A SOH evaluation method for lithium-ion batteries of electric vehicles based on network model

Abstract

The present invention relates to the technical field of battery evaluation, and discloses a SOH evaluation method for a lithium-ion battery of an electric vehicle based on a network model, comprising the following steps: step S01: collecting target data; step S02: preprocessing the target data; step S03: selecting a domain; step S04: extracting features; step S05: building a prediction model; step S06: forming a feature sequence with the collected real-time target data and historical target data, taking the feature sequence as the input of a battery health status evaluation network model, estimating through the constructed battery health status evaluation network model, and outputting a battery health status evaluation value; by providing steps S03, step S04 and step S05, a plurality of source domain data similar to the target domain are jointly trained through the network model, so as to solve the domain difference problem between working condition data and meet the actual application needs.

Description

Technical Field

[0001] The present invention relates to the technical field of battery evaluation, and more specifically to a SOH evaluation method for a lithium-ion battery of an electric vehicle based on a network model. Background Art

[0002] With the continuous development of electric vehicles, power batteries face huge challenges in terms of degradation during use and mileage limitations. During the battery life cycle, the degradation process is manifested as a gradual attenuation of capacity and a gradual increase in internal resistance. This series of changes directly affects the performance and remaining service life of the battery. The degradation mechanisms of power batteries mainly include cycle aging and chronological aging. Cycle aging is mainly caused by chemical reactions in the battery during the charging and discharging process. With the increase in the number of charge and discharge cycles, the active materials inside the battery are gradually consumed, resulting in a decrease in battery capacity and an increase in internal resistance. Chronological aging is the gradual performance degradation of the battery during storage, affected by factors such as ambient temperature and storage time. Even if the battery is not working, it will age over time.

[0003] Battery state of health (SOH) is an important indicator for evaluating the degree of degradation of power batteries. Accurately evaluating the battery's state of charge (SOC) and health status and timely replacing low-life batteries will reduce the possibility of failure, which is of great significance for maintaining the overall performance of lithium batteries and ensuring the safety of vehicles and personnel. Traditional battery health status evaluation uses equivalent circuit method and other non-data-driven methods. However, due to the complexity of the internal structure and strong dynamic characteristics of lithium-ion batteries, traditional evaluation methods have limitations in modeling accuracy and online estimation capabilities, and are difficult to meet actual application needs. In addition, data-driven methods often show good estimation effects in experimental data estimation, but when applied to vehicle operating data, due to the large difference between actual operating data and laboratory data estimation, the features extracted from the laboratory may fail when applied to operating data. Operating data usually lacks real SOH labels, which makes it difficult for traditional supervised learning methods that rely on a large amount of labeled data to meet the needs.

[0004] In view of this, the present invention proposes a SOH evaluation method for lithium-ion batteries of electric vehicles based on a network model, which solves the domain difference problem between operating condition data through a deep belief network and an attention repository to meet the needs of practical applications and improve the estimation accuracy and applicability. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a SOH evaluation method for lithium-ion batteries of electric vehicles based on a network model to solve the problems existing in the above-mentioned background technology.

[0006] The present invention provides the following technical solution: a method for evaluating the SOH of a lithium-ion battery of an electric vehicle based on a network model, comprising the following steps:

[0007] Step S01: Collect target data: collect real-time target data and historical target data;

[0008] Step S02: preprocessing the target data: preprocessing the target data collected in step S01 to obtain target data that can be directly used;

[0009] Step S03: Perform domain selection: Use a domain selection algorithm based on multi-source transfer learning to obtain the target domain;

[0010] Step S04: Perform feature extraction: establish a deep belief network model, add a content error mask mechanism to learn target data features, and perform feature extraction;

[0011] Step S05: constructing a prediction model: combining the deep belief network model with the content error mask mechanism added in step S04 with the attention repository to form a battery health status assessment network model;

[0012] Step S06: The collected real-time target data and historical target data are combined into a feature sequence, and the feature sequence is used as the input of the battery health status assessment network model. The battery health status assessment network model is used to perform estimation and output a battery health status assessment value.

