A method for rapid estimation of state of health of retired battery based on transfer learning

Abstract

This invention discloses a rapid estimation method for the health status of retired batteries based on transfer learning. The method first collects source domain data of retired power batteries with different charging and aging states, and simultaneously collects a small batch of similar data of retired power batteries in the field to construct a target domain dataset. Through a three-stage transfer learning strategy of source domain data pre-training, source domain and target domain data feature transfer training, and target domain data fine-tuning, the method achieves rapid estimation of the health status of retired power batteries in scenarios without historical data, effectively accelerating the on-site screening process of retired power batteries and improving the efficiency and on-site adaptability of retired power battery health status diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of battery energy storage technology, and in particular relates to a method for rapid estimation of the health status of retired batteries based on transfer learning. Background Technology

[0002] With the global energy transition and the rapid development of the new energy vehicle industry, a large-scale wave of retired power batteries has arrived. The massive volume of retired batteries brings severe pressure on recycling and processing, but also holds enormous value for resource recycling. Retired power batteries are not "waste products"; they typically retain 70% to 80% of their remaining energy storage capacity, which can be used in energy storage systems, backup power for communication base stations, and low-speed electric vehicles. Batteries with lower remaining value can be recycled to extract scarce metal resources such as lithium, cobalt, and nickel, alleviating the pressure of dependence on foreign sources for strategic resources.

[0003] Battery state of health (SOH), as a core indicator characterizing the degree of battery aging, directly determines the processing path and utilization value of retired power batteries, and is a key prerequisite for achieving large-scale and standardized recycling. Accurate SOH assessment can effectively avoid safety risks in cascade utilization, improve resource allocation efficiency, and reduce overall processing costs. However, current SOH estimation technology for retired power batteries still faces many bottlenecks, making it difficult to meet the needs of large-scale industry development.

[0004] Specifically, existing SOH estimation techniques have significant shortcomings: traditional detection methods such as direct discharge are complex and time-consuming, and may damage the battery; electrochemical impedance spectroscopy relies on specialized equipment and has poor practicality; chemical analysis methods can damage the battery structure. Model-driven methods are difficult to model and have complex parameter identification, resulting in poor adaptability to dynamic operating conditions. While data-driven methods do not require precise modeling, they heavily rely on high-quality training data. When faced with the large individual differences in retired batteries and the complex aging mechanisms, the model's generalization ability is severely insufficient, and some algorithms have long training cycles and poor real-time performance.

[0005] Against this backdrop, the industry urgently needs to develop a rapid SOH estimation technology that combines speed, accuracy, and practicality to address the problems of low efficiency, high cost, and poor adaptability of existing methods. This will provide core technical support for the large-scale promotion of the cascade utilization and recycling of retired power batteries, and help achieve multiple goals of resource recycling, environmental protection, and economic benefits. Summary of the Invention

[0006] This invention provides a rapid method for estimating the state of health (SOH) of decommissioned batteries based on transfer learning, aiming to solve the technical problems of lack of historical data for on-site diagnosis of decommissioned batteries and low efficiency in SOH estimation. This method utilizes knowledge learned from source domain data and transfers it to the target domain, reducing the amount of data required in the target domain and improving the accuracy of SOH estimation.

[0007] To achieve the above-mentioned objectives, the present invention specifically adopts the following technical solution:

[0008] In a first aspect, the present invention provides a method for rapid estimation of the health status of decommissioned batteries based on transfer learning, comprising the following steps:

[0009] S1. Select several retired power batteries of the same specification as source domain batteries, obtain source domain data of source domain batteries, and form source domain dataset; select several retired power batteries of the same specification as target domain batteries, obtain target domain data of target domain batteries, and form target domain dataset; then, preprocess the source domain dataset and target domain dataset.

[0010] S2. Pre-train the first encoder, the first predictor, and the first reconstructor using the pre-processed source domain dataset;

[0011] S3. Fix the pre-trained weight parameters of the first encoder, and use the pre-processed source domain dataset and target domain dataset to perform feature transfer training on the second encoder, the second reconstructor, and the domain discriminator; wherein, the domain discriminator is connected to the first encoder and the second encoder through a gradient inversion layer;

[0012] S4. The third encoder and the second predictor are fine-tuned using the preprocessed target domain dataset, and the fine-tuned third encoder and the second predictor constitute a health status estimation model for retired power batteries.

[0013] S5. Retired power batteries recovered from industrial sites and awaiting health status estimation are used as target detection batteries. Voltage response data of each target detection battery under rapid current pulse excitation is collected to obtain the feature matrix corresponding to the target detection battery. The feature matrix of the target detection battery is input into the retired power battery health status estimation model, and the SOH estimate of the target detection battery is output.

