A Hybrid Model-Based Data Augmentation Method for Monitoring the Health Status of Lithium-ion Batteries

Abstract

This invention relates to a data-augmented hybrid model-based lithium battery health status monitoring method, belonging to the field of battery detection technology. The method uses a multi-head attention mechanism and a temporal convolutional hybrid model as its base model, and incorporates generative adversarial methods into the hybrid neural network model. It designs an adversarial model for time-series data generation, adding supervised learning to the embedding and generation modules on top of the traditional unsupervised generative adversarial module. The parallel and serial connection of the hybrid neural network overcomes the unsupervised nature of generative adversarial learning in traditional neural network methods and the limitations of single-model data feature extraction, compensating for information loss due to insufficient training sets, and achieving high-precision lithium battery health status monitoring based on limited datasets. This invention can provide sufficient data through data augmentation, even with limited abnormal operating condition battery data, to achieve high-precision battery health status prediction.

Description

Technical Field

[0001] This invention belongs to the field of lithium-ion battery testing technology, and relates to a data-augmented hybrid model-based method for monitoring the health status of lithium batteries. Background Technology

[0002] With the increasing demand in the new energy vehicle market, breakthroughs in key technologies for electric vehicles are urgently needed. As the power source for electric vehicles, the development of lithium-ion batteries determines their future. To ensure battery safety and reliability, the Battery Management System (BMS), a crucial component of electric vehicles, undertakes functions such as data acquisition, equalization management, and state estimation. Among these, the State of Health (SOH) is one of the most important indicators of the BMS, making accurate diagnosis of the battery's SOH essential.

[0003] In State of Health (SOH) diagnostic technology, model-based methods are frequently employed, such as equivalent circuit models, electrochemical models, and mathematical models. In addition, neural network models are widely used in lithium-ion battery health status monitoring. Neural network models achieve a direct mapping from input to output by learning the nonlinear relationships between data. Some traditional neural networks, such as Artificial Neural Networks (ANNs), Probabilistic Neural Networks (PNNs), Convolutional Neural Networks (CNNs), prior knowledge-based neural networks, Gated Logic Regulators (GRUs), and Long Short-Term Memory (LSTM) neural networks, have been widely used in battery SOH monitoring. Furthermore, among numerous hybrid models, neural networks are the most widely used single model in combination with other models. A hybrid network based on Convolutional Neural Networks (GRU-CNN) extends CNNs using GRU-RNN submodules. The CNN convolutional blocks utilize a shared weight structure to reduce the number of weights and attempt to find shared information from measured charging voltage, current, and temperature data. On the other hand, GRU-RNN blocks use their internal states to learn characteristics and temporal dependencies from sequential data. In effect, this combined structure can simultaneously leverage the advantages of CNNs and GRU-RNN networks to capture the shared spatiotemporal characteristics of charging data. Similarly, ensemble algorithms combining ELM and LSTM can capture the underlying correspondences between State of Health (SOH). Furthermore, numerous studies have demonstrated the feasibility of combining the advantages of CNN and LSTM for lithium-ion battery health prediction.

[0004] However, most of the currently available data comes from batteries operating under normal conditions, while data from a small number of abnormal conditions, such as extreme situations (e.g., battery swelling, runaway, combustion, explosion), are extremely scarce. Repeatedly conducting battery-related experiments to collect this data requires expensive equipment and a significant amount of time. Furthermore, the lack of diversity in publicly available datasets also limits research in this field. Therefore, how to process the limited amount of lithium battery data under abnormal conditions and improve the accuracy of model predictions is an urgent problem for lithium battery health monitoring. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a data-augmented hybrid model for monitoring the health status of lithium batteries. The method uses a multi-head attention mechanism and a temporal convolutional hybrid model as the base model, integrates generative adversarial methods, and combines unsupervised and supervised learning methods to make up for the lack of information due to insufficient training set, thereby achieving high-precision monitoring of the health status of lithium batteries based on a limited dataset.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A data-augmented hybrid model-based method for monitoring the health status of lithium batteries is proposed. This method uses a hybrid neural network model based on multi-head attention mechanism and temporal convolutional neural network as the base model, while integrating temporal generative adversarial methods, embedding supervised learning methods, and combining regression models to predict the health status of lithium batteries. The implementation steps of this method are as follows:

[0008] S1. Collect the training dataset and randomly generate a noisy dataset;

[0009] S2. Using temporal generative adversarial networks as the basis for the data augmentation model, a data augmentation model is constructed using the training dataset and a noisy dataset to generate the augmented dataset.

