Battery core temperature acquisition method and device, and storage medium

Abstract

The invention discloses a battery core temperature acquisition method and device, and a storage medium. The method comprises the following steps: acquiring sample time sequence data of a sample battery in charge and discharge cycles under different working conditions; splitting the sample time sequence data into a target number of initial sub-signals by using a splitting model, wherein the initial sub-signals have different bandwidths; determining sub-signals in the initial sub-signals according to the activeness of each initial sub-signal; fusing the sub-signals to obtain fused time sequence data corresponding to the sample time sequence data; training the initial calculation model by using the fused time sequence data to obtain a target calculation model; and in response to the temperature measurement and calculation instruction, inputting application time sequence data of the application battery in the current charge-discharge cycle into the target calculation model, so that the target calculation model calculates the core temperature of the application battery according to the application time sequence data. According to the technical scheme, the core temperature of the battery is accurately obtained, and safe and stable operation of the battery system is guaranteed.

Description

Technical Field

[0001] This application belongs to the field of battery data processing, specifically relating to methods and devices for obtaining battery core temperature, and storage media. Background Technology

[0002] With the rapid development of the new energy industry, lithium-ion batteries, with their advantages of high energy density and recyclability, have been widely used in key fields such as energy storage, electric vehicles, and aerospace. Temperature, as a core indicator affecting the safe and efficient operation of batteries, is directly related to cycle life, reliability, and safety. However, the only directly measurable temperature is the battery surface temperature, which has a significant temperature difference and time delay compared to the core temperature, failing to accurately reflect the internal thermal state of the battery and posing a severe challenge to battery management.

[0003] To obtain the core temperature, the industry has developed a variety of technical solutions, but all of them have obvious limitations: direct measurement methods require the deployment of measurement devices inside the battery, which is costly and can easily damage the battery structure; numerical thermal modeling methods are complex and computationally burdensome; electrochemical impedance spectroscopy analysis methods rely on offline calibration relationships and are easily affected by actual operating conditions; traditional data-driven methods do not require complex modeling, but they have weak noise resistance and the models are difficult to accurately capture the nonlinear and time-varying characteristics of the core temperature, resulting in insufficient estimation accuracy and stability.

[0004] Therefore, how to efficiently, accurately, and in accordance with actual working conditions estimate the core temperature of lithium-ion batteries, and achieve accurate acquisition of the core temperature of the battery to ensure the safe and stable operation of the battery system is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] The purpose of this application is to estimate the core temperature of lithium-ion batteries efficiently, accurately, and in accordance with actual operating conditions, so as to achieve accurate acquisition of the core temperature of the battery and ensure the safe and stable operation of the battery system.

[0006] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0007] According to one aspect of the embodiments of this application, a method for obtaining the core temperature of a battery is provided, the method comprising: Obtain sample time-series data of the sample battery under different operating conditions for charge-discharge cycles; The sample time series data is split into a target number of initial sub-signals using a splitting model, each of which has a different bandwidth; Based on the activity level of each initial sub-signal, a sub-signal is determined from the initial sub-signals; The sub-signals are fused to obtain the fused time series data corresponding to the sample time series data; The initial computational model is trained using the fused time-series data described above to obtain the target computational model; In response to the temperature measurement command, the application timing data of the battery in the current charge-discharge cycle is input into the target calculation model, so that the target calculation model can calculate the core temperature of the battery based on the application timing data.

[0008] According to one aspect of the embodiments of this application, obtaining sample time-series data of a sample battery under different operating conditions for charge-discharge cycles includes: Obtain coarse time-series data of the sample battery under various operating conditions for charge-discharge cycles; For each type of parameter in each coarse time series data, divide it into multiple segments of data to be denoised according to the time series. A multinomial fitting is applied to each time step of the data to be denoised, and the fitted value corresponding to the time step data is used to replace the data at that time step to obtain clean data. The purified data are spliced ​​together according to time sequence to obtain the sample time sequence data.

[0009] According to one aspect of the embodiments of this application, the method further includes: Obtain the value range of various adjustment parameters in the initial splitting model; Take one value from the range of each of the aforementioned adjustment parameters to obtain the initial parameters; Obtain multiple different initial parameters and calculate the splitting error of the initial splitting model under each of the initial parameters; The splitting model is determined based on the splitting error.

[0010] According to one aspect of the embodiments of this application, determining a splitting model based on the splitting error includes: The initial parameter whose splitting error is less than the first threshold is used as the initial target parameter; The adjustment range of the adjustment parameter is determined based on the number of iterations of the initial target parameter; the number of iterations is negatively correlated with the size of the adjustment range. Based on the adjustment range, within the value range of the adjustment parameter, sub-value ranges are determined respectively with each of the initial target parameters as a reference; In each of the sub-value intervals, the initial parameters are redefined so that the sub-value intervals are updated again based on multiple different initial parameters; The process continues until the number of times the initial parameters are selected reaches a first value, or the number of times the splitting error increases reaches a second value, at which point the initial splitting model with the smallest splitting error is selected as the splitting model.

[0011] According to one aspect of the embodiments of this application, an initial computational model is trained using the fused time-series data to obtain a target computational model, including: The fused time series data are divided using time windows to obtain multiple initial sub-data. Each of the initial sub-data is optimized to obtain the target sub-data. The optimization process includes at least one of data feature enhancement and data temporal weakening. Based on the time series of the target sub-data in each of the fused time series data, the initial computational model is trained sequentially using the target sub-data to obtain the target computational model.

[0012] According to one aspect of the embodiments of this application, each of the initial sub-data is optimized to obtain target sub-data, including: For a portion of the time step data in the initial sub-data, position swapping is performed within a set time series range; Randomly select from the time step data during location swapping to obtain at least one winning result; The winning data is replaced with random data to obtain the target sub-data. The random number is obtained based on the initial sub-data in which the time step data is located.

[0013] According to one aspect of the embodiments of this application, optimizing each of the initial sub-data to obtain target sub-data further includes: Obtain the correlation between each type of data and the core temperature in the fused time series data; Data types with a correlation greater than a set correlation threshold are considered strongly correlated data. The target sub-data is obtained by copying and retaining the strongly correlated data in the initial sub-data a set number of times.

[0014] According to one aspect of the embodiments of this application, an initial computational model is trained sequentially using the target sub-data in each of the fused time series data to obtain a target computational model, including: If the target sub-data is the starting data in its respective fused time series data, then the target sub-data is used as the input data; If the target sub-data is not the starting data in its fused time series data, then the reference data is determined based on the target sub-data, and the reference data and the target sub-data are used as input data. The reference data refers to historical sub-data, as well as the predicted and actual temperature values ​​corresponding to the historical sub-data; the historical sub-data and the target sub-data belong to the same fused time series data, and the historical sub-data refers to the target sub-data that has been input into the initial calculation model for prediction; The initial computational model is trained based on the input data to obtain the target computational model.

[0015] According to one aspect of the embodiments of this application, a device for acquiring battery core temperature is provided, including a memory, a processor, and a readable program stored in the memory, wherein the processor executes the readable program to implement the method described in any of the above.

[0016] According to one aspect of the embodiments of this application, a readable storage medium is provided, on which a readable program / instruction is stored, which, when executed by a processor, implements the method described in any one of the above-described embodiments.

[0017] In this application, sample time-series data of a sample battery under different operating conditions for charge-discharge cycles are obtained; the sample time-series data are split into a target number of initial sub-signals using a splitting model, each initial sub-signal having a different bandwidth; a sub-signal is determined in each initial sub-signal based on the activity of each initial sub-signal; the sub-signals are fused to obtain fused time-series data corresponding to the sample time-series data; the initial calculation model is trained using the fused time-series data to obtain a target calculation model; in response to a temperature measurement command, the application time-series data of the applied battery in the current charge-discharge cycle is input into the target calculation model, so that the target calculation model calculates the core temperature of the applied battery based on the application time-series data.

[0018] In this embodiment, by collecting time-series sample data related to core temperature under different operating conditions, the comprehensiveness of model training data is ensured, and the adaptability to operating conditions is improved. Initial sub-signals of different bandwidths are split and filtered according to activity, effectively removing redundant information and noise while retaining key features. The filtered sub-signals are fused to simplify the training dimension, improving the training efficiency and learning accuracy of the initial computational model to obtain the target computational model. Finally, the target computational model processes the application time-series data in real time, achieving rapid and accurate acquisition of battery core temperature without damaging the battery structure, ensuring safe battery operation, and meeting the real-time and reliability requirements for battery core temperature acquisition in practical applications.

[0019] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0022] Figure 1 A schematic diagram of a method for obtaining the core temperature of a battery according to an embodiment of this application is shown.

[0023] Figure 2 A flowchart illustrating the acquisition of sample timing data of a sample battery under different operating conditions during charge-discharge cycles, according to an embodiment of this application, is shown.

[0024] Figure 3 A flowchart illustrating the determination of a splitting model according to an embodiment of this application is shown.

[0025] Figure 4 A flowchart illustrating the determination of a splitting model based on splitting error according to an embodiment of this application is shown.

[0026] Figure 5 A flowchart is shown illustrating a process of training an initial computational model using fused time-series data according to an embodiment of this application to obtain a target computational model.

[0027] Figure 6 A flowchart illustrating the optimization processing of each initial sub-data according to an embodiment of this application to obtain target sub-data is shown.

[0028] Figure 7 A flowchart illustrating the optimization processing of each initial sub-data according to another embodiment of this application to obtain target sub-data is shown.

[0029] Figure 8 The flowchart illustrates a process according to an embodiment of this application, in which an initial computational model is trained sequentially using the target sub-data in each fused time series data to obtain a target computational model.

[0030] Figure 9 A block diagram of a computer device for performing a method for obtaining battery core temperature according to an embodiment of this application is shown. Detailed Implementation

[0031] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art.

