Lithium ion battery fire prediction and control method and system

Abstract

The application provides a lithium ion battery fire prediction and control method and system, relates to the technical field of battery monitoring, and regards each battery cell in a battery module as a graph node, constructs an initial adjacency matrix, and dynamically updates edge weights in combination with multi-dimensional feature data to form a dynamic adjacency matrix; the dynamic graph attention convolution network is trained, the feature data and the dynamic adjacency matrix in a historical time window are combined, the thermal runaway probability and the heat release rate of each battery cell at the next moment are predicted, the deviation probability of the current feature data is calculated based on the Gaussian distribution parameters under the health state, and then the multi-modal risk entropy and the comprehensive abnormality degree are obtained; the comprehensive fire risk index is constructed based on the comprehensive abnormality degree, the thermal runaway probability prediction value and the heat release rate prediction value, the actual risk index is determined in combination with the module overall risk index, different control strategies are executed according to the actual risk index, and early prediction and differentiated intervention of the battery fire risk are realized.

Description

Technical Field

[0001] This invention relates to the field of battery monitoring technology, specifically to a method and system for predicting and controlling lithium-ion battery fires. Background Technology

[0002] With the widespread application of electric vehicles and energy storage systems, lithium-ion batteries have become the mainstream choice due to their high energy density and long cycle life. However, lithium-ion batteries are highly susceptible to thermal runaway under extreme conditions such as overcharging and internal short circuits, which can lead to fires or even explosions. Traditional battery management systems mainly rely on single threshold judgments to trigger alarms or protection actions, which cannot accurately predict the evolution trend of thermal runaway and ignore the thermal-electric coupling relationship between cells within the battery module and the dynamic characteristics of fault propagation. This makes it difficult to achieve early warning and graded intervention before a fire occurs, resulting in delayed warnings or frequent false alarms and missed alarms. Therefore, how to construct a dynamic and accurate fire prediction model based on the spatiotemporal correlation characteristics between cells and formulate differentiated control strategies accordingly has become a key problem that urgently needs to be solved in current lithium-ion battery fire prevention and control technology.

[0003] In the prior art, CN119087227A discloses a method, device, and medium for predicting and controlling lithium-ion battery fires, including: extracting features from the operating data and environmental data within the battery module, and merging the extracted features in a time and space sequence; determining a battery evaluation model using a machine learning algorithm based on a constructed lithium-ion battery operating feature database; evaluating the battery using the evaluation model; identifying lithium-ion batteries that deviate from the normal range using an anomaly detection algorithm based on the evaluation results; continuing to monitor the monitored object, and when the evaluation result deviates from the normal range by more than a set threshold, issuing an alarm and controlling the opening of the solenoid valve in the lithium battery fire suppression device to cut off the power to the lithium-ion battery.

[0004] The main problems with the above solutions are: they do not consider the mutual influence between different cells within the battery module due to factors such as thermal diffusion and uneven current distribution; they do not clearly define how to quantify the distance between cells and the thermal-electric coupling relationship; essentially, they still treat each cell as an independent entity, failing to capture the impact of thermal runaway on the entire battery module; they only use anomaly detection algorithms to identify lithium-ion batteries that deviate from the normal range, ignoring the distribution differences of different characteristics under healthy conditions, and it is difficult to quantify the joint probability of deviations of various characteristics; when the evaluation result deviates from the normal range by more than a set threshold, they directly issue an alarm and control the solenoid valve to cut off power, ignoring different levels of fire risk, which may lead to over-intervention or under-intervention.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for predicting and controlling lithium-ion battery fires, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting and controlling lithium-ion battery fires, comprising the following steps: Step 1: Treat each cell in the battery module to be predicted as a graph node, establish an initial adjacency matrix based on the Manhattan distance between cells, set a sampling window, and obtain multi-dimensional feature data of the surface of each cell at a fixed sampling frequency within the sampling window to construct a multi-dimensional feature vector, and update the edge weights between graph nodes accordingly to construct a dynamic adjacency matrix. Step 2: Train a dynamic graph attention convolutional network to concatenate the multidimensional feature data of each cell at the current sampling time and several past sampling times in time, and input the dynamic adjacency matrix into the dynamic graph attention convolutional network to output the predicted thermal runaway probability and thermal release rate of each cell at the next sampling time. Step 3: Obtain the Gaussian distribution parameters of the multidimensional feature data of the battery module in a healthy state. Based on the multidimensional feature data collected by each cell at the current sampling time, calculate the deviation probability of each feature data in the Gaussian distribution in a healthy state, and then calculate the comprehensive anomaly degree. Step 4: Construct a comprehensive fire risk index for each cell based on the comprehensive anomaly degree of each cell at the current sampling time, the predicted value of thermal runaway probability at the next sampling time, and the predicted value of heat release rate. Calculate the module fire risk index based on the dynamic adjacency matrix and the predicted value of thermal runaway probability. Determine the actual risk index based on the comprehensive fire risk index and the module fire risk index, and then select the corresponding control strategy based on the actual risk index.

[0008] Furthermore, the principle for establishing the initial adjacency matrix is ​​as follows: Calculate the Manhattan distance between any two battery cells and set a distance threshold; The edge weights between each graph node are calculated to determine the values ​​of each element in the initial adjacency matrix. For any two graph nodes, if the Manhattan distance between their corresponding cells is equal to 0 or greater than the distance threshold, the edge weight between these two graph nodes is 0. If the Manhattan distance between their corresponding cells is between 0 and the distance threshold, the edge weight between these two graph nodes is set to an exponential function value with the natural constant as the base and half the negative value of the Manhattan distance between them as the exponent. Then, the initial adjacency matrix is ​​constructed based on all edge weights.

[0009] Furthermore, the principle of constructing a dynamic adjacency matrix is ​​as follows: Multidimensional feature data includes the surface temperature, terminal voltage, temperature rise rate, voltage drop rate, and CO gas concentration of each cell; The sampling frequency is set to 10Hz. Several sampling times are divided within the sampling window according to the sampling frequency. At each sampling time, the surface temperature, terminal voltage and CO gas concentration of each cell are collected. The sampling window is a time interval formed by tracing back a preset time from the current time as the end point. The principle for calculating the temperature rise rate is as follows: For any sampling time, the surface temperature of each cell at that sampling time is subtracted from the surface temperature of the previous sampling time, and then the result of the difference is divided by the sampling time interval to obtain the temperature rise rate of each cell at that sampling time. The principle of calculating the voltage drop rate is as follows: For any sampling time, the difference between the terminal voltage of the cell at that sampling time and the previous sampling time is calculated, and then the result of the difference is divided by the sampling time interval to obtain the voltage drop rate of each cell at that sampling time. At each sampling time, the surface temperature, terminal voltage and CO gas concentration of each cell are collected, and the corresponding temperature rise rate and voltage drop rate are calculated. Based on the multidimensional feature data of each cell at each sampling time, several multidimensional feature vectors are constructed, and each multidimensional feature vector corresponds to the multidimensional feature data of a cell at a sampling time. The maximum value of the temperature rise rate of all cells at each sampling time within the sampling window is extracted as the maximum absolute temperature rise. The dynamic weight between graph nodes is calculated based on the temperature rise rate of each cell and the maximum absolute temperature rise. The calculation principle is as follows: For any two graph nodes, calculate the absolute difference of the temperature rise rate of the corresponding cells at any sampling time, and then divide the absolute difference of the temperature rise rate by the sum of the maximum absolute temperature rise and the small coefficient. Multiply the result, the adjustment coefficient and the edge weight of the two nodes in sequence to obtain the dynamic weight between the two nodes at that sampling time. Based on the dynamic weights between nodes in the graph at each sampling time, a dynamic adjacency matrix is ​​constructed for each sampling time, where the small coefficient is 10. -6 The adjustment coefficient ranges from 0.3 to 0.5.

