A safety energy storage cabinet thermal failure early warning method and system

Abstract

The present application relates to the technical field of battery management, and discloses a safe energy storage cabinet thermal failure early warning method and system, the method comprising: collecting energy storage cabinet battery electric heating characteristic data and extracting real-time battery state; the data is denoised, and the instantaneous vector is deconstructed, the monitoring characteristics are determined by recombining and mapping when the coupling correlation threshold is satisfied; the electric heating coupling strength is calculated, the optimal hyperplane is determined by mapping to the high-dimensional space, and the abnormal vector is determined; the parameter channel is locked, the trend is modeled, the thermal energy accumulation is calculated, the thermal failure risk level is quantitatively determined; the history snapshot is retrieved, the similar state is screened, the topological network is constructed, and the evolution path is generated; the gait cycle logic time sequence prediction is used, the monitoring parameters are reconstructed according to the alignment error, the interaction mode is optimized; finally, the cross-correlation value of the interaction mode and the real-time battery state is calculated to generate a fusion matrix, the maximum eigenvalue is extracted to quantitatively score and output a pulse early warning signal. The present application can solve the problem of poor timeliness of energy storage cabinet thermal failure early warning in the prior art.

Description

Technical Field

[0001] This invention relates to the field of battery management technology, and in particular to a method and system for early warning of thermal failure in a safe energy storage cabinet. Background Technology

[0002] Currently, with the large-scale application of energy storage systems in areas such as power peak shaving and new energy grid connection, the safety of the battery packs inside the energy storage cabinet has become the core of ensuring the stable operation of the energy system. Especially under high-load operation or complex environmental conditions, the complex electrochemical reactions inside the battery and the external heat transfer can easily induce thermal runaway accidents. Therefore, there is an urgent need to introduce advanced fault prediction and health management technologies to conduct in-depth monitoring and risk assessment of battery operating status.

[0003] In existing technologies, energy storage cabinet early warning systems often employ threshold monitoring schemes based on a single physical quantity. These systems collect battery voltage or temperature data via sensors and trigger alarms when the monitored values ​​exceed preset static safety limits. The system typically treats the collected raw measurement data streams as independent time-series sequences, performing logical judgments only on single-dimensional characteristic fluctuations, and then outputs alarm signals according to a fixed control strategy. However, during dynamic operation, battery thermal failure often manifests as a deep nonlinear coupling between electrical and thermal characteristics. For example, under high load conditions, abnormal fluctuations in the battery's internal voltage are accompanied by a rapid accumulation of temperature gradients. Single-dimensional monitoring, by ignoring the interaction patterns between different physical quantities, makes it difficult for the system to accurately detect early thermal runaway risks when faced with multiple overlapping risk factors.

[0004] Therefore, existing technologies suffer from poor timeliness in early warning of thermal failure in energy storage cabinets. Summary of the Invention

[0005] This invention provides a method and system for early warning of thermal failure of energy storage cabinets, which addresses the problem of poor timeliness in early warning of thermal failure of energy storage cabinets in existing technologies.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a method for early warning of thermal failure of a safe energy storage cabinet, comprising: Acquire real-time measurement data, which includes voltage, current, and temperature data of individual battery cells; The real-time measurement data is denoised and deconstructed to obtain an instantaneous state vector. The coupling correlation value of the instantaneous state vector is calculated, a feature matrix is ​​constructed, and a mapping process is performed to determine the monitoring features. The monitoring features are used to calculate dynamic correlation indicators, and the dynamic correlation indicators are mapped to the feature space to obtain the optimal classification hyperplane. The anomaly vector is determined based on the geometric distance deviation between the monitoring features and the optimal classification hyperplane. The incremental features are obtained by performing a difference operation on the anomaly vector. A trend model is constructed based on the incremental features and integrated to obtain the heat energy accumulation. The heat energy accumulation is mapped to a preset risk classification table to determine the thermal failure risk level. Based on the thermal failure risk level, similar sequences that meet the preset spatial proximity conditions with the anomaly vector are selected and a topological network is constructed. An evolution path is generated by combining the propagation probability distribution of the topological network. A time-series prediction is performed on the evolution path to obtain a prediction result. The alignment error between the prediction result and the real-time measurement data is calculated. A time-series calibration is performed based on the alignment error to determine an optimized interaction mode. The cross-correlation value between the optimized interaction mode and the real-time measurement data is calculated, the maximum feature value of the cross-correlation result is extracted and quantified to obtain a risk score. If the risk score exceeds a preset score threshold, a pulse code sequence is generated and an early warning signal is output.

[0007] Preferably, the real-time measurement data is denoised and deconstructed to obtain an instantaneous state vector. The coupling correlation value of the instantaneous state vector is calculated, a feature matrix is ​​constructed, and mapping processing is performed to determine the monitoring features, including: The real-time measurement data is subjected to denoising and deconstruction processing to obtain the instantaneous state vector; Calculate the coupling correlation value of the instantaneous state vector. If the coupling correlation value is greater than a preset correlation threshold, combine the instantaneous state vector according to the time series to generate a feature matrix. The feature matrix is ​​reorganized using a sliding window, and the reorganized result is spatially transformed to obtain the monitoring features.

[0008] Preferably, a dynamic correlation index is calculated using the monitoring features, and the dynamic correlation index is mapped to the feature space to obtain the optimal classification hyperplane. An anomaly vector is determined based on the geometric distance deviation between the monitoring features and the optimal classification hyperplane, including: The dynamic correlation index of the voltage, current and temperature data of the battery cell is calculated using the monitoring characteristics to obtain the coupling strength data; The coupling strength data is processed using a preset nonlinear mapping function to obtain the optimal classification hyperplane and the maximum margin distance in the feature space; Calculate the geometric distance between the monitored feature and the optimal classification hyperplane. If the geometric distance is less than the maximum interval distance, determine the anomaly vector based on the deviation of the geometric distance.

[0009] Preferably, the incremental features are obtained by performing a difference operation on the anomaly vector, a trend model is constructed based on the incremental features and integrated to obtain the accumulated heat energy, and the accumulated heat energy is mapped to a preset risk classification table to determine the thermal failure risk level, including: The anomaly vector is subjected to a difference operation to obtain incremental features, and a model is constructed based on the incremental features to obtain a trend model; The cumulative amount of heat energy is obtained by performing integral calculations using the aforementioned trend model. The accumulated heat energy is mapped to a preset risk classification table to determine the risk level of thermal failure.

[0010] Preferably, based on the thermal failure risk level, similar sequences that satisfy a preset spatial proximity condition with the anomaly vector are selected and a topology network is constructed. An evolution path is generated by combining the propagation probability distribution of the topology network, including: Extract historical state snapshots corresponding to the thermal failure risk level, map the historical state snapshots to a multi-dimensional space, and filter out similar sequences that satisfy the preset spatial proximity condition with the anomaly vector; The similar sequences are used as nodes to construct a topology network, and the probability of each state branch in the topology network evolving into a thermal failure state is calculated to determine the propagation probability distribution. By matching the propagation probability distribution with the real-time gradient direction of the anomaly vector, the evolution branch with the highest probability is locked, and an evolution path is generated.

[0011] Preferably, the process includes performing time-series prediction on the evolution path to obtain prediction results, calculating the alignment error between the prediction results and the real-time measurement data, performing time-series calibration based on the alignment error, and determining an optimized interaction mode, including: The node attributes of the evolution path are input into a pre-trained long short-term memory network. The cell state is updated using the forget gate and the input gate, and the hidden layer state reflecting the temporal dependency is determined by combining the output gate. The hidden layer state is mapped and regressed using the fully connected layer of the Long Short-Term Memory network to generate a thermal failure risk probability curve and obtain the prediction result. The alignment error is obtained by comparing the values ​​of key time points in the prediction results with the corresponding times in the real-time measurement data. If the alignment error exceeds the preset error range, then timestamp realignment and feature value interpolation are performed on the measurement data according to the alignment error to reconstruct the optimized monitoring features; Calculate the correlation coefficients of different dimensional attributes in the optimized monitoring features, construct the interaction correlation matrix, and obtain the optimized interaction mode.

