Multi-mechanism constrained energy storage battery anti-decoupling characterization and health estimation method

Abstract

The invention discloses an adversarial decoupling characterization and health estimation method for an energy storage battery constrained by multiple mechanisms, and the method comprises the steps: firstly carrying out the decoupling characterization of an input feature as a health state related feature and a working condition related domain feature through the fusion application of adversarial decoupling characterization learning and a mechanism equation in a neural network; secondly, mechanism parameters are generated through a mechanism parameter generator, and the problem that mechanism equation parameters are different under different working conditions is solved; finally, through continuous circulation physical constraint, the health state is restrained to decline along with the increase of the number of cycles, the problem of change generated by mechanism equation parameters is solved, and then generalization estimation of the health state of the unseen working condition is achieved. According to the method, multiple mechanism constraints are fused in decoupling characterization learning of the energy storage battery for the first time, generalization modeling of unseen working conditions is achieved under the condition that target working condition data are not needed, then the health state estimation precision of the model under the unseen working conditions is improved, and necessary support is provided for safe operation and maintenance of the energy storage battery.

Description

Technical Field

[0001] The present invention relates to the technical field of health management of energy storage power stations, especially to the technical field of continuous degradation problems of lithium-ion battery stacks during operation in energy storage power stations, and particularly relates to a method for anti-decoupling characterization and health estimation of energy storage batteries with multiple mechanism constraints. Background Art

[0002] With the rapid development of renewable energy and the popularization of electric vehicles, energy storage batteries are used more and more widely. Energy storage systems not only play a key role in the dispatching and optimization of power grids, but also play an indispensable role in the low-carbon economy. The state of health (SOH) of energy storage batteries is an important indicator for evaluating their performance and life, directly affecting the reliability and economy of energy storage systems. To ensure the efficient operation of the system in the long term, it is crucial to accurately evaluate the health state of energy storage batteries. However, due to the diversity of the battery usage environment and working conditions, there are significant differences even among batteries of the same model. This poses higher requirements for the further research and development of health state estimation (SOH Estimation) technology to achieve generalization ability from one type of battery to another.

[0003] Although there have been many studies on the health assessment methods for lithium batteries, each type of method has its own limitations. The direct measurement method is only suitable for use in the laboratory. Model-based methods often have difficulty ensuring the accuracy of the model or have too many parameters to be accurately identified. Therefore, more and more research tends to use data-driven methods to explore the aging mechanism and state inside the battery. However, the vast majority of data-driven assessment methods rely on the distribution of training data. Although existing research uses transfer learning to solve this problem, partial information in the target domain, i.e., the real application scenario, is still required. How to achieve domain generalization of the model without the operation data of energy storage batteries under real working conditions is a major problem. Representation learning is a means of domain generalization, but general representation methods align features between different domains. Due to the lack of mechanism constraints, only similar features that can be aligned can be found in the known source domain, and a good effect cannot be directly achieved. Therefore, how to accurately generalize and estimate the health state of energy storage batteries under unseen working conditions remains an urgent problem to be solved. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for anti-decoupling characterization and health estimation of energy storage batteries with multiple mechanism constraints, aiming at the deficiencies in the prior art that the health state estimation of energy storage batteries lacks mechanism constraints and requires operation data under target working conditions. The present invention can estimate the health state of energy storage batteries.

[0005] The object of the present invention is achieved by the following technical solutions: A method for anti-decoupling characterization and health estimation of an energy storage battery with multiple mechanism constraints, comprising the following steps: (1) Obtain the complete full-life operation data of multiple energy storage batteries under different test conditions, including the current, voltage, sampling time, cycle number, and corresponding cycle capacity data of each charge-discharge cycle, so as to construct a multi-source domain dataset; (2) Preprocess the complete full-life operation data of each charge-discharge cycle in each source domain dataset to extract health features and the health status label of each charge-discharge cycle, and merge the health features with the cycle number as input data; at the same time, initialize the parameters of the mechanism equation for each source domain and save them as mechanism parameter labels; construct a training set according to the input data and its corresponding health status label, domain label, and mechanism parameter label; (3) Construct a neural network model, including an encoder, a health status estimator, a domain discriminator, and a mechanism parameter generator; and use the training set to train the neural network model. During the training process, minimize the total loss function of the neural network model as the optimization target, and adjust the parameters of the neural network model to obtain a trained encoder and health status estimator; (4) During on-line application, obtain the charge-discharge cycle data of the energy storage battery under the operating conditions, including the operating current, voltage, sampling time, and cycle number; extract the health features based on the on-line operating current, voltage, and sampling time, normalize them, and then merge them with the cycle number as input data, and input them into the trained encoder to obtain encoded features. After splitting the encoded features in half, obtain the health status features and domain features. Input the health status features into the trained health status estimator to obtain the health status value of the current loss, and further obtain the current health status estimation value of the energy storage battery.

