Method and system for predicting health state of battery under small amount of data
By applying transfer learning on special-purpose battery models, the prediction model of battery models in a wide application scenario is migrated to the battery models in a narrow application scenario, which solves the problem of low prediction accuracy of battery health status under a small amount of data and achieves higher predictions Accuracy and computational efficiency.
Patent Information
- Application Number
- CN202411332542.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-24
- Publication Date
- 2025-02-07
AI Technical Summary
Due to the narrow use range of special-purpose battery models, it is difficult to accurately predict their health status under a small amount of data, which reduces the accuracy of the battery health status prediction model.
Using the battery health status prediction method based on transfer learning, the source domain task model is trained using the full life data of battery models with wide application scenarios (type B batteries) under the same type, and migrating it to the battery model with narrow application scenarios (A ) fine-tune to improve prediction accuracy.
Through transfer learning, the accuracy of the health status prediction model of A-type battery can be significantly improved under a small amount of data, the computing efficiency and model performance are improved, and the problem of low training model accuracy caused by small samples is solved.
Smart Images
Figure CN119395582A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of battery health management, and in particular to a method and system for predicting the health status of a battery using a small amount of data. Background Art
[0002] The current battery health status prediction methods mainly include SOH (State of Health) prediction methods based on battery principle models, SOH prediction methods based on statistical algorithm models, and SOH prediction methods based on data-driven models. Among them, the data-driven method does not require the establishment of a complex battery physical model, but only reflects the current SOH information of the battery and its possible evolutionary laws based on the battery's state data such as voltage, current, temperature, etc. during operation. These data not only reflect the impact of the working environment and working mode on the battery, but also reflect the differences and individual states of different battery cells, and have been widely used.
[0003] The accuracy of the prediction model of the data-driven battery health status prediction method depends largely on the richness and effectiveness of the data samples. However, for some special-purpose battery models, due to their narrow scope of use, the amount of full-life usage data that can be collected for this type of battery is likely to be low, making it difficult to effectively train the battery health status prediction model, resulting in reduced prediction accuracy. Summary of the invention
[0004] In view of the limitations of some special-purpose battery models due to their narrow scope of use and low number of users, the prediction of battery health status based on data-driven methods is not accurate enough. A battery health status prediction method based on transfer learning with a small amount of data is proposed to improve the prediction accuracy of the health status of such batteries.
[0005] Based on the above technical ideas, the technical solution adopted by the present invention is:
[0006] One object of the present invention is to provide a method for predicting the health status of a battery with a small amount of data. The battery model with a narrow application scenario under the same type is a type A battery, and the battery model with a wide application scenario under the same type is a type B battery. The prediction model of the health status of the type A battery is used as the target domain task model, and the prediction model of the health status of the type B battery is used as the source domain task model. The prediction method for the health status of the type A battery includes:
[0007] Step 1: Use the full life cycle usage data of type B batteries as source domain sample data, use the source domain sample data to train the source domain task model, and obtain the structure and parameters of the type B battery health status prediction model;
[0008] Step 2: Load the saved source domain task model as the target domain task model, and directly migrate or randomly initialize the weight parameters of the target domain task model according to different migration schemes;
[0009] Step 3: Use the full life cycle usage data of type A batteries as the target domain sample data and preprocess the target domain sample data;
[0010] Step 4: According to the migration plan, the back propagation algorithm is used to fine-tune and freeze the parameters of each layer of the target domain task model using the target domain sample data to complete the training of the target domain task model;
[0011] Step 5: Use the trained target domain task model to predict and output the type A battery health status test samples.
[0012] On the basis of the above technical solution, in step 1, a CNN-BiLSTM-Attention model is established, a CNN-BiLSTM-Attention model is established, the CNN-BiLSTM-Attention model is used as the source domain task model, and the source domain sample data is preprocessed, and training is performed using the preprocessed source domain sample data, and the structure and parameters of the trained B-type battery health status prediction model are saved.
[0013] On the basis of the above technical solution, further, in step 2, a migration scheme of the source domain task model is designed according to whether the source domain task model is added with an adaptive layer. A total of 6 migration schemes are designed.
