Battery multi-scene state of health estimation method based on domain adaptation migration

Abstract

The application discloses a kind of based on domain self-adapting migration's battery multi-scene health state estimation method, comprising steps 1, deduce the health state quantity of battery;Step 2, filter out and battery health state related coefficient absolute value is greater than or equal to the feature of set threshold as model input;Step 3, source domain data and target domain data are extracted;Step 4, calculate the distribution distance of source domain data and target domain data in the high-level feature output of convolutional neural network model, and the total loss function is obtained by weighted summation to the loss function and mean square error loss function, the total loss function is optimized by adaptive matrix estimation optimizer, realize the domain self-adapting training of unlabeled battery data;Step 5, the performance of model is evaluated using validation set, and it is applied to test set to output health state estimation result.The application effectively solves the estimation precision decline problem caused by data distribution difference under cross-domain scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of battery state of health estimation, in particular to a battery multi-scene state of health estimation method based on domain adaptation transfer. BACKGROUND

[0002] Batteries are widely used in electric vehicles, energy storage and portable devices, and accurate estimation of their state of health is crucial for safety and longevity. Traditional state of health estimation methods are mainly based on laboratory data, but in practical applications, factors such as temperature, humidity and charging strategies cause the battery degradation process to be nonlinear and scene-dependent, resulting in differences in data distribution and affecting the model's cross-scene generalization ability. Although transfer learning and domain adaptation techniques provide a way to address data distribution bias, existing methods still have shortcomings in feature alignment and temporal information mining, making it difficult to meet the demand for accurate estimation in complex scenarios. Therefore, there is an urgent need for a multi-scene state of health estimation method based on domain adaptation transfer to improve the adaptability and estimation accuracy of batteries in diverse scenarios. SUMMARY

[0003] The present application aims to overcome the shortcomings of existing battery state of health estimation methods, which have large errors in cross-scene use and cannot adapt to the use requirements of multiple scenarios. The present application provides a battery multi-scene state of health estimation method based on domain adaptation transfer, which uses source domain data for pre-training and transfers to the target domain to improve the estimation accuracy of the model on samples with significant data distribution differences. The method effectively utilizes unlabeled data through a convolutional neural network-multi-kernel maximum mean discrepancy (CNN-MKMMD) to improve the model's adaptability to complex scenarios and reduce data costs. This method can achieve high-precision state of health estimation with a small amount of data and effectively handle data distribution differences, with strong generalization ability.

[0004] The present application aims to achieve the following technical solutions:

[0005] The battery multi-scene state of health estimation method based on domain adaptation transfer comprises the following steps:

[0006] Step 1: Collect the dynamic process parameter signals of the monitored battery during the charging and discharging phase, and collect the state parameters reflecting the state of health during the battery aging phase, and derive the state of health of the battery;

[0007] Step 2: Extract the health factors representing battery degradation from the charging and discharging data and electrochemical impedance spectroscopy analysis results; perform Pearson correlation coefficient analysis on the extracted health factors, and select features with an absolute value of the correlation coefficient with the battery state of health greater than or equal to a set threshold as the model input;

[0008] Step 3, a convolutional neural network based on residual structure is constructed as a high-dimensional feature extractor of the model, mean square error is selected as the optimization function, and source domain data and target domain data are subjected to feature extraction;

[0009] Step 4, a multi-core maximum mean difference loss function is introduced, the distribution distance of the source domain data and the target domain data in the high-level feature output of the convolutional neural network model is calculated, and the loss function and the mean square error loss function are summed by weighting to obtain a total loss function, an adaptive matrix estimation optimizer is applied to optimize the total loss function, and domain adaptive training of the unlabeled battery data is realized;

[0010] Step 5, the model performance is evaluated using the validation set, and it is applied to the test set to output the state of health estimation result, and the state of health estimation of the storage battery in the cross-domain scene is realized.

[0011] As preferred, in step 1, the dynamic process parameter signal includes: current signal , charging voltage signal and charging temperature signal .

[0012] As preferred, the state of health quantity is:

[0013]

[0014] wherein, represents the maximum discharge capacity of the battery in each cycle during the aging process, represents the current cycle number, is the nominal capacity of the battery.