[0010] S3. Extract the effective temporal and spatial features of battery data using a hybrid neural network model;

[0011] S4. Integrate time-effective features and space-effective features, and input the integrated feature vector into the regression model to predict the health status of lithium batteries.

[0012] Furthermore, in step S1, the training dataset is represented as The noisy dataset is represented as Where T represents the sequence length, n∈{1,...,N} represents the index of a single sample, and N represents the number of samples.

[0013] Furthermore, in step S2, the data augmentation model includes a generator G and a discriminator D. The generator G is used to generate effective data that approximates the real data, and the discriminator D is used to classify the real data and the effective data as true or false.

[0014] Furthermore, the data augmentation model construction process is as follows: a generator G is trained using the training dataset and the noisy dataset, and the generator G that minimizes the maximum value is obtained through adversarial evaluation; at the same time, the discriminator D is optimized through JS divergence to maximize the classification accuracy of the discriminator D; during the construction of the data augmentation model, the discriminator D maximizes the classification accuracy of the training data and the effective data, and the generator G minimizes the maximum value. When the maximization of the discriminator D and the minimization of the generator G reach Nash equilibrium, the data augmentation model construction is completed.

[0015] Furthermore, in step S3, an encoder with a multi-head attention mechanism is used to extract the time-effective features of the lithium battery data;

[0016] The encoder's input A (l) for:

[0017] A (l) =PE (l) +X (l)

[0018] In the formula, X (l) Represents the input sample of the Transformer, PE (l) This represents the position code of the input sample, where the position code is as follows:

[0019]

[0020] In the formula, pos represents the position of the current element in the entire vector, i represents the index of each element in the vector, and d represents the dimension of the vector;

[0021] Encoder output B (l) for:

[0022]

[0023] In the formula, h represents the number of multi-head attention mechanisms, C is the concatenation function, and W... q W k and W v Both represent weight matrices, W i Q W i K and W i V Both represent the multi-head mapping parameter matrix, W O This represents the fusion matrix.

[0024] Furthermore, in step S3, 1d-TCN is used to extract the spatially effective features of the lithium battery data;

[0025] For the nth convolutional layer, its input X n and output Y n They are represented as follows:

[0026]

[0027]

[0028] In the formula, L and W represent the length and width of the input data, respectively; the correspondence between input and output is shown in the following formula:

[0029]

[0030] In the formula, σ represents the activation function. d represents the kernel bias, s represents the kernel stride, and s represents the kernel stride. v c v k represents the position coordinates of the features extracted by the convolution kernel. v Represents the kernel ordinal number; Y n Inputting the max pooling layer yields the pooling vector:

[0031]

[0032]

[0033] In the formula, P n Represents the pooling vector, x l y w The position coordinates of the pooling output are represented by d, and the pooling step size is represented by d.

[0034] Furthermore, the length of the input data is the length of the sample, and the width of the input data is the lithium battery data sample type.

[0035] Furthermore, in step S4, a regression model based on a weight allocation mechanism is used to predict the health status of the lithium battery;

[0036] For the integrated feature vector The weight allocation mechanism is as follows:

[0037]

[0038] In the formula, d model′ ×n′ represents the feature size of the output of the mixture model, and W1′ and W2′ represent the parameter matrices. denoted as the Hadamard product dot product, and h′ represents the weight correction coefficient.