[0032] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0033] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0034] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0035] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0036] Please see Figure 1 , Figure 1 A schematic diagram of a method for obtaining battery core temperature according to an embodiment of this application is shown. This application embodiment provides the execution steps of a method for obtaining battery core temperature, including: Step S110: Obtain sample time-series data of the sample battery under different operating conditions for charge-discharge cycles; Step S120: Use a splitting model to split the sample time series data into a target number of initial sub-signals, each of which has a different bandwidth. Step S130: Determine sub-signals from the initial sub-signals based on the activity level of each initial sub-signal; Step S140: Fuse the sub-signals to obtain the fused time series data corresponding to the sample time series data; Step S150: Train the initial computational model using the fused time series data to obtain the target computational model; Step S160: In response to the temperature calculation command, the application timing data of the battery in the current charge-discharge cycle is input into the target calculation model, so that the target calculation model can calculate the core temperature of the battery based on the application timing data.

[0037] The above six steps are described in detail below.

[0038] In step S110, the sample battery (which may be a lithium battery) is used to provide training data. The operating conditions cover various working conditions encountered in actual battery use, including static conditions with a fixed charge / discharge rate and dynamic conditions with fluctuating charge / discharge power (such as electric vehicle acceleration, sudden power supply from energy storage devices, etc.). A charge / discharge cycle is the complete process of discharging the battery from a fully charged state to the cutoff voltage and then recharging it back to a fully charged state. The sample time-series data consists of continuously recorded battery operating parameters in chronological order, including current, voltage, surface temperature, etc.

[0039] For example, charge-discharge cycles are performed using different charge-discharge rates for various operating conditions. Alternatively, under various operating conditions, charge-discharge cycles are performed at different set temperatures and charge-discharge rates to collect data from the sample batteries, resulting in multiple sample time-series data. In some embodiments, one sample time-series data is generated for each charging cycle, or discharging cycle, or charge-discharge cycle.

[0040] For the sample battery, various working conditions in actual use are simulated, allowing it to complete multiple charge-discharge cycles. At the same time, the above parameters are continuously recorded at fixed time intervals (e.g., 1 second / time) using professional testing equipment to form a complete sample time series dataset, ensuring that the data covers all possible working states of the battery.

[0041] In some embodiments, the sample battery is of the same type as the application battery to be tested.

[0042] In some embodiments, core temperature correlation tests are performed on various types of data, and data types with a core temperature correlation greater than a correlation threshold are taken as target types. Data of the target type in the charge-discharge cycle of the sample battery are collected as sample time-series data.

[0043] First, the sample battery underwent multiple charge-discharge cycles under different operating conditions, and raw parameter data were collected during the charge-discharge process. The data generated from each charge-discharge cycle were then organized in chronological order to form sample time-series data, ensuring that the sample time-series data can effectively characterize the core temperature change pattern and provide reliable input for subsequent model training.

[0044] In step S120, the splitting model refers to an algorithmic model with signal decomposition capabilities, used to decompose sample time-series data. The initial sub-signal refers to the signal segment obtained after splitting, possessing independent frequency characteristics. The frequency coverage range of each initial sub-signal, and the bandwidths of different sub-signals, do not completely overlap, or do not overlap at all. The target number refers to the number of split sub-signals preset or automatically optimized by the algorithm.

[0045] By decomposing the data along the frequency dimension, the relevant signals of the core temperature are separated. The decomposition model decomposes the sample time-series data obtained in step S110 into multiple initial sub-signals according to a preset or automatically optimized target number. Each initial sub-signal corresponds to a different bandwidth (e.g., low frequency corresponds to the influence of ambient temperature, and mid-frequency corresponds to the normal charging and discharging temperature change). Through this decomposition, the complex raw data is transformed into simple signal segments with single frequency characteristics, laying the foundation for subsequent screening of core information and avoiding mutual interference between signals of different frequencies.

[0046] The target number can be preset or obtained from parameters trained in the splitting model.

[0047] In step S130, since all types of data in the sample time series data are related to the core temperature, the activity of the initial sub-signal can be used as a quantitative indicator to measure the correlation between the initial sub-signal and the core temperature (e.g., calculated by dispersive entropy). The higher the value, the stronger the correlation. A sub-signal refers to a highly active signal that is strongly correlated with the core temperature, selected from the initial sub-signals. Based on the activity threshold, initial sub-signals that meet the requirements are selected as sub-signals.

[0048] The activity level of each initial sub-signal is calculated using a preset algorithm (such as the dispersion entropy algorithm). This indicator directly reflects the degree of correlation between the initial sub-signal and the core temperature. Then, an activity threshold is set, and initial sub-signals with activity levels higher than the threshold are identified as sub-signals. Redundant signals with low activity levels and unrelated to the core temperature (such as signals corresponding to equipment measurement noise) are removed. This focuses on key information, reduces the computational burden and interference of subsequent model training, and improves training efficiency.

[0049] In some embodiments, for each initial sub-signal, all data points of the initial sub-signal are divided into 'a' intervals based on their numerical values, and a corresponding interval label is assigned to each data point to discretize the initial sub-signal, where 'a' is a positive integer greater than 1. The numerical discretization process uses an equidistant interval division method, that is, based on the difference between the maximum and minimum values ​​of the initial sub-signal, 'a' consecutive and non-overlapping numerical intervals are obtained, with each numerical interval corresponding to an interval label. For the discretized initial sub-signal, the interval labels of adjacent 'm' data points are combined in the order of consecutive 'm' data points to form all dispersion modes of the initial sub-signal (the total number of dispersion modes is 'a'). m(Number of occurrences); count the number of times each dispersion mode appears in the initial sub-signal; the ratio of the number of occurrences of each dispersion mode to the total number of dispersion modes is the probability corresponding to each dispersion mode, and the sum of all probabilities is 1; the probability p corresponding to each dispersion mode i (i=1,2,…,a) m ), and all p i The sum is 1.

[0050] The activity of the initial sub-signal is calculated using the following formula:

[0051] In some embodiments, This refers to the activity level of the initial sub-signal. The preferred value for m (embedding dimension) is 2, and the preferred value for the number of categories a is 3. The activity level of each initial sub-signal is obtained, and the initial sub-signal with an activity level greater than the activity level threshold is taken as the sub-signal.

[0052] First, equidistant division refers to dividing the initial sub-signal data into 'a' consecutive, non-overlapping intervals based on the difference between the maximum and minimum values, ensuring a uniform and unbiased division standard. The number of intervals is a positive integer greater than 1, which can be flexibly set according to the range of the initial sub-signal data; that is, 'a' is a positive integer greater than 1. Discretization involves converting the continuously changing initial sub-signal data into discrete interval labels, simplifying subsequent computational complexity. A unique identifier (such as an integer from 1 to 'a') is assigned to each interval to characterize the numerical value of the data point; then, based on the specific value of each data point, it is assigned to the corresponding interval and given a label for that interval.

[0053] The process involves obtaining the maximum and minimum values ​​of all data points in the initial sub-signal and calculating their difference. This difference is then divided into a continuous and non-overlapping numerical intervals (i.e., a gears). Finally, each data point is assigned a corresponding gear label based on its interval (e.g., if the value is in interval 3, the label is "3"). This operation transforms continuous time-series data into a discrete label sequence, preserving the numerical distribution characteristics of the data while simplifying subsequent combination calculations and ensuring the processing is objective and unbiased.

[0054] Secondly, the embedding dimension *m* is used to determine the number of adjacent data points of the gear position labels (which can be 2), thus determining the length of the dispersion pattern. The dispersion pattern is a feature sequence formed by concatenating *m* consecutive gear position labels in chronological order (the total number of features is *a*). m (For example, when a=3 and m=2, the total number of modes is 9). The gear labels of m adjacent data points are sequentially extracted and concatenated according to the time sequence of the initial sub-signals.

[0055] For the initial sub-signal that has been discretized, a sliding combination is performed with an embedding dimension of m as the window length, following a time sequence: starting with the gear position label of the first data point of the initial sub-signal, it is concatenated with the gear position labels of the subsequent m-1 data points to form the first dispersion pattern; then the window slides forward one data point, repeating the concatenation operation until all data points of the initial sub-signal are covered. For example, if the gear position label sequence after discretization of the initial sub-signal is "1, 2, 3, 2", when m=2, the resulting dispersion patterns are "1-2", "2-3", and "3-2". This process effectively captures the temporal correlation between the data points of the initial sub-signal, providing core feature units for subsequent activity calculations.

[0056] Next, record the specific number of times each dispersion mode appears in all generated mode sequences. The number of times a certain dispersion mode appears is proportional to the total number of dispersion modes (a m The ratio of the number of occurrences to the total number of modes reflects the frequency of occurrence of that mode in the initial sub-signal. The probability is obtained by dividing the number of occurrences by the total number of modes, and the sum of the probabilities of all dispersive modes is 1.

[0057] Summarize all generated dispersion patterns and count the occurrences of each unique pattern (e.g., pattern "1-2" appears 5 times); then calculate the sum of the occurrences of each pattern and the total number of dispersion patterns (a m The probability p corresponding to this pattern is obtained by comparing the ratio of the individual values ​​(p, p). i (If the total number of patterns is 20, and "1-2" appears 5 times, then p) i =5 / 20=0.25). This operation transforms the occurrence of dispersion patterns into standardized probability values, ensuring the objectivity and comparability of subsequent activity calculations.

[0058] Based on the information, the activity level of the initial sub-signal is calculated using the formula described above.

[0059] Finally, a preset activity threshold is established (set based on historical data or experimental requirements, such as a critical value determined based on core temperature correlation analysis); then the activity DE of each initial sub-signal is determined. k The activity level is compared with an activity threshold; finally, initial sub-signals with activity levels greater than the threshold are identified as sub-signals, while redundant signals with activity levels below the threshold (such as noise sub-signals unrelated to core temperature) are removed. This screening process, based on quantitative indicators, avoids the subjectivity of human judgment, ensures high correlation of sub-signals, and provides accurate materials for subsequent fusion of time-series data and model training.