[0013] The target data includes battery voltage, current, temperature, and charging and discharging battery data. The real-time target data is the current target data of the lithium-ion battery, and the historical target data is the previous target data of the lithium-ion battery.

[0014] The target domain is the real-time target data domain where battery health status assessment is required.

[0015] Preferably, the formula of the field selection algorithm is expressed as: , where Dis is the distance of the target area, D w is the distance the bulldozer moves, D o is the Euclidean distance;

[0016] The specific method for calculating the moving distance of the bulldozer is:

[0017] Assuming there are probabilities P1 and P2, then by the formula: , the moving distance of the bulldozer is calculated, where W(P1,P2) represents the bulldozer distance, γ is the joint distribution of P1 and P2, is the distribution of all joint combinations of the two distributions P1 and P2, x and y are samples sampled from γ, is the distance between x and y, inf represents the infimum, that is, the lower bound of the expected value, represents the expected value of the distance between x and y when x and y are sampled from γ. Indicates that x and y are sampled from γ, Indicates that γ is Sampling is carried out in

[0018] The specific method of calculating the Euclidean distance is:

[0019] The feature matrix of the source domain is labeled A, the feature matrix of the target domain is labeled B, and the eigenvalue of the i-th row in the A matrix is ​​represented as a i , denote the eigenvalue of the jth row in the B matrix as b j , n is the total number of rows in matrix A, m is the total number of rows in matrix B, then i=1, 2, 3, ..., n, j=1, 2, 3, ..., m; then by formula: , where D o (A, B) is the Euclidean distance between matrix A and matrix B; the source domain is the data domain in the historical target data used for battery health status assessment.

[0020] Preferably, the deep belief network model consists of an input layer, a visible layer, a hidden layer and an output layer, the visible layer and the hidden layer form a restricted Boltzmann machine, the deviation of the visible layer is represented as a, the deviation of the hidden layer is represented as b, and the weight matrix between the visible layer and the hidden layer is represented as w;

[0021] Assumptions , , where v is the visible layer, h is the hidden layer, N is the total number of nodes in the visible layer, M is the total number of nodes in the hidden layer, and w N×M is the weight matrix between the visible layer and the hidden layer, and the model parameter vector , v c is the cth node in the visible layer, h d is the dth node in the hidden layer, c=1, 2, 3, …, N, d=1, 2, 3, …, M;

[0022] Calculate the energy function:

[0023] , where E(v,h) is the energy function, w cd represents the weight between the visible layer node c and the hidden layer node d, a c represents the deviation of the visible layer node c, b d represents the deviation of the visible layer node d, c=1, 2, 3, ..., N, d=1, 2, 3, ..., M;

[0024] Based on the energy function, the joint probability distribution is obtained, and the formula is expressed as: ,in, is a normalization function, P(v,h) is a joint probability distribution, θ is a model parameter vector, and e is a natural constant. The calculation formula of the normalization function is expressed as follows: ; Among them, ∑ v To sum the states of the visible layer v, ∑ h To sum the hidden layer h states;

[0025] The independent probability distribution calculation formula of the visible layer is: , where p(v) is the independent probability distribution of the visible layer; since there is no connection between nodes in the same layer, the conditional probability distributions of the visible layer and the hidden layer can be deduced as follows:

[0026] ;

[0027] ;

[0028] Among them, p(h d =1|v) is the conditional probability distribution of the visible layer, p(v c =1|h) is the conditional probability distribution of the hidden layer, and σ is the sigmoid function.

[0029] Preferably, the specific process of the content error mask is:

[0030] Extract features from the source domain and the target domain, denote the features extracted from the source domain as F1, and denote the features extracted from the target domain as F2. Then the content error mask can be expressed as: , where UM represents the content error mask, 1 represents valid data, 0 represents data to be masked, |F1-F2| represents the absolute value of the difference between F1 and F2, To indicate whether the error value reflects the threshold of possible content information inconsistency caused by domain shift, the final extracted feature is expressed as: , where F´ is the final extracted feature, F is either F1 or F2, and Dropout is the Dropout function.

