SOC fast estimation system and method combined with lightweight neural network

Abstract

The present invention discloses a system and method for rapid SOC estimation in combination with a lightweight neural network, which relates to the technical field related to battery SOC calculation. The system includes: a lightweight sampling data determination module for preprocessing and feature extraction of electrical data sampled from the target battery operation to determine lightweight sampling data; a lightweight estimation model construction module for dividing battery management into charging and discharging conditions and constructing a lightweight estimation model using a distillation training method; and a battery state of charge estimation module for transmitting the lightweight sampling data to the lightweight estimation model, activating the target estimation sub-model based on the battery condition, performing battery state of charge estimation, and determining the SOC data of the target battery. This system solves the technical problems of high SOC calculation complexity and resource consumption, insufficient estimation accuracy, and poor real-time performance in the prior art, achieving the technical effects of improving estimation accuracy, reducing calculation complexity, reducing resource consumption, and enhancing real-time performance.

Description

Technical Field

[0001] The present application relates to the technical field related to battery SOC calculation, and specifically to a system and method for rapid SOC estimation combined with a lightweight neural network. Background Art

[0002] In electric vehicles, energy storage systems, and portable electronic devices, battery condition monitoring and management are critical for ensuring device performance, extending battery life, and ensuring safe operation. Accurately estimating the battery's state of charge (SOC), a key parameter describing the remaining charge in a battery, is crucial for improving energy efficiency, preventing overcharge and overdischarge, and optimizing battery management systems. SOC refers to the ratio of the battery's current stored charge to the charge it can store in a fully charged state. Accurate SOC estimation helps users understand the battery's remaining useful life, guide charging strategies, avoid deep discharge or overcharging, and thus extend battery cycle life. However, while existing SOC estimation methods, such as open-circuit voltage, ampere-hour integration, and Kalman filtering, can handle noise and uncertainty, they are computationally complex and require high hardware resources, making efficient SOC estimation difficult. Furthermore, when processing battery data under dynamic conditions, estimation accuracy and real-time performance are poor, which in turn impacts battery life and increases costs.

[0003] Therefore, in the current related technologies, there are technical problems such as high SOC calculation complexity and resource consumption, insufficient estimation accuracy and poor real-time performance. Summary of the Invention

[0004] This application solves the technical problems of high SOC calculation complexity and resource consumption, insufficient estimation accuracy and poor real-time performance in the existing technology by providing a SOC fast estimation system and method combined with a lightweight neural network, achieving the technical effects of improving estimation accuracy, reducing calculation complexity, reducing resource consumption and enhancing real-time performance.

[0005] The present application provides a SOC rapid estimation system combined with a lightweight neural network, the system including: a lightweight sampling data determination module, used to preprocess and extract features of electrical data sampled from the target battery operation, and determine the lightweight sampling data, wherein the preprocessing includes horizontal processing of the feature attribute dimension and vertical processing of the time dimension; a lightweight estimation model construction module, used to divide battery management into charging conditions and discharging conditions, and construct a lightweight estimation model using a distillation training method, wherein the lightweight estimation model includes a charging estimation sub-model and a discharging estimation sub-model, and uses a distillation loss function as a training constraint; a battery state of charge estimation module, used to transmit the lightweight sampling data to the lightweight estimation model, activate the target estimation sub-model based on the battery condition, perform battery state of charge estimation, and determine the SOC data of the target battery, which includes the state of charge estimation of the single cell and the overall state of charge estimation of the battery pack.

[0006] In a possible implementation, the lightweight sampling data determination module further performs the following processing: setting an SOC correlation coefficient; constructing a preprocessing level based on the SOC correlation coefficient, wherein the preprocessing level includes a first preprocessing level and a second preprocessing level, the first preprocessing level performs electrical data screening, and the second preprocessing level performs vertical and horizontal data processing; based on the first preprocessing level and the second preprocessing level, preprocessing the electrical data.

[0007] In a possible implementation, the lightweight estimation model construction module also performs the following processing: obtaining an estimation sample based on the battery pack and performing lightweight processing, the estimation sample including sample sampling data and sample SOC data; based on the estimation sample, supervised training of the estimation large model; using the output of the estimation large model as soft training data, combined with the estimation sample, performing distillation training processing on the estimation large model to obtain the lightweight estimation model.

[0008] In a possible implementation, the lightweight estimation model construction module also performs the following processing: for the charging condition, determining a first distillation loss function, the first distillation loss function characterizing the logic loss and data loss under the charging condition; for the discharging condition, determining a second distillation loss function, the second distillation loss function characterizing the logic loss and data loss under the discharging condition; based on the first distillation loss function and the second distillation loss function, performing distillation processing on the estimation large model to determine the charging estimation sub-model and the discharging estimation sub-model; an estimation sub-model parallel unit, used to parallelize the charging estimation sub-model and the discharging estimation sub-model as the lightweight estimation model.

