A method for SOH estimation and RUL prediction of lithium-ion batteries based on TTSNet

Abstract

The present invention discloses a method for SOH estimation and RUL prediction of lithium-ion batteries based on TTSNet, comprising the following steps: S1, constructing a TTSNet network model; S2, extracting features of the voltage, current and temperature data in the preprocessed battery data through the corresponding extraction branches through the feature extraction network of the time series transformer, and obtaining rich semantic information features; S3, adaptively fusing the extracted semantic information features through the feature fusion network of AGFF, and obtaining the prediction output of the neural network; S4, SOH estimation and RUL prediction based on the TTSNet network model. This method can not only adaptively extract and interact with useful information in the input time series of lithium-ion batteries, but also consider the correlation between the voltage, current, temperature and capacity of the long time series, pay attention to the time series dimension information related to the previous and next steps, and improve the ability of the network model to study the long-term dependency of the time series. Finally, real-time and high-precision SOH estimation and RUL prediction results are obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of power battery management, and in particular to a method for SOH estimation and RUL prediction of lithium-ion batteries based on TTSNet. Background Art

[0002] Existing research on SOH estimation and RUL prediction can be divided into two categories: model-based methods and data-driven methods. Model-based methods reflect the internal degradation process of the battery by establishing a physical model. However, due to the complexity of the internal reaction of the battery and the great influence of external conditions, it is difficult to establish an accurate model, and the calculation is complex and the practicality is weak. Data-driven methods ignore the internal mechanism and directly establish a mapping relationship with the battery capacity through historical data. They are divided into traditional machine learning methods and artificial neural network methods. Artificial neural network methods have attracted widespread attention from researchers due to their good generalization performance and strong feature extraction capabilities. Among them, recurrent neural networks are a type of neural network specifically used to process time series information. They are the most common neural network used to estimate SOH and RUL. However, its ability to process relatively long sequence information is limited, and it cannot fully process the previous and next correlation information.

[0003] Based on the defects of the above methods, the present invention proposes a new temporal transformer-based sequence network (TTSNet) model to solve the SOH estimation and RUL prediction problems. Summary of the invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and propose a lithium-ion battery SOH estimation and RUL prediction method based on TTSNet, which can not only adaptively extract and fuse useful information in the lithium-ion battery input sequence, but also consider the correlation between long-term series voltage, current, temperature and capacity, pay attention to the relevant information of the previous and next steps, and improve the ability of the network model to study long-term dependencies. Finally, real-time and high-precision SOH estimation and RUL prediction results are obtained.

[0005] In order to solve the above technical problems, the technical solution of the present invention is:

[0006] A method for estimating SOH and predicting RUL of lithium-ion batteries based on TTSNet, comprising the following steps:

[0007] S1, collecting and preprocessing battery data, and making a data set from the preprocessed battery data, wherein the battery data includes lithium-ion battery terminal voltage, current, temperature data and corresponding battery capacity data;

[0008] S2. Construct a TTSNet network model, which includes a feature extraction network of a timing transformer and a feature fusion network of AGFF. The feature extraction network of the timing transformer includes three specific data feature extraction branches, which are used to extract output features of voltage, current, and temperature data, respectively. Each of the data feature extraction branches consists of a feature embedding block and a VIT timing transformer. The feature fusion network of AGFF consists of three convolutional neural networks with different receptive fields and a spatial attention mechanism.

[0009] S3, through the feature extraction network of the time series transformer, feature extraction is performed on the voltage, current and temperature data in the preprocessed battery data through the corresponding extraction branches to obtain rich semantic information features;

[0010] S4, adaptively fuse the extracted semantic information features through the feature fusion network of AGFF to obtain the prediction output of the neural network;

[0011] S5. SOH estimation and RUL prediction based on TTSNet network model

[0012] For SOH estimation, select the voltage-time, current-time, and temperature-time in the dataset as the network input, and the corresponding capacity data as the input label to calculate the SOH prediction error:

[0013]

[0014] Among them, C i and C0 represent the battery capacity of the ith cycle and the initial capacity of the battery, respectively. For RUL prediction, the capacity values ​​from the initial capacity to the end of life are used as the capacity data set, and the capacity values ​​obtained by the SOH estimation model are input into the trained RUL prediction model to perform RUL multi-step prediction:

[0015] RUL=n EOL -n t

[0016] Among them, n EOL and n t They represent the charge and discharge cycle when the battery reaches the end of its life and the charge and discharge cycle at time t respectively.