[0015] As preferred, step 2 specifically includes the following sub-steps:

[0016] Step 2a, the isochronous voltage drop time interval of the constant current charging stage voltage curve, the time integral of the constant voltage charging stage current decay curve, and the temperature change rate in the charging and discharging cycle are extracted from the charging and discharging data; at the same time, the system characteristic parameters extracted from the electrochemical impedance spectrum analysis results are combined to form an initial health factor;

[0017] Step 2b, the linear correlation between the initial health factor and the state of health of the battery is calculated by the Pearson correlation coefficient, and the features with an absolute value of the correlation coefficient greater than or equal to a set threshold value are selected as the model input;

[0018] Step 2c, the processed data is divided into a training set, a validation set and a test set, and the labels are the corresponding state of health sequences.

[0019] As preferred, in step 3, the convolutional neural network includes:

[0020] The input layer receives time sequence health factor data; the residual convolution module is composed of a plurality of one-dimensional convolution layers, each layer comprising a convolution kernel, a batch normalization layer, an activation function and a residual connection; the pooling layer accesses a global average pooling layer after the last convolution layer, and outputs a high-dimensional feature vector; and the full connection layer adopts a linear activation function to map the features after the pooling to a health state prediction value.

[0021] Preferably, in step 4, the multi-core maximum mean difference loss function comprises a regression loss and a distribution alignment loss.

[0022] Preferably, step 5 is specifically as follows:

[0023] The mean absolute error and the root mean square error are calculated on the validation set, and the optimal model parameters are selected; the trained model is applied to the target domain unlabeled data to output health state estimation values, and health state prediction in the cross-condition and cross-batch scene is realized.

[0024] A computer storage medium, the storage medium comprising computer instructions, when running on a computer, make the computer execute the above-mentioned method.

[0025] An electronic device, the electronic device comprising: a memory, a processor and a computer program stored on the memory and executable on the processor; wherein the processor implements the above-mentioned method when executing the program.

[0026] The beneficial effects of the present application are: the innovative method of the convolutional neural network-multi-core maximum mean difference model proposed by the present application effectively solves the problem of estimation accuracy decline caused by data distribution difference in cross-domain scene; through pre-training of source domain data and effective use of unlabeled data, the dependence on a large amount of labeled data is significantly reduced; the present application can still maintain high-precision health state estimation under the condition of a small amount of samples, and has strong practical application value. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is a flowchart of the present application;

[0028] Figure 2 is a voltage and current data curve graph during the charging and discharging process of the storage battery of the present application;

[0029] Figure 3 is a health state degradation curve graph under the charging and discharging strategy of the present application;

[0030] Figure 4 is a schematic diagram of the neural network structure of the present application;

[0031] Figure 5 and Figure 6 is an error analysis graph of the experimental estimation value; DETAILED DESCRIPTION

[0032] Example implementations are now described with reference to the drawings; however, these descriptions are not intended to limit the scope of the application, but are intended to provide example examples, which, while possibly being the best modes of practicing the application at the time, were developed using design constraints that can have subsequently been loosened or eliminated. It should be appreciated that specific examples can result in less than optimal designs, but which will be understood to reside within the scope of the application.

[0033] Furthermore, described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a thorough understanding of embodiments of the application. One skilled in the relevant art will recognize, however, that the application can be practiced without one or more of the specific details, or with other methods, components, materials, and so forth. In other instances, well-known structures, devices, and so forth are shown in block diagram form or are not shown or described in order to avoid obscuring the application.

[0034] The block diagrams in the drawings show functions and functionality, which can be implemented in software executed by digital hardware, such as a general purpose computer, special purpose computer, ASIC, etc. The software implementations can be downloaded to the hardware in one or more ways. Alternatively, these functions and functionality can be implemented in hardware, such as integrated circuits, ASICs, etc.

[0035] The flow diagrams depicted herein are examples of sequences of operations that can be performed, for example, by a computer system. The depicted examples are not meant to limit the scope of the application, but are merely intended to provide examples of sequences of operations that can be performed by a computer system. In some examples, the operations can be performed in a different order, or additional or fewer operations can be performed.

[0036] Embodiments:

[0037] A battery multi-scene state of health estimation method based on domain adaptation migration, as shown in Figure 1 , includes the following steps:

[0038] S100: Real-time monitoring of dynamic process parameter signals such as voltage, current and temperature of the battery in different working conditions, synchronous acquisition of the maximum discharge capacity of the battery in the aging stage reflecting the state of health, and further derivation of the state of health of the battery;

[0039] As shown in Figures 2-3 , in this step, the dynamic process parameter signals include current signals , voltage signals and temperature signals recorded during the charging and discharging process, the lengths of which are all .

