A lithium battery SOH intelligent estimation method and system

Abstract

A lithium battery SOH intelligent estimation method and system, including: sensors collect lithium battery operation data of different types and working conditions; construct a VAE-GAN combination model, perform data enhancement, and generate more lithium battery operation data; according to the lithium battery operation data of different types and working conditions, all models in the model library are modeled separately, source models are selected and weighted superimposed into a stacking model; the operation data of the lithium battery to be evaluated is input, and the ART network adaptively weights each source model in the stacking model to obtain a prediction model and output a preliminary prediction result; establish an error correction model based on random forest, complete the error correction of the preliminary prediction result, and output the final prediction value of the lithium battery SOH. The present invention constructs a new VAE-GAN combination model, which reduces the difficulty of collecting multi-source data; constructs a stacking model based on multiple source models, which increases the generalization ability of the model; and performs error correction on the prediction result, which improves the accuracy of the model prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium battery health status assessment, and in particular to a lithium battery SOH intelligent estimation method and system. Background Art

[0002] As a widely used energy storage device, lithium batteries have always attracted wide attention in various fields. Lithium batteries have the advantages of high energy density, long life and no memory effect. However, with the increase of charging cycles and complex usage of lithium batteries, the lithium ion activity of lithium batteries will also decrease, and the power will also decrease, which will also bring safety problems, so the health status (SOH) evaluation of lithium batteries is very important.

[0003] Most of the current health assessment methods for lithium batteries are to conduct SOH assessments on specific types of lithium batteries or lithium batteries in specific usage environments. They cannot be flexibly applied to the operating data of lithium batteries of different types and working conditions, and the proposed health assessment methods have weak adaptability. Most model-based health assessment methods for lithium batteries lack error correction parts, resulting in an uncertainty as to whether the final prediction results have achieved the expected effect and whether the final prediction value is accurate. Summary of the invention

[0004] Purpose of the invention: To provide a lithium battery SOH intelligent estimation method and system to collect lithium battery operating data of different types and working conditions, and propose a VAE-GAN data enhancement combination model with a new structure, so as to generate more lithium battery data of different types and working conditions, so that the final model is more adaptable and can be applied in multiple scenarios; through the constructed model library, multi-model modeling is carried out, and the source model is selected according to the requirements of accuracy, etc., to enhance the prediction ability of the model; and the prediction result of the model is corrected for errors, so as to make the final prediction value more accurate.

[0005] Technical solution: The present invention discloses a lithium battery SOH intelligent estimation method, comprising:

[0006] Step (1) The sensor collects operating data of lithium batteries of different types and working conditions, including voltage, current, temperature, number of charging cycles, temperature and time data;

[0007] Step (2) constructing a variational autoencoder-generative adversarial network VAE-GAN combined model, inputting the collected lithium battery operation data in step (1) for data enhancement, and generating more lithium battery data of different types and working conditions;

[0008] Step (3) constructing a model library, modeling all models in the model library according to the lithium battery data of different types and working conditions in step (2), weighing the accuracy and calculation time according to the amount of lithium battery modeling data, taking the model that meets the accuracy requirements and time threshold as the source model, and weighting all the source models to integrate a stacked model;

[0009] Step (4) inputting the operating data of the lithium battery to be evaluated, using the adaptive resonance theory ART network to adaptively assign weights to each source model in the stacking model, obtaining the final SOH prediction model of the battery to be evaluated, and outputting a preliminary prediction result;

[0010] Step (5) extracts the maximum information coefficient MIC feature of the lithium battery operation data as the input variable, takes the error sequence of the difference between the preliminary prediction result and the actual SOH value of the battery as the target variable, establishes an error correction model based on random forest RF, realizes the prediction of future errors and the correction of prediction results, and outputs the final prediction value of the lithium battery SOH.

[0011] Furthermore, in step (2), a variational autoencoder-generative adversarial network VAE-GAN combined model is constructed, and the collected lithium battery operation data is input for data enhancement to generate more lithium battery data of different types and working conditions, specifically:

[0012] The VAE-GAN combined model is composed of an encoder-generator-decoder-determinator structure. The encoder learns the probability distribution of the input lithium battery operation data, that is, captures the key features of the input data, and maps the input data to a representation of a latent space, that is, latent variables, and outputs it to the generator; the generator receives the latent variables to generate sample data close to the real data, and transmits it to the decoder; the decoder maps the sample data back to the original data space, reconstructs the data, and generates more sample data; the determiner compares and identifies the sample data output by the decoder with the real data, so as to help the generated sample data be closer to the real data. Using the VAE-GAN combined model, more available data is generated based on the collected lithium battery operation data of different types and working conditions.