[0017] Preferably, in step S1, the method for collecting battery data is: performing uninterrupted charge and discharge experiments on the lithium-ion battery to obtain the lithium-ion battery terminal voltage, current, temperature data and corresponding battery capacity data.

[0018] Preferably, the method for preprocessing the battery data collected in step S1 is:

[0019] According to the error range in the data acquisition process, adaptive Gaussian noise is introduced to achieve data enhancement operation. The enhanced time series data expression is:

[0020] X aug =X i +α gauss G(X max -X min )

[0021] Among them, X i is the original input sequence, i.e., voltage, current, and temperature timing information; α gauss is the adaptive Gaussian noise coefficient; X max and X min are the maximum and minimum values ​​of the input sequence respectively; G(·) is the introduced Gaussian white noise;

[0022] The data after data enhancement is normalized by the min-max method. The expression of the normalized time series data is:

[0023]

[0024] The input data is preprocessed using a sliding window, and the original data is X after data enhancement and normalization. aug&normal =(X1,X2,X3,...,X t ) T ,in is the pth feature of the sample at the tth moment, and the sliding window size is s. The data mapping relationship after sliding window processing is expressed as:

[0025]

[0026] Among them, n is the total number of features of the sample, Y n Represents the true value of the nth sample to be predicted.

[0027] Preferably, in step S1, the preprocessed battery data Divided into in and They represent three branch inputs of voltage, current and temperature respectively, and are made into a data set, which is divided into a training set, a validation set and a test set in a ratio of 5:1:1.

[0028] Preferably, in step S3, a method for extracting voltage data features of a branch is:

[0029] S3-1, first voltage data The initial features are obtained through a one-dimensional convolutional neural network with a convolution kernel size of 1 and a step size of 1:

[0030]

[0031] Where δ represents the activation layer;

[0032] S3-2. Initial features of voltage data through feature embedding block To perform feature embedding operation, the calculation steps of the feature embedding block are as follows:

[0033]

[0034] First, the features The information is effectively learned through the long short-term memory network, that is, the feature embedding operation, and the feature LSTM(·) represents the LSTM module, and the output of this layer is Where K and D represent the time window and the number of channels respectively;

[0035] S3-3. Output embedding data through temporal transformer Perform feature extraction.

[0036] Preferably, the step S3-3 includes the following sub-steps:

[0037] S3-3-1. First, the input dimension of the time series transformer is exchanged. This operation swaps the time dimension and the spatial dimension of the input feature, so that the next step VIT time series transformer processes the time series dimension information. The expression is as follows:

[0038]

[0039] S3-3-2, then pass the features through the multi-head self-attention module. Define the three scalars in the attention mechanism, namely Query (Q), Key (K) and Value (V), the expressions are as follows:

[0040]

[0041]

[0042] Among them, H i =Attention(Q i ,K i ,V i ), i represents the number of attention heads, d k represents the feature dimension, softmax(·) calculates the weight, and Concat(·) represents the element addition operation;

[0043] S3-3-3. The original features and the features processed by the attention module are summed and normalized at the same time. The expression is as follows:

[0044]

[0045] S3-3-4, then the features are processed by the feedforward neural network, and the expression is as follows:

[0046]

[0047] Where W ffn is the convolutional layer weight of the feedforward neural network, b ffn is bias;

[0048] S3-3-5. Finally, the features processed by the feedforward neural network are Compared with the features without feedforward neural network processing Perform summation and normalization operations, the expression is as follows:

[0049]

[0050] Initial features and After three parallel improved VIT temporal transformer branches, we get rich semantic information features X V , X I and X T .

[0051] Preferably, in step S2, the methods for extracting current data features and temperature data features of the other two extraction branches are the same as the method for extracting voltage data features.

[0052] Preferably, the step S4 includes the following sub-steps:

[0053] S4-1. First, sum the features of the three dimensions. The expression is as follows:

[0054]

[0055] S4-2, then the features are added after passing through three one-dimensional convolutional networks with increasing receptive fields. and They are different features obtained through the convolution layer. The convolution kernel sizes of the three convolutional neural networks are 1, 3, and 5 respectively; the expansion rates are 1, 3, and 9 respectively. The added feature X′ integrates the information of multiple receptive fields, and the expression is as follows:

[0056]

[0057] S4-3, the added features are passed through the global average pooling layer, and then the feature dimension reduction operation is realized through a simple fully connected layer. The expression is as follows:

[0058]

[0059] Among them, F fc (·) represents the global average pooling operation, and MLP(·) represents the fully connected layer. Finally, the three branches output different weights. and

[0060] S4-4. Finally, multiply the weight by the feature and add the three channel data to get the output. The expression is as follows:

[0061]

[0062] Out i Represents the output of the i-th dimension. This operation needs to be performed on the features of each dimension to finally obtain the predicted output of the neural network.

