Battery health state prediction method and electronic device based on SOC coding

Abstract

This application belongs to the field of battery management technology, specifically relating to a battery health state prediction method, electronic device, and storage medium based on SOC encoding. The method includes: acquiring time-series data of the battery during charging / discharging; using the SOC value as a location encoding parameter, dividing the time-series data into multiple consecutive SOC intervals according to the SOC value, and calculating feature data within each SOC interval; if the time-series data does not meet a preset integrity condition, inputting the location encoding and feature data corresponding to the SOC interval into a pre-trained deep learning data completion model to obtain the completed feature data; inputting the complete feature data into a pre-trained battery health state (SOH) prediction model, and outputting the predicted SOH value. The method of this application can complete missing feature data even when charging / discharging data is incomplete, significantly improving the model's adaptability and reliability in harsh edge environments.

Description

Technical Field

[0001] This application belongs to the field of battery management technology, specifically relating to a battery health state prediction method based on SOC coding, an electronic device, and a storage medium. Background Technology

[0002] With the rapid development of new energy technologies, lithium-ion batteries, due to their advantages such as high energy density, long cycle life, and low self-discharge rate, are widely used in electric vehicles, portable energy storage devices, and distributed energy systems. Battery State of Health (SOH), as a crucial indicator of the degree of battery performance degradation, directly relates to the safety, reliability, and lifespan of the battery system. Therefore, accurately assessing battery SOH and implementing effective battery management accordingly has become one of the core tasks of a Battery Management System (BMS).

[0003] Existing SOH estimation methods mainly include model-based methods and data-driven methods. Model-based methods (such as equivalent circuit models and electrochemical models) rely on accurate modeling of the battery's internal mechanisms, but the battery aging process is complex and nonlinear, model parameters are difficult to calibrate accurately, and generalization ability is limited. Data-driven methods (such as support vector machines, neural networks, and long short-term memory networks) predict SOH by learning patterns from historical data, and have better adaptability.

[0004] However, existing data-driven methods still face many challenges in practical applications. Especially in edge scenarios such as outdoor portable energy storage devices or electric vehicles, the time-series data (voltage, current, temperature, etc.) collected during battery charging / discharging often suffers from missing, abnormal, or discontinuous sampling due to factors such as communication bandwidth limitations, poor sensor reliability, and harsh operating environments. When the input data is incomplete, the prediction accuracy of existing deep learning models drops significantly, and they may even fail to output valid results, making it difficult to meet the demands of high reliability and intelligent battery management in practical applications.

[0005] Therefore, there is an urgent need for a technical solution that can maintain high SOH prediction accuracy even when battery timing data is incomplete, in order to meet the urgent needs of edge energy storage devices for high reliability and intelligent battery management. Summary of the Invention

[0006] The purpose of this application is to provide a battery health state prediction method, electronic device, and storage medium based on SOC coding, so as to solve the problem of decreased SOH prediction accuracy due to incomplete charge / discharge data in the prior art.

[0007] In a first aspect, embodiments of this application provide a battery health state prediction method based on SOC coding, the method comprising: Acquire timing data of the battery during the charging / discharging process, the timing data including the SOC value and state parameters of the charging / discharging process; Using the SOC value as the location encoding parameter, the time series data is divided into multiple consecutive SOC intervals according to the SOC value, and the feature data in each SOC interval is calculated. If the time series data does not meet the preset integrity conditions, the position code corresponding to the SOC interval and the feature data are input into a pre-trained deep learning data completion model to obtain the completed feature data. The complete feature data is input into a pre-trained battery health state (SOH) prediction model, which outputs the predicted SOH value.

[0008] Optionally, the preset integrity condition is: within a charging / discharging process, the number of SOC intervals corresponding to the timing data is equal to the total number of SOC interval divisions.

[0009] Optionally, after outputting the predicted SOH value, the following is also included: The linear regression coefficients of the interval characteristics of each SOC interval and SOH are calculated based on the offline experimental data, and the linear regression coefficients are normalized to obtain the weight coefficients of each SOC interval. The predicted SOH value is adaptively corrected based on the weighting coefficients of each SOC interval to obtain the corrected SOH estimate.