[0013] Furthermore, in step (3), a model library is constructed. According to the lithium battery data of different types and working conditions in step (2), all models in the model library are modeled respectively. According to the amount of lithium battery modeling data, the accuracy and calculation time are weighed. The model that meets the accuracy requirements and time threshold is used as the source model. All source models are weighted and superimposed to integrate a stacking model, which is specifically:

[0014] A lithium battery health status assessment model library with various structures and modeling mechanisms is constructed. The models in the model library are divided into three categories. The first category has an advantage in small sample processing. The second category has a simple model structure and a fast running speed. The third category has a complex model and is used to capture the long-term characteristics of lithium battery operation data. Lithium battery data of different types and working conditions generated by the VAE-GAN combined model are input, and each model in the model library is modeled and trained separately. The accuracy and calculation time are weighed according to the amount of lithium battery modeling data. The model that meets the accuracy requirements and time threshold is used as the source model, and the source models are weighted and superimposed to form a stacked model.

[0015] Furthermore, in step (4), the operating data of the lithium battery to be evaluated is input, and the adaptive resonance theory ART network is used to perform adaptive weight allocation on each source model in the stacking model, specifically:

[0016] The ART network includes: input layer, node layer, reset mechanism, resonance threshold, learning rules and output layer. The process is as follows:

[0017] 4.1 Initialize the parameters of the ART network, including the resonance threshold β, the weight matrix, and the learning rate; the weight matrix includes a bottom-up weight vector and a top-down weight vector, and assigns initial weights to each source model, giving them the same weight coefficient;

[0018] The bottom-up weight vector is a vector stored inside the network to represent an existing pattern, and is used to measure the degree of matching between new input data and the existing pattern;

[0019] The top-down weight vector is a weight vector adjusted from the high-level structure to the bottom level according to the learning rules and resonance mechanism of the ART network after resonance occurs, that is, when the input data highly matches the existing pattern;

[0020] The resonance threshold refers to a set value used to determine the degree of match, ie, similarity, between input data and existing patterns in the network.

[0021] 4.2 Input the operating data of the lithium battery to be evaluated into the stacking model, and output the prediction results of each source model in the stacking model. The prediction results are Y = [y1, y2, ..y i ..,y n ],y i Represents the prediction result vector of the i-th source model;

[0022] 4.3 The input layer receives the prediction results of each source model as the input of the ART network and transmits it to the node layer;

[0023] 4.4 Each node in the node layer is connected by a bottom-up weight vector t jConnected to the input layer, these weight vectors participate in calculating the similarity between the input data and the node, thereby affecting the activation state of the node; assign each node to each source model, then for each node j, its bottom-up weight vector is t j =[t 1j ,t 2j ,...,t mj ]; m, n are positive integers;

[0024] 4.5 Calculate the matching degree between the input vector and each node by taking the dot product of the input vector and the bottom-up weight vector. The formula is as follows:

[0025] w=||y i ·t mj || / ||y i ||||t mj ||

[0026] In the formula, w represents the matching degree;

[0027] 4.6 Compare the matching degree w with the set resonance threshold β. When w<β, the reconstruction mechanism of the ART network will suppress the current node and search other nodes to find a better match. When w>β, the network enters a resonance state. In the resonance state, the top-down weight vector will use the ART network's own learning rules and update and adjust according to the input data to learn and strengthen this similarity.

[0028] 4.7 According to the output results of the ART network, the weight coefficient of each source model is adjusted, and the final prediction result of each source model is multiplied by its weight coefficient, and then weighted superposition is performed to output the preliminary prediction result.