[0039] The beneficial effects of this invention are as follows: This invention integrates generative adversarial methods into a hybrid neural network model, designs an adversarial model for time-series data generation, and adds supervised learning of embedding and generation modules on the basis of traditional unsupervised generative adversarial modules. This breaks through the limitations of unsupervised generative adversarial learning and single model in data feature extraction in traditional neural network methods, makes up for the information loss caused by insufficient training sets, and compensates for the deficiencies of data imbalance and difficulty in obtaining experimental data in lithium battery research, thereby achieving high-precision monitoring of lithium battery health status based on limited datasets.

[0040] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0041] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0042] Figure 1 This is a schematic diagram of the framework model of the present invention. Detailed Implementation

[0043] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0044] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0045] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0046] This invention addresses the lack of abnormal operating condition samples in lithium battery health status prediction tasks. It utilizes deep learning methods to learn the shared information and time dependencies of discharge curves, achieving accurate prediction of battery health status. Figure 1 As shown, on the data side, a temporal generative adversarial network model is constructed to process battery usage data under abnormal operating conditions, and a data-augmented training set is obtained; during the modeling process, an end-to-end high-performance learning model is constructed by studying the multi-head attention mechanism and the parallel hybrid model of temporal convolutional neural networks; on the prediction side, the output of the hybrid model is connected, corrected and the prediction result is given.

[0047] The specific contents of this invention are as follows:

[0048] 1. Battery anomaly data augmentation based on temporal generative adversarial model

[0049] Obtain the original training dataset Randomly generated noisy dataset Where T represents the sequence length, n∈{1,...,N} represents the index of a single sample, and N represents the number of samples. Define the feature space. Distribution p(X) 1:T Generator G distribution Using the training dataset Training should be as close as possible to the feature space distribution p(X) 1:T Generation distribution of ) This leads to the generation of augmented datasets. The process is as follows:

[0050] Data on lithium batteries Based on the temporal characteristics of the training data, a temporal generative adversarial network (GAN) is used as the data augmentation model to generate effective data that approximates the training data as closely as possible. In this model, the feature space is implemented through an embedding function e. and potential space Completely invertible mapping between them:

[0051]

[0052] The latent representation h is recovered by the r function. 1:T Reconstructed into a precise representation of the feature space

[0053]

[0054] The process includes the embedding function e, the reconstruction function r, and the mapping-reconstruction loss function. It can be represented as:

[0055]

[0056]

[0057] In the formula, h t Representing the potential space The output vector, x t This represents the original input data. Let p represent the input vector to the prediction network, and p represent the feature space distribution.

[0058] During the augmentation data generation phase, the original training dataset and noisy datasets Used for training generator G. During training, generator G operates in two states: open-loop mode and closed-loop mode.

[0059] In open-loop mode (unsupervised learning), the discriminator D is derived from the latent space. Extract the input and output the true / false classification of the data.

[0060]

[0061] The generator function and the discriminator function are as follows:

[0062]

[0063]

[0064] In the formula, For generator functions, Here, g() is the discriminator function, g() is the generator network, d() is the discriminator network, and z is the discriminator network. t This is noisy data.

[0065] The ultimate goal of the generator can be represented as:

[0066] G * =argminmaxV(G,D)

[0067] The goal is to evaluate the differences between datasets adversarially and obtain a generator G that minimizes the maximum value. Therefore, the JS divergence metric is used to measure the vectors obtained in two new feature spaces. The JS divergence theory is expressed as follows:

[0068]

[0069] For a given The generated sample is obtained through G. Similarly, the output feature vector X is generated by embedding a portion of the network E. i =E(x) i ) to obtain x i ={x 1 ,x 2 ,...,x n From this, we can know that:

[0070]

[0071] Where, p date It's real data, p g It's fake data. Optimize the discriminator. Update via gradient descent

[0072]

[0073] In the formula, m represents the total number of data samples. Finally, the adversarial unsupervised loss function between the generator G and the discriminator D can be expressed as:

[0074]

[0075] In the formula, y t This represents the output feature of the discriminator, where t represents the label.

