Method and device for predicting thermal runaway of power battery, storage medium and electronic equipment

Abstract

The application provides a thermal runaway prediction method and device of a power battery, a storage medium and an electronic device. The method comprises: acquiring a time sequence of a plurality of operating parameters of the power battery in a first historical time period, and determining a plurality of embedding matrices according to the time sequence, wherein the plurality of operating parameters comprise a voltage and a temperature of the power battery; acquiring a static text matrix and a dynamic text matrix of the power battery; obtaining a fusion data matrix according to the embedding matrices, the static text matrix and the dynamic text matrix, and inputting the fusion data matrix into a target large language model to predict an operating parameter prediction value of the power battery in a future time period, so as to predict a thermal runaway state of the power battery, wherein the operating parameter prediction value comprises a voltage prediction value and a temperature prediction value of the power battery. The method realizes all-round perception and risk assessment of the battery operating state through steps such as time sequence data reprogramming, cross-modal information fusion and intelligent analysis of the large language model.

Description

Technical Field

[0001] This application relates to the field of automotive technology, and more specifically, to a method for predicting thermal runaway of a power battery, a device for predicting thermal runaway of a power battery, a computer-readable storage medium, and an electronic device. Background Technology

[0002] With the rapid development of new energy vehicles and energy storage systems, the safety of power batteries has received increasing attention. Thermal runaway, as the most serious safety accident involving power batteries, is characterized by its suddenness, rapid spread, and significant harm, potentially leading to severe consequences. Therefore, establishing an efficient, accurate, and real-time thermal runaway early warning mechanism is crucial for ensuring the safe operation of battery systems. Current thermal runaway early warning methods mostly rely on single sensor data (such as temperature or voltage), making it difficult to accurately capture the complex signs preceding thermal runaway. These methods suffer from high false alarm rates, response delays, and poor interpretability, limiting the intelligence and practicality of early warning systems. Summary of the Invention

[0003] The main objective of this application is to provide a method, device, computer-readable storage medium, and electronic device for predicting thermal runaway of power batteries, in order to at least solve the problems of high false alarm rate and false negative rate, poor real-time performance, and poor accuracy of existing thermal runaway early warning methods for power batteries.

[0004] To achieve the above objectives, according to one aspect of this application, a method for predicting thermal runaway of a power battery is provided, comprising: acquiring time series of multiple operating parameters of the power battery within a first historical time period, and determining multiple embedding matrices based on the time series, wherein the multiple operating parameters include the voltage and temperature of the power battery, one operating parameter corresponds to one time series, and one embedding matrix corresponds to a time segment in one time series; acquiring a static text matrix and a dynamic text matrix of the power battery, wherein the static text matrix includes text information of the operating state and operating parameters of the power battery, and the dynamic text matrix includes text information of the start and end times of different time segments in the time series; obtaining a fused data matrix based on the embedding matrix, the static text matrix, and the dynamic text matrix, and inputting the fused data matrix into a target large language model to predict the predicted values ​​of the operating parameters of the power battery in a future time period, so as to predict the thermal runaway state of the power battery, wherein the predicted values ​​of the operating parameters include the predicted values ​​of the voltage and temperature of the power battery.

[0005] Optionally, determining multiple embedding matrices based on the time series includes: reversibly normalizing the time series to obtain a normalized time series; dividing the normalized time series into multiple time segments based on a sliding window; inputting all the time segments into a linear layer for dimensionality transformation to obtain the embedding matrix corresponding to each time segment, wherein each embedding matrix includes the data content of a corresponding time segment.

[0006] Optionally, before dividing the normalized time series into multiple time segments based on a sliding window, the method further includes: determining the difference between the length of the time series and a preset time length as a target difference, wherein the preset time length is a preset length of the time segment; determining the ratio of the target difference to the window sliding step size as a target ratio; and determining the number of time segments by the sum of the target ratio and a preset value, wherein the preset value is 2.

[0007] Optionally, obtaining the fused data matrix based on the embedding matrix, the static text matrix, and the dynamic text matrix includes: obtaining text prototype vectors from a semantic subset of the word embedding matrix of the pre-trained language model, and determining a reprogramming matrix based on the text prototype vectors and the embedding matrix; obtaining a fused input matrix based on the reprogramming matrix and the static text matrix, the fused input matrix including all data information of the embedding matrix and the static text matrix; and summing the fused input matrix and the dynamic text matrix to obtain the fused data matrix, wherein one fused data matrix corresponds to one time segment.

[0008] Optionally, determining the reprogramming matrix based on the text prototype vector and the embedding matrix includes: determining a vector space based on the text prototype vector, wherein the vector space at least represents the vocabulary size of the pre-trained language model and the hidden dimension of the backbone model of the pre-trained language model; unifying the dimensions of the embedding matrix and the vector space using a multi-head cross-attention mechanism, and weighting the semantics of the embedding matrix and the vector space to obtain the reprogramming matrix.

[0009] Optionally, the static text matrix further includes cross-mapping information of the operating parameters of the power battery. The cross-mapping information includes the mapping relationship between different operating parameters. Based on the reprogramming matrix and the static text matrix, a fusion input matrix is ​​obtained, including: embedding and concatenating the static text matrix corresponding to a time segment as a prefix with the corresponding multiple reprogramming matrices to obtain the fusion input matrix. The multiple reprogramming matrices are reprogramming matrices corresponding to different operating parameters, and a time segment corresponds to multiple reprogramming matrices.

[0010] Optionally, before inputting the fused data matrix into the target large language model to predict the predicted operating parameters of the power battery in the future time period, the method further includes: acquiring an initial large language model and multiple sets of training data for the initial large language model, each set of training data including: a historical fused data matrix and the historical predicted operating parameters corresponding to the historical fused data matrix acquired in a second historical time period, the second historical time period being earlier than the first historical time period; training and adjusting the initial large language model using a quantized low-rank optimization method based on a loss function and the multiple sets of training data to obtain the target large language model, the target large language model including two fully connected layers.

[0011] According to another aspect of this application, a thermal runaway prediction device for a power battery is provided, comprising: a first acquisition unit, configured to acquire time series of multiple operating parameters of the power battery within a first historical time period, and determine multiple embedding matrices based on the time series, wherein the multiple operating parameters include the voltage and temperature of the power battery, one operating parameter corresponds to one time series, and one embedding matrix corresponds to a time segment in one time series; a second acquisition unit, configured to acquire a static text matrix and a dynamic text matrix of the power battery, wherein the static text matrix includes text information of the operating state and operating parameters of the power battery, and the dynamic text matrix includes text information of the start and end times of different time segments in the time series; and a prediction unit, configured to obtain a fused data matrix based on the embedding matrix, the static text matrix, and the dynamic text matrix, and input the fused data matrix into a target large language model to predict the predicted values ​​of the operating parameters of the power battery within a future time period, so as to predict the thermal runaway state of the power battery, wherein the predicted values ​​of the operating parameters include the predicted values ​​of the voltage and temperature of the power battery.

[0012] According to another aspect of this application, a computer-readable storage medium is provided, the computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to perform any of the aforementioned thermal runaway prediction methods for power batteries.

