Lithium battery state of health estimation method based on joint weighted domain adversarial network

Abstract

The present application relates to a lithium battery state of health estimation method based on a joint weighted domain adversarial network, and belongs to the technical field of new energy. The method comprises the following steps: S1: similarity evaluation between different time series; S2: feature extraction of a stacked bidirectional long short-term memory neural network, then lithium battery charge estimation based on an improved domain adversarial network, and finally specific information retention based on an InfoNCE module. The present application aims at the problems of great difficulty in obtaining lithium ion battery label data, complex experimental procedures, various environmental tests and the like, realizes sharing of lithium ion battery data information, in the face of various battery test procedures and various battery aging test conditions, considers how to realize automatic extraction of excellent aging features while improving the accuracy of lithium ion battery state of health prediction, to a certain extent, makes up for the limitation of insufficient data information, and is very meaningful for realizing lithium ion battery state of health prediction under complex environments.

Description

Technical Field

[0001] This invention belongs to the field of new energy technology and relates to a method for estimating the health status of lithium batteries based on joint weighted domain adversarial networks. Background Technology

[0002] The aging process of lithium-ion batteries is a dynamic, coupled process with highly complex internal chemical mechanisms, which increases the difficulty of understanding their internal structure. During long-term operation, lithium-ion battery performance degrades due to reduced lithium reserves, decomposition of active materials, structural changes, and increased internal resistance caused by the growth of the solid electrolyte interphase (SEI). However, the state of health of lithium-ion batteries cannot be directly measured, which is a significant challenge for battery management systems (BMS) in predicting the state of health of lithium-ion batteries and a major problem that needs to be solved in the current lithium-ion battery field. Existing machine learning methods, which heavily rely on labeled data, are becoming increasingly difficult to use for predicting the state of health of lithium-ion batteries under different operating conditions and environmental test conditions.

[0003] In order to solve the problem of accurately estimating SOH in practice, researchers at home and abroad have proposed many effective methods. Gaussian process regression (GPR)

[23] , kernel ridge regression, support vector machine (SVM), support vector regression

[26] and other methods have been successfully applied to battery capacity estimation. A framework based on GPR is used to simultaneously predict capacity and quantify the uncertainty of the predicted value, and to estimate battery SOH by using health features extracted from partial incremental capacity curves. Taking four batteries in the NASA battery degradation dataset as examples, the accuracy, robustness and effectiveness of the method are verified. By using the online SOH estimation method based on SVM, the characteristics of the battery charging curve under different SOH are used to establish a support vector machine model offline. Then, by comparing the characteristics of the measured charging voltage segment and the storage model, the model is used to estimate SOH online. Combine kernel ridge regression and transfer learning to estimate battery health status to improve prediction accuracy. In addition, the CNN-LSTM method has also been widely used and has achieved good prediction results. Existing transfer learning-based prediction models, such as LSTM and RNN models, combine SOH prediction with ensemble learning frameworks to estimate capacity using voltage, current, and charging capacity as network inputs. Clearly, these methods show great potential for battery capacity estimation, but their performance is heavily dependent on the size of the training dataset. Only models trained with sufficient data can achieve satisfactory accuracy. However, collecting a large battery degradation dataset requires extensive iterative testing, which is very time-consuming and expensive. Therefore, incorporating transfer learning techniques into these methods to improve estimation performance on small datasets is highly meaningful; weighted domain adversarial networks have shown good results. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method for estimating the health status of lithium batteries based on joint weighted domain adversarial networks.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A lithium battery health state estimation method based on joint weighted domain adversarial networks includes the following steps:

[0007] S1: Similarity assessment between different time series;

[0008] S2: Perform feature extraction using a stacked bidirectional long short-term memory neural network, then estimate the state of health of the lithium battery based on a joint weighted domain adversarial network, and finally retain specific information based on the information-noise contrast estimation Info-NCE module.

[0009] Optionally, S1 specifically includes:

[0010] First, initialize a temporary average sequence T of a time series set. Calculate the Dynamic Time Warping (DTW) path between the time series of each sample and the temporary average sequence to be refined. Iteratively update each coordinate of the average sequence to its associated barycentric coordinate M = DTW(T, A / B), where T is the time series after each update, A is the time series of the first sample in the source domain data, and B is the time series of the next sample. Calculate the square of the difference between all points in A and B to finally obtain the barycentric average sequence T′ of the entire set.