[0060] This application's embodiments possess significant technical advantages through a closed-loop process of discretization, pattern construction, probability statistics, activity calculation, and threshold screening: The method discretizes the initial sub-signals using an equidistant grading approach, ensuring a unified and objective grading standard, avoiding human bias, and guaranteeing the accuracy of converting continuous data into discrete labels. By combining the grading labels of adjacent m data points to form a dispersion pattern, it effectively captures the temporal correlation features of the initial sub-signals, reflecting dynamic changes related to battery core temperature more effectively than single data point analysis. The probability is calculated based on the frequency of dispersion pattern occurrences, and the activity is then solved using the entropy formula, resulting in rigorous quantification logic. The activity value accurately distinguishes the strength of the correlation between the initial sub-signals and core temperature, making it more scientific than traditional qualitative screening. Setting an activity threshold to screen the initial sub-signals accurately eliminates redundant interference signals, retaining initial sub-signals with high-value core features, reducing the computational burden of subsequent time-series data fusion construction and model training, and improving training efficiency. The overall process is simple to operate, highly repeatable, requires no complex equipment support, and balances computational efficiency with correlation identification accuracy, laying a solid foundation for accurate feature screening in subsequent battery core temperature estimation.

[0061] In some embodiments, the optimal parameters are obtained using the ADMM (Alternating Direction Method of Multipliers) algorithm, thereby determining the target splitting model.

[0062] In step S140, fusion refers to the process of integrating multiple sub-signals through an algorithm, extracting core features, and eliminating redundant information between signals. Fusion time-series data refers to the integrated data that centrally reflects the core temperature correlation patterns. It is important to clarify that one sample time-series data corresponds to one fusion time-series data.

[0063] Key features are integrated to form high-quality training data. Using a suitable fusion algorithm (such as principal component analysis), the multiple sub-signals selected in step S130 are integrated, extracting the core features of each sub-signal, eliminating redundant information and repetitive content between signals, and obtaining fused time-series data. This data retains key features strongly correlated with core temperature, simplifies the data structure, and ensures that each sample time-series data has a unique corresponding fused time-series data, providing efficient and accurate input material for subsequent model training.

[0064] In step S150, the initial computational model refers to a basic model (such as a VMD model) that has not been trained but has the potential to predict temperature; the target computational model refers to a model that has been trained with fused time-series data, has optimized parameters, and can accurately predict core temperature. The process involves inputting the fused time-series data into the initial computational model and iteratively adjusting the parameters through error comparison to improve the model's prediction accuracy.

[0065] The initial computational model is optimized using data to enable it to predict the core temperature. In some embodiments, multiple sets of fused time-series data obtained in step S140 are input into the initial computational model. The initial computational model first outputs the predicted core temperature, compares the calculation error with the actual core temperature, and then iteratively adjusts the model parameters (such as the weights and biases of the neural network) through algorithms such as backpropagation, repeatedly optimizing until the error stabilizes within a preset range; finally, the target computational model is obtained, which has the ability to accurately predict the battery core temperature using input data.

[0066] In some embodiments, cyclic data of charge-discharge cycles of sample batteries under different operating conditions are collected. The cyclic data is divided into sample time-series data and validation data according to a set ratio. The target calculation model is validated using the validation data. If the error is less than a set error threshold, the target calculation model is put into use. If the error threshold is greater than or equal to the set error threshold, the target calculation model is redefined.

[0067] The set ratio is a pre-defined allocation ratio between the two types of data, determined based on the total amount of data and model requirements; the sample time series data is the core data used to train the initial computational model and needs to be of sufficient scale to support model learning; the validation data is a dataset independent of the sample time series data, used to test the model's generalization ability and avoid model overfitting.

[0068] For example, the total amount of collected cyclic data (e.g., 410 cycles) is separated into two groups using a random partitioning algorithm according to a set ratio (usually 7:3 or 8:2, and 9:1 when the data volume is small). During partitioning, the operating condition distribution and battery health status distribution of the two types of data are consistent to avoid data bias. For example, 1000 sets of cyclic data are partitioned in a 7:3 ratio to obtain 700 sets of sample time-series data (for early model training) and 300 sets of validation data (for subsequent model performance testing). After partitioning, they are stored separately to avoid cross-contamination.

[0069] Validation involves inputting validation data into the target computational model and evaluating its performance by comparing the predicted core temperature with the actual core temperature. Error is the deviation between the core temperature output by the target computational model and the actual core temperature corresponding to the validation data (obtained through capacity calibration cycle measurements). It can be expressed as absolute error, relative error, or root mean square error (RMSE). The error threshold is a qualified limit determined based on actual application requirements (e.g., prediction error of power batteries ≤ ±3%).

[0070] Input 300 sets of validation data into the target computational model according to a preset format. The model outputs the predicted core temperature corresponding to each set of data. Calculate the error between the predicted core temperature and the actual core temperature (e.g., mean absolute percentage error). If the error is less than a set threshold (e.g., 2.33%), the model performance is deemed satisfactory and can be put into practical application. If the error is greater than or equal to the set threshold (e.g., exceeding 5%), return to the data optimization (e.g., reprocessing the cyclic data), model training (e.g., adjusting the network structure), or parameter optimization (e.g., iteratively optimizing and adjusting parameters) steps, and re-execute the training process to obtain a new target computational model until the validation error meets the requirements.

[0071] This application embodiment collects charge-discharge cycle data of sample batteries under different operating conditions, covering complex scenarios commonly encountered in actual use such as temperature, discharge rate, and dynamic load. This ensures the comprehensiveness and representativeness of the data, providing high-fidelity basic data for the verification of the target calculation model and avoiding the problem of insufficient generalization ability of the target calculation model caused by single-condition data. The sample time-series data and verification data are randomly divided according to a set ratio. This ensures that the sample time-series data is large enough to support the initial calculation model's full learning, while the independent verification data objectively tests the predictive performance of the target calculation model, effectively avoiding the risk of overfitting and ensuring the authenticity of the evaluation results. The assessment is based on a set error threshold. The decision to deploy the target computational model clarifies the performance qualification standard for the model. This threshold aligns with industry standards and practical application needs, ensuring that the deployed model possesses sufficient prediction accuracy to accurately output the core temperature of lithium-ion batteries. When the error fails to meet the standard, the target computational model is redefined, forming a closed-loop optimization mechanism of training, verification, and iteration. This further enhances the reliability and adaptability of the target computational model, eliminating the need for complex battery physics modeling, lowering the implementation threshold, and ensuring that the model can handle diverse operating conditions in real-world applications. This significantly improves the engineering practicality and application value of the entire data-driven lithium-ion battery core temperature prediction method.

[0072] In step S160, the applied battery refers to the battery for which core temperature monitoring is actually required (such as batteries used in electric vehicles or energy storage devices). The applied time-series data refers to parameter data recorded over time during the current charge-discharge cycle of the applied battery, consistent with the sample time-series data type. The temperature calculation command refers to the signal that triggers core temperature calculation (such as a real-time monitoring command triggered based on a custom setting or a trigger event). Upon receiving the temperature calculation command, the core temperature calculation process is initiated. Core temperature refers to the true temperature of the central region (or the critical region where the core active material is located) inside the (lithium-ion) battery. It is a core parameter reflecting the internal thermal state of the battery, distinct from the directly measurable surface temperature of the battery.

[0073] When the target computing model receives the temperature measurement command, it immediately responds and starts the calculation process. The application time-series data (consistent with the data type of the sample time-series data) collected in the current charge-discharge cycle of the application battery is input into the target computing model. The model quickly analyzes the core temperature-related features in the data through the trained parameters and logic, and outputs the core temperature of the application battery in real time, realizing non-destructive and efficient core temperature monitoring.

[0074] The application timing data can be the timing data from the start of the current charge / discharge cycle of the battery to the moment the temperature measurement command is triggered. In this case, the length of the application timing data is equal to the length of all historical data of the battery in the current charge / discharge cycle.

[0075] Application timing data can also be obtained by tracing back from the temperature measurement command trigger time to obtain battery data within a set duration. In this case, the length of the application timing data can be the same as the initial sub-data below. Alternatively, application timing data can be obtained by continuing to run the application battery for a set duration from the temperature measurement command trigger time. In this case, the length of the application timing data can also be the same as the initial sub-data below. That is, in the latter two cases, the historical timing data is not all the historical data of the battery in the current charge-discharge cycle, but only a small portion of the total historical data of the battery in the current charge-discharge cycle.

[0076] The target computational model analyzes input data and outputs the core temperature of the applied battery through learned mapping relationships. Upon receiving a temperature calculation command, it collects application time-series data of the current discharge cycle of the applied battery. The target computational model quickly analyzes the data characteristics based on the application time-series data and accurately outputs the core temperature of the applied battery based on the patterns learned during training.

[0077] In some embodiments, a target calculation model corresponding to the applied battery is determined based on the type or model of the applied battery. The core temperature of the applied battery is then calculated using the target calculation model.

[0078] This application's embodiments possess significant technical advantages through a closed-loop process of data acquisition, splitting, filtering, fusion, training, and calculation. This method eliminates the need for implanting sensors inside the battery, employing a non-destructive approach to collect externally measurable data. This avoids damage to the battery structure, reduces hardware costs and implementation difficulty, and overcomes the drawbacks of traditional invasive temperature measurement. By acquiring time-series data covering different operating conditions, combined with the steps of decomposing initial sub-signals of different bandwidths using a splitting model and filtering highly correlated sub-signals based on activity, measurement noise and redundant information are effectively eliminated, focusing on key core temperature characteristics. Furthermore, the target calculation model is trained using fused time-series data, giving the model strong generalization capabilities, adapting to various complex usage scenarios, and improving the accuracy of core temperature prediction. Once the target computational model is trained, it only needs to input the real-time time-series data of the applied battery (application time-series data) to quickly respond and output the core temperature, meeting the real-time monitoring requirements. It can promptly capture temperature anomalies, reduce battery chemical aging and capacity loss, prevent thermal runaway and fire accidents, and extend battery life. At the same time, the method is simple to operate, does not rely on complex thermal models or laboratory conditions, and is adaptable to different types of lithium-ion batteries, providing key technical support for battery safety management in energy storage, electric vehicles and other fields.