[0012] Preferably, the cross-correlation value between the optimized interaction mode and the real-time measurement data is calculated, the maximum eigenvalue of the cross-correlation result is extracted and quantified to obtain a risk score. If the risk score exceeds a preset score threshold, a pulse-coded sequence is generated and a warning signal is output, including: A nonlinear mapping is performed between the optimized interaction mode and the real-time measurement data to obtain an interaction vector and a state vector. Calculate the cross-correlation coefficient between the interaction vector and the state vector to obtain the cross-correlation value; The cross-correlation values ​​are arranged in a matrix according to the time index to obtain the fusion matrix; The risk score is obtained by extracting the largest eigenvalue from the fusion matrix and performing normalization. If the risk score exceeds a preset score threshold, a pulse code is generated based on a preset combination of pulse width and interval, and a warning signal is output.

[0013] Secondly, the present invention provides a thermal failure early warning system for a safe energy storage cabinet, comprising: The data acquisition module is used to acquire real-time measurement data, which includes the voltage, current and temperature data of individual battery cells. The feature extraction module is used to denoise and deconstruct the real-time measurement data to obtain an instantaneous state vector, calculate the coupling correlation value of the instantaneous state vector, construct a feature matrix and perform mapping processing to determine the monitoring features; An anomaly detection module is used to calculate dynamic correlation indicators using the monitoring features, map the dynamic correlation indicators to the feature space to obtain the optimal classification hyperplane, and determine the anomaly vector based on the geometric distance deviation between the monitoring features and the optimal classification hyperplane. The risk classification module is used to perform differential operations on the anomaly vector to obtain incremental features, construct a trend model based on the incremental features and perform integration to obtain the heat energy accumulation, and map the heat energy accumulation to a preset risk classification table to determine the thermal failure risk level. The path evolution module is used to select similar sequences that meet the preset spatial proximity conditions with the anomaly vector according to the thermal failure risk level and construct a topology network, and generate an evolution path by combining the propagation probability distribution of the topology network. The interaction optimization module is used to perform time-series prediction on the evolution path to obtain prediction results, calculate the alignment error between the prediction results and the real-time measurement data, perform time-series calibration based on the alignment error, and determine the optimized interaction mode. The early warning output module is used to calculate the cross-correlation value between the optimized interaction mode and the real-time measurement data, extract the maximum feature value of the cross-correlation result and perform quantitative scoring to obtain a risk score. If the risk score exceeds a preset score threshold, a pulse code sequence is generated and an early warning signal is output.

[0014] Thirdly, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the thermal failure early warning method for the safe energy storage cabinet described in any one of the above.

[0015] Fourthly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform the thermal failure early warning method for the safety energy storage cabinet described in any one of the above.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention performs noise reduction and deconstruction processing on the measurement data collected by the sensor array, extracts the instantaneous vector reflecting the electrothermal characteristics and reconstructs it into monitoring features according to the coupling correlation value; through this preprocessing process, the system can effectively filter out random noise interference in complex electromagnetic environment, improve the correlation accuracy between the original measurement data and the real-time state of the battery, and provide a highly reliable data foundation for subsequent accurate analysis of electrothermal coupling behavior.

[0017] (2) The present invention calculates the electrothermal coupling strength based on monitoring features and uses kernel functions to map it to a high-dimensional feature space to determine the optimal classification hyperplane and geometric distance deviation, thereby determining the anomaly vector. This high-dimensional mapping process can capture the complex nonlinear interaction between electrical and thermal parameters, solve the technical problem that single-dimensional monitoring is difficult to reflect the real risks inside the battery, and significantly improve the sensitivity and accuracy of identifying critical abnormal states.

[0018] (3) The present invention calculates the magnitude of the abnormal vector and locks the specific parameter channel. It calculates the deviation area through differential operation and trend modeling, and then obtains the thermal energy accumulation and risk level. This process quantifies the abnormal state from the dimension of dynamic energy accumulation, so that the system can capture the dynamic deterioration trend of thermal runaway before the battery is substantially damaged, which enhances the foresight of the early warning system and effectively reduces the probability of missed detection of thermal failure risk.

[0019] (4) This invention retrieves historical state snapshots that are of the same origin as the risk level, constructs a topological network based on spatial proximity, and generates an evolution path by combining the propagation probability distribution and real-time gradient direction. By comparing the current abnormal situation with the historical evolution pattern in multiple dimensions, the system can identify high-risk evolution directions in advance and deduce the potential threat evolution trajectory, providing a scientific path reference for operation and maintenance intervention and improving the energy storage system's response capability to sudden multiple impacts.

[0020] (5) This invention uses gated loop logic to perform time-series prediction on evolutionary features, and reconstructs monitoring features in real time based on the alignment error between the prediction results and the measurement data to optimize the interaction mode, and finally integrates real-time status to output early warning signals; this feedback optimization mechanism can dynamically compensate for the lag deviation of model prediction, and realize the whole process management from anomaly identification, evolution deduction to real-time closed-loop correction. While ensuring the real-time performance of early warning signals, it greatly improves the accuracy and environmental adaptability of the intelligent early warning system under complex operating conditions. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of a method for early warning of thermal failure of a safe energy storage cabinet provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of a thermal failure early warning system for a safe energy storage cabinet provided in the second embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Reference Figure 1 The first embodiment of the present invention provides a method for early warning of thermal failure of a safe energy storage cabinet, comprising the following steps: S11, acquire real-time measurement data, which includes the voltage, current and temperature data of the battery cells; S12, the real-time measurement data is denoised and deconstructed to obtain an instantaneous state vector, the coupling correlation value of the instantaneous state vector is calculated, a feature matrix is ​​constructed and mapped, and the monitoring features are determined. S13, calculate the dynamic correlation index using the monitoring features, and map the dynamic correlation index to the feature space to obtain the optimal classification hyperplane. Determine the anomaly vector based on the geometric distance deviation between the monitoring features and the optimal classification hyperplane. S14, perform differential operation on the anomaly vector to obtain incremental features, construct a trend model based on the incremental features and perform integration to obtain the heat energy accumulation, map the heat energy accumulation to a preset risk classification table, and determine the thermal failure risk level. S15, Based on the thermal failure risk level, similar sequences that satisfy the preset spatial proximity condition with the anomaly vector are selected and a topological network is constructed. An evolution path is generated by combining the propagation probability distribution of the topological network. S16, perform time-series prediction on the evolution path to obtain prediction results, calculate the alignment error between the prediction results and the real-time measurement data, perform time-series calibration based on the alignment error, and determine the optimized interaction mode; S17, calculate the cross-correlation value between the optimized interaction mode and the real-time measurement data, extract the maximum feature value of the cross-correlation result and perform quantitative scoring to obtain a risk score. If the risk score exceeds a preset score threshold, generate a pulse code sequence and output a warning signal.

[0024] In step S11, real-time measurement data is acquired, which includes the voltage, current and temperature data of individual battery cells.

[0025] First, raw measurement data of individual battery cells are collected using a sensor array deployed inside the energy storage cabinet. This data includes the voltage, current, and temperature of each cell. During data acquisition, voltage is obtained using a high-precision voltage sensor with a preset sampling frequency of 1kHz to capture millivolt-level voltage fluctuations during dynamic charging and discharging. Current is monitored in real-time using a Hall effect current sensor to capture ampere-level load intensity. Temperature is measured using a distributed thermistor array to obtain the degree Celsius difference on the surface of each battery cell and between adjacent cells, thereby determining the temperature gradient.

[0026] To achieve deep coupling analysis of multi-dimensional parameters, the data acquisition process must ensure strict time alignment. In this step, a preset sampling period is set. Preset sampling period The sampling interval is based on the difference between the battery's electrochemical kinetic response rate and thermal conduction rate. Because the electrical responses of voltage and current are relatively fast, their sampling interval is set within the range of 1 ms to 10 ms; while temperature changes have thermal inertia, and their sampling interval is set within the range of 100 ms to 500 ms. A unified timestamp management mechanism maps these data streams of different frequencies onto the same time axis. If data points are not completely synchronized, a linear interpolation algorithm is used for alignment and completion, thereby determining the original time-aligned multidimensional measurement data stream.