[0009] In a possible implementation, the lightweight estimation model construction module also performs the following processing: for the estimation sample, extracting a first estimation sample and determining a first large model sample, wherein the first large model sample is the output of the decision logic data of the estimation large model based on the first estimation sample; based on the first large model sample and the first distillation loss function, determining a first probability distribution, wherein the first probability distribution represents the transmission probability of the decision logic data; and training the charging estimation sub-model once based on the first large model sample, the first probability distribution and the first estimation sample.

[0010] In a possible implementation, the battery state of charge estimation module further performs the following processing: determining the battery operating condition of the target battery, wherein the battery operating condition is a charging condition or a discharging condition; based on the battery operating condition, activating the target estimation sub-model within the lightweight estimation model, inputting the lightweight sampling data, and performing state of charge estimation of the single cell and the state of charge estimation of the battery pack as the SOC data of the target battery.

[0011] In a possible implementation, the battery state of charge estimation module further performs the following processing: identifying the single cell SOC data based on the SOC data of the target battery, performing abnormal state of charge judgment, and determining a first judgment result, wherein the charge and discharge critical value of the single cell is used as the judgment criterion; identifying the battery pack SOC data, performing abnormal charge distribution judgment, and determining a second judgment result, wherein the battery pack SOC data represents the battery pack, and the SOC fluctuation between single cells is used as the judgment criterion; and performing an abnormal operating condition warning of the target battery based on the first judgment result and the second judgment result.

[0012] The present application also provides a method for rapid SOC estimation combined with a lightweight neural network, the method comprising: preprocessing and feature extraction of electrical data sampled from the target battery operation to determine lightweight sampled data, wherein the preprocessing includes horizontal processing of the feature attribute dimension and vertical processing of the time dimension; dividing battery management into charging conditions and discharging conditions, and constructing a lightweight estimation model using a distillation training method, wherein the lightweight estimation model includes a charging estimation sub-model and a discharging estimation sub-model, and a distillation loss function is used as a training constraint; transmitting the lightweight sampled data to the lightweight estimation model, activating the target estimation sub-model based on the battery condition, performing battery state of charge estimation, and determining the SOC data of the target battery, which includes state of charge estimation of a single cell and overall state of charge estimation of the battery pack.

[0013] The proposed SOC rapid estimation system and method in combination with a lightweight neural network in this application includes a lightweight sampling data determination module for preprocessing and feature extraction of the electrical data sampled from the target battery operation to determine the lightweight sampling data; a lightweight estimation model construction module for dividing battery management into charging and discharging conditions, and constructing a lightweight estimation model using distillation training; and a battery state of charge estimation module for transmitting the lightweight sampling data to the lightweight estimation model, activating the target estimation sub-model based on the battery condition, performing battery state of charge estimation, and determining the SOC data of the target battery. This solves the technical problems of high SOC calculation complexity and resource consumption, insufficient estimation accuracy, and poor real-time performance in the prior art, achieving the technical effects of improving estimation accuracy, reducing calculation complexity, reducing resource consumption, and enhancing real-time performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] To more clearly illustrate the technical solutions of the embodiments of the present disclosure, the accompanying drawings of the embodiments of the present disclosure are briefly introduced below. Flowcharts are used in this application to illustrate the operations performed by the systems according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed in precise order. Instead, various steps may be processed in reverse order or simultaneously as needed. Furthermore, other operations may be added to these processes, or one or more operations may be removed from these processes.

[0015] Figure 1 A schematic diagram of the structure of a fast SOC estimation system combined with a lightweight neural network provided in an embodiment of the present application;

[0016] Figure 2 A flowchart of a method for rapid SOC estimation using a lightweight neural network provided in an embodiment of the present application.

[0017] Description of reference numerals: lightweight sampling data determination module 10 , lightweight estimation model construction module 20 , battery state of charge estimation module 30 . DETAILED DESCRIPTION

[0018] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below.

[0019] In order to make the purpose, technical solutions and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limiting this application. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0020] In the following description, reference is made to “some embodiments”, which describes a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict, and the terms “first\second” involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. The terms “including” and “having” and any variations are intended to cover non-exclusive inclusions. For example, a process, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or modules that are not clearly listed or that are inherent to these processes, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application belongs. The terms used herein are for the purpose of describing the embodiments of this application only.

[0021] The embodiment of the present application provides a SOC fast estimation system combined with a lightweight neural network, such as Figure 1 As shown, the system includes:

[0022] The lightweight sampling data determination module 10 is used to preprocess and extract features from the electrical data sampled from the target battery operation to determine lightweight sampling data, wherein the preprocessing includes horizontal processing of feature attribute dimensions and vertical processing of time dimensions.