[0010] Optional, also includes: The observed energy of each SOC interval is extracted from the complete feature data, the reference energy of each SOC interval of the current battery is obtained from the offline experimental data, and the correction coefficient of the current battery is calculated based on the observed energy and reference energy of each SOC interval. The predicted SOH value is adaptively corrected based on the correction coefficient and the weight coefficient of each SOC interval to obtain the corrected SOH estimate.

[0011] Optionally, the deep learning data completion model is trained using a masking training method, which includes: randomly masking all or part of the feature data in multiple SOC intervals in each training iteration, and using cross-validation to alternately mask different SOC intervals in different iterations, using the true value of the masked data as a supervision signal to train the model to complete the masked data.

[0012] Optionally, the SOH prediction model includes: The state change sensing module is used to perform differential or comparative modeling of the feature data of adjacent SOC intervals in the complete feature data to extract the change patterns during the charging / discharging process. The feature encoding module is used to map the feature data of different physical quantities to a unified feature space; The attention module is used to assign weights to the changing features of different SOC intervals and to filter out changing patterns that are sensitive to SOH.

[0013] Optionally, the deep learning data completion model includes: An embedding layer is used to map the position encoding corresponding to the SOC interval to a first position vector; A fully connected layer is used to linearly transform the feature data into a feature vector; The fusion layer is used to add the first position vector to the feature vector to obtain the fusion vector; The Transformer layer is used to perform temporal modeling on the fused vector and output contextual features; The output layer is used to output the completed feature data based on the context features.

[0014] Optionally, the time-series data further includes a cycle number, and the input of the deep learning data completion model further includes a position code corresponding to the cycle number. The embedding layer is further configured to map the position code corresponding to the cycle number into a second position vector, and the fusion layer is further configured to add the first position vector, the second position vector, and the feature vector to obtain the fusion vector.

[0015] Secondly, embodiments of this application provide an electronic device, including at least one processor and a memory communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the method described above.

[0016] Thirdly, embodiments of this application provide a computer storage medium storing instructions or programs that, when executed by at least one processor, cause the at least one processor to perform the method described above.

[0017] This application provides a battery health state prediction method based on SOC encoding. First, time-series data of the battery during charging / discharging is acquired. Using the SOC value as a location encoding parameter, the time-series data is divided into multiple consecutive SOC intervals, and feature data within each SOC interval is calculated. When the time-series data is incomplete, the location encoding and feature data corresponding to the SOC interval are input into a pre-trained deep learning data completion model to obtain complete feature data. Then, the complete feature data is input into a pre-trained battery health state (SOH) prediction model to output the predicted SOH value. This method, by introducing a deep learning data completion model, can complete missing feature data even when charging / discharging data is incomplete, providing complete input for subsequent SOH prediction and significantly improving the model's adaptability and reliability in harsh edge environments. Attached Figure Description

[0018] Figure 1 An exemplary flowchart of a battery health state prediction method based on SOC coding is shown; Figure 2 An exemplary schematic diagram of the structure of a deep learning data completion model is shown; Figure 3 An exemplary schematic diagram of the SOH prediction model is shown. Figure 4 An example is shown in the flowchart of adaptive correction; Figure 5 An exemplary diagram illustrating the training method of a deep learning data completion model is shown. Figure 6 An exemplary schematic diagram of the hardware structure of an electronic device is shown. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] The technical terms used in this application are explained below: 1. SOC State of Charge (SOC) is an indicator describing the remaining capacity of a battery, defined as the ratio of the battery's current remaining capacity to its current rated capacity.

[0021] 2. SOH State of Health (SOH) refers to the percentage of a battery's current rated capacity relative to its initial rated capacity. It is an important indicator for measuring battery performance and aging.

[0022] 3. LSTM Long Short-Term Memory (LSTM) networks are a variant of Recurrent Neural Networks (RNNs) that use memory units to better capture long-term dependencies.

[0023] 4. CNN A convolutional neural network is a type of neural network used for image and time series data processing, which extracts local features through convolutional layers.

