A lithium ion battery safety valve opening and failure early warning method based on expansion force

Abstract

The application provides a lithium ion battery safety valve opening and failure early warning method based on expansion force, and belongs to the technical field of lithium ion batteries.The application collects battery expansion force and cycle data under different pre-tightening force conditions, extracts statistical characteristics to construct a state characteristic set, divides health state groups by using a fuzzy clustering algorithm, establishes a segmented nonlinear mapping model of expansion force and internal pressure, carries out wavelet denoising and robust differential calculation on the expansion force signal, optimizes an initial expansion force derivative threshold value by analyzing time dispersion under different heating rates, corrects the threshold value by using a multi-scale feature fusion algorithm based on hierarchical attention aggregation to construct a state adaptive early warning model, and sets a four-level early warning mechanism to monitor the safety valve opening and failure state, thereby solving the technical problem that the safety valve opening time cannot be accurately predicted under different battery health states and pre-tightening force working conditions and adaptive early warning cannot be realized.

Description

Technical Field

[0001] This invention belongs to the field of lithium-ion battery technology, and specifically relates to a method for opening and failure warning of a lithium-ion battery safety valve based on expansion force. Background Technology

[0002] The pressure relief via a safety valve in lithium-ion batteries under thermal abuse conditions is a crucial mechanism to prevent the spread of thermal runaway. Existing technologies, such as using a fixed expansion force threshold or pressure sensors to monitor internal pressure changes to determine the safety valve's opening status, are widely used in the safety monitoring of power batteries and energy storage systems. However, existing methods using a uniform threshold ignore the influence of battery aging and external constraints on the evolution of expansion force, leading to a significant deviation between the warning time and the actual safety valve opening. Furthermore, a fixed threshold cannot adapt to the time differences between abnormal expansion force and safety valve opening under different heating rates, easily resulting in false alarms or delayed alarms in rapid thermal abuse scenarios. In other words, existing technologies suffer from the technical problem of failing to accurately predict the safety valve opening time and achieve adaptive warnings under different battery health states and preload conditions. Summary of the Invention

[0003] In view of this, the present invention provides a method for early warning of opening and failure of lithium-ion battery safety valve based on expansion force, which can solve the technical problem in the prior art that it is impossible to accurately predict the opening time of the safety valve and achieve adaptive early warning under different battery health states and pre-tightening force conditions.

[0004] This invention is implemented as follows: It provides a method for early warning of safety valve opening and failure in lithium-ion batteries based on expansion force. This method collects historical expansion force data, cycle count data, and capacity decay data of batteries under different pre-tightening force conditions. It extracts statistical features of expansion force and combines these with cycle decay factors and capacity retention rates to construct a battery state feature set. A fuzzy C-means clustering algorithm is used to cluster the battery state feature set to obtain healthy state groups. Multiple thermal abuse experiments with different external heating rates are conducted to collect expansion force and internal pressure signals, establishing a piecewise nonlinear mapping model. Wavelet multi-scale decomposition is performed on the expansion force signal to denoise it, and the expansion force derivative is calculated. The dispersion of the time interval from the expansion force derivative reaching a candidate threshold to the safety valve opening under different heating rates is analyzed to select an initial expansion force derivative threshold. The initial expansion force derivative threshold is corrected using a state-adaptive early warning model. When the expansion force derivative exceeds the corrected threshold, a first-level early warning is triggered, and an early warning window timer is started. After the early warning window ends, the safety valve opening window begins. If the expansion force exceeds the expansion force failure threshold after the safety valve opening window ends, a third-level early warning is triggered to determine that the safety valve has failed.

[0005] Specifically, the extraction of expansion force statistical features involves calculating the mean, standard deviation, difference between maximum and minimum values, and skewness coefficient of historical expansion force data within a set time window, and arranging them in sequence as four feature components to form an expansion force feature vector.

[0006] The execution of the fuzzy C-means clustering algorithm specifically involves initializing the number of cluster centers to an integer value between 2 and 8, randomly initializing the membership matrix, calculating the weighted average of the battery sample coordinates based on the membership matrix, updating the membership degree to the negative fuzzy exponential power ratio of the distance between the sample and the cluster center, calculating the objective function as the weighted sum of squares of the distances from the sample to the cluster center, repeating the update until the change in the objective function is less than 0.001, and finally selecting the optimal number of clusters using the silhouette coefficient.

[0007] Specifically, the piecewise nonlinear mapping model is established by dividing the internal pressure range into three pressure segments: 0 to 0.3 MPa, 0.3 to 0.6 MPa, and 0.6 to 1.0 MPa. Within each segment, a cubic polynomial is used to fit the relationship between the expansion force and the internal pressure. Temperature correction factors and state of charge correction factors are introduced as multiplicative correction terms and superimposed on the basic polynomial model. The polynomial coefficients and correction factor weights are iteratively adjusted using a Bayesian optimization algorithm to minimize the root mean square error.

[0008] Specifically, wavelet multi-scale decomposition denoising involves selecting the Daubechies4 wavelet as the mother wavelet function to decompose the original expansion force signal into 5 layers, calculating the energy proportion of detail coefficients in each layer, applying soft thresholding to layers with an energy proportion greater than 5%, and setting the soft threshold to the standard deviation of detail coefficients multiplied by the logarithmic square root of the signal length; and setting the energy proportion of layers with an energy proportion less than 5% to zero. The denoised expansion force signal is then reconstructed using inverse wavelet transform.

[0009] The calculation of the expansion force derivative adopts the sliding window robust differential algorithm. Specifically, the length of the sliding window is set to one percent of the number of sampling points and not less than 5 sampling points. Within each sliding window, the least squares method is used to fit a linear function to the denoised expansion force signal. The slope of the linear function is extracted as the estimated value of the expansion force derivative at the center point of the window. The sliding window is moved and the fitting is repeated to obtain the full-time domain expansion force derivative sequence.

[0010] Specifically, the selection of the initial expansion force derivative threshold involves analyzing the time interval from the expansion force derivative reaching the candidate threshold to the opening of the safety valve under different external heating rates for each battery health state group and pre-tightening force condition analysis. The standard deviation of the time interval is calculated as the dispersion index, and the candidate threshold that minimizes the dispersion index is selected as the initial expansion force derivative threshold.

[0011] Among them, the state adaptive early warning model adopts a multi-scale feature fusion algorithm based on hierarchical attention aggregation. It captures multi-level feature dependencies by constructing parallel attention branches with different receptive fields, and uses learnable scale weights to adaptively weight and fuse the outputs of each parallel attention branch. It also eliminates semantic bias between features of different resolutions through a cross-scale feature alignment mechanism.

[0012] The structure of the state-adaptive early warning model is as follows: the input layer receives a feature vector composed of the initial expansion force derivative threshold, battery health status grouping identifier, pre-tightening force value, and temperature value; the feature extraction layer contains three parallel attention branches that use convolutional kernel sizes of 3, 5, and 7 to obtain local features of different receptive fields; each parallel attention branch has a self-attention module that calculates the correlation weight between each position and other positions in the feature map and performs weighted aggregation of the features; the scale fusion layer spatially aligns the feature maps output by the three parallel attention branches and then concatenates them according to the channel dimension; a learnable scale weight vector is introduced to perform weighted summation of the concatenated feature channels to obtain the fused features; and the fully connected layer maps the fused features to the corrected expansion force derivative threshold output.

[0013] In this module, the query matrix, key matrix, and value matrix of the self-attention module are all generated from the input features through a linear transformation layer. The attention weights are calculated using a scaled dot product attention mechanism with the scaling factor being the reciprocal of the square root of the feature dimension. The learnable scale weight vector is initialized with a uniform distribution and updated through backpropagation during training. The learnable scale weight vector is then normalized by Softmax and used for weighted fusion.

[0014] The training dataset for the state-adaptive early warning model is established by extracting samples from the historical thermal abuse experiment database, which include different battery health state groups, preload conditions, temperature conditions, and external heating rates. Each sample contains the initial expansion force derivative threshold as an input feature and the actual optimal expansion force derivative threshold as a label. The dataset is augmented by superimposing Gaussian noise with an amplitude of 10% of the feature standard deviation onto the original samples to generate new samples. The augmented dataset is then divided into a training set and a validation set in an 8:2 ratio.