[0013] According to another aspect of this application, an electronic device is provided, comprising: one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including methods for performing thermal runaway prediction of any of the aforementioned power batteries.

[0014] Applying the technical solution of this application, the above-mentioned method for predicting thermal runaway of a power battery first obtains the time series of multiple operating parameters of the power battery within a first historical time period, and determines multiple embedding matrices based on the time series. The multiple operating parameters include the voltage and temperature of the power battery, with one operating parameter corresponding to one time series and one embedding matrix corresponding to a time segment in one time series. Then, it obtains a static text matrix and a dynamic text matrix of the power battery. The static text matrix includes text information about the operating state and operating parameters of the power battery, while the dynamic text matrix includes text information about the start and end times of different time segments in the time series. Finally, it obtains a fused data matrix based on the embedding matrix, the static text matrix, and the dynamic text matrix, and inputs the fused data matrix into a target large language model to predict the predicted values ​​of the operating parameters of the power battery within future time periods, thereby predicting the thermal runaway state of the power battery. The predicted operating parameter values ​​include the predicted voltage and temperature values ​​of the power battery. This method achieves comprehensive perception and risk assessment of battery operating status through steps such as time-series data reprogramming, cross-modal information fusion, and intelligent analysis of large language models. It effectively improves the accuracy and timeliness of thermal runaway early warning, reduces the probability of false alarms and missed alarms, and solves the problems of high false alarm and missed alarm rates, poor real-time performance, and poor accuracy of existing thermal runaway early warning methods for power batteries. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0016] Figure 1 A hardware structure block diagram of a mobile terminal for performing a thermal runaway prediction method for a power battery, according to an embodiment of this application, is shown.

[0017] Figure 2 A schematic flowchart of a method for predicting thermal runaway of a power battery according to an embodiment of this application is shown.

[0018] Figure 3 A schematic diagram of a method for predicting thermal runaway of a power battery according to an embodiment of this application is shown.

[0019] Figure 4 A flowchart illustrating another method for predicting thermal runaway of a power battery according to an embodiment of this application is shown.

[0020] Figure 5 A flowchart illustrating another method for predicting thermal runaway of a power battery according to an embodiment of this application is shown.

[0021] Figure 6A structural block diagram of a thermal runaway prediction device for a power battery according to an embodiment of this application is shown.

[0022] The above figures include the following reference numerals:

[0023] 102. Processor; 104. Memory; 106. Transmission device; 108. Input / output device. Detailed Implementation

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0026] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0027] For ease of description, the following explains some of the nouns or terms used in the embodiments of this application:

[0028] As described in the background section, most current thermal runaway early warning methods rely on single sensor data (such as temperature or voltage), making it difficult to accurately capture the complex signs preceding thermal runaway. These methods suffer from high false alarm rates, response delays, and poor interpretability, limiting the intelligence and practicality of early warning systems. To address the problems of high false alarm and false negative rates, poor real-time performance, and low accuracy in existing thermal runaway early warning methods for power batteries, embodiments of this application provide a thermal runaway prediction method for power batteries, a thermal runaway prediction device for power batteries, a computer-readable storage medium, and an electronic device.

[0029] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0030] The methods and embodiments provided in this application can be executed on a mobile terminal, computer terminal, or similar computing device. Taking running on a mobile terminal as an example, Figure 1 This is a hardware structure block diagram of a mobile terminal for a power battery thermal runaway prediction method according to an embodiment of the present invention. Figure 1 As shown, a mobile terminal may include one or more ( Figure 1 Only one is shown in the diagram. A processor 102 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.) and a memory 104 for storing data are also shown. The mobile terminal may further include a transmission device 106 for communication functions and an input / output device 108. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the mobile terminal described above. For example, the mobile terminal may also include components that are more... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0031] The memory 104 can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the thermal runaway prediction method for power batteries in this embodiment of the invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, thereby implementing the above-described method. The memory 104 may include high-speed random access memory and non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the mobile terminal via a network. Examples of the aforementioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. The transmission device 106 is used to receive or send data via a network. Specific examples of the aforementioned networks may include wireless networks provided by the mobile terminal's communication provider. In one example, the transmission device 106 includes a network interface controller (NIC), which can be connected to other network devices via a base station to communicate with the Internet. In one example, the transmission device 106 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0032] This embodiment provides a method for predicting thermal runaway of a power battery that runs on a mobile terminal, computer terminal, or similar computing device. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0033] Figure 2 This is a flowchart of a power battery thermal runaway prediction method according to an embodiment of this application. Figure 2 As shown, the method includes the following steps:

[0034] Step S201: Obtain the time series of multiple operating parameters of the power battery within the first historical time period, and determine multiple embedding matrices based on the time series. The multiple operating parameters include the voltage and temperature of the power battery. One operating parameter corresponds to one time series, and one embedding matrix corresponds to a time segment in one time series.

[0035] Specifically, the first historical time period refers to a specific time interval for data collection. In practice, this typically refers to a period in the past, such as the past few hours, days, or weeks, depending on the context length and prediction time span required for thermal runaway warning. Operating parameters specifically refer to measurements that affect battery state or performance, such as battery voltage and temperature. These parameters change over time, forming a series of values—time series data. Each operating parameter (such as voltage or temperature) has a numerical sequence associated with its timestamp. For example, battery voltage might be recorded every minute, forming a sequence of voltage changes over time; similarly, battery temperature also has its own sequence of changes over time. These sequences provide detailed information about the evolution of battery state over time.

[0036] For each time series, and especially for each time segment (or patch) it comprises, an embedding matrix is ​​generated. The embedding matrix is ​​transformed from the numerical time segment data into a low-dimensional vector representation that the model can understand. Specifically, the data for each time segment is converted into a vector, and all these vectors form a matrix, which is the so-called embedding matrix. This means that if a time segment of length L_p is extracted from the time series, then this segment will be transformed into an embedding matrix of a specific size (e.g., R^(P×d_m)). Therefore, for the entire time series, there will be a series of embedding matrices, each corresponding to an independent segment of the time series. This segmentation helps the model understand local features and long-term dependencies in the data.

[0037] Step S202: Obtain the static text matrix and dynamic text matrix of the power battery. The static text matrix includes text information of the operating status and operating parameters of the power battery, and the dynamic text matrix includes text information of the start and end times of different time segments in the time series.

[0038] Specifically, static means that this information is relatively fixed and does not change over time. It provides basic background and attribute descriptions of the battery, such as battery type, chemical composition, maximum voltage and temperature range, and rated capacity. This textual information helps the model build a preliminary understanding of battery characteristics, enabling a certain degree of reasoning and understanding based on textual descriptions even in the absence of specific numerical data.

[0039] Operating status text information may include the battery's charge / discharge state, health status, and current operating mode (such as charging, discharging, or idle). For example, "The battery is charging in constant current mode" or "The battery is in good health with no obvious signs of aging." Operating parameter text information specifically describes key parameters of the battery during operation, such as voltage, current, and temperature. Although these parameters are presented numerically in the time series, they are given in textual descriptions in the static text matrix, such as "The maximum recommended operating temperature of the battery is 45°C" or "The ideal charging voltage range of the battery is 3.5V to 4.2V."

