A method and system for error correction in joint evaluation of battery SOC and SOH

Abstract

This invention belongs to the field of data processing technology and relates to a method and system for error correction in the joint evaluation of battery SOC and SOH. The method includes: S1, preprocessing the acquired multi-source data; S2, constructing a joint SOH-SOC estimation model; S3, segmenting the preprocessed data from different time points to obtain first window data and second window data; S4, obtaining a fused SOC estimate based on the first and second window data using an ampere-hour integral algorithm and a quartile distance group detection algorithm; and S5, dynamically correcting the joint SOH-SOC estimation model using a gradient backpropagation mechanism based on the fused SOC estimate and the SOC value at the end of the second window data to obtain a corrected joint SOH-SOC estimation model. This invention solves the problems of low estimation accuracy and low reliability in existing methods by dynamically correcting the joint SOH-SOC estimation model.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically discloses a method and system for error correction in the joint evaluation of battery SOC and SOH. Background Technology

[0002] With the continuous increase in the growth rate of the lithium battery market, the failure and safety hazards of lithium batteries have attracted attention from all parties. In order to reduce the risks during the charging and discharging process, it is necessary to accurately monitor the battery's state of health (SOH) and state of charge (SOC) through a suitable battery management system (BMS).

[0003] In recent years, artificial intelligence (AI) technology has been widely applied in the field of lithium batteries and has made progress. However, there are still drawbacks in the estimation of state of health (SOH) and state of charge (SOC) using artificial intelligence. In order to capture the long-term dependence of time-series data, most estimation methods mainly use recurrent neural networks (RNN) and long short-term memory networks (LSTM) to build battery state estimation models. Some deep learning models, such as convolutional neural networks (CNN) and deep neural networks (DNN), are also widely used in battery management systems (BMS). These estimation methods can extract key features from battery data through powerful feature extraction capabilities, but they still face the following problems: (1) In practical applications, battery data often has variable length characteristics. Due to differences in sampling frequency or charge / discharge cycles, fixed-length input requirements may lose important information or introduce invalid padding, resulting in low estimation accuracy; (2) Most existing estimation methods focus on estimation and ignore error quantification, which leads to a significant decrease in estimation reliability in the later stages of battery aging or under extreme conditions; (3) Existing studies usually handle the estimation of state of health (SOH) and state of charge (SOC) independently, ignoring the coupling relationship between the two. However, in the actual battery degradation process, there is a complex nonlinear interaction between capacity decay and state of charge, resulting in low estimation accuracy and reliability. Therefore, there is an urgent need for a joint estimation method for the state of health (SOH) and the state of charge (SOC), and to study the error propagation mechanism and online correction method in order to establish a more robust intelligent BMS system. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for error correction in the joint evaluation of battery SOC and SOH, thereby solving the problems of low accuracy and low reliability in existing estimations.

[0005] The specific solution of the present invention is as follows:

[0006] A method for error correction in the joint evaluation of battery SOC and SOH includes the following steps:

[0007] S1. Preprocess the acquired multi-source data;

[0008] S2. Based on the preprocessed data, construct the SOH-SOC joint estimation model through a one-dimensional convolutional neural network layer, a spatial pyramid pooling layer, and a fully connected layer.

[0009] S3. Segment the preprocessed data from different time points to obtain the first window data and the second window data;

[0010] S4. Based on the first window data and the second window data, the fused SOC estimate is obtained through the ampere-hour integral algorithm and the interquartile range group detection algorithm.

[0011] S5. Based on the fused SOC estimate and the SOC value at the end of the second window data, the SOH-SOC joint estimation model is dynamically corrected through the gradient backpropagation mechanism to obtain the corrected SOH-SOC joint estimation model.

[0012] In some embodiments, preprocessing includes:

[0013] S11. Synchronous data is obtained by aligning the acquired multi-source data to a time scale using cubic spline interpolation.

[0014] S12. Set windows of different lengths, slide the synchronized data through windows of different lengths, and generate window data of different lengths.

[0015] S13. Increase the sampling frequency of low SOH data in the window data.

[0016] In some embodiments, step S2 includes:

[0017] Based on the preprocessed data, SOH estimation model and SOC estimation model are constructed by one-dimensional convolutional neural network layer, spatial pyramid pooling layer and fully connected layer, respectively.

