A method for estimating the state of health of lithium batteries when new lithium battery data is unavailable.

Abstract

This invention discloses a method for estimating the health status of lithium batteries under conditions where new lithium battery data is unavailable, belonging to the field of lithium battery health status assessment. The invention first performs simple preprocessing on the original aging data, and then uses a feature extractor composed of convolutional network modules and residual network modules to extract deep features from the aging data. During feature extraction, domain-specific aging features are randomly forgotten. Finally, an adaptive consistency loss is used to enhance the model's robustness to feature fluctuations and promote the model's learning of domain-invariant aging features, thereby enhancing the generalization ability of the health status assessment model under conditions where new lithium battery data is unavailable. This invention specifically addresses health status assessment under conditions where new lithium battery data is unavailable, resulting in a health status assessment model with low error, high generalization ability, and the ability to effectively monitor the health status of lithium batteries in real-world usage scenarios, thus improving equipment safety.

Description

Technical Field

[0001] This invention relates to the field of lithium battery health status assessment, and in particular to a method for estimating the health status of lithium batteries when new lithium battery data is unavailable. Background Technology

[0002] Lithium-ion batteries, as crucial energy devices, are widely used in electric vehicles, mobile communications, and energy storage systems. The health status of a lithium battery reflects its degree of aging relative to its original manufacturing condition. Accurate lithium battery health status assessments can help manufacturers and users better manage and maintain battery products, promote the development of renewable energy and electric vehicles, and foster innovation in battery materials and processes. Given that a large number of electric vehicle fires are caused by battery aging, monitoring battery status through health status assessments to ensure equipment safety is of paramount importance.

[0003] Due to factors such as the inherent characteristics of lithium batteries, their operating environment, and usage frequency, the aging process of different lithium batteries varies significantly. This difference results in inconsistent distribution of aging data collected from different batteries, causing a substantial decrease in model accuracy when a customized model trained for one battery is directly applied to another. Therefore, traditional lithium-ion battery health status assessment algorithms require collecting a large amount of aging data from new lithium batteries and training a model specifically for that battery from scratch.

[0004] The application environments of lithium batteries are diverse, and the specific application environments of different batteries are unknown. Aging data collection is time-consuming and difficult. When data for new batteries cannot be collected or the amount of collected data is insufficient, it becomes difficult to train a health status assessment model for new batteries. In such cases, it is necessary to fully utilize existing lithium battery aging data to train a highly generalizable model that can be applied to all lithium batteries for which aging data is unavailable, ensuring low prediction errors.

[0005] Domain generalization techniques enhance a model's performance on unseen individuals by training it on multiple individual datasets relevant to the target task. In the field of lithium-ion battery health status assessment, this technique can utilize datasets of multiple aged batteries to enhance the model's generalization ability. The trained generalized model can then accurately assess the health status of the target battery when faced with lithium-ion batteries under unknown aging conditions in real-world applications.

[0006] Some studies have expanded the diversity of the source domain by generating pseudo-domains through data augmentation and adversarial examples, enabling the model to learn more generalized feature extraction capabilities from diverse samples. However, these methods have a key problem: the quality of the generated pseudo-samples cannot be guaranteed. In lithium battery health status assessment tasks, the labels of different lithium battery aging data also differ, i.e., there are differences in conditional distribution, which further increases the uncertainty of pseudo-sample quality. When there are too many unreasonable pseudo-samples, it may lead to unstable or degraded model performance.

[0007] Another group of researchers uses strategies such as adversarial learning and meta-learning to learn domain-invariant aging features to align the aging feature distributions of various lithium batteries. Existing research primarily imposes constraints on the entire network to supervise the model's learning of domain-invariant aging features. However, these methods do not focus on the aging features learned at each layer of the network, leading to the model inevitably learning domain-specific features and limiting further improvements in model robustness. Therefore, it is essential to research a lithium battery health state estimation method applicable to conditions where new lithium battery data is unavailable. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a method for estimating the health status of lithium batteries when new lithium battery data is unavailable. This method introduces a domain discriminator in the intermediate layer of the feature extractor to locate domain-specific channels, forgets domain-specific aging features, and enhances the robustness of the model to feature fluctuations through adaptive consistency loss.