[0040] use This represents the quantity signal during the charging and discharging process, where Indicates the current loop number, then The maximum discharge capacity of the battery per cycle during the aging process is... Furthermore, the battery's health status value is expressed as:

[0041]

[0042] in, This refers to the battery's nominal capacity.

[0043] During signal acquisition, the process quantity signals acquired in different cycles are different. As the battery ages, the maximum discharge capacity of the battery in different processes will gradually decrease.

[0044] S200: Extract the constant voltage drop time interval of the voltage curve during the constant current charging stage, the time integral of the current decay curve during the constant voltage charging stage, and the temperature change rate during the charge-discharge cycle from the charge-discharge data; at the same time, combine the system characteristic parameters extracted from the electrochemical impedance spectroscopy analysis results to form an initial health factor; perform Pearson correlation coefficient analysis on the extracted health factor, and select features with an absolute value of correlation coefficient ≥0.6 with the battery health state as model input;

[0045] In this step, the raw charge-discharge data is first processed and visualized to extract indirect health factors that can characterize capacity degradation, such as characteristic sequences of current, voltage, temperature, and time at different stages. Then, system characteristic parameters are calculated from the electrochemical impedance spectroscopy analysis results. Finally, the Pearson correlation coefficients between these health factors and the battery's health state are calculated. The formula for calculating the correlation coefficients can be expressed as:

[0046]

[0047] in, and They represent the first The values ​​of a certain feature and health status in a sample; and These are the mean values ​​of the characteristic and the health status, respectively. Let the covariance of the two be , and These are their standard deviations. Based on the above correlation coefficient analysis, features with an absolute correlation coefficient of not less than 0.6 with health status are selected, and these strongly correlated health factors are then used to construct the model input feature set. This process effectively removes redundant information that is irrelevant to or weakly correlated with battery degradation, laying a robust data foundation for subsequent model training.

[0048] S300: Construct a convolutional neural network based on residual structure as a high-dimensional feature extractor for the model, select mean square error as the optimization function, and extract features from source domain data and target domain data.

[0049] like Figure 4 As shown, in this step, residual units are first introduced into the overall network structure. Skip connections are used to address the gradient vanishing problem that often occurs in deep networks, ensuring that the network can maintain efficient training and effective feature extraction while increasing the number of layers. Then, activation functions and batch normalization layers are added after each convolutional layer to enhance the network's ability to express battery degradation information. Finally, the mean squared error is used as the objective optimization function, which can be expressed as follows:

[0050]

[0051] in and They represent the first Predicted values ​​and actual values ​​for each sample The total number of samples.

[0052] S400: Introduces a multi-kernel maximum mean difference loss function to calculate the distribution distance between the source domain and the target domain in the high-level feature output of the convolutional neural network model. The total loss function is obtained by weighted summation of this loss function and the mean square error loss function. An adaptive moment estimation optimizer is applied to optimize the total loss function to achieve domain adaptive training of unlabeled battery data.

[0053] S500: Uses the validation set to evaluate model performance and applies it to the test set to output health status estimation results, thereby enabling battery health status estimation in cross-domain scenarios.

[0054] In another embodiment, step S400 includes the following steps:

[0055] S401: To adapt to different feature distribution patterns, various kernel functions such as linear kernels and multinomial kernels are selected and weighted together. By calculating the mean difference between high-level features in the source and target domains in the kernel space, the maximum mean difference loss value of the multi-kernel kernel is obtained; assuming and Let them represent the source domain dataset and the target domain dataset, respectively, and let them each conform to... and Distribution. Data sample sizes are respectively and Therefore, the empirical estimate of the maximum mean difference can be expressed as:

[0056]

[0057] in Fis the set of continuous functions in the sample space. When and are determined, the result of the maximum mean discrepancy is determined by the given set of continuous functions . Only when p and q are the same, the value of the maximum mean discrepancy is 0. In order to approach the theoretical value, the set of continuous functions F needs to contain enough elements, but when F approaches infinity, the value of the maximum mean discrepancy becomes infinite and it is difficult to converge, so the set of continuous functions F needs to be restricted. When the set of continuous functions F is restricted in the unit ball of the reproducing kernel Hilbert space, the requirement can be met. The reproducing kernel Hilbert space can be represented by the inner product of the points in the space , as follows:

[0058]

[0059] Using the properties of the reproducing kernel Hilbert space and replacing and with and respectively, the following can be derived, where H represents the reproducing kernel Hilbert space:

[0060]

[0061] Then square both sides of the above equation to get the following derivation:

[0062]

[0063] Replace the inner product with the kernel function , and use the Gaussian kernel function for mapping, as follows:

[0064]

[0065] Through the above derivation, the final maximum mean discrepancy can be expressed as:

[0066]

[0067] The maximum mean discrepancy can be regarded as the distance between two points in the reproducing kernel Hilbert space, which is used to measure the difference between two distributions. In practical use, multiple kernel maximum mean discrepancy is often used, which has stronger representation performance for the mapping function than single kernel maximum mean discrepancy.

