A multi-task learning battery state of health estimation method, system and computer device

Abstract

The application provides a multi-task learning battery health state estimation method, system and computer equipment, relates to the battery health estimation field, and comprises the following steps: obtaining a battery dataset covering different types of batteries and charge-discharge protocols, and performing feature extraction on the battery dataset to obtain a feature dataset; specifically, the feature dataset is input into a shared layer, a feature extraction encoder of the shared layer converts the feature dataset into shared features, a multi-head attention mechanism is used to convert the shared feature representation into task-specific features, and the task-specific features are used to complete a battery health degree estimation task and a prediction confidence evaluation task. The multi-task learning battery health state estimation method, system and computer equipment can accurately predict the health degree of different types of batteries and batteries with different charge-discharge protocols, provide a confidence reference for predicting unobserved feature data, and improve the prediction reliability of SOH.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of battery health estimation, in particular to a multi-task learning battery health state estimation method and system and computer equipment. BACKGROUND

[0002] With the rapid development of electric vehicles and renewable energy systems, the performance requirements of battery management systems are becoming increasingly stringent. The performance of the battery management system directly affects the vehicle's endurance, charging efficiency and safety. Therefore, accurately estimating the state of health (SOH) of the battery is of great significance to improve the performance of the battery management system.

[0003] Traditional battery health state estimation methods mainly rely on a single measurement index, such as voltage, capacity and internal resistance, etc. This method ignores the comprehensive influence of multiple factors on battery health and has high requirements for equipment precision, and the estimation accuracy is not high; the partial differential equation of physical constraint is also used to construct an equivalent circuit model to evaluate the state of health of lithium batteries. This method is simple and practical, and has low complexity, but due to the influence of many uncertain factors, the reliability and accuracy of the model need to be improved.

[0004] With the continuous development of artificial intelligence, some data-driven algorithms have also been applied in battery health state evaluation. Multi-task learning (MTL) is a new machine learning technology that shares knowledge between different tasks and improves the performance of the model on each task. Therefore, the application of multi-task learning in battery health state evaluation is feasible. SUMMARY

[0005] The purpose of the present application is to provide a multi-task learning battery health state estimation method, system and computer equipment, which uses the data sharing capability between tasks through a multi-task learning framework, reduces the number of training iterations and computing power cost, improves the prediction accuracy and transfer learning ability; through the battery health degree estimation task of multi-task learning, the battery degradation dynamic process is simulated, the SOH of the battery is estimated, and the battery health degree prediction of different types and charging and discharging protocols is realized; through the prediction confidence evaluation task, the deep auto-encoding Gaussian mixture model is used to identify the feature data anomaly degree when estimating the battery health degree, and to provide a confidence reference for the prediction of unseen feature data, thereby improving the prediction reliability of SOH.

[0006] To achieve the above purpose, the present application provides a multi-task learning battery health state estimation method, comprising the following steps:

[0007] The battery dataset covering different types of batteries and charge-discharge protocols is obtained, and the battery dataset is input into a deep neural network for feature extraction to obtain a feature dataset, and the feature dataset is input into a multi-task learning model to complete the battery health estimation task and the prediction confidence evaluation task, the multi-task learning model includes a shared layer, a physical information neural network and a deep auto-encoding Gaussian mixture model, and specifically includes

[0008] The feature dataset is input into the shared layer, and the feature extraction encoder of the shared layer converts the feature dataset into shared features, and the shared feature representation is converted into task-specific features using a multi-head attention mechanism, and the battery health estimation task and the prediction confidence evaluation task are completed using the task-specific features.

[0009] In the battery health estimation task, the physical neural network is used to simulate the battery degradation dynamic process to estimate the battery SOH; in the prediction confidence evaluation task, the deep auto-encoding Gaussian mixture model is used to output the confidence of SOH prediction.

[0010] Preferably, the feature dataset package includes battery health features:

[0011] The battery cycle number and the average value, standard deviation, kurtosis, skewness, charging time, cumulative charge, curve slope and curve entropy extracted from the current and voltage curve.

[0012] Preferably, the feature extraction encoder of the shared layer converts the feature dataset into shared features, and the feature extraction function is:

[0013] z=f s (p;θ s )

[0014] Where: z represents the shared representation; p represents the data features in the battery dataset; θ s represents the structure parameters of the multi-layer perceptron.

