Lithium battery thermal runaway early warning method based on multi-physical sensing information

Abstract

The invention discloses a lithium battery thermal runaway early warning method based on multi-physical sensing information, and belongs to the field of lithium battery TR early warning. The method comprises the following steps: S1, collecting monitoring data such as voltage, temperature and strain of a lithium battery, and carrying out linear interpolation and replacing continuous abnormal values; s2, multiple physical sensing signal data abrupt change points are selected, a TR tag column is added to a data set, and data behind the abrupt change points are true; s3, constructing a Transform encoder model based on the time sequence characteristics of the lithium battery, and generating a TR risk vector by fusing multi-physical sensing information; s4, dividing the public data set into a training set and a verification set according to a proportion for pre-training and parameter adjustment, and storing a pre-training model architecture and parameters; based on a pre-trained model and collected three types of characteristic data of different TR inducement lithium batteries, a data set is divided into a training set and a verification set in proportion for fine adjustment and parameter adjustment, thermal runaway data different from the training set and the verification set are used, and TR early warning is directly carried out after the data are processed.

Description

Technical Field

[0001] The present invention belongs to the technical field of battery thermal runaway warning technology, and in particular relates to a battery thermal runaway warning method based on multi-physical sensing information. Background Art

[0002] In order to cope with the energy crisis and climate change, my country has proposed energy transformation goals and a dual-carbon strategy, which has promoted the widespread application of energy storage units based on lithium batteries in energy storage equipment, electric vehicles and other fields. However, due to the frequent occurrence of thermal runaway safety issues of lithium batteries, traditional thermal runaway warning methods have lagged behind, and inaccurate warnings have become prominent.

[0003] Current research on lithium battery thermal runaway warning focuses on early warning of battery thermal runaway by monitoring one or more characteristic parameters, thereby avoiding serious safety accidents and economic losses caused by thermal runaway. It mainly relies on physical signals such as electricity, heat, force and gas. Based on different technical routes, warning methods can be divided into three main warning methods: signal feature analysis, model prediction and reasoning, and data-driven learning.

[0004] Thermal runaway warning methods based on single sensor information, which rely on a single physical quantity, suffer from limitations in the monitored physical quantity, insufficient information fusion, poor adaptability, delayed warning times, and weak anti-interference capabilities. Methods based on signal feature analysis suffer from subjectivity in feature extraction, high signal preprocessing requirements, limited ability to handle non-stationary signals, high computational complexity, poor feature interpretability, and limited ability to detect abnormal signals. These shortcomings may affect the performance and reliability of the models in practical applications. Furthermore, methods based on model prediction and inference rely on model accuracy, poor adaptability, high computational complexity, and extensive expert knowledge to construct models of thermal runaway mechanisms. Data-driven learning-based warning methods, however, offer advantages such as automatic feature extraction, high precision, strong adaptability, and the ability to discover hidden patterns, and are currently a hot research topic.

[0005] However, current data-driven learning-based methods focus on only one type of thermal runaway trigger, mostly constructing and testing thermal runaway warning models under conditions of electrical abuse. The warning models proposed by these methods can only complete warning tasks under a single thermal runaway trigger, and cannot effectively identify thermal runaway warning tasks under scenarios such as thermal abuse and mechanical abuse. There is a lack of a type of warning model that is highly generalizable and applicable to different thermal runaway triggering modes. Furthermore, there is a lack of publicly available datasets related to thermal runaway, and the high cost and long test cycles of thermal runaway experiments limit the number of samples obtained through self-constructed experiments. This results in a lack of a training mechanism suitable for small-sample thermal runaway data in existing research, which would enable the model to complete the thermal runaway warning task with the support of small-sample experimental data.

[0006] Based on the above, there is an urgent need for a lithium battery thermal runaway warning method based on multi-physical sensing information.

[0007] After searching, the application publication number is [unclear]. A lithium battery thermal runaway early warning and protection method and system is found. The lithium battery thermal runaway early warning and protection method includes: performing real-time thermal runaway detection on the lithium battery to obtain first, second, and third detection data sets corresponding to the lithium battery; using the first, second, and third detection data sets to obtain first, second, and third prediction parameters corresponding to the first, second, and third detection data sets; using the first, second, and third prediction parameters to obtain comprehensive parameters, and determining whether to issue a thermal runaway early warning for the lithium battery based on the comprehensive parameters. The system includes modules corresponding to the steps of the method.