[0027] After acquiring real-time measurement data, the real-time battery state is extracted. The real-time battery state includes an initial vector composed of voltage fluctuation amplitude, current load intensity, and temperature gradient. In this process, voltage is uniformly measured in V, current in A, and temperature in °C. All physical quantities are converted into digital signals by the sensor's built-in analog-to-digital converter and arranged according to a preset time sequence, forming the initial input for subsequent noise reduction and deconstruction processing. The real-time measurement data determined in this step has a clear source and can accurately reflect the instantaneous physical state of the battery during charging and discharging, providing highly reliable data support for subsequent identification of electrothermal coupling behavior.

[0028] In step S12, the real-time measurement data is denoised and deconstructed to obtain an instantaneous state vector. The coupling correlation value of the instantaneous state vector is calculated, a feature matrix is ​​constructed, and mapping processing is performed to determine the monitoring features, including: The real-time measurement data is subjected to denoising and deconstruction processing to obtain the instantaneous state vector; Calculate the coupling correlation value of the instantaneous state vector. If the coupling correlation value is greater than a preset correlation threshold, combine the instantaneous state vector according to the time series to generate a feature matrix. The feature matrix is ​​reorganized using a sliding window, and the reorganized result is spatially transformed to obtain the monitoring features.

[0029] First, the original measurement data stream is decomposed into multiple scales using wavelet transform. Then, by selecting appropriate basis functions, threshold shrinkage is applied to the high-frequency coefficients obtained from the decomposition to filter out electromagnetic interference noise generated during sensor acquisition, while retaining low-frequency signals characterizing battery state evolution. Subsequently, principal component analysis (PCA) is used to perform dimensionality reduction on the denoised multidimensional data. Specifically, the covariance matrix of the original data is constructed, its eigenvalues ​​and eigenvectors are solved, and the first three principal components are extracted in descending order of eigenvalue contribution rate, thus projecting the high-dimensional original space into a low-dimensional feature space. Through this decomposition process, the standard deviation of voltage fluctuation, peak current intensity, and maximum temperature gradient are extracted, collectively forming the instantaneous state vector.

[0030] The coupling correlation value is calculated by using the Pearson correlation coefficient between the voltage fluctuation amplitude and the temperature gradient. To determine this, the calculation formula is as follows: ; In the formula, Indicates the first The voltage fluctuation value at each sampling point is expressed in V. The arithmetic mean of voltage fluctuations; This represents the temperature gradient value at the corresponding sampling point, in °C. The arithmetic mean of the temperature gradient; The preset correlation calculation window length.

[0031] The correlation threshold is preset to 0.8. This threshold is based on statistical analysis of 500 sets of thermal runaway-induced experimental samples of 280Ah lithium iron phosphate (LFP) batteries and NCM811 ternary lithium batteries under high-rate charge / discharge, overcharge, and external heating conditions. Experimental results show that under normal operating conditions, the correlation coefficient between controlled heat generation and voltage fluctuation is usually between 0.3 and 0.6; when the correlation coefficient exceeds 0.8, it means that there is a significant local micro-short circuit or abnormal heat generation caused by the initial decomposition of the electrolyte inside the battery, and the electrothermal coupling effect is significantly enhanced. This threshold has strong universality for mainstream battery systems and can serve as a key physical threshold for identifying potential thermal failure risks. If the calculated coupling correlation value is greater than 0.8, the instantaneous state vectors of consecutive time moments are vertically stacked and combined according to the time sampling order to generate a feature matrix. The row dimension of this feature matrix corresponds to different timestamps, and the column dimension corresponds to the standard deviation of voltage fluctuation, peak current intensity, and maximum temperature gradient.

[0032] Then, the feature matrix is ​​reorganized using a sliding window. The preset width of the sliding window is... Step size is .width The sampling period is set to 100 cycles, corresponding to a complete observation interval for capturing dynamic changes in battery state; step size A sampling period of 10 is set to control the update frequency of the monitoring task. By sliding a window along the time axis on the feature matrix, the long sequence data is divided into multiple overlapping sub-block feature sets.

[0033] The recombined results are spatially transformed to obtain monitoring features. A pre-defined orthogonal mapping matrix is ​​then used. The sub-block feature set is projected onto a normalized coordinate space consisting of the battery's electrical and thermal dimensions. The orthogonal mapping matrix is ​​used. This linear transformation operator is determined by performing singular value decomposition on the electrothermal steady-state data of historical energy storage batteries at different charge-discharge rates. The configuration of this matrix aims to extract the maximum linear independence between components of each dimension, eliminating redundant information in the original data. During the spatial transformation process, linear transformations are performed on physical quantities of different dimensions, mapping them to... The dimensionless space within the range eliminates the influence of differences in the magnitude of voltage, current, and temperature data on subsequent model judgments, ultimately determining the monitoring characteristics. These characteristics can track the evolution trend of electrothermal coupling in real time, providing input support for the subsequent classification of abnormal modes.

[0034] In step S13, a dynamic correlation index is calculated using the monitoring features, and the dynamic correlation index is mapped to the feature space to obtain the optimal classification hyperplane. Based on the geometric distance deviation between the monitoring features and the optimal classification hyperplane, an anomaly vector is determined, including: The dynamic correlation index of the voltage, current and temperature data of the battery cell is calculated using the monitoring characteristics to obtain the coupling strength data; The coupling strength data is processed using a preset nonlinear mapping function to obtain the optimal classification hyperplane and the maximum margin distance in the feature space; Calculate the geometric distance between the monitored feature and the optimal classification hyperplane. If the geometric distance is less than the maximum interval distance, determine the anomaly vector based on the deviation of the geometric distance.

[0035] Dynamic correlation indices between battery cell voltage, current, and temperature data are calculated using monitoring features to obtain coupling strength data. Specifically, this is achieved by analyzing the synchronous evolution of voltage fluctuation amplitude, current load intensity, and temperature gradient within the monitoring features. In this process, the dynamic correlation coefficient between voltage fluctuation and temperature gradient is calculated using a sliding window with a preset step size of 1 second. The Pearson correlation coefficient between the voltage and temperature sampling sequences within this window is calculated to quantify the degree of matching between electrical and thermal fluctuations on the time axis. Simultaneously, the contribution rate of current load to battery temperature rise is calculated. This contribution rate is determined by establishing a heat generation model, converting the current load intensity acquired by the current sensor per unit time into Joule thermal power, and calculating the ratio of this power to the total temperature rise rate of the battery cell during that period. This determines the direct driving force of the electrical load on the increase in heat. The dynamic correlation coefficient characterizes the real-time tightness of how electrical parameter fluctuations guide changes in thermal parameters; the contribution rate reflects the driving force of current load intensity on heat accumulation. These parameters together constitute the coupling strength data, used to quantify the complex nonlinear interaction intensity within the electrothermal system.

[0036] The coupling strength data is processed using a pre-defined nonlinear mapping function to obtain the optimal classification hyperplane and maximum margin distance in the feature space. To address the linear inseparability of electrothermal coupling anomaly patterns in the original low-dimensional space, a kernel function is used to map the low-dimensional coupling strength data to a high-dimensional feature space. The kernel function, as a specific implementation of the nonlinear mapping, is pre-defined as a radial basis function. The expression for this kernel function is: ; In the formula, and Represents different feature vectors in the coupling strength data; This represents the preset kernel width parameter. The value range is set to 0.5 to 1.5. The specific determination process for this parameter is as follows: A historical dataset containing 2000 battery charge-discharge cycle samples is trained using 5-fold cross-validation. During cross-validation, the dataset is divided into 5 equal parts, with one part selected alternately as the validation set and the remaining 4 parts as the training set. A grid search algorithm is used to optimize within the range [0.5, 1.5] with a step size of 0.1. Finally, the value that achieves the highest classification accuracy and the most reasonable support vector ratio is selected as the fixed parameter. For example, in experiments targeting lithium iron phosphate battery systems, the final value selected is... In a high-dimensional feature space, the optimal classification hyperplane is determined by solving a quadratic programming problem. The optimal classification hyperplane is determined by the normal vector. and intercept The definition, its expression is: ; In the formula, This represents the high-dimensional feature vector after mapping. Simultaneously, the maximum margin distance is determined based on the geometric constraints between the support vectors and the hyperplane. The maximum margin distance defines the robustness boundary between the normal operating state and the potentially risky state.