[0015] Specifically, the training of the state-adaptive early warning model uses mean squared error as the loss function to measure the squared mean of the difference between the predicted corrected expansion force derivative threshold and the actual optimal expansion force derivative threshold. The regularization term penalizes the sum of squared model weights to prevent overfitting, and the regularization coefficient is set to 0.0001. The Adam optimizer is used for gradient descent optimization with an initial learning rate of 0.001. The batch size is set to 32 samples, and the number of training epochs is set to 200. After each training epoch, the model performance is evaluated on the validation set. When the validation set loss does not decrease for 20 consecutive epochs, an early stopping mechanism is triggered to terminate the training. The model parameters with the minimum validation set loss are saved as the final model.

[0016] The scaling factor of the self-attention module is determined based on the current battery capacity retention rate, temperature value, and preload value. Specifically, the ratio of the capacity retention rate to 80% of the standard capacity retention rate is calculated as the first influencing factor. The absolute value of the difference between the temperature value and the reference temperature of 25℃ is divided by 50℃ and the reciprocal is taken as the second influencing factor. The ratio of the preload value to the standard preload of 0.1MPa is calculated as the third influencing factor. The first, second, and third influencing factors are multiplied together and then multiplied by the reciprocal of the square root of the feature dimension to obtain the adaptive scaling factor.

[0017] The warning window duration is defined as the time from when the expansion force derivative reaches the corrected expansion force derivative threshold to when the expansion force reaches the expansion force value corresponding to the lower limit pressure of the safety valve opening at the maximum external heating rate. The safety valve opening window duration is obtained by adding the heating rate uncertainty time component and the safety valve action uncertainty time component. The heating rate uncertainty time component is the time difference from the start of heating to the opening of the safety valve at the minimum and maximum external heating rates. The safety valve action uncertainty time component is the time it takes for the internal pressure to rise from the lower limit of the safety valve opening pressure to the upper limit of the safety valve opening pressure at the minimum external heating rate.

[0018] Specifically, the determination of the expansion force value corresponding to the lower limit pressure of the safety valve opening, the expansion force failure threshold, and the expansion force rupture threshold is based on the following: the expansion force value corresponding to the lower limit pressure of the safety valve opening is calculated by inputting the lower limit pressure of the safety valve opening with the current temperature and the current state of charge into a piecewise nonlinear mapping model; the expansion force value corresponding to the upper limit pressure of the safety valve opening is calculated by inputting the upper limit pressure of the safety valve opening with the current temperature and the current state of charge into a piecewise nonlinear mapping model; and the expansion force value corresponding to the pressure corresponding to the yield strength of the battery casing material is calculated by inputting the pressure with the current temperature and the current state of charge into a piecewise nonlinear mapping model; and the expansion force value corresponding to the expansion force rupture threshold is calculated by inputting the yield strength of the battery casing material with the current temperature and the current state of charge into a piecewise nonlinear mapping model.

[0019] Furthermore, if the expansion force continues to rise and exceeds the expansion force rupture threshold, a fourth-level warning is triggered to determine that the shell is at risk of rupture. The first-level warning corresponds to the safety valve opening warning window stage, the second-level warning corresponds to the safety valve opening window stage, the third-level warning corresponds to the safety valve failure state, and the fourth-level warning corresponds to the shell rupture risk state.

[0020] This invention uses a fuzzy clustering algorithm to group battery health states and establishes a state feature set by combining the statistical characteristics of historical expansion force data and capacity decay information, thus formulating differentiated early warning strategies for batteries with different aging levels. For each health state group and pre-tightening force condition, the invention analyzes the time dispersion from when the expansion force derivative reaches a candidate threshold to when the safety valve opens in multiple sets of external heating rate experimental data. An initial threshold that minimizes the standard deviation of the time interval is selected. Then, a state-adaptive early warning model is constructed using a multi-scale feature fusion algorithm based on hierarchical attention aggregation to correct the initial threshold. This corrected threshold can capture the expansion force evolution characteristics under different receptive fields and adaptively adjust the weights of each scale. In summary, this invention solves the technical problem mentioned in the background art of being unable to accurately predict the safety valve opening time and achieve adaptive early warning under different battery health states and pre-tightening force conditions. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a schematic diagram of the state-adaptive early warning model.

[0023] Figure 3 This is a graph showing the relationship between internal pressure and expansion force under different preloads.

[0024] Figure 4 This is a graph showing the evolution of the derivative of the expansion force under different heating rates. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0026] like Figure 1 The diagram shown is a flowchart of a lithium-ion battery safety valve opening and failure warning method based on expansion force provided by the present invention. This method includes the following steps:

[0027] S1. Collect historical expansion force data, cycle count data, and capacity decay data of the battery under different pre-tightening force conditions. Extract the expansion force feature vector by statistical feature extraction of the historical expansion force data within a time window. Normalize the cycle count data to the interval of 0 to 1 to obtain the cycle decay factor. Use the ratio of the capacity decay data to the initial capacity data as the capacity retention rate.

[0028] S2. Combine the expansion force feature vector with the cycle decay factor and the capacity retention rate to form a battery state feature set. Use the fuzzy C-means clustering algorithm to perform cluster analysis on the battery state feature set, calculate the membership degree of each battery sample to each cluster center and iteratively update the cluster center position. When the change in the objective function is less than the convergence threshold, terminate the iteration to obtain the battery health state grouping result.

[0029] S3. Conduct multiple sets of thermal abuse experiments with different external heating rates under each pre-tightening force condition, and simultaneously collect the battery expansion force signal and internal pressure signal. Establish a piecewise nonlinear mapping model of expansion force and internal pressure through finite element simulation combined with experimental data, and extend the input parameters of the piecewise nonlinear mapping model to a multidimensional space that includes temperature and state of charge.

[0030] S4. Perform wavelet multi-scale decomposition on the acquired original expansion force signal to decompose the original expansion force signal into wavelet coefficients of different frequency bands. Set an adaptive threshold according to the energy distribution characteristics of the wavelet coefficients at each scale for denoising. Reconstruct the denoised expansion force signal and use the sliding window robust differential algorithm to calculate the expansion force derivative.

[0031] S5. For each battery health state group and pre-tightening force condition, analyze the time interval from the expansion force derivative reaching the candidate threshold to the safety valve opening under different external heating rates, calculate the standard deviation of the time interval as the dispersion index, select the candidate threshold that minimizes the dispersion index as the initial expansion force derivative threshold under the pre-tightening force condition of the battery health state group, and input the initial expansion force derivative threshold into the state adaptive early warning model to obtain the corrected expansion force derivative threshold.

[0032] S6. When the real-time monitored expansion force exceeds the upper limit of the normal operating expansion force, the monitoring process is started. When the expansion force derivative exceeds the corrected expansion force derivative threshold, the first-level warning is triggered and the warning window timer is started. The duration of the warning window is defined as the time from when the expansion force derivative reaches the corrected expansion force derivative threshold to when the expansion force reaches the expansion force value corresponding to the lower limit pressure of the safety valve opening under the maximum external heating rate.

[0033] S7. When the warning window timer ends, the safety valve opening window is entered. The duration of the safety valve opening window is obtained by adding the heating rate uncertainty time component and the safety valve action uncertainty time component. The heating rate uncertainty time component is the time difference from the start of heating to the opening of the safety valve under the minimum external heating rate and the maximum external heating rate. The safety valve action uncertainty time component is the time it takes for the internal pressure to rise from the lower limit of the safety valve opening pressure to the upper limit of the safety valve opening pressure under the minimum external heating rate.

[0034] S8. If the expansion force exceeds the expansion force failure threshold after the safety valve opening window ends and the safety valve does not open, a third-level early warning is triggered to determine that the safety valve has failed.

[0035] S9. If the expansion force continues to rise and exceeds the expansion force rupture threshold, a level 4 warning is triggered to determine that the shell is at risk of rupture.

[0036] The extraction steps of the expansion force feature vector specifically include: calculating the mean of historical expansion force data within a set time window as the first feature component, calculating the standard deviation of the historical expansion force data as the second feature component, calculating the difference between the maximum and minimum values ​​of the historical expansion force data as the third feature component, calculating the skewness coefficient of the historical expansion force data as the fourth feature component, and arranging the first feature component to the fourth feature component in sequence to form the expansion force feature vector.

[0037] The execution steps of the fuzzy C-means clustering algorithm specifically include: initializing the number of cluster centers to an integer value between 2 and 8; randomly initializing the membership matrix of each battery sample to each cluster center, ensuring that the sum of the membership degrees of each battery sample to all cluster centers is 1; calculating the coordinates of each cluster center based on the membership matrix as the weighted average of the coordinates of all battery samples, with the weight being the fuzzy exponent of the membership degree; updating the membership degree of each battery sample to each cluster center as the ratio of the negative fuzzy exponent of the distance between the battery sample and the cluster center to the sum of the negative fuzzy exponents of the distances from the battery sample to all cluster centers; calculating the objective function as the weighted sum of squared distances from all battery samples to their respective cluster centers; repeatedly updating the cluster centers and membership matrix until the change in the objective function is less than 0.001; and using the silhouette coefficient as an effectiveness index to select the number of cluster centers that maximizes the silhouette coefficient as the optimal number of clusters. The fuzzy exponent is set to 2.0.