[0008] The thermal runaway warning method described in the patent above involves the transformation and processing of multi-physics sensor information and the manual setting of weight coefficients. Note that weight coefficients corresponding to different parameters are introduced during the acquisition of the first prediction parameter and the comprehensive parameter. The selection of these values ​​relies on expert knowledge and extensive experimental calculations. Furthermore, multi-physics sensor information under different thermal runaway triggers exhibits different variation characteristics. For the same set of weight coefficients, this method may not be able to account for the thermal runaway warning effect under different triggers, requiring manual testing and modification.

[0009] This paper uses a deep learning approach to optimize the thermal runaway early warning process for lithium batteries. The proposed deep learning model can learn from historical thermal runaway data, automatically extracting and learning complex features and correlations within sensor data, thereby more accurately determining whether a lithium battery is experiencing thermal runaway. It also automatically optimizes and adjusts the weight parameters within the network structure, eliminating the need for expert knowledge and extensive experimental calculations. It effectively performs thermal runaway early warning tasks under different thermal runaway triggers. Summary of the Invention

[0010] The present invention aims to solve the above problems of the prior art. It proposes a lithium battery thermal runaway early warning method based on multi-physical sensing information. The technical solution of the present invention is as follows:

[0011] A lithium battery thermal runaway early warning method based on multi-physical sensing information comprises the following steps:

[0012] S1: Collects monitoring data on voltage, temperature, and strain of lithium batteries from normal operation to thermal runaway (TR). Data collected by sensors with a sampling frequency lower than the set value are linearly interpolated at intervals of one time unit, and continuous abnormal values ​​collected are replaced with a constant value.

[0013] S2: Based on the collected lithium battery data, referring to the definition of lithium battery thermal runaway TR of the national standard and the data itself, selecting the mutation point of multi-physical sensing signal data, adding a TR label column to the data set, and the data after the mutation point is true;

[0014] S3: Based on the time series characteristics of lithium batteries, a Transformer encoder model is constructed, and a TR risk vector is generated by fusing multi-physical sensing information;

[0015] S4: Based on the public data set containing lithium battery multi-physical sensing signal, the training set and the verification set are divided in proportion for pre-training and parameter adjustment, and the pre-trained model architecture and parameters are saved; based on the pre-trained model and the collected three types of feature data of lithium battery thermal runaway under different incentives, the training set and the verification set are divided in proportion for fine-tuning and parameter adjustment, and the fine-tuned model architecture and parameters are saved; based on the fine-tuned early warning model, using thermal runaway data different from the above training set and verification set, the data is processed and directly used for TR early warning.

[0016] Further, the step S1 specifically comprises the following steps:

[0017] S11: Selecting voltage, temperature and strain as three characteristic quantities for early warning.

[0018] S12: Linear interpolation is performed on the data collected by the low sampling frequency device at an interval of one time unit, and a reasonable constant is used to replace the continuous abnormal values collected after the sensor is damaged, so that the subsequent model can use reliable data for early warning.

[0019] Further, the step S2 specifically comprises the following steps:

[0020] S21: Based on the definition of thermal runaway of the national standard and the different factors of the actual battery material, selecting the appropriate physical signal mutation point as the thermal runaway mark;

[0021] S22: According to the signal mutation point, a TR label column is added to the multi-physical sensing information, and the data after the mutation point is true and the data before the mutation point is false, so that the model can learn the change rule of the multi-physical sensing information in the thermal runaway process during the training stage, and verify the thermal runaway early warning ability in the test process.