[0079] Please see Figure 2 , Figure 2 A flowchart illustrating the acquisition of sample time-series data of a sample battery under different operating conditions during charge-discharge cycles, according to an embodiment of this application, is shown. This application embodiment provides step S110 for acquiring sample time-series data of a sample battery under different operating conditions during charge-discharge cycles, including: Step S111: Obtain coarse time-series data of the sample battery under various operating conditions for charge-discharge cycles; Step S112: For each type of parameter in each coarse time series data, divide it into multiple segments of data to be denoised according to the time series. Step S113: Use a multinomial fit on the data at each time step in the data to be denoised, and replace the data at that time step with the fitted value corresponding to the data at that time step to obtain clean data. Step S114: The pure data are spliced ​​together according to the time sequence to obtain the sample time sequence data.

[0080] The above four steps are described in detail below.

[0081] In step S111, coarse time series data refers to the directly collected, unprocessed raw data, which includes noise such as equipment measurement errors and external interference, and serves as the basic material for subsequent data processing.

[0082] Under different actual operating conditions (such as fast charging, slow charging, high and low temperature environments, etc.), the sample battery was allowed to complete multiple full charge-discharge cycles. Parameters related to the core temperature (such as current, voltage, surface temperature, etc.) were recorded in real time in chronological order using professional testing equipment, forming coarse time-series data. This data directly reflects the operating status of the sample battery under different conditions. Although it contains noise, it provides comprehensive and accurate raw material for subsequent data processing. Furthermore, one set of coarse time-series data corresponds to each charge-discharge cycle, ensuring a one-to-one correspondence between the data and the cycle process.

[0083] In step S112, each type of parameter refers to different types of physical quantities (such as current, voltage, etc.) contained in the coarse time series data, and each type of parameter corresponds to an independent time series curve. A single coarse time series data set (corresponding to one charge-discharge cycle) is split into multiple non-overlapping segments to obtain multiple data sets to be denoised, avoiding data overlap or omission. The data sets to be denoised refer to the segmented data obtained after dividing the data into different types of parameters; these have not undergone denoising processing and need to be further processed by polynomial fitting to remove noise.

[0084] For each coarse time-series data set (corresponding to one charge-discharge cycle), each type of parameter (such as current, voltage, etc.) is divided into multiple non-overlapping segments to be denoised, arranged chronologically. For example, one hour of coarse time-series data (3600 time steps) from one charge-discharge cycle is divided into 180 segments of denoised data, each segment consisting of 20 time steps. This division method ensures the independence of each segment while preserving the temporal characteristics of the data, avoiding the loss of local features due to denoising the entire data segment, and laying the foundation for subsequent accurate denoising.

[0085] In step S113, each data point to be denoised includes multiple time-step data. A time-step data point is the basic unit of the sample time-series data, referring to a complete set of battery parameters corresponding to a single time point. Polynomial fitting refers to constructing a low-order polynomial curve that closely matches the changing trends of the data at each time step in the data to be denoised. The theoretical fitting value for each type of data at each time step is obtained through mathematical calculation. The fitted value refers to the theoretical optimal value corresponding to the time-step data after polynomial fitting, after removing noise interference. Clean data refers to the noise-free, stable data obtained after replacing the original time-step data with the fitted value.

[0086] For each segment of data to be denoised, low-order polynomials (such as 1st to 3rd order polynomials) are used to fit the data of each type at multiple time steps. The linear least squares method is used to make the polynomial curves fit the variation patterns of each type of data in the data to be denoised as closely as possible. The fitted values ​​for each type of data at each time step are calculated, and these fitted values ​​replace the original time step data in the data to be denoised. Compared with traditional median filtering and mean filtering, this method can effectively remove measurement noise and external interference while fully preserving the key features of the data (such as voltage spikes and current peaks), ultimately obtaining clean data free from noise interference.

[0087] In step S114, the temporal sequence between the clean data points is the same as the temporal sequence between the data points to be denoised to which each clean data point belongs. That is, the clean data points are concatenated according to the temporal sequence between the data points to be denoised to which each clean data point belongs. The resulting clean data points are concatenated according to their chronological order in the original coarse time series data, ensuring that the concatenated time series data completely corresponds to the charge-discharge cycle process of the sample battery (one charge-discharge cycle corresponds to one sample time series data point). For example, 180 segments of clean data are concatenated in the order of time steps 1-20, 21-40, etc., to restore a complete 1-hour time series data point. This step ensures the continuity and integrity of the data, ultimately yielding clean, stable, and structured sample time series data, providing high-quality input for subsequent steps such as model splitting and model training.

[0088] In some embodiments, each coarse time series data is processed in the manner described in steps S111 to S114 to obtain the sample time series data corresponding to each coarse time series data, that is, each coarse time series data corresponds to one sample time series data.

[0089] This application's embodiments first acquire coarse time-series data of sample batteries under multiple charging and discharging cycles to ensure data coverage of actual battery usage scenarios, providing comprehensive raw materials for initial computational model training. The coarse data is then split into non-overlapping data to be denoised through time-series partitioning, achieving structured data processing and avoiding the loss of local features caused by large-scale denoising. Polynomial fitting denoising is employed, replacing the original time-step data with fitted values. This effectively removes measurement noise and external interference while fully preserving key data features (such as voltage spikes and current peaks), solving the problem of traditional denoising methods easily smoothing out useful signals. Finally, the clean data is spliced ​​together in time sequence to form complete and clean sample time-series data, ensuring data continuity and accuracy. The overall process yields high-quality, low-noise, and feature-complete sample time-series data, providing reliable input for subsequent signal decomposition of the split model and training of the target computational model, avoiding model accuracy degradation caused by invalid data interference.

[0090] Please see Figure 3 , Figure 3A flowchart illustrating the determination of a splitting model according to an embodiment of this application is shown. Embodiments of this application provide steps for determining a splitting model, including: Step S201: Obtain the value range of various adjustment parameters in the initial splitting model; Step S202: Take a value from the range of each adjustment parameter to obtain the initial parameter; Step S203: Obtain multiple different initial parameters and calculate the splitting error of the initial splitting model under each initial parameter; Step S204: Determine the splitting model based on the splitting error.

[0091] The four steps described above are described in detail below.

[0092] In step S201, the initial splitting model (which can be a VMD (Variational Mode Decomposition) model) refers to an algorithm model with basic signal decomposition capabilities but whose parameters are not optimized; it forms the basis for subsequent parameter optimization. Adjustment parameters are key parameters that determine the decomposition effect of the splitting model. Different splitting algorithms have different types of adjustment parameters (e.g., the number of modes K and penalty coefficient α in the VMD algorithm). The legal and valid value range of each adjustment parameter is determined to avoid decomposition failure due to parameters exceeding the range. Reasonable value ranges for each adjustment parameter are determined through algorithm theoretical characteristic analysis, preliminary experimental testing, or industry experience.

[0093] First, clearly define the initial decomposition model used for signal decomposition (such as the VMD algorithm) and identify the core adjustment parameters of the model (such as the number of modes K and the penalty coefficient α). Then, combined with the theoretical requirements of the algorithm (such as the number of modes should not be too small to cause insufficient signal decomposition, and too many to cause redundancy), determine the value range of each adjustment parameter (such as the value range of K being 3-10, and the value range of α being 100-5000). This provides a clear basis for the selection of subsequent initial parameters and avoids low decomposition efficiency or poor results caused by chaotic parameter values.

[0094] In step S202, a specific value is randomly selected or selected at a specific interval from the range of values ​​of each adjustment parameter; a complete set of parameters formed by combining the selected values ​​of each adjustment parameter is used to test the splitting effect of the initial splitting model.

[0095] For each adjustment parameter and its value range determined in step S201, a specific value is extracted from the value range of each parameter using random selection or equal-interval selection. These values ​​are then combined to form a set of initial parameters. For example, if the adjustment parameters are K (target quantity) (value range 3-10) and α (value range 100-5000), selecting K=6 and α=2000 forms a set of initial parameters. Repeating this operation can generate multiple sets of different initial parameters, ensuring that the impact of different parameter combinations on the splitting effect can be comprehensively tested subsequently.

[0096] In step S203, multiple sets of independent parameter combinations are obtained, each set of parameters corresponding to a splitting mode. Each set of initial parameters is input into the initial splitting model to decompose the sample time series data. The decomposed sub-signals are then reconstructed into complete signals. The deviation between the reconstructed signal and the standard reconstructed signal is calculated using a preset error formula (such as the root mean square error formula) to obtain the splitting error of that set of parameters.

[0097] For example, firstly, multiple sets of different initial parameters are collected (e.g., 50 independent parameter combinations); then, each set of initial parameters is sequentially input into the initial splitting model, which is used to decompose the sample time series data to obtain multiple initial sub-signals under the corresponding parameters; next, these initial sub-signals are recombined and reconstructed into a complete signal, and the deviation between the reconstructed signal and the standard reconstructed signal (i.e., splitting error) is calculated using the error formula; finally, the splitting error corresponding to each set of initial parameters is obtained. The smaller the splitting error, the better the fit between the initial parameter and the initial splitting model.

[0098] In step S204, the splitting error refers to the deviation value corresponding to each set of initial parameters obtained through calculation, and it serves as the basis for judging the quality of the parameters. By comparing the splitting errors corresponding to all initial parameters, the initial parameter with the smallest splitting error is selected, and the initial splitting model corresponding to this parameter is determined as the final usable splitting model.