[0014] Based on the above scheme, each step can be implemented in the following preferred manner.

[0015] As a preferred embodiment of the first aspect mentioned above, the specific process for obtaining the source domain data and the target domain data in step S1 is as follows:

[0016] S11. Obtain the 1C discharge capacity of the source domain battery and the target domain battery after constant current-constant voltage charging, and at the same time obtain the voltage response data of the source domain battery and the target domain battery under different states of charge and fast current pulse excitation.

[0017] S12. Divide the 1C discharge capacity by the rated capacity of the battery to obtain the SOH of the battery and use it as the true label:

[0018] S13. Subtract the voltage response data at each time step from the initial voltage response data to obtain the polarization voltage data at the corresponding time step. Then, construct the feature matrix of the battery from the voltage response data and polarization voltage data at different times step. Use the real label and feature matrix of the source domain battery as the source domain data, and use the real label and feature matrix of the target domain battery as the target domain data.

[0019] As a preferred embodiment of the first aspect, the specific process for obtaining 1C discharge capacity and voltage response data in step S11 is as follows: First, discharge to the lower cutoff voltage with a 1C current, then charge to the upper cutoff voltage using a constant current-constant voltage mode. After the first resting treatment, discharge to the lower cutoff voltage with a 1C current to obtain the 1C discharge capacity after constant current-constant voltage charging. Subsequently, charge to the specified cutoff voltage using a constant current mode. After the second resting treatment, perform a fast current pulse test, change the cutoff voltage, and repeat the resting treatment and fast current pulse test steps at each cutoff voltage until the battery is charged to the upper cutoff voltage or all cutoff voltage tests are completed, and finally obtain the voltage response data.

[0020] As a preferred embodiment of the first aspect, in step S1, during preprocessing, the feature matrix of the source domain battery is first denormalized using the upper and lower cutoff voltages of the source domain battery, and the feature matrix of the target domain battery is denormalized using the upper and lower cutoff voltages of the target domain battery; then, the maximum and minimum values ​​are found in the two denormalized feature matrices, and the corresponding feature maximum and feature minimum values ​​are obtained; finally, based on the feature maximum and feature minimum values, the maximum-minimum normalization method is used to map the two denormalized feature matrices to the [0,1] interval, thereby completing the preprocessing.

[0021] As a preferred embodiment of the first aspect mentioned above, the specific pre-training process in step S2 is as follows:

[0022] S21. Input the normalized feature matrix of the source domain cell into the first encoder, and output the first deep feature of the source domain from the first encoder.

[0023] S22. Input the first deep feature of the source domain into the first predictor to predict the label and obtain the predicted label of the source domain. Input the first deep feature of the source domain into the first reconstructor to reconstruct the feature and obtain the first reconstructed feature of the source domain. Use the mean squared error loss between the predicted label of the source domain and its true label as the prediction loss. Use the mean squared error loss between the first reconstructed feature of the source domain and the normalized feature matrix of the source domain battery as the first reconstruction loss. Sum the prediction loss and the first reconstruction loss by weight as the total pre-training loss. Update the weight parameters of the first encoder, the first predictor and the first reconstructor based on minimizing the total pre-training loss.

[0024] As a preferred embodiment of the first aspect mentioned above, the specific process of feature transfer training in step S3 is as follows:

[0025] S31. Assign the pre-trained weight parameters of the first encoder to the second encoder as the initial weight parameters of the second encoder, and freeze the weight parameters of the first encoder at the same time.

[0026] S32. Input the normalized feature matrix of the source domain battery into the pre-trained first encoder and output the second deep feature of the source domain; input the normalized feature matrix of the target domain battery into the second encoder and output the first deep feature of the target domain.

[0027] S33. After inputting the second deep features of the source domain and the first deep features of the target domain into the neighborhood discriminator, calculate the neighborhood discrimination loss based on the output of the neighborhood discriminator;

[0028] S34. Calculate feature statistics to measure the loss based on the second deep features of the source domain and the first deep features of the target domain;

[0029] S35. Input the second deep feature of the source domain into the second reconstructor for feature reconstruction to obtain the second reconstructed feature of the source domain. Input the first deep feature of the target domain into the second reconstructor for feature reconstruction to obtain the first reconstructed feature of the target domain. Use the mean square error loss between the second reconstructed feature of the source domain and the normalized feature matrix of the source domain battery as the mean square error loss of the source domain feature reconstruction. Use the mean square error loss between the first reconstructed feature of the target domain and the normalized feature matrix of the target domain battery as the mean square error loss of the target domain feature reconstruction. Add the mean square error loss of the source domain feature reconstruction and the mean square error loss of the target domain feature reconstruction as the second reconstruction loss.