[0018] A training dataset is constructed based on the preprocessed data. The SOH estimation model and the SOC estimation model are then concatenated and spliced ​​to generate an initial SOH-SOC joint estimation model. The initial SOH-SOC joint estimation model and the SOC estimation model are then trained using the training dataset to obtain the SOH-SOC joint estimation model.

[0019] In some embodiments, constructing the SOH estimation model and the SOC estimation model respectively includes:

[0020] The preprocessed data includes preprocessed voltage, current, temperature, SOH, and SOC. The training dataset includes an initial training set and an optimization dataset, both of which include voltage, current, temperature, SOH, and SOC.

[0021] Based on the preprocessed voltage, current, and temperature, an initial SOH estimation model with output SOH is constructed by sequentially connecting a one-dimensional convolutional neural network layer, a spatial pyramid pooling layer, and a fully connected layer. The initial SOH estimation model is then trained with voltage, current, temperature, and SOH from the initial training set and the optimization dataset to obtain the SOH estimation model.

[0022] Based on the preprocessed voltage, current, temperature, and SOH, an initial SOC estimation model with output SOC is constructed by sequentially connecting a one-dimensional convolutional neural network layer, a spatial pyramid pooling layer, and a fully connected layer. The initial SOC estimation model is then trained on the voltage, current, temperature, SOH, and SOC of the initial SOC estimation model using the initial training set and the optimization dataset to obtain the final SOC estimation model.

[0023] In some embodiments, the SOH estimation model uses the L2 loss function for model optimization. The expression for the L2 loss function is:

[0024]

[0025] Among them, L SOH Estimate the output value of the loss function of the SOH model; This is the SOH estimate calculated by the current SOH estimation model based on the i-th input data; SOH i This represents the actual SOH value corresponding to the i-th input data.

[0026] In some embodiments, the SOC estimation model uses the Huber loss function for model optimization. The expression for the Huber loss function is as follows:

[0027]

[0028] Among them, L SOC δ represents the output value of the loss function of the SOC estimation model; δ is the balancing weight. SOC is the estimated SOC value calculated by the current SOC estimation model based on the i-th input data; SOC i Let be the actual SOC value corresponding to the i-th input data.

[0029] In some embodiments, the SOH-SOC joint estimation model employs a weighted loss function for model optimization. The expression for the weighted loss function is as follows:

[0030] L=αL SOH +(1-α)L SOC ,

[0031] Where L is the output value of the weighted loss function; α is the hyperparameter; L SOH To estimate the output value of the loss function of the SOH model; LSOC The output value of the loss function of the SOC estimation model.

[0032] In some embodiments, step S4 includes:

[0033] Based on the first window data, SOC estimates for different end times are obtained using the SOH-SOC joint estimation model. For each start time based on the second window data, SOC estimates for the corresponding time are selected from the SOC estimates for different end times to obtain SOC estimates for different start times. Based on the second window data and the SOC estimates for different start times, the ampere-hour integral value for different start points and the same end point is obtained using the ampere-hour integral algorithm. Based on the ampere-hour integral values ​​for different start points and the same end point, the quartile distance group detection algorithm is used to obtain the fused SOC estimate.

[0034] In some embodiments, step S5 includes:

[0035] The error between the fused SOC estimate and the SOC value at the end of the second window data is calculated to obtain the IQR perception weight. The IQR perception weight is applied to the fully connected layers of the SOH estimation model and the SOC estimation model in the SOH-SOC joint estimation model, respectively. The error result is obtained based on the weighted loss function of the SOH-SOC joint estimation model. The error result is backpropagated to the SOH-SOC joint estimation model in the form of gradient. The fully connected layers of the SOH estimation model and the SOC estimation model in the SOH-SOC joint estimation model are optimized through the error result to obtain the corrected SOH-SOC joint estimation model.

[0036] This invention also relates to a battery SOC and SOH joint evaluation error correction system, comprising:

[0037] The multi-source data acquisition module is used to collect voltage, current, temperature, SOC, and SOH data from multiple batteries under various operating conditions.

[0038] The preprocessing module is used to perform spatiotemporal alignment, window data processing, and data layering processing on the collected multi-source data.

[0039] The SOH estimation model building module is used to build an SOH estimation model based on preprocessed voltage, current and temperature through a one-dimensional convolutional neural network layer, a spatial pyramid pooling layer and a fully connected layer.

