ratio, and the sum of the first-level threshold and the corresponding interval value is determined as the third-level threshold. Therefore, the first-level threshold is less than the second-level threshold, and the second-level threshold is less than the third-level threshold.

[0077] An anomaly score below the first-level threshold is considered normal; an anomaly score reaching the first-level threshold but below the second-level threshold is considered the first warning level; an anomaly score reaching the second-level threshold but below the third-level threshold is considered the second warning level; and an anomaly score reaching the third-level threshold is considered the third warning level. Both the first and second preset ratios are greater than 0 and less than 1, with the first preset ratio being less than the second preset ratio. Therefore, all warning levels are predicated on the anomaly score reaching an adaptive anomaly threshold.

[0078] The battery management system determines the initial anomaly level within each sliding time window. Figure 9In the example, W1, W2, and W3 represent the first, second, and third sliding time windows arranged consecutively in chronological order, respectively. In this example, W1 and W2 both correspond to the first warning level, while W3 corresponds to the normal state. Therefore, when at least two of the three consecutive sliding time windows have the same initial anomaly level, the first warning level is determined as the final anomaly level. To reduce misjudgments caused by occasional fluctuations in a single window, when at least a preset number of sliding time windows have the same initial anomaly level, the corresponding final anomaly level is determined. The sliding time windows with the same initial anomaly level can be adjacent to each other or distributed within the multiple consecutive sliding time windows; in embodiments requiring higher anomaly persistence, the preset number of sliding time windows can be further limited to being adjacent to each other. As an example, at least two sliding time windows having the same initial anomaly level can be used as a confirmation condition within three consecutive sliding time windows.

[0079] When the anomaly level of adjacent windows increases progressively, the highest risk level can be output; when the anomaly level decreases, it can be required that multiple consecutive windows be below the current level threshold before downgrading is performed, to avoid frequent level jumps. The final thermal runaway detection result includes the cell identifier of the abnormal cell, the anomaly level, the anomaly score, the corresponding threshold, and the timestamp.

[0080] like Figure 10 As shown, when the final anomaly level of the target cell reaches the first warning level, the battery management system increases the sampling frequency of the target cell from the basic sampling frequency to the target sampling frequency, or increases it by a preset factor, and records the continuous voltage and temperature data of the target cell; simultaneously, it controls the thermal management system to adjust the battery module temperature to a preset target range. For example, the basic sampling frequency can be 1Hz, the target sampling frequency can be 10Hz, and the target temperature range can be 25℃±5℃.

[0081] When the final anomaly level is the second warning level, the battery management system generates a charge and discharge power limit command to limit the charge and discharge rate and turn off the fast charging function; controls the cooling pump, fan or refrigeration circuit to enter a forced cooling state; and sends an alarm message containing cell identification, anomaly level and anomaly score to the vehicle controller.

[0082] When the final abnormality level is the third warning level, the battery management system first sends a stop charging and discharging command to the charging equipment or drive controller. After confirming that the circuit current has dropped to the allowable range or the emergency disconnection condition has been met, it controls the high-voltage main relay to disconnect. At the same time, it controls the thermal management system to operate at the preset emergency cooling power and sends emergency alarm information to the vehicle controller and emergency call module.

[0083] The battery management system receives instruction execution status reports from each execution unit. When a safety protection instruction fails to execute, the system can reissue the instruction, switch to a backup control channel, or escalate the alarm level. The specific parameters for these actions can be calibrated based on the battery type, vehicle safety strategy, and thermal management capabilities.

[0084] The vehicle controller, fast charging disable function, and emergency call module are primarily applicable to new energy vehicle power battery scenarios. When applied to energy storage battery systems, the vehicle controller can be replaced by an energy storage system controller or a station-level controller, the fast charging disable function can be replaced by limiting the charging and discharging power of the energy storage converter, and the emergency call module can be replaced by a station-level alarm module or a remote operation and maintenance communication module; each replacement object receives corresponding limiting, cooling, isolation, or alarm commands based on the same anomaly level.

[0085] like Figure 11 As shown, the battery management system acquires voltage time-series data, temperature time-series data, operating parameters, battery health status, thermal runaway detection results, and safety protection strategy execution results throughout the battery's lifecycle, according to a preset cycle or after detecting a confirmed abnormal event, to construct a model update sample set. The model update samples may include the cell number, time range, vehicle or energy storage device model, software version, and the confirmation status of the detection results.

[0086] Anomalies in the model update sample set are labeled. Labeling of anomalies can come from maintenance and inspection results, laboratory reproduction results, manual review results, or results confirmed through cross-verification of fault information from multiple sources. Unconfirmed anomalies are not directly added to the training set as normal samples to reduce the risk of the model learning abnormal states as normal states.