[0029] Furthermore, the step (5) performs maximum information coefficient MIC feature extraction on the lithium battery operation data, specifically:

[0030] The operating data of the lithium battery to be evaluated is discretized on a two-dimensional plane, and a number of grid cells are divided on the horizontal and vertical coordinates of the plane, i.e., the x-direction and the y-direction, to construct a grid plane. The joint probability distribution of any two random operating data in these grid cells and their respective probability distributions are calculated, and then the maximum mutual information between them is calculated. The calculation formula of the maximum mutual information value is as follows:

[0031]

[0032] Where p(A,B) represents the joint probability distribution function of operation data A and operation data B; p(A) and p(B) are the marginal probability distribution functions of operation data A and operation data B respectively;

[0033] The maximum mutual information value is normalized. By calculating and comparing the maximum mutual information values ​​at different scales, the maximum value is selected as the maximum information coefficient MIC. The normalized expression is as follows:

[0034]

[0035] Where c and d are the number of grids divided in the x and y directions of the two-dimensional plane, respectively;

[0036] The maximum information coefficient MIC value between each operating data is calculated, and the feature data with a larger MIC value, that is, a higher correlation with the SOH of the lithium battery, is selected as the input variable of the training error correction model.

[0037] Furthermore, in step (5), the error sequence of the difference between the preliminary prediction result and the actual SOH value of the battery is used as the target variable, and an error correction model based on random forest RF is established to realize the prediction of future errors and the correction of prediction results, specifically:

[0038] The initial predicted SOH value is subtracted from the actual SOH value of the lithium battery to obtain error sequence data. The characteristic data is used as the input variable and the error sequence data is used as the output variable to establish an error correction model based on random forest RF. The specific process is as follows:

[0039] (1) Use the Bootstrap method of random forest to randomly select N sample data with replacement from the input feature data to form a sample set. Repeat this operation to generate multiple sample sets with a capacity of N.

[0040] (2) A decision tree is used for training each sample set. Different decision trees use different sample sets, and these sample sets are used as samples of the root node of the decision tree.

[0041] (3) Randomly select s feature data from the N feature data in the sample set, and use the information gain rate to select the optimal features of these s feature data. Each decision tree will split with the optimal feature;

[0042] (4) The splitting process of each node of each decision tree is carried out according to step (3) until the decision tree cannot be split any further;

[0043] (5) Repeat the above steps to combine all trained decision trees to form a random forest.

[0044] The error sequence of the difference between the preliminary prediction result and the actual SOH value of the battery is used as the target variable to train the random forest model, input the characteristic data of the battery to be evaluated into the model, summarize the prediction results of all decision trees, and calculate their average value, output the prediction value, that is, the error sequence, and compare it with the actual error sequence until the expected effect is achieved, that is, the training is completed;

[0045] Input the new battery characteristic data to be evaluated into the trained random forest error correction model again, and the predicted error sequence is directly output. It is added to the preliminary prediction result of the source model to obtain the final lithium battery SOH prediction value.

[0046] The present invention also discloses a lithium battery SOH intelligent estimation system, comprising: a data acquisition module, a data enhancement module, a model library, a stacking model, an ART network and an error correction module;

[0047] The data acquisition module is used to collect operating data of lithium batteries of different types and working conditions, including voltage, current, temperature, number of charging cycles, temperature and time data;

[0048] The data enhancement module is used to construct a VAE-GAN combined model, taking the data collected by the data acquisition module as input and performing data enhancement on it, so as to generate more operating data of different types of lithium batteries under different working conditions;

[0049] The model library is used to use the data enhanced by the data enhancement module as input, model all the models in the model library respectively, weigh the accuracy and calculation time according to the amount of lithium battery modeling data, take the model that meets the progress requirements and time threshold as the source model, perform weighted superposition on them, and integrate a stacking model;

[0050] The ART network is used to adaptively assign weights to each source model in the stacking model to obtain a final SOH prediction model for the battery to be evaluated, thereby outputting a preliminary prediction result;

[0051] The error correction model is based on the error correction model of random forest RF. It uses the lithium battery operation data after maximum information coefficient MIC feature extraction as input variables and the error sequence of the difference between the preliminary prediction result and the actual SOH value of the battery as the target variable for training, so as to realize the prediction of future errors and the correction of prediction results, and output the final prediction value of the lithium battery SOH.

