A lithium battery state of health estimation method fusing unsupervised and supervised learning

Abstract

The application provides a lithium battery health state estimation method fusing unsupervised learning and supervised learning in the technical field of lithium battery detection, and comprises the following steps: S10, a large amount of charging data of a lithium battery is acquired, and a training set and a test set are constructed based on the charging data; S20, the training set and the test set are preprocessed by normalization and length unification; S30, a first estimation model fusing unsupervised learning and supervised learning is created based on an encoding module, a decoding module and an SOH prediction module; S40, the first estimation model is trained and tested by using the preprocessed training set and test set; S50, the encoding module and the SOH prediction module are extracted from the trained first estimation model, and a second estimation model is created based on the extracted encoding module and SOH prediction module; and S60, the second estimation model is used for lithium battery health state estimation. The application has the advantages that the accuracy and the generalization of the lithium battery health state estimation are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium battery detection, and in particular to a lithium battery health state estimation method integrating unsupervised and supervised learning. Background Art

[0002] Lithium batteries are a widely used energy tool due to their high energy density, low self-discharge rate, low price, and long service life. However, as usage increases, their actual available capacity decreases and their remaining lifespan decreases. Therefore, estimating the current state of health (SOH) of lithium batteries is particularly important. In addition, after lithium batteries are put into use, it is difficult to perform long-term charge and discharge cycle tests like in the laboratory. Moreover, the operating conditions of lithium batteries in online use are complex and changeable, which has led to the need for online SOH estimation of lithium batteries.

[0003] When estimating the health status of lithium batteries, especially those used during charging of new energy vehicles, only time-series signals such as voltage, current, and temperature within a limited and non-fixed charging SOC range can be obtained. Therefore, it is impossible to calculate the current capacity of the lithium battery using the ampere-hour integration method or the OCV model, let alone its subsequent aging trend. Therefore, the following methods are generally used traditionally:

[0004] 1. Theoretical modeling method: According to different theories, it can be mainly divided into electrochemical model and equivalent circuit model. The electrochemical model simplifies the lithium battery into a system consisting of positive electrode, negative electrode, diaphragm and electrolyte. The battery model is established according to electrochemical theories such as internal electrochemical reaction, ion diffusion and polarization effect of lithium battery, and then the battery law is analyzed and mastered. The equivalent circuit model refers to the use of basic electrical components such as capacitors, resistors, and voltage sources to describe the dynamic output characteristics of lithium batteries. Equivalent circuit models are mostly created based on Kirchhoff's voltage law and current law.

[0005] However, the theoretical modeling method is difficult to construct models, and it is difficult to obtain high prediction accuracy. In addition, due to the large differences between lithium battery cells, it is impossible to obtain good model generalization.

[0006] 2. Deep learning method: The deep neural network model is a system that uses a multi-layer hidden neuron network to describe the strong nonlinear relationship between input and output. Therefore, the deep neural network can mine the potential internal characteristics of lithium batteries from big data through training or learning.

[0007] Although the deep learning method can obtain higher accuracy and generalization than the theoretical modeling method through the training of big data, the deep learning method is a single supervised learning method, which has the following disadvantages: a, the supervised learning model is not suitable for processing complex tasks because it needs a large amount of labeled data and computing time, and may have overfitting or underfitting problems; b, the supervised learning model is sensitive to the expression form of the input data, and if the test data is different from the training data set distribution, the supervised learning cannot predict the correct output; c, if the training data set is not carefully reviewed, the supervised learning model will lead to disastrous results because it may follow the bias or errors existing in the training data; d, a large amount of unlabeled data is generated during online charging of lithium batteries, and only a small amount of labeled data is available in the laboratory due to limited cost; the supervised learning model can only utilize the small amount of labeled data in the laboratory, while a large amount of unlabeled data is wasted.