[0068] S402: Weighted sum of the multi-core maximum mean difference loss value and the mean square loss to form the final total loss function. The mean square loss term is:

[0069]

[0070] wherein and are the predicted value and the true value of the model respectively, N is the number of samples. The total loss function is defined as:

[0071]

[0072] wherein is a weight coefficient, used to balance between the prediction accuracy of the health state and the distribution alignment effect.

[0073] S403: The adaptive moment estimation optimizer is used to iteratively optimize the total loss function, and the network parameters are constantly updated, so that the network realizes effective domain adaptation training on the unlabeled battery data. The parameter update process of the adaptive moment estimation optimizer is usually given by the following formula:

[0074]

[0075] wherein is the trainable parameter of the network, is the learning rate, and are the first and second moment estimates of the gradient respectively, is a small constant to prevent division by zero. Through continuous iteration, the network minimizes the mean square loss while minimizing the difference between the source domain and the target domain feature distribution, thereby improving the health state prediction accuracy of the unlabeled battery data in the cross-domain scenario and realizing the domain adaptation training of the target domain data.

[0076] By using the training method of the embodiment, the data distribution difference between different domains in the battery field regression task can be reduced, and the accurate estimation of the cross-domain battery health state can be realized.

[0077] Next, by comparing the constructed convolutional neural network-multi-core maximum mean difference model with other model methods, the beneficial effects of the technical solutions of the present disclosure are further proved.

[0078] Specifically, the main evaluation indexes used in this comparative experiment are the mean absolute percentage error (MAE) and the root mean square error (RMSE).

[0079] Since the actual situation considered by the convolutional neural network-multiple kernel maximum mean discrepancy model is the effective use of unlabeled data, there are relatively sufficient unlabeled data in real scenarios, and this does not involve the case of limited data samples. Both the target domain unlabeled data and the source domain labeled data need to select multiple batteries and the number is preferably similar, so 8 batteries of a certain working condition are selected for the source domain, and 8 batteries of another working condition are selected for the target domain, wherein the battery state of health label value is used for the source domain, and the battery state of health label value is not used for the target domain. In the training process of the convolutional neural network-multiple kernel maximum mean discrepancy model, the source domain data are all used as training data and the target domain data are trained together, the target domain data take 1 battery as a test set, the remaining 7 batteries are divided into a validation set and a training set according to 8:2, and 8 times are needed, and each time a different battery is selected as a test set. In order to highlight the performance of the convolutional neural network-multiple kernel maximum mean discrepancy model, a comparative experiment needs to be designed for verification. Since the target domain data are unlabeled, a target domain data direct supervision training comparative experiment group cannot be set. In the experiment in this section, the comparative experiment is designed as follows:

[0080] (1) Method 1 directly predicts using source domain model parameters: directly supervised training using source domain data (8 batteries are all used as training data, and are divided into a training set and a validation set according to 8:2), then the model parameters of the model directly loaded with the target domain data are used to predict the test set of the target domain, to obtain a prediction result and calculate three index values.

[0081] (2) Method 2 predicts based on the convolutional neural network-multiple kernel maximum mean discrepancy model after training: the source domain data and the target domain data are used to train the convolutional neural network-multiple kernel maximum mean discrepancy model, the distribution difference of the output features of the model network layer is reduced, the learned features are suitable for both the source domain and the target domain, and after the training is completed, the prediction is performed on the test set, to obtain a prediction result and calculate three index values.

[0082] Based on the above experimental design, the experimental results are shown in Table 1.

[0083] Table 1

[0084]

[0085] In Figure 5 and Figure 6In the middle, the average absolute error and root mean square error indicators of mode 2 are optimal on all batteries. As can be seen from table 1, the average absolute error indicator of mode 2 has an average of 1.213%, and the root mean square error indicator has an average of 1.533%, which is much better than mode 1 in both average indicators, and performs well. From the above analysis, it can be seen that mode 2 has great improvement compared with mode 1 on almost all batteries, and the precision is greatly improved. This is because the distribution difference between the source domain and the target domain in mode 1 is too large, which leads to poor prediction performance, while in mode 2, the multi-core maximum mean difference is used for optimization, which greatly reduces the feature distribution distance between the source domain and the target domain, and well solves the problem of mode 1.