[0024] 5. Transformer Transformers, a neural network architecture based on self-attention mechanisms, are widely used in natural language processing and time series analysis.

[0025] 6. Cross-validation A data partitioning and validation method for evaluating the generalization performance of a model is proposed. This method divides the sample data into multiple subsets and alternately uses some subsets for model training and the remaining subsets for validation, thereby reducing the evaluation bias caused by a single data partitioning.

[0026] 7. MAE Mean Absolute Error (MAE) is a metric for evaluating the accuracy of a predictive model. It calculates the average of the absolute differences between the predicted and actual values.

[0027] 8. ResNet Residual networks, by introducing skip connections into neural network layers, allow inputs to bypass one or more layers and be added to the output, thereby mitigating the degradation problem of deep networks.

[0028] Please refer to Figure 1 , Figure 1 A flowchart of a battery health state prediction method based on SOC coding is shown. Figure 1 As shown in the figure, this application provides a battery health state prediction method, including the following steps: Step S101: Obtain timing data of the battery during the charging / discharging process.

[0029] Specifically, timing data of the battery during the charging / discharging process is collected through a BMS (Battery Management System). This timing data includes at least the SOC value and state parameters of the charging / discharging process. State parameters include, but are not limited to, voltage, current, temperature, charging / discharging duration, and input / output energy. The discharge current value can be defined as negative, and the charging current value as positive.

[0030] It should be noted that the timing data can be timing parameters for the charging process only, timing parameters for the discharging process only, or timing parameters that include both charging and discharging processes. Preferably, the timing parameters are for the charging process only.

[0031] Step S102: Using the SOC value as the location coding parameter, the time series data is divided into multiple consecutive SOC intervals according to the SOC value, and the feature data in each SOC interval is calculated.

[0032] Traditional LSTM or other methods based on single time series modeling typically only depict the voltage-time relationship, making it difficult to understand the battery electrochemical state information implied by time positions within the charge / discharge process, and also difficult to depict the performance evolution caused by battery aging across charge / discharge cycles. However, for a typical electrochemical system like a battery, the direct determinant of its instantaneous behavior and performance evolution is not time itself, but the current lithium intercalation state (equivalent to SOC). That is, the battery's electrochemical state is mainly determined by SOC. Therefore, this application uses SOC value instead of time as the location encoding basis, uniformly mapping the time series data of the battery's entire life cycle to a standardized SOC coordinate space, shifting the model's focus from "voltage changes over time" to "voltage changes with electrochemical state." This state-aligned representation effectively decouples operating condition fluctuations (such as changes in charge / discharge rate) at the physical level, thereby significantly improving the model's generalization ability to different operating conditions and aging stages.

[0033] Specifically, the SOC from 0% to 100% is divided into N consecutive intervals, where N=100 in this embodiment, meaning each interval corresponds to a 1% change in SOC (e.g., SOC 0-1%, 1-2%, ..., 99-100%). For each SOC interval, based on the original sampled data falling within that interval, the following feature data are calculated: Voltage change value: End voltage of interval minus start voltage of interval; Average current: The average value of all current sampling points within the interval; Charge / discharge duration: The duration of the interval; Temperature average: The average value of all temperature sampling points within the interval; Unit input / output energy within a range: cumulative input / output energy within the range (obtained by integrating voltage and current).

[0034] In this way, each single charge / discharge process is converted into a feature matrix of shape (N, M), where N is the number of SOC intervals (100 in this embodiment) and M is the feature dimension (5 in this embodiment). At the same time, the position code corresponding to each SOC interval is recorded, that is, the interval index 0~99.

[0035] Step S103: If the time series data does not meet the preset integrity conditions, the position encoding and feature data corresponding to the SOC interval are input into the pre-trained deep learning data completion model to obtain the completed feature data.

[0036] In one embodiment, the preset integrity condition is: within a charging / discharging process, the number of SOC intervals corresponding to the time-series data is equal to the total number of SOC interval divisions (i.e., 100). If, during the charging / discharging process, due to communication interruption, sensor failure, or other reasons, some SOC intervals have no data, or the number of SOC intervals corresponding to the collected data is less than 100, then the data is determined to be incomplete.