[0052] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0053] 1. The operating data of lithium batteries of different types and working conditions were collected, and a new structure VAE-GAN combination model was proposed to further enhance the collected data, thereby generating a lot of lithium battery data of different types and working conditions, enhancing the diversity and coverage of the data set, and greatly improving the generalization ability of the prediction model;

[0054] 2. Through the constructed model library, multi-model modeling was carried out, and the models that met the accuracy and computing time requirements were selected as the source models, and the source models were weighted and superimposed; and the ART network was used to adaptively allocate the weights of the superimposed models. The prediction accuracy and ability of the overall model were improved, making the model more flexible in processing different input data to obtain the best prediction effect;

[0055] 3. An error correction model based on random forest was constructed to correct the errors of the preliminary prediction results, making the final prediction results more accurate and closer to the true value. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is the overall flow chart of the present invention. DETAILED DESCRIPTION

[0057] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0058] This embodiment provides a lithium battery SOH intelligent estimation method and system, as shown in the attached Figure 1 As shown, including:

[0059] Step (1) uses sensors to collect operating data of lithium batteries of different types and working conditions, including voltage, current, temperature, number of charging cycles, temperature and time data;

[0060] Step (2) constructing a variational autoencoder-generative adversarial network VAE-GAN combined model, inputting the collected lithium battery operation data in step (1) for data enhancement, and generating more lithium battery data of different types and working conditions;

[0061] Step (3) constructs a model library. According to the lithium battery data of different types and working conditions in step (2), all models in the model library are modeled separately. According to the amount of lithium battery modeling data, the accuracy and calculation time are weighed. The model that meets the accuracy requirements and time threshold is used as the source model. All source models are weighted and superimposed to integrate a stacked model.

[0062] Step (4) inputting the operating data of the lithium battery to be evaluated, using the adaptive resonance theory ART network to adaptively assign weights to each source model in the stacking model, obtaining the final SOH prediction model of the battery to be evaluated, and outputting a preliminary prediction result;

[0063] Step (5) extracts the maximum information coefficient MIC feature of the lithium battery operation data as the input variable, takes the error sequence of the difference between the preliminary prediction result and the actual SOH value of the battery as the target variable, establishes an error correction model based on random forest RF, realizes the prediction of future errors and the correction of prediction results, and outputs the final prediction value of the lithium battery SOH.

[0064] Among them, in step (2), a variational autoencoder-generative adversarial network VAE-GAN combined model is constructed, and the collected lithium battery operation data is input for data enhancement to generate more lithium battery data of different types and working conditions, specifically:

[0065] The VAE-GAN combined model consists of an encoder-generator-decoder-determinator structure. The encoder learns the probability distribution of the input lithium battery operation data, that is, captures the key features of the input data, and maps the input data to a representation of a latent space, that is, latent variables, and outputs it to the generator; the generator receives the latent variables to generate sample data close to the real data, and transmits it to the decoder; the decoder maps the sample data back to the original data space, reconstructs the data, and generates more sample data; the determiner compares and identifies the sample data output by the decoder with the real data, so as to help the generated sample data be closer to the real data. The VAE-GAN combined model is used to generate more available data based on the collected lithium battery operation data of different types and working conditions.

[0066] In step (3), a model library is constructed. According to the lithium battery data of different types and working conditions in step (2), all models in the model library are modeled respectively. According to the amount of lithium battery modeling data, the accuracy and calculation time are weighed. The model that meets the accuracy requirements and time threshold is used as the source model. All source models are weighted and superimposed to integrate a stacking model, which is specifically:

[0067] A lithium battery health status assessment model library with various structures and modeling mechanisms is constructed. The models in the model library are divided into three categories. The first category has an advantage in small sample processing, including support vector machine SVM, Gaussian regression model GPR and other models. The second category has a simple model structure and fast running speed, including random forest RF, gradient boosting decision tree GBDT and other models; the third category has complex models, which are used to capture the long-term characteristics of lithium battery operation data, including long short-term memory network LSTM, Pathformer and other models. The lithium battery data of different types and working conditions generated by the VAE-GAN combination model are input, and each model in the model library is modeled and trained respectively. The accuracy and calculation time are weighed according to the amount of lithium battery modeling data. The model that meets the accuracy requirements and time threshold is selected from the three types of models as the source model, and the source models are weighted and superimposed to form a stacked model.