[0031] Preferably, the attention repository includes a query item, a key item, and a value item; given an input feature, the input feature is the query item, firstly, an attention score between the input feature and the key item is calculated, and then the value item and the attention score are used as a linear combination of weights to obtain a reconstructed feature map;

[0032] The formula is: , where A is the attention score, Q is the query item, V is the key item, R is the dimension, and T is the transposed representation; the query item is used to determine the input part, the key item is used to match the query item and determine the location of the relevant information, and the value item is the actual stored information.

[0033] Preferably, the battery health status assessment network model includes an input layer, a feature layer, a regression layer and an output layer;

[0034] The input layer is used to input data, the feature layer is composed of a deep belief network model and a content error mask, the regression layer is composed of an attention repository and a softmax function, and the attention repository is used to capture long-term and short-term dependencies and integrate the high-dimensional features extracted in step S04 into time series modeling;

[0035] The specific training process of the battery health status assessment network model is as follows:

[0036] The real-time target data at the current moment and the historical target data form time series data, and the time series data is input into the input layer of the battery health status assessment network model. The data features are extracted through the feature layer, and the deep belief network model extracts the high-dimensional features of the data and learns them. The content error mask further divides the high-dimensional information learned by the deep belief network model into domain-related and domain-independent information, and then transmits it to the regression layer. The battery health status assessment value is calculated in the shared feature space according to the features learned by the deep belief network model, and the ReLU activation function is introduced; the regression layer outputs a regression loss value, which is the absolute difference between the estimated value and the true value.

[0037] Preferably, the regression loss is obtained in the following manner:

[0038] The source domain data at time t is input into the battery health status assessment network model, and after passing through the feature layer and regression layer, the predicted battery health status assessment value is output. The target domain data at time t is input into the battery health status assessment network model, and after passing through the feature layer and regression layer, the true battery health status assessment value is output. The absolute difference between the predicted battery health status assessment value and the true battery health status assessment value is used as the regression loss value, and the regression loss values ​​at multiple times are calculated, and their average is used as the regression loss of the final model.

[0039] Technical effects and advantages of the present invention:

[0040] The present invention is provided with step S03, step S04 and step S05, which is conducive to solving the problem of lack of SOH estimation labels under working conditions by using a deep belief network, and introduces an attention repository and a content error mask to solve the problem of poor generalization in the SOH estimation algorithm using the transfer learning idea. The SOH estimation model based on the deep belief network and the attention repository can have a good performance effect on working vehicles, and has obvious accuracy advantages over the comparison algorithm, improves the accuracy of SOH estimation under working conditions, can meet the needs of SOH estimation, and can be widely used in vehicle-mounted battery management systems; through the battery health status assessment network model, multiple source domain data similar to the target domain are jointly trained to solve the problem of domain differences between working condition data and meet actual application needs. The transfer learning method based on the deep belief network and the attention repository performs better on working condition data than the traditional method, and its SOH estimation accuracy is significantly improved; while meeting the daily applicability of the battery management system, it shows a strong promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 The present invention is a flow chart of the SOH evaluation method of lithium-ion batteries for electric vehicles based on a network model.

[0042] Figure 2 This is a flow chart of SOH estimation of the present invention. DETAILED DESCRIPTION

[0043] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the present invention. In addition, the forms of the various structures recorded in the following embodiments are merely illustrative. The network model-based electric vehicle lithium-ion battery SOH evaluation method involved in the present invention is not limited to the various structures recorded in the following embodiments. All other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0044] Example 1

[0045] like Figure 1 As shown, the present invention provides a method for evaluating the SOH of a lithium-ion battery of an electric vehicle based on a network model, comprising the following steps:

[0046] Step S01: collecting target data: collecting real-time target data and historical target data, wherein the target data includes but is not limited to battery data such as battery voltage, current, temperature, and charge and discharge, wherein the real-time target data is the target data of the current lithium-ion battery, and the historical target data is the target data of the previous lithium-ion battery; the purpose is to evaluate the health status of the lithium-ion battery by collecting real-time target data, and to train by collecting historical target data, so as to lay a data foundation for subsequent evaluation of the battery health status;

[0047] Step S02: preprocessing the target data: preprocessing the target data collected in step S01 to obtain target data that can be used directly; the preprocessing operation includes but is not limited to outlier processing, missing value processing, removal of duplicate values ​​and noise, etc.;

[0048] Step S03: Performing domain selection: Based on multi-source transfer learning, a domain selection algorithm is used to obtain a target domain. The target domain is a domain where knowledge and experience are expected to be acquired through transfer learning technology, that is, a real-time target data domain where battery health status assessment is required; the domain selection algorithm is composed of a combination of bulldozer moving distance and Euclidean distance, which can take advantage of the bulldozer moving distance and take into account the coverage of the health status range of the source domain and the target domain. The source domain is a domain with a large amount of labeled data in transfer learning and a domain containing useful information for the target domain, that is, a data domain used for battery health status assessment in historical target data; its purpose is to use a large amount of data in the source domain and a small amount of data in the target domain to learn the model, so as to minimize the generalization error in the target domain and improve the accuracy of subsequent battery health status assessment; domain selection is to select one or more domains that are most similar to the target domain from multiple source domains, so as to transfer knowledge therefrom, to identify and utilize the most relevant domains, thereby improving the performance of transfer learning, reducing the difficulty of migration, reducing the subsequent prediction network model training time and consumption of computing resources, and improving the generalization ability of the model;

[0049] Step S04: feature extraction: establish a deep belief network model, add a content error mask mechanism, learn the target data features, and perform feature extraction; feature extraction is a key step in realizing battery health state prediction. Selecting more relevant and practical features as inputs of subsequent prediction models can significantly improve the accuracy and robustness of the prediction model, thereby improving the accuracy and effect of battery health state assessment;

[0050] Step S05: Constructing a prediction model: combining the deep belief network model with the content error mask mechanism added in step S04 with the attention repository to form a battery health status assessment network model; the deep belief network model can effectively extract deep feature representations from high-dimensional data, and its hierarchical training mechanism helps to capture the global structural features of the data, thereby providing a richer feature basis for downstream tasks. The attention repository can capture long-term and short-term dependencies, and integrate the high-dimensional features extracted by the deep belief network model into time series modeling, which helps to improve the prediction ability of complex time series data. The combination of the two can better balance data information, improve the performance of transfer learning, and better estimate the battery health status of the target domain;

[0051] Step S06: The collected real-time target data and historical target data are formed into a feature sequence, and the feature sequence is used as the input of the battery health status assessment network model. The historical data features can be fully extracted, and the battery health status is estimated and predicted in combination with the real-time target data features, thereby improving the accuracy and estimation effect. The battery health status assessment network model is used to perform estimation and output the battery health status assessment value.

[0052] In this embodiment, it should be specifically explained that the formula of the field selection algorithm is expressed as: , where Dis is the distance of the target area, D w is the distance the bulldozer moves, D o is the Euclidean distance; the bulldozer distance is a distance measurement method based on probability distribution, which is expressed as the minimum cost required to transform one probability distribution into another probability distribution, and is defined as the minimum conversion cost between two probability distributions;

[0053] The specific method for calculating the moving distance of the bulldozer is:

[0054] Assuming there are probabilities P1 and P2, then by the formula: , the moving distance of the bulldozer is calculated, where W(P1,P2) represents the bulldozer distance, γ is the joint distribution of P1 and P2, is the distribution of all joint combinations of the two distributions P1 and P2, x and y are samples sampled from γ, is the distance between x and y, inf represents the infimum, that is, the lower bound of the expected value, represents the expected value of the distance between x and y when x and y are sampled from γ. Indicates that x and y are sampled from γ, Indicates that γ is Sampling is carried out in