[0011] JS divergence measures the similarity between two probability distributions, addressing the asymmetry issue of KL divergence. After alignment, JS divergence is used to calculate the distribution similarity between data from two different domains. The JS divergence calculation process is as follows:

[0012]

[0013] n is the number of samples, P1 is the source domain data, P2 is the target domain data, q1 is the probability of obtaining sample x in Q1, and q2 is the probability of obtaining sample x in Q2.

[0014] Optionally, S2 specifically includes:

[0015] S21: Using lithium battery current, voltage, and sampling time series as model inputs, a three-layer bidirectional long short-term memory neural network is adopted. The fully connected layer and the degenerate layer constitute a regression neural network model. The long short-term memory neural network feature extractor is used to represent the multivariate time series as a single vector implicit representation, and the LSTM network is represented as multiple sequential feedforward layers.

[0016] By combining a forward loop LSTM and a backward hidden LSTM, the generated forward hidden state and backward hidden state are passed together to the output:

[0017]

[0018] Where t is the current time, x t It is the input at the current moment, y t It is the output at the current moment. These are the weights of the forward propagation neural network. These are the weights of the backpropagation neural network; H cell Receive the input X at the current moment t The hidden unit of the previous moment Reverse the hidden unit of the previous moment And the memory unit c from the previous moment t-1 The memory unit c at the current moment t The output is h after the current time step is updated.t and c t ;

[0019] S22: The optimization of a weighted domain adversarial network consists of three parts: the parameters of the feature extractor f, and the first domain discriminator D. z And a second discriminator D0; using features from the source domain dataset and features from the target domain dataset, train the domain category recognition network to obtain the importance weights of the source domain samples in the source domain dataset, and obtain the normalized importance weights of the source domain samples based on the importance weights; use an adversarial game strategy to iteratively train the domain classifier and the importance-weighted domain adversarial adaptive network until all network parameters converge; the weighted domain adaptive algorithm is conceived as follows: preprocess the battery data under two different working conditions to obtain the source domain training dataset and the target domain training dataset, obtain the features of the input signal through the feature extractor, and the loss function of the domain category recognition network is expressed as:

[0020]

[0021] The function is to calculate the mean of the samples in the source domain; To obtain the mean of the samples in the target domain; log(·) is the domain category label prediction probability vector of the normalized network samples; D is the discriminator network; G is the generator network; N is the total number of samples in the original data; i is the number of each sample; x i Feature data for each sample; d i Let L represent the binary label of the i-th sample, used to indicate whether the sample belongs to the source domain or the target domain; L is the training loss of the adversarial network.

[0022] By constructing a new dataset, the source samples are labeled as 1 and the target samples are labeled as 0:

[0023]

[0024] The importance weights of source domain samples are designed as follows:

[0025]

[0026] The sig mod (·) function maps the input to the range 0 and 1; θ dz Representative domain classifier D z Network parameters; F s This is the output of the source feature extractor;

[0027] The importance weights of the normalized source domain samples are:

[0028]

[0029] x s It is source domain sample data;

[0030] The domain category label prediction results are obtained by training with weighted source and target samples:

[0031]

[0032] F t It is the output of the target feature extractor;

[0033] For a given F s And D, each sample F t The optimizer D0 of a weighted adversarial network is represented as:

[0034]

[0035] p0 is the domain category label prediction probability vector of the normalized network sample; θ d0 The network parameters represent the domain classifier D0; ultimately, the optimization objective of the entire network can be obtained.