[0006] Further, the step (1) specifically includes: Obtain the current, voltage, sampling time, cycle number, and cycle capacity data during the standard charging period of each charge-discharge cycle of multiple energy storage batteries from the start of the test to the end of their life cycle under different test conditions; use the charge-discharge rate and the ambient temperature set in the test as the basis for dividing different test conditions, and divide the collected current, voltage, sampling time, cycle number, and cycle capacity data into different source domains according to different charge-discharge rates and ambient temperatures, and merge the charge-discharge rate and ambient temperature as the domain label to construct a multi-source domain dataset.

[0007] Further, the step (2) specifically includes the following sub-steps: (2.1) For the complete life-cycle operation data of each charge-discharge cycle in each source domain data set, extract health features based on the current, voltage and sampling time of each charge-discharge cycle; wherein the health features include the mean value, standard deviation, kurtosis, skewness, curve slope and curve entropy of the voltage and current during the cyclic charging period, as well as the time and accumulated charge of the two stages of constant current charging and constant voltage charging, a total of 16 health features; (2.2) For the health features extracted in step (2.1), first use the 3-σ criterion to eliminate outliers in the health features, then use the normalization method to normalize the health features to obtain the normalized health features, and combine the number of cycles as input data; (2.3) Divide the cycle capacity data of each charge and discharge cycle by the rated capacity of the corresponding energy storage battery to calculate the health status label of each charge and discharge cycle; (2.4) Using the particle swarm optimization algorithm, the parameters of the mechanism equation of each source domain are initialized and saved as mechanism parameter labels; (2.5) Construct a training set based on the input data and its corresponding health state labels, domain labels, and mechanism parameter labels.

[0008] Furthermore, the mechanism equation of the source domain is an equation derived from the Arrhenius formula, and its differential form is:

[0009] In the formula, u represents the health status value of the loss, u=1-y, y represents the health status label; t represents the current cycle number; e is a natural constant; T represents the ambient temperature set for the test; are two parameters of the mechanism equation; is the differential symbol.

[0010] Furthermore, in the neural network model, the encoder includes an input layer, a hidden layer and an output layer, the input layer includes a fully connected layer and an activation function, the hidden layer is a single-layer base network layer, and the output layer is a fully connected layer; the health state estimator, the domain discriminator and the mechanism parameter generator all include a single-layer base network layer and a fully connected layer; wherein the single-layer base network layer includes a fully connected layer, an activation function and a random loss layer, and the activation functions are all sinusoidal functions; The encoder is used to perform feature encoding on the input data input into the neural network model to obtain encoded features; then the encoded features are split in half to obtain two features, with the former feature used as the health state feature and the latter feature used as the domain feature; the health state estimator is used to map the health state feature to a loss of health state value; the domain discriminator is used to discriminate the working condition category of the input domain feature; the encoded features are simultaneously input into the mechanism parameter generator to be used for generating the parameters of the mechanism equation corresponding to each sample.

[0011] Further, the neural network model is trained using the training set. During the training process, with the goal of minimizing the total loss function of the neural network model, the parameters of the neural network model are adjusted to obtain the trained encoder and health state estimator, which specifically includes: Input the input data of the cyclic samples in the training set into the neural network model, and obtain encoded features through the encoder; after splitting the encoded features in half, obtain the health state feature and the domain feature; the health state feature passes through the health state estimator to obtain the loss of health state value output by the health state estimator; the domain feature passes through the domain discriminator to obtain the working condition category output by the domain discriminator for the domain feature; the health state feature passes through the gradient reversal layer and the domain discriminator in sequence to obtain the working condition category output by the domain discriminator for the health state feature; the encoded features pass through the mechanism parameter generator to obtain the parameters of the mechanism equation output by the mechanism parameter generator; Calculate the health label loss based on the loss of health state value output by the health state estimator and its corresponding true loss of health state value, where the true loss of health state value is obtained by subtracting the corresponding health state label in the training set from 1; calculate the domain discrimination loss based on the working condition category output by the domain discriminator for the domain feature, the working condition category output by the domain discriminator for the health state feature, and the domain label corresponding to it in the training set; calculate the parameter generation loss based on the parameters of the mechanism equation output by the mechanism parameter generator and the corresponding mechanism parameter label in the training set; calculate the physical equation loss based on the loss of health state value output by the health state estimator, the parameters of the mechanism equation output by the mechanism parameter generator, and the corresponding cycle number and environmental temperature in the training set; calculate the continuous cycle physical constraint loss based on the loss of health state value output by the health state estimator and the parameters of the mechanism equation output by the mechanism parameter generator; calculate the total loss function of the neural network model based on the health label loss, the domain discrimination loss, the parameter generation loss, the physical equation loss, and the continuous cycle physical constraint loss; With the goal of minimizing the total loss function of the neural network model, use the gradient descent method for iterative optimization to adjust the parameters of the encoder, the health state estimator, the domain discriminator, and the mechanism parameter generator until the preset number of training rounds is reached to obtain the trained encoder and health state estimator.