[0008] Therefore, how to provide a lithium battery state of health estimation method combining unsupervised and supervised learning to improve the accuracy and generalization of lithium battery state of health estimation has become a technical problem to be solved. SUMMARY

[0009] The technical problem to be solved by the present application is to provide a lithium battery state of health estimation method combining unsupervised and supervised learning to improve the accuracy and generalization of lithium battery state of health estimation.

[0010] The present application is implemented as follows: a lithium battery state of health estimation method combining unsupervised and supervised learning, comprising the following steps:

[0011] Step S10, a large amount of charging data of the lithium battery is obtained, and a training set and a test set are constructed based on each of the charging data;

[0012] Step S20, the training set and the test set are preprocessed by normalization and length unification;

[0013] Step S30, a first estimation model combining unsupervised learning and supervised learning is created based on an encoding module, a decoding module and an SOH prediction module;

[0014] Step S40, the first estimation model is trained and tested using the preprocessed training set and test set;

[0015] Step S50, the encoding module and the SOH prediction module are extracted from the trained first estimation model, and a second estimation model is created based on the extracted encoding module and SOH prediction module;

[0016] Step S60, the second estimation model is used for lithium battery state of health estimation.

[0017] Furthermore, the step S10 specifically includes:

[0018] Step S11, obtaining a large amount of charging data of the lithium battery including at least voltage, current, temperature, SOC, rated capacity and capacity decay rate;

[0019] Step S12: reconstructing a multivariate time series based on the voltage, current, temperature, and SOC, and using the capacity decay rate corresponding to the multivariate time series as a label for SOH prediction;

[0020] Step S13: randomly sampling a SOC interval from the charging data, taking the voltage, current, temperature, and SOC within the SOC interval as input, and taking the capacity decay rate and multivariate time series as output, to construct a charging data set;

[0021] Step S14: Divide the charging data set into a training set and a test set based on a preset ratio.

[0022] Furthermore, in step S20, the preprocessing of normalizing the training set and the test set is specifically as follows:

[0023] The voltage, current and temperature are obtained from the training set and the test set, the voltage is divided by the rated voltage to normalize to the interval [0, 1], the current is divided by the rated capacity to normalize to the charge rate curve, and the temperature is divided by 100 to normalize to the interval [0, 1].

[0024] Furthermore, in step S20, the preprocessing of the training set and the test set to unify the length is specifically as follows:

[0025] A time series length is set, and data in the training set and the test set whose length is longer than the time series length is truncated, and data whose length is shorter than the time series length is padded with 0 at the end.

[0026] Furthermore, in step S30, the encoding module is constructed based on Transformer; the decoding module is constructed based on CNN and Transformer for unsupervised learning; and the SOH prediction module is constructed based on a fully connected layer for supervised learning.

[0027] Furthermore, in step S40, during the training process, the first estimation model continuously optimizes parameters through a back-propagation algorithm so that the loss value calculated by the set loss function is less than a set loss threshold.

[0028] Furthermore, the formula of the loss function is:

[0029] Loss = α*MSE + β*MAE;

[0030]

[0031]

[0032] Among them, Loss represents the loss value; MSE represents the mean square error loss function; MAE represents the mean absolute value loss function; α and β represent weight coefficients; f(x) represents the predicted value; y represents the target value; n represents the total data volume; i represents the data number.

[0033] Furthermore, the step S60 is specifically as follows:

[0034] The charging data to be estimated of the lithium battery is obtained, the charging data to be estimated is pre-processed and then input into a second estimation model, and the health status of the lithium battery is estimated using the second estimation model.

[0035] The advantages of the present invention are:

[0036] A training set and a test set are constructed by acquiring various charging data, and the training set and the test set are normalized and preprocessed to unify the length; then, a first estimation model integrating unsupervised learning and supervised learning is created based on the encoding module, decoding module and SOH prediction module, and the first estimation model is trained and tested using the training set and test set. Then, the encoding module and the SOH prediction module are extracted from the trained first estimation model to create a second estimation model, and finally, the health status of the lithium battery is estimated using the second estimation model; the encoding module is constructed based on Transformer; the decoding module is constructed based on CNN and Transformer for unsupervised learning; the SOH prediction module is constructed based on the fully connected layer. The proposed method is built for supervised learning. Since CNN has translation invariance and local correlation, while Transformer can capture long-distance dependencies, the combination of CNN and Transformer enables the network structure to inherit the advantages of CNN and Transformer, and retain the global and local battery characteristics to the greatest extent. In addition, the supervised learning and unsupervised learning methods constrain each other and effectively prevent the model from falling into the local optimal solution. Compared with a single supervised learning model, it can be closer to the representation of the internal mechanism of the battery. Unsupervised learning can utilize a large amount of unlabeled charging data, directly increasing the data volume of the training set, and ultimately greatly improving the accuracy and generalization of lithium battery health status estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0038] Figure 1 This is a flow chart of a lithium battery health status estimation method that integrates unsupervised and supervised learning in the present invention.

[0039] Figure 2 It is a structural diagram of the first estimation model of the present invention.

[0040] Figure 3 It is a structural diagram of the encoding module of the present invention.

[0041] Figure 4 It is a structural diagram of the decoding module of the present invention. DETAILED DESCRIPTION

[0042] The technical solution in the embodiment of the present application has the following overall idea: a first estimation model that integrates unsupervised learning and supervised learning is created through an encoding module, a decoding module and an SOH prediction module; the encoding module is constructed based on Transformer; the decoding module is constructed based on CNN and Transformer, and is used for unsupervised learning; the SOH prediction module is constructed based on a fully connected layer, and is used for supervised learning; the combination of CNN and Transformer enables the network structure to retain global and local battery characteristics to the greatest extent, and the two methods of supervised learning and unsupervised learning constrain each other and avoid falling into local optimal solutions. Compared with a single supervised learning model, it can be closer to the representation of the internal mechanism of the battery; and unsupervised learning can utilize a large amount of unlabeled charging data, thereby improving the accuracy and generalization of lithium battery health status estimation.

[0043] Please refer to Figures 1 to 4 As shown, a preferred embodiment of the present invention is a method for estimating the health status of a lithium battery that integrates unsupervised and supervised learning, comprising the following steps:

[0044] Step S10: obtaining a large amount of charging data of lithium batteries, and constructing a training set and a test set based on the charging data;

[0045] Step S20: performing normalization and length unification preprocessing on the training set and the test set;

[0046] Step S30: creating a first estimation model integrating unsupervised learning and supervised learning based on the encoding module, the decoding module, and the SOH prediction module;

[0047] Step S40: training and testing the first estimation model using the preprocessed training set and test set;

[0048] Step S50: extracting a coding module and a SOH prediction module from the trained first estimation model, and creating a second estimation model based on the extracted coding module and SOH prediction module; deploying the trained coding module and SOH prediction module on the application side to perform lithium battery health status estimation;

[0049] Step S60: Estimating the health status of the lithium battery using the second estimation model.

[0050] The step S10 specifically includes:

[0051] Step S11: obtaining a large amount of charging data of the lithium battery, including at least voltage, current, temperature, SOC, rated capacity, and capacity decay rate; in specific implementation, the charging data can be obtained from online charging scenarios and laboratory scenarios;

[0052] Step S12: reconstructing a multivariate time series based on the voltage, current, temperature, and SOC, and using the capacity decay rate corresponding to the multivariate time series as a label for SOH prediction;

[0053] Step S13: randomly sampling a SOC interval from the charging data, taking the voltage, current, temperature, and SOC within the SOC interval as input, and taking the capacity decay rate and multivariate time series as output, to construct a charging data set;

[0054] Step S14: Divide the charging data set into a training set and a test set based on a preset ratio, for example, dividing the charging data set into a training set and a test set in a ratio of 8:2 or 7:3.

[0055] In step S20, the preprocessing of normalizing the training set and the test set is specifically as follows:

[0056] The voltage, current and temperature are obtained from the training set and the test set, the voltage is divided by the rated voltage to normalize to the interval [0, 1], the current is divided by the rated capacity to normalize to the charge rate curve, and the temperature is divided by 100 to normalize to the interval [0, 1].