[0063] Preferably, the hyperparameters of the TTSNet network model constructed in step S2 are determined by a cross-validation method, specifically including sliding window size, learning rate, number of iterations and batch size.

[0064] The present invention has the following characteristics and beneficial effects:

[0065] By adopting the above technical solution, the present invention uses the timing transformer to fully extract the features of the voltage, current, temperature, and time series directly measured by the battery, and uses AGFF to adaptively fuse the three-dimensional information. Through the method based on the new TTSNet network, the disadvantages of the traditional network in capturing the long sequence information of the original input of the battery are avoided, and the information on the timing of the previous and next steps is paid attention to, so as to improve the ability of the network model to study the correlation of long time series information. The measurement process is simple, and the SOH and RUL of lithium-ion batteries can be estimated in a timely and accurate manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0067] Figure 1This is a framework diagram of a lithium-ion battery SOH estimation and RUL prediction method based on TTSNet in an embodiment of the present invention;

[0068] Figure 2 Schematic diagram of the TTSNet network structure in an embodiment of the present invention. DETAILED DESCRIPTION

[0069] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0070] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and the like are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, features defined as "first", "second", and the like may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0071] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood by specific circumstances.

[0072] The present invention provides a method for estimating SOH and predicting RUL of lithium-ion batteries based on TTSNet. Figure 1 As shown, the following steps are included:

[0073] Step (1): Battery data collection and preprocessing.

[0074] (a) Perform uninterrupted charge and discharge experiments on lithium-ion batteries to obtain the raw data required by the lithium-ion battery SOH and RUL joint estimation network model, including lithium-ion battery terminal voltage, current, temperature data and corresponding battery capacity data.

[0075] (b) According to the error range in the process of acquiring the data set, adaptive Gaussian noise is introduced to achieve data enhancement. The enhanced time series data expression is:

[0076] X aug =X i +α gauss G(X max -X min )

[0077] Among them, X i is the original input sequence, i.e., voltage, current, and temperature timing information; α gauss is the adaptive Gaussian noise coefficient; X max and X min are the maximum and minimum values ​​of the input sequence respectively; G(·) is the introduced Gaussian white noise. The data after data enhancement is normalized by the min-max method. The expression of the normalized time series data is:

[0078]

[0079] (c) Use sliding window to preprocess the input data. The original data is X after data enhancement and normalization. aug&normal =(X1,X2,X3,…,X t ) T ,in is the pth feature of the sample at the tth moment. Let the sliding window size be s. The data mapping relationship expression after sliding window processing is:

[0080]

[0081] Among them, n is the total number of features of the sample, Y n Represents the true value of the nth sample to be predicted.

[0082] Step (2) is based on the feature extraction structure of the improved temporal transformer, such as Figure 2 As shown in the figure, in the feature extraction stage, three branches are set for voltage, current, and temperature respectively, and each branch is dedicated to extracting features for specific data. Each branch consists of a feature embedding block and a VIT timing transformer. Specifically:

[0083] Raw data Divided into in and Represent the three branch inputs of voltage, current and temperature respectively, and the dimension of each branch data is the same. For example, we first obtain the initial features through a one-dimensional convolutional neural network with a convolution kernel size of 1 and a step size of 1:

[0084]

[0085] Where δ represents the activation layer.

[0086] The feature embedding block structure is constructed. For the voltage, current, and temperature branches, the designed feature embedding structure is the same, with the voltage branch input For example, the calculation steps of each feature embedding block are as follows:

[0087]

[0088] First, the features The information is effectively learned through the long short-term memory network, that is, the feature embedding operation, and the feature Where LSTM(·) represents the LSTM module. The output of this layer is Where K and D represent the time window and the number of channels respectively.

[0089] The temporal transformer structure is constructed, and the embedded data features are extracted through this module.