[0038] The steps for establishing the piecewise nonlinear mapping model specifically include: dividing the internal pressure range into three pressure segments: 0 to 0.3 MPa, 0.3 to 0.6 MPa, and 0.6 to 1.0 MPa; fitting the relationship between the expansion force and the internal pressure using a cubic polynomial in each pressure segment; introducing a temperature correction factor (the difference between the temperature and the reference temperature of 25℃ divided by 100℃ plus 1); introducing a state of charge correction factor (the state of charge value divided by 100% plus 0.5); superimposing the temperature correction factor and the state of charge correction factor as multiplicative correction terms into the basic polynomial model; and iteratively adjusting the polynomial coefficients and correction factor weights using a Bayesian optimization algorithm to minimize the root mean square error between the experimental data and the model predictions.

[0039] The denoising steps of the wavelet multi-scale decomposition specifically include: selecting the Daubechies4 wavelet as the mother wavelet function, decomposing the original expansion force signal into 5 layers to obtain the approximation coefficients and detail coefficients of each layer, calculating the energy proportion of the detail coefficients of each layer as the sum of squares of the detail coefficients divided by the sum of squares of the wavelet coefficients of all layers, applying soft thresholding to layers with an energy proportion greater than 5% and setting the soft threshold to the square root of the logarithm of the length of the original expansion force signal multiplied by the standard deviation of the detail coefficients of the layer, and setting the energy proportion of layers with an energy proportion less than 5% to zero, and reconstructing the denoised expansion force signal by performing inverse wavelet transform on the processed wavelet coefficients of each layer.

[0040] The calculation steps of the sliding window robust differential algorithm specifically include: setting the sliding window length to 1% of the number of sampling points and not less than 5 sampling points; fitting a linear function to the denoised expansion force signal using the least squares method within each sliding window; extracting the slope of the linear function as the estimated value of the expansion force derivative at the center point of the sliding window; moving the sliding window according to the sampling time sequence and repeating the above fitting process to obtain the full-time domain expansion force derivative sequence.

[0041] The state-adaptive early warning model adopts a multi-scale feature fusion algorithm based on hierarchical attention aggregation. It captures multi-level feature dependencies by constructing parallel attention branches with different receptive fields, adaptively weights and fuses the outputs of each parallel attention branch using learnable scale weights, and eliminates semantic bias between features of different resolutions through a cross-scale feature alignment mechanism.

[0042] like Figure 2 As shown, the specific structure of the state-adaptive early warning model is as follows: The input layer receives a feature vector composed of an initial expansion force derivative threshold, a battery health state grouping identifier, a preload force value, and a temperature value. The feature extraction layer contains three parallel attention branches, each using different convolutional kernel sizes of 3, 5, and 7 to obtain local features with different receptive fields. Each parallel attention branch internally sets up a self-attention module to calculate the correlation weight between each position in the feature map and other positions, and performs weighted aggregation of the features. The scale fusion layer spatially aligns the output feature maps of the three parallel attention branches and concatenates them according to the channel dimension. A learnable scale weight vector is introduced to perform weighted summation on the concatenated feature channels to obtain fused features. The fully connected layer maps the fused features to the corrected expansion force derivative threshold output. The query matrix, key matrix, and value matrix of the self-attention module are all generated from the input features through a linear transformation layer. The attention weight calculation adopts a scaled dot product attention mechanism, and the scaling factor is the reciprocal of the square root of the feature dimension. The learnable scale weight vector is initialized with a uniform distribution and updated through backpropagation during training. The learnable scale weight vector is normalized by Softmax and then used for weighted fusion.

[0043] The steps for establishing the training dataset of the state-adaptive early warning model specifically include: extracting samples from the historical thermal abuse experiment database that contain different battery health state groups, pre-tightening conditions, temperature conditions, and external heating rates. Each sample contains an initial expansion force derivative threshold as an input feature and an actual optimal expansion force derivative threshold as a label. The actual optimal expansion force derivative threshold is determined by posterior analysis of the time dispersion from the expansion force derivative reaching different candidate thresholds to the opening of the safety valve under the sample. The dataset is augmented by superimposing Gaussian noise with an amplitude of 10% of the feature standard deviation onto the original samples to generate new samples. The augmented dataset is divided into a training set and a validation set in an 8:2 ratio.

[0044] The specific training steps of the state adaptive early warning model include: using mean squared error as the loss function to measure the squared mean of the difference between the corrected expansion force derivative threshold predicted by the state adaptive early warning model and the actual optimal expansion force derivative threshold; introducing an L2 regularization term to penalize the sum of squares of the weights of the state adaptive early warning model to prevent overfitting, with the regularization coefficient set to 0.0001; using the Adam optimizer for gradient descent optimization with an initial learning rate set to 0.001; setting the batch size to 32 samples; setting the number of training cycles to 200 rounds; evaluating the performance of the state adaptive early warning model on the validation set after each round of training; triggering an early stopping mechanism to terminate training when the validation set loss does not decrease for 20 consecutive rounds; and saving the parameters of the state adaptive early warning model with the minimum validation set loss as the final model.

[0045] The scaling factor of the self-attention module in the state adaptive early warning model is determined based on the current battery capacity retention rate, temperature value, and preload value. The specific steps are as follows: calculate the ratio of the capacity retention rate to the standard capacity retention rate of 80% as the first influencing factor; calculate the absolute value of the difference between the temperature value and the reference temperature of 25℃, divide it by 50℃ and take the reciprocal as the second influencing factor; calculate the ratio of the preload value to the standard preload of 0.1MPa as the third influencing factor; multiply the first influencing factor, the second influencing factor, and the third influencing factor together, and then multiply by the reciprocal of the square root of the feature dimension to obtain the adaptive scaling factor.

[0046] A multi-scale feature fusion algorithm based on hierarchical attention aggregation can simultaneously capture both short-term rapid changes and long-term trends in expansion force signals over time by setting parallel attention branches with different receptive fields, overcoming the limitation of single-scale analysis in not being able to take into account both local mutations and overall evolution. The self-attention module establishes long-range dependencies between feature locations within each parallel attention branch, enabling the state-adaptive early warning model to identify key patterns related to thermal runaway in the expansion force derivative sequence while suppressing irrelevant noise interference. A cross-scale feature alignment mechanism ensures semantic consistency of features extracted by different parallel attention branches through spatial interpolation and channel dimension concatenation, avoiding feature mismatch due to resolution differences. A learnable scale weight vector adaptively adjusts the contribution of each parallel attention branch based on training data, automatically strengthening the most effective feature scale for early warning while weakening redundant information under different operating conditions. The multi-scale feature fusion algorithm based on hierarchical attention aggregation significantly improves the prediction accuracy of the state-adaptive early warning model for the initial expansion force derivative threshold under different battery health state groups, pre-tightening conditions, and temperature conditions. This enables the corrected expansion force derivative threshold to more accurately adapt to the actual thermal runaway evolution characteristics of individual batteries, thereby reducing the uncertainty range of the early warning window duration and lowering the false alarm rate and false alarm rate, providing a more reliable time reference for the timely intervention of subsequent safety measures.

[0047] The upper limit of the normal operating expansion force is defined as the measured value of the expansion force applied to the battery by the clamp when the battery is charged to 100% of its rated capacity at room temperature (25°C).

[0048] The lower limit pressure for opening the safety valve is the lowest internal pressure value allowed to open in the safety valve design specifications, which is 0.6 MPa for lithium iron phosphate batteries and 0.8 MPa for ternary lithium batteries. The upper limit pressure for opening the safety valve is the highest internal pressure value allowed to open in the safety valve design specifications, which is 0.8 MPa for lithium iron phosphate batteries and 1.0 MPa for ternary lithium batteries.

[0049] The steps for determining the expansion force value corresponding to the lower limit pressure of the safety valve opening, the expansion force failure threshold, and the expansion force rupture threshold specifically include: calculating the expansion force value corresponding to the lower limit pressure of the safety valve opening based on the piecewise nonlinear mapping model, the current temperature, and the current state of charge; calculating the expansion force value corresponding to the upper limit pressure of the safety valve opening based on the piecewise nonlinear mapping model, the current temperature, and the current state of charge; and calculating the expansion force value corresponding to the yield strength of the battery casing material based on the piecewise nonlinear mapping model, the current temperature, and the current state of charge. For aluminum alloy casing batteries, the yield strength corresponding to the pressure is taken as 1.0 MPa to 1.2 MPa.