[0099] Specifically, firstly, the calculated sets of initial parameters and their corresponding splitting errors are sorted and ranked. Then, based on the error ranking results, the set of initial parameters with the smallest splitting error (e.g., the parameter combination with the smallest root mean square error) is selected. Finally, this optimal set of parameters is incorporated into the initial splitting model. This initial splitting model is the splitting model, which possesses the best signal decomposition capability. It can effectively split the sample time-series data while minimizing signal distortion, providing high-quality decomposition results for subsequent sub-signal selection and fusion. Furthermore, the target quantity can be determined based on the parameters of the splitting model.

[0100] This application's embodiments improve parameter optimization efficiency by first clearly defining the range of adjustment parameters, avoiding invalid tests caused by blind parameter selection. By generating multiple sets of different initial parameters and calculating the splitting error, the adaptability of different parameter combinations can be comprehensively evaluated, avoiding the omission of optimal parameters caused by single-parameter testing. The splitting model is determined based on the principle of minimizing splitting error, ensuring the optimality of the parameters in the final splitting model, minimizing distortion and redundancy in the signal decomposition process, and improving the accuracy and reliability of sub-signal decomposition. This method eliminates the need for manual trial and error adjustment of parameters, reducing human error and technical barriers, and achieving efficient and accurate determination of the splitting model. It provides a high-quality signal decomposition foundation for subsequent sub-signal selection, fusion, and training of time-series sample data, thereby ensuring the overall accuracy and stability of battery core temperature estimation.

[0101] Please see Figure 4 , Figure 4 A flowchart illustrating the determination of a splitting model based on splitting error according to an embodiment of this application is shown. This embodiment provides step S204 of determining a splitting model based on splitting error, including: Step S301: Use the initial parameters whose splitting error is less than the first threshold as the initial target parameters; Step S302: Determine the adjustment range of the adjustment parameter based on the number of iterations of the initial target parameter; the number of iterations is negatively correlated with the size of the adjustment range. Step S303: Based on the adjustment range, within the range of values ​​for the adjustment parameters, determine the sub-ranges based on each initial target parameter. Step S304: Redetermine the initial parameters in each sub-value interval, so as to update the sub-value intervals again according to multiple different initial parameters; Step S305: Continue until the number of times the initial parameters are selected reaches the first value, or the number of times the splitting error increases reaches the second value, and then use the initial splitting model with the smallest splitting error as the splitting model.

[0102] The above five steps are described in detail below.

[0103] In step S301, the first threshold refers to a pre-set error critical value used to distinguish the quality of initial parameters (e.g., set to 0.01). Initial parameters that meet the condition that the splitting error is less than the first threshold are designated as initial target parameters. The initial target parameters refer to the high-quality initial parameters selected by the first threshold, which are the basis for subsequent accurate optimization.

[0104] Specifically, by comparing the splitting errors corresponding to all initial parameters, initial parameters with splitting errors less than a first threshold are selected and determined as initial target parameters. This selection process eliminates poorly suited initial parameters, focuses on high-quality parameter regions, avoids wasting resources in the invalid parameter range during subsequent optimization, and improves parameter optimization efficiency.

[0105] In step S302, the number of iterations refers to the number of iterations for parameter optimization (e.g., the 1st iteration, the 5th iteration), reflecting the depth of optimization. The adjustment range refers to the search interval for adjusting the parameters. The more iterations, the smaller the adjustment range (e.g., the initial iteration adjustment range is 100-500, and subsequent iterations reduce it to 200-300), realizing the transition from large-scale exploration to small-scale refinement.

[0106] Record the number of iterations corresponding to the initial target parameters (1 for the first iteration, and increment the number of iterations by 1 for each update of the sub-value interval); then, based on the negative correlation between the number of iterations and the adjustment range, determine the adjustment range for the current iteration. In the early stages of iteration, fewer iterations and a larger adjustment range can cover more potential high-quality parameters; in the later stages of iteration, more iterations and a smaller adjustment range focus on the discovered high-quality parameter regions for fine-tuning. Through this dynamic adjustment, the coarseness of the initial large-range search and the limitations of the later small-range search are avoided, thus improving the targeting of parameter optimization.

[0107] In step S303, the adjustment range refers to the search interval for the adjustment parameter determined in step S302, which serves as the basis for the range of the sub-value interval. The value interval of the adjustment parameter refers to the total legal value range of the adjustment parameter (e.g., α∈[100, 5000]). The initial target parameter refers to the high-quality parameter selected in step S301, which serves as the benchmark point for the sub-value interval.

[0108] Using the initial target parameter as the center or benchmark, a local search interval is defined within the total value interval as a sub-value interval (e.g., if the initial target parameter z=2000 and the adjustment range is ±500, then the sub-value interval is [1500, 2500]).

[0109] Specifically, based on the adjustment range determined in step S302, each initial target parameter is used as a benchmark to divide the corresponding sub-range within the total range of the adjustment parameters. For example, if the total range of the adjustment parameter K is [3, 10], and an initial target parameter K=6 with an adjustment range of ±1, then the sub-range is [5, 7]. If the sub-range exceeds the total range, then the boundary of the total range is used as the limit (e.g., when the adjustment range is ±2, the sub-range is [3, 8]). Through this operation, the search for adjustment parameters is narrowed from the total range to the surrounding local range of each high-quality parameter, laying the foundation for subsequent precise optimization.

[0110] In step S304, the sub-value interval refers to the defined local search interval, which is the range for reselecting initial parameters. Within the sub-value interval, new initial parameters are extracted by random selection or equal-interval selection; the range of the sub-value interval is adjusted according to the splitting error of the newly selected initial parameters (e.g., updating the sub-value interval according to steps S301 to S303). In some embodiments, the first threshold gradually decreases as the number of iterations increases, that is, the first threshold is negatively correlated with the number of iterations.

[0111] As the number of iterations increases, the adjustment range becomes smaller and smaller, so the final result is more targeted, avoiding aimless parameter experiments.

[0112] In some embodiments, within each defined sub-value interval, multiple different sets of initial parameters are redefined in a preset manner (repeatedly taking multiple data points or taking one data point per sub-interval). Then, the sub-value intervals are continuously updated according to steps S301 to S304.

[0113] In some embodiments, if the splitting error is smaller, the adjustment range is reduced; if the splitting error is larger, the original adjustment range is retained, or the adjustment range is appropriately expanded. For example, within each defined sub-range, multiple different initial parameters are redefined according to a preset method (one data point per sub-range); these new initial parameters are input into the initial splitting model, and their respective splitting errors are calculated; the sub-ranges are updated based on the error results—if the new parameter error is less than the original initial target parameter, the sub-range is reduced based on the new parameter; if the error is larger, the sub-range is appropriately expanded or the original range is retained; through this cycle of redefined parameters, error calculation, and interval updates, the optimal parameters are gradually approximated.

[0114] In step S305, the first value refers to the preset maximum number of times the initial parameters can be selected (e.g., 50 times). The first value is a positive integer greater than 1 to avoid infinite iteration. The second value refers to the preset upper limit of the number of times the splitting error continuously increases (e.g., 3 times) to prevent excessive iteration. The second value is a positive integer greater than 1.

[0115] Repeat steps S301 to S304 until either of the following stopping conditions is met: 1) the total number of times initial parameters are selected reaches a first value (e.g., 50 sets of initial parameters have been selected); 2) the number of times the splitting error continuously increases reaches a second value (e.g., the error of 3 consecutive sets of new parameters is greater than the previous set). At this point, stop the iteration and select the model with the smallest splitting error from all the initial splitting models corresponding to the tested initial parameters. This model is then determined as the final splitting model. This model has the optimal signal decomposition capability, can minimize signal distortion, and provides high-quality support for the subsequent decomposition of sample time series data.

[0116] This application first filters initial target parameters using a first threshold to quickly eliminate parameters with poor fit, reducing invalid searches and improving optimization efficiency. Based on the negative correlation between the number of iterations and the adjustment range, a dynamic transition from large-scale exploration to small-scale refinement is achieved, avoiding the coarseness of the initial search while ensuring the accuracy of the later search. Sub-value intervals are defined based on the initial target parameters, focusing on the range around high-quality parameters to further narrow the search radius and reduce computational burden. Through multiple rounds of iteration, parameters are redefined and sub-value intervals are updated to gradually approach the optimal parameters. At the same time, dual stopping conditions are set (the number of selections reaches a first value, and the number of times the error increases reaches a second value) to balance optimization sufficiency and efficiency. Finally, the splitting model with the smallest splitting error is determined to ensure the accuracy and reliability of signal decomposition, providing a high-quality decomposition foundation for subsequent sub-signal selection, fusion, and time-series data training. This ensures the overall accuracy and stability of battery core temperature estimation and avoids decomposition distortion and increased temperature estimation error caused by poor parameter fit.

[0117] Please see Figure 5 , Figure 5 A flowchart illustrating the process of training an initial computational model using fused time-series data to obtain a target computational model, according to an embodiment of this application, is shown. This embodiment provides step S150 of training an initial computational model using fused time-series data to obtain a target computational model, including: Step S151: Divide each fused time series data into multiple initial sub-data using time windows; Step S152: Optimize each initial sub-data to obtain the target sub-data. The optimization process includes at least one of data feature enhancement and data temporal weakening. Step S153: Based on the time series of the target sub-data in each fused time series data, the initial computational model is trained sequentially using the target sub-data to obtain the target computational model.

[0118] The above three steps are described in detail below.

[0119] In step S151, the fused time-series data refers to the comprehensive feature data formed after the sub-signals are fused, which is the object of division. The time window is a preset fixed-length time segment (such as 10 minutes); division is to cut a single continuous long time-series data segment (such as 60 minutes of data in one charge-discharge cycle) into multiple overlapping or continuous short data blocks, i.e., initial sub-data, according to the time window length.