[0012] Furthermore, the calculation formula of the total loss function of the neural network model is as follows:

[0013] In the formula, represents the total loss function of the neural network model, represents the health label loss, represents the domain discrimination loss, represents the parameter generation loss, represents the physical equation loss, represents the continuous cyclic physical constraint loss, and respectively represent the fixed weight parameters corresponding to the health label loss and the domain discrimination loss corresponding fixed weight parameters, is a dynamic weight parameter, and its dynamic change formula is:

[0014] In the formula, p is a dynamic parameter that linearly increases from 0 to 1 with the number of training rounds; The calculation formula of the health label loss is:

[0015] In the formula, n represents the total number of working condition categories, represents the total number of energy storage batteries under the i-th working condition category, represents the total number of cyclic samples of the j-th energy storage battery, represents the health state value of the loss corresponding to the output of the health state estimator for the k-th cyclic sample of the j-th energy storage battery under the i-th working condition category, represents the true health state value of the loss corresponding to the k-th cyclic sample of the j-th energy storage battery under the i-th working condition category; The calculation formula of the domain discrimination loss is:

[0016] In the formula, represents the domain feature of the k-th cyclic sample of the j-th energy storage battery under the i-th working condition category the working condition category output by the domain discriminator, represents the health state feature of the k-th cyclic sample of the j-th energy storage battery under the i-th working condition category the working condition category output by the domain discriminator, represents the domain label corresponding to the k-th cyclic sample of the j-th energy storage battery under the i-th working condition category in the training set; The calculation formula of the parameter generation loss is:

[0017] In the formula, represents the parameter of the mechanism equation output by the mechanism parameter generator corresponding to the k-th cycle sample of the j-th energy storage battery under the i-th working condition category, represents the mechanism parameter label corresponding to the k-th cycle sample of the j-th energy storage battery under the i-th working condition category in the training set; The calculation formula of the loss of the physical equation is:

[0018] In the formula, represents the number of cycles of the k-th cycle sample of the j-th energy storage battery under the i-th working condition category, represents the ambient temperature of the working condition corresponding to the k-th cycle sample of the j-th energy storage battery under the i-th working condition category; The calculation formula of the continuous cycle physical constraint loss is:

[0019] In the formula, is the continuous cycle physical constraint loss, represents the rectified linear unit function.

[0020] The beneficial effects of the present invention are as follows: The present invention first integrates multiple mechanism constraints into the decoupled representation learning of energy storage batteries, and realizes the generalization modeling of unseen working conditions without the need for target working condition data, thereby improving the health state estimation accuracy of the model under unseen working conditions and providing necessary support for the safe operation and maintenance of energy storage batteries. Description of the Drawings

[0021] Figure 1 is a flowchart of the method for anti-decoupled representation and health estimation of energy storage batteries with multiple mechanism constraints of the present invention; Figure 2 is a schematic diagram of the network structure of the neural network model of the present invention; Figure 3 is an experimental result diagram of the method of the present invention. Detailed Embodiments

[0022] Here, the exemplary embodiments will be described in detail, and the examples are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. On the contrary, they are only examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0023] The terms used in the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The singular forms "a", "said" and "the" used in the present invention and the appended claims are also intended to include the plural forms unless the context clearly dictates otherwise. It should also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0024] It should be understood that although the terms first, second, third, etc. may be used in the present invention to describe various information, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present invention, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "when" or "while" or "in response to determining".

[0025] The present invention will be described in detail below with reference to the accompanying drawings. Without conflict, the features in the following embodiments and implementation manners may be combined with each other.

[0026] See Figure 1 , the method for anti-decoupling characterization and health estimation of an energy storage battery with multiple mechanism constraints of the present invention specifically includes the following steps:

[0027] (1) Obtain the complete full-life operation data of multiple energy storage batteries under different test conditions, including the current, voltage, sampling time, cycle number, and corresponding cycle capacity data of each charge-discharge cycle, so as to construct a multi-source domain dataset.

[0028] Specifically, obtain the current, voltage, sampling time, cycle number, and cycle capacity data during the standard charging period of each charge-discharge cycle of multiple energy storage batteries from the start of the test to the end of their life cycles under different test conditions; use the charge-discharge rate and the ambient temperature set in the test as the basis for dividing different test conditions, and divide the collected current, voltage, sampling time, cycle number, and cycle capacity data into different source domains according to different charge-discharge rates and ambient temperatures. The energy storage batteries under the same test condition are divided into the same source domain, and the charge-discharge rate and ambient temperature are combined as the domain label d to construct a multi-source domain dataset , where n represents the total number of source domain datasets.