[0055] The specific method of calculating the Euclidean distance is:

[0056] The feature matrix of the source domain is labeled A, the feature matrix of the target domain is labeled B, and the eigenvalue of the i-th row in the A matrix is ​​represented as a i , denote the eigenvalue of the jth row in the B matrix as b j , n is the total number of rows in matrix A, m is the total number of rows in matrix B, then i=1, 2, 3, ..., n, j=1, 2, 3, ..., m; then by formula: , where D o (A,B) is the Euclidean distance between matrix A and matrix B.

[0057] In this embodiment, it should be specifically explained that the deep belief network model can effectively capture the deep nonlinear characteristics in the battery operating condition data, adapt to complex data distribution, and reduce noise interference through layer-by-layer training. At the same time, combined with multi-source transfer learning, it can better solve the distribution difference problem of operating condition data, and pre-train multiple source domain data at the same time, extract common features of the target domain, and reduce the feature distribution difference between the source domain and the target domain through the content error mask mechanism, further improve the estimation accuracy of the battery health status evaluation, and effectively meet the actual needs under operating conditions;

[0058] The deep belief network model consists of an input layer, a visible layer, a hidden layer and an output layer. The visible layer and the hidden layer form a restricted Boltzmann machine. The deviation of the visible layer is represented as a, the deviation of the hidden layer is represented as b, and the weight matrix between the visible layer and the hidden layer is represented as w. The restricted Boltzmann machine is an energy-based model, assuming , , where v is the visible layer, h is the hidden layer, N is the total number of nodes in the visible layer, M is the total number of nodes in the hidden layer, and w N×M is the weight matrix between the visible layer and the hidden layer, and the model parameter vector , v c is the cth node in the visible layer, h d is the dth node in the hidden layer, c=1, 2, 3, …, N, d=1, 2, 3, …, M;

[0059] Calculate the energy function:

[0060] , where E(v,h) is the energy function, w cd represents the weight between the visible layer node c and the hidden layer node d, a c represents the deviation of the visible layer node c, b d represents the deviation of the visible layer node d, c=1, 2, 3, ..., N, d=1, 2, 3, ..., M;

[0061] Based on the energy function, the joint probability distribution is obtained, and the formula is expressed as: ,in, is a normalization function, P(v,h) is a joint probability distribution, θ is a model parameter vector, and e is a natural constant. The calculation formula of the normalization function is expressed as follows: ; Among them, ∑ v To sum the states of the visible layer v, ∑ h To sum the hidden layer h states;

[0062] The independent probability distribution calculation formula of the visible layer is: , where p(v) is the independent probability distribution of the visible layer; since there is no connection between nodes in the same layer, the conditional probability distributions of the visible layer and the hidden layer can be deduced as follows:

[0063] ;

[0064] ;

[0065] Among them, p(h d =1|v) is the conditional probability distribution of the visible layer, p(v c =1|h) is the conditional probability distribution of the hidden layer, and σ is the sigmoid function.

[0066] In this embodiment, it should be specifically explained that the content error mask is a processing mechanism for noise or erroneous information in the input data, which can dynamically shield the noise features to avoid interference with the learning of the model. The introduction of the content error mask can further improve the prediction accuracy of the battery health status assessment, and perform better in the case of low data quality and outliers. At the same time, the weights can be adaptively adjusted to improve the robustness and overall performance of the features; the content error mask adaptively learns the weights of the input features, and when there are noise or outliers in the data, it effectively reduces the impact of these interfering information on the model, thereby improving the purity and representation quality of the features, highlighting the importance of key features, and making the model more focused on the most relevant information for SOH prediction, improving the overall prediction accuracy and data utilization efficiency, and integrating the content error mask into the deep belief network model, which significantly enhances the robustness and generalization ability of the model;

[0067] The specific process is:

[0068] Extract features from the source domain and the target domain, denote the features extracted from the source domain as F1, and denote the features extracted from the target domain as F2. Then the content error mask can be expressed as: , where UM represents the content error mask, 1 represents valid data, 0 represents data to be masked, |F1-F2| represents the absolute value of the difference between F1 and F2, To indicate whether the error value reflects the threshold of possible content information inconsistency caused by domain shift, the final extracted feature is expressed as: , where F´ is the final extracted feature, F is either F1 or F2, and Dropout is the Dropout function.