[0023] Preferably, the electrical data sampled from the target battery operation is preprocessed, including horizontal processing of the feature attribute dimension and vertical processing of the time dimension. Specifically, the horizontal processing of the feature attribute dimension refers to the comprehensive processing and screening of multiple attributes (such as voltage, current, temperature, internal resistance, etc.) of the battery operation sampling data, and the extraction of key features related to SOC estimation, including data denoising and normalization (filtering the noise in the electrical data to ensure data quality and mapping the numerical range of different physical quantities to a unified scale). The vertical processing of the time dimension refers to the dynamic analysis of the changes in the battery operation sampling data in the time series to explore the impact of historical data on the current SOC estimation, including selecting a reasonable time window according to the dynamic characteristics of the battery operation, such as selecting the voltage and current data within a certain time range. Then, feature extraction is performed on the preprocessed electrical data to determine lightweight sampling data, specifically including extracting key features related to SOC estimation, that is, extracting the features that are most sensitive to SOC changes from multiple attributes, such as determining key variables such as voltage change rate and current average value through correlation analysis, fusing multi-attribute data according to certain rules, simplifying data dimensions, such as calculating comprehensive indicators through specific weighting methods, improving data quality, and providing optimized input for subsequent modeling; it also includes extracting dynamic characteristics from historical data, such as the rate of change of voltage over time, current fluctuation characteristics, etc., analyzing the charging and discharging trends of the battery over a period of time, such as charging or discharging rate, to help more accurately estimate the current SOC, and using the changing patterns in the time series to enhance the dynamic response capability of the SOC estimation model.

[0024] The lightweight estimation model construction module 20 is used to divide battery management into charging conditions and discharging conditions, and to construct a lightweight estimation model using a distillation training method, wherein the lightweight estimation model includes a charging estimation sub-model and a discharging estimation sub-model, and uses a distillation loss function as a training constraint.

[0025] Preferably, the battery management is divided into charging conditions and discharging conditions, which means that the SOC change rules and characteristics of the battery are different under different working conditions, including charging conditions and discharging conditions. The charging condition refers to the operating state of the battery during charging, mainly focusing on the voltage rising trend, current decreasing characteristics and battery capacity recovery process. The discharging condition refers to the operating state of the battery during discharge, mainly focusing on the voltage falling trend, current load characteristics and remaining capacity changes. This division allows the SOC estimation model to be modeled separately for each operating condition, thereby improving the accuracy and adaptability of the estimation; then, a lightweight estimation model is constructed using distillation training. Among them, distillation training is a transfer learning method, which is usually used to transfer the knowledge of a complex "large model" (called a teacher model) to a simpler "small model" (called a student model). Specifically, a high-precision complex model (teacher model) is first trained, which has strong representation capabilities and can accurately predict SOC. Then a lightweight model (student model) is designed for actual deployment, which has a smaller parameter scale and higher computational efficiency. By designing the distillation loss function, the probability distribution and feature representation of the teacher model output are used as learning objectives to guide the training of the student model so that its performance is close to that of the teacher model.

[0026] Preferably, the distillation training method allows the lightweight model to maintain performance close to that of the complex model while significantly reducing the model size and inference time. The distillation loss function is used to constrain the learning behavior of the student model during training to ensure that it can reproduce the ability of the teacher model as much as possible when predicting SOC. It usually contains dual constraints, namely hard target loss and soft target loss, which can more effectively learn the feature representation and prediction logic in the teacher model. The hard target loss measures the gap between the predicted value of the student model and the true SOC value (such as mean square error), and the soft target loss measures the difference between the output distribution of the student model and the output distribution of the teacher model. The gap between them (such as KL divergence), the lightweight estimation model consists of two sub-models (charging estimation sub-model and discharging estimation sub-model), which are optimized for different working conditions respectively. The charging estimation sub-model is used to process the SOC estimation under the charging condition, and its training data mainly comes from the sampling data of the charging process. The discharging estimation sub-model is used to process the SOC estimation under the discharging condition, and its training data comes from the sampling data of the discharging process. Through this division, the model can be optimized separately for the charging and discharging characteristics, ensuring that the lightweight estimation model can significantly improve the computational efficiency while making the SOC estimation have high accuracy and real-time performance under different working conditions.

[0027] The battery state of charge estimation module 30 is used to transmit the lightweight sampling data to the lightweight estimation model, activate the target estimation sub-model based on the battery operating condition, perform battery state of charge estimation, and determine the SOC data of the target battery, including the state of charge estimation of the single cell and the overall state of charge estimation of the battery pack.

[0028] Preferably, the lightweight sampled data is optimized data generated after preprocessing and feature extraction. The data undergoes horizontal feature processing (such as voltage, current, and temperature screening) and vertical time series processing (such as dynamic trend analysis) to be optimized into a structured form suitable for model input. The data dimension and redundancy are significantly reduced, and only the most relevant features for SOC estimation are retained, reducing the input complexity and computational burden of the model. The lightweight sampled data is transmitted to the lightweight estimation model as the basis for model operation. Specifically, based on the battery operating condition, the target estimation sub-model is activated to estimate the battery state of charge. This means that the lightweight estimation model dynamically selects an appropriate sub-model based on the battery operating condition (charging or discharging condition). That is, by detecting characteristics such as the current direction and voltage change trend of the input data, it is determined whether the current condition is charging or discharging. Based on the dynamic activation mechanism, the charging estimation sub-model or the discharging estimation sub-model is activated based on the judgment result. The sub-models separately modeled for different operating conditions can better capture the characteristics of SOC changes under the operating condition and determine the SOC data of the target battery, thereby improving estimation accuracy and efficiency and avoiding redundant calculations.