[0036] S23: When using a target feature extractor on the input X t The corresponding feature f is then obtained. T To preserve target-specific characteristics during the domain alignment task, maximize X t and f T The mutual information between them is defined by a density ratio function Φ at each time step. k :

[0037]

[0038] By maximizing the potential target feature f T and the original input features X t The interactions between them, preserving their common latent variables; for calculating Φ k A fully connected network θ is used to map the latent features f. T and input features X t To the same dimension; by transforming the eigenvector q k =θ k (f t The density ratio is estimated by performing a dot product with the original input vector k; to maximize the density ratio function, the contrastive estimation loss of the target feature extractor and the fully connected network θ is jointly optimized; the Info NCE loss maximizes mutual information by comparing positive and negative samples, and the optimal probability p(d=k|X) for the noise contrastive estimation loss NCEloss is obtained. t ,f T ) is represented as:

[0039]

[0040] The mutual information is ultimately represented as:

[0041]

[0042] x k It is the original input at the k-th time step, f t It is the input feature of a fully connected neural network, φ k It is a dot product operation, where K is a total time step. By continuously reducing NCE loss during the model training process, we find the optimal solution that maximizes mutual information and obtain the best network model.

[0043] The final optimized solution is expressed as follows:

[0044]

[0045]

[0046] The features extracted from the above feature extractor are fed into a multi-layer fully connected network, and an end-to-end method is used for training using the mean squared error loss between the predicted SOH and the actual SOH:

[0047]

[0048] The beneficial effects of this invention are as follows:

[0049] This invention focuses on providing a new direction for predicting the health status of lithium-ion batteries under different environmental testing and operating conditions when aging label data is lacking. Addressing the challenges of obtaining lithium-ion battery label data, complex experimental procedures, and diverse environmental testing environments, this invention utilizes transfer learning to explore how to share lithium-ion battery data information using partial label sample data. Furthermore, considering various battery testing procedures and aging test conditions, this invention explores how to automatically extract desirable aging characteristics while improving the accuracy of lithium-ion battery health status prediction. This, to a certain extent, compensates for the limitations of insufficient data and is highly significant for predicting the health status of lithium-ion batteries under complex environments.

[0050] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0051] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0052] Figure 1 This is a schematic diagram of the present invention. Detailed Implementation

[0053] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0054] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0055] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0056] Figure 1 This is a schematic diagram of the invention. Addressing the difficulty in assessing the validity of source domain data in cross-data domain lithium battery health estimation, this invention analyzes the differences in data distribution among different lithium batteries, studies the transmission process of different sequences, and combines the inherent information reflected in the data waveforms to establish a clear standard for evaluating the similarity between time series. An adversarial domain adaptation network is designed to achieve feature alignment of lithium battery data across different domains. Furthermore, addressing the issue that general adversarial adaptive methods can only perform overall inter-domain migration and may lose important information in the target features, the network is further improved to achieve effective sample selection and partial migration. A mutual information module is combined to establish a connection between the input and output of the target feature extractor, preserving the feature information of the target domain during the domain alignment task and improving the performance of the target feature extractor model.

[0057] (1) Similarity assessment between different time series

[0058] Extracting effective source domain data based on dynamic time warping and JS divergence.

[0059] First, initialize a temporary average sequence T of a time series set. Calculate the DTW path between the time series of each sample and the temporary average sequence to be refined. Iteratively update each coordinate of the average sequence to the associated centroid coordinate M = DTW(T, A / B). Finally, obtain the centroid average sequence T′ of the entire set.

[0060] JS divergence measures the similarity between two probability distributions, resolving the asymmetry issue of KL divergence. Therefore, after alignment, JS divergence is used to further calculate the distribution similarity between the two series. The JS divergence calculation process is as follows:

[0061]

[0062] (2) Lithium-ion battery health status estimation based on joint weighted domain adversarial network

[0063] 1) Feature extraction from stacked bidirectional long short-term memory neural networks

[0064] This study uses lithium battery current, voltage, and sampling time series as model inputs. A three-layer bidirectional long short-term memory neural network is used, with fully connected layers and dropout layers forming a regression neural network model. An LSTM feature extractor is used to represent the multivariate time series as a single-vector implicit representation. Specifically, the LSTM network can be represented as multiple sequential feedforward layers.

[0065] By combining a forward loop LSTM and a backward hidden LSTM, the generated forward hidden state and backward hidden state are passed together to the output:

[0066]

[0067] Where H cell Receive the input X at the current moment t The hidden unit h from the previous moment t-1 , and the memory unit c from the previous moment t The output is h after the current time step is updated. t and c k .