[0040] The SOC estimation model building module is used to build an SOC estimation model based on preprocessed voltage, current, temperature and SOH through a one-dimensional convolutional neural network layer, a spatial pyramid pooling layer and a fully connected layer.

[0041] The SOH-SOC joint estimation model construction module is used to construct the SOH-SOC joint estimation model based on the SOH estimation model and the SOC estimation model.

[0042] The dual-window segmentation module is used to segment the preprocessed data from different time points to obtain the first window data and the second window data;

[0043] The ampere-time-IQR fusion module is used to obtain the fused SOC estimate based on the first window data and the second window data through the ampere-time integral algorithm and the interquartile range group detection algorithm.

[0044] The dynamic correction module is used to dynamically correct the SOH-SOC joint estimation model based on the fused SOC estimate and the SOC value at the end of the second window data through the gradient backpropagation mechanism, so as to obtain the corrected SOH-SOC joint estimation model.

[0045] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0046] 1. This invention constructs SOH and SOC estimation models from preprocessed multi-source data, respectively, and then concatenates the SOH and SOC estimation models to form a joint SOH-SOC estimation model. First and second window data are obtained from the preprocessed multi-source data. Based on the first and second window data, a fused SOC estimate is obtained using the ampere-hour integral algorithm and the interquartile range group detection algorithm. The joint SOH-SOC estimation model is dynamically corrected using the fused SOC estimate and the SOC value at the end of the second window data through a gradient backpropagation mechanism, resulting in a corrected SOH-SOC joint estimation model. Using the corrected SOH-SOC joint estimation model for SOH and SOC estimation improves estimation accuracy and reliability while reducing estimation errors.

[0047] 2. Lightweight error quantification is performed by sensing weights through IQR. Based on the error results, the SOH-SOC joint estimation model is dynamically corrected through gradient backpropagation mechanism, which can effectively improve the estimation reliability under complex extreme conditions and in the later stages of battery aging.

[0048] 3. By dynamically correcting the SOH-SOC joint estimation model using first and second window data obtained based on the variable length characteristics of battery data, the model can adapt to various operating conditions with different sampling frequencies or different charge and discharge cycles in practical applications, thereby improving the estimation accuracy.

[0049] 4. By using one-dimensional convolutional neural networks and spatial pyramid pooling, the dynamic characteristics of multi-source data such as voltage, current, and temperature are captured. This fully leverages the advantages of local feature extraction and computational efficiency, improving estimation accuracy and providing stronger support for efficient BMS evaluation. Attached Figure Description

[0050] Figure 1 This is a flowchart of a battery SOC and SOH joint evaluation error correction method in an embodiment of the present invention.

[0051] Figure 2 This is a block diagram of the corrected SOH-SOC joint estimation model in an embodiment of the present invention.

[0052] Figure 3 This is a block diagram of a battery SOC and SOH joint evaluation error correction system according to an embodiment of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0054] A method for error correction in the joint evaluation of battery SOC and SOH, such as Figure 1 As shown, it includes the following steps:

[0055] S1. Acquire multi-source data and preprocess the multi-source data;

[0056] Data such as voltage, current, temperature, SOC, and SOH of multiple batteries under various operating conditions are collected. These data are time series data, and the various operating conditions include constant current charge and discharge and non-constant current charge and discharge.

[0057] Preprocessing includes:

[0058] S11. Spatiotemporal alignment processing: The acquired multi-source data is aligned to the time scale using the cubic spline interpolation method to obtain synchronized data, namely voltage, current, temperature, SOC, SOH and other data with the same timestamp.

[0059] S12. Window Data Processing: Set windows of different lengths, slide windows of different lengths on the synchronous data, and trim windows of different lengths to generate window data of different lengths. Window data of different lengths includes voltage, current, and temperature data of different lengths, as well as SOC and SOH at the last moment of the window. The window length can be 20-40.

[0060] S13. Data Stratification: Increase the sampling frequency of low SOH data in the window data, i.e., increase the amount of low SOH data, while satisfying the following ratio: Low SOH data sampling frequency : High SOH data sampling frequency = High SOH data amount : Low SOH data amount. Since the amount of low SOH data is much smaller than the amount of high SOH data, it may have a significant impact on the training and performance of the estimation model. Therefore, increasing the sampling frequency of low SOH data can alleviate the problem of data imbalance. Here, low SOH data refers to data with an SOH value less than 80%, and the amount of low SOH data refers to the quantity of low SOH data. High SOH data refers to data with an SOH value greater than or equal to 80%, and the amount of high SOH data refers to the quantity of high SOH data.