[0068] In step (4), the operating data of the lithium battery to be evaluated is input, and the adaptive resonance theory ART network is used to perform adaptive weight allocation on each source model in the stacking model, specifically:

[0069] The ART network consists of: input layer, node layer, reset mechanism, resonance threshold, learning rules and output layer. The process is as follows:

[0070] 4.1 Initialize the parameters of the ART network, including the resonance threshold β, the weight matrix, and the learning rate; the weight matrix includes a bottom-up weight vector and a top-down weight vector, and assigns initial weights to each source model, giving them the same weight coefficient;

[0071] Bottom-up weight vector: a vector stored inside the network to represent an existing pattern, used to measure the degree of match between new input data and the existing pattern;

[0072] Top-down weight vector: According to the learning rules and resonance mechanism of the ART network, the weight vector is adjusted from the high-level structure to the bottom level after resonance occurs, that is, when the input data highly matches the existing pattern;

[0073] Resonance threshold: A set value used to determine the degree of match, or similarity, between input data and existing patterns in the network.

[0074] 4.2 Input the operating data of the lithium battery to be evaluated into the stacking model, and output the prediction results of each source model in the stacking model. The prediction results are Y = [y1, y2, ..y i ..,y n ],y i Represents the prediction result vector of the i-th source model;

[0075] 4.3 The input layer receives the prediction results of each source model as the input of the ART network and transmits it to the node layer;

[0076] 4.4 Each node in the node layer is connected by a bottom-up weight vector t j Connected to the input layer, these weight vectors participate in calculating the similarity between the input data and the node, thereby affecting the activation state of the node; assign each node to each source model, then for each node j, its bottom-up weight vector is t j =[t 1j ,t 2j ,...,t mj ];

[0077] 4.5 Calculate the matching degree between the input vector and each node by taking the dot product of the input vector and the bottom-up weight vector. The formula is as follows:

[0078] w=||y i ·t mj || / ||y i ||||t mj ||

[0079] In the formula, w represents the matching degree;

[0080] 4.6 Compare the matching degree w with the set resonance threshold β. When w<β, the reconstruction mechanism of the ART network will suppress the current node and search other nodes to find a better match. When w>β, the network enters a resonance state. In the resonance state, the top-down weight vector will use the ART network's own learning rules and update and adjust according to the input data to learn and strengthen this similarity.

[0081] 4.7 According to the output results of the ART network, the weight coefficient of each source model is adjusted, and the final prediction result of each source model is multiplied by its weight coefficient, and then weighted superposition is performed to output the preliminary prediction result.

[0082] Among them, step (5) extracts the maximum information coefficient MIC feature of the lithium battery operation data, specifically:

[0083] The operating data of the lithium battery to be evaluated is discretized on a two-dimensional plane, and a number of grid cells are divided on the horizontal and vertical coordinates of the plane, i.e., the x-direction and the y-direction, to construct a grid plane. The joint probability distribution of any two random operating data in these grid cells and their respective probability distributions are calculated, and then the maximum mutual information between them is calculated. The calculation formula of the maximum mutual information value is as follows:

[0084]

[0085] Where p(A,B) represents the joint probability distribution function of operation data A and operation data B; p(A) and p(B) are the marginal probability distribution functions of operation data A and operation data B respectively;

[0086] The maximum mutual information value is normalized. By calculating and comparing the maximum mutual information values ​​at different scales, the maximum value is selected as the maximum information coefficient MIC. The normalized expression is as follows:

[0087]

[0088] Where c and d are the number of grids divided in the x and y directions of the two-dimensional plane, respectively;

[0089] The maximum information coefficient MIC value between each operating data is calculated, and the feature data with a larger MIC value, that is, a higher correlation with the SOH of the lithium battery, is selected as the input variable of the training error correction model.

[0090] Among them, step (5) takes the error sequence of the difference between the preliminary prediction result and the actual SOH value of the battery as the target variable, establishes an error correction model based on random forest RF, and realizes the prediction of future errors and the correction of prediction results, specifically:

[0091] The initial predicted SOH value is subtracted from the actual SOH value of the lithium battery to obtain error sequence data. The characteristic data is used as the input variable and the error sequence data is used as the output variable to establish an error correction model based on random forest RF. The specific process is as follows:

[0092] (1) Use the Bootstrap method of random forest to randomly select N sample data with replacement from the input feature data to form a sample set. Repeat this operation to generate multiple sample sets with a capacity of N.

[0093] (2) A decision tree is used for training each sample set. Different decision trees use different sample sets, and these sample sets are used as samples of the root node of the decision tree.