[0030] S36. The domain discrimination loss, feature statistics measurement loss, and second reconstruction loss are weighted and summed to form the total loss for feature transfer training. The weight parameters of the second encoder, second reconstructor, and domain discriminator are updated based on minimizing the total loss for feature transfer training.

[0031] As a preferred embodiment of the first aspect mentioned above, the specific process of fine-tuning in step S4 is as follows:

[0032] S41. Transfer the pre-trained weight parameters of the first predictor to the second predictor as the initial weight parameters of the second predictor; transfer the weight parameters of the second encoder trained by feature transfer to the third encoder as the initial weight parameters of the third encoder;

[0033] S42. Input the normalized feature matrix of the target battery into the third encoder, and the third encoder outputs the second deep feature of the target domain.

[0034] S43. Input the second deep feature of the target domain into the second predictor to predict the label and obtain the predicted label of the target domain. Use the mean squared error loss between the predicted label of the target domain and its true label as the total loss for fine-tuning, and update the weight parameters of the third encoder and the second predictor based on minimizing the total loss for fine-tuning.

[0035] In a second aspect, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can implement the rapid estimation method for the health status of decommissioned batteries based on transfer learning as described in any of the solutions in the first aspect above.

[0036] Thirdly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for rapid estimation of the health status of decommissioned batteries based on transfer learning as described in any of the solutions of the first aspect above.

[0037] Fourthly, the present invention provides a computer electronic device, which includes a memory and a processor;

[0038] The memory is used to store computer programs;

[0039] The processor is configured to, when executing the computer program, implement the fast estimation method for the health status of decommissioned batteries based on transfer learning as described in any of the solutions of the first aspect above.

[0040] Compared with the prior art, the present invention has the following advantages:

[0041] This invention does not rely on historical operating data of retired batteries. It can complete model training and SOH estimation by collecting rapid current pulse discharge data of a small batch of retired batteries in the field. This solves the industry pain point of lacking historical data of retired batteries in the field and improves the field adaptability of the method.

[0042] This invention designs a transfer learning strategy that integrates domain feature alignment and model fine-tuning: by pre-training with source domain data, the feature extraction capability is transferred; then, by aligning the features of the source domain and the target domain, the distribution difference between the domains is reduced; finally, the model is fine-tuned using a small amount of target domain data, which effectively improves the accuracy of SOH estimation in scenarios without historical data.

[0043] This invention combines rapid current pulse testing technology to collect battery data, replacing the traditional cumbersome charge and discharge tests. This significantly shortens the data collection time, effectively accelerates the on-site screening process of retired batteries, and reduces the testing costs for reuse. Attached Figure Description

[0044] Figure 1This is a schematic diagram of the core steps of the method of the present invention;

[0045] Figure 2 This is a schematic diagram of the fast current pulse test data provided in this embodiment;

[0046] Figure 3 This is a schematic diagram of the architecture of the first encoder and the first reconfigurator provided in this embodiment;

[0047] Figure 4 This is an expanded diagram showing the implementation details of the first encoder and the first reconstructor provided in this embodiment;

[0048] Figure 5 This is a schematic diagram illustrating the SOH estimation effect of the retired power battery according to the present invention. Detailed Implementation

[0049] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.

[0050] In the description of this invention, it should be understood that the terms "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features.

[0051] like Figure 1 As shown, in a preferred embodiment of the present invention, the above-mentioned fast estimation method for the health status of decommissioned batteries based on transfer learning includes the following steps S1 to S5. The specific implementation process of each step will be described in detail below.

[0052] S1. Select several retired power batteries of the same specification as source domain batteries, obtain source domain data of source domain batteries, and form source domain dataset; select several retired power batteries of the same specification as target domain batteries, obtain target domain data of target domain batteries, and form target domain dataset; then, preprocess the source domain dataset and target domain dataset.

[0053] It should be noted that in step S1 of this invention, the source domain battery and the target domain battery can be two different batteries from different manufacturers, or batteries collected from different industrial sites. This invention does not impose any restrictions.

[0054] It should be noted that the specific process of obtaining the source domain data and the target domain data in step S1 of this invention is as follows:

[0055] S11. Obtain the 1C discharge capacity of the source domain battery and the target domain battery after constant current-constant voltage charging, and at the same time obtain the voltage response data of the source domain battery and the target domain battery under different states of charge and fast current pulse excitation.