[0061] S2. Construct an SOH estimation model based on the preprocessed data;

[0062] An initial training set is constructed based on the voltage, current, temperature, SOC, and SOH data under preprocessed constant current charging and discharging conditions. An optimization dataset is constructed based on the voltage, current, temperature, SOC, and SOH data under preprocessed non-constant current charging and discharging conditions.

[0063] Based on the preprocessed voltage, current, and temperature, an initial SOH estimation model with output SOH is constructed by sequentially connecting a one-dimensional convolutional neural network layer, a spatial pyramid pooling layer, and a fully connected layer. The initial SOH estimation model is then trained sequentially using voltage, current, temperature, and SOH from the initial training set and the tuning dataset to obtain the final SOH estimation model.

[0064] By using one-dimensional convolutional neural networks and spatial pyramid pooling, the dynamic characteristics of multi-source data such as voltage, current, and temperature are captured. This fully leverages the advantages of local feature extraction and computational efficiency, improving estimation accuracy and providing stronger support for efficient BMS evaluation.

[0065] The calculation formula for the SOH estimation model is as follows:

[0066] F conv =ReLU(W conv *X)

[0067] F spp =SPP(F conv ) = [Pool4(F conv )||Pool2(F conv )||Pool1(F conv )]

[0068] SOH = sigmoid(W fc F spp +b)

[0069] Among them, Fconv W is the output of the one-dimensional convolutional neural network layer in the SOH estimation model. conv For the SOH estimation model, X represents the parameters of the one-dimensional convolutional neural network layer, where X is the preprocessed voltage U, current I, and temperature T; F spp This is the output of the spatial pyramid pooling layer in the SOH estimation model. The spatial pyramid pooling layer is formed by splicing together multiple pooling layers. l (F conv ) represents the output of the l-th pooling layer in the SOH estimation model, where l = 1, 2, 4; SOH represents the output of the fully connected layer in the SOH estimation model, i.e., the output of the SOH estimation model itself. fc denoted as , where b is the parameter of the fully connected layer in the SOH estimation model, and b is the bias of the fully connected layer in the SOH estimation model.

[0070] The SOH estimation model uses the L2 loss function for model optimization. The expression for the L2 loss function is as follows:

[0071]

[0072] Among them, L SOH This is the output value of the L2 loss function, which is the output value of the loss function of the SOH estimation model; This is the SOH estimate calculated by the current SOH estimation model based on the i-th input data; SOH i This represents the actual SOH value corresponding to the i-th input data.

[0073] S3. Construct an SOC estimation model based on the preprocessed data;

[0074] Based on the preprocessed voltage, current, temperature, and SOH, an initial SOC estimation model with output SOC is constructed by sequentially connecting a one-dimensional convolutional neural network layer, a spatial pyramid pooling layer, and a fully connected layer. The initial SOC estimation model is then trained sequentially using voltage, current, temperature, SOH, and SOC from the initial training set and the tuning dataset to obtain the final SOC estimation model.

[0075] The formula for calculating the SOC estimation model is as follows:

[0076] F conv2 =ReLU(W conv2 *[X,SOH])

[0077] F spp2 =SPP2(F conv2 ) = [Pool4(F conv2 )||Pool2(F conv2 )||Pool1(F conv2 )]

[0078] SOC = Sigmoid(W fc2 F spp2 +b2),

[0079] Among them, F conv2 W is the output of the one-dimensional convolutional neural network layer in the SOC estimation model. conv2 Here, X represents the preprocessed voltage U, current I, and temperature T, and SOH represents the preprocessed SOH in the SOC estimation model; F spp2 This refers to the output of the spatial pyramid pooling layer in the SOC estimation model. The spatial pyramid pooling layer is formed by splicing together multiple pooling layers. l (F conv2 W represents the output of the l-th pooling layer in the SOC estimation model, where l = 1, 2, 4; SOC represents the output of the fully connected layer in the SOC estimation model, i.e., the output of the SOC estimation model itself. fc2 b1 represents the parameters of the fully connected layer in the SOC estimation model, and b2 represents the bias of the fully connected layer in the SOC estimation model.