[0014] On the basis of the above technical solutions, the six migration solutions are as follows:
[0015] Solution 1: Do not add the adaptive layer, freeze the convolutional layer, BiLSTM layer, and attention layer in the source domain task model, and fine-tune and train the remaining network layers;
[0016] Solution 2: Add an adaptive layer, freeze the convolutional layer, BiLSTM layer, and attention layer in the source domain task model, and fine-tune and train the remaining network layers;
[0017] Solution 3: Do not add the adaptive layer, migrate and freeze the parameters of the convolutional layer, BiLSTM layer, and attention layer in the source domain task model, and randomly initialize the parameters of the fully connected layer before training;
[0018] Solution 4: Add an adaptive layer, migrate and freeze the parameters of the convolutional layer, BiLSTM layer, and attention layer in the source domain task model, randomly initialize the parameters of the fully connected layer, and then train the parameters of the fully connected layer and the adaptive layer;
[0019] Solution 5: Do not add the adaptive layer and fine-tune the parameter values of all network layers in the source domain task model;
[0020] Solution 6: Add an adaptive layer to fine-tune the parameter values of all network layers in the source domain task model.
[0021] On the basis of the above technical solution, further, in step 3, the preprocessed target domain sample data is divided into training samples and test samples of the target domain; in step 4, the training samples are used to complete the training of the target domain task model; in step 5, the prediction accuracy of the target domain prediction model is verified by the test samples.
[0022] Another object of the present invention is to provide a battery health status prediction system with a small amount of data applied to the above prediction method, the prediction system comprising:
[0023] The source domain task model training module is used to establish the CNN-BiLSTM-Attention model. It is trained through source domain data samples to obtain the structure and parameters of the source domain task model, and then the trained source domain task model is imported into the target domain prediction model establishment module.
[0024] The target domain prediction model establishment module selects the migration scheme and uses the target domain sample data through the back propagation algorithm to fine-tune and freeze the parameters of each layer of the target domain task model, complete the training of the source domain task model, and obtain the target domain prediction model;
[0025] The target battery health status prediction module calls the target domain prediction model to predict and output the target battery health status.
[0026] Compared with the prior art, the present invention has the following advantages:
[0027] 1. The present invention migrates and applies the source domain task model, eliminating the need for retraining with huge sample data, thereby solving the problem of low training model accuracy caused by a small number of samples. The network structure of the source domain model is used as the basis of the target domain model, and only the source domain network model and parameters need to be fine-tuned according to the task characteristics. Compared with building and training a new network with random initialization, the present invention has higher computational efficiency and better model performance.
[0028] 2. Through transfer learning, the parameters of each layer of the target prediction model of the present invention can transfer the empirical knowledge of the source domain to the target domain, which helps to enhance the prediction ability of the target domain model. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 Schematic diagram of the transfer learning process.
[0030] Figure 2 This is the structure diagram of the transfer learning model corresponding to Scheme 1.
[0031] Figure 3 Flowchart of battery health status prediction based on transfer learning. DETAILED DESCRIPTION
[0032] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.
[0033] For different models of batteries of the same type (such as lithium batteries, but with different capacities), their working principles are the same and the laws of characteristic changes are similar. The variable data structures and characteristics recorded by each model of battery are basically the same (such as the same sampling frequency for collecting current, voltage, temperature and other parameters). Therefore, although the usage data of batteries of different models are not exactly the same, there is a strong correlation. We can consider fully utilizing the usage data information value of various models of batteries of the same type through transfer learning, and fully explore the common feature information contained in the usage data of different models of batteries, thereby improving the accuracy of the SOH prediction model of various models of batteries, especially those for special purposes.
[0034] Transfer learning refers to applying the experiential knowledge gained through learning and training in the old environment to the new model based on the correlation between data, tasks or models in the new and old environments, so as to improve the learning effect of the model in the new environment. The process diagram is as follows Figure 1 shown.
[0035] The basic components of transfer learning include: domain and task. The acquired data and model knowledge and experience are called source domain, and the source domain dataset is represented as: s =[X s ,P(X s )], where X s represents the source domain data sample, P(X s ) represents the probability distribution of the source domain data; the data and model knowledge to be learned are called the target domain, and the target data set is represented as: D t =[X t ,P(X t )], where X t represents the target domain data sample, P(X t ) represents the probability distribution of the target domain data. The model used to solve the problem is called a task, which is expressed as T = [y, f(x)], where y represents the label space and f(x) represents the target prediction function.