[0056] In step S11 of this invention, the specific process of obtaining 1C discharge capacity and voltage response data is as follows: First, discharge to the lower cutoff voltage with a 1C current, then charge to the upper cutoff voltage using a constant current-constant voltage mode. After the first resting treatment, discharge to the lower cutoff voltage with a 1C current to obtain the 1C discharge capacity after constant current-constant voltage charging. Then, charge to the specified cutoff voltage using a constant current mode. After the second resting treatment, perform a fast current pulse test, change the cutoff voltage, and repeat the resting treatment and fast current pulse test steps at each cutoff voltage until the battery is charged to the upper cutoff voltage or all cutoff voltage tests are completed, and finally obtain the voltage response data.

[0057] S12. Based on 1C discharge capacity and the rated capacity of the battery Calculate the SOH of the battery using the following formula and use it as the actual label:

[0058] ;

[0059] In this embodiment S12, several retired lithium iron phosphate (LFP) batteries of the same specification are selected as the source batteries mentioned above. Retired nickel-cobalt-manganese (NCM) power batteries (with core cathode material) recovered on-site are also used. As the target domain battery described above, source domain and target domain data acquisition is carried out according to the process described in this invention. Taking source domain data acquisition as an example, in this embodiment, the battery is discharged to 2.0V with a 1C current, charged to 3.6V using CC-CV mode, left to stand for 10 minutes, and then discharged to 2.0V with a 1C current. Subsequently, it is charged to 3.25V with a constant current, left to stand for 2 hours, and then subjected to a fast current pulse test to acquire voltage response data. The resting and fast current pulse test steps are repeated to complete the tests at cutoff voltages of 3.30V, 3.35V, and 3.40V in sequence to obtain the voltage response data of the source domain battery. In this embodiment, the fast current pulse data and voltage response data are as follows: Figure 2 As shown. The data acquisition process for the target domain battery is similar, except that different cutoff voltages are used: 3.65V, 3.70V, 3.75V, 3.80V, and 3.85V, which will not be elaborated here.

[0060] S13. Subtract the voltage response data at each time step from the initial voltage response data to obtain the polarization voltage data at the corresponding time step. Then, construct the feature matrix of the battery from the voltage response data and polarization voltage data at different times step. Use the real label and feature matrix of the source domain battery as the source domain data, and use the real label and feature matrix of the target domain battery as the target domain data.

[0061] In this embodiment S13, with Taking a specific time as an example, the voltage response data at that time... and initial voltage response data Subtract, and you get Polarization voltage data at time The final constructed feature matrix As shown below:

[0062]

[0063]

[0064] It should be noted that the feature matrices of both the source domain cell and the target domain cell can be calculated using the above formula. For ease of description below, the feature matrix of the source domain cell will be denoted as follows. The feature matrix of the target domain battery is denoted as .

[0065] It should be noted that in step S1 of this invention, during preprocessing, the feature matrix of the source domain battery is first denormalized using the upper and lower cutoff voltages of the source domain battery, and the feature matrix of the target domain battery is denormalized using the upper and lower cutoff voltages of the target domain battery; then, the maximum and minimum values ​​are found in the two denormalized feature matrices, and the corresponding feature maximum and feature minimum values ​​are obtained; finally, based on the feature maximum and feature minimum values, the maximum-minimum normalization method is used to map the two denormalized feature matrices to the interval [0,1], thereby completing the preprocessing.

[0066] In the preprocessing of this embodiment, the upper and lower cutoff voltages of the source domain battery are 2V and 3.65V, respectively, and the upper and lower cutoff voltages of the target domain battery are 4.2V and 2.5V, respectively. Based on this, the present invention denormalizes the feature matrices of the two types of batteries. Then, the maximum value in the two feature matrices is taken as the feature maximum and the minimum value is taken as the feature minimum, thereby performing max-min normalization on the two feature matrices.

[0067]

[0068]

[0069]

[0070]

[0071] in, These represent the upper and lower cutoff voltages of the source domain cell and the upper and lower cutoff voltages of the target domain cell, respectively. This represents the characteristic matrix of the source domain cell after denormalization. This represents the feature matrix of the target domain battery after denormalization; It is the minimum value of the characteristic; It is the maximum value of the characteristic; This is the feature matrix of the source domain cell after normalization. This is the feature matrix of the target domain battery after normalization.

[0072] S2. Pre-train the first encoder, the first predictor, and the first reconstructor using the pre-processed source domain dataset.

[0073] It should be noted that the specific pre-training process in step S2 of this invention is as follows:

[0074] S21. The feature matrix of the normalized source region cell. Input first encoder The first deep feature of the source domain is output by the first encoder. .