[0080] The SOC estimation model uses the Huber loss function for model optimization. The expression for the Huber loss function is as follows:

[0081]

[0082] Among them, L SOC δ is the output value of the Huber loss function, which is the output value of the loss function of the SOC estimation model; δ is the balancing weight, which is set to 0.05 to balance the smoothness of the Huber loss function and its sensitivity to outliers. SOC is the estimated SOC value calculated by the current SOC estimation model based on the i-th input data; SOC i Let be the actual SOC value corresponding to the i-th input data.

[0083] S4. Construct a joint SOH-SOC estimation model based on the SOH estimation model and the SOC estimation model;

[0084] The SOH estimation model and the SOC estimation model are concatenated and spliced ​​together, with the output of the SOH estimation model used as the input of the SOC estimation model to generate an initial SOH-SOC joint estimation model. The initial SOH-SOC joint estimation model is then trained sequentially using the initial training set and the optimization dataset to fine-tune the parameters of the SOC estimation model, thus obtaining the SOH-SOC joint estimation model.

[0085] The SOH-SOC joint estimation model, constructed by using the output of the SOH estimation model as the input of the SOC estimation model, can solve the problem of the interaction between capacity decay and state of charge, as well as the problem of variable length characteristics of battery data. It can adapt to various operating conditions with different sampling frequencies or different charge and discharge cycles in practical applications, thereby improving the estimation accuracy.

[0086] The SOH-SOC joint estimation model uses a weighted loss function for model optimization. The expression for the weighted loss function is as follows:

[0087] L=αL SOH +(1-α)L SOC ,

[0088] Where L is the output value of the weighted loss function; α is a hyperparameter, set to 0.5, used to balance the relative learning degree of the loss functions of the SOH estimation model and the SOC estimation model; L SOH To estimate the output value of the loss function of the SOH model; L SOC The output value of the loss function of the SOC estimation model.

[0089] S5. Segment the preprocessed data from different time points to obtain the first window data and the second window data;

[0090] The preprocessed data is segmented at different time points to obtain a first window of data that starts at the same time but ends at different times, and a second window of data that starts at different times but ends at the same time.

[0091] S6. Based on the first window data and the second window data, the fused SOC estimate is obtained through the ampere-hour integral algorithm and the interquartile range group detection algorithm.

[0092] Based on the first window data, the SOC estimates for different end times are obtained through the SOH-SOC joint estimation model. Different end times refer to any time t. The first window data is input into the SOH-SOC joint estimation model to first obtain the SOH estimate at time t, and then the SOC estimate at time t is obtained based on the SOH estimate at time t.

[0093] For each start time of the second window data, the SOC estimate for the corresponding time is selected from the SOC estimates for different end times to obtain the SOC estimates for different start times, i.e., the initial SOC values. Based on the second window data and the SOC estimates for different start times, the ampere-hour integration algorithm is used to obtain the ampere-hour integration values ​​for different start points and the same end point, i.e., the ampere-hour integration is performed on any data in the second window data to obtain the ampere-hour integration value. Based on the ampere-hour integration values ​​for different start points and the same end point, the interquartile range group detection algorithm is used to find the upper quartile and the lower quartile, and the mean of the interquartile range between the upper quartile and the lower quartile is calculated to obtain the fused SOC estimate.

[0094] The formula for calculating the integral value of ampere-hours is:

[0095]

[0096] Among them, f Ah (w) represents the ampere-hour integral value of any data point w in the second window, C max I represents the maximum usable capacity of the battery, t represents the current, t represents the time series of data w, and SOC0 represents the initial SOC value.

[0097] The formula for calculating the SOC estimate at different start times is as follows:

[0098] SOC Ah =[…,f Ah (w),...],

[0099] Among them, SOC Ah f represents the SOC estimate for the same end time calculated from data with different start times. Ah (w) is the ampere-hour integral value of any data w in the second window data.

[0100] The formula for calculating the fused SOC estimate is as follows:

[0101]

[0102] Among them, SOC IQR To integrate the SOC estimates, Q3 represents the upper four-digit score, and Q1 represents the lower four-digit score. Ah (i) represents the SOC estimate at the i-th starting time, and n IQR denoted as the number of non-discrete points, and k is a hyperparameter, k = 0.1.

[0103] S7. Based on the fused SOC estimate and the SOC value at the end of the second window data, the SOH-SOC joint estimation model is dynamically corrected through the gradient backpropagation mechanism to obtain the corrected SOH-SOC joint estimation model.