[0029] It should be understood that the cycle number is the cycle number sequence number of the charge-discharge cycle. That is, for the same energy storage battery, the cycle number of the first charge-discharge cycle is 1, the cycle number of the second charge-discharge cycle is 2, the cycle number of the third charge-discharge cycle is 3, and similarly, the cycle number is the cycle number sequence number of the charge-discharge cycle.

[0030] It should be noted that the multiple energy storage batteries described in the present invention all follow a standard charging scheme: first, they are charged to 4.2V by constant current, and then charged at a constant voltage of 4.2V until a cut-off current of 0.05C is reached. The current rate during constant current charging is the charging rate. In this embodiment, the current rate during constant current charging ranges from 0.25C to 1C, and the ambient temperature set in the experiment is 25°C to 45°C.

[0031] (2) The complete life-cycle operation data of each charge-discharge cycle in each source domain dataset is preprocessed to extract health features and the health status label of each charge-discharge cycle, and the health features are combined with the number of cycles as input data; at the same time, the parameters of the mechanism equation of each source domain are initialized and saved as mechanism parameter labels; a training set is constructed based on the input data and its corresponding health status labels, domain labels, and mechanism parameter labels.

[0032] (2.1) For the complete life-cycle operation data of each charge-discharge cycle in each source domain dataset, health features are extracted based on the current, voltage and sampling time of each charge-discharge cycle; the health features include the mean, standard deviation, kurtosis, skewness, curve slope and curve entropy of the voltage and current during the cyclic charging, as well as the time and accumulated charge of the constant current charging and constant voltage charging stages, a total of 16 health features.

[0033] (2.2) For the health features extracted in step (2.1), the 3-σ criterion is first used to eliminate outliers in the health features, and then the normalization method is used to normalize the health features to obtain the normalized health features, and the number of cycles is combined as input data, for a total of 17 input features.

[0034] It should be understood that the 3σ criterion means that in a normal distribution, most of the data will be concentrated around the mean. The range of the data point is 10 times the standard deviation, that is, 99.73% of the data are within this interval. If a data point exceeds this range, it is considered an outlier or bad value and can be considered for removal.

[0035] Furthermore, common normalization methods include Min-Max normalization method, Z-Score normalization method, etc., which can be selected according to actual needs.

[0036] (2.3) The health status label of each charge and discharge cycle is calculated by dividing the cycle capacity data of each charge and discharge cycle by the rated capacity of the corresponding energy storage battery. The lost health status value u=1-y, where y represents the health status label.

[0037] (2.4) Initialize the parameters of the mechanism equation for each source domain using the Particle Swarm Optimization (PSO). In this embodiment, there are n source domains, so there are a total of n groups of parameters, which are saved as mechanism parameter labels.

[0038] It should be understood that PSO is a meta-heuristic optimization algorithm based on swarm intelligence, which simulates the social behavior of bird flocks or fish schools and finds the optimal solution through the cooperation of individuals and the group.

[0039] Further, the mechanism equation of the source domain is an equation derived from the Arrhenius formula, and its differential form is:

[0040] In the formula, u represents the value of the lost health state, u = 1 - y, where y represents the health state label; t represents the current number of iterations; e is the natural constant; T represents the environmental temperature set in the experiment; are two parameters of the mechanism equation, and are initialized and calculated for these two parameters through PSO; This is the differential symbol. is the differential symbol.

[0041] It should be understood that the Arrhenius equation is a classical formula that describes the relationship between the reaction rate constant and temperature in a chemical reaction.

[0042] (2.5) Construct a training set based on the input data, its corresponding health state label, domain label, and mechanism parameter label.

[0043] (3) Construct a neural network model, including an encoder, a health state estimator, a domain discriminator, and a mechanism parameter generator; and use the training set to train the neural network model. During the training process, minimize the total loss function of the neural network model as the optimization objective, and adjust the parameters of the neural network model to obtain a trained encoder and health state estimator.

[0044] In this embodiment, the structure of the neural network model is as Figure 2 shown. The encoder includes an input layer, a hidden layer, and an output layer. The input layer includes a fully connected layer and an activation function. The hidden layer is a single-layer basic network layer, and the output layer is a fully connected layer; the health state estimator, domain discriminator, and mechanism parameter generator all include a single-layer basic network layer and a fully connected layer; among them, the single-layer basic network layer includes a fully connected layer, an activation function, and a dropout layer, and the activation function is a sine function. Specifically, the encoder is used to perform feature encoding on the input data x input to the neural network model to obtain encoded features; then split the encoded features in half to obtain two features, and use the first feature as the health state feature Take the latter feature as the domain feature ; The health state estimator is used to map the health state feature to the lost health state value u; The domain discriminator is used to discriminate the working condition category of the input domain feature ; The encoded feature is input into the mechanism parameter generator at the same time for generating the parameters of the mechanism equation corresponding to each sample.