[0022] Further, the step S3 based on the time series characteristics of lithium batteries constructs a Transformer encoder model, and generates a TR risk vector by fusing multi-physical sensing information; specifically comprising the following steps:

[0023] S31: First, through the projection layer conversion, the multi-physical sensing information is mapped from the original feature space to the required dimension of the model. At this time, the unified dimension transformation is performed on different physical quantity characteristics, so that they can be transformed and processed in a unified space-time and prepared for subsequent coding. Then, the position encoding layer is used to obtain the position encoding, which provides the position information of the time sequence for the model, so that the model can distinguish the characteristics of different time steps and enhance the perception of sequence information. Then, the Transformer encoder is used to realize feature extraction and coding output. The output of the encoder enters the activation function and the Dropout layer to introduce nonlinearity and prevent overfitting. Finally, the output layer is used to map the coded features to the target prediction space, and the feature vector for thermal runaway early warning is output. Through the feature vector, the model's early warning result for TR at each time step can be known.

[0024] S32: Building a Transformer encoder.

[0025] Further, the S32: building a Transformer encoder specifically includes:

[0026] For each input sample X∈R^(3×L), where 3 represents four features-voltage, ambient temperature, strain, and time, and L represents the sequence length, the data has been normalized. The data is linearly mapped to the range of [0, 1] by using the deviation standardization method, and the specific formula is as follows:

[0027]

[0028] In the formula: is the normalized value of the nth feature at the ith time, is the original data of the nth feature at the ith time, and are the minimum and maximum values of the nth feature, respectively. Then, the feature data is position encoded and high-dimensional feature extraction is performed to facilitate subsequent processing in the encoder, and the formula is as follows:

[0029] X e = XW i + W pos

[0030] In the formula: X is the normalized feature data, represents a high-dimensional embedding vector with position information, d model is the size of the high-dimensional vector, represents the parameter matrix of the mapping input, is a learnable position matrix used to provide position information to the input. Then, X e enters the self-attention mechanism module, and the expression is as follows:

[0031]

[0032] In the formula: is a query vector and key, value vectors, d k ,d v The dimension of Q, K, V, The encoder follows the idea of multi-head attention, and uses multiple attention heads to extract different subspace features. The multi-head self-attention mechanism can be represented as:

[0033] MultiHead(Q,K,V)=concat(head1,…,head n )W O

[0034]

[0035] In the formula: the parameter matrix of the mapping is and h is the number of attention heads; finally, the vector representation after the multi-head attention mechanism is obtained:

[0036] X A =MultiHead(X e )

[0037] X A After residual connection and layer normalization, we get:

[0038]

[0039] Input into the feedforward neural network layer and apply residual connection and layer normalization again to get:

[0040]

[0041] In the formula: b1 and b2 are the biases of the network, d ff is the dimension of the feedforward layer, d ff is twice d model ; the above is the vector representation after 1 layer of encoder, and after n layers of such encoder, the final vector The vector is subsequently introduced into nonlinearity by the activation function GELU to enhance the expression ability of the model; Dropout is used to prevent overfitting and improve the generalization ability of the model; Time step average compresses the sequence information into global feature representation; Finally, the output layer maps the features to the dimension of the number of categories to obtain the final classification result.

[0042] Further, the step S4 specifically comprises the following steps:

[0043] S41: In the pre-training phase, the public dataset is randomly masked, that is, the mask flag 0 is used to replace certain parts of the original data, and the model is trained for regression, so that it can process the masked input data, predict the masked parts, and achieve data reconstruction; the collected and processed feature data is divided into a training set and a validation set to train and adjust the parameters of the constructed model; to fix the input interpolation and output TR label, it is only necessary to set the encoder's d_model feature dimension and the number of encoder layers, and set the parameters such as the learning rate and number of iterations during training;

[0044] The training set data is input into the model with constructed parameters, and the back propagation algorithm is used to update the parameters. Specifically, the mean square error loss function MSE is used to calculate the loss between the predicted RUL and the true RUL after each iteration. The calculation formula is as follows:

[0045]

[0046] Among them, y i is the true value of the masked physical signal in the i-th cycle, is the predicted value of the masked physical signal in the i-th cycle; then the modified adaptive moment estimation RAdam optimizer is used, and the cosine warm-up method is used to dynamically adjust the learning rate and iteratively update the parameters in the model; after the loss accuracy obtained after multiple iterations meets the standard, the model framework structure and model parameters are saved;