[0050] The first level warning corresponds to the safety valve opening warning window stage, at which point the internal self-heating reaction of the battery has started but the safety valve has not yet reached the opening condition. The second level warning corresponds to the safety valve opening window stage, at which point the internal pressure has approached or reached the safety valve opening pressure range, and the safety valve is at risk of opening at any time. The third level warning corresponds to the safety valve failure state, at which point the safety valve fails to open normally within the design pressure range, indicating a loss of pressure relief function. The fourth level warning corresponds to the casing rupture risk state, at which point the internal pressure has exceeded the strength limit of the battery casing material, posing an immediate risk of rupture.

[0051] The steps for calculating the silhouette coefficient are as follows: for each battery sample, calculate the average distance between the battery sample and other battery samples in the same cluster as the cohesion, calculate the average distance between the battery sample and all battery samples in the nearest other cluster as the separation, and the silhouette coefficient of the battery sample is the separation minus the cohesion and then divided by the larger value between the separation and the cohesion. The average of the silhouette coefficients of all battery samples is the silhouette coefficient of the clustering scheme.

[0052] The initial capacity data is the rated capacity value stated at the time of battery manufacture. The current temperature is the real-time measured surface temperature of the battery. The current state of charge is the percentage of the battery's remaining charge relative to its rated capacity, calculated in real time.

[0053] Optionally, the present invention also provides a method for implementing a lithium-ion battery safety valve opening and failure early warning system based on expansion force through a computer. The computer is provided with a readable storage medium, which stores program instructions. When the program instructions are run in the computer, they execute the above-described method.

[0054] The specific implementation methods of the above steps are described in detail below.

[0055] The specific implementation of step S1 involves first continuously collecting expansion force data of the battery during charge-discharge cycles under different preload conditions using a force sensor as historical expansion force data. Simultaneously, the number of charge-discharge cycles for each battery is recorded as cycle count data. The difference between the current battery capacity and the factory nominal capacity is periodically measured using a capacity testing device as capacity decay data. When extracting statistical features from the historical expansion force data, a time window of 100 hours is set. Within this window, the arithmetic mean, standard deviation, range, and skewness coefficient of the expansion force data are calculated. These four statistical quantities are then arranged in order to form a four-dimensional vector, which is the expansion force feature vector. The cycle count data is divided by the battery's designed life cycle count of 5000 cycles to obtain a normalized value between 0 and 1, which serves as the cycle decay factor. The capacity decay data is divided by the initial capacity data to obtain the capacity retention rate. The initial capacity data is the battery's nominal rated capacity value at the time of manufacture. The purpose of this step is to transform the battery's multi-dimensional historical operating state into a standardized feature representation, providing a numerical basis for subsequent cluster analysis.

[0056] The specific implementation of step S2 involves combining the four components of the expansion force feature vector with the cyclic decay factor and capacity retention rate to form a six-dimensional vector as the battery state feature set. A fuzzy C-means clustering algorithm is used to group the battery state feature set. First, the number of cluster centers is initialized and iterated between 2 and 8. For each cluster center, a membership matrix is ​​randomly generated. The cluster center coordinates and membership matrix are iteratively calculated and updated until the change in the objective function is less than 0.001, at which point the iteration stops. The fuzzy C-means clustering algorithm allows each battery sample to belong to multiple categories with different membership degrees, which better reflects the gradual changes in battery health status compared to hard clustering. By calculating the silhouette coefficient of each clustering scheme, the cluster center number corresponding to the scheme with the largest silhouette coefficient is selected as the optimal number of clusters, resulting in the battery health status grouping result. The fuzzy index is set to 2.0, and the convergence threshold is set to 0.001. The purpose of this step is to perform refined grouping based on the differences in battery health status and to formulate targeted early warning strategies for different groups.

[0057] The specific implementation of step S3 involves designing a thermal abuse test scheme for each preload condition. Five typical operating points are selected within the preload range of 0.02 to 0.2 MPa, and three to five different external heating rates are set for each operating point, with heating rates ranging from 1 to 10 °C / min. During the experiment, the internal pressure signal of the battery is simultaneously measured by an embedded pressure sensor, and the expansion force signal is collected by a force sensor on the fixture, with the sampling frequency set to 100 Hz. A three-dimensional model of the battery is established using finite element simulation software to simulate the shell deformation and stress distribution under different preload and internal pressure conditions. The experimental data and simulation data are combined to establish cubic polynomial relationships describing the mapping between expansion force and internal pressure in three pressure ranges: 0 to 0.3 MPa, 0.3 to 0.6 MPa, and 0.6 to 1.0 MPa. Temperature correction factors and state-of-charge correction factors are introduced as multiplicative correction terms, and the polynomial coefficients are adjusted using a Bayesian optimization algorithm to minimize the root mean square error, with the reference error threshold set at 5%. The purpose of this step is to establish a high-precision mapping model that considers the coupling effects of multiple factors, so as to provide a theoretical basis for subsequent internal pressure estimation based on expansion force.

[0058] The specific implementation of step S4 is as follows: First, the Daubechies4 wavelet function is selected as the mother wavelet, and the original expansion force signal is decomposed into five layers of wavelet decomposition to obtain approximate coefficients and detail coefficients for each layer. The energy proportion of each layer's detail coefficients is calculated. For layers with an energy proportion greater than 5%, soft thresholding is used for denoising. The soft threshold value is taken as the standard deviation of the detail coefficients of that layer multiplied by the square root of the logarithm of the signal length. Layers with an energy proportion less than 5% are directly set to zero to eliminate high-frequency noise. Wavelet transform can separate the frequency components of signal and noise at different scales, avoiding phase delay problems compared to traditional low-pass filtering. The denoised wavelet coefficients are then reconstructed using inverse transform to obtain the denoised expansion force signal. When calculating the expansion force derivative using the sliding window robust differential algorithm, the sliding window length is set to 1% of the total number of sampling points and no less than 5 points. Within each window, a linear function is fitted once using the least squares method, and the slope is extracted as the derivative estimate of the window center point. The complete expansion force derivative sequence is obtained by sliding the window sequentially. The purpose of this step is to extract the true expansion force change trend and its derivative characteristics from the noisy original signal, providing reliable monitoring data for early warning criteria.

[0059] The specific implementation of step S5 involves extracting the expansion force derivative curves and safety valve opening times under different external heating rates from historical experimental data for each battery health state group and preload condition. A series of candidate thresholds are set within the range of 0.05 to 0.10 kPa / s. For each candidate threshold, the time interval from when the expansion force derivative first exceeds the threshold to when the safety valve actually opens is calculated. The time intervals under all heating rate conditions are statistically analyzed, and the standard deviation is calculated as a dispersion index. The candidate threshold that minimizes the dispersion index is selected as the initial expansion force derivative threshold. This threshold provides a relatively stable early warning lead time under different thermal abuse rates. The initial expansion force derivative threshold, along with the battery health state group identifier, preload value, and temperature value, is input into the state-adaptive early warning model. The model outputs the corrected expansion force derivative threshold. The state-adaptive early warning model learns the influence of different conditions on the early warning threshold from historical data to achieve intelligent correction of the threshold. The purpose of this step is to determine a dynamic early warning threshold that can adapt to different thermal runaway response rates, improving the predictability of the early warning window duration.

[0060] The specific implementation of step S6 involves real-time monitoring of the expansion force signal during battery operation. When the expansion force exceeds the upper limit of the normal operating expansion force, i.e., the measured expansion force value when fully charged, the thermal runaway monitoring process is initiated, and the expansion force derivative is calculated. When the expansion force derivative exceeds the corrected expansion force derivative threshold, it is determined that the internal self-heating reaction of the battery has started, and the system immediately triggers the first-level warning and starts the warning window timer. The warning window duration is defined as the time from when the expansion force derivative reaches the threshold until the expansion force reaches the expansion force value corresponding to the lower limit pressure of the safety valve opening under extreme conditions of maximum external heating rate. This time duration is obtained through historical data statistics, with a reference range of 15 to 25 minutes. The purpose of this step is to provide a first-level warning in the early stage of thermal runaway, buying time for intervention measures such as cooling or isolation.