[0045] Specifically, the 17 input features of the input data obtained in step (2.2) are input into the neural network model. First, they enter the encoder and pass through the input layer, hidden layer and output layer of the encoder in sequence to obtain the encoded feature. Among them, the input dimension of the input layer is 17, the dimension of the hidden layer is 60, and the output dimension of the output layer is 32, that is, the domain feature and the health state feature both have a dimension of 16. The health state estimator, domain discriminator and mechanism parameter generator all include a single-layer basic network layer and a fully connected layer, and their hidden dimensions are all 32, only their specific parameters and dimensions are different. Among them, the input dimension of the health state estimator is 16 and the output dimension is 1; the input dimension of the domain discriminator is 16 and the output dimension is 2; the input dimension of the mechanism parameter generator is 32 and the output dimension is 2.

[0046] In this embodiment, the neural network model is trained using a training set. During the training process, with the goal of minimizing the total loss function of the neural network model, the parameters of the neural network model are adjusted to obtain a trained encoder and a health state estimator. Specifically, it includes: inputting the input data of the cyclic samples in the training set into the neural network model, and obtaining encoded features through the encoder; splitting the encoded features in half to obtain health state features and domain features; passing the health state features through the health state estimator to obtain the health state value of the loss output by the health state estimator; passing the domain features through the domain discriminator to obtain the working condition category output by the domain discriminator for the domain features; passing the health state features through the gradient reversal layer and then through the domain discriminator to obtain the working condition category output by the domain discriminator for the health state features; passing the encoded features through the mechanism parameter generator to obtain the parameters of the mechanism equation output by the mechanism parameter generator. Calculate the health label loss based on the health state value of the loss output by the health state estimator and its corresponding true health state value of the loss, where the true health state value of the loss is obtained by subtracting the corresponding health state label in the training set from 1; calculate the domain discrimination loss based on the working condition category output by the domain discriminator for the domain features, the working condition category output by the domain discriminator for the health state features, and the corresponding domain label in the training set. The adversarial training and feature alignment of the neural network model are achieved through the gradient reversal layer; calculate the parameter generation loss based on the parameters of the mechanism equation output by the mechanism parameter generator and the corresponding mechanism parameter label in the training set; calculate the physical equation loss based on the health state value of the loss output by the health state estimator, the parameters of the mechanism equation output by the mechanism parameter generator, and the corresponding cycle number and environmental temperature in the training set; calculate the continuous cycle physical constraint loss based on the health state value of the loss output by the health state estimator and the parameters of the mechanism equation output by the mechanism parameter generator; calculate the total loss function of the neural network model based on the health label loss, domain discrimination loss, parameter generation loss, physical equation loss, and continuous cycle physical constraint loss. With the goal of minimizing the total loss function of the neural network model, use the gradient descent method for iterative optimization to adjust the parameters of the encoder, health state estimator, domain discriminator, and mechanism parameter generator until the preset number of training rounds is reached to obtain a trained encoder and a health state estimator. The trained encoder and health state estimator can be constructed into a health state estimation model finally used for any working condition.

[0047] Furthermore, the calculation formula for the total loss function of the neural network model is:

[0048] In the formula, represents the total loss function of the neural network model, represents the health label loss, represents the domain discrimination loss, Represents the parameter generation loss, Represents the physical equation loss, Represents the continuous cycle physical constraint loss, And Represents the healthy label loss And the domain discrimination loss The corresponding fixed weight parameter, Is a dynamic weight parameter, and its dynamic change formula is:

[0049] In the formula, p is a dynamic parameter that linearly increases from 0 to 1 with the number of training rounds.

[0050] Furthermore, the calculation formula of the healthy label loss is:

[0051] In the formula, Is the healthy label loss, n represents the total number of working condition categories, Represents the total number of energy storage batteries under the i-th working condition category, Represents the total number of cycle samples of the j-th energy storage battery, Represents the healthy state value of the loss corresponding to the output of the healthy state estimator of the k-th cycle sample of the j-th energy storage battery under the i-th working condition category, Represents the true healthy state value of the loss corresponding to the k-th cycle sample of the j-th energy storage battery under the i-th working condition category, , Represents the healthy state label corresponding to the k-th cycle sample of the j-th energy storage battery under the i-th working condition category in the training set.

[0052] Furthermore, the calculation formula of the domain discrimination loss is:

[0053] In the formula, Is the domain discrimination loss, Represents the domain feature of the k-th cycle sample of the j-th energy storage battery under the i-th working condition category The working condition category output by the domain discriminator, Represents the healthy state feature of the k-th cycle sample of the j-th energy storage battery under the i-th working condition category The working condition category output by the domain discriminator, Represents the domain label corresponding to the k-th cycle sample of the j-th energy storage battery under the i-th working condition category in the training set.

[0054] Furthermore, the calculation formula of the parameter generation loss is:

[0055] In the formula, is the parameter generation loss, represents the parameter of the mechanism equation output by the mechanism parameter generator corresponding to the k-th cycle sample of the j-th energy storage battery under the i-th working condition category, represents the mechanism parameter label corresponding to the k-th cycle sample of the j-th energy storage battery under the i-th working condition category in the training set.