[0076] In closed-loop mode (supervised learning), noisy datasets Transform into the latent space via generator

[0077]

[0078] To further capture the distribution of the original data, a supervised learning method is introduced. In this mode, the generator receives the embedding sequence h of the actual data. 1:t-1 Therefore, the embedding latent distribution p(H) t |H 1:t-1 and generating potential distributions The difference distribution between them can be represented by the maximum likelihood function as:

[0079]

[0080] Throughout the construction of the data augmentation model, the discriminator attempts to maximize its classification accuracy on both the training and generated data, while the generator attempts to minimize it. The data augmentation model is completed when the two reach a Nash equilibrium.

[0081] 2. A lithium battery health status monitoring model based on a multi-head attention mechanism and a temporal convolutional hybrid model

[0082] 1) Battery data feature extraction

[0083] In the hybrid neural network model, 1d-TCN and an encoder with multi-head attention mechanism are used as feature extractors to simultaneously extract spatial and temporal dependency information from lithium battery data.

[0084] ① Targeting the characteristics of one-dimensional lithium battery sequences, an encoder with a multi-head attention mechanism uses a global information acquisition method to extract the temporally effective features of the lithium battery data (including raw data and augmented data) in parallel. For the l-th layer (l∈1,2,...,L) encoder model, let... and Let be the input and output of the l-th layer encoder, where D represents the dimension of the vector. Specifically, and These are the input samples and position codes of the Transformer, respectively. This is the input to the first-layer encoder after position embedding. Position encoding preserves the relative and absolute positions of each element in the input sequence. The position encoding is calculated as follows:

[0085] A (1) =PE (1) +X (1)

[0086]

[0087]

[0088] In the formula, pos represents the position of the current element in the entire vector; i represents the index of each element in the vector, i.e., 2i represents an even position, and 2i+1 represents an odd position; d represents the dimension of the vector, which is consistent with the encoding vector. By the properties of trigonometric functions:

[0089]

[0090] available:

[0091]

[0092] Where k represents the relative position vector, it can be seen that the pos+k position vector can be represented as a linear combination of the pos and k position vectors. This linear combination means that the absolute position vector contains the relative position vector, thus reflecting the temporal sequence of the lithium battery sequence. After position encoding, A (l) Input to the encoder to get the encoder output:

[0093]

[0094] In the formula, h represents the number of multi-head attention mechanisms, and C is the concatenation function. and Both represent weight matrices. and All represent multi-head mapping parameter matrices. This represents the fusion matrix.

[0095] ② As a supplement to the multi-head attention mechanism, a 1d-TCN is introduced as a feature extractor to extract the spatially effective features of the battery data. For the nth layer (n∈1,2,...,N) TCN model, let... and These represent the input and output of a one-dimensional convolutional layer, and the convolutional kernel, respectively. The number is k m Here, L and W represent the length and width of the input data, respectively. Since lithium battery data is a one-dimensional time series, the length of the input data X1 is the sampling length, and the width is the lithium battery data sample type, such as current, voltage, temperature, sampling time, capacity, etc. Therefore, for a given input X... n-1 Output Y n It can be represented as:

[0096]

[0097]

[0098] In the formula, σ represents the activation function. d represents the kernel bias, s represents the kernel stride, and s represents the kernel stride. v c v k represents the position coordinates of the features extracted by the convolution kernel. v Indicates the kth v One convolutional kernel.

[0099] After convolution, Y n The input is fed into the max pooling layer to obtain:

[0100]

[0101]

[0102] In the formula, P n Here, x is the pooling vector, and O represents the pooling kernel; l y w The position coordinates of the pooling output are represented by d; the pooling step size is represented by d.