[0120] In some embodiments, each fused time series data is overlapped and cut according to a preset time window length. For example, a 60-minute long data is divided into 10-minute windows to obtain 51 initial sub-data that are superimposed sequentially from the 1st to the 10th minute, the 2nd to the 11th minute, and so on.

[0121] Specifically, for each fused time series dataset, a fixed-length time window is used for partitioning: the time window contains d time steps (e.g., m=15). Starting from the initial time step of the fused time series dataset, d consecutive time steps are extracted as initial sub-data. Subsequent windows slide forward one time step (or a set step size) until the entire fused time series dataset is covered, resulting in multiple initial sub-data. For example, a fused time series dataset with 100 time steps can be partitioned into 86 initial sub-data by using a 15-time-step window. This partitioning method preserves the local temporal correlation of the data while breaking down long time series data into short segments that the model can learn efficiently, laying the foundation for subsequent optimization and training.

[0122] In step S152, the optimization process is a process of improving data quality and adapting it to model training. Data feature enhancement highlights the role of data strongly correlated with core temperature, while data temporal weakening reduces the model's dependence on fixed data temporal sequences to avoid overfitting; the target sub-data is high-quality data that meets the model training requirements after optimization.

[0123] In some embodiments, data feature enhancement can retain only parameters with strong correlations to the battery core temperature based on the correlation between each type of parameter in the initial sub-data and the battery core temperature.

[0124] Data feature enhancement can also be achieved through mathematical methods, calculating the correlation between various parameters in the initial sub-data and the core battery temperature. Strongly correlated data (such as voltage and battery surface temperature) with correlations exceeding a set threshold are selected and replicated a set number of times to reinforce core information.

[0125] The correlation between the various types of parameters in the initial sub-data and the battery core temperature can be calculated in the following way.

[0126]

[0127] Where x represents the core temperature of the battery, and y represents a type of parameter in the initial sub-data. i x represents a parameter of type i in the i-th initial sub-data. i Indicates y i The corresponding core temperature. 'r' refers to the correlation between the two. This method allows for the calculation of the correlation between each type of parameter in the initial subdata and the core temperature.

[0128] In some embodiments, the data in the fused time-series data exists in an ordered temporal sequence, with each time unit called a time step, and each time step corresponding to data, called time step data. Data temporal weakening refers to shuffling the order of the time step data in the initial sub-data.

[0129] In other words, time step data is the basic unit of fused time-series data and the initial sub-data. It refers to a complete set of data corresponding to a single time point (each time step, with the size of the time step being the same as the fixed time interval) when battery operating parameters are continuously recorded at fixed time intervals (such as 1 second, 10 seconds). For example, if the battery status is recorded once every 1 second, the voltage, current, and other data recorded at each moment, such as the 1st second, the 2nd second, etc., are each a time step data. Multiple consecutive time step data are strung together to form a complete initial sub-data. Multiple initial sub-data make up the fused time-series data, which is also the basic unit for subsequent time window division and data optimization processing.

[0130] In some embodiments, data temporal weakening can involve swapping the positions of some time-step data within a set temporal range of the initial sub-data, and then randomly selecting some winning data from the swapped data and replacing it with a random number within the numerical range of the corresponding type parameter in the initial sub-data, thereby ensuring data rationality while breaking fixed temporal dependencies.

[0131] Please see Figure 6 , Figure 6 A flowchart illustrating the optimization processing of each initial sub-data according to an embodiment of this application to obtain target sub-data is shown. This embodiment of the application provides step S152, which includes optimizing each initial sub-data to obtain target sub-data, comprising: Step S401a: For a portion of the time step data in the initial sub-data, position swapping is performed within a set time series range; Step S402a: Randomly select from the time step data of the location exchange to obtain at least one winning data; Step S403a: Replace the winning data with random data to obtain target sub-data. The random number is obtained based on the initial sub-data in which the time step data is located.

[0132] The above three steps are described in detail below.

[0133] In step S401a, the initial sub-data is a short data block (such as a 10-minute sequence of voltage, current, and other parameters) obtained after dividing the data into time windows, which is the original object for optimization processing. Time step data is the basic unit of time series data, referring to a complete set of battery parameters (voltage, current, battery surface temperature, etc.) corresponding to a single time point (such as the 3rd second). Setting the time series range is a pre-defined time interval for partial time step data that can be swapped (such as 20% before and after the time step data or no more than 15% of the data block length), to avoid excessive data disruption that could cause the time series features to fail. Position swapping is the operation of interchange the arrangement order of partial time step data within the set range.

[0134] For each initial sub-data, first randomly select a portion of time step data from that range (such as the 3rd, 5th, and 8th time steps), then determine a preset time series range (such as a data block containing 100 time steps, with the range set before and after a predetermined number of time steps (such as 15)), and swap their positions (such as swapping the data of the 3rd and 9th time steps, or swapping the data of the 5th and 12th time steps), breaking the fixed temporal arrangement of the data and reducing the dependence of the target computational model on a specific order.

[0135] In step S402a, the time step data for position swapping refers to all time step data for which position swapping was completed in step S401a. Random sampling is the process of randomly selecting from the swapped time step data according to a preset proportion or quantity. The selected data are the time step data that are chosen and subsequently require replacement.

[0136] The total amount of time step data that completed the position exchange in step S401a is counted (e.g., a total of 8 time step data were exchanged). The data is then randomly selected according to preset rules (e.g., 2 are selected at a fixed rate or 30% is selected (rounded down or up, or if the rounded down result is zero, then the rounded up result is selected)). The selected time step data is the winning data (e.g., 2 are selected from 8), and the target of the subsequent replacement operation is determined.

[0137] In step S403a, the random data is virtual data used to replace the winning data, and the values ​​of various parameters in it must meet the requirements of rationality. Specifically, the value range of the random data must be limited to the maximum and minimum values ​​of the corresponding data type in the initial sub-data to which the winning data belongs (e.g., if the voltage range of an initial sub-data is 3.0-3.6V, the random number of the voltage data in the replacement data must be within this range), ensuring the logical coherence of the data.

[0138] For each winning data, first extract the numerical range of the corresponding type of data in its initial sub-data (e.g., the current data range is 0.5-2.0A, and the SOC range is 30%-70%). Then, generate corresponding random numbers in the numerical range of each type of data using a random algorithm. Use these random numbers to completely replace the original parameters in the winning data, and finally form the target sub-data with weakened time-series dependency.

[0139] This embodiment first swaps the positions of some time-step data within a set time range, slightly disrupting the strict temporal dependencies of the data to avoid the model overfitting to a fixed temporal pattern. Then, it randomly selects winning data from the swapped data to ensure the randomness and unbiasedness of the optimization process, avoiding feature distortion caused by human intervention. Finally, it generates suitable random data to replace the winning data based on the characteristics of the initial sub-data. This ensures the distributional adaptability of the random data to the original data, further weakens the temporal dependencies, and does not destroy key features related to core temperature. The overall process achieves temporal weakening optimization of the initial sub-data, effectively reducing the risk of overfitting during the training of the target computational model, improving the model's adaptability to battery data under different operating conditions, and enabling the trained target computational model to have stronger generalization and anti-interference capabilities. It can more accurately capture the nonlinear mapping relationship between battery core temperature and input parameters, providing high-quality training data support for accurate estimation of battery core temperature.

[0140] Please see Figure 7 , Figure 7 A flowchart illustrating the optimization processing of each initial sub-data according to another embodiment of this application to obtain target sub-data is shown. This embodiment of the application provides step S152, which involves optimizing each initial sub-data to obtain target sub-data, including: Step S401b: Obtain the correlation between each type of data and the core temperature in the fused time series data; Step S402b: Data of each type with a correlation greater than the set correlation threshold are classified as strongly correlated data; Step S403b: Copy and retain the strongly correlated data in the initial sub-data a set number of times to obtain the target sub-data.

[0141] The above three steps are described in detail below.

[0142] In step S401b, correlation is a quantitative indicator that measures the degree of correlation between a certain type of data and the core temperature. The value ranges from -1 to 1, and the closer the absolute value is to 1, the stronger the correlation.

[0143] A correlation analysis algorithm is invoked to analyze each type of parameter (such as voltage data, current data, and battery surface temperature) in the fused time-series data. The correlation between each parameter and the known actual core temperature of the battery is calculated, providing a basis for subsequent selection of key data. A higher correlation value indicates a stronger correlation.

[0144] In step S402b, the relevant threshold is a predefined standard for judging whether data is critical (e.g., 0.8, which can be adjusted according to actual training needs); strongly correlated data refers to a type of data that is closely related to the core temperature of the battery and has a significant impact on the prediction results of the target calculation model.

[0145] The correlation values ​​of various data obtained in step S401b are compared with the set correlation threshold. Data types with correlation values ​​greater than the threshold are selected as strongly correlated data (for example, the correlation values ​​of voltage and battery surface temperature data are 0.85 and 0.82 respectively, both greater than the threshold of 0.8, which are determined to be strongly correlated data). The core data that the model training needs to focus on is then identified.

[0146] In step S403b, the initial sub-data is a short data block divided by a time window, which is the original object for optimization processing. The set number of replications refers to a predetermined number of replications (e.g., 2 or 3 times, set according to data characteristics and model requirements) used to strengthen the effect of strongly correlated data. Replication retention involves repeatedly adding strongly correlated data to the original initial sub-data without deleting existing data. The target sub-data is the final data that, after being strengthened by replicating strongly correlated data, highlights core information and is suitable for model training.

[0147] For each initial sub-data set, identify the strongly correlated data columns (such as voltage and surface temperature columns). Copy these data columns a predetermined number of times, and then merge the copied data with the original initial sub-data set (for example, if the original data contains three columns: voltage, current, and surface temperature, with voltage and surface temperature being strongly correlated, after two copies, the data columns become voltage, voltage, voltage, SOC, current, surface temperature, surface temperature, and surface temperature). This ultimately forms the target sub-data set with enhanced core information. Each initial sub-data set corresponds to one target sub-data set.