[0037] The structure of a deep learning data completion model is as follows: Figure 2 As shown, it includes: The embedding layer is used to map the position encoding corresponding to the SOC interval to the first position vector, with the vector dimension being d_model (a hyperparameter that can be adjusted flexibly).

[0038] The fully connected layer (Linear Layer) is used to linearly transform feature data (5 dimensions) into feature vectors, which also have the dimension d_model.

[0039] The fusion layer is used to add the first position vector to the feature vector to obtain the fused vector.

[0040] The Transformer layer is used to perform temporal modeling on the fused vector, capture the dependencies between different SOC intervals, and output contextual features.

[0041] The output layer is a fully connected layer used to map the context features back to the original feature dimensions (5 dimensions), and outputs the completed feature data.

[0042] In one embodiment, the time-series data also includes a cycle number, indicating which charge / discharge cycle the current process is in. The battery's health state does not change randomly, but exhibits a monotonically decreasing non-linear trend with increasing cycle count (e.g., rapid degradation in the early stages, gradual degradation in the middle stages, and accelerated degradation in the later stages). Traditional time-series models (such as LSTM and Transformer) are based solely on "time steps," making it difficult to distinguish whether the "characteristic differences within the same SOC range at different cycle counts" stem from fluctuations in operating conditions or from aging. By introducing cycle number position encoding, the model can perceive which stage of the battery's lifespan (early, middle, or late) the current charge / discharge process is in, thus considering the physical characteristic changes caused by aging when completing the data, further improving the completion accuracy.

[0043] In step S103 of this embodiment, the input of the deep learning data completion model also includes the position code corresponding to the loop number. The embedding layer is also used to map the position code corresponding to the loop number to a second position vector. The fusion layer adds the first position vector, the second position vector and the feature vector to obtain the fusion vector.

[0044] Training methods for deep learning data completion models, such as Figure 5 As shown, a masking training strategy is adopted: in each training iteration, a portion of the SOC interval is randomly selected as the mask interval, and all or part of the feature data in the interval is marked as missing (e.g., set to -1); the feature data and position encoding of the unmasked SOC interval are used as input, and the original feature data of the masked interval is used as the prediction target. Cross-validation is used to alternately mask the data, so that the model learns to accurately complete the data when there is a large area of ​​missing data; at the same time, the MAE loss function is used to effectively preserve the "feature peak" information that is crucial for SOH prediction.

[0045] Step S104: Input the complete feature data into the pre-trained battery health state (SOH) prediction model and output the predicted SOH value.

[0046] In one embodiment, the structure of the SOH prediction model is as follows: Figure 3 As shown, it includes: The state change sensing module is used to perform differential or comparative modeling of feature data from adjacent SOC intervals to extract change patterns during the charging / discharging process. For example, it calculates the difference between feature data from adjacent intervals to form a change feature sequence, and learns the change patterns of voltage, current, temperature, and energy during the charging / discharging process.

[0047] The feature encoding module adopts the ResNet architecture and uses a one-dimensional convolutional network (1D-CNN) to jointly encode the feature data of different physical quantities, mapping features of different dimensions such as voltage change value, average current, and average temperature to a unified feature space, thus eliminating the difference in dimensions.

[0048] The attention module employs a Transformer encoder structure and, based on a multi-head attention mechanism, assigns weights to the changing characteristics of different SOC ranges, automatically selecting the changing patterns most sensitive to SOH. For example, the voltage rise rate change at the end of charging (high SOC range) often more effectively reflects battery aging, and the attention module assigns a higher weight to this range.

[0049] The output layer is a fully connected layer used to map the high-dimensional feature vector output by the attention module to the SOH prediction scalar, thereby achieving regression output.

[0050] This design of the SOH prediction model fully leverages the physical characteristics of lithium batteries, namely that battery health is often not reflected in a single moment's value, but rather in the change patterns (such as increased polarization) during specific stages (e.g., the high SOC range). In particular, the aging mechanism of lithium batteries in energy storage scenarios is a long-term and continuous process, and traditional solutions (such as AI algorithms like LSTM) have limited capabilities in handling extremely long periods. By adopting the Transformer encoder architecture proposed in this application, it is possible to capture the "context" over near-infinite time periods, thereby discovering the impact of minute long-term electrochemical changes on the quantified battery aging stages (i.e., SOH).