[0086] The application further provides a computer storage medium, which comprises computer instructions, and when the computer instructions are run on a computer, the computer is caused to execute the method.

[0087] An electronic device comprises a memory, a processor and a computer program stored on the memory and executable on the processor; wherein the processor implements the above method when executing the program.

[0088] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the embodiments disclosed herein. Variations and modifications of the present application are possible in light of the above teachings, and it is therefore intended to cover within the scope of the application any and all such variations and modifications. It is intended to include the equivalents of what has been described and to cover all such modifications and variations as long as they fall within the scope of the appended claims.

[0089] It should be understood that the application is not limited to the precise construction that has been described above and illustrated in the accompanying drawings, and that various modifications and changes can be made by those skilled in the art without departing from the scope of the application. The scope of the application should be limited only by the appended claims.

Claims

1. A method for estimating the state of health of a battery in multiple scenarios based on domain adaptation migration, characterized in that, The method comprises the following steps: Step 1, collecting the dynamic process parameter signals of the monitored battery charging and discharging stage, and collecting the state parameters of the battery aging stage reflecting the health state, and deriving the health state quantity of the battery; Step 2, extracting the health factors representing the degradation of the battery from the charging and discharging data and the electrochemical impedance spectrum analysis results; Step 3, constructing a convolutional neural network based on residual structure as a high-dimensional feature extractor of the model, selecting mean square error as the optimization function, and extracting features from the source domain data and the target domain data; Step 4, introducing a multi-kernel maximum mean difference loss function, calculating the distribution distance of the high-level feature output of the source domain data and the target domain data in the convolutional neural network model, and obtaining a total loss function by weighted summation of the loss function and the mean square error loss function, and applying an adaptive matrix estimator optimizer to optimize the total loss function, realizing domain adaptive training of unlabeled battery data; Step 5, calculating the mean absolute error and the root mean square error on the validation set, selecting the optimal model parameters; applying the trained model to the target domain unlabeled data to output the health state estimation value, and realizing the health state prediction under the cross-condition and cross-batch scene. The health state quantity is:

2. The method of claim 1, wherein the method is characterized by, The dynamic process parameter signals in step 1 include: current signals recorded during charging , charging voltage signals , and charging temperature signals .

3. The method of claim 1, wherein the method further comprises: Step 2 specifically comprises the following sub-steps: wherein, represents the maximum discharge capacity of the battery per cycle during the aging process, represents the current cycle number, is the nominal capacity of the battery.

4. The method of claim 1, wherein the domain adaptation migration based battery multi-scenario state of health estimation method is characterized by, Step 2a, extracting the constant-voltage drop time interval of the constant-current charging stage voltage curve, the time integral of the constant-voltage charging stage current decay curve, and the temperature change rate in the charging and discharging cycle from the charging and discharging data; and combining the system characteristic parameters extracted from the electrochemical impedance spectrum analysis results to form initial health factors; Step 2b, calculating the linear correlation between the initial health factors and the battery health state through the Pearson correlation coefficient, and selecting the features with an absolute value of the correlation coefficient greater than or equal to a set threshold as the model input; Step 2c, dividing the processed data into a training set, a validation set and a test set, and the label is the corresponding health state sequence. In step 3, the convolutional neural network comprises:

5. The method of claim 1, wherein the domain adaptation migration based battery multi-scenario state of health estimation method is characterized by, An input layer receives time series health factor data; a residual convolution module is composed of a plurality of one-dimensional convolution layers, each layer including a convolution kernel, a batch normalization layer, an activation function and a residual connection; a pooling layer is connected to a global average pooling layer after the last convolution layer, and outputs a high-dimensional feature vector; a fully connected layer uses a linear activation function to map the pooled features to a health state prediction value. In step 4, the multi-kernel maximum mean difference loss function includes a regression loss and a distribution alignment loss.

6. The method of claim 1, wherein the domain adaptation migration based battery multi-scenario state of health estimation method is characterized by, The storage medium comprises computer instructions which, when executed on a computer, cause the computer to perform the method of any one of claims 1-6.

7. A computer storage medium, characterized by, The electronic device comprises a memory, a processor and a computer program stored on the memory and executable on the processor; wherein the processor implements the method of any one of claims 1-6 when executing the program.

8. An electronic device, characterized by comprising: ​

Images