[0069] In this embodiment, it should be specifically explained that the attention storage library significantly improves the context processing capability of the deep learning model through external storage, especially in long time series tasks, with lower computational complexity, and is suitable for tasks with high real-time requirements. The traditional attention storage library is designed as a memory library that updates memory vectors separately, and each memory vector corresponds to a specific category, so it cannot be applied to regression tasks because a limited number of memory vectors cannot cover continuous density values. In this embodiment, each feature vector is reconstructed into an attention pair memory vector, so that any density value can be represented as a linear combination of representative values ​​in the memory library.

[0070] The attention repository includes query items, key items, and value items. Given an input feature, which is the query item, firstly, the attention score between the input feature and the key item is calculated, and then the value item and the attention score are used as a linear combination of weights to obtain a reconstructed feature map.

[0071] The formula is: , where A is the attention score, Q is the query term, V is the key term, R is the dimension, and T is the transposed representation;

[0072] The query item is used to determine the input part, the key item is used to match the query item and determine the location of the relevant information, and the value item is the actual stored information and is the source of the weighted result.

[0073] In this embodiment, it should be specifically explained that the battery health status assessment network model includes an input layer, a feature layer, a regression layer and an output layer;

[0074] The input layer is used to input data, the feature layer is composed of a deep belief network model and a content error mask, and the regression layer is composed of an attention repository and a softmax function. The attention repository is used to capture long-term and short-term dependencies, and integrate the high-dimensional features extracted in step S04 into time series modeling, which helps to improve the prediction ability of complex time series data;

[0075] The specific training process of the battery health status assessment network model is as follows:

[0076] The real-time target data at the current moment and the historical target data form time series data, and the time series data is input into the input layer of the battery health status assessment network model. The data features are extracted through the feature layer, and the deep belief network model extracts the high-dimensional features of the data and learns them. The content error mask further divides the high-dimensional information learned by the deep belief network model into domain-related and domain-independent information, and then transmits it to the regression layer. The battery health status assessment value is calculated according to the features learned by the deep belief network model in the shared feature space, and the ReLU activation function is introduced so that the network can learn and represent more complex functions; the regression layer outputs a regression loss value, which is the absolute difference between the estimated value and the true value, and the regression loss value includes the regression loss of each source field;

[0077] The method for obtaining the regression loss is specifically as follows:

[0078] The source domain data at time t is input into the battery health status assessment network model, and after passing through the feature layer and regression layer, the predicted battery health status assessment value is output. The target domain data at time t is input into the battery health status assessment network model, and after passing through the feature layer and regression layer, the true battery health status assessment value is output. The absolute difference between the predicted battery health status assessment value and the true battery health status assessment value is used as the regression loss value, and the regression loss values ​​at multiple times are calculated, and their average is used as the regression loss of the final model.

[0079] Example 2

[0080] like Figure 2 As shown, the present invention provides a SOH estimation process for using a network model-based SOH evaluation method for a lithium-ion battery of an electric vehicle proposed in Example 1, comprising the following steps:

[0081] Step S11: inputting the collected battery target data;

[0082] Step S12: cleaning the battery target data and extracting features;

[0083] Step S13: Calculate the domain distance and perform domain selection;

[0084] Step S14: extracting high-dimensional features using a deep belief network model with content error mask added;

[0085] Step S15: Use the attention repository to process features and obtain regression loss;

[0086] Step S16: Output the battery health status evaluation value.