[0036] According to the concepts of domain and task, the transfer learning process can be expressed as: for the source domain space Ds And the corresponding trained model T s , when the target domain space D t ≠D s or T t ≠T s Under the condition of s and T s Useful experience knowledge in the target domain can improve the task T t The prediction function f t (x) performance, and improve the prediction ability of the target task. In this patent example, the full life cycle usage data of a type A battery of the same battery type is used as the source domain sample data D s , taking the full life usage data of type B batteries as the target domain sample data D t , taking the prediction model of A-type battery SOH as the source domain task T s , taking the prediction model of B-type lead-acid battery SOH as the target domain task T t , through transfer learning of source domain tasks, the accuracy of the target domain model is improved.
[0037] In the process of transfer learning, the design of the migration scheme is very important to improve the effect of transfer learning. The migration operations of the source domain model in the migration process mainly include "fine-tuning", "freezing" and adding adaptive layers. The selection of which operation method is more conducive to improving the prediction accuracy has a complex relationship with data similarity, data structure, data quality, etc.
[0038] "Fine-tuning" and "freezing" are common operations for parameters of each layer of the target domain model. Freezing means that during the training of the target domain model, the parameters of a certain layer in the source domain model network are fixed. Its main advantage is to retain the experiential knowledge learned by the corresponding layer during the source domain training process. Fine-tuning means that the parameters of a certain layer of the target domain model are not randomly initialized, but the parameters of the source domain model are directly migrated, and are continuously updated and optimized as the model is trained. Its advantages are fast convergence speed, short training time, and the trained model has strong robustness and generalization ability. For the adaptive layer added to the migration model, it means adding a fully connected layer before the output layer of the source domain model, and randomly initializing the weights of this layer.
[0039] Based on the above research, a battery health status prediction method based on transfer learning with a small amount of data is proposed, which mainly includes the following aspects:
[0040] First: Selection of source domain model method. In the process of deep learning, the learned feature knowledge can be divided into general features and abstract features. The shallow network of deep learning can often effectively extract the general features of sample data, while the deep network mainly extracts the abstract features of sample data. Therefore, the selection of the source domain model requires the use of a deep learning model. Here, the CNN-BiLSTM-Attention model is taken as an example. The source domain model contains three functional layers, namely the convolution layer, the BiLSTM layer, and the attention layer, which perform calculations in sequence.
[0041] Second: Design of the migration plan: Based on the selected source domain model, formulate a migration learning plan for the target domain model.
[0042] Depending on whether an adaptive layer is added, the training scheme used for the parameters of each layer is "fine-tuning" or "freezing" and other operations. The following six transfer learning schemes are designed for the source domain prediction model. Adding an adaptive layer is equivalent to adding a fully connected layer after the attention layer, which can be called an adaptive layer.
[0043] Solution 1: Do not add the adaptive layer, freeze the convolutional layer, BiLSTM layer and attention layer, and fine-tune and train the remaining network layers. The transfer learning model structure corresponding to Solution 1 is as follows: Figure 2 shown.
[0044] Solution 2: Add an adaptive layer, freeze the convolutional layer, BiLSTM layer, and attention layer, and fine-tune and train the remaining network layers.
[0045] Solution 3: Migrate and freeze the parameters of the convolutional layer, BiLSTM layer, and attention layer, do not add the adaptive layer, and randomly initialize the parameters of the fully connected layer before training.
[0046] Solution 4: Migrate and freeze the parameters of the convolutional layer, BiLSTM layer, and attention layer, add an adaptive layer, randomly initialize the parameters of the fully connected layer, and then train the parameters of the fully connected layer and the adaptive layer.
[0047] Solution 5: Do not add the adaptive layer and fine-tune the parameter values of all layers of the model.
[0048] Solution 6: Add an adaptive layer to fine-tune the parameter values of all layers of the model.
[0049] Third: Training and prediction of target domain models according to different transfer learning schemes.