[0057] In step S20, the preprocessing of the training set and the test set to unify the length is specifically as follows:

[0058] A time series length is set, and data in the training set and the test set whose length is longer than the time series length is truncated, and data whose length is shorter than the time series length is padded with 0 at the end.

[0059] For example, the time series length is set to 2048, the part longer than 2048 is truncated and discarded, and the part shorter than 2048 is padded with 0. The dimension of the input variable is [batch_size, 4, 2048], where batch_size is the batch size entering the model training.

[0060] In step S30, the encoding module is constructed based on Transformer; the decoding module is constructed based on CNN and Transformer for unsupervised learning; and the SOH prediction module is constructed based on a fully connected layer (FC) for supervised learning.

[0061] The structure of the encoding module is as follows Figure 3 As shown in the figure, the input embedding represents a multivariate time series consisting of voltage, current, temperature, and SOC [4, 2048]. The positional embedding is the position encoding of the input embedding. After the output of three series-connected attention blocks, the attention block module extracts features through multi-head attention and FFN. The mathematical representation of multi-head attention and FFN is:

[0062]

[0063] Z2=FFN(Z1)=max(0,ZW1+b1)W2+b2;

[0064] Among them, K, Q, V are the same matrix, and the matrix shape is [4,512]; d k is the arithmetic square root of 512; FFN is a linear projection layer, which ultimately transforms the [2048,512] matrix into a [2048,1024] matrix.

[0065] The structure of the decoding module is as follows Figure 4 As shown in the figure, the input is the encoded feature with a shape of [2048, 1024], and the output is the reconstructed multivariate time series. The decoding module consists of a one-dimensional convolution layer Conv1d and three series-connected Attention Blocks. The mathematical expression of one-dimensional convolution is:

[0066] y(i)=∑(w(k)*x(i+k));

[0067] Where y(i) represents the i-th element of the output sequence; w represents the weight; x represents the input sequence; and k represents the position of the convolution window. Through one-dimensional convolution, the feature vector of shape [2048, 1024] is converted to [2048, 512]. This is then reconstructed back into a multivariate time series of shape [2048, n] through three consecutive Attention Blocks.

[0068] The SOH prediction module is a fully connected layer Linera, which is expressed as follows: the linear projection layer performs linear calculation on the sampled data to predict the SOH, which is mathematically expressed as:

[0069] Y=LN(Z3)=Z3W+b;

[0070] b represents the bias term; the encoded feature is [2048, 1024], and then the data is stretched into a 1-dimensional vector by straightening, and finally converted into a [1, n] matrix shape by the above method, where n is the number of points in Linera regression.

[0071] In step S40, during the training process, the first estimation model continuously optimizes parameters through the back propagation algorithm so that the loss value calculated by the set loss function is less than the set loss threshold.

[0072] The formula of the loss function is:

[0073] Loss = α*MSE + β*MAE;

[0074]

[0075]

[0076] Among them, Loss represents the loss value; MSE represents the mean square error loss function; MAE represents the mean absolute value loss function; α and β represent weight coefficients; f(x) represents the predicted value; y represents the target value; n represents the total data volume; i represents the data number.

[0077] The step S60 is specifically as follows:

[0078] The charging data to be estimated of the lithium battery is obtained, the charging data to be estimated is pre-processed and then input into a second estimation model, and the health status of the lithium battery is estimated using the second estimation model.