[0040] Dynamic text matrices focus on the temporal attributes of time-series data, specifically recording the start and end times of each time patch. "Dynamic" means that this information changes over time, reflecting the dynamic characteristics of battery state over time.

[0041] Step S203: Obtain a fused data matrix based on the above-mentioned embedding matrix, the above-mentioned static text matrix and the above-mentioned dynamic text matrix, and input the above-mentioned fused data matrix into the target large language model to predict the predicted values ​​of the operating parameters of the power battery in the future time period, so as to predict the thermal runaway state of the power battery. The predicted values ​​of the operating parameters include the predicted values ​​of the voltage and temperature of the power battery.

[0042] Specifically, the embedding matrix contains features extracted from the power battery time-series data. Each time segment has been transformed into a low-dimensional vector representation, and these vectors are combined into a matrix reflecting the battery's state at different time points. The static text matrix contains static attribute information about the power battery, such as battery type, chemical composition, maximum voltage, and temperature range. This information does not change over time but is crucial for understanding the battery's operating environment and potential state. The dynamic text matrix contains timestamp information for each time segment in the time series, i.e., the start and end times. This information tells the model the position of the current data segment on the time axis, helping the model understand the dynamic changes and historical evolution of the time series. The fused data matrix, by combining the information from the embedding matrix, static text matrix, and dynamic text matrix, creates a more comprehensive input representation that includes the numerical features, static attributes, and temporal location of the battery state. This fusion ensures that the model can not only understand the current data point but also make predictions based on a broader knowledge background and temporal context.

[0043] A target large language model refers to an LLM (Limited Language Model) fine-tuned to meet the specific needs of power battery forecasting. For example, GPT-2, fine-tuned using QLoRA, is specifically designed for time series forecasting in the power battery field. The fused data matrix is ​​input into the target LLM, leveraging its powerful sequence understanding and long-term dependency capture capabilities to perform deep analysis and inference on the information within the matrix. The model outputs predictions of battery operating parameters for future time periods, including predicted voltage and temperature values. These predictions are based on current and historical data, as well as the potential evolutionary patterns learned during model training.

[0044] Upon predicting the thermal runaway state of the power battery, the vehicle control system will take a series of preventative and reactive measures to ensure passenger safety and minimize vehicle damage. For example, the warning system will immediately issue a thermal runaway warning to the driver and vehicle control center, alerting the driver to safety and reporting the specific situation to the control center for further action. The vehicle's Energy Management System (EMS) will adjust the battery's charging and discharging strategy based on the warning information, such as reducing the battery's charging and discharging power, stopping battery charging, or switching to a safer battery operating mode to reduce the risk of thermal runaway. The Thermal Management System (TMS) will activate emergency cooling protocols, such as increasing coolant flow, turning on additional cooling fans, or using the air conditioning to force-cool the battery area to control battery temperature and prevent thermal runaway. In extreme cases, the vehicle control system may disconnect the battery from the vehicle's electrical system to prevent electrical fires or explosions caused by thermal runaway. The control system may automatically reduce the vehicle's speed or guide the vehicle to a safe stop to prevent the vehicle from being in a high-speed driving state during thermal runaway, which would increase the danger. If the thermal runaway warning level reaches a certain threshold, the vehicle control system will issue passenger evacuation instructions via in-vehicle displays or voice systems to ensure passengers can leave the vehicle promptly and safely. The warning system may automatically contact emergency services, such as the fire department, providing vehicle location and warning information so that professionals can respond quickly and take measures. The vehicle control system will record relevant data from this warning event, including battery voltage, temperature, and charge / discharge status. This data will be used for subsequent fault analysis and system optimization to improve the accuracy of the warning and the overall safety of the system.

[0045] By comprehensively applying these measures, vehicles can take timely action when the risk of thermal runaway of the power battery is predicted, minimizing potential hazards and protecting the safety of passengers and the vehicle. These control strategies are designed with the suddenness and severity of thermal runaway in mind, ensuring a rapid response and effective countermeasures when a warning is triggered.

[0046] The thermal runaway prediction method for power batteries described in this application first obtains time series of multiple operating parameters of the power battery within a first historical time period, and determines multiple embedding matrices based on the time series. The multiple operating parameters include the voltage and temperature of the power battery, with one operating parameter corresponding to one time series and one embedding matrix corresponding to a time segment within one time series. Next, it obtains a static text matrix and a dynamic text matrix of the power battery. The static text matrix includes textual information about the operating state and operating parameters of the power battery, while the dynamic text matrix includes textual information about the start and end times of different time segments in the time series. Finally, it obtains a fused data matrix based on the embedding matrix, the static text matrix, and the dynamic text matrix, and inputs the fused data matrix into a target large language model to predict the predicted values ​​of the operating parameters of the power battery within future time periods, thereby predicting the thermal runaway state of the power battery. The predicted operating parameter values ​​include the predicted voltage and temperature values ​​of the power battery. This method achieves comprehensive perception and risk assessment of battery operating status through steps such as time-series data reprogramming, cross-modal information fusion, and intelligent analysis of large language models. It effectively improves the accuracy and timeliness of thermal runaway early warning, reduces the probability of false alarms and missed alarms, and solves the problems of high false alarm and missed alarm rates, poor real-time performance, and poor accuracy of existing thermal runaway early warning methods for power batteries.

[0047] The above embodiments, based on the powerful contextual understanding and reasoning capabilities of large language models, combine data reprogramming and cross-modal fusion techniques to address the limitations of existing technologies in knowledge preservation and adaptability, thereby meeting the practical needs of fault early warning models for long-term knowledge memory, meta-knowledge extraction, and rapid parameter adjustment. The essence of long-term prediction for thermal runaway early warning is to use multivariate time series inputs to achieve long-term multivariate prediction of the future. The problem of long-term multivariate prediction enhanced by natural language can be described as follows: Given a historical review window with H time steps and K variables... The goal is to generate future insights with F time steps and K variables. In addition, timestamps are used. As a covariate at each time point Alignment, using This represents a natural language description that includes general knowledge and cross-references. The core of this invention is training an LLM-based time series predictor. To achieve the above transformation, the specific formula is as follows: ,in, Indicates by function The predicted target sequence tensor.

[0048] The powerful generalization performance of the method proposed in the above embodiments is driven by its use of unique domain knowledge embedded in the pre-trained LLM. A schematic diagram of the method can be found below. Figure 3The left section depicts the workflow of text modality processing, while the right section focuses on power battery parameter processing and cross-modal fusion. To minimize training costs, this method uses an existing, fully trained base LLM. All LLM parameters are frozen to eliminate the need for fine-tuning the backbone model.

[0049] In some embodiments, such as Figure 4 As shown, multiple embedding matrices are determined based on the above time series, including the following steps:

[0050] Step S2011: Perform reversible normalization on the above time series to obtain the normalized time series;

[0051] Step S2012: Divide the normalized time series into multiple time segments based on a sliding window.

[0052] Step S2013: Input all the above time segments into the linear layer for dimensionality transformation to obtain the above embedding matrix corresponding to each of the above time segments, wherein each of the above embedding matrices includes the data content of a corresponding time segment.