[0087] The battery management system can perform incremental training when the vehicle is parked, the energy storage system is under low load, or an external power source is available and the processor temperature is below a preset upper limit. During incremental training, historical representative samples are mixed with new samples in a preset ratio, and the model parameters and attention weights of the temporal transform encoder model are updated using the labeled model update sample set. The number of training epochs, learning rate, and sample batches are set according to the available computing power and validation set performance to reduce the risk of a decrease in the ability to recognize historical operating conditions due to new samples.

[0088] The updated time-transform encoder model is used to re-extract the normal cell features under different operating conditions. Based on the re-extracted normal cell features, the corresponding isolated forest sub-model is retrained, isolated trees are added, or sub-sampling data is adjusted. Then, based on the abnormal scores output by the calibrated isolated forest sub-model and the distribution of normal samples and confirmed thermal runaway precursor samples, the benchmark abnormal thresholds corresponding to different battery health states are calibrated.

[0089] The model update sample set is divided into a training part and an independent validation part. The detection accuracy of the updated time-transform encoder model, the calibrated isolated forest sub-model, and the calibrated baseline anomaly threshold is verified. Preset validation conditions may include a recall rate of thermal runaway precursor samples not lower than that before the update, a false alarm rate of normal samples not higher than that before the update, and inference latency and memory usage not exceeding the preset upper limit of the battery management system. When the preset validation conditions are met, the updated model and thresholds are used for subsequent cell thermal runaway detection; when they are not met, the model parameters and threshold parameters before the update are retained or restored.

[0090] like Figure 12 As shown, this embodiment provides a cell thermal runaway detection device driven by CCS data acquisition. This device is applied to a battery management system and includes a data acquisition module, a spatial gradient determination module, an anomaly scoring determination module, a threshold determination module, and a detection result determination module.

[0091] The data acquisition module acquires voltage and temperature time-series data from multiple cells synchronously collected by the CCS, and obtains spatial arrangement information, operating parameters, and battery health status. The spatial gradient determination module determines the voltage and temperature spatial gradients of each cell relative to its adjacent cells based on the spatial arrangement information. The voltage spatial gradient characterizes the deviation of each cell's voltage data from its adjacent cells at the same sampling time, while the temperature spatial gradient characterizes the deviation of each cell's temperature data from its adjacent cells at the same sampling time. The anomaly scoring determination module determines anomaly scores based on the voltage and temperature time-series data, voltage spatial gradient, and temperature spatial gradient. The threshold determination module determines a baseline anomaly threshold based on the battery health status and corrects it according to the operating parameters. The detection result determination module compares the anomaly score with the adaptive anomaly threshold and outputs the thermal runaway detection result.

[0092] The aforementioned modules can be implemented using software, hardware, or a combination of both. Multiple modules can be integrated into the same processor, or they can be implemented collaboratively by the main control processor and neural network acceleration unit in the battery management system. The device may also include a memory and a communication interface. The memory is used to store model parameters, threshold parameters, spatial arrangement information, and programs. The communication interface is used to receive CCS data and transmit information to the thermal management system, vehicle controller, high-voltage main relay control unit, and model training equipment.

[0093] Those skilled in the art will understand that, without conflict, the technical features in the above embodiments can be combined with each other. Hardware interfaces, vehicle bus, and basic battery management control processes not described in detail can be implemented using methods conventional in the art.

[0094] The above description is merely a preferred embodiment of this application and is not intended to limit this application. All equivalent substitutions, improvements, or combinations made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A cell thermal runaway detection method based on CCS data acquisition, characterized in that, Applied to a battery management system, the method includes: The system acquires voltage and temperature timing data of multiple cells within the battery module, which are synchronously collected by the cell connection component (CCS), and obtains the spatial arrangement information, operating parameters, and battery health status of the multiple cells. The voltage spatial gradient and temperature spatial gradient of each cell relative to its adjacent cells are determined based on the spatial arrangement information. The voltage spatial gradient characterizes the deviation of the voltage data of each cell from the voltage data of its adjacent cells at the same sampling time, and the temperature spatial gradient characterizes the deviation of the temperature data of each cell from the temperature data of its adjacent cells at the same sampling time. The determination of the voltage spatial gradient and temperature spatial gradient of each cell relative to its adjacent cells based on the spatial arrangement information includes: determining the set of adjacent cells corresponding to each cell based on the spatial arrangement information; for each sampling time, determining the neighborhood voltage reference value corresponding to each cell based on the voltage data of each cell and the voltage data of each adjacent cell in the corresponding set of adjacent cells, and determining the voltage spatial gradient of each cell based on the deviation of the voltage data of each cell from the neighborhood voltage reference value; determining the neighborhood temperature reference value corresponding to each cell based on the temperature data of each cell and the temperature data of each adjacent cell in the corresponding set of adjacent cells, and determining the temperature spatial gradient of each cell based on the deviation of the temperature data of each cell from the neighborhood temperature reference value. Anomaly scores for each cell are determined based on the voltage timing data, the temperature timing data, the voltage spatial gradient, and the temperature spatial gradient. A baseline anomaly threshold is determined based on the battery health status, and the baseline anomaly threshold is corrected based on the operating condition parameters to obtain an adaptive anomaly threshold. The thermal runaway detection result of each cell is determined based on the comparison between the abnormal score of each cell and the adaptive abnormal threshold.