[0094] (3) Randomly select s feature data from the N feature data in the sample set, and use the information gain rate to select the optimal features of these s feature data. Each decision tree will split with the optimal feature;

[0095] (4) The splitting process of each node of each decision tree is carried out according to step (3) until the decision tree cannot be split any further;

[0096] (5) Repeat the above steps to combine all trained decision trees to form a random forest.

[0097] The error sequence of the difference between the preliminary prediction result and the actual SOH value of the battery is used as the target variable to train the random forest model, input the characteristic data of the battery to be evaluated into the model, summarize the prediction results of all decision trees, and calculate their average value, output the prediction value, that is, the error sequence, and compare it with the actual error sequence until the expected effect is achieved, that is, the training is completed;

[0098] Input the new battery characteristic data to be evaluated into the trained random forest error correction model again, and the predicted error sequence is directly output. It is added to the preliminary prediction result of the source model to obtain the final lithium battery SOH prediction value.

[0099] The present invention also discloses a lithium battery SOH intelligent estimation system, comprising: a data acquisition module, a data enhancement module, a model library, a stacking model, an ART network and an error correction module;

[0100] Data acquisition module, used to collect operating data of lithium batteries of different types and working conditions, including voltage, current, temperature, number of charging cycles, temperature and time data;

[0101] The data enhancement module is used to build a VAE-GAN combination model, taking the data collected by the data acquisition module as input and performing data enhancement on it, so as to generate more operating data of different types of lithium batteries under different working conditions;

[0102] The model library is used to take the data enhanced by the data enhancement module as input, model all the models in the model library separately, weigh the accuracy and calculation time according to the amount of lithium battery modeling data, take the models that meet the progress requirements and time threshold as the source models, perform weighted superposition on them, and integrate a stacking model;

[0103] The ART network is used to adaptively assign weights to each source model in the stacking model to obtain the final SOH prediction model of the battery to be evaluated, thereby outputting a preliminary prediction result;

[0104] The error correction model is based on the random forest RF error correction model. It uses the lithium battery operation data after maximum information coefficient MIC feature extraction as the input variable and the error sequence of the difference between the preliminary prediction result and the actual SOH value of the battery as the target variable for training. It realizes the prediction of future errors and the correction of prediction results, and outputs the final prediction value of the lithium battery SOH.

[0105] The present invention also includes a computing device, comprising: one or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing a lithium battery SOH intelligent estimation method.

[0106] The present invention also includes a computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, which, when executed by a computing device, enable the computing device to perform a lithium battery SOH intelligent estimation method.

Claims

1. A lithium battery SOH intelligent estimation method, characterized in that: include: Step (1) The sensor collects operating data of lithium batteries of different types and working conditions, including voltage, current, temperature, number of charging cycles, temperature and time data; Step (2) constructing a variational autoencoder-generative adversarial network VAE-GAN combined model, inputting the lithium battery operation data collected in step (1) for data enhancement, and generating more lithium battery operation data of different types and working conditions; Step (3) constructing a model library, modeling all models in the model library according to the operating data of lithium batteries of different types and working conditions in step (2), weighing the accuracy and calculation time according to the amount of lithium battery modeling data, taking the model that meets the accuracy requirements and time threshold as the source model, and weighting all the source models to integrate a stacked model; Step (4) inputting the operating data of the lithium battery to be evaluated, using the adaptive resonance theory ART network to adaptively assign weights to each source model in the stacking model, obtaining the final SOH prediction model of the battery to be evaluated, and outputting a preliminary prediction result; Step (5) extracts the maximum information coefficient MIC feature of the lithium battery operation data as the input variable, takes the error sequence of the difference between the preliminary prediction result and the actual SOH value of the battery as the target variable, establishes an error correction model based on random forest RF, realizes the prediction of future errors and the correction of prediction results, and outputs the final prediction value of the lithium battery SOH.

2. A lithium battery SOH intelligent estimation method according to claim 1, characterized in that: The variational autoencoder-generative adversarial network VAE-GAN combined model is composed of an encoder-generator-decoder-determinator structure. The encoder learns the probability distribution of the input lithium battery operation data, that is, captures the key features of the input data, and maps the input data to a representation of a latent space, that is, latent variables, and outputs it to the generator; the generator receives the latent variables to generate sample data close to the real data, and transmits it to the decoder; the decoder maps the sample data back to the original data space, reconstructs the data, and generates more sample data; The determiner compares and identifies the sample data output by the decoder with the real data, so that the generated sample data is closer to the real data.