[0029] Preferably, the battery state of charge estimation includes two levels: single cell state of charge estimation and battery pack overall state of charge estimation. Single cell state of charge estimation refers to the estimation model directly calculating the SOC based on the input lightweight sampling data for the single cell, which is used for the management and monitoring of the single cell; the overall state of charge estimation of the battery pack refers to the use of weighted average, balance calculation and other calculations based on the SOC data of multiple single cells in the battery pack to derive the overall SOC of the battery pack, and estimate the battery state of charge of the battery pack as a whole. This hierarchical estimation strategy can not only meet the refined management needs at the single cell level, but also support macro battery management through overall estimation, taking into account the accuracy of single cells and the integrity of the battery pack, and providing an efficient solution for battery management in different application scenarios.

[0030] The SOC rapid estimation system incorporating a lightweight neural network, according to an embodiment of the present invention, is designed to address the technical issues of high SOC calculation complexity and resource consumption, insufficient estimation accuracy, and poor real-time performance in the prior art. This system achieves the technical benefits of improving estimation accuracy, reducing computational complexity, minimizing resource consumption, and enhancing real-time performance. The system includes a lightweight sampling data determination module 10, a lightweight estimation model construction module 20, and a battery state of charge estimation module 30.

[0031] The specific configuration of the lightweight sampled data determination module 10 will be described in detail below. The lightweight sampled data determination module 10 may further include: an SOC correlation coefficient setting unit for setting an SOC correlation coefficient; a preprocessing level construction unit for constructing a preprocessing level based on the SOC correlation coefficient, wherein the preprocessing level includes a first preprocessing level and a second preprocessing level, wherein the first preprocessing level performs electrical data screening and the second preprocessing level performs vertical and horizontal data processing; and an electrical data preprocessing unit for preprocessing the electrical data based on the first and second preprocessing levels.

[0032] Preferably, the SOC correlation coefficient is an indicator that measures the relationship between electrical data attributes and SOC changes. The specific goals of setting the SOC correlation coefficient include evaluating the correlation between multiple battery operating data attributes such as voltage, current, temperature, internal resistance, etc. and SOC changes. For example, statistical methods (such as Pearson correlation coefficient) or machine learning techniques (such as feature importance analysis) are used, and then according to the size of the correlation, the data attributes are sorted according to their contribution to SOC estimation, and data screening and feature extraction are performed based on this. By setting the SOC correlation coefficient, the data attributes that are most valuable for SOC estimation can be effectively identified, thereby simplifying the subsequent data processing process and reducing the computational complexity; then, according to the set SOC correlation coefficient, the battery is automatically evaluated. Relationship number, construct preprocessing levels, including the first preprocessing level and the second preprocessing level. Specifically, the first preprocessing level performs electrical data screening, that is, it is used to perform preliminary screening of the original sampled data to filter out irrelevant or low-correlated electrical data, including noise filtering (removing outliers and interference noise in the sampled data to ensure data quality), attribute screening (according to the SOC correlation coefficient, screening out highly correlated electrical data attributes, such as priority retention of key data such as voltage and current that directly affect SOC estimation), and dimensionality simplification (through feature selection or dimensionality reduction technology, eliminating redundant data attributes that do not significantly contribute to SOC estimation) to reduce the data dimension of the collected electrical data and improve the efficiency of subsequent data processing and model training.

[0033] Preferably, the second preprocessing stage performs vertical and horizontal data processing, that is, further processing of the electrical data after being filtered by the first stage includes horizontal processing based on attribute dimension analysis and vertical processing based on time dimension analysis. Specifically, horizontal processing is to compare and combine multiple electrical data attributes (such as voltage, current, temperature, etc.) horizontally to extract comprehensive features. For example, by calculating composite features such as voltage change rate and current average value, the feature characterization capability of SOC estimation is enhanced; vertical processing is to analyze the time series changes of data, such as tracking the dynamic change trend of voltage or current over time, extracting time series features such as charge and discharge rate, short-term change pattern or long-term accumulation characteristics, and using moving window or time sliding technology to extract the impact of historical data on the current SOC state. The information of the data in the attribute dimension and time dimension is comprehensively utilized to provide a more comprehensive input for the lightweight estimation model. Finally, the electrical data that has passed the first and second preprocessing stages completes preprocessing, retains the features most closely related to SOC, improves the validity, structure and real-time performance of the data, thereby improving the estimation accuracy and reducing the usage of computing resources.

[0034] The specific configuration of the lightweight estimation model construction module 20 will be described in detail below. The lightweight estimation model construction module 20 may further include: an estimation sample acquisition unit for acquiring and lightweighting estimation samples based on the battery pack, wherein the estimation samples include sample sampling data and sample SOC data; an estimation large model training unit for supervised training of a large estimation model based on the estimation samples; and a distillation training processing unit for performing distillation training on the large estimation model using the output of the large estimation model as soft training data and combining it with the estimation samples to obtain the lightweight estimation model.