[0010] Furthermore, the principle of training a dynamic graph attention convolutional network is as follows: Seven experimental conditions were designed, specifically: Condition 1: Normal charge and discharge; Condition 2: Normal rest; Condition 3: Overcharge, triggered by charging with a constant current of 1C until thermal runaway; Condition 4: Internal short circuit, triggered by implanting a low-melting-point alloy between the positive and negative electrodes of the battery module; Condition 5: External heating, triggered by attaching a heating film to the cell in the middle of the battery module and heating it at a rate of 2℃ / s; Condition 6: Low temperature plus overcharge, triggered by charging at a constant current of 1C at -10℃ until thermal runaway; Condition 7: Overcharge after aging, triggered by performing 500 charge and discharge cycles first, then charging with a constant current of 1C until thermal runaway. Multi-source feature data of battery modules of the same type as the battery module to be predicted under the seven experimental conditions were obtained from historical data. Specifically, for battery modules of the same type as the battery module to be predicted, the duration from start to end of each experimental condition was determined, and within the duration, several experimental moments were divided at a frequency of 10Hz. At each experimental moment, the surface temperature, terminal voltage, CO concentration, and heat release rate of each cell were acquired, and the temperature rise rate and voltage drop rate of each cell at each experimental moment were calculated. The data of each cell during the duration of each experimental condition were recorded. At the moment of thermal runaway of a cell, for any cell at any experimental moment, if a thermal runaway event occurs within 1 second after the experimental moment, the first thermal runaway event that occurs after the experimental moment is taken as the associated thermal runaway event of the cell at that experimental moment. At the same time, the thermal runaway probability label of the cell at that experimental moment is set to 1, and the heat release rate label of the cell at that experimental moment is set to the heat release rate when the associated thermal runaway event of the cell occurs at that experimental moment. Otherwise, the thermal runaway probability label and heat release rate label of the cell at that experimental moment are both set to 0. Each five consecutive experimental time points are used as a prediction window. For each prediction window, the surface temperature, terminal voltage, CO concentration, temperature rise rate, and voltage drop rate of each cell are acquired at each experimental time point within the window and normalized using Z-Score. The dynamic adjacency matrix at each experimental time point is calculated. The surface temperature, terminal voltage, CO concentration, temperature rise rate, voltage drop rate, and dynamic adjacency matrix of each cell at each experimental time point after normalization within the prediction window are used as inputs, and the thermal runaway probability label and thermal release rate label of each cell at the next experimental time point after the prediction window are used as outputs to train a dynamic graph attention convolutional network.

[0011] Furthermore, the principle for obtaining the predicted values ​​of thermal runaway probability and heat release rate is as follows: Starting from the current sampling time, four sampling times are traced back to form a historical prediction window. The dynamic adjacency matrix of each sampling time within the historical prediction window, as well as the normalized surface temperature, terminal voltage, CO concentration, temperature rise rate, and voltage drop rate of each cell, are obtained. The dynamic adjacency matrix of each sampling time within the historical prediction window, as well as the normalized surface temperature, terminal voltage, CO concentration, temperature rise rate, and voltage drop rate of each cell, are used as inputs to the dynamic graph attention convolutional network. The thermal runaway probability and heat release rate of each cell at the next sampling time after the historical prediction window are used as labels. The predicted values ​​of thermal runaway probability and heat release rate of each cell at the next sampling time are output through the dynamic graph attention convolutional network.

[0012] Furthermore, the principle for calculating the overall anomaly degree is as follows: The Gaussian distribution parameters are the mean and variance of each feature data under the healthy state. The healthy state means that at 25°C, the battery module is first powered on and charged to half of its rated capacity with a constant current of 0.5C. During the duration of the healthy state, each feature data is collected at a frequency of 10Hz, and the mean and variance of each feature data of each cell are calculated as the mean and variance of each feature data of each cell under the healthy state. For any given cell, calculate the deviation probability of its multidimensional feature data in the Gaussian distribution of the health state at the current sampling time. The calculation principle is as follows: for each feature data of the cell at the current sampling time, obtain the probability that the feature data collected at the current sampling time is less than or equal to the cumulative distribution function of the Gaussian distribution, and use it as the deviation probability of the various feature data of the cell at the current sampling time. For any given battery cell, the principle for calculating its multimodal risk entropy at the current sampling time based on deviation probabilities is as follows: The deviation probabilities of various feature data of the battery cell at the current sampling time are summed to obtain the total deviation probability. Then, the deviation probabilities of each feature data of the battery cell at the current sampling time are divided by the total deviation probability to obtain the normalized weights of each feature data. For each feature data, the natural logarithm of the normalized weights is calculated. The natural logarithm is multiplied by the normalized weights to obtain the risk components of each feature data of the battery cell, and these components are accumulated. Finally, the accumulated result is multiplied by the negative reciprocal of ln5 to obtain the multimodal risk entropy of the battery cell at the current sampling time. For any given cell, calculate the average deviation probability of various characteristic data of the cell at the current sampling time, and then multiply the average deviation probability by the multimodal risk entropy to obtain the comprehensive anomaly degree of the cell at the current sampling time.

[0013] Furthermore, the principle for constructing a comprehensive fire risk index is as follows: A reference heat release rate is set, and the predicted heat release rate of each cell at the next sampling time is divided by the reference heat release rate to make it dimensionless. The dimensionless heat release rate prediction, thermal runaway probability prediction, and comprehensive anomaly of each cell are weighted and summed to obtain the comprehensive fire risk index of each cell. The weight of the thermal runaway probability prediction is greater than the weight of the dimensionless heat release rate prediction, which is greater than the weight of the comprehensive anomaly, and the sum of the weights of the three is 1. The principle of calculating the module fire risk index is as follows: multiply the weight of each edge in the dynamic adjacency matrix at the current sampling time with the predicted value of the heat release rate of the corresponding cell of the graph node, and then divide the result of the multiplication by the number of graph nodes to obtain the module fire risk index of the battery module at the current sampling time. The maximum value between the fire risk index of the module and the comprehensive fire risk index of all cells at the current sampling time is taken as the actual risk index of the battery module at the current sampling time.

[0014] Furthermore, the principle for selecting the appropriate control strategy based on the actual risk index is as follows: The actual risk index range corresponding to different risk levels is determined based on the expert scoring method, and based on the value of the actual risk index at the current sampling time, it is classified into one of the low, medium and high risk levels. If the battery module is at a low risk level, no fire suppression or power-off operations are performed, and the system maintains its normal sampling frequency. If the battery module is at a medium risk level, select all graph nodes whose comprehensive fire risk index values ​​are within the medium risk range and record them as a risk node set. Then, perform local cooling treatment on the areas where each graph node in the risk node set is located. If the battery module is at a high risk level, disconnect the power input and output circuits of the entire battery module and activate the fire extinguishing device to extinguish the fire in the battery module.

[0015] This invention also provides a lithium-ion battery fire prediction and control system, the system being used to implement the above-mentioned lithium-ion battery fire prediction and control method, specifically including: The feature acquisition and calculation module is used to treat each cell in the battery module to be predicted as a graph node, establish an initial adjacency matrix based on the Manhattan distance between cells, set a sampling window, and acquire multi-dimensional feature data of the surface of each cell at a fixed sampling frequency within the sampling window to construct a multi-dimensional feature vector, thereby updating the edge weights between graph nodes to construct a dynamic adjacency matrix. The data prediction module is used to train the dynamic graph attention convolutional network. It splices the multi-dimensional feature data of each cell at the current sampling time and several past sampling times in time, and combines it with the dynamic adjacency matrix to input into the dynamic graph attention convolutional network. It outputs the predicted value of thermal runaway probability and the predicted value of heat release rate of each cell at the next sampling time. The risk calculation module is used to obtain the Gaussian distribution parameters of the multi-dimensional feature data of the battery module in a healthy state. Based on the multi-dimensional feature data collected by each cell at the current sampling time, it calculates the deviation probability of each feature data in the Gaussian distribution in a healthy state, and then calculates the comprehensive anomaly degree. The integrated control module is used to construct the comprehensive fire risk index of each cell based on the comprehensive anomaly degree of each cell at the current sampling time, the predicted value of thermal runaway probability at the next sampling time, and the predicted value of heat release rate. It calculates the module fire risk index based on the dynamic adjacency matrix and the predicted value of thermal runaway probability, determines the actual risk index based on the comprehensive fire risk index and the module fire risk index, and then selects the corresponding control strategy according to the actual risk index.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention establishes an initial adjacency matrix based on the Manhattan distance between battery cells and dynamically updates edge weights in real time based on characteristic data such as the temperature rise rate of the battery cells. This enables the model to capture the propagation direction and intensity of thermal runaway in real time, improving the prediction accuracy of thermal runaway propagation path and speed, and capturing minute abnormal signs before irreversible thermal runaway occurs in the battery cells. Through a dynamic graph attention convolutional network, the multidimensional feature data of each battery cell at the current and past multiple times are spliced ​​in time and simultaneously input into the dynamic adjacency matrix. This allows the network to synchronously learn the evolution trend of the same battery cell features over time and the thermal runaway propagation mode transmitted between different battery cells through dynamic edge weights. This avoids misjudging random noise in a single battery cell as thermal runaway and also avoids ignoring slowly propagating thermal runaway.