[0068] 2) Lithium-ion battery health status estimation based on joint weighted domain adversarial network

[0069] Weighted domain adversarial networks primarily optimize three parts: the parameters of the feature extractor f, the first domain discriminator D, and the first domain discriminator D. zAnd a second discriminator D0. Using features from the source domain dataset and the target domain dataset, a domain category recognition network is trained to obtain the importance weights of the source domain samples in the source domain dataset. Based on these importance weights, normalized importance weights of the source domain samples are obtained. An adversarial game strategy is used to iteratively train the domain classifier and the importance-weighted domain adversarial adaptive network until all network parameters converge. The weighted domain adaptive algorithm is conceived as follows: Battery data from two different operating conditions are preprocessed to obtain source domain training datasets and target domain training datasets. Features of the input signal are obtained through a feature extractor. The loss function of the domain category recognition network can be expressed as:

[0070]

[0071] The function is to calculate the mean of the samples in the source domain; To obtain the mean of the target domain samples; log(·) is the domain category label prediction probability vector of the normalized network samples.

[0072] By constructing a new dataset, the source samples are labeled as 1 and the target samples are labeled as 0:

[0073]

[0074] The importance weights of source domain samples are designed as follows:

[0075]

[0076] The importance weights of the normalized source domain samples are:

[0077]

[0078] The sigmod(·) function maps the input to the range 0 and 1; θ dz Representative domain classifier D z Network parameters.

[0079] The domain category label prediction results are obtained by training with weighted source and target samples:

[0080]

[0081] For a given F s And D, each sample F t The optimizer D0 of a weighted adversarial network is represented as:

[0082]

[0083] p0 is the domain category label prediction probability vector of the normalized network sample; θ d0The network parameters represent the domain classifier D0; ultimately, the optimization objective of the entire network can be obtained.

[0084] 3) Specific information retention based on the InfoNCE (Information-NCE) module

[0085] When the target feature extractor is applied to the input X t The corresponding feature f is then obtained. T In order to preserve as many target-specific characteristics as possible during the domain alignment task and maximize X t and f T The mutual information between them is defined by a density ratio function Φ at each time step. k :

[0086]

[0087] By maximizing the potential target feature f T and the original input features X t The interactions between them can preserve their common latent variables. To calculate Φ k A fully connected network θ is used to map the latent features f. T and input features X t To achieve the same dimension. Therefore, by applying the transformed eigenvector q... k =θ k (f t The density ratio is estimated by performing a dot product between the target feature extractor and the original input vector k. To maximize the density ratio function, the contrastive estimation loss of the target feature extractor and the fully connected network θ is jointly optimized. The NCE loss maximizes mutual information by comparing positive and negative samples; the optimal probability of NCE loss is p(d = k|X). t ,f T ) is represented as:

[0088]

[0089] The mutual information is ultimately represented as:

[0090] I(X k ,f T ) = log(K) - L InfoNCE (14)

[0091] By continuously reducing NCE loss during model training, the optimal solution that maximizes mutual information is found, and the best network model is obtained.

[0092] The final optimized solution of the entire method can be expressed as:

[0093]

[0094] The features extracted from the feature extractor are fed into a multi-layer fully connected network, and an end-to-end approach is used to train the network using the mean squared error loss between the predicted SOH and the actual SOH.

[0095]