[0061] The specific implementation of step S7 is as follows: when the warning window timer reaches the warning window duration, the system automatically enters the safety valve opening window stage and triggers a second-level warning. The duration of the safety valve opening window is determined by the sum of two uncertainty components. The heating rate uncertainty time component reflects the impact of heating rate differences under different thermal abuse scenarios on the valve opening time. It is obtained by statistically analyzing the time difference from the start of heating to the opening of the safety valve under the minimum and maximum heating rates, with a reference value of 3 to 7 minutes. The safety valve action uncertainty time component reflects the valve opening time uncertainty caused by fluctuations in the safety valve opening pressure within the design range. It is obtained by calculating the time it takes for the internal pressure to rise from the lower limit to the upper limit of the safety valve opening pressure under the minimum heating rate, with a reference value of 12 to 20 minutes. The sum of the two components yields the safety valve opening window duration, with a reference value of 15 to 27 minutes. The purpose of this step is to quantify and predict the safety valve opening time window, transforming the originally uncertain valve opening moment into a predictable risk period.

[0062] The specific implementation of step S8 involves the system continuously monitoring the expansion force signal after the safety valve opening window ends, and calculating the estimated internal pressure corresponding to the current expansion force using a piecewise nonlinear mapping model. If the expansion force exceeds the expansion force failure threshold and the safety valve still shows no signs of opening, the system determines that the safety valve has failed due to blockage, jamming, or mechanical damage, and triggers a third-level warning. The expansion force failure threshold is calculated by substituting the upper limit of the safety valve opening pressure into the piecewise nonlinear mapping model; the reference value is 400 to 450 kPa for lithium iron phosphate batteries and 500 to 550 kPa for ternary lithium batteries. The purpose of this step is to identify the safety valve failure state and prevent more serious accident consequences due to the loss of pressure relief function.

[0063] The specific implementation of step S9 is as follows: if the expansion force continues to rise beyond the expansion force rupture threshold after the safety valve failure warning, it is determined that the internal pressure borne by the battery casing has exceeded the material strength limit, and the system triggers a fourth-level warning. The expansion force rupture threshold is calculated by substituting the yield strength of the battery casing material corresponding to the pressure into a piecewise nonlinear mapping model; the reference value for aluminum alloy casing batteries is 520 to 600 kPa. The purpose of this step is to provide an early warning of the immediate risk of battery casing rupture, indicating the most critical safety state so that extreme measures such as emergency evacuation can be taken.

[0064] It should be noted that one of the key technical ideas of this invention is to establish a battery health state grouping mechanism based on fuzzy C-means clustering. By allowing battery samples to belong to multiple categories with different membership degrees, it more accurately describes the gradual characteristics of battery health state, avoiding the misjudgment problem at state boundaries in hard clustering methods. This allows different health state groups to use optimized early warning models, significantly improving the adaptability of the early warning threshold to individual battery characteristics and reducing false alarms and missed alarms caused by a uniform threshold. The second key technical idea is to adopt a signal processing strategy combining wavelet multi-scale decomposition and a sliding window robust differential algorithm. Wavelet transform separates signal and noise at different scales, avoiding the phase delay of traditional filtering. Sliding window least squares fitting suppresses the interference of quantization error and abrupt noise on derivative calculation in local regions, making the expansion force derivative features extracted from the noisy original signal more stable and reliable, providing high-quality monitoring data for subsequent threshold judgment. The third key technical approach involves constructing a state-adaptive early warning model and introducing a multi-scale feature fusion algorithm based on hierarchical attention aggregation. This model simultaneously captures short-term abrupt changes and long-term trends in expansion force signals through parallel attention branches. A self-attention mechanism identifies key patterns related to thermal runaway and suppresses irrelevant noise. Learnable scale weights adaptively adjust the contribution of features at each scale according to different operating conditions, enabling the corrected early warning threshold to accurately adapt to the thermal runaway evolution characteristics of individual batteries under different temperatures, states of charge, and aging levels, significantly reducing the uncertainty range of the early warning window. The synergistic effect of these technical approaches lies in firstly, achieving refined grouping management of batteries through fuzzy clustering, providing more targeted training samples and application scenarios for the state-adaptive early warning model; secondly, providing high-quality derivative feature input through a wavelet sliding window algorithm; and finally, the multi-scale feature fusion model synthesizing multi-dimensional information to output the optimal early warning threshold. These three elements form a complete early warning chain from data grouping to signal processing to intelligent decision-making. Compared to traditional fixed threshold methods, this approach exhibits stronger robustness and predictive accuracy in complex and variable thermal abuse scenarios, providing a more reliable time benchmark and safety guarantee for early intervention in lithium-ion battery thermal runaway.

[0065] It should be noted that this invention also solves the following technical problem: In existing technologies, the expansion force signal is susceptible to measurement noise and battery mechanical vibration interference, leading to unstable derivative calculations. This invention employs a five-level multi-scale decomposition using the Daubechies4 wavelet. An adaptive soft threshold is set based on the energy proportion of the detail coefficients in each level to suppress high-frequency noise. Levels with an energy proportion less than 5% are directly set to zero to filter out random interference. The reconstructed signal uses a sliding window robust differential algorithm to fit a linear function within a local window to extract the slope, avoiding the sensitivity of single-point differences to noise. This significantly improves the stability and anti-interference capability of the expansion force derivative sequence, providing reliable feature input for subsequent threshold judgment, thus solving the problem of false alarm triggering caused by unstable derivative calculations. Furthermore, in existing technologies, the opening time of the safety valve is affected by the heating rate and valve body manufacturing tolerances, resulting in an uncertain window that is difficult to quantify. This invention defines the duration of the safety valve opening window as the sum of the time components of the heating rate uncertainty and the valve body action uncertainty. The former is determined by the time difference from the start of heating to the opening of the safety valve under the minimum and maximum external heating rates, while the latter is determined by the time it takes for the internal pressure to rise from the lower opening limit to the upper opening limit under the minimum heating rate. By decoupling and quantifying the two types of uncertainty sources, and combining the temperature and state of charge correction factors in the piecewise nonlinear mapping model, the expansion force value corresponding to the opening of the safety valve is dynamically calculated. This allows the warning window to cover the variation range of the actual opening time, reducing the risk of missed alarms caused by manufacturing tolerances and operating condition fluctuations.

[0066] Specifically, the principle of this invention is as follows: The reason why this invention can solve the technical problem is that the derivative of the expansion force reflects the gas generation rate and casing deformation rate inside the battery. By optimizing the dispersion index, a critical point of the derivative threshold with a stable time correlation with the opening time of the safety valve can be found. The state-adaptive early warning model uses parallel attention branches with different convolutional kernel sizes to simultaneously extract short-term mutation features and long-term trend features. The self-attention mechanism establishes long-range dependencies between feature locations to identify key patterns related to thermal runaway. The learnable scale weight vector adaptively adjusts the contribution of each branch according to the training data, enabling the model to output personalized correction thresholds for specific battery health states, preload, and temperature conditions. The piecewise nonlinear mapping model correlates the expansion force with the internal pressure and introduces temperature and state of charge correction factors to ensure the accuracy of the safety valve opening window calculation under different operating conditions, thereby achieving a precise correspondence between the early warning time and the actual safety valve action.

[0067] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0068] The specific implementation of step S1 is as follows: Collect historical expansion force data of the battery under different preload conditions. Loop count data and capacity decay data ,in This is a time series of expansion forces, in units of... , This refers to the number of charge-discharge cycles the battery has completed, measured in cycles. This is the current measured capacity value, in units of... Within the set time window Statistical features were extracted from historical expansion force data. Experience points: 600 Expansion force eigenvector The formula is expressed as follows:

[0069] ;

[0070] In the formula, The first characteristic component, in units of The calculation formula is: ; This is the second characteristic component, with units of The calculation formula is: ; The third characteristic component, in units of The calculation formula is: ; The fourth characteristic component is dimensionless, and its calculation formula is: ; This represents the number of sampling points within the time window, expressed in units of individual samples. For the first Each sampling time, in units of ; The maximum value of the expansion force within the time window, in units of ; The minimum expansion force within the time window, in units of Cyclic decay factor The formula is expressed as follows:

[0071] ;

[0072] In the formula, The battery's design life is measured in cycles, with an empirical value of 3000 cycles. This is a dimensionless parameter. Capacity retention rate. The formula is expressed as follows:

[0073] ;

[0074] In the formula, Initial capacity data, defined as the rated capacity value stated at the time of battery manufacture, in units of... , It is a dimensionless parameter.

[0075] The specific implementation of step S2 is as follows: The expansion force feature vector... Cyclic decay factor and capacity retention Combining features to form a battery state feature set The formula is expressed as follows:

[0076] .