[0075] S22. Input the first deep feature from the source domain into the first predictor. Perform label prediction to obtain the predicted labels from the source domain, and input the first deep features from the source domain into the first reconstructor. Feature reconstruction is performed to obtain the first reconstructed feature of the source domain. The mean squared error loss between the predicted label and the true label of the source domain is used as the prediction loss. The mean squared error loss between the first reconstructed feature of the source domain and the normalized feature matrix of the source domain battery is used as the first reconstruction loss. The prediction loss and the first reconstruction loss are weighted and summed to obtain the total pre-training loss. The weight parameters of the first encoder, the first predictor and the first reconstructor are updated based on minimizing the total pre-training loss.

[0076] In this embodiment S22, the predicted label is based on the source domain. The first reconstruction feature of the source domain Calculate the predicted loss accordingly and the first reconstruction loss The two losses are summed to obtain the total pre-training loss. :

[0077]

[0078]

[0079]

[0080] in, Indicates the number of samples in the current batch; The first in the source domain of the current batch A real label, The first in the source domain of the current batch One predicted label; The first in the current batch The normalized feature matrix of the source domain cell The first in the source domain of the current batch The first reconstructed feature.

[0081] In this embodiment, the source domain dataset is further randomly divided into a source domain training set based on the battery number. Source domain test set The first encoder, first predictor, and first reconstructor are pre-trained on the source domain training set. After pre-training, the performance is verified using the source domain test set. At the same time, the weight parameters of the first encoder, first predictor, and first reconstructor are saved to provide a pre-training basis for the subsequent transfer learning process.

[0082] It should also be noted that in step S2 of this invention, the specific network structures of the first encoder, the first predictor, and the first reconstructor can be selected by those skilled in the art according to actual needs. As a preferred implementation, the first predictor employs a two-layer fully connected network, and the first encoder is constructed using a depthwise separable one-dimensional convolutional neural network and residual connections. Furthermore, the first encoder and the first reconstructor form a U-net architecture, such as... Figure 3 As shown. Figure 3 The implementation methods of each component module belong to existing technologies, and their specific structural forms are as follows: Figure 4 As shown: Figure 4 (a) in the diagram represents the DWS 1-D CNN module; Figure 4 (b) in the diagram represents the computation process of the DWS 1-D CNN. Figure 4 (c) in the text refers to the DWS Resblock with the identity connection, i.e., DWSResblock 1; Figure 4 In the diagram, (d) represents the DWS Resblock for projection connection, namely DWS Resblock 2, which will not be elaborated further.

[0083] S3. Fix the pre-trained weight parameters of the first encoder, and use the pre-processed source domain dataset and target domain dataset to perform feature transfer training on the second encoder, the second reconstructor and the domain discriminator; wherein, the domain discriminator is connected to the first encoder and the second encoder through a gradient reversal layer (GRL).

[0084] It should be noted that the specific process of feature transfer training in step S3 of this invention is as follows:

[0085] S31. Assign the pre-trained weight parameters of the first encoder to the second encoder as the initial weight parameters of the second encoder, and freeze the weight parameters of the first encoder at the same time.

[0086] S32. The feature matrix of the normalized source region cell. Input the pre-trained first encoder Output the second deep features of the source domain The feature matrix of the target domain battery after normalization. Input to the second encoder Output the first deep feature of the target domain .

[0087] S33. After inputting the second deep features of the source domain and the first deep features of the target domain into the neighborhood discriminator, the neighborhood discrimination loss is calculated based on the output of the neighborhood discriminator.

[0088] In this embodiment S33, the neighborhood discriminator is used to determine whether the input feature originates from the source domain or the target domain; correspondingly, the neighborhood discrimination loss... The calculation method is as follows:

[0089]

[0090] in, Indicates taking the expected value; This represents the domain discriminator.

[0091] S34. Calculate feature statistics to measure loss based on the second deep features of the source domain and the first deep features of the target domain.

[0092] In this embodiment S34, the feature statistics measure the loss. The result is obtained by adding the multi-core maximum mean difference (MK-MMD) loss and the correlation alignment (CORAL) loss. The implementation methods of the multi-core maximum mean difference loss and the correlation alignment loss are existing technologies. To facilitate understanding by those skilled in the art, the calculation methods of the two losses are briefly explained below.

[0093] 1) Multi-core maximum mean difference loss Calculation process:

[0094]

[0095]

[0096] in, and These represent the number of samples in the source domain and the number of samples in the target domain, respectively. It is a nonlinear mapping function that maps the deep feature matrix space to the regenerated Hilbert space (RKHS); The first in the current batch The second deep feature of each source domain; The first in the current batch The first deep feature of the target domain; This represents the L2 norm operation in the regenerated Hilbert space (RKHS); The combination of multiple Gaussian radial basis function (RBF) kernels is calculated as follows:

[0097]

[0098]

[0099] in, Indicates the first Gaussian radial basis functions Its bandwidth; It is the Frobenius norm for matrix operations.