2. The method as described in claim 1, characterized in that, The step of determining the anomaly score for each cell based on the voltage time-series data, the temperature time-series data, the voltage spatial gradient, and the temperature spatial gradient includes: The voltage timing difference of each cell is determined based on the voltage data of each cell at adjacent sampling times, and the temperature timing difference of each cell is determined based on the temperature data of each cell at adjacent sampling times. According to the sampling time order, the voltage time series data, temperature time series data, voltage time series difference, temperature time series difference, voltage spatial gradient and temperature spatial gradient of each cell are combined to construct the multi-dimensional time series feature matrix corresponding to each cell. The feature vectors of each battery cell are obtained by performing time-series feature extraction on the multi-dimensional time-series feature matrix, and the anomaly score of each battery cell is determined based on the feature vectors of each battery cell.

3. The method as described in claim 2, characterized in that, The step of extracting time-series features from the multi-dimensional time-series feature matrix to obtain the feature vectors of each cell includes: The multidimensional temporal feature matrix is ​​mapped to feature dimensions, and positional codes representing the sampling time order are injected into the mapped multidimensional temporal feature matrix. The multi-dimensional temporal feature matrix injected with the position code is input into the temporal transform encoder model. The multi-head self-attention network of the temporal transform encoder model is used to determine the correlation between different sampling times and different feature dimensions. The output of the multi-head self-attention network is processed by the feedforward network, residual connection, and layer normalization of the temporal transform encoder model to obtain temporal coding features; The time-series encoded features are pooled to obtain the feature vectors of each battery cell, and attention weight vectors corresponding to the feature dimensions of each battery cell's feature vectors are determined.

4. The method as described in claim 3, characterized in that, The process of determining the anomaly score for each battery cell based on its feature vector includes: The attention weight vector is used to weight each feature dimension in the feature vector of the corresponding battery cell to obtain a weighted feature vector. The operating condition category of each cell is determined based on the operating condition parameters, and a target isolated forest sub-model corresponding to the operating condition category is determined from multiple pre-trained isolated forest sub-models. The different isolated forest sub-models are trained based on normal cell samples under different operating condition categories. The weighted feature vector is input into the target isolated forest sub-model to determine the path length of the weighted feature vector in multiple isolated trees of the target isolated forest sub-model; The anomaly score of the corresponding battery cell is determined based on the path length of the weighted feature vector in the multiple isolated trees, wherein the shorter the path length, the higher the corresponding anomaly score.

5. The method according to any one of claims 1-4, characterized in that, The step of determining a baseline anomaly threshold based on the battery health status and correcting the baseline anomaly threshold based on the operating condition parameters to obtain an adaptive anomaly threshold includes: Based on the battery health status characterized by the battery health status, the current life cycle state of the battery is determined as a healthy stage, a degraded stage, or an aging stage, and a benchmark abnormality threshold corresponding to the current life cycle state is determined, wherein the benchmark abnormality threshold decreases as the battery health status decreases. The operating parameters include charge / discharge rate, state of charge, and ambient temperature. A rate correction parameter is determined based on the charge / discharge rate, a state of charge correction parameter is determined based on whether the state of charge is within a preset state of charge boundary range, and a temperature correction parameter is determined based on whether the ambient temperature is within a preset low temperature range or a preset high temperature range. The operating condition correction coefficient is determined based on the ratio correction parameter, the state of charge correction parameter, and the temperature correction parameter, and the benchmark anomaly threshold is corrected using the operating condition correction coefficient to obtain the adaptive anomaly threshold.