[0087] In the present embodiment, it should be specifically explained that the difference between the present embodiment and the prior art lies mainly in that the present embodiment has step S03, step S04 and step S05, uses a deep belief network to solve the problem of lack of SOH estimation labels under working conditions, and introduces an attention repository and a content error mask to solve the problem of poor generalization in the SOH estimation algorithm using the transfer learning idea. The SOH estimation model based on a deep belief network and an attention repository can have better performance on working vehicles, has obvious accuracy advantages over the comparison algorithm, improves the accuracy of SOH estimation under working conditions, can meet the needs of SOH estimation, and can be widely used in vehicle battery management systems.

[0088] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

[0089] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for evaluating SOH of lithium-ion batteries of electric vehicles based on a network model, characterized in that: The following steps are involved: Step S01: Collect target data: collect real-time target data and historical target data; Step S02: preprocessing the target data: preprocessing the target data collected in step S01 to obtain target data that can be directly used; Step S03: Perform domain selection: Use a domain selection algorithm based on multi-source transfer learning to obtain the target domain; Step S04: Perform feature extraction: establish a deep belief network model, add a content error mask mechanism to learn target data features, and perform feature extraction; Step S05: constructing a prediction model: combining the deep belief network model with the content error mask mechanism added in step S04 with the attention repository to form a battery health status assessment network model; Step S06: The collected real-time target data and historical target data are combined into a feature sequence, and the feature sequence is used as the input of the battery health status assessment network model. The battery health status assessment network model is used to perform estimation and output a battery health status assessment value. The target data includes battery voltage, current, temperature, and charging and discharging battery data. The real-time target data is the current target data of the lithium-ion battery, and the historical target data is the previous target data of the lithium-ion battery. The target domain is the real-time target data domain where battery health status assessment is required.

2. The SOH evaluation method for lithium-ion batteries of electric vehicles based on a network model according to claim 1 is characterized in that: The formula of the domain selection algorithm is expressed as: , where Dis is the distance of the target area, D w is the distance the bulldozer moves, D o is the Euclidean distance; The specific method for calculating the moving distance of the bulldozer is: Assuming there are probabilities P1 and P2, then by the formula: , the moving distance of the bulldozer is calculated, where W(P1,P2) represents the bulldozer distance, γ is the joint distribution of P1 and P2, is the distribution of all joint combinations of the two distributions P1 and P2, x and y are samples sampled from γ, is the distance between x and y, inf represents the infimum, that is, the lower bound of the expected value, represents the expected value of the distance between x and y when x and y are sampled from γ. Indicates that x and y are sampled from γ, Indicates that γ is Sampling is carried out in The specific method of calculating the Euclidean distance is: The feature matrix of the source domain is labeled A, the feature matrix of the target domain is labeled B, and the eigenvalue of the i-th row in the A matrix is ​​represented as a i , denote the eigenvalue of the jth row in the B matrix as b j , n is the total number of rows in matrix A, m is the total number of rows in matrix B, then i=1, 2, 3, ..., n, j=1, 2, 3, ..., m; then by formula: , where D o (A, B) is the Euclidean distance between matrix A and matrix B; the source domain is the data domain in the historical target data used for battery health status assessment.