[0037] Calculate the geometric distance between the monitored features and the optimal classification hyperplane. Project the real-time generated monitored features into a high-dimensional feature space, determine the corresponding feature point positions, and calculate the geometric distance from each feature point to the optimal classification hyperplane. The calculation formula is as follows: ; In the formula, The norm of the normal vector.

[0038] If the geometric distance is less than the maximum interval distance, the anomaly vector is determined based on the deviation of the geometric distance. When a feature point falls within the interval band defined by the maximum interval distance, it indicates that the battery state has left the absolutely safe region and entered a critical anomaly state. At this time, a mode state label, such as "critical anomaly" or "potential risk," is assigned to the feature point. Subsequently, a fractional mapping calculation is performed on the geometric distance to determine the geometric distance deviation. The calculation formula is as follows: ; In the formula, The calculated geometric distance; This is the preset maximum interval distance. Deviation. The range of values ​​is The closer the value is to 1, the closer the feature point is to the classification boundary.

[0039] Determine the degree of deviation Does it exceed the preset safety boundary threshold? The preset safety boundary threshold is 0.7. The logic behind determining this threshold is based on using a logistic regression model to analyze the evolutionary probability of critical outliers in historical experimental data. Statistical analysis revealed that when the deviation reaches 0.7, the probability of a feature point transforming into an actual thermal runaway node within 300 seconds follows a 90% confidence level of the cumulative distribution function (CDF). That is, under this threshold, the risk of missed detection is compressed to less than 10%, balancing the timeliness and accuracy of early warning. If the deviation... If the value exceeds 0.7, the key dimension combination causing the geometric deviation at that moment is extracted, including the instantaneous voltage offset value (in V), current surge value (in A), and temperature jump value (in °C), and this is reconstructed to determine the anomaly vector. This vector accurately pinpoints the key features of electrothermal coupling runaway, providing physically meaningful feature inputs for subsequent trend analysis.

[0040] In step S14, the anomaly vector is differentially processed to obtain incremental features. A trend model is constructed based on the incremental features and integrated to obtain the accumulated heat energy. The accumulated heat energy is then mapped to a preset risk classification table to determine the thermal failure risk level, including: The anomaly vector is subjected to a difference operation to obtain incremental features, and a model is constructed based on the incremental features to obtain a trend model; The cumulative amount of heat energy is obtained by performing integral calculations using the aforementioned trend model. The accumulated heat energy is mapped to a preset risk classification table to determine the risk level of thermal failure.

[0041] First, obtain the anomaly vector and calculate its magnitude. Modulus length The calculation formula is as follows: ; In the formula, The anomaly vector is represented in the th... The feature values ​​in each dimension include voltage deviation, current surge, and temperature change. This represents the total dimensions of the outlier vectors. Determine the magnitude. Is it greater than the preset modulus threshold? The preset modulus threshold is 2.0. This value is set based on the standard deviation of the electrothermal parameter fluctuation intensity recorded in the accelerated aging test of the energy storage cabinet. In the dimensionless feature space, 2.0 corresponds to a distance of approximately 2 standard deviations (2-sigma) from the cluster center under normal operating conditions, statistically representing that the current fluctuation has exceeded 95% of the normal random noise range. If the modulus... If the value is greater than 2.0, the corresponding electrothermal parameter channel is locked, and the time series data within that electrothermal parameter channel for a certain period of time is extracted.

[0042] Secondly, the time-series data within the electrothermal parameter channels are weighted and fused to determine the fused information sequence. A first-order difference operation is then performed on the fused information sequence to obtain the incremental features. The calculation process of the first-order difference operation is as follows: ; In the formula, This represents the fused information value at the current moment; This represents the fused information value from the previous sampling time. A preset sampling interval, with dimensions in seconds (s), is used. A dynamic trend model is constructed based on the obtained incremental features. The trend model is established through polynomial fitting or linear regression equations to capture the rate and direction of changes in the battery's electrothermal state.

[0043] Then, the cumulative heat energy is obtained by performing geometric area integration using a trend model. First, a baseline is predefined to characterize the ideal state of normal system operation. The baseline is defined based on the stable operating trajectory of the battery under the same state of charge (SOC) and the same load current intensity, according to a historical health operation database. This is achieved by extracting the average battery temperature rise rate from a large number of normal operating condition samples. steady-state average value of voltage fluctuation amplitude A weighted fusion algorithm was used to construct a time function curve reflecting the ideal electrothermal equilibrium state. This baseline serves as a comparison reference, representing the heat generation level of the battery under abnormal heat release conditions such as no internal short circuits or electrolyte decomposition. Next, the trend model curve was calculated. With baseline The area of ​​deviation within the observation time window. To address the issue of dimensional inconsistency, the area of ​​deviation is converted into a physically meaningful cumulative amount of thermal energy. The calculation process incorporates an equivalent heat conversion coefficient. This coefficient, determined through simulation experiments in an adiabatic calorimeter (ARC), maps the voltage drop area in the electrical dimension to the temperature rise area in the thermal dimension into a unified energy unit (joule). Its specific integration and conversion formulas are as follows: ; In the formula, This represents the accumulated thermal energy, measured in joules (J). This value quantifies the degree of excess energy accumulation inside the battery due to abnormal side reactions (such as micro-internal short circuits and SEI film decomposition). This is the equivalent heat conversion coefficient. In the calculation process, if the trend function represents the temperature rise deviation, then this coefficient includes the specific heat capacity of the individual battery cells. With quality The product; if it represents a deviation in electrical characteristics, then this coefficient is derived from the Ohm's heat loss equation. It is derived from [a specific source]. Its dimensions are based on [a specific source]. The physical quantity definitions are dynamically matched to ensure that the final result has the dimension J; This is a function of the trend model over time after feature fusion processing; This is a function of the baseline defined above changing over time; and : These represent the start and end times of the integration, respectively, with dimensions in seconds; the accumulated heat energy. It can accurately reflect the dynamic deterioration of the electrothermal state.

[0044] Finally, the accumulated heat energy Mapping to a pre-defined risk grading table, the thermal failure risk level is determined. This risk grading table is established through the following steps: First, accelerated aging experiments and overcharge heat dissipation experiments under extreme conditions are conducted on individual energy storage battery cells at different charge / discharge rates. During the experiments, key nodes of the internal chemical reactions of the battery are recorded simultaneously, such as the moment when the solid electrolyte interphase (SEI) membrane begins to decompose, the separator melts and shrinks, and the moment when the electrolyte undergoes severe oxidation. The total energy accumulated by the battery at these physical nodes is calculated using the energy balance equation, serving as the original basis for determining the risk boundary. Second, based on the cumulative heat distribution characteristics extracted from a large number of experimental samples, a normal distribution probability model is used to determine the quantification range of different risk levels. The first-level risk threshold is set based on the critical energy value at which the battery begins to generate self-generated heat but can still achieve thermal equilibrium through an external heat dissipation system; the second-level risk threshold is set based on the key energy value at which battery side reactions accelerate significantly, the temperature rise rate changes abruptly, and the heat is difficult to eliminate by conventional means; the third-level risk threshold corresponds to the initial energy value at which the battery enters the irreversible thermal runaway evolution stage.

[0045] In this embodiment, the risk classification table presets multiple cumulative ranges. The specific division logic is as follows: if the cumulative heat energy... In If it falls within the specified range, it is classified as a Level 1 risk; if the accumulated heat energy... In If it falls within the specified range, it is classified as a level two risk; if the accumulated heat energy... A value exceeding 500 is classified as Level 3 risk, with the accumulated heat energy measured in J. This classification method visually represents the severity of the abnormal state and provides a logical basis for subsequent precise interventions based on risk levels. For example, when the system is determined to be at Level 1 risk, a strategy of reducing operating power can be adopted; when it enters Level 2 or Level 3 risk, an emergency shutdown or fire suppression command is immediately triggered. By establishing this quantitative risk assessment standard, the early warning logic has shifted from qualitative judgment to quantitative analysis.