[0015] Preferably, the shared feature representation is converted into task-specific features through a multi-head attention mechanism:

[0016] The multi-head attention mechanism includes four heads, wherein the first head and the second head focus on the voltage curve segment and the current response mode to capture the electrochemical characteristics; the third head processes the cycle time dependence; and the fourth head focuses on the long-term degradation trend through the history capacity change.

[0017] Preferably, in the battery health estimation task, the physical neural network is used to simulate the battery degradation dynamic process to estimate the battery SOH, and the physical information neural network includes a feature-SOH prediction network and a battery degradation dynamic simulation network, including the following steps:

[0018] (1) The feature-SOH prediction network uses a multi-layer neural network structure to compress and process task-specific features to capture periodic patterns and monotonous degradation trends in voltage-current features, thereby achieving SOH mapping estimation.

[0019] (2) The degradation dynamic simulation network simulates the degradation rate of SOH by using a neural network with the same structure and adding physical constraints to the structural state space equation, thereby constraining the accuracy of the mapping estimation of SOH.

[0020] Preferably, in the prediction confidence assessment task, a deep autoencoder Gaussian mixture model is used to output the confidence level of the SOH prediction. The deep autoencoder Gaussian mixture model includes a compression network and an estimation network, and includes the following steps:

[0021] The compressed network reduces the dimensionality of the input task-specific features through a deep autoencoder, reconstructs the original data using a decoder, and extracts a low-dimensional latent representation of the task-specific features from the reduced dimensionality space and reconstruction error features.

[0022] The network receives a low-dimensional latent representation. A Gaussian mixture model is used to model the density distribution of the low-dimensional latent representation. Density estimation and energy calculation are performed, and the energy is used to measure the confidence of the SOH prediction.

[0023] Preferably, the compressed network reduces the dimensionality of the input task-specific features using a deep autoencoder, reconstructs the original data using a decoder, and extracts a low-dimensional latent representation of the task-specific features from the reduced dimensionality space and reconstruction error features, including:

[0024] By using deep autoencoders to gradually reduce the attention weighting of features through multiple nonlinear transformations, dimensionality reduction of task-specific features can be achieved.

[0025] The decoder reconstructs the original input space through a symmetric extension layer, thereby reconstructing the original feature data.

[0026] The reconstruction error features are obtained by calculating the relative reconstruction error and feature similarity index, and together with the low-dimensional latent representation in the dimensionality reduction space, they constitute the input estimation network of the low-dimensional latent representation of the complete task-specific features.

[0027] Preferably, the low-dimensional latent representation received by the network is estimated, and the density distribution of the low-dimensional latent representation is modeled using a Gaussian mixture model. Density estimation and energy calculation are then performed, including:

[0028] Density estimation: Estimating the low-dimensional latent representation output z received by the network from the compressed network. new The density distribution of γ is modeled using a Gaussian mixture model with K components;

[0029] γ = softmax(MLP(z) new ));

[0030] wherein Softmax(·) represents a normalization function;

[0031] The mixing parameter is calculated according to the soft assignment γ of the estimated network output of each low-dimensional latent representation;

[0032] Energy calculation: the sample energy value is calculated according to the likelihood function constructed according to the mixing parameter, to measure the confidence of the prediction.

[0033] A multi-task learning battery state of health estimation system comprises:

[0034] A data set acquisition module is configured to acquire a battery data set covering different types of batteries and charge-discharge protocols, and input the battery data set into a deep neural network for feature extraction to obtain a feature data set.

[0035] A battery state of health estimation module is configured to input the feature data set into a multi-task learning model to complete a battery state of health estimation task and a prediction confidence evaluation task.

[0036] A computer device comprises a memory and a processor; the memory stores a computer program, and the processor implements the steps of the method when executing the computer program.

[0037] Therefore, the multi-task learning battery state of health estimation method, system and computer device provided by the present application can reduce the number of training iterations and the cost of computing power by using the data sharing capability between tasks in a multi-task learning framework, improve the prediction accuracy and transfer learning ability, simulate the battery degradation dynamic process through the battery state of health estimation task of multi-task learning, estimate the battery SOH, have good transfer learning ability and generalization, and realize the battery state of health prediction of different types and charge-discharge protocols. Through the prediction confidence evaluation task, the deep auto-encoding Gaussian mixture model is used to identify the feature data abnormality when estimating the battery state of health, provide a confidence reference for the prediction of unseen feature data, and improve the prediction reliability of the battery state of health (SOH). BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The flowchart of the multi-task learning battery state of health estimation method of the present application. DETAILED DESCRIPTION

[0039] The technical solutions of the present application are further described below by means of the accompanying drawings and examples.