[0047] S42: In the fine-tuning stage, the collected and processed feature data is divided into a training set and a validation set to train and adjust the parameters of the built model; the input interpolation and output TR label are fixed, the encoder's d_model feature dimension and the number of encoder layers are set, and the parameters including the learning rate and number of iterations during training are set;

[0048] The training set data is input into the model with constructed parameters, and the back propagation algorithm is used to update the parameters. Specifically, the cross entropy loss function CE is used to calculate the loss of the predicted TR label and the true TR label after each iteration. The calculation formula is as follows:

[0049]

[0050] Among them, y i is the true value of the lithium battery TR label at the i-th cycle, is the probability that lithium is predicted by the model at the i-th cycle, which means the probability that the sample belongs to category 1;

[0051] Then, the modified adaptive moment estimation RAdam optimizer is used, and the cosine warm-up method is used to dynamically adjust the learning rate and iteratively update the parameters in the model. After the loss accuracy obtained after multiple iterations meets the standard, the model framework structure and model parameters are saved.

[0052] S43: Based on the trained network model, the lithium battery data that needs TR warning, i.e., the test set data, is processed and input into the model to directly obtain the TR output.

[0053] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the lithium battery TR warning method based on multi-physical sensing information as described in any one of the above is implemented.

[0054] A non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the lithium battery TR early warning method based on multi-physical sensing information as described in any one of the claims.

[0055] The advantages and beneficial effects of the present invention are as follows:

[0056] The present invention proposes a TR warning method based on multi-physical sensor information. This method comprehensively utilizes multi-physical sensor information and adopts a data-driven approach to perform TR warning, and the model has good generalization and can complete the TR warning task under small samples of different TR inducements through the "pre-training-fine-tuning" mechanism. This method is based on the encoder structure of Transformer. By integrating the voltage, temperature and strain signals of the thermal runaway process of the lithium battery, it adopts a data-driven approach to achieve thermal runaway warning, and with the help of the "pre-training-fine-tuning" training mechanism, the model can well realize the thermal runaway warning function under small sample conditions of different thermal runaway inducements to achieve zero false alarms.

[0057] S3 is based on the Transformer encoder

[0058] The innovative Transformer encoder is used to fuse multi-physics sensor information to generate a TR risk vector. Feature extraction and encoding are achieved through modules such as projection layers, position encoding, and a multi-head self-attention mechanism, effectively enhancing the model's ability to perceive multi-physics sensor time series data. It can automatically mine complex features and correlations in the data, reducing reliance on human experience. Furthermore, the model's adaptive learning capabilities enable it to quickly adapt to different types of lithium batteries and diverse operating conditions, further improving the accuracy and reliability of early warnings.

[0059] S4 phased training strategy

[0060] A training strategy combining pre-training and fine-tuning is adopted. Pre-training is first performed on a public dataset to enhance the model's generalization ability, and then the model is fine-tuned and optimized using specific thermal runaway data. This step not only enables the model to adapt to a variety of thermal runaway situations, but also significantly improves the accuracy and universality of early warning. It also solves the problem of small sample sizes in the thermal runaway early warning training process. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a multi-physical sensor signal data diagram of a battery under different thermal runaway inducements in a preferred embodiment of the present invention;

[0062] Figure 2 This is a structural diagram of the encoder according to the present invention;

[0063] Figure 3 This is a diagram of the model structure of the present invention;

[0064] Figure 4 This is the overall flow chart of the present invention. DETAILED DESCRIPTION

[0065] The following will describe the technical solutions in the embodiments of the present invention in detail with reference to the accompanying drawings. The described embodiments are only a part of the embodiments of the present invention.

[0066] The technical solution of the present invention to solve the above technical problems is:

[0067] like Figure 4 As shown, a lithium battery thermal runaway warning method based on multi-physical sensing information specifically includes the following steps:

[0068] S1: Collect monitoring data such as voltage, temperature, and strain of lithium batteries from normal operation to the entire process of thermal runaway. Perform linear interpolation on the data collected by sensors with lower sampling frequency at intervals of one time unit, and replace the continuous abnormal values ​​collected due to sensor damage with a reasonable constant value. The specific steps include:

[0069] S11: In the process of lithium battery status detection, voltage and temperature are indispensable basic monitoring parameters, forming the basic signal framework for thermal runaway monitoring; strain sensors can be well applied to the three thermal runaway experiments due to their small size, and they can accurately reflect the changes in battery expansion pressure and realize non-invasive monitoring of external shape and volume changes. Therefore, voltage, temperature and strain are selected as the three characteristic quantities for early warning.