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1.A lithium battery state of health estimation method based on a joint weighted domain adversarial network, characterized in that: The method comprises the following steps: S1: similarity evaluation between different time sequences; S2: feature extraction of stacked bidirectional long short-term memory neural network, lithium battery state of health estimation based on joint weighted domain adversarial network, and finally specific information reservation based on information noise contrast estimation Info-NCE module; Specifically: S21: using lithium battery current, voltage and sampling time sequence as model input, adopting three-layer bidirectional long short-term memory neural network, full connection layer and degradation layer to constitute regression neural network model, using long short-term memory neural network feature extractor to represent multivariate time sequence as single vector implicit representation, and using LSTM network to represent multiple sequential feedforward layers; By combining a forward recurrent LSTM and a backward hidden LSTM, the generated forward hidden state and backward hidden state are transmitted to the output together: (3) (4) (5) wherein t is the current time step, is the input at the current time step, is the output at the current time step, is the forward propagated neural network weight, is the backward propagated neural network weight; is the input at the current time step , the forward hidden unit at the previous time step , the backward hidden unit at the previous time step , and the memory unit at the previous time step , the memory unit at the current time step , the output is the updated and at the current time step; S22: Weighted Domain Adversarial Network Optimization: Three Parts - Feature Extractor f The parameters are: the first domain discriminator and the second discriminator; using features from the source domain dataset and the target domain dataset, a domain category recognition network is trained to obtain the importance weights of the source domain samples in the source domain dataset, and the normalized importance weights of the source domain samples are obtained based on the importance weights; an adversarial game strategy is used to iteratively train the domain classifier and the importance-weighted domain adversarial adaptive network until all network parameters converge; the weighted domain adaptive algorithm is conceived as follows: preprocessing battery data from two different operating conditions to obtain the source domain training dataset and the target domain training dataset, obtaining the features of the input signal through a feature extractor, and the loss function of the domain category recognition network is expressed as: (6) function for obtaining the mean of the source domain samples; function for obtaining the mean of the target domain samples; function for normalizing the domain class label prediction probability vector of the network samples; D is a discriminator network; G is a generator network; N is the total number of samples of the original data; i is the number of each sample; is the feature data of each sample; represents the binary label of the i th sample, which is used to indicate whether the sample belongs to the source domain or the target domain; L is the adversarial network training loss; By constructing a new data set, the source sample is marked as 1 and the target sample is marked as 0: (7) The importance weight of the source domain sample is designed as: (8) The function represents a mapping of the input to between 0, 1; Network parameters representing the domain classifier; For the output of the source feature extractor; The normalized importance weight of the source domain sample is: (9) is source domain sample data; The weighted source domain sample and the target sample are used for training, and the results of domain category label prediction are obtained: (10) is the output of the target feature extractor; For a given and D each sample The optimizer of the weighted adversarial network is represented as: (11) to normalize the domain class label prediction probability vector of the network sample; network parameters representing the domain classifier; the optimization objective of the entire network can be finally obtained; S23: in the case of a target feature extractor acting on the input at the current time step yields the corresponding features , i.e. the output features of the target feature extractor, preserve the target-specific properties during the domain adaptation task, maximizing the mutual information between and a density ratio function is defined at each time step (12) By maximizing the output features of the target feature extractor and the input at the current moment The interactions between them, preserving their common latent variables; for computation Using a fully connected network Output features of the mapping target feature extractor and the input at the current moment To the same dimension; through the transformed eigenvectors With the original input vector The density ratio is obtained by performing dot product estimation; to maximize the density ratio function, the target feature extractor and the fully connected network are jointly optimized. Contrastive estimation loss; Info NCE loss maximizes mutual information by comparing positive and negative samples, and the optimal probability of noise-contrastive estimation loss NCE loss. It is represented as: (13) denotes the probability distribution of the given condition under the given Finally, the mutual information is expressed as: (14) is the original input of the k th time step, is the input feature of the fully connected neural network, is a dot product operation, K is a total time step, and NCEloss is continuously reduced through the process of model training to find the optimal solution that maximizes mutual information and obtain the best network model. The final optimization scheme is expressed as: (15) The features extracted from the above feature extractor are put into a multi-layer fully connected network, and an end-to-end method is used to train using the mean square error loss between the predicted SOH and the actual SOH: (16)。 2.The lithium battery state of health estimation method based on the joint weighted domain adversarial network according to claim 1, wherein: The S1 is specifically: First, initialize a temporary average sequence T of a time series set, calculate the dynamic time warping DTW path between each sample time series and the temporary average sequence to be refined, and update each coordinate of the average sequence to the associated barycenter coordinate M = DTW(T, A / B), wherein T is the time series after each update, A is the first sample time series of the source domain data, and B is the next sample time series, calculate the square of the difference of all points in A and B, and finally obtain the barycenter average sequence T of the entire set ; JS divergence measures the similarity of two probability distributions, and solves the problem of non-symmetry of KL divergence; After alignment, the distribution similarity between the data of two different domains is calculated using JS divergence; The JS divergence calculation process is: (1) (2) n For the sample size, It is source domain data. It is target domain data. Is Samples obtained from x The probability, Is Samples obtained from x The probability of.

Images