[0040] Unless otherwise defined, the technical terms or scientific terms used in the present application shall have the usual meanings understood by those skilled in the art to which the present application belongs.

[0041] Example 1

[0042] As Figure 1 shown, a multi-task learning battery state of health estimation method, comprising the following steps:

[0043] Step one, obtain a battery dataset covering different types of batteries and charge-discharge protocols and input the battery dataset into a deep neural network for feature extraction to obtain a feature dataset;

[0044] The feature dataset package includes battery health features:

[0045] Battery cycle number and average, standard deviation, kurtosis, skewness, charge time, cumulative charge, curve slope and curve entropy extracted from current and voltage curves.

[0046] Step two, input the feature dataset into the multi-task learning model to complete the battery health degree estimation task and the prediction confidence evaluation task, the multi-task learning model includes a shared layer, a physical information neural network and a deep auto-encoding Gaussian mixture model, specifically including

[0047] The feature dataset is input into the shared layer, and the feature extraction encoder of the shared layer converts the feature dataset into shared features, uses a multi-head attention mechanism to convert the shared feature representation into task-specific features, and uses the task-specific features to complete the battery health degree estimation task and the prediction confidence evaluation task;

[0048] The feature extraction encoder of the shared layer converts the feature dataset into shared features, and the feature extraction function is:

[0049] z=f s (p;θ s )

[0050] Where: z represents the shared representation; p represents the data features in the battery dataset; θ s represents the structure parameters of the multi-layer perceptron MLP, and p is 17-dimensional and z is 32-dimensional.

[0051] The shared feature representation is converted into task-specific features through a multi-head attention mechanism:

[0052] The multi-head attention mechanism includes four heads, wherein the first head and the second head focus on the voltage curve segment and the current response mode to capture the electrochemical characteristics; the third head processes the cycle time dependence; and the fourth head focuses on the long-term degradation trend through the history capacity change.

[0053] In the battery health degree estimation task, the physical neural network is used to simulate the battery degradation dynamic process to estimate the battery SOH, including the following steps:

[0054] (1) Feature-SOH prediction network is used to compress task-specific features through a multi-layer neural network structure to capture periodic patterns and monotonic degradation trends in voltage-current characteristics, and realize the mapping estimation of SOH.

[0055] (2) Degradation dynamic simulation network uses a neural network with the same structure to simulate the SOH degradation rate by adding physical constraints to the structure state space equation, and constrain the accuracy of the mapping estimation of SOH.

[0056] The modeling function of the feature-SOH mapping network can be represented as:

[0057] SOH = f(t, x);

[0058] Where t represents the cycle number of the battery; x represents the task-specific features of the estimation task;

[0059] The modeling function of the battery degradation dynamic simulation network can be represented as:

[0060]

[0061] Where u represents the SOH of the battery; g(.) is a nonlinear function describing the dynamic degradation of the battery; u t ,u x are the partial derivatives of u with respect to t and x, respectively;

[0062] In the prediction confidence evaluation task, the confidence of SOH prediction is output by the deep auto-encoding Gaussian mixture model, including the following steps:

[0063] The compression network uses a deep auto-encoder to reduce the dimensionality of the input task-specific features, and uses the decoder to reconstruct the original data. The low-dimensional latent representation of the task-specific features is extracted from the reduced dimensionality space and the reconstruction error feature, including:

[0064] The deep auto-encoder uses multiple nonlinear transformations to gradually reduce the attention weighting of the features, achieving dimensionality reduction of the task-specific features;

[0065] The decoder reconstructs the original input space through a symmetric expansion layer, achieving reconstruction of the original feature data;

[0066] The relative reconstruction error and the feature similarity index are calculated to obtain the reconstruction error feature, which together with the low-dimensional latent representation of the reduced dimensionality space forms the complete low-dimensional latent representation of the task-specific features input to the estimation network.