[0070] S12: Since data acquisition devices may have inconsistent sampling frequencies, linear interpolation is performed on the data collected by devices with lower sampling frequencies at intervals of one time unit. At the same time, thermal runaway has the characteristic of violent reaction, and the sensor is easily damaged during the data detection process. The continuous abnormal values ​​collected after the sensor is damaged are replaced with a reasonable constant value to ensure that the subsequent model can use reliable data for early warning.

[0071] S2: Based on the collected lithium battery data, referring to the national standard definition of TR and the data itself, select the mutation point of the multi-physical sensor signal data and add a TR label column to the data set. The data after the mutation point is true. The specific steps include:

[0072] S21: Based on the national standard's definition of thermal runaway and factors such as the differences in actual battery materials, select appropriate physical signal mutation points as thermal runaway signs.

[0073] S22: Add a TR label column to the multi-physics sensor information based on the signal mutation point. The value after the mutation point is true, and the value before the mutation point is false. This allows the model to learn the changing pattern of the multi-physics sensor information during the thermal runaway process during the training phase, and to verify its thermal runaway warning capability during the testing process.

[0074] S3: Build a Transformer encoder model based on the time series characteristics of lithium batteries and generate a TR risk vector by fusing multi-physical sensor information. The specific steps include:

[0075] S31: To achieve multi-physics information fusion during the lithium battery thermal runaway process, the multi-physics sensor information is first transformed through the projection layer, mapping it from the original feature space to the dimensions required by the model. At this point, the different physical quantity features undergo a unified dimensional transformation, allowing them to be transformed and processed in a unified time and space, preparing for subsequent encoding. Next, the position encoding layer obtains the positional encoding, providing the model with temporal position information, enabling the model to distinguish features at different time steps, enhancing its perception of sequence information and enabling it to better understand the temporal variation of physical quantities. The Transformer encoder then flows through the data, extracting features and encoding them for output. The encoder output enters the activation function and Dropout layer, introducing nonlinearity and preventing overfitting. Finally, the output layer maps the encoded features to the target prediction space, outputting a feature vector for thermal runaway warning. This feature vector provides the model's warning result for TR at each time step.

[0076] S32: Build the Transformer encoder. For each input sample X∈R^(3×L), where 3 represents four features—voltage, ambient temperature, strain, and time—and L represents the sequence length (which can also be considered as L data points after data processing). In order to allow the model to treat each feature more fairly and prevent certain features from dominating the training process due to their large dimensions, the data is normalized before entering the model. The deviation normalization method is used to linearly map the data to the range [0,1]. The specific formula is as follows:

[0077]

[0078] Where: is the normalized value of the nth feature at the i-th moment, is the original data of the nth feature at the i-th moment, and are the minimum and maximum values ​​under the nth feature respectively. The feature data is then position-encoded and high-dimensional features are extracted so that it can be fed into the encoder for subsequent processing. The formula is as follows:

[0079] X e =XW i +W pos

[0080] Where: X is the normalized feature data, Represents a high-dimensional embedding vector with position information, d model is the high-dimensional vector size (also called model size), represents the parameter matrix of the mapping input, is a learnable position matrix that provides input position information. Then, X e Enter the self-attention mechanism module, its formula is:

[0081]

[0082] Where: is the query vector and the key and value vectors, d k ,d v are the dimensions of Q, K, and V, The encoder follows the idea of ​​multi-head attention and uses multiple attention heads to extract features from different subspaces. The multi-head self-attention mechanism can be expressed as:

[0083] MultiHead(Q,K,V)=concat(head1,…,head n )W O

[0084]

[0085] Where: The parameter matrix of the mapping is as well as h is the number of attention heads. The final vector representation obtained after the multi-head attention mechanism is:

[0086] X A =MultiHead(X e )

[0087] X A After residual connection and layer normalization, we get:

[0088]

[0089] Input to the feedforward neural network layer and apply residual connection and layer normalization again to get:

[0090]

[0091] Where: b1 and b2 are the bias of the network, d ff is the dimension of the feedforward layer. In this invention, d ff It is d model The above is the vector representation after 1 layer of encoder. Similarly, after n layers of encoder, the final vector is obtained. The vector is then passed through the activation function GELU to introduce nonlinearity and enhance the expressiveness of the model; Dropout is used to prevent overfitting and improve the generalization ability of the model; time-step averaging compresses the sequence information into a global feature representation; finally, the output layer maps the features to the dimension of the number of categories to obtain the final classification result.

[0092] S4: Based on a public dataset containing multi-physical sensor signals of lithium batteries, divide it into a training set and a validation set in proportion for pre-training and parameter adjustment, and save the pre-trained model architecture and parameters; based on the pre-trained model and the three types of characteristic data collected for different causes of lithium battery thermal runaway as a dataset, divide it into a training set and a validation set in proportion for fine-tuning and parameter adjustment, and save the fine-tuned model architecture and parameters; based on the fine-tuned warning model, use thermal runaway data different from the above training set and validation set, process the data and directly perform TR warning, which specifically includes the following steps:

[0093] S41: In the pre-training stage, the public data set is randomly masked, that is, the mask flag 0 is used to replace certain parts of the original data, and the model is regressed to enable it to process the masked input data, predict the masked part, and achieve data reconstruction, so that the model has the ability to predict the multi-physical signals inside the lithium battery. The collected and processed feature data is divided into a training set and a validation set to train and adjust the parameters of the built model. In order to allow the model to learn more data knowledge, the ratio is generally set to 8:2 or 9:1. At the same time, since the input length (interpolation) and output length (ie TR label) have been fixed, it is only necessary to set the encoder's d_model (feature dimension), the number of encoder layers, and set the learning rate, number of iterations and other parameters during training.

[0094] The training set data is input into the model with constructed parameters, and the parameters are updated by combining the back propagation algorithm. Specifically, the mean squared error loss function (MSE) is used to calculate the loss between the predicted RUL and the true RUL after each iteration. The calculation formula is as follows:

[0095]

[0096] Among them, y i is the true value of the masked physical signal in the i-th cycle, is the predicted value of the masked physical signal at the i-th cycle. The model parameters are then iteratively updated using the Rectified Adaptive Moment Estimation (RAdam) optimizer and a cosine warmup method to dynamically adjust the learning rate. After achieving the required loss accuracy over multiple iterations, the model framework and parameters are saved.

[0097] S42: During the fine-tuning phase, the collected and processed feature data is divided into a training set and a validation set to train and adjust the parameters of the constructed model. In order to allow the model to learn more data knowledge, the ratio is generally set to 8:2 or 9:1. At the same time, since the input length (interpolation) and output length (i.e. TR label) are fixed, only the encoder d_model (feature dimension) and the number of encoder layers need to be set, as well as the learning rate, number of iterations, and other parameters during training.

[0098] The training set data is input into the model with constructed parameters, and the back-propagation algorithm is used to update the parameters. Specifically, the cross-entropy loss function (CE) is used to calculate the loss between the predicted TR label and the true TR label after each iteration. The calculation formula is as follows:

[0099]

[0100] Among them, y i is the true value of the lithium battery TR label at the i-th cycle, is the probability that lithium is predicted by the model at the i-th cycle (indicating the probability that the sample belongs to category 1).

[0101] The model parameters are then iteratively updated using the Rectified Adaptive Moment Estimation (RAdam) optimizer and a cosine warmup method to dynamically adjust the learning rate. After the loss accuracy reaches the target after multiple iterations, the model framework and parameters are saved.

[0102] S43: Based on the trained network model, the lithium battery data (i.e., the test set) that requires TR warning is processed and input into the model to directly obtain the TR output.

[0103] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions.