[0056] Furthermore, the calculation formula for the physical equation loss is:

[0057] In the formula, is the physical equation loss, represents the number of cycles of the k-th cycle sample of the j-th energy storage battery under the i-th working condition category, represents the ambient temperature of the working condition corresponding to the k-th cycle sample of the j-th energy storage battery under the i-th working condition category, which can be specifically obtained from the training set.

[0058] Furthermore, the calculation formula for the continuous cycle physical constraint loss is:

[0059] In the formula, is the continuous cycle physical constraint loss, represents the rectified linear unit function. RELU is used as an activation function in the neural network. Here, RELU is simply used as a part of the loss function.

[0060] In this embodiment, the Adam optimizer is used to optimize the parameters of the neural network model. The parameter settings of the Adam optimizer are: the exponential decay rate of the first moment estimate is 0.9, and the exponential decay rate of the second moment estimate is 0.999; the learning rate is set to 1e -3 , the number of training rounds is set to 200 rounds, and the size of each batch is 512. and take the values of 100 and 1 respectively.

[0061] (4)During online application, obtain the charge and discharge cycle data of the energy storage battery under operating conditions, including the operating current, voltage, sampling time, and number of cycles; the online operating conditions are different from all the source domain conditions in the training; extract the health features based on the online operating current, voltage, and sampling time, normalize them, and then combine the number of cycles as input data. Input the input data into the trained encoder to obtain the encoded features. After splitting the encoded features in half, obtain the health state features and domain features. Input the health state features into the trained health state estimator to obtain the current loss of the health state value, and then obtain the current health state estimation value of the energy storage battery. Among them, since the output of the health state estimator is u = 1 - y, subtracting the current loss of the health state value from 1 can obtain the current health state estimation value y of the energy storage battery, which represents the health state of the energy storage battery. Usually, the health state of the energy storage battery will be evaluated according to different scenarios. For example, in some scenarios, when y is greater than 80% or 70%, the energy storage battery is considered healthy.

[0062] In summary, the present invention estimates the domain-generalized health state of the energy storage battery by adopting a multi-mechanism-constrained adversarial decoupling representation method. Among them, the multi-mechanism constraints include two constraints: the physical equation loss and the continuous cycle physical constraint loss; the adversarial decoupling representation means splitting the encoded features into two features and realizing the decoupling representation through adversarial training using a gradient reversal layer. When estimating the domain-generalized health state of the energy storage battery, first, through the fusion application of adversarial decoupling representation learning and mechanism equations in the neural network, decouple and represent the input features as health state-related features and domain features related to the operating conditions; second, use the mechanism parameter generator as an auxiliary task to generate mechanism parameters to solve the problem of different mechanism equation parameters under different operating conditions; finally, through continuous cycle physical constraints, constrain the health state to decrease with the increase in the number of cycles and solve the variation problem of the mechanism equation parameters, thereby realizing the generalization estimation of the health state of unseen operating conditions. Compared with the existing battery health estimation methods, the present invention first integrates multi-mechanism constraints into the decoupling representation learning of energy storage batteries, and realizes the generalization modeling of unseen operating conditions without the need for target operating condition data, thereby improving the accuracy of the health state estimation of the model under unseen operating conditions and providing necessary support for the safe operation and maintenance of energy storage batteries.

[0063] In this embodiment, an open-source dataset is selected for experimental verification to verify the effectiveness of the method described in the present invention. The batteries in this open-source dataset are commercially available 18650 NCA (nickel cobalt aluminum) ternary lithium batteries, with a cut-off voltage of 2.65 - 4.2V and a rated capacity of 3.5Ah. Set different ambient temperatures and charging rates, with a total of five operating conditions, and the discharge rate is 1C for all of them. The specific data is shown in Table 1. Select operating condition 4 as the target domain and other operating conditions as the source domains. The results of the experiment using the method described in the present invention are as Figure 3As shown, the abscissa is the true value of the battery's state of health, and the ordinate is the predicted value of the battery's state of health. It can be seen that the predicted value of the battery's state of health is not much different from the true value, which proves the effectiveness of the method described in the present invention.

[0064] Table 1: Detailed Table of Set Data for Batteries under Different Operating Conditions

[0065] To more clearly demonstrate the superiority of the method described in the present invention in estimating the state of health under unseen operating conditions, the deep neural network (DNN) and the domain generalization paradigm of the deep adversarial network (DANN) were also used for comparison with the method described in the present invention. The experimental results are shown in Table 2. In the experiment, the evaluation criteria selected were the root mean square error (RMSE) and the mean absolute percentage error (MAPE). The encoder and the structure of the state-of-health estimator of the two comparison methods were the same as those of the method described in the present invention, and the other training parameters were also the same.