[0017] This invention also calculates the deviation probability of multi-dimensional characteristic data of each cell based on the Gaussian distribution parameters under the health state of the battery module, and then constructs a multimodal risk entropy and comprehensive anomaly degree. Combined with the predicted thermal runaway probability and thermal release rate calculated in the previous steps, it forms a complement in the time dimension. Combining current data with future predicted data improves the accuracy of risk assessment. It integrates the comprehensive anomaly degree at the current moment, the predicted thermal runaway probability and thermal release rate at the next moment to generate a comprehensive fire risk index, reflecting the immediate and future comprehensive risk of a single cell. Based on the dynamic adjacency matrix and the predicted thermal runaway probability of each cell, it calculates the coupling propagation risk of the entire battery module and generates a module fire risk index to reflect the diffusion trend of anomalies among cells. Finally, the maximum value of the two is taken as the actual risk index, which will not be diluted by the overall normal due to severe local anomalies, nor will it ignore individual high-risk cells due to high overall coupling risk, thus improving the comprehensiveness of risk assessment. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the method flow of an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the variation of deviation probability with the surface temperature of the battery cell in an embodiment of the present invention. Figure 3 This is a schematic diagram of the system modules in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0020] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0021] Example: Please see Figures 1 to 2 The present invention provides a technical solution: A method for predicting and controlling lithium-ion battery fires, comprising the following steps: Step 1: Treat each cell in the battery module to be predicted as a graph node, establish an initial adjacency matrix based on the Manhattan distance between cells, set a sampling window, and obtain multi-dimensional feature data of the surface of each cell at a fixed sampling frequency within the sampling window to construct a multi-dimensional feature vector, and update the edge weights between graph nodes accordingly to construct a dynamic adjacency matrix. In this embodiment, the principle for establishing the initial adjacency matrix is ​​as follows: Calculate the Manhattan distance between any two battery cells and set a distance threshold; The edge weights between each graph node are calculated to determine the values ​​of each element in the initial adjacency matrix. For any two graph nodes, if the Manhattan distance between their corresponding cells is equal to 0 or greater than the distance threshold, the edge weight between these two graph nodes is 0. If the Manhattan distance between their corresponding cells is between 0 and the distance threshold, the edge weight between these two graph nodes is set to an exponential function value with the natural constant as the base and half the negative value of the Manhattan distance between them as the exponent. Then, the initial adjacency matrix is ​​constructed based on all edge weights.

[0022] Each cell in the battery module is considered a graph node. The elements in the initial adjacency matrix represent the initial edge weights between any two graph nodes. The specific calculation process is as follows: The formula for calculating the Manhattan distance is: ,in, Indicates the first The and the first Manhattan distance between graph nodes Indicates the index of a graph node. Indicates the first The coordinates of each graph node. Indicates the first The coordinates of the graph nodes; the propagation of thermal runaway between cells mainly depends on thermal conduction and thermal radiation, which are mainly horizontal extensions and diffusions. The thermal propagation path in the vertical direction is very limited, so only the horizontal and vertical coordinates of the graph nodes are considered.

[0023] Calculate the average spacing between each cell, and take twice the average spacing as the distance threshold, set as follows: ; The formula for calculating the edge weight between any two graph nodes is as follows: ;in, Indicates the first The and the first Edge weights between nodes in a graph Represents the natural constant; An initial adjacency matrix is ​​constructed based on the edge weights between any two graph nodes, and the first node in the initial adjacency matrix is... line, number The elements of the column are ; Each cell within the battery module is abstracted as a graph node. ,in , This represents the total number of cells in the battery module, with each graph node representing an independent cell. For any two cells, the Manhattan distance between them is calculated. The Manhattan distance measures the grid distance between two cells on a two-dimensional plane, suitable for scenarios where cells in a battery module are arranged in a matrix, better reflecting the actual heat conduction path. A distance threshold is then set to filter out cell pairs that are too far apart spatially, have weak thermal or electrical coupling, reducing graph density. Edge weights reflect the initial coupling strength between two cells in the event of thermal runaway or abnormal propagation. A smaller Manhattan distance indicates a shorter heat conduction and radiation path between cells, making it easier for heat to spread during thermal runaway. This means thermal runaway is more likely to spread to neighboring cells via heat radiation, conduction, or through the electrolyte and gas, increasing the risk of thermal runaway in the battery module. Under steady-state or quasi-steady-state conditions, for the thermal diffusion of a point heat source on a two-dimensional plane, the temperature increment decays exponentially with the propagation distance. The initial edge weights are essentially the potential thermal influence coefficients between cells; the closer the distance, the stronger the potential heat transfer, and the higher the initial edge weights. At that time, the Manhattan distance between a node and itself corresponds to the edge weight, which describes the probability of thermal runaway propagation between different battery cells. Self-propagation has no physical meaning. When the edge weight is zero; When the distance between two cells exceeds the Manhattan distance threshold, the propagation range of thermal runaway within the battery module is limited. When the cell spacing exceeds 2-3 times the average spacing, the probability of thermal propagation drops sharply. Taking a distance threshold of 2 times the average spacing is a reasonable cutoff distance obtained empirically. Even if there is thermal influence between cells that are too far apart, it will be overwhelmed by temperature fluctuations under normal operating conditions. Therefore, the edge weight between cells exceeding the distance threshold is set to zero to avoid mistakenly treating remote weak coupling as a risk propagation path. When using an exponential decay method, the greater the distance between battery cells, the smaller the impact of thermal runaway propagation.

[0024] The principle of constructing a dynamic adjacency matrix is ​​as follows: Multidimensional feature data includes the surface temperature, terminal voltage, temperature rise rate, voltage drop rate, and CO gas concentration of each cell; The sampling frequency is set to 10Hz. Several sampling times are divided within the sampling window according to the sampling frequency. At each sampling time, the surface temperature, terminal voltage, and CO gas concentration of each cell are collected. The sampling window is a time interval formed by tracing back a preset time from the current time as the end point. The preset time is adjusted according to the size of the battery module, the spacing between cells, the thermal response speed, and the system computing resources. The common value range is 1 to 5 seconds. In this scheme, it is set to 2 seconds, which includes 20 continuous sampling times. This is sufficient to capture the continuous change trend of the temperature, voltage, CO concentration, and other characteristics of the cells in the early stage of thermal runaway.

[0025] Temperature is the most direct characteristic parameter of thermal runaway. Under abnormal operating conditions, lithium-ion batteries will experience temperature rise due to Joule heating or chemical reaction heat. When the temperature exceeds a certain threshold, it may trigger an internal short circuit, which in turn leads to thermal runaway. Real-time temperature monitoring is used to identify abnormal temperature rise trends. T-type thermocouples are installed at the geometric center of each cell surface to monitor the surface temperature of each cell.

[0026] Changes in terminal voltage are used to reflect abnormal chemical reactions inside the battery. When a lithium-ion battery is charging normally, the terminal voltage rises slowly from the full charge voltage. When overcharging occurs, the voltage will continue to exceed the safety limit and show a trend of rising first and then falling before thermal runaway. When an internal short circuit occurs in the battery module, a partial discharge circuit is formed inside the battery, causing an abnormal drop in terminal voltage. The terminal voltage of each cell is collected by the AFE chip.

[0027] The principle for calculating the temperature rise rate is as follows: For any sampling time, the surface temperature of each cell at that sampling time is subtracted from the surface temperature of the previous sampling time, and then the result of the difference is divided by the sampling time interval to obtain the temperature rise rate of each cell at that sampling time. For the first sampling moment in the sampling window, set its temperature rise rate to be equal to the temperature rise rate at the second sampling moment; For each cell, the formula for calculating its temperature rise rate at each sampling time is: ,in, Indicates the first Temperature rise rate at each sampling time, Indicates the index of the sampling time. Indicates the first The cell surface temperature at each sampling time. Indicates the time interval between adjacent sampling times; Before thermal runaway, the temperature change of lithium-ion batteries often exhibits a non-linear accelerating trend. The temperature rise rate is small in the early stages, but increases sharply as thermal runaway approaches. Using differential calculations with short time intervals can capture this trend promptly, avoiding the smoothing out of critical abrupt changes due to excessively long time windows. Under normal charging / discharging or resting conditions, the cell temperature rise rate is usually small. When internal anomalies such as overcharging or internal short circuits occur, the temperature rise rate increases significantly due to the increased heat within the cell. If one cell experiences thermal runaway, the heat generated will spread to other cells through thermal conduction or radiation, causing their temperature rise rates to also increase. When the battery operates under normal high-rate discharge or high-temperature conditions, the temperature may be high, but it remains stable or changes slowly. Before thermal runaway, however, the temperature rises exponentially. Simply looking at the temperature value cannot distinguish between these two situations; however, by reflecting the temperature rise rate, anomalies can be detected as soon as the rapid temperature increase begins.