[0009] The technical solution adopted in this invention is as follows:

[0010] A method for estimating the state of health of lithium batteries when new lithium battery data is unavailable, comprising the following steps:

[0011] Step S1: Preprocess the raw aging data of the lithium battery, and pass the preprocessed aging data x through a feature extractor and a domain discriminator to obtain the domain-invariant features f related to the health status assessment.

[0012] Step S2: During each network training, a residual network module is randomly selected, and specific aging features of different channels in the forgetting domain are randomly selected according to the weighted random selection algorithm; the weighted random selection algorithm is executed twice, and finally two aging features f1(x) and f2(x) are obtained.

[0013] Step S3: Input f1(x) and f2(x) into a classifier network C (i.e., prediction network C) composed of multiple fully connected layers to obtain the health status assessment result and calculate the mean squared error. and adaptive consistency loss

[0014] Step S4: Based on the mean square error L mse And adaptive consistency loss L acl Calculate the total network loss L, and optimize the network parameters based on the total network loss L to obtain a trained lithium battery health state estimation network suitable for conditions where new lithium battery data is unavailable.

[0015] Step S5: Input the real-time aging data x of the new battery into the trained lithium battery health state estimation network. i The corresponding health state e(x) is obtained by sequentially passing through a feature extractor and a classifier network C. i ).

[0016] Furthermore, in step S1, preprocessing includes denoising the original aging data and extracting artificial features.

[0017] The feature extractor comprises a cascaded convolutional network module and several residual network modules. The convolutional network module includes cascaded convolutional layers, batch normalization layers, ReLU activation layers, and max pooling layers. The residual network modules also include cascaded convolutional layers, batch normalization layers, and ReLU activation layers, with short connections between the inputs of the convolutional layers and the outputs of the ReLU activation layers. Specifically, the output of each residual network module is obtained by fusing the inputs of the convolutional layers and the outputs of the ReLU activation layers. During training, the output of each residual network module is also connected to a domain discriminator.

[0018] Furthermore, the domain discriminator consists of cascaded global average pooling layers and fully connected layers. The domain discriminator is used to determine whether the current input features include too many domain-specific aging features. If so, features from some channels are forgotten in a certain proportion.

[0019] Furthermore, the discrimination process of the domain discriminator is as follows:

[0020] Define P c Let P represent the probability that the domain discriminator correctly identifies any feature channel of the aging feature (the input feature of the domain discriminator, i.e., the output feature of a certain layer of residual network module) as originating from a certain domain, i.e., the discrimination success rate. c Threshold P greater than the threshold threshold If so, it is considered that the aging characteristics of the current channel contain too many domain-specific features;

[0021] Calculate the success rate P of discrimination based on the current input features. c The number of channels n, if the channel ratio P n =n / t exceeds a specified threshold, then according to a certain forgetting ratio, n samples are forgotten, resulting in a success rate P. c The features of some channels in the domain discriminator are not forgotten; otherwise, the domain discriminator outputs the input features as is, i.e., without forgetting them.

[0022] Preferably, the ratio threshold is set to Among them, P forget This is the preset forgetting rate.

[0023] Furthermore, the weighted random selection algorithm randomly selects channels based on the key values ​​obtained from the domain discriminator, specifically as follows:

[0024] Get the key values ​​of each channel in, It is the weight of each channel output by each domain discriminator, r j It is a random number between (0,1);

[0025] Select the m channels with the largest key values, set their corresponding masks to 0, and set the masks of the other channels to 1 to obtain the channel mask matrix;

[0026] The aging characteristics after the weighted random selection algorithm is obtained by multiplying the channel mask matrix with the output characteristics of the residual network module.

[0027] The technical solution provided by this invention brings at least the following beneficial effects:

[0028] The health status assessment method of the present invention has high generalization and low error. It does not require aging data of new lithium batteries and is suitable for health status assessment under conditions where aging data of new lithium batteries cannot be collected. It can be easily applied to any health status assessment scenario of lithium batteries. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a schematic diagram illustrating the processing steps of a lithium battery health status estimation method applicable to situations where new lithium battery data cannot be obtained, as provided in an embodiment of the present invention.