[0050] Select the training method of the source domain in the transfer learning process (taking the CNN-BiLSTM-Attention model as an example), use the source domain data to train the source domain model, train the target domain model according to different transfer learning schemes and target domain data, and carry out predictions in the target domain. Battery models with narrower applications under the same type are recorded as type A batteries, and the usage data that can be collected is relatively small. Battery models with more application scenarios under the same type are recorded as type B batteries, and the usage data that can be collected is relatively rich. The prediction model of the health status of type A batteries is used as the target domain task, and the prediction model of the health status of type B batteries is used as the source domain task. The health status prediction process of type A batteries based on transfer learning methods is proposed as follows: Figure 3 The specific steps are as follows:
[0051] (1) Preprocess the source domain data samples, establish the CNN-BiLSTM-Attention model, and train it using the source domain data samples.
[0052] (2) Save the structure and parameters of the trained B-type battery health status prediction model.
[0053] (3) Load the saved source domain model structure as the target domain task model, and directly migrate or randomly initialize the weight parameters of each layer according to different migration schemes.
[0054] (4) Preprocess the target domain data samples and divide the target domain into training samples and test samples.
[0055] (5) According to the migration plan, during the training process, the back propagation algorithm is used to use the training samples of the target domain to perform operations such as fine-tuning and freezing on the parameters of each layer of the target domain model to complete the training of the target domain model.
[0056] (6) The prediction accuracy of the target domain model is verified through the test samples of the target domain, that is, the trained target domain model is used to complete the prediction and output of the health status test samples of type A batteries.
[0057] Fourth: Based on the prediction results of different transfer learning schemes, the most suitable transfer learning scheme is adopted as the prediction method under small amount of battery data according to the needs.
[0058] Case
[0059] Take a certain type of flooded lead-acid large-capacity battery used in industry as an example. This type of battery is mainly used in industrial marine batteries. It has a large capacity, but its scope of use is narrow, and there are differences in battery models. Therefore, the amount of data collected for the same type of battery usage is relatively small. However, for batteries of the same type, their working principles are the same and the laws of characteristic changes are similar. The variable data structures and characteristics recorded by each model of battery are basically the same. Consider using this patented method to complete the health status prediction of such batteries.
[0060] Take one type of battery in this type of battery as type A battery, and the other type of battery as type B battery. The prediction model of the health status of type A battery is used as the target domain task, and the prediction model of the health status of type B battery is used as the source domain task. A health status prediction model of type B battery based on the CNN-BiLSTM-Attention method is established in the source domain, and the transfer learning process adopts six designed transfer schemes.
[0061] Type A batteries do not use transfer learning methods, and directly carry out training and prediction based on the CNN-BiLSTM-Attention method according to the usage data of Type A batteries, which is recorded as Scheme X; Type A batteries use the six transfer learning methods described in this patent at a time, which are recorded as Schemes 1 to 6, and the mean absolute error (MAE), mean absolute percentage error (MAPE) and root mean square error (RMSE) are used to measure the prediction performance of the algorithm.
[0062] in:
[0063]
[0064] where x i and t i They represent the predicted value and actual value of the i-th sample respectively, and N represents the number of samples.
[0065] The prediction accuracy under different schemes is shown in Table 1:
[0066] Table 1 Prediction accuracy of SOH of type A battery under different prediction schemes
[0067] Prediction accuracy Plan X Solution 1 Solution 2 Solution 3 Solution 4 Solution 5 Solution 6 MAE 0.2565 0.1247 0.1196 0.1644 0.1333 0.1021 0.0925 MAPE 0.0351 0.0169 0.0162 0.0225 0.0183 0.0150 0.0127 MSE 0.0976 0.0218 0.0202 0.0364 0.0268 0.0157 0.0133
[0068] Through the prediction results, the following conclusions are drawn: The prediction accuracy of the A-type battery health status prediction model based on transfer learning has been significantly improved, indicating that through transfer learning, the parameters of each layer of the model can transfer the empirical knowledge of the source domain to the target domain, which helps to enhance the prediction ability of the target domain model. Among the six transfer learning schemes, the sixth scheme has the highest prediction accuracy and is more suitable for the data of the battery in this example. The transfer learning method based on this scheme is selected as the prediction method of battery SOH.