3. The SOH evaluation method for lithium-ion batteries of electric vehicles based on a network model according to claim 2 is characterized in that: The deep belief network model consists of an input layer, a visible layer, a hidden layer and an output layer. The visible layer and the hidden layer form a restricted Boltzmann machine. The deviation of the visible layer is represented as a, the deviation of the hidden layer is represented as b, and the weight matrix between the visible layer and the hidden layer is represented as w; Assumptions , , where v is the visible layer, h is the hidden layer, N is the total number of nodes in the visible layer, M is the total number of nodes in the hidden layer, and w N×M is the weight matrix between the visible layer and the hidden layer, and the model parameter vector , v c is the cth node in the visible layer, h d is the dth node in the hidden layer, c=1, 2, 3, …, N, d=1, 2, 3, …, M; Calculate the energy function: , where E(v,h) is the energy function, w cd represents the weight between the visible layer node c and the hidden layer node d, a c represents the deviation of the visible layer node c, b d represents the deviation of the visible layer node d, c=1, 2, 3, ..., N, d=1, 2, 3, ..., M; Based on the energy function, the joint probability distribution is obtained, and the formula is expressed as: ,in, is a normalization function, P(v,h) is a joint probability distribution, θ is a model parameter vector, and e is a natural constant. The calculation formula of the normalization function is expressed as follows: ; Among them, ∑ v To sum the states of the visible layer v, ∑ h To sum the hidden layer h states; The independent probability distribution calculation formula of the visible layer is: , where p(v) is the independent probability distribution of the visible layer; since there is no connection between nodes in the same layer, the conditional probability distributions of the visible layer and the hidden layer can be deduced as follows: ； ； Among them, p(h d =1|v) is the conditional probability distribution of the visible layer, p(v c =1|h) is the conditional probability distribution of the hidden layer, and σ is the sigmoid function.

4. The SOH evaluation method for lithium-ion batteries of electric vehicles based on a network model according to claim 3 is characterized in that: The specific process of content error mask is as follows: Extract features from the source domain and the target domain, denote the features extracted from the source domain as F1, and denote the features extracted from the target domain as F2. Then the content error mask can be expressed as: , where UM represents the content error mask, 1 represents valid data, 0 represents data to be masked, |F1-F2| represents the absolute value of the difference between F1 and F2, To indicate whether the error value reflects the threshold of possible content information inconsistency caused by domain shift, the final extracted feature is expressed as: , where F´ is the final extracted feature, F is either F1 or F2, and Dropout is the Dropout function.

5. The SOH evaluation method for lithium-ion batteries of electric vehicles based on a network model according to claim 4 is characterized in that: The attention repository includes query items, key items, and value items. Given an input feature, which is the query item, firstly, the attention score between the input feature and the key item is calculated, and then the value item and the attention score are used as a linear combination of weights to obtain a reconstructed feature map. The formula is: , where A is the attention score, Q is the query item, V is the key item, R is the dimension, and T is the transposed representation; the query item is used to determine the input part, the key item is used to match the query item and determine the location of the relevant information, and the value item is the actual stored information.

6. The method for evaluating SOH of a lithium-ion battery of an electric vehicle based on a network model according to claim 5 is characterized in that: The battery health status assessment network model includes an input layer, a feature layer, a regression layer and an output layer; The input layer is used to input data, the feature layer is composed of a deep belief network model and a content error mask, the regression layer is composed of an attention repository and a softmax function, and the attention repository is used to capture long-term and short-term dependencies and integrate the high-dimensional features extracted in step S04 into time series modeling; The specific training process of the battery health status assessment network model is as follows: The real-time target data at the current moment and the historical target data form time series data, and the time series data is input into the input layer of the battery health status assessment network model. The data features are extracted through the feature layer, and the deep belief network model extracts the high-dimensional features of the data and learns them. The content error mask further divides the high-dimensional information learned by the deep belief network model into domain-related and domain-independent information, and then transmits it to the regression layer. The battery health status assessment value is calculated in the shared feature space according to the features learned by the deep belief network model, and the ReLU activation function is introduced; the regression layer outputs a regression loss value, which is the absolute difference between the estimated value and the true value.

7. The method for evaluating SOH of a lithium-ion battery of an electric vehicle based on a network model according to claim 6 is characterized in that: The method for obtaining the regression loss is specifically as follows: The source domain data at time t is input into the battery health status assessment network model, and after passing through the feature layer and regression layer, the predicted battery health status assessment value is output. The target domain data at time t is input into the battery health status assessment network model, and after passing through the feature layer and regression layer, the true battery health status assessment value is output. The absolute difference between the predicted battery health status assessment value and the true battery health status assessment value is used as the regression loss value, and the regression loss values ​​at multiple times are calculated, and their average is used as the regression loss of the final model.

Images