[0148] In some embodiments, only the strongly correlated data in the initial sub-data are retained to obtain the target sub-data corresponding to the initial sub-data.

[0149] In this embodiment, by acquiring the correlation between various types of data and the core temperature of the battery, the core data that plays a crucial role in predicting the core temperature can be accurately identified, avoiding irrelevant or weakly correlated data from interfering with the model's learning direction and solving the problem of insufficient key information in the initial sub-data. Setting relevant thresholds and filtering strongly correlated data provides clear and quantifiable standards for data filtering, ensuring the consistency and rationality of the filtering results and avoiding bias caused by subjective judgment. Copying and retaining strongly correlated data a set number of times can significantly strengthen the weight of core information in the target sub-data, making it easier for the model to capture the core mapping pattern between data and core temperature during training, improving model learning efficiency and training accuracy. The final target sub-data has prominent core features and higher data quality, meeting the training requirements of data-driven models and providing solid data support for the high-precision prediction of the subsequent target calculation model.

[0150] In some embodiments, the initial sub-data may first undergo data feature enhancement, and then data temporal sequence weakening may be performed to obtain the target sub-data. Alternatively, the initial sub-data may first undergo data temporal sequence weakening, and then data feature enhancement may be performed to obtain the target sub-data.

[0151] For example, the correlation between various types of data and the core temperature of the battery is obtained in the fused time series data; data types with a correlation greater than a set correlation threshold are identified as strongly correlated data; the strongly correlated data in each initial sub-data are copied and retained a set number of times to obtain the initial optimized data corresponding to each initial sub-data; then, some time step data in the initial optimized data are swapped within a set time series range; random sampling is performed on the time step data where the position is swapped to obtain multiple winning data; the winning data are replaced with random data, and the random number is obtained according to the initial sub-data in which the time step data is located.

[0152] Alternatively, for a portion of the time-step data in the initial sub-data, position swaps are performed within a set time series range; random sampling is then performed on the position-swapped time-step data to obtain multiple winning data; these winning data are then replaced with random data to obtain the initial optimized data. Then, based on the correlation between each type of data in the initial optimized data and the battery core temperature, data types with a correlation greater than a set correlation threshold are considered strongly correlated data; the strongly correlated data in each initial optimized data set are copied and retained a set number of times to obtain the target sub-data corresponding to the initial optimized data.

[0153] In step S153, the target sub-data are input into the initial computational model one by one in the temporal order, without disrupting the order. The initial computational model refers to an untrained basic model (such as a NARX (Nonlinear Autoregressive With Exogeneous Inputs)-MHA (Multi-Head Attention) neural network) that has basic learning capabilities.

[0154] There's no need to sort the different fused time-series data. Simply input the target sub-data into the initial computational model sequentially according to its original time sequence within each fused time-series dataset. Performing this operation on the target sub-data in each fused time-series dataset allows for full utilization of the temporal nature of the fused time-series data during initial model training, resulting in more accurate predictions from the target computational model.

[0155] In some embodiments, the optimized target sub-data is input into the initial calculation model. The initial calculation model extracts features from the data, calculates the predicted core temperature corresponding to each target sub-data, and then uses the error between the predicted core temperature and the actual core temperature as feedback to continuously adjust the internal parameters of the model until the model prediction accuracy reaches the set requirements, thus forming a stable and usable target calculation model.

[0156] This application employs a time window approach to divide and fuse time-series data, breaking down long-term time-series data into structured initial sub-data. This preserves the local temporal correlation of the data while reducing the learning difficulty of the model and improving training efficiency. By strengthening the core temperature-related features through data feature enhancement and weakening the time sequence of the data to reduce the risk of model overfitting, these two optimization methods jointly enhance the training value of the target sub-data and improve the model's generalization and anti-interference capabilities. The model is trained sequentially according to the time sequence of the target sub-data, which aligns with the dynamic characteristics of battery core temperature changes. It is particularly suitable for models with time-series dependency capture capabilities. By combining historical and current data for training, the model learns the nonlinear mapping relationship between core temperature and input parameters more accurately, reducing prediction errors. The entire process does not require complex manual parameter adjustments, and there is no need to sort the fused time-series data. Only the temporal sequence of the internal target sub-data needs to be guaranteed. The operation is simple and efficient, and the final target calculation model has high prediction accuracy and strong stability, providing reliable model support for accurate estimation of battery core temperature.

[0157] Please see Figure 8 , Figure 8This document illustrates a flowchart illustrating how, according to an embodiment of the present application, an initial computational model is trained sequentially using the target sub-data from each fused time series dataset to obtain a target computational model. The embodiment of the present application provides step S153, which involves training an initial computational model sequentially using the target sub-data from each fused time series dataset to obtain a target computational model. This step includes: Step S501: If the target sub-data is the starting data in its respective fused time series data, then the target sub-data is used as the input data; Step S502: If the target sub-data is not the starting data in its fused time series data, then the reference data is determined based on the target sub-data, and the reference data and the target sub-data are used as input data; wherein, the reference data refers to the historical sub-data, as well as the predicted temperature value and the actual temperature value corresponding to the historical sub-data; the historical sub-data and the target sub-data belong to the same fused time series data, and the historical sub-data refers to the target sub-data that has been input into the initial calculation model for prediction; Step S503: Train the initial computational model based on the input data to obtain the target computational model.

[0158] The above three steps are described in detail below.

[0159] In step S501, the starting data refers to the first target sub-data in the fused time series data, without any preceding data.

[0160] First, it is determined whether the current target sub-data is the starting data (i.e., the first target sub-data to be processed) in its fused time series data. If it is determined to be the starting data, since there is no historical sub-data or related temperature values ​​already input to the model in the fused time series data at this time, the starting target sub-data is directly used as the only input data. At this time, the input data only contains the starting target sub-data. For example, if a fused time series data is divided into 10 target sub-data, and the first target sub-data is the starting data, it is directly used as the input data to the initial calculation model, providing the model with initial training material and laying the foundation for the basic training direction.

[0161] In step S502, non-starting data refers to the target sub-data in the fused time series data other than the first one, for which there is preceding data for reference. Based on the affiliation of the current non-starting target sub-data (its belonging to the fused time series data), the corresponding historical sub-data and related temperature values ​​are selected as reference data. Reference data refers to historical training information associated with the non-starting target sub-data, including historical sub-data already input into the model within the same fused time series data, as well as the corresponding predicted and actual temperature values, used to reflect the temporal correlation of the data.

[0162] At this point, the input data consists of a comprehensive dataset composed of reference data, parameter data, and current non-initial target sub-data.

[0163] For example, the current target sub-data is determined to be non-starting data (such as the 2nd to 10th target sub-data in a fused time series dataset). Then, based on the fused time series dataset to which the target sub-data belongs, reference data is determined—that is, historical sub-data in the fused time series dataset that has been pre-input into the initial calculation model for prediction, as well as the predicted and actual temperature values ​​corresponding to the historical sub-data. Finally, the reference data, parameter data related to the core temperature, and the current non-starting target sub-data are used together as input data. By incorporating historical training information, the model can capture the temporal dependencies of the data, adapt to the dynamic characteristics of battery core temperature changes, and improve training targeting.

[0164] In step S503, the input data determined in step S501 or S502 is input into the initial calculation model, and the model outputs a predicted core temperature value based on the input data. Then, the predicted value is compared with the actual core temperature value corresponding to the target sub-data, and the error (such as root mean square error) is calculated. The model parameters (such as the weights and biases of the neural network) are then adjusted through the backpropagation algorithm to optimize the model's learning of the mapping relationship between the input data and the core temperature. The above process is repeated by inputting all target sub-data in the fusion time series data according to the temporal order of the target sub-data. If the input data corresponding to the target sub-data of a fusion time series data is exhausted, the input data corresponding to the target sub-data of that fusion time series data is used. This process is repeated until any stopping condition is met. In the first embodiment, either the number of times the model outputs the predicted core temperature reaches the fourth value, or the error value is less than the second threshold. In this case, the current initial calculation model is directly used as the target calculation model. In the second embodiment, either the number of predictions reaches the fourth value, or the number of times the error value continuously increases reaches the fifth value. In this case, the initial calculation model with the smallest error in all iterations is selected, its parameters are fixed, and the target calculation model is formed. The initial computational model is trained to obtain the target computational model.

[0165] This embodiment determines the input data based on whether the target sub-data is the starting data. The starting data directly uses itself as input, while non-starting data incorporates reference data and parameter data. This ensures the basic effectiveness of the initial training and fully utilizes time-series correlation information, aligning with the dynamic characteristics of battery core temperature changes and helping the model more accurately capture the nonlinear mapping relationship between data. The reference data focuses on the historical sub-data and corresponding temperature values ​​of the fused time-series data, ensuring the relevance and effectiveness of the time-series correlation and avoiding interference from irrelevant data. Through iterative training by inputting the target sub-data one by one in time series, and adjusting the model parameters with error feedback, the initial calculation model is gradually optimized. The final target calculation model has high prediction accuracy and strong stability, effectively adapting to battery data under different operating conditions, reducing the risk of overfitting, and providing reliable model support for accurate estimation of battery core temperature. At the same time, this process requires no complex manual intervention, is highly efficient, and adapts to the real-time temperature measurement needs in subsequent practical applications.

[0166] In some embodiments of this application, the initial computational model is trained on the target sub-data based on the input data to obtain the target computational model, including: The MHA (Multi-Head Attention) mechanism projects the input data into multiple spaces. The input data in each space is then transformed into query vectors, key vectors, and value vectors. For each space, attention weights are calculated based on the similarity between the query vector and the key vector (by calculating the dot product of the query vector and the key vector and then converting it to attention weights using softmax). The product of the attention weights and the value vectors in each space is used as the output of that space. The outputs from each space are concatenated and transformed using a projection matrix to obtain the input features corresponding to the input data. The input features of each target sub-data are then input into the initial computational model until the initial computational model outputs the fourth number of times it predicts the core temperature, or until the error value of the initial computational model is less than the second threshold. At this point, the initial computational model is used as the target computational model.