[0051] The SOH prediction model is trained using laboratory full-lifecycle aging data to ensure that the model learns the objective change patterns of standard lithium-ion batteries.

[0052] In one embodiment, if the time series data meets the preset integrity condition (i.e., the time series data covers all 100 SOC intervals), the feature data obtained after dividing according to the SOC intervals is directly input into the SOH prediction model, and the predicted SOH value is output.

[0053] In reality, a standard charging process from 0-100% SOC is rarely observed. Furthermore, the actual input energy during charging often differs significantly from the theoretical input energy due to various factors (inflow measurement errors, incomplete charging process, etc.). The same applies to the discharge process and output energy. In addition, incomplete charge / discharge data, even after being supplemented by a deep learning data completion model, can introduce prediction bias. To reduce the impact of real-world data noise and the bias in the data supplemented by the deep learning model on SOH, this application introduces a correction coefficient to standardize and correct the SOH value predicted by the SOH prediction model, thereby better matching the actual conditions of the current battery and outputting a more effective SOH.

[0054] Furthermore, since the feature data of different SOC intervals contribute differently to SOH, this application embodiment also introduces a weighting coefficient to adjust the impact of SOC intervals on SOH prediction.

[0055] Therefore, in the SOH prediction model, after outputting the predicted SOH value, the following steps are also taken: adaptively correcting the predicted SOH value according to the correction coefficient and weight coefficient to obtain the final SOH estimate.

[0056] In one embodiment, the adaptive correction process is as follows: Figure 4 As shown, it specifically includes: Step S401: Extract the observed energy of each SOC interval from the complete feature data, obtain the reference energy of each SOC interval of the current battery from the offline experimental data, and calculate the correction coefficient of the current battery based on the observed energy and reference energy of each SOC interval.

[0057] The complete feature data refers to the complete feature data after being completed by the deep learning data completion model, or the feature data corresponding to the time series data that meets the integrity condition. The observed energy of each SOC interval is the actual input / output energy (actual value) of each interval unit sampled when the integrity condition is met, or the predicted input / output energy (completed value) predicted by the deep learning data completion model when the integrity condition is not met.

[0058] By conducting full-cycle charge-discharge experiments on the battery cells, characteristic data for each State of Charge (SOC) range during each cycle were measured, including voltage, current, temperature, and input / output energy. Simultaneously, the State of Harm (SOH) for each charge or discharge cycle could be determined using the capacity method. Based on this offline experimental data, reference input / output energy for each SOC range of the current battery can be obtained.

[0059] In one embodiment, the correction coefficient The calculation formula is:

[0060] in, =0,1,2,……, N-1 , which are the SOC position parameters. The first one obtained based on offline experimental data Reference energy for each SOC interval, For the first The observation energy of each SOC interval.

[0061] Here, the input / output energy values ​​are chosen for calculating the correction coefficient because, among the data from a large number of existing outdoor energy storage devices, the standard deviation of the input / output energy values ​​across batteries is relatively small compared to other battery characteristics. This allows them to reflect certain differences between different batteries without being excessively affected by noise. Those skilled in the art can also use other battery characteristics with small standard deviations across batteries to calculate the correction coefficient.

[0062] Step S402: Calculate the interval characteristics of each SOC interval and the linear regression coefficient of SOH based on the offline experimental data, and normalize the linear regression coefficients to obtain the weight coefficients of each SOC interval.

[0063] Specifically, the linear regression coefficients The calculation formula is:

[0064] in, For the first Feature data for each SOC interval This is the corresponding battery health status value (obtained from offline experimental data using the capacity method). yes and The covariance reflects the direction and strength of the linear correlation between the two. yes The variance reflects the fluctuation range of the feature itself. This is achieved by calculating the variance of each SOC interval. This allows us to quantitatively determine which SOC intervals are most valuable for SOH prediction, thus providing a scientific basis for subsequent weight allocation.