[0031] Figure 2 This is a schematic diagram of the discrimination process used in the embodiments of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described in detail and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Generally, the components of the embodiments of the present invention described and shown in the accompanying drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present invention.

[0033] This invention provides a method for estimating the health status of lithium batteries under conditions where new lithium battery data is unavailable. First, the original aging data undergoes simple preprocessing. Then, a feature extractor composed of convolutional and residual network modules is used to extract deep features from the aging data. During feature extraction, domain-specific aging features are randomly forgotten. Finally, an adaptive consistency loss method is used to enhance the model's robustness to feature fluctuations and promote the model's learning of domain-invariant aging features, thereby improving the generalization ability of the health status assessment model under conditions where new lithium battery data is unavailable. The method proposed in this invention is specifically designed for health status assessment under conditions where new lithium battery data is unavailable, resulting in a health status assessment model with low error, high generalization ability, and the ability to effectively monitor the health status of lithium batteries in real-world usage scenarios, thus improving equipment safety.

[0034] See Figure 1 The lithium battery health status estimation method provided in this embodiment of the invention, applicable to situations where new lithium battery data is unavailable, includes the following steps:

[0035] Step S1: Preprocess the raw aging data of the lithium battery, and pass the preprocessed aging data x through a feature extractor and a domain discriminator to obtain the domain-invariant features f related to the health status assessment.

[0036] Specifically, the preprocessing includes denoising the original aging data and extracting artificial features. The artificial features include constant current charging time, constant voltage charging time, time interval of the same voltage, maximum value of incremental capacity analysis curve and capacity of constant current charging stage, to obtain the input aging data x of the model.

[0037] Specifically, the feature extractor includes one convolutional network module and three residual network modules:

[0038] The feature extractor consists of an aged data x input to a convolutional network module, which passes through a convolutional layer, a batch normalization layer, a ReLU activation layer, and a max pooling layer. This input is then passed to a first residual network module, which also passes through a convolutional layer, a batch normalization layer, and a ReLU activation layer. The inputs to the convolutional layers and the outputs of the ReLU activation layers in the residual network module are set as short connections (i.e., skip connections). The first residual network module yields feature f1(x). During training, the features of each channel of f1(x) are input to a domain discriminator composed of a global average pooling layer and a fully connected layer. If the domain discriminator determines that f1(x) contains too many domain-specific aged features, then a certain proportion of the features from some channels of f1(x) are forgotten to obtain h1(x).

[0039] The specific discrimination process of the domain discriminator is as follows: Figure 2 As shown, it includes: using P c The hyperparameter P represents the probability that the domain discriminator correctly identifies an aging feature as originating from a certain domain. threshold Indicates the discrimination threshold; when P c Greater than the set threshold P threshold At that time, it was considered that the features of the channel contained too many domain-specific features; if the network layer has n channels, the discrimination success rate P c Greater than P threshold , using P n =n / t represents the proportion of this part of the feature channels to the total feature channels (t), P forget =m / t represents the number of channels to be forgotten (preset value). For example, when If this happens, it's considered that too many channels in the network layer have learned domain-specific features, and channels containing too many domain-specific features in that layer need to be forgotten to ensure the network learns more domain-invariant features during forward propagation. At that time, the aging characteristics of the network layer should not be forgotten.

[0040] The aging data x is input into the first residual module to obtain feature f1(x), and then feature f1(x) is input into the domain discriminator to obtain the forgetting domain specific aging feature to obtain h1(x); then h1(x) is input into the second residual network module to obtain feature f2(x), and feature f2(x) is input into the domain discriminator to obtain the forgetting domain specific aging feature to obtain h2(x); then h2(x) is input into the third residual network module to obtain f3(x), and f3(x) is input into the domain discriminator to obtain the forgetting domain specific aging feature to obtain the final aging feature f.

[0041] Step S2: During each network training iteration, a residual network module is randomly selected, and specific aging features of different channels in the forgetting domain are randomly selected using a weighted random selection algorithm. The weighted random selection algorithm is executed twice, resulting in two aging features f1(x) and f2(x).