[0104] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0105] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0106] The above embodiments should be understood as merely illustrating the present invention and not as limiting the scope of protection of the present invention. After reading the contents of the present invention, technicians may make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A lithium battery thermal runaway early warning method based on multi-physical sensing information, characterized in that: The following steps are involved: S1: Collects monitoring data on voltage, temperature, and strain of lithium batteries from normal operation to thermal runaway (TR). Data collected by sensors with a sampling frequency lower than the set value are linearly interpolated at intervals of one time unit, and continuous abnormal values ​​collected are replaced with a constant value. S2: Based on the collected lithium battery data, referring to the national standard definition of lithium battery thermal runaway TR and the data itself, select the mutation point of the multi-physical sensor signal data and add a TR label column to the data set. The data after the mutation point is true; S3: Build a Transformer encoder model based on the time series characteristics of lithium batteries and generate TR risk vectors by fusing multi-physical sensing information; S4: Based on a public dataset containing multi-physical sensor signals of lithium batteries, divide it into training sets and validation sets in proportion for pre-training and parameter adjustment, and save the pre-trained model architecture and parameters; based on the pre-trained model and the three types of characteristic data collected for different causes of lithium battery thermal runaway as a dataset, divide it into training sets and validation sets in proportion for fine-tuning and parameter adjustment, and save the fine-tuned model architecture and parameters; based on the fine-tuned warning model, use thermal runaway data different from the above training sets and validation sets, process the data and directly perform TR warning.

2. The lithium battery thermal runaway early warning method based on multi-physical sensing information according to claim 1, characterized in that: The step S1 specifically includes the following steps: S11: Select voltage, temperature and strain as three characteristic quantities for early warning. S12: Linearly interpolate the data collected by devices with lower sampling frequency at intervals of one time unit, and replace the continuous abnormal values ​​collected after the sensor is damaged with a reasonable constant value to ensure that the subsequent model can use reliable data for early warning.

3. The lithium battery thermal runaway early warning method based on multi-physical sensing information according to claim 1, characterized in that: The step S2 specifically includes the following steps: S21: Based on the definition of thermal runaway in the national standard and the different factors of actual battery materials, select the appropriate physical signal mutation point as the thermal runaway sign; S22: Add a TR label column to the multi-physics sensor information based on the signal mutation point. The value after the mutation point is true, and the value before the mutation point is false. This allows the model to learn the changing pattern of the multi-physics sensor information during the thermal runaway process during the training phase, and to verify its thermal runaway warning capability during the testing process.

4. The lithium battery thermal runaway early warning method based on multi-physical sensing information according to claim 1, characterized in that: The step S3 constructs a Transformer encoder model based on the time series characteristics of the lithium battery and generates a TR risk vector by fusing multi-physical sensor information; specifically, it includes the following steps: S31: First, the multi-physical sensing information is transformed from the original feature space to the dimension required by the model through the projection layer. At this time, the different physical quantity features are uniformly transformed in dimension, so that they can be transformed and processed in a unified time and space and prepare for subsequent encoding; then the position encoding is obtained through the position encoding layer, which provides the model with the position information of the time series, so that the model can distinguish the features of different time steps and enhance the perception of sequence information; then comes the Transformer encoder, through which the data flows to realize feature extraction and encode output; the output of the encoder enters the activation function and Dropout layer, which introduces nonlinearity and prevents overfitting; finally, the output layer maps the encoded features to the target prediction space and outputs the feature vector for thermal runaway warning; through this feature vector, the warning result of the model for TR at each time step can be known. S32: Build the Transformer encoder.