[0035] Preferably, estimation samples based on the battery pack are obtained and lightweight processed, wherein the estimation samples are the basic data set for training the SOC estimation model, including sample sampling data (multi-dimensional electrical data collected during the operation of the battery pack, such as voltage, current, temperature, internal resistance, etc.) and sample SOC data (the actual SOC value of the battery pack at the corresponding time, usually obtained through high-precision measurement equipment or offline analysis methods, as label data for supervised learning). The lightweight processing includes feature selection or dimensionality reduction on the original sample data, retaining the data attributes most relevant to SOC estimation, removing redundant information, and then normalizing or standardizing the sample data to make it suitable for subsequent model training, improve data quality and computational efficiency, thereby obtaining optimized sample data. Then, a prediction model is constructed based on a deep learning model (such as a multi-layer neural network or a deep recurrent neural network). The lightweight sample sampling data is used as the model input and the sample SOC data is used as the target. The error (such as the mean square error) between the model prediction value and the actual SOC value is minimized. The prediction model is supervised and trained by gradient descent, etc., and the model parameters are continuously adjusted to gradually improve the model's prediction ability, thereby obtaining a high-precision estimation model as a "teacher model" for knowledge distillation.

[0036] Preferably, the output of the estimated large model is used as soft training data, and the estimated large model is distilled and trained in combination with the estimated samples. Specifically, the output data of the estimated large model is a set of "soft labels" that are more informative than hard labels (true SOC values), i.e., soft training data, which is usually a continuous probability distribution (such as the logarithmic probability of the predicted value), which can capture subtle changes and potential patterns of SOC, or the internal feature expression of the large model, such as the comprehensive understanding and processing logic of the input samples; then, in combination with the estimated samples, the knowledge of the large model is transferred to the lightweight estimation model, i.e., the estimated large model is distilled and trained, including designing a "student model" with a simpler structure and fewer parameters, and then inputting the lightweight sample data. At the same time, the soft labels of the large model and the true SOC data are used as supervision targets, and a distillation loss function (including soft target loss and hard target loss) is designed for training and optimizing the parameters of the lightweight model so that it can reproduce the prediction ability of the large model while maintaining a simple structure. Finally, a lightweight estimation model is obtained, which can be close to the large model in performance while meeting real-time and computing resource requirements.

[0037] The specific configuration of the lightweight estimation model construction module 20 will be described in detail below. The lightweight estimation model construction module 20 may further include: a first distillation loss function determination unit, configured to determine a first distillation loss function for a charging condition, wherein the first distillation loss function represents the logic loss and data loss under the charging condition; a second distillation loss function determination unit, configured to determine a second distillation loss function for a discharging condition, wherein the second distillation loss function represents the logic loss and data loss under the discharging condition; an estimation sub-model determination unit, configured to perform distillation processing on the estimation large model based on the first distillation loss function and the second distillation loss function to determine the charging estimation sub-model and the discharging estimation sub-model; and an estimation sub-model parallelization unit, configured to parallelize the charging estimation sub-model and the discharging estimation sub-model as the lightweight estimation model.

[0038] Preferably, a first distillation loss function is used to optimize lightweight model training under charging conditions. Its function is to measure the effectiveness of the lightweight model in learning the knowledge of the large model under charging conditions. The first distillation loss function includes two parts: logic loss and data loss. For example, the class probability distribution transferred from the large model to the lightweight model determines whether a certain part of the training logic is transferred to the lightweight model based on the probability distribution value. This is used to remove charging-irrelevant or relevant information, which can be regarded as lightweight processing of the large model. Specifically, the logic loss measures the ability of the lightweight model to reproduce the predicted behavior of the large model under charging conditions. It is quantified by calculating the difference between the lightweight model output and the soft label of the large model (for example, using KL divergence or cross entropy). The data loss measures the prediction error of the lightweight model for the real SOC data under charging conditions. For example, the mean squared error (MSE) or mean absolute error (MAE) is used to evaluate the numerical deviation between the predicted value and the real SOC value. This ensures that the charging estimation sub-model can both capture the core logical pattern under charging conditions (the predicted behavior of the large model) and accurately fit the real SOC value.

[0039] Preferably, the second distillation loss function is used to optimize the lightweight model under the discharge condition, and its definition is similar to that of the first distillation loss function, including two parts: logic loss and data loss, so as to remove the discharge-irrelevant or relevant parts, and ensure that the discharge estimation sub-model can achieve dual optimization of logical reasoning and numerical fitting under the discharge condition; then, based on the first distillation loss function and the second distillation loss function, the estimation large model is distilled to generate two sub-models, namely the charging estimation sub-model and the discharge estimation sub-model. The charging estimation sub-model is designed for the charging condition, has a strong charging characteristic characterization capability, and processes the charging process. The discharge estimation sub-model is designed for discharge conditions and can accurately estimate the SOC during discharge, adapting to dynamic characteristics such as battery load changes and discharge voltage attenuation. The two estimation sub-models maintain high efficiency while approaching the prediction accuracy of the large model. Finally, the charging estimation sub-model and the discharge estimation sub-model are deployed in parallel to obtain a lightweight estimation model, that is, the two sub-models run independently, handling the SOC estimation tasks under charging and discharging conditions respectively. They can be called separately at different times, or run in parallel in a multi-threaded / multi-tasking environment to improve overall processing efficiency and accuracy.