[0090] First, the input dimension of the time series transformer is swapped. This operation swaps the time dimension and the spatial dimension of the input feature, so that the next step VIT time series transformer processes the time series dimension information:

[0091]

[0092] The features are then passed through a multi-head self-attention module. Define the three scalars in the attention mechanism, namely Query (Q), Key (K) and Value (V):

[0093]

[0094]

[0095] Among them, H i =Attention(Q i ,K i ,V i ), i represents the number of attention heads, d k Represents feature dimension. Softmax(·) calculates weights, and Concat(·) represents element addition operation;

[0096] At the same time, the original features and the features processed by the attention module are summed and normalized:

[0097]

[0098] Then the features are processed by a feedforward neural network, where W ffn is the convolutional layer weight of the feedforward neural network, b ffn For bias:

[0099]

[0100] Finally, the features processed by the feedforward neural network are Compared with the features without feedforward neural network processing Perform summation and normalization operations.

[0101]

[0102] Initial features and After three parallel improved VIT temporal transformer branches, we get rich semantic information features X V , X I and X T .

[0103] Step (3): Based on the AGFF feature fusion structure, the extracted features are deeply adaptively fused. The neural network in the fusion stage consists of three convolutional neural networks with different receptive fields and an improved spatial attention mechanism. Specifically:

[0104]

[0105] First, the features of the three dimensions are summed.

[0106]

[0107] The features are then added together after passing through three one-dimensional convolutional networks with increasing receptive fields. and They are different features obtained through the convolution layer. The convolution kernel sizes of the three convolutional neural networks are 1, 3, and 5 respectively; the expansion rates are 1, 3, and 9 respectively. The added feature X′ integrates the information of multiple receptive fields.

[0108]

[0109] Taking one of the branches as an example, the added features are passed through the global average pooling layer, and then the feature dimension reduction operation is achieved through a simple fully connected layer. fc (·) represents the global average pooling operation, and MLP(·) represents the fully connected layer. Finally, the three branches output different weights. and

[0110]

[0111] Finally, the weight is multiplied by the feature, and the three channel data are added together to get the output. i Represents the output of the i-th dimension. This operation needs to be performed on the features of each dimension to finally obtain the predicted output of the neural network.

[0112] Step (4) SOH estimation and RUL prediction based on the TTSNet network model.

[0113] First, determine the hyperparameters of the TTSNet network model. The cross-validation method is used to determine the network hyperparameters, including sliding window size, learning rate, number of iterations, and batch size.

[0114] Then, the dataset is divided into training set, validation set and test set in a ratio of 5:1:1.

[0115] For SOH estimation, select voltage-time, current-time, and temperature-time as network inputs, and the corresponding capacity data as input labels to calculate the SOH prediction error:

[0116]

[0117] Among them, C i and C0 represent the battery capacity of the ith cycle and the initial capacity of the battery, respectively.

[0118] For RUL prediction, the capacity values ​​from initial capacity to end of life are used as capacity data sets, and the capacity values ​​obtained by the SOH estimation model are input into the trained RUL prediction model to perform RUL multi-step prediction:

[0119] RUL=n EOL -n t

[0120] Among them, n EOL and n t They represent the charge and discharge cycle when the battery reaches the end of its life and the charge and discharge cycle at time t respectively.

[0121] Step (5) The present invention conducts a comparative experiment on the SOH and RUL estimation effects of lithium-ion batteries on the public data set Oxford battery aging data set, wherein the comparison method adopts the currently widely used CNN-LSTM method, and the test results are shown in the following table:

[0122] Estimation Model MAE(%) RMSE(%) SOH Error(%) RUL Error(items) CNN-LSTM 0.204 0.399 2.2 -5 This method 0.032 0.071 0.3 1

[0123] Table 1

[0124] Among them, MAE and RMSE are mean absolute error and root mean square error loss functions, respectively, SOH Error is the error of online estimation of SOH of lithium battery, and RUL Error is the error of online prediction of RUL of lithium battery. It can be seen from Table 1 that the SOH estimation and RUL prediction method of lithium-ion battery based on TTSNet proposed in the present invention can well realize the SOH estimation and RUL prediction tasks of lithium battery, and can ensure high accuracy.

[0125] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments including components are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