[0046] In step S15, based on the thermal failure risk level, similar sequences that satisfy a preset spatial proximity condition with the anomaly vector are selected and a topology network is constructed. An evolution path is generated by combining the propagation probability distribution of the topology network, including: Extract historical state snapshots corresponding to the thermal failure risk level, map the historical state snapshots to a multi-dimensional space, and filter out similar sequences that satisfy the preset spatial proximity condition with the anomaly vector; The similar sequences are used as nodes to construct a topology network, and the probability of each state branch in the topology network evolving into a thermal failure state is calculated to determine the propagation probability distribution. By matching the propagation probability distribution with the real-time gradient direction of the anomaly vector, the evolution branch with the highest probability is locked, and an evolution path is generated.

[0047] First, based on the determined thermal failure risk level, the historical operating status database is retrieved, and a set of historical state snapshots from the same source is extracted. The retrieval process is a precise data filtering based on risk level; the system prioritizes retrieving operating segments with the same risk level as the current one. For example, when the current risk level is determined to be level two, the system filters out historical records of voltage, current, and temperature operations marked as level two risk for analysis reference.

[0048] Subsequently, the historical state snapshot set is mapped to a multi-dimensional state space. The mapping process consists of two stages: data standardization and coordinate transformation. In the data standardization stage, the voltage values ​​in the historical state snapshots are... Current and temperature The data undergoes max-min normalization, linearly scaling the original values ​​of each physical quantity to dimensionless values ​​within the range [0,1]. During the coordinate transformation stage, the normalized voltage, current, and temperature values ​​at each moment are defined as three independent axis vectors in an orthogonal coordinate system. The electrothermal state at each sampling moment is converted into the coordinates of a geometric point in a three-dimensional state space, thus transforming the temporal historical state snapshot set into a state point cloud with geometric distribution characteristics in space.

[0049] Secondly, similar sequences that satisfy the preset spatial proximity condition with the anomaly vector are selected. The geometric distance between the current anomaly vector and each historical state point in the state point cloud is calculated using Euclidean distance. The calculation formula is as follows: ; In the formula, Indicates the current anomaly vector at the th position. Feature values ​​in each dimension; Indicates the historical state point at the 1st Feature values ​​in each dimension; This represents the total number of feature dimensions. In this embodiment, The value corresponds to the number of extracted electrothermal features (such as voltage offset, current surge, temperature change, etc.).

[0050] The preset spatial proximity condition refers to the geometric distance. The distance threshold is less than a preset threshold of 0.5. This threshold is set based on the analysis results of DBSCAN (density-based spatial clustering algorithm) applied to historical data. First, mixed point cloud data containing historical normal operation samples and thermal failure evolution samples is retrieved. The DBSCAN algorithm is used to cluster the point cloud. This algorithm automatically identifies high-density regions in the data by setting the neighborhood radius Eps and the minimum number of core points MinPts. Cluster analysis reveals that historical "normal fluctuation" data is usually concentrated in several high-density core clusters, while "abnormal evolution" data shows characteristics of spreading towards the cluster edge or forming isolated low-density noise points. Next, the average radius of the normal core clusters and the minimum separation distance between the normal clusters and the thermal failure initiation point are calculated. Statistical analysis shows that in the normalized three-dimensional state space, 95% of the normal fluctuation points are less than 0.3 from their cluster centers, while when a state point enters the evolution risk zone, its geometric distance to the normal core cluster increases significantly. To effectively distinguish between "normal fluctuations" and "abnormal evolution," the distance threshold is set to 0.5. This value is an equilibrium point determined through optimization using the ROC curve (Receiver Operating Characteristic curve). At this threshold, the system can accurately capture all known historical thermal failure precursor sequences with a low false alarm rate. Choosing 0.5 as the threshold ensures that the identified historical state sequences have the highest degree of physical evolution similarity to the current anomalous pattern.

[0051] Using the selected similar sequences as basic data, a topological network characterizing the evolution logic of the electrothermal state is constructed. In this process, each independent electrothermal coupled state is defined as a node in the network. These nodes contain a combination of voltage fluctuations, current loads, and temperature gradients at a specific moment. To improve computational efficiency, a clustering algorithm is used to center the state points in the similar sequences, merging spatially close state points into a representative state node. By analyzing the evolution order of historical similar sequences on the time axis, if state A in the historical record transitions to state B at the next sampling moment, a directed edge is established in the topological network from node A to node B. This directed edge represents the physical trajectory of the battery's electrothermal state evolution over time. By statistically analyzing the frequency of each node evolving to different subsequent nodes in the historical data, the conditional probability of state transition is calculated. The specific transition probabilities are... The calculation formula is as follows: ; In the formula: This indicates that in a historically similar sequence, the state... Directly transfer to state The number of times; Indicates from state The total number of all possible evolutionary branches starting from the state; the denominator represents the number of branches starting from the state. The total number of transitions from the starting point. Through the above calculations, a probability weight is assigned to each directed edge in the topology network, thus constructing a complete propagation probability distribution matrix. In this matrix, several endpoint nodes representing thermal failure states are predefined; these endpoint nodes correspond to the extreme electrothermal characteristic parameters when the battery experiences thermal runaway. Taking a specific secondary risk evolution as an example, if the currently detected abnormal state corresponds to a node in the topology network… By retrieving the propagation probability distribution, it was found that the node has three evolutionary branches: Branch 1 points to a severely anomalous state node, accounting for 70% of the total transitions, with a propagation probability of 0.7; Branch 2 points to a risk-mitigated state node, accounting for 20% of the transitions, with a probability of 0.2; and Branch 3 points to a node maintaining the status quo, with a probability of 0.1. Through this probability quantification method, the system can clearly identify the degree of danger of the current anomalous state evolving towards extreme thermal failure.

[0052] Finally, the constructed topology network and its corresponding propagation probability distribution are used as the basis for matching, combined with the real-time gradient direction of the anomaly vector. The gradient direction is determined by calculating the rate of change vectors of current voltage, current, and temperature. By calculating the cosine of the angle between the gradient vector and each evolutionary branch in multidimensional space, the path with the highest directional consistency and the highest propagation probability is selected, ultimately generating the potential threat evolution path. This path not only has empirical support from historical data but also incorporates the current transient trend, providing a structured evolutionary trajectory for subsequent accurate time-series prediction using a long short-term memory network.

[0053] In step S16, a time-series prediction is performed on the evolution path to obtain a prediction result. The alignment error between the prediction result and the real-time measurement data is calculated. Time-series calibration is performed based on the alignment error to determine an optimized interaction mode, including: The node attributes of the evolution path are input into a pre-trained long short-term memory network. The cell state is updated using the forget gate and the input gate, and the hidden layer state reflecting the temporal dependency is determined by combining the output gate. The hidden layer state is mapped and regressed using the fully connected layer of the Long Short-Term Memory network to generate a thermal failure risk probability curve and obtain the prediction result. The alignment error is obtained by comparing the values ​​of key time points in the prediction results with the corresponding times in the real-time measurement data. If the alignment error exceeds the preset error range, then timestamp realignment and feature value interpolation are performed on the measurement data according to the alignment error to reconstruct the optimized monitoring features; Calculate the correlation coefficients of different dimensional attributes in the optimized monitoring features, construct the interaction correlation matrix, and obtain the optimized interaction mode.

[0054] The node attributes of the evolution path are extracted and input into a pre-trained Long Short-Term Memory (LSTM) network. The LSTM network model used in this embodiment is a deep recurrent neural network containing an input layer, hidden layers, and a fully connected output layer. The specific structure is as follows: the input layer receives a sequence of node attributes with three channels; the hidden layer contains two stacked LSTM units, each with 64 hidden nodes, using recurrent connections to capture the nonlinear temporal dependencies in the electrothermal evolution process; a dropout layer with a dropout rate of 0.2 is connected after the hidden layer to prevent overfitting during training; finally, a fully connected layer is connected to map high-dimensional features to single-dimensional risk probability values.