[0067] The estimation network receives the low-dimensional latent representation and uses a Gaussian mixture model to model the density distribution of the low-dimensional latent representation, performs density estimation and energy calculation, including:

[0068] Density estimation: the estimation network receives the low-dimensional latent representation output znew The density distribution of gamma is modeled using a mixture of K components of Gaussian models;

[0069] gamma = softmax(MLP(z new ));

[0070] where Softmax(·) denotes a normalization function.

[0071] The mixture parameters are computed from the estimated network output soft assignments gamma of each low-dimensional latent representation;

[0072]

[0073] where N denotes the number of mini-batches, phi k denotes the mixing weights, mu k , Sigma k denote the mean and covariance matrix of the k-th component, gamma i denotes the soft assignment of the i-th low-dimensional latent representation of gamma, denotes the i-th low-dimensional latent representation, and T denotes the transpose;

[0074] Energy computation: the sample energy value is computed from the likelihood function constructed from the mixture parameters to judge the fitting degree of the new sample to the learned density distribution, measure the confidence of SOH prediction, if the energy is high, the prediction deviates from the normal mode, the confidence is low,

[0075]

[0076] A multi-task learning battery state of health estimation system comprises:

[0077] A data set acquisition module is configured to acquire a battery data set covering different types of batteries and charge-discharge protocols, and input the battery data set into a deep neural network to extract features and obtain a feature data set.

[0078] A battery health estimation module is configured to input the feature data set into a multi-task learning model to complete a battery health estimation task and a prediction confidence evaluation task.

[0079] The terminal device provided by the embodiment includes a processor, a memory, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the steps in each of the above method embodiments. Alternatively, the processor executes the computer program to implement the functions of each module / unit in each of the above device embodiments.

[0080] The computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present application.

[0081] The terminal device can be a desktop computer, a notebook computer, a palm computer, a cloud server and the like. The terminal device can include, but is not limited to, a processor, a memory.

[0082] The processor can be a central processing unit (CPU), a graphics processing unit (GPU), and can also be other general-purpose processors and the like.

[0083] The memory can be used to store the computer program and / or modules, and the processor realizes various functions of the terminal device by running or executing the computer program and / or modules stored in the memory, and calling data stored in the memory.

[0084] The modules / units integrated in the terminal device, if realized in the form of software function units and sold or used as independent products, can be stored in a computer readable storage medium. Based on this understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. When the processor executes the computer program, the steps of the above-mentioned various method embodiments can be realized. The computer program includes computer program code, which can be in the form of source code, object code, executable file or some intermediate form, etc. The computer readable medium can include any entity or device, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. that can carry the computer program code. It should be noted that the content included in the computer readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction, for example, in some jurisdictions, according to legislation and patent practice, the computer readable medium does not include electric carrier signals and telecommunication signals.

[0085] Therefore, the application adopts the above-mentioned multi-task learning battery health state estimation method, system and computer device, reduces the training iteration number and the computing power cost by using the data sharing ability between tasks through the multi-task learning framework, and improves the prediction accuracy and the transfer learning ability; the battery health degree estimation task of multi-task learning simulates the battery degradation dynamic process and estimates the battery SOH, has good transfer learning ability and generalization, and realizes the battery health degree prediction of different types and charge-discharge protocols; through the prediction confidence evaluation task, the feature data anomaly degree is identified when the battery health degree is estimated by using the deep self-encoding Gaussian mixture model, the prediction confidence of the unobserved feature data is provided, and the prediction reliability of the SOH is improved.