Claims

1. A lithium-ion battery SOH estimation and RUL prediction method based on TTSNet, characterized in that: The steps include: S1. Construct a TTSNet network model, which includes a feature extraction network of a timing transformer and a feature fusion network of AGFF. The feature extraction network of the timing transformer includes three specific data feature extraction branches, which are used to extract output features of voltage, current, and temperature data, respectively. Each of the data feature extraction branches consists of a feature embedding block and a VIT timing transformer. The feature fusion network of AGFF consists of three convolutional neural networks with different receptive fields and a spatial attention mechanism. S2, through the feature extraction network of the time series transformer, feature extraction is performed on the pre-processed voltage, current and temperature data through the corresponding extraction branches to obtain rich semantic information features; S3, adaptively fuse the extracted semantic information features through the feature fusion network of AGFF to obtain the prediction output of the neural network; S4. SOH estimation and RUL prediction based on TTSNet network model For SOH estimation, select the voltage-time, current-time, and temperature-time in the dataset as the network input, and the corresponding capacity data as the input label to calculate the SOH prediction error: Among them, C i and C0 represent the battery capacity of the ith cycle and the initial capacity of the battery, respectively. For RUL prediction, the capacity values ​​from the initial capacity to the end of life are used as the capacity data set, and the capacity values ​​obtained by the SOH estimation model are input into the trained RUL prediction model to perform RUL multi-step prediction: RUL=n EOL -n t Among them, n EOL and n t They represent the charge and discharge cycle when the battery reaches the end of its life and the charge and discharge cycle at time t respectively.

2. According to claim 1, a lithium-ion battery SOH estimation and RUL prediction method based on TTSNet is characterized in that: In step S2, a method for extracting branch voltage data features is: S2-1, first voltage data The initial features are obtained through a one-dimensional convolutional neural network with a convolution kernel size of 1 and a step size of 1: Where δ represents the activation layer; S2-2. Initial features of voltage data through feature embedding block To perform feature embedding operation, the calculation steps of the feature embedding block are as follows: First, the features The information is effectively learned through the long short-term memory network, that is, the feature embedding operation, and the feature LSTM(·) represents the LSTM module, and the output of this layer is Where K and D represent the time window and the number of channels respectively; S2-3. Output embedding data through temporal transformer Perform feature extraction.

3. A lithium-ion battery SOH estimation and RUL prediction method based on TTSNet according to claim 2, characterized in that: The step S2-3 includes the following sub-steps: S2-3-1. First, the input dimension of the time series transformer is exchanged. This operation swaps the time dimension and the spatial dimension of the input feature, so that the next step VIT time series transformer processes the time series dimension information. The expression is as follows: S2-3-2, then pass the features through the multi-head self-attention module. Define the three scalars in the attention mechanism, namely Query (Q), Key (K) and Value (V), and the expressions are as follows: Among them, H i =Attention(Q i ,K i ,V i ), i represents the number of attention heads, d k represents the feature dimension, softmax(·) calculates the weight, and Concat(·) represents the element addition operation; S2-3-3. Sum and normalize the original features and the features processed by the attention module at the same time. The expression is as follows: S2-3-4, then the features are processed by the feedforward neural network, and the expression is as follows: Where W ffn is the convolutional layer weight of the feedforward neural network, b ffn is bias; S2-3-5. Finally, the features processed by the feedforward neural network are Compared with the features without feedforward neural network processing Perform summation and normalization operations, the expression is as follows: Initial features and After three parallel improved VIT temporal transformer branches, we get rich semantic information features X V , X I and X T .

4. A method for SOH estimation and RUL prediction of lithium-ion batteries based on TTSNet according to claim 3, characterized in that: In step S1, the methods for extracting current data features and temperature data features of the other two extraction branches are the same as the method for extracting voltage data features.

5. The method for SOH estimation and RUL prediction of lithium-ion batteries based on TTSNet according to claim 4, characterized in that: The step S3 includes the following sub-steps: S3-1. First, sum the features of the three dimensions. The expression is as follows: S3-2, then the features are added after passing through three one-dimensional convolutional networks with increasing receptive fields. and They are different features obtained through the convolution layer. The convolution kernel sizes of the three convolutional neural networks are 1, 3, and 5 respectively; the expansion rates are 1, 3, and 9 respectively. The added feature X′ integrates the information of multiple receptive fields, and the expression is as follows: S3-3, the added features are passed through the global average pooling layer, and then the feature dimension reduction operation is realized through a simple fully connected layer. The expression is as follows: Among them, F fc (·) represents the global average pooling operation, and MLP(·) represents the fully connected layer. Finally, the three branches output different weights. and S3-4. Finally, multiply the weight by the feature and add the three channel data to get the output. The expression is as follows: Out i Represents the output of the i-th dimension. This operation needs to be performed on the features of each dimension to finally obtain the predicted output of the neural network.

6. A method for estimating SOH and predicting RUL of lithium-ion batteries based on TTSNet according to any one of claims 1 to 5, characterized in that: The hyperparameters of the TTSNet network model constructed in step S1 are determined by a cross-validation method, specifically including sliding window size, learning rate, number of iterations and batch size.

Images