[0077] Fuzzy C-means clustering algorithm is used to perform cluster analysis on the battery state feature set, and the number of cluster centers is initialized. The membership matrix consists of integer values ​​between 2 and 8, dimensionless, and randomly initialized. And the constraints are satisfied. ,in Indicates the first The battery sample pair The membership degree of each cluster center is dimensionless and ranges from 0 to 1. Cluster center coordinates. The calculation formula is expressed as follows:

[0078] ;

[0079] In the formula, This represents the total number of battery samples, expressed in units of individual cells. The fuzzy index is dimensionless and set to 2.0. For the first A set of state features for each battery sample; For the first Each cluster center coordinate vector has a dimension of . The same applies. The membership update formula is expressed as follows:

[0080] ;

[0081] In the formula, For the first The battery sample and the first Euclidean distance between cluster centers For the first Cluster center coordinate vectors. Objective function. The formula is expressed as follows:

[0082] .

[0083] Repeatedly update the cluster center and membership matrix until the objective function changes. The process terminates when the value is less than 0.001. The difference between two adjacent iterations of the objective function is dimensionless. Profile coefficient. The calculation formula is expressed as follows:

[0084] ;

[0085] In the formula, The cohesion is calculated using the following formula: ; The resolution is calculated using the following formula: ; For the first The number of battery samples in each cluster, in units of; For the first The number of battery samples in each cluster, in units of; for and The larger value in; This is a dimensionless parameter, ranging from -1 to 1. The average profile coefficient of all battery samples is... , For dimensionless parameters, choose to make The maximum number of cluster centers is taken as the optimal number of clusters.

[0086] The specific implementation of step S3 is as follows: Under each preload condition, conduct multiple sets of thermal abuse experiments with different external heating rates, and simultaneously collect the battery expansion force signal. With internal pressure signals ,in This is the time series of expansion force measured in the experiment, in units of... , The internal gas pressure of the battery is expressed in MPa. The internal pressure range is divided into three segments: 0–0.3 MPa, 0.3–0.6 MPa, and 0.6–1.0 MPa. Within each pressure segment, a cubic polynomial is used to fit the relationship between the expansion force and the internal pressure. A piecewise nonlinear mapping model is employed. The formula is expressed as follows:

[0087] ;

[0088] In the formula, The temperature correction factor is dimensionless and is calculated using the following formula: ; The state-of-charge correction factor is dimensionless and is calculated using the following formula: ; The coefficient of the zero-degree term, in units of ; These are the polynomial coefficients, all in units of 1. ; This represents the internal pressure of the battery, measured in MPa. The reference pressure is set at 1 MPa. The current temperature is defined as the real-time measured surface temperature of the battery, in °C. The current state of charge (SBC) is defined as the percentage of the battery's remaining charge relative to its rated capacity, calculated in real time, expressed as a percentage (%). The expansion force value obtained from the mapping calculation is in units of The polynomial coefficients and correction factor weights are iteratively adjusted using a Bayesian optimization algorithm to reduce the root mean square error. The minimum value is expressed by the following formula:

[0089] ;

[0090] In the formula, This represents the total number of experimental sampling points, expressed in units of [number]. Units are .

[0091] The specific implementation of step S4 is as follows: Select the Daubechies4 wavelet as the mother wavelet function, and convert the original expansion force signal... Decomposed to 5 layers to obtain the first Approximation coefficients of the layer With detail coefficient ,in This is the raw, unprocessed expansion force measurement signal, in units of... , represents the wavelet decomposition level, taking values ​​from 1 to 5. Energy percentage of layer detail factor The formula is expressed as follows:

[0092] ;

[0093] In the formula, For the first Layer wavelet coefficient length, in units; For the first Layer Individual detailed coefficients, dimensionless; For the 5th floor One approximation coefficient, dimensionless; This is a dimensionless parameter, ranging from 0 to 1. Soft thresholding is used for layers with an energy percentage greater than 5% to denoise. The formula is expressed as follows:

[0094] ;

[0095] In the formula, For the first Standard deviation of layer detail coefficient, dimensionless; The length of the original expansion force signal, in units of; The parameter is dimensionless. Layers with an energy percentage less than 5% are directly set to zero. The processed wavelet coefficients of each layer are then reconstructed using inverse wavelet transform to obtain the denoised expansion force signal. The unit is The robust differential algorithm for sliding windows sets the sliding window length. The number of sampling points is 1% of the total number of sampling points, but not less than 5 sampling points, in units of 1 / 5. Within each sliding window... Fitting a linear function using the least squares method Extracting the slope The estimated value of the derivative of the expansion force at the center point of the sliding window ,in The fitted linear function is in units of . , This is the intercept, in units of , Units are .

[0096] The specific implementation of step S5 is as follows: group each battery health state and pre-tightening force condition, and analyze different external heating rates. The derivative of the expansion force reaches the candidate threshold. Time interval until the safety valve opens ,in For the first External heating rate of the group of experiments, in °C , The threshold value for the derivative of the candidate expansion force is given in units of 1. , For the first The time interval between groups of experiments, in units of Calculate the standard deviation of the time interval. As an index of dispersion, the formula is expressed as follows:

[0097] ;

[0098] In the formula, The number of external heating rate samples, expressed in units of individual samples. This is the average value over time intervals, in units of... The calculation formula is: ; Units are Choose to make The smallest candidate threshold is used as the initial expansion force derivative threshold. The unit is ,Will The input state adaptive early warning model obtains the corrected expansion force derivative threshold. The unit is The input-output relationship of the state-adaptive early warning model is expressed by the following formula:

[0099] ;

[0100] In the formula, This is the mapping function for the state-adaptive early warning model. This function is a multi-scale feature fusion neural network based on hierarchical attention aggregation, and the output is a dimensionless correction coefficient. The battery health status group identifier is a dimensionless integer, ranging from 1 to the optimal cluster number. This is the preload value, in MPa; The standard expansion force derivative threshold, in units of Experience value: 50 ; The standard preload is 0.1 MPa. The standard temperature is 25℃. The calculation method is to use The dimensionless correction factor of the output is multiplied by get.

[0101] The specific implementation of step S6 is: the upper limit of normal expansion force. Defined as the measured value of the expansion force applied to the battery by the clamp when the battery is charged to 100% of its rated capacity at room temperature (25°C), in units of... When the expansion force is monitored in real time Exceed The monitoring process is initiated at the appropriate time. The expansion force value is measured in real time, and the unit is... When the derivative of the expansion force Exceeding the modified expansion force derivative threshold The first-level warning is triggered and the warning window timer starts. The derivative of the real-time expansion force with respect to time is given by units of 1 / 2π. The lower limit pressure for opening the safety valve corresponds to the expansion force value. The calculation formula is expressed as follows:

[0102] ;

[0103] In the formula, The lower limit pressure for the safety valve opening is given in MPa. For lithium iron phosphate batteries, it is 0.6 MPa, and for ternary lithium batteries, it is 0.8 MPa. Units are Warning window duration The formula is expressed as follows:

[0104] ;

[0105] In the formula, The value of the expansion force at the moment the warning is triggered, in units of ; Maximum external heating rate, in °C ; The derivative of the expansion force at the moment the warning is triggered, in units of ; The standard expansion force is set to 1000. ; The standard heating rate is set at 1℃. ; Units are .

[0106] The specific implementation of step S7 is as follows: after the warning window timeout ends, the safety valve opening window is entered, and the expansion force failure threshold is reached. The calculation formula is expressed as follows:

[0107] ;

[0108] In the formula, This is the upper limit of the safety valve opening pressure, in MPa. For lithium iron phosphate batteries, it is 0.8 MPa, and for ternary lithium batteries, it is 1.0 MPa. Units are Duration of the safety valve opening window The formula is expressed as follows:

[0109] ;

[0110] In the formula, The time component of the heating rate uncertainty is given, in units of 1. The calculation formula is: ; The time component of the safety valve's actuation uncertainty, in units of... The calculation formula is: ; Minimum external heating rate, in °C ; The derivative of the expansion force within the opening window of the safety valve, in units of ; Units are .

[0111] The specific implementation method of step S8 is the same as described above, and will not be repeated in detail here.

[0112] The specific implementation of step S9 is: expansion force rupture threshold. The calculation formula is expressed as follows:

[0113] ;

[0114] In the formula, The pressure corresponding to the yield strength of the battery casing material is expressed in MPa. For aluminum alloy casing batteries, the pressure ranges from 1.0 MPa to 1.2 MPa. Units are If the expansion force continues to rise beyond... If this is detected, a Level 4 warning will be triggered, indicating a risk of shell rupture.