[0040] The specific configuration of the lightweight estimation model construction module 20 will be described in detail below. The lightweight estimation model construction module 20 may further include: a first large model sample determination unit, configured to extract a first estimation sample for the estimation sample and determine a first large model sample, wherein the first large model sample is the output of the decision logic data of the estimation large model based on the first estimation sample; a first probability distribution determination unit, configured to determine a first probability distribution based on the first large model sample and the first distillation loss function, wherein the first probability distribution represents the transmission probability of the decision logic data; and a charging estimation sub-model training unit, configured to train the charging estimation sub-model once based on the first large model sample, the first probability distribution, and the first estimation sample.

[0041] Preferably, the charging estimation sub-model is trained once using the estimated samples and the soft targets (such as probability distribution) generated by the large model. Specifically, a data subset related to the charging condition is extracted from the overall estimated sample as the first estimated sample, which includes the sampling data of the battery under the charging condition (such as charging voltage, current, etc.) and the corresponding real SOC data. The first estimated sample is processed using the estimated large model to determine the first large model sample, that is, after receiving the first estimated sample, the large model generates output data according to its complex decision logic, and outputs decision logic data, including the model's estimated value of the SOC under the charging condition, and the expression of internal hidden features; then, based on the first large model sample and the first distillation loss function, the first large model sample is calculated by the first distillation loss function to generate a probability distribution (first probability distribution), which characterizes the large model's prediction confidence in the charging condition data and the transmission probability of the decision logic data (for example, the distribution of possible values ​​of SOC).

[0042] Preferably, the charging estimation sub-model is trained once based on the first large model sample, the first probability distribution and the first estimation sample. Specifically, by comparing the probability distribution output by the lightweight model with the first probability distribution, whether the sub-model has learned the logic and knowledge of the large model is measured, and the loss is usually calculated by methods such as KL divergence; by comparing the predicted SOC value of the sub-model with the actual SOC value in the first estimation sample, the actual estimation ability of the model is optimized, and the loss is usually calculated by mean square error (MSE); by jointly optimizing soft targets and hard targets, it is ensured that the lightweight charging estimation sub-model can reproduce the logic of the large model and meet the actual estimation needs; then the initial lightweight A quantified charging estimation sub-model is constructed, and the sampling data of the first estimation sample is input into the sub-model. At the same time, the output of the first large model sample and the first probability distribution are used as guidance information. The first distillation loss function is used to calculate the gap between the model predicted output and the actual SOC value (hard target) and the first probability distribution (soft target). The backpropagation algorithm and optimizer (such as SGD or Adam) are used to update the parameters of the sub-model according to the loss function. The training process is repeated until the loss function value reaches the preset standard or the model performance is no longer significantly improved. The charging estimation sub-model is obtained, which can capture the main laws of SOC changes under charging conditions and has efficient and high-precision SOC estimation capabilities.

[0043] The specific configuration of the battery SOC estimation module 30 will be described in detail below. The battery SOC estimation module 30 may further include: a battery operating condition determination unit, configured to determine the battery operating condition of the target battery, where the battery operating condition is either a charging condition or a discharging condition; and a SOC estimation unit, configured to activate a target estimation sub-model within the lightweight estimation model based on the battery operating condition, input the lightweight sampled data, and perform SOC estimation for both the single cell and the battery pack as the SOC data for the target battery.

[0044] Preferably, determining the battery operating condition of the target battery refers to determining whether the target battery is in a charging condition or a discharging condition, including determining whether the battery is in a charging or discharging state by collecting real-time current and voltage data, or determining the operating condition based on the current direction (positive and negative signs) and voltage change trend through preset rules or simple logic, and then dynamically selecting the corresponding sub-model according to the determined operating condition, activating the target estimation sub-model within the lightweight estimation model, that is, if it is a charging condition, activating the charging estimation sub-model, receiving the lightweight sampling data as input data to perform state of charge estimation of the single cell and the state of charge estimation of the battery pack. Specifically, the state of charge estimation of the single cell is that the charging estimation sub-model estimates the SOC based on the lightweight sampling data of the single cell, captures the individual characteristics of the single cell, and is used to monitor the state of the single cell to prevent abnormal conditions (such as overcharging). The state of charge estimation of the battery pack is based on the single cell SOC data and the overall sampling data, combined with weight calculation, averaging method or other algorithms to estimate the SOC of the battery pack for overall management of the operating status of the battery pack.