[0028] The principle of calculating the voltage drop rate is as follows: For any sampling time, the difference between the terminal voltage of the cell at that sampling time and the previous sampling time is calculated, and then the result of the difference is divided by the sampling time interval to obtain the voltage drop rate of each cell at that sampling time. For the first sampling moment in the sampling window, set its voltage drop rate to be equal to the voltage drop rate at the second sampling moment; For each cell, the formula for calculating the voltage drop rate at each sampling time is: ,in, Indicates the first The voltage drop rate at each sampling time. Indicates the first Terminal voltage at each sampling time; The voltage drop rate is the rate of change of voltage over time. Under normal operating conditions, the voltage of a lithium-ion battery changes slowly, but under abnormal conditions such as internal short circuits or overcharging, the voltage may change drastically in a short period of time. A voltage drop rate close to 0 indicates that the voltage is more stable, the cell is in a relatively balanced state, and there is no obvious abnormal charging or discharging process. When the voltage drop rate is positive and large, it indicates that the voltage is rising rapidly and abnormal charging may occur. When the voltage drop rate is negative and large in absolute value, it indicates that the voltage is falling rapidly and internal short circuits, external short circuits, or severe overheating may occur.

[0029] Before thermal runaway, lithium-ion batteries produce CO and other harmful substances due to electrolyte decomposition. , The system collects CO gas concentration data at the location of each battery cell, including gases such as CO, which is an early characteristic gas. CO concentration begins to increase even before the battery temperature rises sharply, providing an early warning capability.

[0030] At each sampling time, the surface temperature, terminal voltage and CO gas concentration of each cell are collected, and the corresponding temperature rise rate and voltage drop rate are calculated. Based on the multidimensional feature data of each cell at each sampling time, several multidimensional feature vectors are constructed, and each multidimensional feature vector corresponds to the multidimensional feature data of a cell at a sampling time. For each cell at each sampling time, a multi-dimensional feature vector is formed, specifically: ,in, Indicates the first At the sampling time, the first For each battery cell's multidimensional feature vector, since there's a one-to-one correspondence between graph nodes and battery cells, the indices of the graph nodes are the same as the indices of the battery cells. Indicates the first At the sampling time, the first The surface temperature of each battery cell Indicates the first At the sampling time, the first The terminal voltage of each battery cell Indicates the first At the sampling time, the first CO gas concentration in each battery cell Indicates the first At the sampling time, the first Temperature rise rate of each battery cell Indicates the first At the sampling time, the first Voltage drop rate of each cell.

[0031] The multidimensional feature vector comprehensively reflects the thermal state, electrochemical state, electrolyte decomposition degree, temperature change rate and voltage change rate of each cell at each sampling time. The typical manifestations of abnormality are: temperature rises close to the thermal runaway threshold, voltage continues to rise or drops sharply, CO concentration continues to rise, temperature rise rate increases sharply and voltage fluctuates violently.

[0032] The maximum value of the temperature rise rate of all cells at each sampling time within the sampling window is extracted as the maximum absolute temperature rise. The dynamic weight between graph nodes is calculated based on the temperature rise rate of each cell and the maximum absolute temperature rise. The calculation principle is as follows: For any two graph nodes, calculate the absolute difference of the temperature rise rate of the corresponding cells at any sampling time, and then divide the absolute difference of the temperature rise rate by the sum of the maximum absolute temperature rise and the small coefficient. Multiply the result, the adjustment coefficient and the edge weight of the two nodes in sequence to obtain the dynamic weight between the two nodes at that sampling time. Based on the dynamic weights between nodes in the graph at each sampling time, a dynamic adjacency matrix is ​​constructed for each sampling time, where the small coefficient is 10. -6 The adjustment coefficient ranges from 0.3 to 0.5.

[0033] The dynamic adjacency matrix describes the dynamic change in the coupling strength of thermal runaway propagation between different cells within a battery module at different sampling times. Compared to the initial adjacency matrix based solely on Manhattan distance, the dynamic adjacency matrix introduces the temperature rise rate difference, reflecting the impact of thermal runaway on the thermal influence intensity between cells during actual propagation. The maximum value of the temperature rise rate of all cells at each sampling time within the sampling window is extracted as the maximum absolute temperature rise. The formula for calculating dynamic weights is: ,in, Indicates the first The nodes at each sampling time are shown in the figure below. and Dynamic weights between them Distribution as graph nodes and The corresponding cell is in the first Temperature rise rate at each sampling time, This represents the adjustment coefficient, with a value ranging from 0.3 to 0.5. Represents a small coefficient, with values ​​ranging from 1 to 10. This is used to prevent the denominator from being zero.

[0034] During thermal runaway, heat is transferred from the abnormal cell to surrounding cells. Heat flux density is proportional to the temperature difference, and the rate of temperature rise reflects the severity of the temperature change within the cell. The greater the difference in the rate of temperature rise between two cells, the more inconsistent their heating or absorbing states are, usually indicating that one of them is undergoing an abnormal thermal process. To quantify the degree of thermal imbalance between the two cells, and by dividing by To perform normalization, so that The range is ; This is a dynamic adjustment term used to adjust the edge weights based on the temperature rise rate difference between the cells. When the temperature rise rates of the two cells are the same, the dynamic adjustment term is 0, and the dynamic weight is equal to the initial edge weight. The same temperature rise rate means that the thermal state changes at the same rate, indicating that even if the temperature itself is high, there is no significant difference in heat flow driving between them. In this case, it is not necessary to amplify the coupling strength through dynamic weights, otherwise it would introduce the risk of false alarms. When the temperature rise rate difference is at its maximum... The result is close to 1, indicating that one cell's temperature rise rate reaches the highest value among all cells, while the other cells have very low temperature rise rates. This creates a huge driving force for heat propagation, causing heat to rapidly transfer from the high-temperature rate cell to the low-temperature rate cell. The risk of thermal runaway propagation is significantly higher than predicted by static spatial distance. (In the context of edge weights...) Enlarged based on The adjustment factor is typically set between 0.3 and 0.5, which amplifies the abnormal propagation signal without causing drastic changes in weights due to noise. In this scheme, the adjustment factor is set to 0.5. At this point, the dynamic weight is appropriately amplified, but not exceeding 1.5 times the edge weight. This highlights the risk of thermal runaway propagation while avoiding numerical instability caused by excessive amplification. In summary, the greater the difference in temperature rise rate, the higher the amplification factor of the dynamic weight relative to the edge weight, in order to enhance the graph connectivity strength of thermal runaway propagation.

[0035] Step 2: Train a dynamic graph attention convolutional network to concatenate the multidimensional feature data of each cell at the current sampling time and several past sampling times in time, and input the dynamic adjacency matrix into the dynamic graph attention convolutional network to output the predicted thermal runaway probability and thermal release rate of each cell at the next sampling time. In this embodiment, the principle of training the dynamic graph attention convolutional network is as follows: Seven experimental conditions were designed, specifically: Condition 1: Normal charge and discharge; Condition 2: Normal rest; Condition 3: Overcharge, triggered by charging with a constant current of 1C until thermal runaway; Condition 4: Internal short circuit, triggered by implanting a low-melting-point alloy between the positive and negative electrodes of the battery module; Condition 5: External heating, triggered by attaching a heating film to the cell in the middle of the battery module and heating it at a rate of 2℃ / s; Condition 6: Low temperature plus overcharge, triggered by charging at a constant current of 1C at -10℃ until thermal runaway; Condition 7: Overcharge after aging, triggered by performing 500 charge and discharge cycles first, then charging with a constant current of 1C until thermal runaway. Seven experimental conditions were designed to construct a comprehensive training dataset that closely reflects real-world usage scenarios and extreme failure modes. This dataset will be used to train a dynamic graph attention convolutional network, enabling it to accurately predict the probability of thermal runaway and the rate of heat release from battery cells under different conditions. In actual use, batteries may face multiple stresses such as overcharging, internal short circuits, external heating, low temperatures, and aging. The triggering mechanisms and propagation characteristics of thermal runaway differ under different conditions. An effective fire prediction model cannot be trained solely based on normal operating condition data; it must include multiple typical fault triggering methods to allow the model to learn the characteristic evolution patterns before, during, and after thermal runaway. Operating condition 1 is used to acquire the feature distribution under healthy conditions, allowing the model to learn features under normal charging and discharging conditions. Operating condition 2 is used to acquire the noise level and feature fluctuation range under static conditions, preventing the model from misjudging normal fluctuations as abnormalities. Operating condition 3 is used to simulate charging management failure scenarios, capturing features such as abnormal voltage rise, accelerated temperature rise, and gas release caused by overcharging. Operating condition 4 is used to simulate local short circuits caused by puncture of the battery's internal separator or metal deposition, characterized by a sudden voltage drop and local temperature rise. Operating condition 5 is used to simulate thermal propagation scenarios, testing the propagation path and speed of thermal runaway in the module and verifying the effectiveness of the dynamic adjacency matrix. Operating condition 6 is used to simulate charging behavior in low-temperature environments, where the risk of lithium deposition is high and overcharging is more likely to trigger thermal runaway, characterized by a sudden change after a slow abnormal voltage rise. Operating condition 7 is used to simulate the decrease in safety after battery aging, where aged batteries have increased internal resistance, poor thermal stability, and a lower thermal runaway trigger threshold. These seven operating conditions cover the complete battery failure process from normal to extreme, from single faults to combined stress, and from new batteries to aged batteries.