[0042] Specifically, forgetting all the features learned by the three residual network modules may cause the model to fail to fit. Using a random training strategy, specific aging features of the residual network modules to be forgotten are randomly selected during each training session. The weighted random selection algorithm randomly selects modules based on the key values ​​of each channel obtained by the domain discriminator.

[0043] The key value of each channel is

[0044] in, It is the weight of each channel output by each domain discriminator, r j It is a random number between (0,1).

[0045] After calculating the key values ​​for all channels, select the m channels with the largest key values ​​and set their corresponding masks to 0.

[0046]

[0047] Where Max(v1,v2,…,v) t ) m This represents the m largest key values ​​out of all t channels. (mask) j =0 indicates that the characteristics of that channel have been forgotten, mask j =1 indicates that the characteristics of the channel are preserved.

[0048] Because in the above weighted random selection algorithm, r j The key value is a random number between (0,1), and the key value of each channel will have a certain degree of randomness. The weighted random selection algorithm will be executed twice, and finally two different aging features will be obtained. The two different aging features will be fed into the next layer of the network, and after passing through the last layer of the feature extractor, features f1(x) and f2(x) will be obtained.

[0049] Step S3: Input the deep feature f into a classifier network C composed of multiple fully connected layers to obtain the predicted feature class. Use Softmax cross-entropy to calculate the loss L between the predicted class and the true class. ce The center loss L is obtained by calculating the feature center of each category. center :

[0050] Specifically, the cross-entropy loss is:

[0051] Where n represents the number of samples in the training set, and I(i = yi) is a binary function, which is used when predicting categories i and y. i The value is 1 if they are equal, otherwise it is 0, p(f) i ) Calculate feature f i The classification probability output is p(·) = softmax(C(·)).

[0052] The central loss is:

[0053] Among them, f i (fc) Let c be the output feature of the second-to-last fully connected layer in the classifier for the i-th sample. yi c represents the center point of all sample features of the category corresponding to sample i. yi The dimension of the matrix is ​​the same as the dimension of the feature. Constraining the distance of each feature to its corresponding center causes features within a class to move closer to the center, becoming more clustered, and thus increasing the distance between different classes.

[0054] Specifically, the average features of each class in the mini-batch are used to approximate the average features of all samples. When calculating the class center derivative, only the sample features of a specific class in the current batch are used to obtain the update amount of that class center. The change in each class center is calculated using only the sample features belonging to that class. In particular, to prevent learning perturbations caused by mislabels, a parameter α is used to control the class center c. yi The learning rate. L center The gradient is expressed as:

[0055]

[0056] The denominator is incremented by 1 to prevent the denominator from being zero if there are no samples of a certain class in the mini-batch.

[0057] Step S3: Input f1(x) and f2(x) into the predictor network C, which consists of multiple fully connected layers, to obtain the health status assessment result. Figure 1 The figure shows a curve comparison between predicted and measured values, and the mean square error is calculated. and adaptive consistency loss

[0058] Specifically, the mean square error is:

[0059] Where, n k It represents the number of aging data samples for the k-th lithium battery. This is the true label corresponding to the sample, P(F(f) i k )1 and P(F(f i k )2 are the health status assessment results corresponding to features f1(x) and f2(x), respectively.

[0060] As shown in steps S1 and S2, different aging features will be extracted from the aging data of the same lithium battery. This feature fluctuation caused by the forgetting domain-specific aging features will further lead to different predictor outputs. Adaptive Consistency Loss (ACL Loss) is used to enhance the model's robustness to feature fluctuations and promote the learning of domain-invariant aging features:

[0061]

[0062] Here, ξ is used to determine the emphasis of the adaptive consistency loss at different training stages, and its value is an empirical value. However, when |P(F(x)1)-P(F(x)2)|>ξ, the model error is relatively large, and the ACL loss can be approximated as the mean absolute error, ensuring that the model updates parameters at a relatively fast speed. When |P(F(x)1)-P(F(x)2)|≤ξ, the ACL loss can be approximated as the mean squared error, which can detect outliers with large differences in the prediction results, ensuring that the model searches for the global optimum with lower precision.