[0115] Adaptive scaling factor of self-attention module in state-adaptive early warning model The formula is expressed as follows:

[0116] ;

[0117] In the formula, The feature dimension is dimensionless; This is a dimensionless parameter. In this formula... The impact of feature capacity retention rate on feature weights The term characterizes the effect of temperature deviation on feature weights. The term represents the effect of preload on feature weights. The three terms are multiplied together and then divided by the square root of the feature dimension to obtain the adaptive scaling factor.

[0118] To better understand and implement this invention, a specific application scenario, Example 2, is provided below: A technical team uses the method described in this invention to develop and apply a safety valve opening and failure early warning system for a batch of lithium iron phosphate square batteries. The nominal capacity of the lithium iron phosphate square batteries is 50... The rated voltage is 3.2. The safety valve is designed to open at a pressure range of 0.6 to 0.8 MPa, and the battery casing is made of aluminum alloy.

[0119] The technical team first collected historical operational data from 100 batteries at different cycle stages. For each battery, under five typical operating conditions within a preload range of 0.02 to 0.2 MPa, expansion force data were continuously collected during charge-discharge cycles, with a sampling frequency set at 10. The continuous monitoring period is 100 hours as the time window. The cumulative cycle count of each battery is recorded simultaneously; the designed lifespan of the battery is 5000 cycles. The current battery capacity is periodically measured using capacity testing equipment to obtain capacity decay data. Statistical analysis is performed on historical expansion force data within the 100-hour window to calculate four characteristic components: mean, standard deviation, range, and skewness coefficient, forming an expansion force characteristic vector. The cycle decay factor is obtained by dividing the cycle count by 5000 cycles. The current capacity is then compared with the nominal capacity by 50... The ratio of these factors is used as the capacity retention rate. The four components of the expansion force eigenvector are combined with the cycle decay factor and the capacity retention rate to form a six-dimensional battery state feature set.

[0120] Fuzzy C-means clustering was used to analyze the state feature set of 100 batteries. The technical team set the number of cluster centers to be between 2 and 8, the fuzzy index to be 2.0, and the convergence threshold to be 0.001. The membership matrix and cluster center coordinates were updated iteratively, and the iteration terminated when the change in the objective function was less than the convergence threshold. The silhouette coefficient of each clustering scheme was calculated, and the optimal number of cluster centers was finally determined to be 4. The 100 batteries were divided into four battery health status groups: good health, slight degradation, moderate degradation, and severe degradation.

[0121] For each health status group, the technical team designed thermal abuse tests at five operating points: preload force 0.02, 0.05, 0.1, 0.15, and 0.2 MPa. For each operating point, five sets of experiments were set up with external heating rates of 1, 3, 5, 7, and 10 °C / min, respectively. An electric heating plate was used to uniformly heat the battery, with an initial temperature of 25 °C. During the experiments, the internal pressure of the battery was measured using an embedded pressure sensor, and the expansion force signal was collected using a force sensor on the fixture. Both sensors had a sampling frequency of 100 kHz. A three-dimensional battery model was established using finite element simulation software to simulate shell deformation and stress distribution under different preload and internal pressure conditions. Based on experimental data, cubic polynomial relationships were established for three pressure ranges: 0–0.3 MPa, 0.3–0.6 MPa, and 0.6–1.0 MPa. Temperature correction factors and state-of-charge correction factors were introduced, and the model parameters were adjusted using a Bayesian optimization algorithm. This reduced the root mean square error between the experimental data and the model predictions to 4.2%, establishing a piecewise nonlinear mapping model.

[0122] When processing the acquired raw expansion force signal, the technical team selected the Daubechies4 wavelet function for a 5-level decomposition to obtain the approximation coefficients and detail coefficients for each level. The energy percentage of each level's detail coefficients was calculated. For the 3rd and 4th levels, where the energy percentage was greater than 5%, soft thresholding was applied for denoising, with soft thresholds set to 0.035 kPa and 0.028 kPa, respectively. The 1st, 2nd, and 5th levels, where the energy percentage was less than 5%, were directly set to zero. The processed wavelet coefficients were then reconstructed using an inverse transform to obtain the denoised expansion force signal. A sliding window robust differential algorithm was used to calculate the expansion force derivative. The sliding window length was set to 120 sampling points. Within each window, a linear function was fitted using the least squares method, and the slope was extracted as the estimated value of the expansion force derivative at the window's center point.

[0123] For the group in good health condition under a preload of 0.05 MPa, the technical team extracted the expansion force derivative curves and safety valve opening times at different heating rates from historical experimental data. Candidate thresholds were set within the range of 0.055 to 0.075 kPa / s, iterating in steps of 0.002 kPa / s. For each candidate threshold, the time interval from when the expansion force derivative first exceeded the threshold to when the safety valve opened was calculated, and the standard deviation of the time intervals under five heating rates was statistically analyzed. When the candidate threshold was 0.063 kPa / s, the minimum standard deviation of the time interval was 2.8 minutes, which was determined as the initial expansion force derivative threshold for this group under this condition. The initial expansion force derivative threshold of 0.063 kPa / s, along with the battery health condition group identifier 1, the preload value of 0.05 MPa, and the temperature value of 25°C, were input into the adaptive warning model. The model output a corrected expansion force derivative threshold of 0.061 kPa / s.

[0124] The state-adaptive early warning model developed by the technical team employs a multi-scale feature fusion algorithm based on hierarchical attention aggregation. The model's input layer receives a four-dimensional feature vector. The feature extraction layer contains three parallel attention branches with kernel sizes of 3, 5, and 7, each with its own self-attention module. An adaptive scaling factor is calculated based on the current battery capacity retention rate of 92%, temperature of 25℃, and preload of 0.05MPa. The first influence factor is 1.15, the second is 1.0, the third is 0.5, and the overall adaptive scaling factor is 0.128. The scale fusion layer spatially aligns and concatenates the outputs of the three branches. The learnable scale weight vectors, after training, have values ​​of 0.35, 0.42, and 0.23, which are used to weight and fuse the concatenated features. The fully connected layer maps the fused features to the corrected expansion force derivative threshold output. The model training uses mean squared error as the loss function, with a regularization coefficient of 0.0001, an initial learning rate of 0.001, a batch size of 32, and the validation set loss converges to 0.0023 after 200 training rounds.

[0125] like Figure 3 As shown, the relationship between internal pressure and expansion force under different preload conditions is calculated using a piecewise nonlinear mapping model. Under a preload of 0.05 MPa, when the internal pressure is the lower limit of the safety valve opening pressure (0.6 MPa), the corresponding expansion force value is 398.5 kPa. When the internal pressure is the upper limit of the safety valve opening pressure (0.8 MPa), the corresponding expansion force failure threshold is 435.6 kPa. When the internal pressure reaches the yield strength of the aluminum alloy shell (1.0 MPa), the corresponding expansion force rupture threshold is 564.9 kPa. The key warning thresholds under different preload conditions are shown in Table 1.

[0126] Table 1 Key warning thresholds under different preloads

[0127]

[0128] In practical applications, the technical team selected a battery from a healthy group and conducted a thermal abuse simulation test under a pre-tightening force of 0.05 MPa, with an external heating rate of 3°C / min. During the heating process, the expansion force signal was monitored in real time. When the expansion force exceeded the normal operating expansion force limit of 42.8 kPa, the monitoring process was initiated to calculate the expansion force derivative. Figure 4 As shown, after 38 minutes of heating, the expansion force derivative reached 0.061 kPa / s, exceeding the corrected expansion force derivative threshold. The system then triggered a first-level warning and started the warning window timing. Based on historical data, at a maximum heating rate of 10℃ / min, the time from the expansion force derivative reaching the threshold to the expansion force reaching 398.5 kPa was 18 minutes; therefore, the warning window duration was determined to be 18 minutes.

[0129] After the warning window timeout period ends, the system enters the safety valve opening window phase, triggering the second-level warning. The time component of the heating rate uncertainty was calculated as 5 minutes by statistically analyzing the time difference from the start of heating to the opening of the safety valve at the minimum heating rate of 1℃ / min and the maximum heating rate of 10℃ / min. The time component of the safety valve actuation uncertainty was calculated as 17 minutes by calculating the time it takes for the internal pressure to rise from 0.6MPa to 0.8MPa at the minimum heating rate of 1℃ / min. The safety valve opening window duration is 22 minutes. In this test, the safety valve opened normally 14 minutes after the start of the second-level warning, corresponding to an expansion force of 412.3 kPa, verifying the effectiveness of the warning window.