[0036] Multi-source feature data of battery modules of the same type as the battery module to be predicted under the seven experimental conditions were obtained from historical data. Specifically, for battery modules of the same type as the battery module to be predicted, the duration from start to end of each experimental condition was determined, and within the duration, several experimental moments were divided at a frequency of 10Hz. At each experimental moment, the surface temperature, terminal voltage, CO concentration, and heat release rate of each cell were acquired, and the temperature rise rate and voltage drop rate of each cell at each experimental moment were calculated. The data of each cell during the duration of each experimental condition were recorded. At the moment of thermal runaway of a cell, for any cell at any experimental moment, if a thermal runaway event occurs within 1 second after the experimental moment, the first thermal runaway event that occurs after the experimental moment is taken as the associated thermal runaway event of the cell at that experimental moment. At the same time, the thermal runaway probability label of the cell at that experimental moment is set to 1, and the heat release rate label of the cell at that experimental moment is set to the heat release rate when the associated thermal runaway event of the cell occurs at that experimental moment. Otherwise, the thermal runaway probability label and heat release rate label of the cell at that experimental moment are both set to 0. For any experimental moment, if a cell in the battery module experiences thermal runaway within 1 second after that experimental moment, the first thermal runaway event occurring within 1 second is taken as the associated thermal runaway event of that cell at that experimental moment; the heat release rate of the cell at the time of the associated thermal runaway event is measured using a battery adiabatic calorimeter.

[0037] Each five consecutive experimental time points are used as a prediction window. For each prediction window, the surface temperature, terminal voltage, CO concentration, temperature rise rate, and voltage drop rate of each cell are acquired at each experimental time point within the window and normalized using Z-Score. The dynamic adjacency matrix at each experimental time point is calculated. The surface temperature, terminal voltage, CO concentration, temperature rise rate, voltage drop rate, and dynamic adjacency matrix of each cell at each experimental time point after normalization within the prediction window are used as inputs, and the thermal runaway probability label and thermal release rate label of each cell at the next experimental time point after the prediction window are used as outputs to train a dynamic graph attention convolutional network.

[0038] A dynamic graph attention convolutional network is used to capture the spatiotemporal propagation dependencies of thermal runaway among individual cells within a battery module. The network's input layer receives two types of information: normalized surface temperature, terminal voltage, CO concentration, temperature rise rate, voltage drop rate, and corresponding dynamic adjacency matrix of each cell at each sampling time within the historical prediction window. A binary cross-entropy loss function is used. The surface temperature, terminal voltage, CO concentration, temperature rise rate, voltage drop rate, and dynamic adjacency matrix of each cell at each experimental time within each prediction window are used as the training set. A certain percentage of the total number of prediction windows are randomly selected from each operating condition. 10% is used as the validation set. The input layer of the graph attention convolutional network takes the surface temperature, terminal voltage, CO concentration, temperature rise rate, voltage drop rate, and dynamic adjacency matrix of each cell at each experimental time point within the prediction window of each graph node as input. The output layer outputs the thermal runaway probability and heat release rate at the next experimental time point. The network is validated using the validation set. The deviation between the network's predicted thermal runaway probability and the actual thermal runaway probability label is measured by the binary cross-entropy loss function. Training stops when the binary cross-entropy loss function stops decreasing or starts to increase. The model with the lowest loss on the validation set is selected as the final judgment model.

[0039] In this embodiment, the principle for obtaining the predicted values ​​of thermal runaway probability and heat release rate is as follows: Starting from the current sampling time, four sampling times are traced back to form a historical prediction window. The dynamic adjacency matrix of each sampling time within the historical prediction window, as well as the normalized surface temperature, terminal voltage, CO concentration, temperature rise rate, and voltage drop rate of each cell, are obtained. The dynamic adjacency matrix of each sampling time within the historical prediction window, as well as the normalized surface temperature, terminal voltage, CO concentration, temperature rise rate, and voltage drop rate of each cell, are used as inputs to the dynamic graph attention convolutional network. The thermal runaway probability and heat release rate of each cell at the next sampling time after the historical prediction window are used as labels. The predicted values ​​of thermal runaway probability and heat release rate of each cell at the next sampling time are output through the dynamic graph attention convolutional network.

[0040] The historical prediction window contains the current sampling time and four sampling times tracing back from the current sampling time. The normalized dynamic adjacency matrix of each sampling time in the historical prediction window, as well as the surface temperature, terminal voltage, CO concentration, temperature rise rate, and voltage drop rate of each cell, are input to the input layer of the dynamic graph attention network. The thermal runaway probability prediction value and heat release rate prediction value of the next sampling time in the historical prediction window are output through the output layer.

[0041] Step 3: Obtain the Gaussian distribution parameters of the multidimensional feature data of the battery module in a healthy state. Based on the multidimensional feature data collected by each cell at the current sampling time, calculate the deviation probability of each feature data in the Gaussian distribution in a healthy state, and then calculate the comprehensive anomaly degree. In this embodiment, the principle for calculating the overall anomaly degree is as follows: The Gaussian distribution parameters are the mean and variance of each feature data under the healthy state. The healthy state means that at 25°C, the battery module is first powered on and charged to half of its rated capacity with a constant current of 0.5C. During the duration of the healthy state, each feature data is collected at a frequency of 10Hz, and the mean and variance of each feature data of each cell are calculated as the mean and variance of each feature data of each cell under the healthy state. Regarding the first A characteristic data point for each battery cell, the sequence of which is collected under healthy conditions, is as follows: ,in, The index representing the sampling time within the duration of the health status. This represents the number of sampling moments within the duration of the health status, and the mean of this feature data is: The variance is: Similarly, for all characteristic data of all cells, the mean and variance are calculated respectively, and expressed as follows: , ,in, An index representing the type of feature data, and , , representing the total number of categories of feature data. Indicates the first The first battery cell The mean of the characteristic data, Indicates the first The first battery cell The variance of the characteristic data.

[0042] For any given cell, calculate the deviation probability of its multidimensional feature data in the Gaussian distribution of the healthy state at the current sampling time. The calculation principle is as follows: for each feature data of the cell at the current sampling time, obtain the probability that the feature data collected at the current sampling time is less than or equal to the cumulative distribution function of the Gaussian distribution, and use it as the deviation probability of the various feature data of the cell at the current sampling time.

[0043] For any given battery cell, the principle for calculating its multimodal risk entropy at the current sampling time based on deviation probabilities is as follows: The deviation probabilities of various feature data of the battery cell at the current sampling time are summed to obtain the total deviation probability. Then, the deviation probabilities of each feature data of the battery cell at the current sampling time are divided by the total deviation probability to obtain the normalized weights of each feature data. For each feature data, the natural logarithm of the normalized weights is calculated. The natural logarithm is multiplied by the normalized weights to obtain the risk components of each feature data of the battery cell, and these components are accumulated. Finally, the accumulated result is multiplied by the negative reciprocal of ln5 to obtain the multimodal risk entropy of the battery cell at the current sampling time. For sampling time is At that time, the first The first cell Type of feature data The formula for calculating its deviation probability is: ,in, Indicates the sampling time as hour, The deviation probability, The cumulative distribution function of a Gaussian distribution is represented by the cumulative distribution function. Let denote the deviation value. For a random variable Z, if it follows a Gaussian distribution, then its cumulative distribution function is: , which represents the probability that the value of the random variable Z is less than or equal to z.