[0066] Table 2: State-of-Health Estimation Errors of Different Methods

[0067] As can be seen from Table 2, the method described in the present invention was compared with the other two methods, which proves the effectiveness of the method described in the present invention. When compared with the DNN method without domain generalization, the performance of the DANN method without the fusion of mechanism equations decreased. The method described in the present invention did not show a performance decrease and achieved the best results in both metrics.

[0068] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for characterizing and estimating the health of energy storage batteries with multiple mechanism constraints, characterized in that: The following steps are involved: (1) Obtain the complete life cycle operation data of multiple energy storage batteries under different test conditions, including the current, voltage, sampling time, number of cycles and corresponding cycle capacity data of each charge and discharge cycle, to construct a multi-source domain dataset; (2) Preprocess the complete life-cycle operation data of each charge-discharge cycle in each source domain dataset to extract health features and the health status label of each charge-discharge cycle, and combine the health features with the number of cycles as input data; at the same time, initialize the parameters of the mechanism equation of each source domain and save them as mechanism parameter labels; construct a training set based on the input data and its corresponding health status labels, domain labels, and mechanism parameter labels; (3) Constructing a neural network model, including an encoder, a health state estimator, a domain discriminator, and a mechanism parameter generator; and using the training set to train the neural network model. During the training process, the parameters of the neural network model are adjusted with the goal of minimizing the total loss function of the neural network model to obtain the trained encoder and health state estimator; (4) When used online, obtain the charge and discharge cycle data of the energy storage battery under operating conditions, including operating current, voltage, sampling time and number of cycles; The health features are extracted based on the online current, voltage and sampling time, and are normalized and combined with the number of cycles as input data. The data is input into the trained encoder to obtain the encoding features, which are split in half to obtain the health status features and domain features. The health status features are input into the trained health status estimator to obtain the health status value of the current loss, and then the current health status estimation value of the energy storage battery is obtained.

2. The method for characterizing and estimating the resistance decoupling of energy storage batteries and health under multiple mechanism constraints according to claim 1 is characterized in that: The step (1) specifically includes: The current, voltage, sampling time, number of cycles and cycle capacity data of the standard charging period of each charge-discharge cycle of multiple energy storage batteries from the time of testing to the end of their life cycle under different test conditions are obtained; the charging rate and the ambient temperature set in the test are used as the basis for dividing different test conditions, and the collected current, voltage, sampling time, number of cycles and cycle capacity data are divided into different source domains according to different charging rates and ambient temperatures, and the charging rate and ambient temperature are combined as domain labels to construct a multi-source domain dataset.

3. The method for characterizing and estimating the resistance decoupling of energy storage batteries and health under multiple mechanism constraints according to claim 1 is characterized in that: The step (2) specifically includes the following sub-steps: (2.1) For the complete life-cycle operation data of each charge-discharge cycle in each source domain data set, extract health features based on the current, voltage and sampling time of each charge-discharge cycle; wherein the health features include the mean value, standard deviation, kurtosis, skewness, curve slope and curve entropy of the voltage and current during the cyclic charging period, as well as the time and accumulated charge of the two stages of constant current charging and constant voltage charging, a total of 16 health features; (2.2) For the health features extracted in step (2.1), first use the 3-σ criterion to eliminate outliers in the health features, then use the normalization method to normalize the health features to obtain the normalized health features, and combine the number of cycles as input data; (2.3) Divide the cycle capacity data of each charge and discharge cycle by the rated capacity of the corresponding energy storage battery to calculate the health status label of each charge and discharge cycle; (2.4) Using the particle swarm optimization algorithm, the parameters of the mechanism equation of each source domain are initialized and saved as mechanism parameter labels; (2.5) Construct a training set based on the input data and its corresponding health state labels, domain labels, and mechanism parameter labels.

4. The method for characterizing and estimating the resistance decoupling of energy storage batteries and health under multiple mechanism constraints according to claim 1 or 3, characterized in that: The mechanism equation of the source domain is derived from the Arrhenius formula, and its differential form is: ； In the formula, u represents the health status value of the loss, u=1-y, y represents the health status label; t represents the current cycle number; e is a natural constant; T represents the ambient temperature set for the test; are two parameters of the mechanism equation; is the differential symbol.

5. The method for characterizing and estimating the resistance decoupling of energy storage batteries and health under multiple mechanism constraints according to claim 1 is characterized in that: In the neural network model, the encoder includes an input layer, a hidden layer and an output layer, the input layer includes a fully connected layer and an activation function, the hidden layer is a single-layer basic network layer, and the output layer is a fully connected layer; the health state estimator, the domain discriminator and the mechanism parameter generator all include a single-layer basic network layer and a fully connected layer; wherein the single-layer basic network layer includes a fully connected layer, an activation function and a random loss layer, and the activation functions are all sine functions; The encoder is used to perform feature encoding on the input data input to the neural network model to obtain the encoded features; the encoded features are then split in half to obtain two features, the former feature is used as the health state feature, and the latter feature is used as the domain feature; the health state estimator is used to map the health state feature to the health state value of the loss; the domain discriminator is used to discriminate the working condition category of the input domain feature; the encoded features are simultaneously input to the mechanism parameter generator to generate the parameters of the mechanism equation corresponding to each sample.