3. A lithium battery SOH intelligent estimation method according to claim 1, characterized in that: The step (3) is specifically: A lithium battery health status assessment model library with various structures and different modeling mechanisms is constructed, and lithium battery operation data of different types and working conditions generated by the variational autoencoder-generative adversarial network VAE-GAN combined model is input. Each model in the model library is modeled and trained separately, and the accuracy and calculation time are weighed according to the amount of lithium battery modeling data. The model that meets the accuracy requirements and time threshold is used as the source model, and the source models are weighted and superimposed to form a stacked model.

4. A lithium battery SOH intelligent estimation method according to claim 1, characterized in that: In step (4), the operating data of the lithium battery to be evaluated is input, and the adaptive resonance theory ART network is used to perform adaptive weight allocation on each source model in the stacking model, specifically: The ART network includes: input layer, node layer, reset mechanism, resonance threshold, learning rules and output layer. The process is as follows: 4.1 Initialize the parameters of the ART network, including the resonance threshold β, the weight matrix, and the learning rate; the weight matrix includes a bottom-up weight vector and a top-down weight vector, and assigns initial weights to each source model, giving them the same weight coefficient; The bottom-up weight vector is a vector stored inside the network to represent an existing pattern, and is used to measure the degree of matching between new input data and the existing pattern; The top-down weight vector is a weight vector adjusted from the high-level structure to the bottom level according to the learning rules and resonance mechanism of the ART network after resonance occurs, that is, when the input data highly matches the existing pattern; The resonance threshold refers to a set value used to determine the degree of match, i.e., similarity, between input data and existing patterns in the network; 4.2 Input the operating data of the lithium battery to be evaluated into the stacking model, and output the prediction results of each source model in the stacking model. The prediction results are Y = [y1, y2, ..y i ..,y n ],y i Represents the prediction result vector of the i-th source model; 4.3 The input layer receives the prediction results of each source model as the input of the ART network and transmits it to the node layer; 4.4 Each node in the node layer is connected by a bottom-up weight vector t j Connected to the input layer, these weight vectors participate in calculating the similarity between the input data and the node, thereby affecting the activation state of the node; Assign each node to each source model, then for each node j, its bottom-up weight vector is t j =[t 1j ,t 2j ,...,t mj ]; m, n are positive integers; 4.5 Calculate the matching degree between the input vector and each node by taking the dot product of the input vector and the bottom-up weight vector. The formula is as follows: w=||y i ·t mj || / ||y i ||||t mj || In the formula, w represents the matching degree; 4.6 Compare the matching degree w with the set resonance threshold β. When w<β, the reconstruction mechanism of the ART network will suppress the current node and search other nodes to find a better match. When w>β, the network enters a resonance state. In the resonance state, the top-down weight vector will use the ART network's own learning rules and update and adjust according to the input data to learn and strengthen this similarity. 4.7 According to the output results of the ART network, the weight coefficient of each source model is adjusted, and the final prediction result of each source model is multiplied by its weight coefficient, and then weighted superposition is performed to output the preliminary prediction result.

5. A lithium battery SOH intelligent estimation method according to claim 1, characterized in that: The step (5) extracts the maximum information coefficient MIC feature from the lithium battery operation data, specifically: The operating data of the lithium battery to be evaluated is discretized on a two-dimensional plane, and a number of grid cells are divided on the horizontal and vertical coordinates of the plane, i.e., the x-direction and the y-direction, to construct a grid plane. The joint probability distribution of any two random operating data in these grid cells and their respective probability distributions are calculated, and then the maximum mutual information between them is calculated. The calculation formula of the maximum mutual information value is as follows: Where p(A,B) represents the joint probability distribution function of operation data A and operation data B; p(A) and p(B) are the marginal probability distribution functions of operation data A and operation data B respectively; The maximum mutual information value is normalized. By calculating and comparing the maximum mutual information values ​​at different scales, the maximum value is selected as the maximum information coefficient MIC. The normalized expression is as follows: Where c and d are the number of grids divided in the x and y directions of the two-dimensional plane, respectively; The maximum information coefficient MIC value between each operating data is calculated, and the feature data with a larger MIC value, that is, a higher correlation with the SOH of the lithium battery, is selected as the input variable of the training error correction model.