[0055] The training process for this Long Short-Term Memory (LSTM) network is supervised learning. Training data is derived from a historical operational database of energy storage systems under various conditions, containing over 5000 complete battery cycle samples covering normal, degradation, and thermal runaway evolution. During training, node features from the historical evolution path are used as input sequences, and the corresponding real thermal failure risk labels, evaluated by experts or verified experimentally, are used as the target output. Mean Squared Error (MSE) is employed as the loss function, and the Adam optimizer is used for iterative training on a large dataset until the model converges. The trained model can output the risk evolution trend within future time steps based on the input evolution path features.

[0056] Node attributes include the rate of temperature change, voltage fluctuation amplitude, and duration of current anomalies at each node along the evolutionary path. Long Short-Term Memory (LSTM) networks process attribute sequences step-by-step, using a forget gate to control the retention of the previous cell state and an input gate to control the weighting of current input information into the cell state, thereby updating the current cell state. Its cell state... The update logic is as follows: ; In the formula, Forgetting gate vector; The input gate vector; This represents the cell state at the previous moment; The candidate cell state to be written; This represents element-wise multiplication. Subsequently, the updated cell states are filtered using the output gating vector to determine the hidden layer states that reflect long-term dependencies and short-term fluctuations.

[0057] Mapping regression is performed on the hidden layer states using the fully connected layers of a Long Short-Term Memory (LSTM) network. First, hidden state vectors reflecting the electrothermal temporal dependency are extracted from the terminal hidden layers of the LSM network. The hidden state vector is a 64-dimensional feature set that highly condenses the nonlinear correlation information of voltage fluctuations, current load, and temperature gradient evolution of the battery over a period of time. This hidden state vector is then input into a predefined linear transformation operator. This linear transformation operator consists of a weight matrix... With bias term The structure involves performing a compression transformation on the high-dimensional latent features in space, projecting them onto a single-dimensional regression space. In this process, the weight matrix... The dimension is preset as Bias term It is a one-dimensional scalar. Its specific mathematical calculation logic is as follows: ; In the formula, It represents the risk probability value determined at the current prediction step size, and is a dimensionless percentage value. This represents the Sigmoid activation function, which forces the output of the linear transformation to be within the range [0,1], thus conforming to the definition of a probability distribution. This represents the weight matrix, where each element represents the weight coefficient of the contribution of different dimensions of features in the hidden state to the final thermal failure risk. This represents a 64-dimensional state vector output from the hidden layer that reflects the electrothermal timing characteristics. This represents the bias term, used to perform baseline offset correction on the prediction results. The weight matrix above... With bias term The specific value was determined during the model pre-training stage by using the mean squared error (MSE) as the loss function and training through error backpropagation on a historical database containing a large number of battery thermal runaway samples.

[0058] By repeatedly performing the above mapping regression calculation over multiple preset time steps (e.g., 60 consecutive steps), the system can output a continuous thermal failure risk probability curve. This risk probability curve characterizes the probability evolution trend of the battery's internal temperature exceeding the preset safety limit or thermal runaway occurring within a specific future time period (e.g., 30 minutes or 60 minutes).

[0059] The alignment error is determined by comparing the predicted results with the values ​​of real-time measurement data. The alignment error reflects the deviation between the predicted trend and the actual sensor data stream in the time dimension. Specifically, it is calculated by extracting key time points (such as the point of maximum temperature rise slope or the predicted peak point) from the predicted probability curve and retrieving the corresponding actual observed values ​​from the real-time measurement data stream. The difference between the two is then calculated to obtain the time lag bias value. For example, if the predicted peak occurs later than the actual observed peak, a time delay bias is determined, with dimensions in seconds (s).

[0060] If the alignment error exceeds the preset error range, feedback calibration is performed on the evolution path based on this alignment error, rather than modifying the original real-time measurement data. The preset error range is determined based on the system response tolerance and is preset to 1 second. When the hysteresis deviation exceeds the range, a predicted phase compensation vector is generated. First, the calculated alignment error is extracted. This is used as a time-axis correction scalar. Simultaneously, the ratio of the peak slope of the predicted risk probability curve to the actual peak slope of characteristic parameters (such as the temperature rise rate) in the real-time measurement data is calculated to determine the amplitude scaling factor of the predicted sequence. The time-axis correction scalar and the amplitude scaling factor are combined to construct a predicted phase compensation vector. This vector is used to perform phase lead or lag compensation on the evolution path nodes of the Long Short-Term Memory network output. Specifically, this is achieved by using the time index of the predicted sequence... Unified subtraction of alignment error This allows for the translation of the predicted evolution trend along the time axis, thereby aligning the prediction results with physical reality on a time scale without altering the accuracy of real-time measurement data. During calibration, if non-uniform sampling points occur, a linear interpolation algorithm is used to complete the predicted sequence. The calculation logic involves filling in the missing time points... Find two adjacent prediction points before and after. and Through the formula: ; The calibrated optimized evolutionary sequence was calculated. This process eliminated prediction bias caused by model response delay, ensuring baseline consistency for subsequent correlation analyses.

[0061] Subsequently, the correlation coefficients of different dimensional attributes between the optimized evolution sequence and the real-time measurement data are calculated to construct an interaction correlation matrix. First, the real-time measurement data stream containing voltage, current, and temperature components is extracted. Using a preset sliding analysis window, the Pearson correlation coefficients of any two dimensional attributes between the real-time measurement sequence and the calibrated optimized evolution sequence are calculated. During this process, the length of the sliding analysis window is set. For a voltage sequence, the sampling period is 50. With temperature sequence Its correlation coefficient The calculation formula is as follows: ; In the formula, and These are the arithmetic mean of the voltage and temperature eigenvalues ​​within the window, respectively. By performing the same calculation on all dimension pairs, a symmetric interaction correlation matrix is ​​constructed. .

[0062] For this interaction correlation matrix, the optimized interaction pattern is determined by identifying highly correlated combinations. To eliminate the abrupt change between the "matrix" and the "vector" descriptions in subsequent steps, this embodiment clarifies that the optimized interaction pattern is a weighted feature expression derived from the interaction correlation matrix. Specifically, the determination process involves first analyzing the matrix... The non-diagonal elements in the equation. If a certain element (such as the correlation coefficient between voltage and temperature)... If the absolute value exceeds the preset coupling enhancement threshold of 0.85, which is a statistical verification result based on the correlation between strongly coupled events and final thermal failure in historical data, then the dominant influence path under the current operating condition is identified. Based on this path, the system generates a set of dynamic weight vectors. Dimensions related to the dominant path are assigned higher weights. Finally, this dynamic weight vector is used to perform a weighted mapping on the real-time measurement data, calculated as follows: ; In the formula, This is the original real-time measurement vector. Through the product operation of this matrix and the vector, the complex interaction correlation matrix is ​​transformed into an interaction vector containing dynamic coupling information. .

[0063] In step S17, the cross-correlation value between the optimized interaction mode and the real-time measurement data is calculated, the maximum eigenvalue of the cross-correlation result is extracted and quantified to obtain a risk score. If the risk score exceeds a preset score threshold, a pulse-coded sequence is generated, and a warning signal is output, including: A nonlinear mapping is performed between the optimized interaction mode and the real-time measurement data to obtain an interaction vector and a state vector. Calculate the cross-correlation coefficient between the interaction vector and the state vector to obtain the cross-correlation value; The cross-correlation values ​​are arranged in a matrix according to the time index to obtain the fusion matrix; The risk score is obtained by extracting the largest eigenvalue from the fusion matrix and performing normalization. If the risk score exceeds a preset score threshold, a pulse code is generated based on a preset combination of pulse width and interval, and a warning signal is output.