[0086] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, but not to limit them, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: the technical solutions of the present application can still be modified or replaced by equivalents, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A multi-task learning method for estimating battery health status, characterized in that, Includes the following steps: A battery dataset covering different types of batteries and charging / discharging protocols is acquired and input into a deep neural network for feature extraction, resulting in a feature dataset. This feature dataset is then fed into a multi-task learning model to complete battery health estimation and prediction confidence assessment tasks. The multi-task learning model includes a shared layer, a physical information neural network, and a deep autoencoder Gaussian mixture model, specifically including... The feature dataset is input into the shared layer. The feature extraction encoder of the shared layer transforms the feature dataset into shared features. The shared feature representation is transformed into task-specific features using a multi-head attention mechanism. The task-specific features are used to complete the battery health estimation task and the prediction confidence assessment task. The task-specific features include features based on voltage curve segments, current response patterns, cycle time dependence, and long-term degradation trends. In the battery health estimation task, a physical neural network is used to simulate the dynamic process of battery degradation and estimate the battery SOH; in the prediction confidence assessment task, a deep autoencoder Gaussian mixture model is used to output the confidence of the SOH prediction. The feature dataset includes battery health features: Battery cycle count and average, standard deviation, kurtosis, skewness, charging time, cumulative charge, curve slope, and curve entropy extracted from the current and voltage curves; The shared-layer feature extraction encoder transforms the feature dataset into shared features. The feature extraction function is: in: z Representation shared; p Represents the data characteristics in the battery dataset; The structural parameters of a multilayer perceptron; The shared feature representation is transformed into task-specific features through a multi-head attention mechanism: The multi-head attention mechanism comprises four heads. The first and second heads focus on voltage curve segments and current response patterns to capture electrochemical properties; the third head addresses cycle time dependence; and the fourth head focuses on long-term degradation trends through historical capacity changes. In the battery health estimation task, a physical information neural network is used to simulate the dynamic process of battery degradation and estimate the battery's state of health (SOH). The physical information neural network includes a feature-SOH prediction network and a battery degradation dynamic simulation network, and includes the following steps: The feature-SOH prediction network uses a multi-layer neural network structure to compress task-specific features to capture periodic patterns and monotonically degrading trends in voltage-current features, thereby achieving SOH mapping estimation. The degradation dynamic simulation network simulates the degradation rate of SOH by using a neural network with the same structure and adding physical constraints to the structural state space equation, thereby improving the accuracy of the mapping estimation of the constrained SOH. In the prediction confidence assessment task, a deep autoencoder Gaussian mixture model is used to output the confidence of the SOH prediction. The deep autoencoder Gaussian mixture model includes a compression network and an estimation network, and includes the following steps: The compressed network reduces the dimensionality of the input task-specific features through a deep autoencoder, reconstructs the original data using a decoder, and extracts a low-dimensional latent representation of the task-specific features from the reduced dimensionality space and reconstruction error features. The network receives a low-dimensional latent representation, and a Gaussian mixture model is used to model the density distribution of the low-dimensional latent representation. Density estimation and energy calculation are then performed, and the energy is used to measure the confidence of the SOH prediction. The compressed network reduces the dimensionality of the input task-specific features using a deep autoencoder, reconstructs the original data using a decoder, and extracts low-dimensional latent representations of the task-specific features from the reduced dimensionality space and reconstruction error features, including: By using deep autoencoders to gradually reduce the attention weighting of features through multiple nonlinear transformations, dimensionality reduction of task-specific features can be achieved. The decoder reconstructs the original input space through a symmetric extension layer, thereby reconstructing the original feature data. The reconstruction error features are obtained by calculating the relative reconstruction error and feature similarity index, and together with the low-dimensional latent representation in the dimensionality reduction space, they constitute the input estimation network of the low-dimensional latent representation of the complete task-specific features.

2. The multi-task learning battery health state estimation method according to claim 1, characterized in that, The network receives a low-dimensional latent representation. A Gaussian mixture model is used to model the density distribution of this low-dimensional latent representation, and density estimation and energy calculation are performed, including: Density estimation: Estimating the low-dimensional latent representation output received by the network from the compressed network. ,use K Gaussian mixture model with individual components Model the density distribution; ； In the formula, Softmax (·) denotes the normalization function; Based on the soft assignment of each low-dimensional latent representation of the estimated network output. Calculate the mixture parameters; Energy calculation: The sample energy value is calculated by constructing a likelihood function based on the mixture parameters, which is used to measure the confidence of the prediction.

3. A multi-task learning battery health state estimation system, characterized in that, A method for performing a multi-task learning battery health state estimation method as described in claim 1 includes: The dataset acquisition module is used to acquire battery datasets covering different types of batteries and charging / discharging protocols, and input the battery datasets into a deep neural network for feature extraction to obtain a feature dataset; The battery health estimation module is used to input the feature dataset into the multi-task learning model to complete the battery health estimation task and the prediction confidence assessment task.

4. A computer device, characterized in that, It includes a memory and a processor; the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the method as described in any one of claims 1-2.

Images