[0130] To verify the safety valve failure warning function, the technical team selected a battery whose safety valve had been manually blocked for testing. After the safety valve's opening window closed, the expansion force continued to rise to 438.2 kPa, exceeding the expansion force failure threshold of 435.6 kPa. Since no signs of opening were observed in the safety valve, the system triggered a third-level warning indicating safety valve failure. If heating continued without intervention, the expansion force reached 568.7 kPa after 5 minutes, exceeding the expansion force rupture threshold of 564.9 kPa. The system then triggered a fourth-level warning indicating a risk of casing rupture. At this point, heating was immediately stopped and emergency cooling measures were implemented, preventing a casing rupture accident.

[0131] The technological advancements of this invention compared to traditional fixed-threshold early warning methods are reflected in several aspects. Traditional methods use a uniform expansion force or temperature threshold for all batteries, failing to consider the changes in thermal runaway evolution characteristics caused by differences in battery health states. This easily leads to false alarms for decaying batteries or missed alarms for healthy batteries. This invention uses a fuzzy C-means clustering algorithm to finely group batteries according to their health states, allowing samples to belong to multiple categories with different membership degrees. This more accurately describes the gradual changes in battery state and allows for the development of early warning strategies for different groups, significantly improving the targeting and accuracy of early warnings. Traditional methods directly calculate derivatives from noisy raw signals or use simple low-pass filtering, resulting in drastic fluctuations in derivative characteristics or phase delays, affecting the timeliness and reliability of early warning judgments. This invention uses wavelet multi-scale decomposition to separate signals and noise in different frequency bands, avoiding the phase delay problem of traditional filtering. Combined with a sliding window robust differential algorithm, it suppresses quantization errors and abrupt noise in local regions, making the extracted expansion force derivative features more stable and reliable. The early warning window duration of traditional methods is difficult to predict quantitatively, only providing an empirical approximate range, and cannot provide a precise time reference for subsequent safety measures. This invention analyzes the impact of different heating rates and the uncertainty of safety valve action on valve opening time, decomposes the warning window into quantifiable time components, and dynamically corrects the warning threshold based on current battery capacity retention rate, temperature, preload, and other state parameters through a state-adaptive warning model. This significantly improves the predictability of the warning window and provides a more reliable decision-making basis for timely intervention measures such as cooling and isolation.

[0132] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0133] Table 2. Variable Explanation Table (Part 1)

[0134]

[0135] Table 3. Variable Explanation Table (Part Two)

[0136]

[0137] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention 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 the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for early warning of opening and failure of a lithium-ion battery safety valve based on expansion force, characterized in that, Historical expansion force data, cycle count data, and capacity decay data of batteries under different preload conditions were collected. Statistical features of expansion force were extracted and combined with cycle decay factor and capacity retention rate to construct a battery state feature set. Fuzzy C-means clustering algorithm was used to cluster the battery state feature set to obtain healthy state groups. Multiple sets of thermal abuse experiments with different external heating rates were conducted to collect expansion force and internal pressure signals, establishing a piecewise nonlinear mapping model. Wavelet multi-scale decomposition was performed on the expansion force signal for noise reduction, and the expansion force derivative was calculated. The dispersion of the time interval from the expansion force derivative reaching the candidate threshold to the safety valve opening under different heating rates was analyzed to select the initial expansion force derivative threshold. The initial expansion force derivative threshold is corrected by a state-adaptive early warning model. When the expansion force derivative exceeds the corrected threshold, the first-level early warning is triggered and the early warning window timer is started. After the early warning window ends, the safety valve opening window is entered. If the expansion force exceeds the expansion force failure threshold after the safety valve opening window ends, the third-level early warning is triggered to determine that the safety valve has failed. If the expansion force continues to rise and exceeds the expansion force rupture threshold, the fourth-level early warning is triggered to determine that the shell has a risk of rupture. The first-level early warning corresponds to the safety valve opening early warning window stage, the second-level early warning corresponds to the safety valve opening window stage, the third-level early warning corresponds to the safety valve failure state, and the fourth-level early warning corresponds to the shell rupture risk state.

2. The method according to claim 1, characterized in that, The extraction of statistical features of expansion force involves calculating the mean, standard deviation, difference between maximum and minimum values, and skewness coefficient of historical expansion force data within a set time window, and arranging them in order to form an expansion force feature vector.

3. The method according to claim 2, characterized in that, The execution of the fuzzy C-means clustering algorithm is as follows: the number of cluster centers is initialized to an integer value between 2 and 8; the membership matrix is ​​randomly initialized; the weighted average of the battery sample coordinates is calculated based on the membership matrix; the membership degree is updated to the negative fuzzy exponential power ratio of the distance between the sample and the cluster center; the objective function is calculated as the weighted sum of squares of the distances from the sample to the cluster center; the update is repeated until the change in the objective function is less than 0.001, at which point the algorithm terminates; and the silhouette coefficient is used to select the optimal number of clusters.

4. The method according to claim 3, characterized in that, The piecewise nonlinear mapping model is established by dividing the internal pressure range into three pressure segments: 0 to 0.3 MPa, 0.3 to 0.6 MPa, and 0.6 to 1.0 MPa. Within each segment, a cubic polynomial is used to fit the relationship between the expansion force and the internal pressure. Temperature correction factors and state of charge correction factors are introduced as multiplicative correction terms and superimposed on the basic polynomial model. The polynomial coefficients and correction factor weights are iteratively adjusted using a Bayesian optimization algorithm to minimize the root mean square error.

5. The method according to claim 4, characterized in that, Wavelet multi-scale decomposition denoising specifically involves selecting the Daubechies4 wavelet as the mother wavelet function to decompose the original expansion force signal into 5 layers, calculating the energy proportion of detail coefficients in each layer, applying soft thresholding to layers with an energy proportion greater than 5%, and setting the soft threshold to the standard deviation of detail coefficients multiplied by the logarithmic square root of the signal length; and setting the energy proportion of layers with an energy proportion less than 5% to zero. The denoised expansion force signal is then reconstructed using inverse wavelet transform.

6. The method according to claim 5, characterized in that, The calculation of the expansion force derivative adopts the sliding window robust differential algorithm. Specifically, the length of the sliding window is set to one percent of the number of sampling points and not less than 5 sampling points. Within each sliding window, the least squares method is used to fit a linear function to the denoised expansion force signal. The slope of the linear function is extracted as the estimated value of the expansion force derivative at the center point of the window. The sliding window is moved and the fitting is repeated to obtain the full-time domain expansion force derivative sequence.

7. The method according to claim 6, characterized in that, The selection of the initial expansion force derivative threshold involves analyzing the time interval from the expansion force derivative reaching the candidate threshold to the opening of the safety valve under different external heating rates for each battery health state group and pre-tightening force condition analysis. The standard deviation of the time interval is calculated as the dispersion index, and the candidate threshold that minimizes the dispersion index is selected as the initial expansion force derivative threshold.

8. The method according to claim 7, characterized in that, The state-adaptive early warning model adopts a multi-scale feature fusion algorithm based on hierarchical attention aggregation. It captures multi-level feature dependencies by constructing parallel attention branches with different receptive fields, and uses learnable scale weights to adaptively weight and fuse the outputs of each parallel attention branch. It also eliminates semantic bias between features of different resolutions through a cross-scale feature alignment mechanism.

9. The method according to claim 8, characterized in that, The structure of the state-adaptive early warning model is as follows: The input layer receives a feature vector composed of the initial expansion force derivative threshold, battery health status grouping identifier, preload force value, and temperature value. The feature extraction layer contains three parallel attention branches that use convolutional kernel sizes of 3, 5, and 7 to obtain local features of different receptive fields. Each parallel attention branch has a self-attention module that calculates the correlation weight between each position and other positions in the feature map and performs weighted aggregation of the features. The scale fusion layer spatially aligns the feature maps output by the three parallel attention branches and concatenates them according to the channel dimension. A learnable scale weight vector is introduced to perform weighted summation of the concatenated feature channels to obtain the fused features. The fully connected layer maps the fused features to the corrected expansion force derivative threshold output.

10. The method according to claim 9, characterized in that, The query matrix, key matrix, and value matrix of the self-attention module are all generated from the input features through a linear transformation layer. The attention weights are calculated using a scaled dot product attention mechanism with the scaling factor being the reciprocal of the square root of the feature dimension. The learnable scale weight vector is initialized with a uniform distribution and updated through backpropagation during training. The learnable scale weight vector is then normalized by Softmax and used for weighted fusion.

Images