[0053] In this embodiment, the technical solution for early warning of thermal runaway in power batteries is further refined. Specific steps include processing the voltage and temperature time-series data of the power battery. First, the original data is reversibly normalized to ensure zero mean and unit standard deviation, thereby mitigating the problem of inconsistent distribution across samples. Then, a sliding window strategy is used to segment the normalized time-series data into a series of overlapping or non-overlapping time segments, forming a patch sequence. These time segments are processed by a linear layer and converted into their low-dimensional representation, namely an embedding matrix, where each row represents a pseudo-token representation of a time segment, facilitating subsequent processing. Next, through a multi-head cross-attention module, the embedding matrix of each time segment is reprogrammed into a semantically weighted combination of text prototype vectors. This step gives the time-series data a structure compatible with natural language processing, making it easier for large language models to understand. Following this, to enhance the model's understanding of the power battery characteristics, static and dynamic semantic information is added through a knowledge-hint structure. Static hints include background knowledge of the input variables and task instructions, while dynamic hints cover the changing trends and statistical attributes of the input data. This information is integrated into the time series in the form of sequence embedding vectors, further enriching the data's representation and meaning. Finally, the preprocessed and knowledge-fused time-series data is fed into a large language model for fine-tuning. The QLoRA mechanism is used to adapt the model to the time-series prediction requirements of power batteries. The prediction results are generated through two fully connected layers, realizing the prediction of future changes in power battery voltage and temperature. This provides a key information basis for early warning of thermal runaway, thereby improving the accuracy and timeliness of the early warning.

[0054] This series of processing steps not only improves the model's prediction accuracy and reduces the false alarm rate, but also significantly enhances the intelligence and adaptability of the early warning system by combining the powerful reasoning capabilities of the large language model and the in-depth understanding of cross-modal data. Overall, the solution in this embodiment can capture subtle changes in the operation of the power battery more comprehensively and meticulously, predicting potential thermal runaway risks in advance, and providing strong support for the safety management of new energy vehicles and energy storage systems. In terms of implementation effectiveness, through precise analysis and cross-modal fusion of the collected data, the early warning system can make accurate judgments under complex operating conditions, gaining valuable time for timely measures to prevent accidents. Of course, the choice of the large language model can be flexibly adjusted according to the specific application environment and performance requirements, and is not limited to GPT-2. Other models such as LLaMA-3-8B and Qwen can also be used as alternatives to achieve the best early warning effect.

[0055] In some embodiments, before dividing the normalized time series into multiple time segments based on a sliding window, the method further includes the following steps:

[0056] Step S301: Determine the difference between the length of the time series and the preset time length as the target difference, where the preset time length is the preset length of the time segment.

[0057] Step S302: Determine the ratio of the target difference to the window sliding step as the target ratio.

[0058] Step S303: The sum of the target ratio and the preset value is determined as the number of time segments, where the preset value is 2.

[0059] In this embodiment, to ensure the integrity of time series data and the consistency of the prediction model's input, the method further includes a preprocessing step for the time series length. Specifically, the difference between the time series length and a preset time length is determined as the target difference, where the preset time length is the length of the sliding window, i.e., the length of a preset time segment. Subsequently, the ratio of the target difference to the window sliding step size is determined as the target ratio. Finally, the sum of the target ratio and a preset value is determined as the number of time segments, typically 2. This preprocessing effectively solves the problem of unstable model input that may be caused by inconsistent data lengths. By adjusting the number of time segments, it ensures that the input data matches the format required by the model, providing a reliable foundation for subsequent data reprogramming and cross-modal fusion. During the model training and prediction stages, this design helps the model better learn the potential characteristics and evolution patterns of the time series, thereby improving the accuracy of predictions and the model's generalization ability. Of course, in other embodiments, different strategies can be used to handle the differences in time series length, such as data padding or pruning. However, this embodiment provides a more refined and flexible solution that can adapt to time series data of different lengths, ensuring the standardization and normalization of model input, and further improving the stability and predictive performance of the power battery thermal runaway early warning technology based on a large language model. In other embodiments, the number of time segments can also be optimized by dynamically adjusting the window sliding step size or a preset value to adapt to more complex data scenarios and improve the model's adaptability and flexibility.

[0060] In some embodiments, each input channel First, the data is reversibly normalized individually to have zero mean and unit standard deviation, mitigating distribution drift across samples. Then, the normalized time series is divided into several segments of length [missing information - likely a number]. Overlapping or non-overlapping patches (i.e., time segments), with a window sliding step size of . S This ultimately forms a sequence of patches, with a total number of patches. These patches are embedded into a low-dimensional representation by a simple linear layer, resulting in an embedding matrix. Each line is equivalent to a pseudo token.

[0061] In some embodiments, obtaining the fused data matrix based on the above-described embedding matrix, the above-described static text matrix, and the above-described dynamic text matrix includes the following steps:

[0062] Step S401: Obtain text prototype vectors from the semantic subset of the word embedding matrix of the pre-trained language model, and determine the reprogramming matrix based on the text prototype vectors and the embedding matrix.

[0063] Step S402: Based on the above reprogramming matrix and the above static text matrix, a fusion input matrix is ​​obtained. The fusion input matrix includes all data information of the above embedding matrix and the above static text matrix.

[0064] Step S403: Summing the above-mentioned fused input matrix and the above-mentioned dynamic text matrix to obtain the above-mentioned fused data matrix, wherein one of the above-mentioned fused data matrices corresponds to one of the above-mentioned time segments.

[0065] In this embodiment, the generation of the fused data matrix involves the comprehensive processing of the embedding matrix, static text matrix, and dynamic text matrix. First, text prototype vectors are selected from the word embedding matrix of the pre-trained language model. Based on these vectors and the normalized time-series embedding matrix, a reprogramming matrix is ​​determined to achieve semantic reconstruction of the time-series data. Then, the embedding matrix is ​​fused with the static text matrix containing domain knowledge and task instructions to generate a fused input matrix, ensuring the model can acquire comprehensive data information and background knowledge. Finally, the fused input matrix and the dynamic text matrix are combined element-wise to form a fused data matrix, with one such matrix corresponding to each time segment, facilitating deep learning and long-term prediction during subsequent fine-tuning of the language model. This series of processing steps effectively improves the model's ability to perceive the operating state of the power battery. Through cross-modal information fusion, the model can more accurately capture signs of impending thermal runaway, achieving early warning and thus enhancing the safety and reliability of the battery system. In other embodiments not described, the choice of the large language model can be extended to LLaMA-3-8B, Qwen, etc., to explore the impact of different models on prediction performance and further optimize the practicality of the early warning system. This technical solution not only improves prediction accuracy but also enhances the model's generalization ability, effectively addressing the complexity and uncertainty of battery operating conditions.

[0066] In some embodiments, the reprogramming matrix is ​​determined based on the text prototype vector and the embedding matrix described above, including the following steps:

[0067] Step S4011: Determine the vector space based on the above text prototype vectors. The vector space at least represents the vocabulary size of the above pre-trained language model and the hidden dimension of the backbone model of the above pre-trained language model.

[0068] Step S4012: The dimensions of the above embedding matrix and the above vector space are unified by a multi-head cross-attention mechanism, and the semantics of the above embedding matrix and the above vector space are weighted and combined to obtain the above reprogramming matrix.