[0100] 2) Correlation Alignment Loss Calculation process:

[0101]

[0102]

[0103]

[0104] in, The number of feature channels output by the encoder; and As an auxiliary variable; Represents a column vector in which all elements are equal to 1; Indicates transpose; This is the feature matrix of the source domain cell after normalization. This is the feature matrix of the target battery after normalization.

[0105] S35. Input the second deep feature of the source domain into the second reconstructor. Perform feature reconstruction to obtain the second reconstructed features of the source domain, and input the first deep features of the target domain into the second reconstructor. Feature reconstruction is performed to obtain the first reconstructed feature of the target domain. The mean square error loss between the second reconstructed feature of the source domain and the normalized feature matrix of the source domain battery is used as the mean square error loss of the source domain feature reconstruction. The mean square error loss between the first reconstructed feature of the target domain and the normalized feature matrix of the target domain battery is used as the mean square error loss of the target domain feature reconstruction. The mean square error loss of the source domain feature reconstruction and the mean square error loss of the target domain feature reconstruction are added together to obtain the second reconstruction loss.

[0106] S36. Combine the domain discrimination loss, the feature statistics measurement loss, and the second reconstruction loss. The weighted sum is used as the total loss for feature transfer training, and the weight parameters of the second encoder, second reconstructor, and domain discriminator are updated based on minimizing the total loss for feature transfer training.

[0107] In this embodiment S36, the total loss of feature transfer training Specifically as follows:

[0108]

[0109] It should be noted that in step S3 of this invention, the second encoder and the first encoder must have the same structure, the difference being that the weight parameters obtained by the two encoders after training are different. The second reconstructor and the first reconstructor have the same structure. The neighborhood discriminator is an additional structure introduced in this invention during feature transfer training. It is necessary to calculate the neighborhood discrimination loss based on it, thereby updating the weight parameters of the second encoder, the second reconstructor, and the neighborhood discriminator. However, after the feature transfer training is completed, only the weight parameters of the second encoder are needed for the subsequent fine-tuning process. As a preferred implementation, the neighborhood discriminator adopts two fully connected layers, which are connected to the first encoder and the second encoder through gradient inversion layers, as specifically implemented as follows:

[0110] Forward propagation:

[0111]

[0112] Backpropagation:

[0113]

[0114]

[0115] in, This indicates a gradient inversion layer; This is the input to the gradient inversion layer, i.e., the output of the two encoders; This is the output of the gradient inversion layer; To determine the loss in the domain; Indicates the parameters of the second encoder; It is a non-negative coefficient; To obtain the partial derivative.

[0116] S4. Use the preprocessed target domain dataset to fine-tune the third encoder and the second predictor, and use the fine-tuned third encoder and the second predictor to form a health status estimation model for retired power batteries.

[0117] It should be noted that the specific process of fine-tuning in step S4 of this invention is as follows:

[0118] S41. Transfer the pre-trained weight parameters of the first predictor to the second predictor as the initial weight parameters of the second predictor; transfer the weight parameters of the second encoder trained by feature transfer to the third encoder as the initial weight parameters of the third encoder.

[0119] S42. The normalized feature matrix of the target battery Input third encoder The second deep feature of the target domain is output by the third encoder. .

[0120] S43. Input the second deep feature of the target domain into the second predictor. Perform label prediction to obtain the predicted label of the target domain. Use the mean squared error loss between the predicted label of the target domain and its true label as the total loss for fine-tuning. Update the weight parameters of the third encoder and the second predictor based on minimizing the total loss for fine-tuning.

[0121] It should also be noted that in step S4 of this invention, the third encoder and the second encoder must have the same structure, and the second predictor and the first predictor must have the same structure. The difference lies in the weight parameters obtained by the two predictors and the two encoders after training. After fine-tuning, the weight parameters of the third encoder and the second predictor are saved for the estimation of the health status of actual decommissioned power batteries.

[0122] S5. Retired power batteries recovered from industrial sites and awaiting health status estimation are used as target detection batteries. Voltage response data of each target detection battery under rapid current pulse excitation is collected to obtain the feature matrix corresponding to the target detection battery. The feature matrix of the target detection battery is input into the retired power battery health status estimation model, and the SOH estimate of the target detection battery is output.

[0123] It should be noted that in step S5 of the present invention, after collecting the voltage response data of the target detection battery under fast current pulse excitation according to the process of S11, the feature matrix of the target detection battery is obtained according to the process of S13, and the SOH estimation of the target detection battery is realized by using the matrix as the model input.