6. A lithium battery SOH intelligent estimation method according to claim 1, characterized in that: In step (5), the error sequence of the difference between the preliminary prediction result and the actual SOH value of the battery is used as the target variable, and an error correction model based on random forest RF is established to realize the prediction of future errors and the correction of prediction results, specifically: The initial predicted SOH value is subtracted from the actual SOH value of the lithium battery to obtain error sequence data. The characteristic data is used as the input variable and the error sequence data is used as the output variable to establish an error correction model based on random forest RF. The specific process is as follows: (1) Use the Bootstrap method of random forest to randomly select N sample data with replacement from the input feature data to form a sample set. Repeat this operation to generate multiple sample sets with a capacity of N. (2) A decision tree is used for training each sample set. Different decision trees use different sample sets, and these sample sets are used as samples of the root node of the decision tree. (3) Randomly select s feature data from the N feature data in the sample set, and use the information gain rate to select the optimal features of these s feature data. Each decision tree will split with the optimal feature; (4) The splitting process of each node of each decision tree is carried out according to step (3) until the decision tree cannot be split any further; (5) Repeat the above steps to combine all trained decision trees to form a random forest. The error sequence of the difference between the preliminary prediction result and the actual SOH value of the battery is used as the target variable to train the random forest model, input the characteristic data of the battery to be evaluated into the model, summarize the prediction results of all decision trees, and calculate their average value, output the prediction value, that is, the error sequence, and compare it with the actual error sequence until the expected effect is achieved, that is, the training is completed; Input the new battery characteristic data to be evaluated into the trained random forest error correction model again, and the predicted error sequence is directly output. It is added to the preliminary prediction result of the source model to obtain the final lithium battery SOH prediction value.

7. A lithium battery SOH intelligent estimation system, characterized in that: include: Data acquisition module, data enhancement module, model library, stacking model, ART network and error correction module; The data acquisition module is used to collect operating data of lithium batteries of different types and working conditions, including voltage, current, temperature, number of charging cycles, temperature and time data; The data enhancement module is used to construct a variational autoencoder-generative adversarial network VAE-GAN combined model, which takes the data collected by the data acquisition module as input and performs data enhancement on it, thereby generating more operating data of different types of lithium batteries under different working conditions; The model library is used to use the data enhanced by the data enhancement module as input, model all the models in the model library respectively, weigh the accuracy and calculation time according to the amount of lithium battery modeling data, take the model that meets the progress requirements and time threshold as the source model, perform weighted superposition on them, and integrate a stacking model; The ART network is used to adaptively assign weights to each source model in the stacking model to obtain a final SOH prediction model for the battery to be evaluated, thereby outputting a preliminary prediction result; The error correction model is based on the error correction model of random forest RF. It uses the lithium battery operation data after maximum information coefficient MIC feature extraction as input variables and the error sequence of the difference between the preliminary prediction result and the actual SOH value of the battery as the target variable for training, so as to realize the prediction of future errors and the correction of prediction results, and output the final prediction value of the lithium battery SOH.

8. A lithium battery SOH intelligent estimation system according to claim 7, characterized in that: The variational autoencoder-generative adversarial network VAE-GAN combined model is composed of an encoder-generator-decoder-determinator structure. The encoder learns the probability distribution of the input lithium battery operation data, that is, captures the key features of the input data, and maps the input data to a representation of a latent space, that is, latent variables, and outputs it to the generator; the generator receives the latent variables to generate sample data close to the real data, and transmits it to the decoder; the decoder maps the sample data back to the original data space, reconstructs the data, and generates more sample data; The determiner compares and identifies the sample data output by the decoder with the real data, thereby helping to make the generated sample data closer to the real data, and using the VAE-GAN combined model to generate more usable data based on the collected lithium battery operation data of different types and working conditions.

9. A computing device, characterized in that include: One or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing any of the methods according to claims 1 to 6.

10. A computer-readable storage medium storing one or more programs, characterized in that: The one or more programs include instructions which, when executed by a computing device, cause the computing device to perform any one of the methods according to claims 1 to 6.

Images