[0069] In this embodiment, based on the time-series data of power battery voltage and temperature, a vector space is first determined by employing text prototype vectors and a multi-head cross-attention mechanism. This space at least represents the vocabulary of the pre-trained language model and the hidden dimensions of the backbone model, thereby achieving effective semantic parsing of the time-series data. Subsequently, the dimensions of the embedding matrix and the vector space are unified using the multi-head cross-attention mechanism, and the semantic information of the embedding matrix and the vector space is weighted and combined to obtain the reprogramming matrix. The key to this technical step is to convert numerical time-series data into a semantic representation that the language model can understand. Through the cooperation of the vector space and the multi-head attention mechanism, the numerical data can be reorganized and encoded, retaining its key features while being converted into a form that can be processed by a large language model. This not only improves the semantics of the data but also helps the model more accurately capture the potential correlations between data, thereby improving the accuracy and timeliness of power battery thermal runaway early warning. During model training, this reprogramming mechanism enables the model to effectively learn long-term dependencies in the time-series data and fine-tune potential patterns in the data. Through deep learning, the evolution of the battery state can be accurately predicted, thereby assisting the decision-making of the early warning system. In other embodiments, the choice of large language model can be adjusted based on specific application scenarios and resource constraints, such as switching from GPT-2 to LLaMA-3-8B or Qwen to optimize model performance. This flexibility is particularly important in practical deployments, ensuring that the early warning system maintains high efficiency and accuracy under different conditions.

[0070] In some embodiments, patch embedding reconstruction is performed after obtaining the initial patch embeddings. Since language models can only understand "linguistic embedding representations," patch vectors cannot be directly input into the language model; they must be "disguised" as language tokens. Based on this, a set of learnable text prototype vectors is introduced. These vectors are taken from a semantic subset of the word embedding matrix of the pre-trained language model, forming a new vector space. ,in V Indicates vocabulary size. D This is the hidden dimension of the backbone model. It is embedded in each patch through a multi-head cross-attention module. It will be reprogrammed into a semantically weighted combination of the prototype set, thus outputting a new patch representation. Its dimensions match the token embedding dimensions required by the language model.

[0071] After completing the patch embedding reprogramming, the text is not directly input into the language model. Instead, a knowledge hint structure is added, primarily consisting of static and dynamic semantic knowledge hints. The static hints integrate basic background information about the input variables, task instructions, and input statistics. The other part includes cross-hints. Due to independent channel configuration, this invention uses cross-hints to achieve cross-channel modeling, including statistical analysis of relevant variables to understand their impact on the input variables. This text content is then processed by the language model's tokenizer and embedding module, transforming it into sequence embedding vectors. ,Will As a prefix, it is concatenated with the time series patch embedding to form the input sequence:

[0072] Specifically, for predicting thermal runaway of power batteries, battery voltage and temperature are selected as two input variables. Then, the present invention further uses text to represent dynamic semantic prompts. Furthermore, it uses positional embedding to align different time points, thus not increasing the context length. The fusion method can be expressed by the following equation: In this case, summation is an element-wise addition. Representing each time segment The input token is used as a cue word. Cue words, acting as context, help the language model better understand the structure of the input data and the task objective. It is a key enhancement mechanism for activating LLM reasoning capabilities.

[0073] In some embodiments, the static text matrix further includes cross-mapping information of the operating parameters of the power battery. The cross-mapping information includes the mapping relationship between different operating parameters. Based on the reprogramming matrix and the static text matrix, a fusion input matrix is ​​obtained, including: embedding and concatenating the static text matrix corresponding to one time segment with the corresponding multiple reprogramming matrices to obtain the fusion input matrix. The multiple reprogramming matrices are reprogramming matrices corresponding to different operating parameters, and one time segment corresponds to multiple reprogramming matrices.

[0074] In this embodiment, the static text matrix further includes not only the basic background information, task instructions, and statistical characteristics of the power battery, but also the cross-mapping information between different operating parameters, i.e., the mutual influence relationship between multiple operating parameters such as voltage and temperature. This design enriches the semantic understanding of time-series data, enabling the capture of more dimensions of battery state changes. In specific implementation, the reprogramming matrix of each time segment is fused with the corresponding static text matrix through prefix concatenation to form a fused input matrix, where each reprogramming matrix corresponds to a battery operating parameter. Through this fusion, the model can simultaneously learn the evolution pattern of individual parameters and the interaction between parameters, making the prediction results more accurate and enabling earlier identification of potential risks of thermal runaway. The application of the multi-head attention mechanism further strengthens the model's understanding of cross-parameter dependencies, thereby improving the overall performance and reliability of the early warning system. In other embodiments, although not described in detail, it is foreseeable that by adjusting parameter settings and model structure, such as changing the construction method of the reprogramming matrix or using different large language models for fine-tuning, accurate early warning of power battery thermal runaway can also be achieved, demonstrating the flexibility and scalability of this technical solution.

[0075] In some embodiments, before inputting the fused data matrix into the target large language model to predict the predicted operating parameters of the power battery for the future time period, the method further includes the following steps:

[0076] Step S501: Obtain the initial large language model and multiple sets of training data of the initial large language model. Each set of training data includes the historical fusion data matrix and the historical running parameter prediction values ​​corresponding to the historical fusion data matrix obtained in the second historical time period. The second historical time period is earlier than the first historical time period.

[0077] Step S502: The initial large language model is trained and adjusted using the quantized low-rank optimization method and based on the loss function and the above-mentioned multiple sets of training data to obtain the target large language model, which includes two fully connected layers.

[0078] In this embodiment, to further optimize the performance of the large language model in power battery thermal runaway early warning, the technical solution includes an additional training phase. This phase involves acquiring an initial large language model and multiple sets of training data. Each set of training data includes a historical fusion data matrix collected within a second historical time period and corresponding predicted values ​​of historical operating parameters. Notably, the second historical time period precedes the first historical time period. The Quantized Low-Rank Optimization (QLoRA) method is employed, and based on the loss function and the aforementioned multiple sets of training data, the initial large language model is fine-tuned to obtain the target large language model. The target large language model consists of two fully connected layers, enabling the model to learn the potential evolution patterns of power battery voltage and temperature from historical data, thereby improving the accuracy and reliability of predicting the future operating state of the battery.

[0079] In some embodiments, the assembled sequence The data will be fed into a large language model (GPT-2 in this embodiment) for complete context processing and semantic modeling. Since the language model itself has powerful cross-token reasoning and long-term dependency modeling capabilities, this invention uses the Quantized Low-Rank Optimization (QLoRA) method to fine-tune it to suit the field of power battery time series prediction, and relies on the cue context and patch expression to complete the understanding of the dynamic changes in voltage and temperature time series.

[0080] The final step employs a two-layer fully connected structure as the predictor, independently projecting each prediction token onto the time segment to be predicted, thus obtaining the long-term predicted values ​​of the actual voltage and temperature of the power battery. During the training phase, the mean squared error (MSE) loss function is used to calculate the error for each segment separately. Ultimately, based on the predicted voltage and temperature values, the thermal runaway state of the power battery can be assessed and warned in advance.