[0065] Weighting coefficient The calculation formula is:

[0066] Weighting coefficient The value of is a real number between [0,1], and its sum is 1. This parameter reflects the importance of different SOC intervals to SOH prediction; the stronger the correlation, the higher the weight of the interval.

[0067] Step S403: Adaptively correct the predicted SOH value based on the correction coefficient and the weight coefficient of each SOC interval to obtain the corrected SOH estimate.

[0068] in, The SOH value output by the SOH prediction model is divided equally into each SOC interval and used for weighted calculation with the weighting coefficients.

[0069] It is understood that those skilled in the art can also adaptively correct the predicted SOH value using only the correction coefficient or only the weighting coefficient. When correcting only using the correction coefficient, the formula for calculating the corrected SOH estimate is:

[0070] When corrected only by weighting coefficients, the formula for calculating the corrected SOH estimate is:

[0071] By introducing correction coefficients and weighting coefficients to adaptively correct the predicted SOH value, the data bias caused by noise in real data and deep learning data completion model is effectively compensated, and the contribution of key SOC intervals to SOH is emphasized, further improving the prediction accuracy.

[0072] This application provides a method for predicting battery health status. First, time-series data of the battery during charging / discharging is acquired. Using the State of Charge (SOC) value as a location encoding parameter, the time-series data is divided into multiple consecutive SOC intervals, and feature data within each SOC interval is calculated. When the time-series data is incomplete, the location encoding and feature data corresponding to the SOC interval are input into a pre-trained deep learning data completion model to obtain complete feature data. Then, the complete feature data is input into a pre-trained battery health status (SOH) prediction model, outputting the predicted SOH value. This method, by introducing a deep learning data completion model, can complete missing feature data even when charging / discharging data is incomplete, providing complete input for subsequent SOH prediction and significantly improving the model's adaptability and reliability in harsh edge environments. Furthermore, when charging / discharging data is complete, the deep learning data completion model can be skipped for direct prediction, reducing computational overhead and improving computational efficiency while maintaining accuracy.

[0073] According to an embodiment of this application, an electronic device is provided, such as... Figure 6 The diagram shown is a hardware structure schematic of an electronic device according to an embodiment of this application. The electronic device 100 includes a processor 10, a memory 20, and a communication interface 30. The processor 10, memory 20, and communication interface 30 are connected via lines. Figure 6 In the embodiment shown, the processor 10, memory 20, and communication interface 30 are connected to each other via a bus.

[0074] The memory 20 is used to store software programs, computer-executable program instructions, etc. The memory 20 may include a program storage area and a data storage area, wherein the program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the electronic device 100, etc.

[0075] The memory 20 can be a read-only memory (ROM), or other types of static storage devices that can store static information and instructions, or random access memory (RAM), or other types of dynamic storage devices that can store information and instructions, or electrically erasable programmable read-only memory (EEPROM). The specific type is not limited here.

[0076] For example, the aforementioned memory 20 can be Double Data Rate Synchronous Dynamic Random Access Memory (DDR SDRAM). This memory 20 can exist independently but is connected to the processor 10. Optionally, the memory 20 can also be integrated with the processor 10, for example, integrated within one or more chips.

[0077] In some embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and this remote memory may be connected to the electronic device 100 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0078] The processor 10 connects various parts of the electronic device 100 using various interfaces and lines. By running or executing software programs stored in the memory 20 and calling data stored in the memory 20, it performs various functions of the electronic device 100 and processes data, such as implementing the methods described in any embodiment of this application.

[0079] The processor 10 can be a field programmable gate array (FPGA), a digital signal processor (DSP), a central processing unit (CPU), or the like.

[0080] Processor 10 can be a single-core processor or a multi-core processor. For example, processor 10 can be composed of multiple FPGAs or multiple DSPs. Furthermore, processor 10 can refer to one or more devices, circuits, and / or processing cores for processing data (e.g., computer program instructions). Processor 10 can be a standalone semiconductor chip or integrated with other circuits into a single semiconductor chip. For example, it can form a system-on-a-chip (SoC) with other circuits (such as encoding / decoding circuits, hardware acceleration circuits, or various bus and interface circuits), or it can be integrated as a built-in processor within an application-specific integrated circuit (ASIC). This ASIC with integrated processor can be packaged separately or together with other circuits.