[0064] The interaction mode and real-time measurement data are processed and optimized using a pre-defined nonlinear mapping function to obtain interaction vectors and state vectors. The optimized interaction mode reflects the dynamic coupling strength between different sensors, while the real-time measurement data includes the actual voltage, current, and temperature of the battery during operation. During processing, the hyperbolic tangent function tanh is used as the nonlinear mapping function to project the multidimensional interaction influence mode into a high-dimensional interaction vector, while simultaneously projecting the battery state data into a state vector. This nonlinear transformation amplifies the minute fluctuations of electrothermal parameters near the critical failure point, avoids feature erasure caused by linear superposition, and thus more completely preserves key nonlinear state information.

[0065] The cross-correlation coefficient between the interaction vector and the state vector is calculated to obtain the cross-correlation value. The cross-correlation value characterizes the similarity and phase delay relationship between the interaction vector and the state vector over time. Specifically, the cross-correlation coefficient between the two vectors at different time delays is calculated using a preset sliding time window. The calculation process follows the formula below: ; In the formula, Represents the interaction vector at the current moment; Indicates the time offset The state vector below, where The preset value range is One sampling period; This represents the total number of sampling points within the sliding window, preset to 100. A higher cross-correlation value indicates a closer morphological similarity between the current battery state and the potential thermal failure mode predicted by the model, i.e., a stronger match.

[0066] The calculated cross-correlation values ​​are arranged into a matrix according to the time index to obtain the fusion matrix. Specifically, the different offsets within each sliding window are... The cross-correlation results are used as column elements, and the calculation results of continuous time steps are used as row elements to construct a fusion matrix. The fusion matrix comprehensively integrates the correlation information between electrothermal interaction characteristics and instantaneous battery operating state in the spatiotemporal dimension, and the numerical distribution pattern within the matrix directly reflects the evolution intensity of abnormal energy on the time axis.

[0067] The risk score is obtained by extracting the largest eigenvalue from the fusion matrix and normalizing it. First, singular value decomposition (SVD) is performed on the fusion matrix to identify the set of eigenvalues ​​that represent the energy distribution of the principal components of the matrix. The largest eigenvalue is then extracted from this set. It is then mapped to a numerical range of 0 to 100 using a preset normalization function, as shown in the following formula: ; In the formula, The eigenvalue benchmark is the eigenvalue corresponding to the system background noise. The characteristic value standard is the preset threshold for triggering thermal runaway. The resulting risk score is the final quantitative indicator for measuring the threat of thermal failure.

[0068] It is worth noting that the eigenvalue benchmark The method for determining the eigenvalue is as follows: Measurement data from the energy storage system during historical healthy operation phases (i.e., charge-discharge cycles without abnormal warnings and where temperature fluctuations are within the normal ambient temperature rise range) are retrieved. These health data are then subjected to the same cross-correlation and matrix fusion processing as in the current step, extracting the average or median of the largest eigenvalue of the fusion matrix from a large number of healthy samples. This value represents the characteristic energy of the noise floor generated by sensor measurement errors, system environmental noise, and normal electrothermal fluctuations. Eigenvalue Standard The determination method is based on the full life cycle data recorded from experiments on thermal runaway triggering of individual energy storage battery cells (such as overcharge experiments, external nail penetration or heating experiments); by analyzing the experimental data, the starting moment when the battery enters the irreversible exothermic stage is identified (e.g., the sudden change in temperature rise rate exceeds [a certain threshold]). The moment of the critical moment is used to extract the fusion matrix corresponding to that critical moment and calculate its maximum eigenvalue. This value represents the energy intensity when the internal heat generation and electrical anomaly of the battery reach the failure critical point, and serves as the upper limit benchmark for normalization calculation.

[0069] If the risk score exceeds a preset threshold, a pulse code is generated based on a preset combination of pulse width and interval, outputting the final warning signal. The preset threshold is 85 points, a value determined based on thermal runaway criticality experimental data of the energy storage system at different aging stages. When the risk score reaches or exceeds 85 points, the system selects the corresponding physical pulse coding logic based on the risk score's range. Level 1 Warning (85-90 points): Generates three consecutive short pulses, each with a width of 100ms and an interval of 200ms, indicating a slight anomaly and prompting the monitoring backend to pay closer attention.

[0070] Level 2 warning (90-95 points): Generates two long pulses and one short pulse. The long pulse width is 500ms and the short pulse width is 100ms, with an interval of 200ms. This indicates a moderate risk, and the system will automatically trigger a power reduction operation command.

[0071] Level 3 warning (95 points or above): five consecutive long pulses with equal intervals are generated, with a pulse width of 800ms and an interval of 300ms, representing an emergency thermal failure risk.

[0072] This pulse coding method directly acts on the hard-wired control loop of the energy storage cabinet, which facilitates downstream actuators to quickly identify the risk level and take intervention measures such as power limiting, power cut-off, or activation of the fire extinguishing system. This significantly improves the communication reliability and real-time response speed of the entire early warning system in complex electromagnetic environments.

[0073] In summary, this invention discloses a method for early warning of thermal failure in energy storage cabinets. It involves collecting electrothermal characteristic data of the energy storage cabinet batteries and extracting real-time battery status; denoising and deconstructing the data to obtain instantaneous vectors; reconstructing the mapping to determine monitoring features when a coupling correlation threshold is met; calculating the electrothermal coupling strength; mapping to a high-dimensional space to determine the optimal hyperplane and anomaly vectors; locking parameter channels; trend modeling; calculating the accumulated thermal energy to quantify and determine the thermal failure risk level; retrieving historical snapshots to filter similar states; constructing a topological network; generating evolution paths; using gated cyclic logic for time-series prediction; reconstructing monitoring parameters based on alignment errors to optimize the interaction mode; and finally calculating the cross-correlation values ​​between the interaction mode and the real-time battery status to generate a fusion matrix, extracting the maximum eigenvalue for quantification scoring, and outputting a pulse warning signal. This method addresses the problem of poor timeliness in early warning of thermal failure in energy storage cabinets in existing technologies.

[0074] Reference Figure 2 The second embodiment of the present invention provides a thermal failure early warning system for a safe energy storage cabinet, comprising: The data acquisition module is used to acquire real-time measurement data, which includes the voltage, current and temperature data of individual battery cells. The feature extraction module is used to denoise and deconstruct the real-time measurement data to obtain an instantaneous state vector, calculate the coupling correlation value of the instantaneous state vector, construct a feature matrix and perform mapping processing to determine the monitoring features; An anomaly detection module is used to calculate dynamic correlation indicators using the monitoring features, map the dynamic correlation indicators to the feature space to obtain the optimal classification hyperplane, and determine the anomaly vector based on the geometric distance deviation between the monitoring features and the optimal classification hyperplane. The risk classification module is used to perform differential operations on the anomaly vector to obtain incremental features, construct a trend model based on the incremental features and perform integration to obtain the heat energy accumulation, and map the heat energy accumulation to a preset risk classification table to determine the thermal failure risk level. The path evolution module is used to select similar sequences that meet the preset spatial proximity conditions with the anomaly vector according to the thermal failure risk level and construct a topology network, and generate an evolution path by combining the propagation probability distribution of the topology network. The interaction optimization module is used to perform time-series prediction on the evolution path to obtain prediction results, calculate the alignment error between the prediction results and the real-time measurement data, perform time-series calibration based on the alignment error, and determine the optimized interaction mode. The early warning output module is used to calculate the cross-correlation value between the optimized interaction mode and the real-time measurement data, extract the maximum feature value of the cross-correlation result and perform quantitative scoring to obtain a risk score. If the risk score exceeds a preset score threshold, a pulse code sequence is generated and an early warning signal is output.

[0075] It should be noted that the thermal failure early warning system for the safe energy storage cabinet provided in this embodiment of the invention is used to execute all the process steps of the thermal failure early warning method for the safe energy storage cabinet in the above embodiment. The working principle and beneficial effect of the two are one-to-one, so they will not be described again.

[0076] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a thermal failure early warning program for a safe energy storage cabinet. When the processor executes the computer program, it implements the steps described in the various embodiments of the thermal failure early warning method for safe energy storage cabinets, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module in the above system embodiments, such as the data acquisition module.