[0167] In some embodiments, the input features of each target sub-data are input into the initial calculation model until the number of times the initial calculation model outputs the predicted temperature value reaches the fourth value, or the number of times the error value of the initial calculation model increases reaches the fifth value. The parameters of the initial calculation model with the smallest error are then fixed to form the target calculation model.

[0168] First, in some embodiments, multiple spaces are independent representation subspaces set up for multi-dimensional feature mining, and the number can be preset according to feature extraction requirements (e.g., 3-8). Projection refers to mapping the input data to each independent space according to preset projection rules, thereby realizing the multi-dimensional splitting of features. For the input data in the steps, the core logic of the multi-head attention mechanism is adopted to project it into multiple preset independent spaces. Each space corresponds to a feature extraction dimension. Through this operation, the different feature dimensions of the input data are split into various spaces, avoiding the omission of key features caused by single-dimensional extraction, and providing multi-dimensional support for the subsequent accurate capture of core temperature-related features.

[0169] Secondly, in some embodiments, the query vector is used to query key information in the value vector; the key vector is used to match the query vector to determine the location of key information; the value vector is a vector of the core features of the input data in the storage space; and projection refers to the secondary mapping operation performed on the input data in the space.

[0170] The similarity between the query vector and the key vector is calculated by dot product operation to obtain the original matching score. The original score is then input into the softmax function to be converted into attention weights (the higher the weight, the stronger the correlation between the corresponding feature and the core temperature). Finally, the attention weights of this space are multiplied with the corresponding value vector to strengthen the key features and weaken the secondary features, thus obtaining the output results of each space and achieving precise focus on the core features.

[0171] Then, in some embodiments, the output results of all spaces are spliced ​​together according to the feature dimensions to integrate the key features extracted from multiple dimensions and avoid the limitations of single spatial features. Then, the spliced ​​multi-dimensional features are input into a preset projection matrix to convert them into input features with unified dimensions, ensuring that the features can adapt to the input requirements of the initial calculation model, while centrally retaining the core temperature-related information mined from multiple spaces.

[0172] Finally, in some embodiments, the input features corresponding to each target sub-data are input into the initial computational model. The model outputs a predicted core temperature, and the error value between the predicted core temperature and the actual core temperature is calculated. This process is repeated until any stopping condition is met. In the first embodiment, either the model outputs the predicted core temperature four times, or the error value is less than the second threshold; in this case, the current initial computational model is directly used as the target computational model. In the second embodiment, either the number of predictions reaches the fourth value, or the number of times the error value continuously increases reaches the fifth value; in this case, the initial computational model with the smallest error in all iterations is selected, its parameters are fixed, and the target computational model is formed. Both methods can balance training sufficiency and efficiency, ensuring the prediction accuracy and stability of the target computational model.

[0173] This application's embodiments possess significant technical advantages through a closed-loop process of "multi-spatial projection, vector transformation, attention weighting, feature integration, and iterative optimization": The method projects input data into multiple spaces, achieving multi-dimensional feature extraction and avoiding the omission of key information due to a single dimension; through three vector transformations and attention weight calculations, combined with softmax function normalization, it can accurately focus on key features strongly correlated with core temperature, weakening irrelevant interference and improving feature utilization; after concatenating multiple spatial outputs, a projection matrix transformation is performed to form input features of a unified dimension, adapting to model input requirements while integrating multi-dimensional core information; setting dual stopping conditions (number of predictions / error threshold or number of predictions / number of consecutively increasing errors) avoids infinite iteration and wastes resources while ensuring sufficient model training. Two embodiments are adapted to different accuracy and efficiency requirements, ensuring the prediction accuracy and stability of the target computational model by fixing the minimum error model parameters or directly using a compliant model. The entire process requires no complex manual intervention, efficiently mining core temperature-related features in the input data, enabling the trained target computational model to possess strong generalization and accurate prediction capabilities, providing reliable model support for accurate estimation of battery core temperature.

[0174] Figure 9 A block diagram of a computer device for performing a method for obtaining battery core temperature according to an embodiment of this application is shown.

[0175] It should be noted that, Figure 9 The computer device 800 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0176] like Figure 9 As shown, the computer device 800 includes a central processing unit (CPU) 801, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 802 or programs loaded from storage section 808 into random access memory (RAM). The RAM 803 also stores various programs and data required for device operation. The CPU 801, ROM 802, and RAM 803 are interconnected via a bus 804. An input / output interface 805 (I / O interface) is also connected to the bus 804.

[0177] The following components are connected to the input / output interface 805: an input section 806 including a keyboard, mouse, etc.; an output section 807 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a local area network card, modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the input / output interface 805 as needed. A removable medium 811, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 810 as needed so that computer programs read from it can be installed into the storage section 808 as needed.

[0178] Specifically, according to embodiments of this application, the processes described in the various method flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 809, and / or installed from removable medium 811. When the computer program is executed by central processing unit 801, it performs the various functions defined in the device of this application.

[0179] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor device, apparatus, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution device, apparatus, or apparatus. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such transmitted data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution device, apparatus, or apparatus. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0180] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of devices, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based device that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0181] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0182] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, touch terminal, or network device, etc.) to execute the methods according to the embodiments of this application.

[0183] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0184] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for obtaining the core temperature of a battery, characterized in that, The method includes: Obtain sample time-series data of the sample battery under different operating conditions for charge-discharge cycles; The sample time series data is split into a target number of initial sub-signals using a splitting model, each of which has a different bandwidth; Based on the activity level of each initial sub-signal, a sub-signal is determined from the initial sub-signals; The sub-signals are fused to obtain the fused time series data corresponding to the sample time series data; The initial computational model is trained using the fused time-series data described above to obtain the target computational model; In response to the temperature measurement command, the application timing data of the battery in the current charge-discharge cycle is input into the target calculation model, so that the target calculation model can calculate the core temperature of the battery based on the application timing data.

2. The method according to claim 1, characterized in that, Obtain sample time-series data of the sample battery under different operating conditions for charge-discharge cycles, including: Obtain coarse time-series data of the sample battery under various operating conditions for charge-discharge cycles; For each type of parameter in each coarse time series data, divide it into multiple segments of data to be denoised according to the time series. A multinomial fitting is applied to each time step of the data to be denoised, and the fitted value corresponding to the time step data is used to replace the data at that time step to obtain clean data. The purified data are spliced ​​together according to time sequence to obtain the sample time sequence data.

3. The method according to claim 1, characterized in that, The method further includes: Obtain the value range of various adjustment parameters in the initial splitting model; Take one value from the range of each of the aforementioned adjustment parameters to obtain the initial parameters; Obtain multiple different initial parameters and calculate the splitting error of the initial splitting model under each of the initial parameters; The splitting model is determined based on the splitting error.

4. The method according to claim 3, characterized in that, Determining the splitting model based on the splitting error includes: The initial parameter whose splitting error is less than the first threshold is used as the initial target parameter; The adjustment range of the adjustment parameter is determined based on the number of iterations of the initial target parameter; the number of iterations is negatively correlated with the size of the adjustment range. Based on the adjustment range, within the value range of the adjustment parameter, sub-value ranges are determined respectively with each of the initial target parameters as a reference; In each of the sub-value intervals, the initial parameters are redefined so that the sub-value intervals are updated again based on multiple different initial parameters; The process continues until the number of times the initial parameters are selected reaches a first value, or the number of times the splitting error increases reaches a second value, at which point the initial splitting model with the smallest splitting error is selected as the splitting model.

5. The method according to claim 1, characterized in that, The initial computational model is trained using the fused time-series data described above to obtain the target computational model, including: The fused time series data are divided using time windows to obtain multiple initial sub-data. Each of the initial sub-data is optimized to obtain the target sub-data. The optimization process includes at least one of data feature enhancement and data temporal weakening. Based on the time series of the target sub-data in each of the fused time series data, the initial computational model is trained sequentially using the target sub-data to obtain the target computational model.

6. The method according to claim 5, characterized in that, Each of the initial sub-data is optimized to obtain the target sub-data, including: For a portion of the time step data in the initial sub-data, position swapping is performed within a set time series range; Randomly select from the time step data during location swapping to obtain at least one winning result; The winning data is replaced with random data to obtain the target sub-data. The random number is obtained based on the initial sub-data in which the time step data is located.

7. The method according to claim 5, characterized in that, The initial sub-data is optimized to obtain the target sub-data, and the optimization process is further performed on each of the initial sub-data. Obtain the correlation between each type of data and the core temperature in the fused time series data; Data types with a correlation greater than a set correlation threshold are considered strongly correlated data. The target sub-data is obtained by copying and retaining the strongly correlated data in the initial sub-data a set number of times.

8. The method according to claim 5, characterized in that, Based on the temporal sequence of the target sub-data in each of the fused temporal series data, the initial computational model is trained sequentially using the target sub-data to obtain the target computational model, including: If the target sub-data is the starting data in its respective fused time series data, then the target sub-data is used as the input data; If the target sub-data is not the starting data in its fused time series data, then the reference data is determined based on the target sub-data, and the reference data and the target sub-data are used as input data. The reference data refers to historical sub-data, as well as the predicted and actual temperature values ​​corresponding to the historical sub-data; the historical sub-data and the target sub-data belong to the same fused time series data, and the historical sub-data refers to the target sub-data that has been input into the initial calculation model for prediction; The initial computational model is trained based on the input data to obtain the target computational model.

9. A device for acquiring the core temperature of a battery, comprising a memory, a processor, and a readable program stored in the memory, characterized in that, The processor executes the readable program to implement the control method according to any one of claims 1 to 8.

10. A readable storage medium, characterized in that, It stores a readable program / instruction, which, when executed by a processor, implements the control method according to any one of claims 1 to 8.

Images