[0124] To further demonstrate the effectiveness of this invention, this embodiment uses the SOH estimate of the target battery predicted by the retired power battery health state estimation model as the prediction label. The mean absolute error (MAE) and root mean squared error (RMSE) between the predicted label and the true label are calculated, and the prediction performance of the method of this invention is evaluated based on these two error indices. The estimation results of the method of this invention in the source and target domains are as follows: Figure 5 As shown, the MAE value of the present invention is 1.55% and the RMSE is 2.22% in the source domain data; after migrating using 10% of the target domain data, the MAE value of the target domain data is 1.43% and the RMSE is 2.16%. It can be seen that the method of the present invention has high accuracy in both the source and target domains.

[0125] It is understood that the rapid estimation method for the health status of decommissioned batteries based on transfer learning described in S1-S5 above can essentially be implemented by a computer program. Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer program product corresponding to the rapid estimation method for the health status of decommissioned batteries based on transfer learning provided in the above embodiments. This product includes a computer program / instruction that, when executed by a processor, can implement the rapid estimation method for the health status of decommissioned batteries based on transfer learning as described in the above embodiments.

[0126] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer electronic device corresponding to the rapid estimation method for the health status of retired batteries based on transfer learning provided in the above embodiments, which includes a memory and a processor;

[0127] The memory is used to store computer programs;

[0128] The processor is configured to implement the fast estimation method for the health status of decommissioned batteries based on transfer learning in the above embodiments when executing the computer program.

[0129] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0130] Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer-readable storage medium corresponding to the fast estimation method for the health status of decommissioned batteries based on transfer learning provided in the above embodiments. The storage medium stores a computer program, which, when executed by a processor, can implement the fast estimation method for the health status of decommissioned batteries based on transfer learning in the above embodiments.

[0131] Specifically, in the computer-readable storage medium of the above three embodiments, the stored computer program is executed by a processor, which can perform the aforementioned steps S1 to S5.

[0132] It is understood that the aforementioned storage media may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Furthermore, the storage media may also be various media capable of storing program code, such as USB flash drives, external hard drives, magnetic disks, or optical discs.

[0133] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0134] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A fast method for estimating the health status of decommissioned batteries based on transfer learning, characterized in that, Includes the following steps: S1. Select several retired power batteries of the same specifications as source domain batteries, obtain source domain data of source domain batteries, and form source domain dataset; Several retired power batteries of the same specification were selected as target domain batteries, and target domain data of the target domain batteries were obtained to form a target domain dataset; then, the source domain dataset and the target domain dataset were preprocessed. S2. Pre-train the first encoder, the first predictor, and the first reconstructor using the pre-processed source domain dataset; S3. Fix the pre-trained weight parameters of the first encoder, and use the pre-processed source domain dataset and target domain dataset to perform feature transfer training on the second encoder, the second reconstructor, and the domain discriminator; wherein, the domain discriminator is connected to the first encoder and the second encoder through a gradient inversion layer; S4. The third encoder and the second predictor are fine-tuned using the preprocessed target domain dataset, and the fine-tuned third encoder and the second predictor constitute a health status estimation model for retired power batteries. S5. Retired power batteries recovered from industrial sites and awaiting health status estimation are used as target detection batteries. Voltage response data of each target detection battery under rapid current pulse excitation is collected to obtain the feature matrix corresponding to the target detection battery. The feature matrix of the target detection battery is input into the retired power battery health status estimation model, and the SOH estimate of the target detection battery is output.

2. The fast health status estimation method for decommissioned batteries based on transfer learning as described in claim 1, characterized in that, In step S1, the specific process of obtaining the source domain data and the target domain data is as follows: S11. Obtain the 1C discharge capacity of the source domain battery and the target domain battery after constant current-constant voltage charging, and at the same time obtain the voltage response data of the source domain battery and the target domain battery under different states of charge and fast current pulse excitation. S12. Divide the 1C discharge capacity by the rated capacity of the battery to obtain the SOH of the battery and use it as the true label: S13. Subtract the voltage response data at each time step from the initial voltage response data to obtain the polarization voltage data at the corresponding time step. Then, construct the feature matrix of the battery from the voltage response data and polarization voltage data at different times step. Use the real label and feature matrix of the source domain battery as the source domain data, and use the real label and feature matrix of the target domain battery as the target domain data.

3. The fast health status estimation method for decommissioned batteries based on transfer learning as described in claim 2, characterized in that, In step S11, the specific process of obtaining 1C discharge capacity and voltage response data is as follows: First, discharge to the lower cutoff voltage with 1C current, then charge to the upper cutoff voltage using constant current-constant voltage mode. After the first resting process, discharge to the lower cutoff voltage with 1C current to obtain the 1C discharge capacity after constant current-constant voltage charging. The battery is then charged to the specified cutoff voltage using constant current mode. After a second resting period, a fast current pulse test is performed. The cutoff voltage is changed, and the resting period and fast current pulse test steps are repeated at each cutoff voltage until the battery is charged to the upper cutoff voltage or all cutoff voltage tests are completed. Finally, the voltage response data is obtained.