[0077] For example, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0078] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0079] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0080] The memory can be used to store the computer program or module. The processor implements various functions of the electronic device by running or executing the computer program or module stored in the memory and calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0081] If the modules integrated into the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or system capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0082] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0083] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for early warning of thermal failure in a safe energy storage cabinet, characterized in that, include: Acquire real-time measurement data, which includes voltage, current, and temperature data of individual battery cells; The real-time measurement data is denoised and deconstructed to obtain an instantaneous state vector. The coupling correlation value of the instantaneous state vector is calculated, a feature matrix is ​​constructed, and a mapping process is performed to determine the monitoring features. The monitoring features are used to calculate dynamic correlation indicators, and the dynamic correlation indicators are mapped to the feature space to obtain the optimal classification hyperplane. The anomaly vector is determined based on the geometric distance deviation between the monitoring features and the optimal classification hyperplane. The incremental features are obtained by performing a difference operation on the anomaly vector. A trend model is constructed based on the incremental features and integrated to obtain the heat energy accumulation. The heat energy accumulation is mapped to a preset risk classification table to determine the thermal failure risk level. Based on the thermal failure risk level, similar sequences that meet the preset spatial proximity conditions with the anomaly vector are selected and a topological network is constructed. An evolution path is generated by combining the propagation probability distribution of the topological network. A time-series prediction is performed on the evolution path to obtain a prediction result. The alignment error between the prediction result and the real-time measurement data is calculated. A time-series calibration is performed based on the alignment error to determine an optimized interaction mode. The cross-correlation value between the optimized interaction mode and the real-time measurement data is calculated, the maximum feature value of the cross-correlation result is extracted and quantified to obtain a risk score. If the risk score exceeds a preset score threshold, a pulse code sequence is generated and an early warning signal is output.

2. The method for early warning of thermal failure of a safety energy storage cabinet according to claim 1, characterized in that, The process of denoising and deconstructing the real-time measurement data to obtain an instantaneous state vector, calculating the coupling correlation value of the instantaneous state vector, constructing a feature matrix, and performing mapping processing to determine the monitoring features includes: The real-time measurement data is subjected to denoising and deconstruction processing to obtain the instantaneous state vector; Calculate the coupling correlation value of the instantaneous state vector. If the coupling correlation value is greater than a preset correlation threshold, combine the instantaneous state vector according to the time series to generate a feature matrix. The feature matrix is ​​reorganized using a sliding window, and the reorganized result is spatially transformed to obtain the monitoring features.

3. The method for early warning of thermal failure of a safety energy storage cabinet according to claim 1, characterized in that, The process of calculating dynamic correlation indicators using the monitoring features and mapping these indicators to a feature space to obtain an optimal classification hyperplane, and determining anomaly vectors based on the geometric distance deviation between the monitoring features and the optimal classification hyperplane, includes: The dynamic correlation index of the voltage, current and temperature data of the battery cell is calculated using the monitoring characteristics to obtain the coupling strength data; The coupling strength data is processed using a preset nonlinear mapping function to obtain the optimal classification hyperplane and the maximum margin distance in the feature space; Calculate the geometric distance between the monitored feature and the optimal classification hyperplane. If the geometric distance is less than the maximum interval distance, determine the anomaly vector based on the deviation of the geometric distance.

4. The method for early warning of thermal failure of a safety energy storage cabinet according to claim 1, characterized in that, The incremental features are obtained by performing a difference operation on the anomaly vector. A trend model is constructed based on the incremental features and integrated to obtain the accumulated heat energy. The accumulated heat energy is then mapped to a preset risk classification table to determine the thermal failure risk level, including: The anomaly vector is subjected to a difference operation to obtain incremental features, and a model is constructed based on the incremental features to obtain a trend model; The cumulative amount of heat energy is obtained by performing integral calculations using the aforementioned trend model. The accumulated heat energy is mapped to a preset risk classification table to determine the risk level of thermal failure.

5. The method for early warning of thermal failure of a safe energy storage cabinet according to claim 1, characterized in that, The step of selecting similar sequences that satisfy a preset spatial proximity condition with the anomaly vector based on the thermal failure risk level and constructing a topology network, and generating an evolution path by combining the propagation probability distribution of the topology network, includes: Extract historical state snapshots corresponding to the thermal failure risk level, map the historical state snapshots to a multi-dimensional space, and filter out similar sequences that satisfy the preset spatial proximity condition with the anomaly vector; The similar sequences are used as nodes to construct a topology network, and the probability of each state branch in the topology network evolving into a thermal failure state is calculated to determine the propagation probability distribution. By matching the propagation probability distribution with the real-time gradient direction of the anomaly vector, the evolution branch with the highest probability is locked, and an evolution path is generated.

6. The method for early warning of thermal failure of a safety energy storage cabinet according to claim 1, characterized in that, The process of performing time-series prediction on the evolution path to obtain prediction results, calculating the alignment error between the prediction results and the real-time measurement data, performing time-series calibration based on the alignment error, and determining an optimized interaction mode includes: The node attributes of the evolution path are input into a pre-trained long short-term memory network. The cell state is updated using the forget gate and the input gate, and the hidden layer state reflecting the temporal dependency is determined by combining the output gate. The hidden layer state is mapped and regressed using the fully connected layer of the Long Short-Term Memory network to generate a thermal failure risk probability curve and obtain the prediction result. The alignment error is obtained by comparing the values ​​of key time points in the prediction results with the corresponding times in the real-time measurement data. If the alignment error exceeds the preset error range, then timestamp realignment and feature value interpolation are performed on the measurement data according to the alignment error to reconstruct the optimized monitoring features; Calculate the correlation coefficients of different dimensional attributes in the optimized monitoring features, construct the interaction correlation matrix, and obtain the optimized interaction mode.

7. The method for early warning of thermal failure of a safe energy storage cabinet according to claim 1, characterized in that, The process involves calculating the cross-correlation value between the optimized interaction mode and the real-time measurement data, extracting the maximum eigenvalue of the cross-correlation result, and quantifying and scoring it to obtain a risk score. If the risk score exceeds a preset threshold, a pulse-coded sequence is generated, and a warning signal is output, including: A nonlinear mapping is performed between the optimized interaction mode and the real-time measurement data to obtain an interaction vector and a state vector. Calculate the cross-correlation coefficient between the interaction vector and the state vector to obtain the cross-correlation value; The cross-correlation values ​​are arranged in a matrix according to the time index to obtain the fusion matrix; The risk score is obtained by extracting the largest eigenvalue from the fusion matrix and performing normalization. If the risk score exceeds a preset score threshold, a pulse code is generated based on a preset combination of pulse width and interval, and a warning signal is output.

8. A thermal failure early warning system for a safe energy storage cabinet, characterized in that, include: The data acquisition module is used to acquire real-time measurement data, which includes the voltage, current and temperature data of individual battery cells. The feature extraction module is used to denoise and deconstruct the real-time measurement data to obtain an instantaneous state vector, calculate the coupling correlation value of the instantaneous state vector, construct a feature matrix and perform mapping processing to determine the monitoring features; An anomaly detection module is used to calculate dynamic correlation indicators using the monitoring features, map the dynamic correlation indicators to the feature space to obtain the optimal classification hyperplane, and determine the anomaly vector based on the geometric distance deviation between the monitoring features and the optimal classification hyperplane. The risk classification module is used to perform differential operations on the anomaly vector to obtain incremental features, construct a trend model based on the incremental features and perform integration to obtain the heat energy accumulation, and map the heat energy accumulation to a preset risk classification table to determine the thermal failure risk level. The path evolution module is used to select similar sequences that meet the preset spatial proximity conditions with the anomaly vector according to the thermal failure risk level and construct a topology network, and generate an evolution path by combining the propagation probability distribution of the topology network. The interaction optimization module is used to perform time-series prediction on the evolution path to obtain prediction results, calculate the alignment error between the prediction results and the real-time measurement data, perform time-series calibration based on the alignment error, and determine the optimized interaction mode. The early warning output module is used to calculate the cross-correlation value between the optimized interaction mode and the real-time measurement data, extract the maximum feature value of the cross-correlation result and perform quantitative scoring to obtain a risk score. If the risk score exceeds a preset score threshold, a pulse code sequence is generated and an early warning signal is output.

Images