[0104] The error between the fused SOC estimate and the SOC value at the end of the second window data is calculated to obtain the IQR sensing weight. This IQR sensing weight is then applied to the fully connected layers of both the SOH and SOC estimation models in the SOH-SOC joint estimation model. The error result is obtained based on the weighted loss function of the SOH-SOC joint estimation model. This error result is backpropagated to the SOH-SOC joint estimation model in the form of a gradient. The fully connected layers of both the SOH and SOC estimation models are then optimized using the error result to obtain the corrected SOH-SOC joint estimation model. Figure 2 As shown.

[0105] By dynamically correcting the SOH-SOC joint estimation model using first and second window data obtained based on the variable length characteristics of battery data, it can adapt to various operating conditions with different sampling frequencies or different charge and discharge cycles in practical applications, thereby improving the estimation accuracy.

[0106] Lightweight error quantification is achieved by using IQR-sensing weights, and the SOH-SOC joint estimation model is dynamically corrected based on the error results through gradient backpropagation mechanism. This can effectively improve the estimation reliability under complex extreme conditions and in the later stages of battery aging.

[0107] The formula for calculating IQR perceived weights is:

[0108]

[0109] Among them, SOC IQR ε is the SOC estimate for fusion; ε is the Lapss smoothing term, used to ensure that the denominator is not zero and to avoid numerical instability; SOC is the SOC value at the end of the second window of data.

[0110] The optimization formula for the fully connected layer of the SOH estimation model in the SOH-SOC joint estimation model is:

[0111]

[0112] Among them, W fc * For the fully connected layer of the optimized SOH estimation model, W fc The fully connected layer of the SOH estimation model before optimization. This is a fully connected layer of the SOH estimation model with IQR-perceived weights applied, where η is the learning rate, typically set to 0.001.

[0113]

[0114] Among them, Wfc2 * For the fully connected layer of the optimized SOC estimation model, W fc2 The fully connected layer of the SOC estimation model before optimization. This is a fully connected layer of the SOC estimation model with IQR-perceived weights applied, where η is the learning rate, typically set to 0.001.

[0115] The calculation formula for the corrected SOH-SOC joint estimation model is as follows:

[0116] SOH = sigmoid(W fc * F spp +b)

[0117] SOC = sigmoid(W fc2 * F spp2 +b2),

[0118] Among them, W fc * To correct the fully connected layer of the SOH estimation model in the SOH-SOC joint estimation model, F spp To correct the output of the spatial pyramid pooling layer in the SOH-SOC joint estimation model, b is the bias of the fully connected layer in the SOH-SOC joint estimation model; W fc2 * To correct the fully connected layer of the SOC estimation model in the SOH-SOC joint estimation model, F spp2 b2 is used to correct the output of the spatial pyramid pooling layer in the SOH-SOC joint estimation model, and b2 is used to correct the bias of the fully connected layer in the SOH-SOC joint estimation model.

[0119] SOH and SOC estimation models are constructed separately using preprocessed multi-source data. These models are then concatenated to form a joint SOH-SOC estimation model. First and second window data are obtained from the preprocessed multi-source data. Based on these data, a fused SOC estimate is obtained using the ampere-hour integral algorithm and the interquartile range group detection algorithm. The joint SOH-SOC estimation model is dynamically corrected using the fused SOC estimate and the SOC value at the end of the second window data through gradient backpropagation, resulting in a corrected SOH-SOC joint estimation model. Using this corrected model for SOH and SOC estimation improves estimation accuracy and reliability while reducing estimation errors.

[0120] This invention also relates to a battery SOC and SOH joint evaluation error correction system, such as... Figure 3 As shown, it includes:

[0121] The multi-source data acquisition module is used to collect time-series data such as voltage, current, temperature, SOC, and SOH of multiple batteries under various operating conditions, including constant current charge and discharge and non-constant current charge and discharge.

[0122] The preprocessing module performs spatiotemporal alignment, window data processing, and data layering on the acquired multi-source data. The preprocessing module includes a spatiotemporal alignment unit, a window data processing unit, and a data layering unit. The spatiotemporal alignment unit uses cubic spline interpolation to align the acquired multi-source data to a time scale to obtain synchronized data. The window data processing unit sets windows of different lengths and slides these windows across the synchronized data to generate window data of varying lengths. The data layering unit increases the sampling frequency of low-SOH data within the window data, ensuring that the ratio of low-SOH data to high-SOH data in the window data reaches a preset ratio.