[0045] Preferably, if the operating condition is discharge, the discharge estimation sub-model is activated, and lightweight sampling data is received as input data to estimate the state of charge of the single cell and the state of charge of the battery pack. Specifically, the state of charge estimation of the single cell is that the discharge estimation sub-model estimates the SOC based on the lightweight sampling data of the single cell, captures the individual characteristics of the single cell, and is used to monitor the state of the single cell to prevent abnormal conditions (such as over-discharge). The state of charge estimation of the battery pack is based on the single cell SOC data and the overall sampling data, combined with weight calculation, averaging method or other algorithms to estimate the SOC of the battery pack, which is used to manage the overall operating status of the battery pack. The battery state of charge estimation results under these two operating conditions are used as the SOC data of the target battery to reflect the state of charge of the target battery. Through the division of operating conditions and the lightweight model, efficient and accurate SOC estimation is achieved, meeting the requirements of real-time performance and low resource consumption.

[0046] The specific configuration of the battery SOC estimation module 30 will be described in detail below. The battery SOC estimation module 30 may further include: a first determination result determination unit, configured to identify the SOC data of a single cell based on the SOC data of the target battery, perform an abnormal SOC determination, and determine a first determination result, wherein the charge and discharge threshold of the single cell is used as a determination criterion; a second determination result determination unit, configured to identify the battery pack SOC data, perform an abnormal charge distribution determination, and determine a second determination result, wherein the battery pack SOC data represents the SOC fluctuation within the battery pack, and the SOC fluctuation between single cells is used as a determination criterion; and an abnormal operating condition warning unit, configured to issue an abnormal operating condition warning for the target battery based on the first and second determination results.

[0047] Preferably, identifying the single cell SOC data based on the SOC data of the target battery means extracting the SOC value of the single cell from the SOC data of the target battery, and then making an abnormal state of charge determination based on the charge and discharge critical value of the single cell, wherein the charge critical value defines the upper limit of the single cell state of charge SOC, for example, 90% to 95%, and the discharge critical value defines the lower limit of the single cell SOC, for example, 10% to 15%. If the single cell SOC is higher than the charge critical value, it is determined to be in an overcharge state; if the single cell SOC is lower than the discharge critical value, it is determined to be in an over-discharge state, and then a first determination result is determined, marking which single cells are in an abnormal state and their abnormality type (such as overcharge or over-discharge).

[0048] Preferably, identifying the battery pack SOC data refers to extracting the overall SOC state of the battery pack from the SOC data of the target battery, including the SOC distribution of the single cells in the group, and then making an abnormal charge distribution judgment based on the SOC fluctuation between the single cells, wherein the fluctuation range is based on the battery pack design, and the maximum SOC difference allowed between the single cells is preset, for example, 3% to 5%. If the SOC difference between the single cells in the battery pack exceeds the allowable range, it is judged that the charge distribution is abnormal, and then the second judgment result is determined, marking the abnormal distribution of the battery pack and the specific SOC fluctuation limit range.

[0049] Preferably, an abnormal operating condition warning for the target battery is performed based on the first and second determination results. Specifically, the first determination result (single cell status) and the second determination result (battery pack status) are combined to identify whether the overall operating condition of the target battery is abnormal. This includes generating a single cell abnormality warning if a single cell is overcharged or over-discharged, generating a pack distribution abnormality warning if the pack's SOC fluctuation exceeds the allowable range, and generating a severe abnormality warning if both conditions exist. The output warning information includes the abnormality type, impact range, and possible cause. By identifying the abnormal state of single cells and the SOC consistency of the battery pack, highly targeted abnormality warning information is generated to prevent safety accidents (such as thermal runaway and overcharge explosion) and ensure the safety and reliability of battery management.

[0050] In the above, refer to Figure 1 The SOC fast estimation system combined with a lightweight neural network according to an embodiment of the present invention is described in detail. Figure 2 A method for fast SOC estimation combined with a lightweight neural network according to an embodiment of the present invention is described.

[0051] The SOC fast estimation method combined with lightweight neural network, such as Figure 2As shown, the method includes: preprocessing and feature extraction of electrical data sampled from the target battery operation to determine lightweight sampling data, wherein the preprocessing includes horizontal processing of feature attribute dimensions and vertical processing of time dimensions; dividing battery management into charging conditions and discharging conditions, and constructing a lightweight estimation model using a distillation training method, wherein the lightweight estimation model includes a charging estimation sub-model and a discharging estimation sub-model, and uses a distillation loss function as a training constraint; transmitting the lightweight sampling data to the lightweight estimation model, activating the target estimation sub-model based on the battery condition, performing battery state of charge estimation, and determining the SOC data of the target battery, wherein the state of charge estimation of the single cell and the overall state of charge estimation of the battery pack are included.

[0052] The SOC rapid estimation system combined with a lightweight neural network provided in an embodiment of the present invention can execute the SOC rapid estimation method combined with a lightweight neural network provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0053] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules may be used and run on the user terminal and / or server, and the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of the functional units are only for the convenience of distinguishing each other and are not used to limit the scope of protection of the present invention.

[0054] The above specific embodiments do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application shall be included within the scope of protection of this application.