6. The method as described in claim 5, characterized in that, The step of determining the thermal runaway detection result of each battery cell based on the comparison result of the abnormal score of each cell with the adaptive abnormal threshold includes: Based on the adaptive anomaly threshold, multiple warning levels are set with corresponding anomaly scoring intervals. The multiple warning levels include at least a first warning level, a second warning level, and a third warning level that increase sequentially according to the risk of thermal runaway. Based on the abnormal score range of each cell, the preliminary abnormality level of each cell within the current sliding time window is determined. The preliminary abnormality level of each cell within a series of consecutive sliding time windows is obtained. When the target cell has the same preliminary abnormality level within at least a preset number of sliding time windows in the series of consecutive sliding time windows, the same preliminary abnormality level is determined as the final abnormality level of the target cell. The thermal runaway detection result of the target cell is determined based on the final anomaly level, and the cell identifier of the target cell is output.

7. The method as described in claim 6, characterized in that, After determining the same initial anomaly level as the final anomaly level of the target cell, the method further includes: When the final anomaly level is the first warning level, the sampling frequency of the voltage and temperature data of the target battery cell is increased, the voltage time-series data and temperature time-series data of the target battery cell are recorded, and the thermal management system is controlled to pre-adjust the temperature of the battery module. When the final abnormality level is the second warning level, a charging and discharging power limiting command is generated to limit the charging and discharging rate of the battery module, turn off the fast charging function, control the thermal management system to force cooling of the battery module, and send abnormal alarm information to the vehicle controller. When the final abnormality level is the third warning level, a high-voltage power-down command is generated to control the high-voltage main relay to disconnect, thereby cutting off the charging and discharging circuit of the battery module, controlling the thermal management system to operate at the preset emergency cooling power, and sending emergency alarm information to the vehicle controller and emergency call module.

8. The method as described in claim 4, characterized in that, The method further includes: Acquire voltage time-series data, temperature time-series data, operating parameters, battery health status, thermal runaway detection results, and safety protection strategy execution results for each cell during the battery life cycle, and build a model update sample set; Abnormal samples in the model update sample set are labeled, and the temporal transform encoder model is incrementally trained based on the labeled model update sample set to update the model parameters and attention weights of the temporal transform encoder model. The isolated forest sub-model is calibrated based on the model update sample set, and the benchmark anomaly thresholds corresponding to different battery health states are calibrated based on the anomaly scores output by the calibrated isolated forest sub-model. The detection accuracy of the updated timing-transform encoder model, the calibrated isolated forest sub-model, and the calibrated baseline anomaly threshold is verified. When the preset verification conditions are met, the updated model and threshold are used for subsequent cell thermal runaway detection.

9. A cell thermal runaway detection device driven by CCS acquisition data, used to implement the method described in claim 1, characterized in that, Applied to a battery management system, the device includes: The data acquisition module is used to acquire voltage and temperature time series data of multiple cells in the battery module synchronously collected by the cell connection component CCS, and to acquire the spatial arrangement information, operating parameters and battery health status of the multiple cells. A spatial gradient determination module is used to determine the voltage spatial gradient and temperature spatial gradient of each cell relative to its adjacent cells based on the spatial arrangement information. The voltage spatial gradient characterizes the deviation of the voltage data of each cell from the voltage data of its adjacent cells at the same sampling time, and the temperature spatial gradient characterizes the deviation of the temperature data of each cell from the temperature data of its adjacent cells at the same sampling time. The process of determining the voltage spatial gradient and temperature spatial gradient of each cell relative to its adjacent cells based on the spatial arrangement information includes: determining the set of adjacent cells corresponding to each cell based on the spatial arrangement information; for each sampling time, determining the neighborhood voltage reference value corresponding to each cell based on the voltage data of each cell and the voltage data of each adjacent cell in the corresponding set of adjacent cells, and determining the voltage spatial gradient of each cell based on the deviation of the voltage data of each cell from the neighborhood voltage reference value; determining the neighborhood temperature reference value corresponding to each cell based on the temperature data of each cell and the temperature data of each adjacent cell in the corresponding set of adjacent cells, and determining the temperature spatial gradient of each cell based on the deviation of the temperature data of each cell from the neighborhood temperature reference value. An anomaly scoring determination module is used to determine the anomaly score of each cell based on the voltage time-series data, the temperature time-series data, the voltage spatial gradient, and the temperature spatial gradient. The threshold determination module is used to determine a baseline abnormal threshold based on the battery health status, and to correct the baseline abnormal threshold based on the operating condition parameters to obtain an adaptive abnormal threshold. The detection result determination module is used to determine the thermal runaway detection result of each cell based on the comparison result of the abnormal score of each cell and the adaptive abnormal threshold.

Images