[0123] The SOH estimation model building module is used to build an SOH estimation model based on preprocessed voltage, current and temperature through a one-dimensional convolutional neural network layer, a spatial pyramid pooling layer and a fully connected layer.

[0124] The SOC estimation model building module is used to build an SOC estimation model based on preprocessed voltage, current, temperature and SOH through a one-dimensional convolutional neural network layer, a spatial pyramid pooling layer and a fully connected layer.

[0125] The SOH-SOC joint estimation model construction module is used to construct the SOH-SOC joint estimation model based on the SOH estimation model and the SOC estimation model.

[0126] The dual-window segmentation module is used to segment the preprocessed data from different time points to obtain the first window data and the second window data;

[0127] The ampere-time-IQR fusion module is used to obtain the fused SOC estimate based on the first window data and the second window data through the ampere-time integral algorithm and the interquartile range group detection algorithm.

[0128] The dynamic correction module is used to dynamically correct the SOH-SOC joint estimation model based on the fused SOC estimate and the SOC value at the end of the second window data through the gradient backpropagation mechanism, so as to obtain the corrected SOH-SOC joint estimation model.

[0129] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for error correction in the joint evaluation of battery SOC and SOH, characterized in that, Includes the following steps: S1. Preprocess the acquired multi-source data; S2. Based on the preprocessed data, construct a joint SOH-SOC estimation model using a one-dimensional convolutional neural network layer, a spatial pyramid pooling layer, and a fully connected layer; specifically including: Based on the preprocessed data, SOH estimation model and SOC estimation model are constructed by one-dimensional convolutional neural network layer, spatial pyramid pooling layer and fully connected layer, respectively. A training dataset is constructed based on the preprocessed data. The SOH estimation model and the SOC estimation model are then concatenated and spliced ​​to generate an initial SOH-SOC joint estimation model. The initial SOH-SOC joint estimation model is then trained using the training dataset to obtain the SOH-SOC joint estimation model. S3. Segment the preprocessed data from different time points to obtain the first window data and the second window data; S4. Based on the first window data and the second window data, the fused SOC estimate is obtained through the ampere-hour integral algorithm and the interquartile range group detection algorithm. S5. Based on the fused SOC estimate and the SOC value at the end of the second window data, the SOH-SOC joint estimation model is dynamically corrected using a gradient backpropagation mechanism to obtain the corrected SOH-SOC joint estimation model; specifically including: The error between the fused SOC estimate and the SOC value at the end of the second window data is calculated to obtain the IQR perception weight. The IQR perception weight is applied to the fully connected layers of the SOH estimation model and the SOC estimation model in the SOH-SOC joint estimation model, respectively. The error result is obtained based on the weighted loss function of the SOH-SOC joint estimation model. The error result is backpropagated to the SOH-SOC joint estimation model in the form of gradient. The fully connected layers of the SOH estimation model and the SOC estimation model in the SOH-SOC joint estimation model are optimized through the error result to obtain the corrected SOH-SOC joint estimation model.

2. The method for error correction in joint evaluation of battery SOC and SOH according to claim 1, characterized in that, The preprocessing includes: S11. Synchronous data is obtained by aligning the acquired multi-source data to a time scale using cubic spline interpolation. S12. Set windows of different lengths, slide the synchronized data through windows of different lengths, and generate window data of different lengths. S13. Increase the sampling frequency of low SOH data in the window data.

3. The method for joint evaluation error correction of battery SOC and SOH according to claim 1, characterized in that, The construction of the SOH estimation model and the SOC estimation model respectively includes: The preprocessed data includes preprocessed voltage, current, temperature, SOH, and SOC. The training dataset includes an initial training set and an optimization dataset, both of which include voltage, current, temperature, SOH, and SOC. Based on the preprocessed voltage, current, and temperature, an initial SOH estimation model with output SOH is constructed by sequentially connecting a one-dimensional convolutional neural network layer, a spatial pyramid pooling layer, and a fully connected layer. The initial SOH estimation model is then trained with voltage, current, temperature, and SOH from the initial training set and the optimization dataset to obtain the SOH estimation model. Based on the preprocessed voltage, current, temperature, and SOH, an initial SOC estimation model with output SOC is constructed by sequentially connecting a one-dimensional convolutional neural network layer, a spatial pyramid pooling layer, and a fully connected layer. The initial SOC estimation model is then trained sequentially using voltage, current, temperature, SOH, and SOC from the initial training set and the tuning dataset to obtain the final SOC estimation model.