[0044] Table 1 reflects the variation of deviation probability with the measured surface temperature of the battery cell. Under healthy conditions, the mean surface temperature of the battery cell is 25℃, with a standard deviation of 0.5℃. When the measured surface temperature equals the mean surface temperature under healthy conditions, the deviation probability is 0.5, indicating a 50% probability of anomaly. Within the deviation range... The curve is steepest in the middle region and flattens out at both ends, indicating that small fluctuations near the healthy state are mapped to the middle probability region and will not trigger false alarms. When there is a serious deviation, the probability quickly approaches the extreme value, providing a clear abnormal signal. The closer the deviation probability is to 0.5, the closer the feature is to the healthy mean and the lower the probability of abnormality. The closer the deviation probability is to 0 or 1, the more serious the deviation and the higher the probability of abnormality.

[0045] Table 1. Deviation probability as a function of cell surface temperature In a healthy state, the feature data follows a Gaussian distribution. The greater the deviation of the feature data from the mean at each sampling time, the closer its deviation probability in the Gaussian distribution is to 0 or 1, indicating that the feature data is more deviating from the healthy state. The closer the deviation probability is to 0.5, the closer the current feature data is to the healthy mean, and the low probability of anomaly.

[0046] No. Each cell at the sampling time Below, the sum of the deviation probabilities of all features is: The normalized weights for each feature data are: The formula for calculating multimodal risk entropy is: ;in, Indicates the first Each cell at the sampling time Multimodal risk entropy; At the same sampling time, the sum of the deviation probabilities of each characteristic data of the same cell is Dividing each feature data by the sum yields the relative contribution weight of each feature data, which reflects the proportion of contribution of each feature data to the cell anomaly at a certain sampling time. The higher the weight, the more dominant the feature data is in the cell anomaly at the current sampling time.

[0047] Multimodal risk entropy reflects the evenness of the distribution of anomalies among multiple feature data of the same battery cell at a given sampling time. It indicates whether an anomaly occurs simultaneously in multiple feature data points or only a few. When anomalies are dispersed, multiple feature data points interact and exhibit moderate deviations simultaneously; in this case, the multimodal risk entropy is high, close to 1, indicating that the anomaly is occurring synchronously across multiple physical dimensions. When anomalies are concentrated, only a few feature data points deviate while others remain normal; in this case, the multimodal risk entropy is low, indicating that the anomaly originates from single feature noise or localized sporadic disturbances. The maximum entropy value is reached when all feature data have equal weights. Therefore, it needs to be divided by To perform normalization, the deviation probability itself is The values ​​between these values ​​are normalized to eliminate the influence of feature dimensions and absolute values, retaining only the relative magnitude relationship between anomalies and feature data.

[0048] Substitute the feature data acquired at the current sampling time into the above formula to calculate the deviation probability of each feature data at the current sampling time. and multimodal risk entropy , Indicates the current sampling time; The above formula calculates the deviation probability and multimodal risk entropy for any sampling time. By substituting the feature data at the current sampling time into the calculation, the deviation probability and multimodal risk entropy of each feature data at the current sampling time can be obtained.

[0049] For any given cell, calculate the average deviation probability of various characteristic data of the cell at the current sampling time, and then multiply the average deviation probability by the multimodal risk entropy to obtain the comprehensive anomaly degree of the cell at the current sampling time.

[0050] The formula for calculating the overall anomaly degree of each cell at the current sampling time is: ,in, Indicates the first The overall anomaly degree of each cell at the current sampling time; The overall anomaly score reflects the degree to which each cell deviates from its healthy state at the current sampling time. The overall anomaly score combines multimodal risk entropy and deviation probability. Deviation probability measures the severity of the anomaly, while multimodal risk entropy measures the reliability of the anomaly. Combining the two can achieve a balance between sensitivity and false alarm rate. The larger the overall anomaly score, the more severe the overall deviation of the cell from its healthy state in terms of multidimensional features. Moreover, this deviation may be caused by synchronous anomalies in multiple feature data, rather than sporadic noise. The cell is very likely in the early stage of thermal runaway evolution. The smaller the overall anomaly score, the closer the cell is to its healthy state. Even if there is some deviation, it is concentrated in a few features, and the overall risk is low.

[0051] Step 4: Construct a comprehensive fire risk index for each cell based on the comprehensive anomaly degree of each cell at the current sampling time, the predicted value of thermal runaway probability at the next sampling time, and the predicted value of heat release rate. Calculate the module fire risk index based on the dynamic adjacency matrix and the predicted value of thermal runaway probability. Determine the actual risk index based on the comprehensive fire risk index and the module fire risk index, and then select the corresponding control strategy based on the actual risk index.

[0052] In this embodiment, the principle for constructing the comprehensive fire risk index is as follows: A reference heat release rate is set, and the predicted heat release rate of each cell at the next sampling time is divided by the reference heat release rate to make it dimensionless. The dimensionless heat release rate prediction, thermal runaway probability prediction, and comprehensive anomaly of each cell are weighted and summed to obtain the comprehensive fire risk index of each cell. The weight of the thermal runaway probability prediction is greater than the weight of the dimensionless heat release rate prediction, which is greater than the weight of the comprehensive anomaly, and the sum of the weights of the three is 1. The formula for calculating the comprehensive fire risk index of each battery cell at the current sampling time is: in, Indicates the first The comprehensive or risk index of each battery cell at the current sampling time. Indicates the first Predicted thermal runaway probability of each cell at the next sampling time. Indicates the next sampling time. Indicates the first Each cell at the next sampling time Predicted heat release rate Indicates the reference heat release rate. This represents the dimensionless predicted heat release rate. These represent the weighting coefficients for the predicted thermal runaway probability, the dimensionless predicted heat release rate, and the overall anomaly degree, respectively. ,and ; The comprehensive fire risk index reflects the overall fire risk level of each battery cell from its current abnormal state to future thermal runaway evolution at the current sampling time. It considers three different dimensions of risk through three weighted terms. The predicted thermal runaway probability reflects the likelihood of thermal runaway occurring in the near future and is the most direct basis for risk assessment, therefore it has the highest weight. The dimensionless predicted heat release rate reflects the potential intensity of the fire once thermal runaway occurs; even if the probability of thermal runaway is low, a very high heat release rate still poses a high risk, hence its second highest weight. The comprehensive anomaly reflects the overall degree to which the current battery cell deviates from its healthy state in multiple dimensions; it is a current status assessment and does not include future trends, but it can provide early warning signals, therefore its lowest weight. The weights of the three are as follows: , , By weighting and combining these parameters, taking into account both the current degree of anomaly and future risks, early warning can be achieved, and delayed alarms can be avoided by using predictive information.

[0053] The principle of calculating the module fire risk index is as follows: multiply the weight of each edge in the dynamic adjacency matrix at the current sampling time with the predicted value of the heat release rate of the corresponding cell of the graph node, and then divide the result of the multiplication by the number of graph nodes to obtain the module fire risk index of the battery module at the current sampling time. The maximum value between the fire risk index of the module and the comprehensive fire risk index of all cells at the current sampling time is taken as the actual risk index of the battery module at the current sampling time.

[0054] The formula for calculating the fire risk index of the module is: in, This indicates the module fire risk index at the current sampling time. This indicates the graph node at the current sampling time. and Dynamic edge weights between them Indicates the first Predicted thermal runaway probability of each cell at the next sampling time; The module fire risk index reflects the overall risk of future thermal runaway propagating and spreading within the entire battery module at the current moment, driven by the dynamic thermal coupling between cells. Each cell... There is a probability of future thermal runaway. Furthermore, this risk of thermal runaway is not independent, but rather arises through dynamic edge weights. The impact of the risk on the entire battery module is calculated by summing all possible propagation paths to obtain the total impact of all cell risks on the entire battery module, and then divided by the number of cells. This ensures that the fire risk index of the module does not change significantly with the size of the module, facilitating comparisons between battery modules of different sizes. The propagation of thermal runaway in a battery module is essentially a thermally driven process. High-temperature cells heat other cells through heat conduction and radiation, increasing their temperature rise rate and potentially triggering new thermal runaway events.

[0055] The actual risk index is: ; This represents the comprehensive fire risk index of the single cell with the highest risk in the module at the current sampling time, avoiding the masking of a dangerous state where a certain cell is already in the early stages of thermal runaway due to overall averaging. This indicates the potential propagation risk of thermal runaway between battery cells through thermal and electrical coupling. Even without a single extremely high-risk battery cell, if there are strong coupling conditions between multiple battery cells, such as large differences in temperature rise rate, the entire module may still experience cascading thermal runaway. The maximum value between the comprehensive fire risk index and the module fire risk index is taken to ensure that both local severe anomalies and overall propagation risks can be identified in a timely manner and trigger corresponding control strategies.