[0103] In simple terms, training a one-dimensional convolutional network from input to output can be represented as learning the parameters Θ of a composite nonlinear function F(x|Θ):

[0104] Y = F(x|Θ) = f L (L f2(f1(x|θ1)|θ2|θ L )

[0105] 2) Battery health status prediction based on regression model

[0106] A regression model based on a weighted allocation mechanism is proposed to predict the health status of lithium batteries. The input to this regression model is the feature vector V output by a hybrid neural network model, which is the output B of a Transformer. (l) and the output P of 1d-TCN n The feature vectors output by the hybrid model are obtained by concatenation. The weight allocation mechanism can be represented as:

[0107]

[0108] In the formula, d model ×n′ represents the feature size output by the hybrid model. The parameter matrix determined during training, denoted as the Hadamard product dot product, and h′ represents the weight correction coefficient.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A data augmentation based hybrid model lithium battery state of health monitoring method, characterized in that: The method takes a hybrid neural network model based on a multi-head attention mechanism and a time convolutional neural network as a basic model, simultaneously fuses a time sequence generative adversarial network model, embeds a supervised learning method, and combines a regression model to realize prediction of the lithium battery health state, and implementation steps of the method are as follows: S1, collect a training data set, and randomly generate a noise data set; S2, taking a time sequence generative adversarial network as a data enhancement model basis, using the training data set and the noise data set to construct a data enhancement model, and generating an enhanced data set; S3, extracting time effective features and space effective features of the battery data through a hybrid neural network model respectively; An encoder adopting a multi-head attention mechanism is used to extract time effective features of the lithium battery data; The input of the encoder is: wherein denotes the input sample of the Transformer, denotes the positional encoding of the input sample, wherein the positional encoding is given by wherein, pos represents the position of the current element in the whole vector, i represents the index of each element in the vector, d represents the dimension of the vector; Output of the encoder Is: In the formula, h represents the number of heads in the multi-head attention mechanism, C is a concatenation function, , and all represent weight matrices, , and all represent multi-head mapping parameter matrices, represents a fusion matrix; 1d-TCN is used to extract space effective features of the lithium battery data; For the first n The input and output of the layer convolution layer are respectively represented as: wherein L and W respectively represent the length and width of the input data; the correspondence between the input and the output is shown in the following equation: In the formula, represents an activation function, represents a bias of the convolution kernel, represents a step of the convolution kernel, , represents a position coordinate of the extracted feature of the convolution kernel, represents a convolution kernel sequence number; will input the maximum pooling layer to obtain a pooling vector: In the formula, denotes the pooling vector, , denotes the position coordinates of the pooling output, denotes the pooling step size; O denotes the pooling kernel; S4, integrating the time effective features and the space effective features, and inputting the integrated feature vector into a regression model to predict the lithium battery health state.

2. The lithium battery state of health monitoring method of claim 1, wherein: In step S1, the training dataset is denoted as , the noise dataset is denoted as , where T denotes the sequence length, denotes the individual sample index, N denotes the number of samples.

3. The lithium battery state of health monitoring method of claim 1, wherein: In step S2, the data enhancement model includes a generator G and a discriminator D, the generator G is used to generate effective data approximating the real data, and the discriminator D is used for true and false classification of the real data and the effective data.

4. The lithium battery state of health monitoring method of claim 3, wherein: The data enhancement model construction process is: training the generator G using the training data set and the noise data set, and obtaining the generator G with minimum maximum value through adversarial evaluation; and optimizing the discriminator D through JS divergence, and maximizing the classification accuracy of the discriminator D; In the construction process of the data enhancement model, the discriminator D maximizes the classification accuracy of the training data and the effective data, and the generator G minimizes the maximum value, when the maximization of the discriminator D and the minimization of the generator G reach Nash equilibrium, the data enhancement model is constructed.

5. The lithium battery state of health monitoring method of claim 1, wherein: The length of the input data is the length of the sampling, and the width of the input data is the type of lithium battery data sample.

6. The lithium battery state of health monitoring method of claim 1, wherein: In step S4, a regression model based on a weight distribution mechanism is used to predict the health state of the lithium battery; For the integrated eigenvectors The weight distribution mechanism is: wherein denotes the feature size of the mixed model output, , denotes the parameter matrix, denotes the Hadamard product, denotes the weight correction factor.

Images