[0069] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, so that those skilled in the art can understand and apply the present invention. It cannot be determined that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art of the present invention, some simple deductions or substitutions can be made without departing from the concept of the present invention, without creative work. Therefore, simple improvements made to the present invention by those skilled in the art based on the disclosure of the present invention should be within the scope of protection of the present invention.
Claims
1. A method for predicting the health status of a battery using a small amount of data, characterized in that: The battery model with narrow application scenarios under the same type is type A battery, and the battery model with broad application scenarios under the same type is type B battery. The prediction model of the health status of type A battery is used as the target domain task model, and the prediction model of the health status of type B battery is used as the source domain task model. The prediction method of the health status of type A battery includes: Step 1: Use the full life cycle usage data of type B batteries as source domain sample data, use the source domain sample data to train the source domain task model, and obtain the structure and parameters of the type B battery health status prediction model; Step 2: Load the saved source domain task model as the target domain task model, and directly migrate or randomly initialize the weight parameters of the target domain task model according to different migration schemes; Step 3: Use the full life cycle usage data of type A batteries as the target domain sample data and preprocess the target domain sample data; Step 4: According to the migration plan, the back propagation algorithm is used to fine-tune and freeze the parameters of each layer of the target domain task model using the target domain sample data to complete the training of the target domain task model; Step 5: Use the trained target domain task model to predict and output the type A battery health status test samples.
2. The method for predicting the health status of a battery using a small amount of data according to claim 1, characterized in that: In step 1, a CNN-BiLSTM-Attention model is established, and the CNN-BiLSTM-Attention model is used as the source domain task model. The source domain sample data is preprocessed, and training is performed using the preprocessed source domain sample data. The structure and parameters of the trained B-type battery health status prediction model are saved.
3. The method for predicting the health status of a battery using a small amount of data according to claim 2, characterized in that: In step 2, a migration scheme for the source domain task model is designed based on whether the source domain task model has an adaptive layer. A total of 6 migration schemes are designed.
4. The method for predicting the health status of a battery using a small amount of data according to claim 3, characterized in that: The six migration plans are: Solution 1: Do not add the adaptive layer, freeze the convolutional layer, BiLSTM layer, and attention layer in the source domain task model, and fine-tune and train the remaining network layers; Solution 2: Add an adaptive layer, freeze the convolutional layer, BiLSTM layer, and attention layer in the source domain task model, and fine-tune and train the remaining network layers; Solution 3: Do not add the adaptive layer, migrate and freeze the parameters of the convolutional layer, BiLSTM layer, and attention layer in the source domain task model, and randomly initialize the parameters of the fully connected layer before training; Solution 4: Add an adaptive layer, migrate and freeze the parameters of the convolutional layer, BiLSTM layer, and attention layer in the source domain task model, randomly initialize the parameters of the fully connected layer, and then train the parameters of the fully connected layer and the adaptive layer; Solution 5: Do not add the adaptive layer and fine-tune the parameter values of all network layers in the source domain task model; Solution 6: Add an adaptive layer to fine-tune the parameter values of all network layers in the source domain task model.
5. The method for predicting the health status of a battery using a small amount of data according to claim 4, characterized in that: In step 3, the preprocessed target domain sample data is divided into training samples and test samples of the target domain; in step 4, the training samples are used to complete the training of the target domain task model; in step 5, the prediction accuracy of the target domain prediction model is tested through the test samples.
6. A battery health status prediction system with a small amount of data applied to the prediction method according to any one of claims 1 to 5, characterized in that: The system includes: The source domain task model training module is used to establish the CNN-BiLSTM-Attention model. It is trained through source domain data samples to obtain the structure and parameters of the source domain task model, and then the trained source domain task model is imported into the target domain prediction model establishment module. The target domain prediction model establishment module selects the migration scheme and uses the target domain sample data through the back propagation algorithm to fine-tune and freeze the parameters of each layer of the target domain task model, complete the training of the source domain task model, and obtain the target domain prediction model; The target battery health status prediction module calls the target domain prediction model to predict and output the target battery health status.
Citation Information
Cited By
Ship power lithium battery health state prediction method based on hybrid depth model
CN121008189A
Ship power lithium battery state of health prediction method based on hybrid deep model
CN121008189B