4. The fast health status estimation method for decommissioned batteries based on transfer learning as described in claim 2, characterized in that, In step S1, during preprocessing, the feature matrix of the source domain battery is first denormalized using the upper and lower cutoff voltages of the source domain battery, and the feature matrix of the target domain battery is denormalized using the upper and lower cutoff voltages of the target domain battery. Then, the maximum and minimum values ​​are found in the two denormalized feature matrices to obtain the corresponding feature maximum and feature minimum values. Finally, based on the feature maximum and feature minimum values, the maximum-minimum normalization method is used to map the two denormalized feature matrices to the interval [0,1], thereby completing the preprocessing.

5. The fast health status estimation method for decommissioned batteries based on transfer learning as described in claim 4, characterized in that, In step S2, the specific pre-training process is as follows: S21. Input the normalized feature matrix of the source domain cell into the first encoder, and output the first deep feature of the source domain from the first encoder. S22. Input the first deep feature of the source domain into the first predictor to predict the label and obtain the predicted label of the source domain. Input the first deep feature of the source domain into the first reconstructor to reconstruct the feature and obtain the first reconstructed feature of the source domain. Use the mean squared error loss between the predicted label of the source domain and its true label as the prediction loss. Use the mean squared error loss between the first reconstructed feature of the source domain and the normalized feature matrix of the source domain battery as the first reconstruction loss. Sum the prediction loss and the first reconstruction loss by weight as the total pre-training loss. Update the weight parameters of the first encoder, the first predictor and the first reconstructor based on minimizing the total pre-training loss.

6. The fast health status estimation method for decommissioned batteries based on transfer learning as described in claim 4, characterized in that, In step S3, the specific process of feature transfer training is as follows: S31. Assign the pre-trained weight parameters of the first encoder to the second encoder as the initial weight parameters of the second encoder, and freeze the weight parameters of the first encoder at the same time. S32. Input the normalized feature matrix of the source domain battery into the pre-trained first encoder and output the second deep feature of the source domain. The normalized feature matrix of the target domain battery is input into the second encoder, and the first deep feature of the target domain is output. S33. After inputting the second deep features of the source domain and the first deep features of the target domain into the neighborhood discriminator, calculate the neighborhood discrimination loss based on the output of the neighborhood discriminator; S34. Calculate feature statistics to measure the loss based on the second deep features of the source domain and the first deep features of the target domain; S35. Input the second deep feature of the source domain into the second reconstructor for feature reconstruction to obtain the second reconstructed feature of the source domain. Input the first deep feature of the target domain into the second reconstructor for feature reconstruction to obtain the first reconstructed feature of the target domain. Use the mean square error loss between the second reconstructed feature of the source domain and the normalized feature matrix of the source domain battery as the mean square error loss of the source domain feature reconstruction. Use the mean square error loss between the first reconstructed feature of the target domain and the normalized feature matrix of the target domain battery as the mean square error loss of the target domain feature reconstruction. Add the mean square error loss of the source domain feature reconstruction and the mean square error loss of the target domain feature reconstruction as the second reconstruction loss. S36. The domain discrimination loss, feature statistics measurement loss, and second reconstruction loss are weighted and summed to form the total loss for feature transfer training. The weight parameters of the second encoder, second reconstructor, and domain discriminator are updated based on minimizing the total loss for feature transfer training.

7. The fast health status estimation method for decommissioned batteries based on transfer learning as described in claim 4, characterized in that, In step S4, the specific process of fine-tuning is as follows: S41. Transfer the pre-trained weight parameters of the first predictor to the second predictor as the initial weight parameters of the second predictor; transfer the weight parameters of the second encoder trained by feature transfer to the third encoder as the initial weight parameters of the third encoder; S42. Input the normalized feature matrix of the target battery into the third encoder, and the third encoder outputs the second deep feature of the target domain. S43. Input the second deep feature of the target domain into the second predictor to predict the label and obtain the predicted label of the target domain. Use the mean squared error loss between the predicted label of the target domain and its true label as the total loss for fine-tuning, and update the weight parameters of the third encoder and the second predictor based on minimizing the total loss for fine-tuning.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it can implement the fast estimation method for the health status of retired batteries based on transfer learning as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the rapid estimation method for the health status of decommissioned batteries based on transfer learning as described in any one of claims 1 to 7.

10. A computer electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the fast estimation method for the health status of decommissioned batteries based on transfer learning as described in any one of claims 1 to 7.

Images