[0063] Step S4: Based on the mean square error L mse And adaptive consistency loss L acl The total network loss is calculated, and the network parameters are optimized based on the total loss L to obtain a lithium battery health state estimation network (a cascaded feature extractor and predictor network C) suitable for conditions where new lithium battery data is unavailable. Specifically, the total loss...

[0064] Step S5: Input the real-time aging data x of the new battery into the trained lithium battery health state estimation network. i The corresponding health status e(x) is obtained. i ).

[0065] Specifically, the lithium battery health state estimation network can be represented as:

[0066] e(x i )=fc(f(x i ))

[0067] Where f(·) represents x i The deep features extracted by the feature extractor, fc represents the output obtained by inputting the deep features into the predictor network C.

[0068] In this embodiment, the optimizer used during network optimization is ADAM, and the learning rate search range is 10^ (-4) Up to 10^ (-1) .

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0070] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for estimating the state of health of lithium batteries when new lithium battery data is unavailable, characterized in that: Includes the following steps: Step S1: Preprocess the original aging data of lithium batteries, and use the preprocessed aging data x through a feature extractor and a domain discriminator to obtain the domain invariant features f related to the health status assessment. The domain discriminator consists of a cascaded global average pooling layer and a fully connected layer. The domain discriminator is used to determine whether the current input features include too many domain-specific aging features. If so, some of the features of its channels are forgotten in a certain proportion. Step S2: During each training session, a residual network module is randomly selected, and specific aging features of different channels in the forgetting domain are randomly selected according to the weighted random selection algorithm; the weighted random selection algorithm is executed twice, and finally two aging features f1(x) and f2(x) are obtained. Step S3: Input f1(x) and f2(x) into a classifier network C composed of multiple fully connected layers to obtain the health status assessment results and calculate the mean squared error. mse and adaptive consistency loss acl ; Step S4: Based on the mean square error L mse And adaptive consistency loss L acl Calculate the total network loss L, and optimize the network parameters based on the total network loss L to obtain a trained lithium battery health state estimation network suitable for conditions where new lithium battery data is unavailable. Step S5: Input the real-time aging data x of the new battery into the trained lithium battery health state estimation network. i The corresponding health state e(x) is obtained by sequentially passing through a feature extractor and a classifier network C. i ).

2. The method as described in claim 1, characterized in that, In step S1, preprocessing includes denoising the original aging data and extracting artificial features.

3. The method as described in claim 1, characterized in that, The feature extractor comprises a cascaded convolutional network module and several residual network modules. The convolutional network module includes cascaded convolutional layers, batch normalization layers, ReLU activation layers, and max pooling layers. The residual network modules also include cascaded convolutional layers, batch normalization layers, and ReLU activation layers, with short connections between the inputs of the convolutional layers and the outputs of the ReLU activation layers. During training, the output of each residual network module is also fed into a domain discriminator.

4. The method as described in claim 1, characterized in that, The discrimination process of the domain discriminator is as follows: Define P c This represents the probability that the domain discriminator correctly identifies any feature channel of the input feature of the domain discriminator as originating from a certain domain, i.e., the discrimination success rate; Calculate the success rate P of discrimination based on the current input features. c The number of channels n, if the channel ratio If the threshold ratio is exceeded, n discrimination success rates P will be forgotten according to a certain forgetting ratio. c The characteristics of some channels in the middle.

5. The method as described in claim 4, characterized in that, The ratio threshold is set to ,in, This is the preset forgetting rate.

6. The method as described in claim 1, characterized in that, The weighted random selection algorithm randomly selects channels based on the key values ​​obtained from the domain discriminator, specifically as follows: Get the key values ​​of each channel in, These are the weights of each channel output by each domain discriminator. It is a random number between (0, 1); Select the m channels with the largest key values, set their corresponding masks to 0, and set the masks of the other channels to 1 to obtain the channel mask matrix; The aging characteristics after the weighted random selection algorithm is obtained by multiplying the channel mask matrix with the output characteristics of the residual network module.

Images