[0081] In addition, some embodiments also provide a power battery thermal runaway early warning device based on data reprogramming and cross-modal fusion, including four main modules: acquisition module, construction module, execution module, and early warning module. The specific functions of each module are as follows: Acquisition module: acquires n consecutive voltage and temperature time series of the power battery, each time series containing index values ​​at multiple time points, where n is a positive integer greater than 1; Construction module: performs slicing and knowledge semantic prompting on the n time series to obtain an input sequence with a unified representation; Execution module: inputs the n time series into a pre-trained power battery thermal runaway prediction model to obtain the index data of the time series prediction model for the τ time series following the n time series of equal length, where τ is a positive integer greater than or equal to 1, and the length of each time series is equal; Early warning module: based on the thermal runaway early warning conditions, determines whether the predicted voltage and temperature index data for the next τ time series exceed the early warning threshold, and outputs the early warning result.

[0082] The advantages of the above embodiments include the following aspects:

[0083] 1. Time-series data reprogramming: The multi-dimensional time-series data such as voltage and temperature collected during the operation of the power battery are semantically parsed and described in natural language using the prototype vector and multi-head attention mechanism in this paper. This gives the numerical data a semantic structure, which is convenient for large language models to understand and reason, and realizes the leap from "digital signal" to "knowledge representation".

[0084] 2. Cross-modal information fusion: Knowledge of the power battery field (including its basic background information description, task instructions and statistical features such as mean, rate of change, volatility, etc.) as well as multi-channel cross-hint embedding and timestamps are embedded into the time-series input sequence through prefix and position respectively, so as to realize the organic integration of text knowledge and time-series data, and enhance the credibility and security of the model in actual deployment.

[0085] 3. Integrating the reasoning capabilities of large language models: Leveraging the powerful context modeling, long-term dependency capture, and logical reasoning capabilities of large language models (LLM), and fine-tuning them through QLoRA, the model enables deep learning of the battery state evolution process. The model can capture complex temporal dependencies, identify potential abnormal evolution patterns, and generate a predicted sequence of state evolution over a future period.

[0086] To enable those skilled in the art to better understand the technical solution of this application, the implementation process of the thermal runaway prediction method for power batteries of this application will be described in detail below with reference to specific embodiments.

[0087] This embodiment relates to a specific method for predicting thermal runaway of a power battery, such as... Figure 5As shown, it includes the following steps:

[0088] Step S1: Perform reversible normalization on the power battery voltage and temperature time series, divide it into several time segments through a sliding window, and embed them into a low-dimensional representation;

[0089] Step S2: Reprogram the initial temporal embedding segments using text prototype vectors and a multi-head attention module to obtain new temporal segment representations;

[0090] Step S3: Integrate the knowledge hint structure into the new time segment representation. Static hints are concatenated with time segments as prefixes, while dynamic hints are aligned with different time segment points through position embedding, ultimately resulting in a concatenated sequence.

[0091] Step S4: Input the spliced ​​sequence into the fine-tuning language, rely on the prompt context and time segment expression to complete the understanding of the dynamic changes of voltage and temperature time series, and output the prediction token;

[0092] Step S5: Using a two-layer fully connected structure as a predictor, each prediction token is independently projected onto the time segment to be predicted to obtain the actual long-term predicted values ​​of the voltage and temperature of the power battery.

[0093] The overall principle of this technical solution lies in transforming the voltage and temperature time-series data of the power battery into a semantic form that can be understood by a large language model through data reprogramming. Then, cross-modal fusion technology is used to integrate domain knowledge and time-series information. Finally, a finely tuned GPT-2 model or its alternative is used as the core to achieve accurate prediction of battery thermal runaway. This solution not only overcomes the reliance on a single sensor in traditional methods but also enhances the model's generalization ability and understanding of complex time-series patterns by introducing knowledge hints and semantic reconstruction. This allows the early warning system to identify the risk of thermal runaway earlier and more accurately, thus providing strong support for the safety management of power batteries. In practical applications, the QLoRA-optimized model can better adapt to battery state prediction in different scenarios, reducing false alarm rates and improving early warning efficiency. In subsequent implementations, the setting of the first inclined surface can be flexibly adjusted according to the needs of specific application scenarios. Whether using a plane, curved surface, or other transitional form, precise control of material pressure can be achieved, ensuring the quality and performance of the product. Based on this, by moving the upper mold plate horizontally, a gradual and smooth transition in pressure distribution is further achieved, enhancing the practicality and adaptability of the solution.

[0094] This application also provides a thermal runaway prediction device for power batteries. It should be noted that the thermal runaway prediction device for power batteries in this application can be used to execute the thermal runaway prediction method for power batteries provided in this application. This device is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0095] The following describes the thermal runaway prediction device for power batteries provided in the embodiments of this application.

[0096] Figure 6 This is a schematic diagram of a power battery thermal runaway prediction device according to an embodiment of this application. Figure 6 As shown, the device includes: a first acquisition unit 10, used to acquire time series of multiple operating parameters of the power battery within a first historical time period, and determine multiple embedding matrices based on the time series. The multiple operating parameters include the voltage and temperature of the power battery, with one operating parameter corresponding to one time series and one embedding matrix corresponding to a time segment in one time series; a second acquisition unit 20, used to acquire a static text matrix and a dynamic text matrix of the power battery. The static text matrix includes text information about the operating status and operating parameters of the power battery, and the dynamic text matrix includes text information about the start and end times of different time segments in the time series; and a prediction unit 30, used to obtain a fused data matrix based on the embedding matrix, the static text matrix, and the dynamic text matrix, and input the fused data matrix into a target large language model to predict the predicted values ​​of the operating parameters of the power battery in the future time period, so as to predict the thermal runaway state of the power battery. The predicted operating parameter values ​​include the predicted voltage and temperature values ​​of the power battery.

[0097] This application provides a power battery thermal runaway prediction device, including a first acquisition unit for acquiring time series of multiple operating parameters of the power battery within a first historical time period, and determining multiple embedding matrices based on the time series. The multiple operating parameters include the voltage and temperature of the power battery, ensuring the capture of key information from historical data. A second acquisition unit is used to acquire a static text matrix and a dynamic text matrix of the power battery. The static text matrix includes textual information about the operating state and operating parameters of the power battery, while the dynamic text matrix includes textual information about the start and end times of different time segments in the time series, enhancing the model's understanding of the data through textual descriptions. A prediction unit is used to obtain a fused data matrix based on the embedding matrix, the static text matrix, and the dynamic text matrix, and input the fused data matrix into a target large language model to predict the predicted values ​​of the operating parameters of the power battery within a future time period, thereby predicting the thermal runaway state of the power battery. The predicted operating parameter values ​​include the predicted voltage and temperature values ​​of the power battery, achieving early identification and warning of battery thermal runaway risks. This device integrates historical operating parameters with textual information using deep learning technology, significantly improving the accuracy and timeliness of predictions. It solves the problems of delayed response, high false alarm rate, and weak generalization ability of existing early warning systems to battery thermal runaway events, effectively enhancing the safety and reliability of power battery systems.