6. The method for characterizing and estimating the resistance decoupling of energy storage batteries and health under multiple mechanism constraints according to claim 1 is characterized in that: The neural network model is trained using the training set. During the training process, the parameters of the neural network model are adjusted with minimizing the total loss function of the neural network model as the optimization goal to obtain the trained encoder and health state estimator, which specifically includes: Input the input data of the cycle samples in the training set into the neural network model, and obtain the encoding features through the encoder; the encoding features are split in half to obtain the health state features and domain features; the health state features pass through the health state estimator to obtain the lost health state value output by the health state estimator; the domain features pass through the domain discriminator to obtain the working condition category output by the domain discriminator; the health state features pass through the gradient reversal layer and the domain discriminator in turn to obtain the working condition category output by the domain discriminator; the encoding features pass through the mechanism parameter generator to obtain the parameters of the mechanism equation output by the mechanism parameter generator; The health label loss is calculated based on the health state value of the loss output by the health state estimator and the corresponding real health state value of the loss, wherein the real health state value of the loss is obtained by subtracting the corresponding health state label in the training set from 1; the domain discrimination loss is calculated based on the operating condition category output by the domain discriminator through the domain feature, the operating condition category output by the domain discriminator through the health state feature and the corresponding domain label in the training set; the parameter generation loss is calculated based on the parameters of the mechanism equation output by the mechanism parameter generator and the corresponding mechanism parameter label in the training set; the physical equation loss is calculated based on the health state value of the loss output by the health state estimator, the parameters of the mechanism equation output by the mechanism parameter generator and the corresponding number of cycles and ambient temperature in the training set; the continuous cycle physical constraint loss is calculated based on the health state value of the loss output by the health state estimator and the parameters of the mechanism equation output by the mechanism parameter generator; the total loss function of the neural network model is calculated based on the health label loss, domain discrimination loss, parameter generation loss, physical equation loss and continuous cycle physical constraint loss; Taking minimizing the total loss function of the neural network model as the optimization goal, the gradient descent method is used for iterative optimization to adjust the parameters of the encoder, health state estimator, domain discriminator and mechanism parameter generator until the preset training rounds are reached to obtain the trained encoder and health state estimator.

7. The method for characterizing and estimating the resistance decoupling of energy storage batteries and health under multiple mechanism constraints according to claim 6 is characterized in that: The calculation formula of the total loss function of the neural network model is: ； In the formula, represents the total loss function of the neural network model, represents the loss of health labels, represents the domain discrimination loss, represents the parameter generation loss, represents the physical equation loss, represents the continuous cycle physical constraint loss, and Represent the health label loss respectively and domain discrimination loss The corresponding fixed weight parameters, is a dynamic weight parameter, and its dynamic change formula is: ； Where p is a dynamic parameter that grows linearly from 0 to 1 with the number of training rounds; The calculation formula of the health label loss is: ； In the formula, n represents the total number of working condition categories, represents the total number of energy storage batteries under the i-th operating condition category, represents the total number of cycle samples of the jth energy storage battery, It represents the lost health state value output by the health state estimator corresponding to the k-th cycle sample of the j-th energy storage battery under the i-th operating condition category, Indicates the actual loss of health status value corresponding to the k-th cycle sample of the j-th energy storage battery under the i-th operating condition category; The calculation formula of the domain discrimination loss is: ； In the formula, Represents the domain characteristics of the kth cycle sample of the jth energy storage battery under the ith operating condition category The working condition category output by the domain discriminator, Represents the health status characteristics of the kth cycle sample of the jth energy storage battery under the i-th operating condition category The working condition category output by the domain discriminator, Indicates the domain label corresponding to the k-th cycle sample of the j-th energy storage battery under the i-th operating condition category in the training set; The calculation formula for the parameter generation loss is: ； In the formula, Represents the parameters of the mechanism equation output by the mechanism parameter generator corresponding to the k-th cycle sample of the j-th energy storage battery under the i-th operating condition category, Indicates the mechanism parameter label corresponding to the k-th cycle sample of the j-th energy storage battery under the i-th operating condition category in the training set; The calculation formula of the physical equation loss is: ； In the formula, represents the number of cycles of the kth cycle sample of the jth energy storage battery under the i-th operating condition category, Indicates the ambient temperature of the kth cycle sample of the jth energy storage battery under the ith operating condition category; The calculation formula of the continuous cycle physical constraint loss is: ； In the formula, is the continuous cycle physical constraint loss, represents the rectified linear unit function.

Images