[0056] The principle underlying the determination of different control strategies based on the actual risk index is as follows: The actual risk index range corresponding to different risk levels is determined based on the expert scoring method, and based on the value of the actual risk index at the current sampling time, it is classified into one of the low, medium and high risk levels. The actual risk index is divided into intervals based on expert scoring, and each interval is assigned a corresponding fire severity level. The actual risk index is divided into three intervals according to the value: low risk, medium risk, and high risk. Low risk indicates no obvious abnormalities and an extremely low probability of fire. Medium risk indicates the presence of local abnormalities or early signs of thermal runaway, but has not yet spread globally. High risk indicates the presence of clear thermal runaway phenomena or an imminent thermal runaway event.

[0057] By analyzing the correspondence between the actual risk index in historical data and the operating conditions of battery modules, the range of actual risk index values ​​for different risks is determined. The actual risk index of several battery modules during operation is collected from historical data, and the risk level of each actual risk index in historical data is determined based on expert evaluation. The risk level of the operating conditions is divided into low risk, medium risk and high risk, and then the range of actual risk index corresponding to these risk levels is determined.

[0058] If the battery module is at a low risk level, no fire suppression or power-off operations are performed, and the system maintains its normal sampling frequency. A low-risk level indicates that there are no obvious anomalies or the risk is extremely low. Therefore, unnecessary intervention should be avoided, and monitoring should be maintained to prevent excessive control from affecting the normal operation of the system or introducing additional power consumption.

[0059] If the battery module is at a medium risk level, select all graph nodes whose comprehensive fire risk index values ​​are within the medium risk range and record them as a risk node set. Then, perform local cooling treatment on the areas where each graph node in the risk node set is located. A medium-risk level indicates that there are local anomalies or early signs of thermal runaway, but they have not yet spread. Local cooling can effectively suppress the rate of temperature rise, delay or prevent the spread of thermal runaway, avoid power outages to the entire module or fire suppression of the entire system, reduce the impact on normal cells, and improve the availability and response accuracy of the system.

[0060] If the battery module is at a high risk level, disconnect the power input and output circuits of the entire battery module and activate the fire extinguishing device to extinguish the fire in the battery module.

[0061] A high-risk level indicates that at least one battery cell is about to or has already experienced thermal runaway, or that thermal runaway has spread rapidly within the module. Cutting off the power circuit can prevent the continuous input of external energy, thus avoiding exacerbating the thermal runaway. Activating the fire extinguishing device can quickly suppress the spread of flames, reduce fire losses, and protect the overall safety of the system.

[0062] Please see Figure 3 The present invention also provides a lithium-ion battery fire prediction and control system, the system being used to implement the above-mentioned lithium-ion battery fire prediction and control method, specifically including: The feature acquisition and calculation module is used to treat each cell in the battery module to be predicted as a graph node, establish an initial adjacency matrix based on the Manhattan distance between cells, set a sampling window, and acquire multi-dimensional feature data of the surface of each cell at a fixed sampling frequency within the sampling window to construct a multi-dimensional feature vector, thereby updating the edge weights between graph nodes to construct a dynamic adjacency matrix. The data prediction module is used to train the dynamic graph attention convolutional network. It splices the multi-dimensional feature data of each cell at the current sampling time and several past sampling times in time, and combines it with the dynamic adjacency matrix to input into the dynamic graph attention convolutional network. It outputs the predicted value of thermal runaway probability and the predicted value of heat release rate of each cell at the next sampling time. The risk calculation module is used to obtain the Gaussian distribution parameters of the multi-dimensional feature data of the battery module in a healthy state. Based on the multi-dimensional feature data collected by each cell at the current sampling time, it calculates the deviation probability of each feature data in the Gaussian distribution in a healthy state, and then calculates the comprehensive anomaly degree. The integrated control module is used to construct the comprehensive fire risk index of each cell based on the comprehensive anomaly degree of each cell at the current sampling time, the predicted value of thermal runaway probability at the next sampling time, and the predicted value of heat release rate. It calculates the module fire risk index based on the dynamic adjacency matrix and the predicted value of thermal runaway probability, determines the actual risk index based on the comprehensive fire risk index and the module fire risk index, and then selects the corresponding control strategy according to the actual risk index.

[0063] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0064] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0065] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0066] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for predicting and controlling lithium-ion battery fires, characterized in that, The specific steps include: Step 1: Treat each cell in the battery module to be predicted as a graph node, establish an initial adjacency matrix based on the Manhattan distance between cells, set a sampling window, and obtain multi-dimensional feature data of the surface of each cell at a fixed sampling frequency within the sampling window to construct a multi-dimensional feature vector, and update the edge weights between graph nodes accordingly to construct a dynamic adjacency matrix. Step 2: Train a dynamic graph attention convolutional network to concatenate the multidimensional feature data of each cell at the current sampling time and several past sampling times in time, and input the dynamic adjacency matrix into the dynamic graph attention convolutional network to output the predicted thermal runaway probability and thermal release rate of each cell at the next sampling time. Step 3: Obtain the Gaussian distribution parameters of the multidimensional feature data of the battery module in a healthy state. Based on the multidimensional feature data collected by each cell at the current sampling time, calculate the deviation probability of each feature data in the Gaussian distribution in a healthy state, and then calculate the comprehensive anomaly degree. Step 4: Construct a comprehensive fire risk index for each cell based on the comprehensive anomaly degree of each cell at the current sampling time, the predicted value of thermal runaway probability at the next sampling time, and the predicted value of heat release rate. Calculate the module fire risk index based on the dynamic adjacency matrix and the predicted value of thermal runaway probability. Determine the actual risk index based on the comprehensive fire risk index and the module fire risk index, and then select the corresponding control strategy based on the actual risk index.

2. The lithium-ion battery fire prediction and control method according to claim 1, characterized in that: The principle behind establishing the initial adjacency matrix in step 1 is as follows: Calculate the Manhattan distance between any two battery cells and set a distance threshold; The edge weights between each graph node are calculated to determine the values ​​of each element in the initial adjacency matrix. For any two graph nodes, if the Manhattan distance between their corresponding cells is equal to 0 or greater than the distance threshold, the edge weight between these two graph nodes is 0. If the Manhattan distance between their corresponding cells is between 0 and the distance threshold, the edge weight between these two graph nodes is set to an exponential function value with the natural constant as the base and half the negative value of the Manhattan distance between them as the exponent. Then, the initial adjacency matrix is ​​constructed based on all edge weights.

3. The lithium-ion battery fire prediction and control method according to claim 2, characterized in that: The principle behind constructing the dynamic adjacency matrix in step 1 is as follows: Multidimensional feature data includes the surface temperature, terminal voltage, temperature rise rate, voltage drop rate, and CO gas concentration of each cell; The sampling frequency is set to 10Hz. Several sampling times are divided within the sampling window according to the sampling frequency. At each sampling time, the surface temperature, terminal voltage and CO gas concentration of each cell are collected. The sampling window is a time interval formed by tracing back a preset time from the current time as the end point. The principle for calculating the temperature rise rate is as follows: For any sampling time, the surface temperature of each cell at that sampling time is subtracted from the surface temperature of the previous sampling time, and then the result of the difference is divided by the sampling time interval to obtain the temperature rise rate of each cell at that sampling time. The principle of calculating the voltage drop rate is as follows: For any sampling time, the difference between the terminal voltage of the cell at that sampling time and the previous sampling time is calculated, and then the result of the difference is divided by the sampling time interval to obtain the voltage drop rate of each cell at that sampling time. At each sampling time, the surface temperature, terminal voltage and CO gas concentration of each cell are collected, and the corresponding temperature rise rate and voltage drop rate are calculated. Based on the multidimensional feature data of each cell at each sampling time, several multidimensional feature vectors are constructed, and each multidimensional feature vector corresponds to the multidimensional feature data of a cell at a sampling time. The maximum value of the temperature rise rate of all cells at each sampling time within the sampling window is extracted as the maximum absolute temperature rise. The dynamic weight between graph nodes is calculated based on the temperature rise rate of each cell and the maximum absolute temperature rise. The calculation principle is as follows: For any two graph nodes, calculate the absolute difference of the temperature rise rate of the corresponding cells at any sampling time, and then divide the absolute difference of the temperature rise rate by the sum of the maximum absolute temperature rise and the small coefficient. Multiply the result, the adjustment coefficient and the edge weight of the two nodes in sequence to obtain the dynamic weight between the two nodes at that sampling time. Based on the dynamic weights between nodes in the graph at each sampling time, a dynamic adjacency matrix is ​​constructed for each sampling time, where the small coefficient is 10. -6 The adjustment coefficient ranges from 0.3 to 0.

5.