4. The method for joint evaluation error correction of battery SOC and SOH according to claim 3, characterized in that, The SOH estimation model is optimized using the L2 loss function, the expression of which is: ， in, L SOH Estimate the output value of the loss function of the SOH model; This is the SOH estimate calculated by the current SOH estimation model based on the i-th input data. This represents the actual SOH value corresponding to the i-th input data.

5. The method for joint evaluation error correction of battery SOC and SOH according to claim 4, characterized in that, The SOC estimation model uses the Huber loss function for model optimization. The expression of the Huber loss function is as follows: ， in, L SOC To estimate the output value of the loss function of the SOC model; δ To balance the weights; This is the SOC estimate calculated by the current SOC estimation model based on the i-th input data. Let be the actual SOC value corresponding to the i-th input data.

6. The method for joint evaluation error correction of battery SOC and SOH according to claim 5, characterized in that, The SOH-SOC joint estimation model uses a weighted loss function for model optimization. The expression of the weighted loss function is as follows: ， in, L This is the output value of the weighted loss function; α For hyperparameters; L SOH Estimate the output value of the loss function of the SOH model; L SOC The output value of the loss function of the SOC estimation model.

7. The method for joint evaluation error correction of battery SOC and SOH according to claim 1, characterized in that, Step S4 includes: Based on the first window data, SOC estimates for different end times are obtained using the SOH-SOC joint estimation model. For each start time based on the second window data, SOC estimates for the corresponding time are selected from the SOC estimates for different end times to obtain SOC estimates for different start times. Based on the second window data and the SOC estimates for different start times, the ampere-hour integral value for different start points and the same end point is obtained using the ampere-hour integral algorithm. Based on the ampere-hour integral values ​​for different start points and the same end point, the quartile distance group detection algorithm is used to obtain the fused SOC estimate.

8. A battery SOC and SOH joint evaluation error correction system, characterized in that, include: The multi-source data acquisition module is used to collect voltage, current, temperature, SOC, and SOH data from multiple batteries under various operating conditions. The preprocessing module is used to perform spatiotemporal alignment, window data processing, and data layering processing on the collected multi-source data. The SOH estimation model construction module is used to construct an SOH estimation model based on preprocessed voltage, current, and temperature through a one-dimensional convolutional neural network layer, a spatial pyramid pooling layer, and a fully connected layer; specifically, it includes: Based on the preprocessed data, SOH estimation model and SOC estimation model are constructed by one-dimensional convolutional neural network layer, spatial pyramid pooling layer and fully connected layer, respectively. A training dataset is constructed based on the preprocessed data. The SOH estimation model and the SOC estimation model are then concatenated and spliced ​​to generate an initial SOH-SOC joint estimation model. The initial SOH-SOC joint estimation model is then trained using the training dataset to obtain the SOH-SOC joint estimation model. The SOC estimation model building module is used to build an SOC estimation model based on preprocessed voltage, current, temperature and SOH through a one-dimensional convolutional neural network layer, a spatial pyramid pooling layer and a fully connected layer. The SOH-SOC joint estimation model construction module is used to construct the SOH-SOC joint estimation model based on the SOH estimation model and the SOC estimation model. The dual-window segmentation module is used to segment the preprocessed data from different time points to obtain the first window data and the second window data; The ampere-time-IQR fusion module is used to obtain the fused SOC estimate based on the first window data and the second window data through the ampere-time integral algorithm and the interquartile range group detection algorithm. The dynamic correction module is used to dynamically correct the SOH-SOC joint estimation model based on the fused SOC estimate and the SOC value at the end of the second window data, through a gradient backpropagation mechanism, to obtain the corrected SOH-SOC joint estimation model; specifically, it includes: The error between the fused SOC estimate and the SOC value at the end of the second window data is calculated to obtain the IQR perception weight. The IQR perception weight is applied to the fully connected layers of the SOH estimation model and the SOC estimation model in the SOH-SOC joint estimation model, respectively. The error result is obtained based on the weighted loss function of the SOH-SOC joint estimation model. The error result is backpropagated to the SOH-SOC joint estimation model in the form of gradient. The fully connected layers of the SOH estimation model and the SOC estimation model in the SOH-SOC joint estimation model are optimized through the error result to obtain the corrected SOH-SOC joint estimation model.

Images