[0079] In summary, the advantages of the present invention are:

[0080] A training set and a test set are constructed by acquiring various charging data, and the training set and the test set are normalized and preprocessed to unify the length; then, a first estimation model integrating unsupervised learning and supervised learning is created based on the encoding module, decoding module and SOH prediction module, and the first estimation model is trained and tested using the training set and test set. Then, the encoding module and the SOH prediction module are extracted from the trained first estimation model to create a second estimation model, and finally, the health status of the lithium battery is estimated using the second estimation model; the encoding module is constructed based on Transformer; the decoding module is constructed based on CNN and Transformer for unsupervised learning; the SOH prediction module is constructed based on the fully connected layer. The proposed method is built for supervised learning. Since CNN has translation invariance and local correlation, while Transformer can capture long-distance dependencies, the combination of CNN and Transformer enables the network structure to inherit the advantages of CNN and Transformer, and retain the global and local battery characteristics to the greatest extent. In addition, the supervised learning and unsupervised learning methods constrain each other and effectively prevent the model from falling into the local optimal solution. Compared with a single supervised learning model, it can be closer to the representation of the internal mechanism of the battery. Unsupervised learning can utilize a large amount of unlabeled charging data, directly increasing the data volume of the training set, and ultimately greatly improving the accuracy and generalization of lithium battery health status estimation.

[0081] Although the specific embodiments of the present invention are described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and are not intended to limit the scope of the present invention. Equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A lithium battery health status estimation method integrating unsupervised and supervised learning, characterized by: The steps include: Step S10: obtaining a large amount of charging data of lithium batteries, and constructing a training set and a test set based on the charging data; Step S20: performing normalization and length unification preprocessing on the training set and the test set; Step S30: creating a first estimation model integrating unsupervised learning and supervised learning based on the encoding module, the decoding module, and the SOH prediction module; Step S40: training and testing the first estimation model using the preprocessed training set and test set; During the training process, the first estimation model continuously optimizes parameters through the back propagation algorithm so that the loss value calculated by the set loss function is less than the set loss threshold; the formula of the loss function is: ; ; ; in, Indicates the loss value; represents the mean square error loss function; represents the mean absolute value loss function; and represents the weight coefficient; represents the predicted value; represents the target value; Indicates the total amount of data; Indicates the data number; Step S50: extracting a coding module and an SOH prediction module from the trained first estimation model, and creating a second estimation model based on the extracted coding module and SOH prediction module; Step S60: Estimating the health status of the lithium battery using the second estimation model.

2. The lithium battery health status estimation method integrating unsupervised and supervised learning according to claim 1, characterized in that: The step S10 specifically includes: Step S11, obtaining a large amount of charging data of the lithium battery including at least voltage, current, temperature, SOC, rated capacity and capacity decay rate; Step S12: reconstructing a multivariate time series based on the voltage, current, temperature, and SOC, and using the capacity decay rate corresponding to the multivariate time series as a label for SOH prediction; Step S13: randomly sampling a SOC interval from the charging data, taking the voltage, current, temperature, and SOC within the SOC interval as input, and taking the capacity decay rate and multivariate time series as output, to construct a charging data set; Step S14: Divide the charging data set into a training set and a test set based on a preset ratio.

3. The lithium battery health status estimation method integrating unsupervised and supervised learning according to claim 1, characterized in that: In step S20, the normalization is specifically as follows: The voltage, current and temperature are obtained from the training set and the test set, the voltage is divided by the rated voltage to normalize to the interval [0, 1], the current is divided by the rated capacity to normalize to the charge rate curve, and the temperature is divided by 100 to normalize to the interval [0, 1].

4. The lithium battery health status estimation method integrating unsupervised and supervised learning according to claim 1, characterized in that: In step S20, the length is specifically: A time series length is set, and data in the training set and the test set whose length is longer than the time series length is truncated, and data whose length is shorter than the time series length is padded with 0 at the end.

5. The lithium battery health status estimation method integrating unsupervised and supervised learning according to claim 1, characterized in that: In step S30, the encoding module is constructed based on Transformer; the decoding module is constructed based on CNN and Transformer for unsupervised learning; and the SOH prediction module is constructed based on a fully connected layer for supervised learning.

6. The lithium battery health status estimation method integrating unsupervised and supervised learning according to claim 1, characterized in that: The step S60 is specifically as follows: The charging data to be estimated of the lithium battery is obtained, the charging data to be estimated is pre-processed and then input into a second estimation model, and the health status of the lithium battery is estimated using the second estimation model.

Images