[0098] In some embodiments, the first acquisition unit includes a first processing module, a second processing module, and a third processing module. The first processing module is used to reversibly normalize the time series to obtain a normalized time series. The second processing module is used to divide the normalized time series into multiple time segments based on a sliding window. The third processing module is used to input all the time segments into a linear layer for dimensionality transformation to obtain the embedding matrix corresponding to each time segment, wherein each embedding matrix includes the data content of a corresponding time segment. This enables the prediction of future changes in the voltage and temperature of the power battery, providing a crucial information foundation for early warning of thermal runaway, thereby improving the accuracy and timeliness of the early warning.

[0099] In some embodiments, the above-described apparatus further includes a first determining module, a second determining module, and a third determining module. The first determining module is used to determine a target difference between the length of the time series and a preset time length before dividing the normalized time series into multiple time segments based on a sliding window. The preset time length is the length of a preset time segment. The second determining module is used to determine a target ratio between the target difference and the window sliding step size. The third determining module is used to determine the number of time segments by the sum of the target ratio and a preset value, where the preset value is 2. The number of time segments can be optimized by dynamically adjusting the window sliding step size or the preset value to adapt to more complex data scenarios and improve the adaptability and flexibility of the model.

[0100] In some embodiments, the prediction unit includes a fourth determining module, a fourth processing module, and a fifth processing module. The fourth determining module is used to obtain text prototype vectors from a semantic subset of the word embedding matrix of the pre-trained language model, and determine a reprogramming matrix based on the text prototype vectors and the embedding matrix. The fourth processing module is used to obtain a fusion input matrix based on the reprogramming matrix and the static text matrix, the fusion input matrix including all data information of the embedding matrix and the static text matrix. The fifth processing module is used to sum the fusion input matrix and the dynamic text matrix to obtain the fusion data matrix, wherein one fusion data matrix corresponds to one time segment. This not only improves prediction accuracy but also enhances the model's generalization ability, effectively addressing the complexity and uncertainty of battery operating states.

[0101] In some embodiments, the fourth determining module includes a determining submodule and a first processing submodule. The determining submodule is used to determine a vector space based on the aforementioned text prototype vector, wherein the vector space at least represents the vocabulary size of the pre-trained language model and the hidden dimension of the backbone model of the pre-trained language model. The first processing submodule is used to unify the dimensions of the aforementioned embedding matrix and the aforementioned vector space using a multi-head cross-attention mechanism, and to perform a weighted combination of the semantics of the aforementioned embedding matrix and the aforementioned vector space to obtain the aforementioned reprogramming matrix. This flexibility is particularly important in practical deployment, ensuring that the early warning system maintains high efficiency and accuracy under different conditions.

[0102] In some embodiments, the static text matrix further includes cross-mapping information of the operating parameters of the power battery. This cross-mapping information includes mapping relationships between different operating parameters. The fourth processing module includes a second processing submodule, used to embed and concatenate the static text matrix corresponding to one time segment with multiple corresponding reprogramming matrices to obtain the fused input matrix. The multiple reprogramming matrices are reprogramming matrices corresponding to different operating parameters, and one time segment corresponds to multiple reprogramming matrices. This enables accurate early warning of thermal runaway in power batteries, demonstrating the flexibility and scalability of this technical solution.

[0103] In some embodiments, the apparatus further includes an acquisition module and a training module. The acquisition module is used to acquire an initial large language model and multiple sets of training data for the initial large language model before inputting the fused data matrix into the target large language model to predict the predicted operating parameters of the power battery in the future time period. Each set of training data includes a historical fused data matrix and the corresponding historical predicted operating parameters acquired in a second historical time period, wherein the second historical time period is earlier than the first historical time period. The training module is used to train and adjust the initial large language model using a quantized low-rank optimization method based on a loss function and the multiple sets of training data to obtain the target large language model, which includes two fully connected layers. This enables the model to learn the potential evolution patterns of the power battery voltage and temperature from historical data, thereby improving the accuracy and reliability of predicting the future operating state of the battery.

[0104] The aforementioned thermal runaway prediction device for power batteries includes a processor and a memory. The first acquisition unit and other components are stored as program units in the memory, and the processor executes these program units to achieve the corresponding functions. All of the above modules are located in the same processor; alternatively, the modules may be located in different processors in any combination.

[0105] The processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and adjusting kernel parameters can address the problems of high false alarm and false negative rates, poor real-time performance, and low accuracy in existing thermal runaway early warning methods for power batteries.

[0106] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0107] This invention provides a computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device containing the computer-readable storage medium to perform the thermal runaway prediction method for the power battery.

[0108] This invention provides a processor for running a program, wherein the program executes the thermal runaway prediction method for a power battery.

[0109] This invention provides a device including a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it performs at least the following steps:

[0110] Step S201: Obtain the time series of multiple operating parameters of the power battery within the first historical time period, and determine multiple embedding matrices based on the time series. The multiple operating parameters include the voltage and temperature of the power battery. One operating parameter corresponds to one time series, and one embedding matrix corresponds to a time segment in one time series.

[0111] Step S202: Obtain the static text matrix and dynamic text matrix of the power battery. The static text matrix includes text information of the operating status and operating parameters of the power battery, and the dynamic text matrix includes text information of the start and end times of different time segments in the time series.

[0112] Step S203: Obtain a fused data matrix based on the above-mentioned embedding matrix, the above-mentioned static text matrix and the above-mentioned dynamic text matrix, and input the above-mentioned fused data matrix into the target large language model to predict the predicted values ​​of the operating parameters of the power battery in the future time period, so as to predict the thermal runaway state of the power battery. The predicted values ​​of the operating parameters include the predicted values ​​of the voltage and temperature of the power battery.

[0113] The devices mentioned in this article can be servers, PCs, tablets, mobile phones, etc.

[0114] This application also provides a computer program product, which, when executed on a data processing device, is suitable for executing an initialization program having at least the following method steps:

[0115] Step S201: Obtain the time series of multiple operating parameters of the power battery within the first historical time period, and determine multiple embedding matrices based on the time series. The multiple operating parameters include the voltage and temperature of the power battery. One operating parameter corresponds to one time series, and one embedding matrix corresponds to a time segment in one time series.

[0116] Step S202: Obtain the static text matrix and dynamic text matrix of the power battery. The static text matrix includes text information of the operating status and operating parameters of the power battery, and the dynamic text matrix includes text information of the start and end times of different time segments in the time series.

[0117] Step S203: Obtain a fused data matrix based on the above-mentioned embedding matrix, the above-mentioned static text matrix and the above-mentioned dynamic text matrix, and input the above-mentioned fused data matrix into the target large language model to predict the predicted values ​​of the operating parameters of the power battery in the future time period, so as to predict the thermal runaway state of the power battery. The predicted values ​​of the operating parameters include the predicted values ​​of the voltage and temperature of the power battery.