Claims

1. The SOC fast estimation system combined with lightweight neural network is characterized by: The system comprises: A lightweight sampling data determination module is used to preprocess and extract features from the electrical data sampled from the target battery operation to determine the lightweight sampling data, wherein the preprocessing includes horizontal processing of the feature attribute dimension and vertical processing of the time dimension; A lightweight estimation model construction module is used to divide battery management into charging and discharging conditions and construct a lightweight estimation model using a distillation training method. The lightweight estimation model includes a charging estimation sub-model and a discharging estimation sub-model, and uses a distillation loss function as a training constraint. a battery state-of-charge estimation module, configured to transmit the lightweight sampling data to the lightweight estimation model, activate the target estimation sub-model based on the battery operating condition, perform battery state-of-charge estimation, and determine the SOC data of the target battery, including the state-of-charge estimation of the single cell and the overall state-of-charge estimation of the battery pack; The lightweight sampling data determination module includes: SOC correlation coefficient setting unit, used to set the SOC correlation coefficient; a preprocessing level construction unit, configured to construct a preprocessing level based on the SOC correlation coefficient, wherein the preprocessing level includes a first preprocessing level and a second preprocessing level, the first preprocessing level performing electrical data screening, and the second preprocessing level performing vertical and horizontal data processing; an electrical data preprocessing unit, configured to preprocess the electrical data based on the first preprocessing checkpoint and the second preprocessing checkpoint; The lightweight estimation model building module includes: An estimation sample acquisition unit, configured to acquire an estimation sample based on a battery pack and perform lightweight processing, wherein the estimation sample includes sample sampling data and sample SOC data; An estimation large model training unit, configured to supervise the training of the estimation large model based on the estimation samples; a distillation training processing unit, configured to perform distillation training processing on the estimation large model using the output of the estimation large model as soft training data and combining the estimation samples to obtain the lightweight estimation model; The lightweight estimation model building module includes: a first distillation loss function determining unit, configured to determine a first distillation loss function for a charging condition, wherein the first distillation loss function represents a logic loss and a data loss under the charging condition; a second distillation loss function determining unit, configured to determine a second distillation loss function for a discharge operating condition, wherein the second distillation loss function represents a logic loss and a data loss under the discharge operating condition; an estimation sub-model determining unit, configured to perform distillation processing on the estimation large model based on the first distillation loss function and the second distillation loss function to determine the charge estimation sub-model and the discharge estimation sub-model; The estimation sub-model parallel unit is used to parallelize the charging estimation sub-model and the discharging estimation sub-model as the lightweight estimation model.

2. The SOC fast estimation system combined with a lightweight neural network as claimed in claim 1, characterized in that: The lightweight estimation model building module includes: a first large model sample determining unit, configured to extract a first estimated sample from the estimated sample and determine a first large model sample, wherein the first large model sample is an output of the estimated large model based on the decision logic data of the first estimated sample; a first probability distribution determining unit, configured to determine a first probability distribution based on the first large model sample and the first distillation loss function, wherein the first probability distribution represents a transmission probability of decision logic data; The charging estimation sub-model training unit is configured to train the charging estimation sub-model based on the first large model sample, the first probability distribution and the first estimation sample.

3. The SOC fast estimation system combined with a lightweight neural network as claimed in claim 1, characterized in that: The battery state of charge estimation module includes: a battery operating condition determining unit, configured to determine a battery operating condition of the target battery, wherein the battery operating condition is a charging operating condition or a discharging operating condition; The state of charge estimation unit is used to activate the target estimation sub-model in the lightweight estimation model based on the battery operating condition, input the lightweight sampling data, and perform state of charge estimation of the single cell and the battery pack as the SOC data of the target battery.

4. The SOC fast estimation system combined with a lightweight neural network as claimed in claim 3, characterized in that: The battery state of charge estimation module includes: a first determination result determining unit, configured to identify the SOC data of a single cell based on the SOC data of the target battery, perform abnormal state of charge determination, and determine a first determination result, wherein the charge and discharge threshold of the single cell is used as a determination criterion; a second determination result determining unit, configured to identify battery pack SOC data, perform abnormal charge distribution determination, and determine a second determination result, wherein the battery pack SOC data represents the battery pack, with SOC fluctuations between single cells serving as a determination criterion; The abnormal operating condition warning unit is used to provide an abnormal operating condition warning for the target battery based on the first determination result and the second determination result.

5. A fast SOC estimation method based on a lightweight neural network is characterized by: The method is applied to the SOC fast estimation system combined with a lightweight neural network according to any one of claims 1 to 4, and the method comprises: Preprocess and extract features from the electrical data sampled from the target battery operation to determine lightweight sampled data. Preprocessing includes horizontal processing of feature attribute dimensions and vertical processing of time dimensions. Divide battery management into charging and discharging conditions, and construct a lightweight estimation model using distillation training. The lightweight estimation model includes a charging estimation sub-model and a discharging estimation sub-model, and uses a distillation loss function as a training constraint. The lightweight sampling data is transmitted to the lightweight estimation model, and the target estimation sub-model is activated based on the battery operating condition to perform battery state of charge estimation and determine the SOC data of the target battery, including the state of charge estimation of the single cell and the overall state of charge estimation of the battery pack.

Images