4. The lithium-ion battery fire prediction and control method according to claim 3, characterized in that: The principle behind training the dynamic graph attention convolutional network in step 2 is as follows: Seven experimental conditions were designed, specifically: Condition 1: Normal charge and discharge; Condition 2: Normal rest; Condition 3: Overcharge, triggered by charging with a constant current of 1C until thermal runaway; Condition 4: Internal short circuit, triggered by implanting a low-melting-point alloy between the positive and negative electrodes of the battery module; Condition 5: External heating, triggered by attaching a heating film to the cell in the middle of the battery module and heating it at a rate of 2℃ / s; Condition 6: Low temperature plus overcharge, triggered by charging at a constant current of 1C at -10℃ until thermal runaway; Condition 7: Overcharge after aging, triggered by performing 500 charge and discharge cycles first, then charging with a constant current of 1C until thermal runaway. Multi-source feature data of battery modules of the same type as the battery module to be predicted under the seven experimental conditions were obtained from historical data. Specifically, for battery modules of the same type as the battery module to be predicted, the duration from start to end of each experimental condition was determined, and within the duration, several experimental moments were divided at a frequency of 10Hz. At each experimental moment, the surface temperature, terminal voltage, CO concentration, and heat release rate of each cell were acquired, and the temperature rise rate and voltage drop rate of each cell at each experimental moment were calculated. The data of each cell during the duration of each experimental condition were recorded. At the moment of thermal runaway of a cell, for any cell at any experimental moment, if a thermal runaway event occurs within 1 second after the experimental moment, the first thermal runaway event that occurs after the experimental moment is taken as the associated thermal runaway event of the cell at that experimental moment. At the same time, the thermal runaway probability label of the cell at that experimental moment is set to 1, and the heat release rate label of the cell at that experimental moment is set to the heat release rate when the associated thermal runaway event of the cell occurs at that experimental moment. Otherwise, the thermal runaway probability label and heat release rate label of the cell at that experimental moment are both set to 0. Each five consecutive experimental time points are used as a prediction window. For each prediction window, the surface temperature, terminal voltage, CO concentration, temperature rise rate, and voltage drop rate of each cell are acquired at each experimental time point within the window and normalized using Z-Score. The dynamic adjacency matrix at each experimental time point is calculated. The surface temperature, terminal voltage, CO concentration, temperature rise rate, voltage drop rate, and dynamic adjacency matrix of each cell at each experimental time point after normalization within the prediction window are used as inputs, and the thermal runaway probability label and thermal release rate label of each cell at the next experimental time point after the prediction window are used as outputs to train a dynamic graph attention convolutional network.

5. The lithium-ion battery fire prediction and control method according to claim 4, characterized in that: The principle behind obtaining the predicted thermal runaway probability and heat release rate in step 2 is as follows: Starting from the current sampling time, four sampling times are traced back to form a historical prediction window. The dynamic adjacency matrix of each sampling time within the historical prediction window, as well as the normalized surface temperature, terminal voltage, CO concentration, temperature rise rate, and voltage drop rate of each cell, are obtained. The dynamic adjacency matrix of each sampling time within the historical prediction window, as well as the normalized surface temperature, terminal voltage, CO concentration, temperature rise rate, and voltage drop rate of each cell, are used as inputs to the dynamic graph attention convolutional network. The thermal runaway probability and heat release rate of each cell at the next sampling time after the historical prediction window are used as labels. The predicted values ​​of thermal runaway probability and heat release rate of each cell at the next sampling time are output through the dynamic graph attention convolutional network.

6. The lithium-ion battery fire prediction and control method according to claim 5, characterized in that: The principle for calculating the overall anomaly degree in step 3 is as follows: The Gaussian distribution parameters are the mean and variance of each feature data under the healthy state. The healthy state means that at 25°C, the battery module is first powered on and charged to half of its rated capacity with a constant current of 0.5C. During the duration of the healthy state, each feature data is collected at a frequency of 10Hz, and the mean and variance of each feature data of each cell are calculated as the mean and variance of each feature data of each cell under the healthy state. For any given cell, calculate the deviation probability of its multidimensional feature data in the Gaussian distribution of the health state at the current sampling time. The calculation principle is as follows: for each feature data of the cell at the current sampling time, obtain the probability that the feature data collected at the current sampling time is less than or equal to the cumulative distribution function of the Gaussian distribution, and use it as the deviation probability of the various feature data of the cell at the current sampling time. For any given battery cell, the principle for calculating its multimodal risk entropy at the current sampling time based on deviation probabilities is as follows: The deviation probabilities of various feature data of the battery cell at the current sampling time are summed to obtain the total deviation probability. Then, the deviation probabilities of each feature data of the battery cell at the current sampling time are divided by the total deviation probability to obtain the normalized weights of each feature data. For each feature data, the natural logarithm of the normalized weights is calculated. The natural logarithm is multiplied by the normalized weights to obtain the risk components of each feature data of the battery cell, and these components are accumulated. Finally, the accumulated result is multiplied by the negative reciprocal of ln5 to obtain the multimodal risk entropy of the battery cell at the current sampling time. For any given cell, calculate the average deviation probability of various characteristic data of the cell at the current sampling time, and then multiply the average deviation probability by the multimodal risk entropy to obtain the comprehensive anomaly degree of the cell at the current sampling time.

7. The lithium-ion battery fire prediction and control method according to claim 6, characterized in that: The principle behind constructing a comprehensive fire risk index is as follows: A reference heat release rate is set, and the predicted heat release rate of each cell at the next sampling time is divided by the reference heat release rate to make it dimensionless. The dimensionless heat release rate prediction, thermal runaway probability prediction, and comprehensive anomaly of each cell are weighted and summed to obtain the comprehensive fire risk index of each cell. The weight of the thermal runaway probability prediction is greater than the weight of the dimensionless heat release rate prediction, which is greater than the weight of the comprehensive anomaly, and the sum of the weights of the three is 1. The principle of calculating the module fire risk index is as follows: multiply the weight of each edge in the dynamic adjacency matrix at the current sampling time with the predicted value of the heat release rate of the corresponding cell of the graph node, and then divide the result of the multiplication by the number of graph nodes to obtain the module fire risk index of the battery module at the current sampling time. The maximum value between the fire risk index of the module and the comprehensive fire risk index of all cells at the current sampling time is taken as the actual risk index of the battery module at the current sampling time.

8. The lithium-ion battery fire prediction and control method according to claim 7, characterized in that: The principle behind selecting the appropriate control strategy based on the actual risk index in step 4 is as follows: The actual risk index range corresponding to different risk levels is determined based on the expert scoring method, and based on the value of the actual risk index at the current sampling time, it is classified into one of the low, medium and high risk levels. If the battery module is at a low risk level, no fire suppression or power-off operations are performed, and the system maintains its normal sampling frequency. If the battery module is at a medium risk level, select all graph nodes whose comprehensive fire risk index values ​​are within the medium risk range and record them as a risk node set. Then, perform local cooling treatment on the areas where each graph node in the risk node set is located. If the battery module is at a high risk level, disconnect the power input and output circuits of the entire battery module and activate the fire extinguishing device to extinguish the fire in the battery module.

9. A lithium-ion battery fire prediction and control system, characterized in that: The system is used to implement the lithium-ion battery fire prediction and control method according to any one of claims 1-8, specifically including: The feature acquisition and calculation module is used to treat each cell in the battery module to be predicted as a graph node, establish an initial adjacency matrix based on the Manhattan distance between cells, set a sampling window, and acquire multi-dimensional feature data of the surface of each cell at a fixed sampling frequency within the sampling window to construct a multi-dimensional feature vector, thereby updating the edge weights between graph nodes to construct a dynamic adjacency matrix. The data prediction module is used to train the dynamic graph attention convolutional network. It splices the multi-dimensional feature data of each cell at the current sampling time and several past sampling times in time, and combines it with the dynamic adjacency matrix to input into the dynamic graph attention convolutional network. It outputs the predicted value of thermal runaway probability and the predicted value of heat release rate of each cell at the next sampling time. The risk calculation module is used to obtain the Gaussian distribution parameters of the multi-dimensional feature data of the battery module in a healthy state. Based on the multi-dimensional feature data collected by each cell at the current sampling time, it calculates the deviation probability of each feature data in the Gaussian distribution in a healthy state, and then calculates the comprehensive anomaly degree. The integrated control module is used to construct the comprehensive fire risk index of each cell based on the comprehensive anomaly degree of each cell at the current sampling time, the predicted value of thermal runaway probability at the next sampling time, and the predicted value of heat release rate. It calculates the module fire risk index based on the dynamic adjacency matrix and the predicted value of thermal runaway probability, determines the actual risk index based on the comprehensive fire risk index and the module fire risk index, and then selects the corresponding control strategy according to the actual risk index.

Images