[0118] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0119] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0120] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0121] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0122] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0123] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0124] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0125] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, 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 technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0126] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0127] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0128] As can be seen from the above description, the embodiments of this application achieve the following technical effects:

[0129] 1) The above-mentioned thermal runaway prediction method for power batteries in this application first obtains the time series of multiple operating parameters of the power battery within a first historical time period, and determines multiple embedding matrices based on the time series. The multiple operating parameters include the voltage and temperature of the power battery. Each operating parameter corresponds to a time series, and each embedding matrix corresponds to a time segment in a time series. The static text matrix and dynamic text matrix of the power battery are obtained. The static text matrix includes text information of the operating state and operating parameters of the power battery, and the dynamic text matrix includes text information of the start and end times of different time segments in the time series. A fused data matrix is ​​obtained based on the embedding matrix, the static text matrix, and the dynamic text matrix, and the fused data matrix is ​​input into the target large language model to predict the predicted values ​​of the operating parameters of the power battery in the future time period, so as to predict the thermal runaway state of the power battery. The predicted values ​​of the operating parameters include the predicted values ​​of the voltage and temperature of the power battery. This method achieves comprehensive perception and risk assessment of battery operating status through steps such as time-series data reprogramming, cross-modal information fusion, and intelligent analysis of large language models. It effectively improves the accuracy and timeliness of thermal runaway early warning, reduces the probability of false alarms and missed alarms, and solves the problems of high false alarm and missed alarm rates, poor real-time performance, and poor accuracy of existing thermal runaway early warning methods for power batteries.

[0130] 2) This application provides a power battery thermal runaway prediction device, including a first acquisition unit for acquiring time series of multiple operating parameters of the power battery within a first historical time period, and determining multiple embedding matrices based on the time series. The multiple operating parameters include the voltage and temperature of the power battery, ensuring the capture of key information from historical data. A second acquisition unit is used to acquire a static text matrix and a dynamic text matrix of the power battery. The static text matrix includes text information about the operating status and operating parameters of the power battery, and the dynamic text matrix includes text information about the start and end times of different time segments in the time series, enhancing the model's understanding of the data through textual descriptions. A prediction unit is used to obtain a fused data matrix based on the embedding matrix, the static text matrix, and the dynamic text matrix, and input the fused data matrix into a target large language model to predict the predicted values ​​of the operating parameters of the power battery within a future time period, so as to predict the thermal runaway state of the power battery. The predicted operating parameters include the predicted voltage and temperature values ​​of the power battery, realizing early identification and warning of battery thermal runaway risk. This device integrates historical operating parameters with textual information using deep learning technology, significantly improving the accuracy and timeliness of predictions. It solves the problems of delayed response, high false alarm rate, and weak generalization ability of existing early warning systems to battery thermal runaway events, effectively enhancing the safety and reliability of power battery systems.

[0131] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for predicting thermal runaway of a power battery, characterized in that, include: The time series of multiple operating parameters of the power battery within a first historical time period are obtained, and multiple embedding matrices are determined based on the time series. The multiple operating parameters include the voltage and temperature of the power battery. Each operating parameter corresponds to a time series, and each embedding matrix corresponds to a time segment in a time series. Obtain the static text matrix and dynamic text matrix of the power battery. The static text matrix includes the operating status of the power battery, text information of operating parameters, cross-mapping information of operating parameters of the power battery, basic background information of the power battery, task instructions, and statistical characteristics of the power battery. The dynamic text matrix includes the text information of the start and end times of different time segments in the time series. The cross-mapping information includes the mapping relationship between different operating parameters. A fusion data matrix is ​​obtained based on the embedding matrix, the static text matrix, and the dynamic text matrix. The fusion data matrix is ​​then input into the target large language model to predict the predicted operating parameters of the power battery in the future time period, so as to predict the thermal runaway state of the power battery. The predicted operating parameters include the predicted voltage and temperature of the power battery. A fused data matrix is ​​obtained based on the embedding matrix, the static text matrix, and the dynamic text matrix, including: Obtain text prototype vectors from a semantic subset of the word embedding matrix of a pre-trained language model; A vector space is determined based on the text prototype vector, and the vector space at least represents the vocabulary size of the pre-trained language model and the hidden dimension of the backbone model of the pre-trained language model. A multi-head cross-attention mechanism is used to unify the dimensions of the embedding matrix and the vector space, and the semantics of the embedding matrix and the vector space are weighted and combined to obtain the reprogramming matrix; The static text matrix corresponding to a time segment is used as a prefix and embedded and concatenated with multiple corresponding reprogramming matrices to obtain a fusion input matrix. The multiple reprogramming matrices are reprogramming matrices corresponding to different running parameters, and a time segment corresponds to multiple reprogramming matrices. The fusion input matrix includes all data information of the embedding matrix and the static text matrix. The fused input matrix and the dynamic text matrix are summed to obtain the fused data matrix, wherein one fused data matrix corresponds to one time segment.

2. The method according to claim 1, characterized in that, Multiple embedding matrices are determined based on the time series, including: The time series is reversibly normalized to obtain a normalized time series; The normalized time series is divided into multiple time segments based on a sliding window. All the time segments are input into a linear layer for dimensionality transformation to obtain the embedding matrix corresponding to each time segment, wherein each embedding matrix includes the data content of a corresponding time segment.

3. The method according to claim 2, characterized in that, Before dividing the normalized time series into multiple time segments based on a sliding window, the method further includes: The difference between the length of the time series and the preset time length is determined as the target difference, where the preset time length is the preset length of the time segment; The ratio of the target difference to the window sliding step size is determined as the target ratio. The sum of the target ratio and a preset value is used to determine the number of time segments, where the preset value is 2.

4. The method according to any one of claims 1 to 3, characterized in that, Before inputting the fused data matrix into the target large language model to predict the predicted operating parameters of the power battery for the future time period, the method further includes: An initial large language model and multiple sets of training data for the initial large language model are obtained. Each set of training data includes a historical fusion data matrix and the predicted historical operating parameters corresponding to the historical fusion data matrix, which are obtained within a second historical time period. The second historical time period is earlier than the first historical time period. The initial large language model is trained and adjusted using a quantized low-rank optimization method based on a loss function and the multiple sets of training data to obtain the target large language model, which includes two fully connected layers.

5. A thermal runaway prediction device for a power battery, characterized in that, The device uses the thermal runaway prediction method for power batteries according to any one of claims 1 to 4 to predict the thermal runaway of the power battery, and the device comprises: The first acquisition unit is used to acquire the time series of multiple operating parameters of the power battery within a first historical time period, and determine multiple embedding matrices based on the time series. The multiple operating parameters include the voltage and temperature of the power battery. One operating parameter corresponds to one time series, and one embedding matrix corresponds to a time segment in one time series. The second acquisition unit is used to acquire the static text matrix and dynamic text matrix of the power battery. The static text matrix includes text information about the operating status and operating parameters of the power battery, and the dynamic text matrix includes text information about the start and end times of different time segments in the time series. The prediction unit is used to obtain a fused data matrix based on the embedding matrix, the static text matrix, and the dynamic text matrix, and input the fused data matrix into the target large language model to predict the predicted values ​​of the operating parameters of the power battery in the future time period, so as to predict the thermal runaway state of the power battery. The predicted operating parameters include the predicted voltage and temperature of the power battery.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device containing the computer-readable storage medium to perform the thermal runaway prediction method for a power battery according to any one of claims 1 to 4.

7. An electronic device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including methods for performing the thermal runaway prediction method for a power battery according to any one of claims 1 to 4.

Images