[0081] The communication interface 30 can use a transceiver device, such as a transceiver, to enable communication between the electronic device 100 and other devices or communication networks.

[0082] This application also provides a non-volatile computer-readable storage medium storing computer-executable instructions that are executed by one or more processors, for example, to perform the battery health state prediction method steps described above.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and not to limit them; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of this application as described above, which are not provided in detail for the sake of brevity; although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A battery state of health prediction method based on SOC coding, characterized by, The method includes: Acquire timing data of the battery during the charging / discharging process, the timing data including the SOC value and state parameters of the charging / discharging process; Using the SOC value as the location encoding parameter, the time series data is divided into multiple consecutive SOC intervals according to the SOC value, and the feature data in each SOC interval is calculated. If the time series data does not meet the preset integrity condition, the position code corresponding to the SOC interval and the feature data are input into the pre-trained deep learning data completion model to obtain the completed feature data. The preset integrity condition is: within a charging / discharging process, the number of SOC intervals corresponding to the time series data is equal to the total number of SOC interval divisions. The complete feature data is input into a pre-trained battery health state (SOH) prediction model, and the predicted SOH value is output. The deep learning data completion model includes: An embedding layer is used to map the position encoding corresponding to the SOC interval to a first position vector; A fully connected layer is used to linearly transform the feature data into a feature vector; A fusion layer is used to add the first position vector to the feature vector to obtain a fusion vector; The Transformer layer is used to perform temporal modeling on the fused vector and output contextual features; The output layer is used to output the completed feature data based on the context features.

2. The method of claim 1, wherein, Following the output of the predicted SOH value, the following is also included: The linear regression coefficients of the interval characteristics of each SOC interval and SOH are calculated based on the offline experimental data, and the linear regression coefficients are normalized to obtain the weight coefficients of each SOC interval. The predicted SOH value is adaptively corrected based on the weighting coefficients of each SOC interval to obtain the corrected SOH estimate.

3. The method of claim 2, wherein, Also includes: The observed energy of each SOC interval is extracted from the complete feature data, the reference energy of each SOC interval of the current battery is obtained from the offline experimental data, and the correction coefficient of the current battery is calculated based on the observed energy and reference energy of each SOC interval. The predicted SOH value is adaptively corrected based on the correction coefficient and the weight coefficient of each SOC interval to obtain the corrected SOH estimate.

4. The method of claim 1, wherein, The deep learning data completion model is trained using a masking training method, which includes: in each training iteration, randomly masking all or part of the feature data in multiple SOC intervals, and using cross-validation to alternately mask different SOC intervals in different iterations, using the true value of the masked data as a supervision signal to train the model to complete the masked data.

5. The method of claim 1, wherein, The SOH prediction model includes: The state change sensing module is used to perform differential or comparative modeling of the feature data of adjacent SOC intervals in the complete feature data to extract the change patterns during the charging / discharging process. The feature encoding module is used to map the feature data of different physical quantities to a unified feature space; The attention module is used to assign weights to the changing features of different SOC intervals and to filter out changing patterns that are sensitive to SOH.

6. The method according to any one of claims 1 to 5, characterized in that, The time-series data also includes a cycle number, and the input of the deep learning data completion model also includes the position code corresponding to the cycle number. The embedding layer is further used to map the position code corresponding to the cycle number into a second position vector. The fusion layer is further used to add the first position vector, the second position vector and the feature vector to obtain the fusion vector.

7. An electronic device, comprising: The method includes at least one processor and a memory communicatively connected to the at least one processor, the memory storing instructions executable by the at least one processor to enable the at least one processor to perform the method as described in any one of claims 1 to 6.

8. A computer storage medium, characterized in that, The computer storage medium stores instructions or programs that, when executed by at least one processor, cause the at least one processor to perform the method as described in any one of claims 1 to 6.

Images