5. The lithium battery thermal runaway early warning method based on multi-physical sensing information according to claim 4, characterized in that: The S32: building a Transformer encoder specifically includes: For each input sample X∈R^(3×L), where 3 represents four features—voltage, ambient temperature, strain, and time, and L represents the sequence length, the data is normalized and the deviation normalization method is used to linearly map the data to the range [0,1]. The specific formula is as follows: Where: is the normalized value of the nth feature at the i-th moment, is the original data of the nth feature at the i-th moment, and are the minimum and maximum values ​​under the nth feature respectively. Then the feature data is positionally encoded and high-dimensional features are extracted so that it can be entered into the encoder for subsequent processing. The formula is as follows: X e =XW i +W pos Where: X is the normalized feature data, Represents a high-dimensional embedding vector with position information, d model is the size of the high-dimensional vector, represents the parameter matrix of the mapping input, is a learnable position matrix, which is used to provide input position information; then, X e Entering the self-attention mechanism module, the expression is: Where: is the query vector and the key and value vectors, d k ,d v are the dimensions of Q, K, and V, The encoder follows the idea of ​​multi-head attention and uses multiple attention heads to extract features from different subspaces. The multi-head self-attention mechanism can be expressed as: MultiHead(Q,K,V)=concat(head1,…,head n )W O Where: The parameter matrix of the mapping is as well as h is the number of attention heads; the final vector representation obtained after the multi-head attention mechanism is: X A =MultiHead(X e ) X A After residual connection and layer normalization, we get: Input to the feedforward neural network layer and apply residual connection and layer normalization again to get: Where: b1 and b2 are the bias of the network, d ff is the dimension of the feedforward layer, d AA It is d model The above is the vector representation after 1 layer of encoder, and the final vector is obtained after n layers of encoder. The vector is then passed through the activation function GELU to introduce nonlinearity and enhance the expressiveness of the model; Dropout is used to prevent overfitting and improve the generalization ability of the model; time-step averaging compresses the sequence information into a global feature representation; finally, the output layer maps the features to the dimension of the number of categories to obtain the final classification result.

6. The lithium battery thermal runaway early warning method based on multi-physical sensing information according to claim 1, characterized in that: The step S4 specifically includes the following steps: S41: In the pre-training phase, the public dataset is randomly masked, that is, the mask flag 0 is used to replace certain parts of the original data, and the model is trained for regression, so that it can process the masked input data, predict the masked parts, and achieve data reconstruction; the collected and processed feature data is divided into a training set and a validation set to train and adjust the parameters of the constructed model; to fix the input interpolation and output TR label, it is only necessary to set the encoder's d_model feature dimension and the number of encoder layers, and set the parameters such as the learning rate and number of iterations during training; The training set data is input into the model with constructed parameters, and the back propagation algorithm is used to update the parameters. Specifically, the mean square error loss function MSE is used to calculate the loss between the predicted RUL and the true RUL after each iteration. The calculation formula is as follows: Among them, y i is the true value of the masked physical signal in the i-th cycle, is the predicted value of the masked physical signal in the i-th cycle; then the modified adaptive moment estimation RAdam optimizer is used, and the cosine warm-up method is used to dynamically adjust the learning rate and iteratively update the parameters in the model; after the loss accuracy obtained after multiple iterations meets the standard, the model framework structure and model parameters are saved; S42: In the fine-tuning stage, the collected and processed feature data is divided into a training set and a validation set to train and adjust the parameters of the built model; the input interpolation and output TR label are fixed, the encoder's d_model feature dimension and the number of encoder layers are set, and the parameters including the learning rate and number of iterations during training are set; The training set data is input into the model with constructed parameters, and the back propagation algorithm is used to update the parameters. Specifically, the cross entropy loss function CE is used to calculate the loss of the predicted TR label and the true TR label after each iteration. The calculation formula is as follows: Among them, y i is the true value of the lithium battery TR label at the i-th cycle, is the probability that lithium is predicted by the model at the i-th cycle, which means the probability that the sample belongs to category 1; Then, the modified adaptive moment estimation RAdam optimizer is used, and the cosine warm-up method is used to dynamically adjust the learning rate and iteratively update the parameters in the model. After the loss accuracy obtained after multiple iterations meets the standard, the model framework structure and model parameters are saved. S43: Based on the trained network model, the lithium battery data that needs TR warning, i.e., the test set data, is processed and input into the model to directly obtain the TR output.

7. An electronic device, characterized in that: It comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the lithium battery thermal runaway warning method based on multi-physical sensing information as claimed in any one of claims 1 to 6 is implemented.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the lithium battery thermal runaway warning method based on multi-physical sensing information as described in any one of claims 1 to 6 is implemented.

Images