Retired lithium battery gradient screening method based on multi-modal information fusion

Abstract

The invention provides a gradient screening method for retired lithium batteries based on multi-modal information fusion, and belongs to the field of battery screening, and the method comprises the steps: S1, obtaining the charging voltage data of the retired lithium batteries in the charge-discharge cycle of the retired lithium batteries; s2, carrying out dimension reduction processing on the charging voltage data to obtain dimension-reduced voltage data, converting the dimension-reduced voltage data into polar coordinate data, converting the polar coordinate data into a coded image by adopting a Gramb angle field, and inputting the coded image into a ViT model to extract image features; s3, inputting the dimension reduction voltage data obtained in the step S2 into an LSTM model to extract time sequence features; and S4, after the image features and the time sequence features are fused to obtain fusion features, according to the fusion features, the types of the lithium batteries are obtained by adopting an SVM, and then gradient screening of the retired lithium batteries is realized. According to the method, complex feature engineering can be avoided, meanwhile, limitation of respective modes is made up, and high-precision gradient screening of the retired lithium batteries is achieved.

Description

Technical Field

[0001] The present invention belongs to the field of battery screening, and in particular relates to a gradient screening method for retired lithium batteries based on multimodal information fusion. Background Art

[0002] Lithium batteries are widely used in new energy vehicles due to their high energy density and low self-discharge. At the end of their life cycle, lithium batteries are facing retirement because they cannot meet the power requirements of electric vehicles such as driving range and acceleration. However, retired lithium batteries can also be used in many other devices with relatively low power and energy requirements. Therefore, the recycling of lithium batteries can create considerable economic and social benefits, and can also alleviate environmental pollution and energy depletion problems.

[0003] Gradient screening of retired lithium batteries is a key step in recycling. At present, the methods widely used for gradient screening of retired lithium batteries can be divided into three categories: direct measurement method, model-based method and data-driven method. These methods all have corresponding defects: the direct measurement method measures the actual capacity of the battery by performing a complete charge and discharge cycle on the lithium battery, or obtains the internal resistance information of the battery through pulse testing, but the direct measurement method is complicated and time-consuming to operate, and the charge and discharge cycle will cause energy waste. The model-based method predicts the health status of the battery through a physical model, thereby realizing the screening and evaluation of retired lithium batteries, but it has high requirements on model accuracy and is not very versatile. At the same time, the implementation process is relatively complex and inefficient. The data-driven method extracts battery data features and then uses machine learning methods to establish a reliable data model to achieve battery screening and evaluation. The screening model generally includes a clustering model and a classification model. However, if the selected cluster center is significantly different from the true center, it will lead to a large error in the category center, which will affect the subsequent screening results. Therefore, the results of the classification model are very dependent on the selection of features. Before the features are input into the model, a very rigorous and complex feature engineering is required to demonstrate the effectiveness of the features. Summary of the invention

[0004] The purpose of the present invention is to propose a gradient screening method for retired lithium batteries based on multimodal information fusion, which can avoid complex feature engineering, while compensating for the limitations of each modality and achieving high-precision gradient screening of retired lithium batteries.

[0005] The present invention is achieved through the following technical solutions: A gradient screening method for retired lithium batteries based on multimodal information fusion comprises the following steps: Step S1, obtaining charging voltage data of retired lithium batteries during the charge and discharge cycle of retired lithium batteries; Step S2, performing dimension reduction processing on the charging voltage data to obtain dimension-reduced voltage data, converting the dimension-reduced voltage data into polar coordinate data, converting the polar coordinate data into a coded image using a Gram angle field, and then inputting the coded image into a ViT model to extract image features; Step S3, inputting the dimension-reduced voltage data obtained in step S2 into the LSTM model to extract time series features; Step S4: after fusing the image features obtained in step S2 and the time series features obtained in step S3 to obtain fused features, SVM is used to obtain the lithium battery category according to the fused features, thereby realizing gradient screening of retired lithium batteries.

[0006] Furthermore, in step S2, the polar coordinate data acquisition process includes the following steps: Step S21: preprocess the charging voltage data to obtain voltage time series data with a length of m. Q = q 1, q 2,…, q m , after compression ratio n = m / k The dimensionality reduction process obtains a length of k Dimensionality-reduced voltage time series data Q '= q' 1, q' 2,…, q' k , the dimension-reduced voltage time series data corresponds to k timestamps, where ; Step S22: Using formula Scale the reduced-dimensional voltage time series data to the interval [-1,1] to obtain the scaled reduced-dimensional voltage time series data ,in, ; Step S23: Using formula The scaled voltage time series data Convert to corresponding polar coordinate data ,in, t i express The corresponding timestamp, N It means to divide the interval [-1,1] into N Equal portions.

[0007] Furthermore, in step S2, the polar coordinate data is converted into a coded image using the Gram angle field. Expressed as ,in, is the vector angle in polar coordinates.

[0008] Furthermore, in step S3, the dimension-reduced voltage time series data is compressed, and the compressed data is scaled to the interval [-1,1] by a minimum-maximum scaler. X Enter the LSTM model.

[0009] Furthermore, in step S3, extracting the time series features includes the following steps: Step S31: According to the formula Calculate the forget gate, where f t is the output of the forget gate, its value range is between 0 and 1, σ is the sigmoid activation function, W f is the weight of the forget gate, b f is the bias of the forget gate, h t-1 is the hidden state of the previous moment, x t is the input at the current moment, x t ∈ X ; Step S32: According to the formula Calculate the input gate and calculate the candidate memory unit according to the formula ,in, i t is the output of the input gate, is the output of the candidate memory cell, W i and W C are the weight matrices of the input gate and candidate memory cells, respectively. b i and b C is bias; Step S33: According to the formula Calculate and update the cell state, where C t Current moment t The cell state, f t To decide what information should be forgotten, It is the current moment t Newly added cell status information; Step S34: According to the formula Compute the output gate, where o t is the output of the output gate, Wo is the weight matrix of the output gate, b o is the bias of the output gate; Step S35: According to the formula Calculate the hidden state at the current moment and use the hidden state as the time series feature, where h t is the hidden state at the current moment, tanh( C t ) is used to scale the cell state.

[0010] Furthermore, obtaining the lithium battery classification in step S4 specifically includes the following steps: Step S41: N' Retired lithium batteries include N' For the sample set of samples, i' samples, and use concat to fuse the image features and time series features to obtain the fusion features x i' ; Step S42: The fusion features of each sample x i' And the classification label y of the sample i' Form sample data and convert the sample data Input SVM for training; Step S43: After the training is completed, SVM learns a set of support vectors And the Lagrange multiplier corresponding to each element of the support vector , predict new samples x When, according to the formula Classify to get lithium battery classification, is the kernel function, which represents the new sample x With support vector x i' The similarity of b is the offset of the classification hyperplane, i '=1,2..., N '.

[0011] Furthermore, in step S42, the optimization goal of SVM is to find the hyperplane that maximizes the interval with the samples in the sample data. The optimization problem is as follows: Indicates that all samples satisfy the constraints ,in, represents the norm of the hyperplane, w represents the normal vector of the hyperplane, b represents the bias, and C represents the regularization coefficient, which is used to control the error term The weight of , ; In order to deal with nonlinear classification problems, the optimization problem is converted into Lagrangian dual form and the kernel function is introduced. The new optimization problem is expressed as , the constraints are ,in, is the Lagrange multiplier, , kernel function Select the radial basis kernel.

[0012] The present invention has the following beneficial effects: The present invention takes the charging voltage data of retired batteries as input. On the one hand, the charging voltage data is first subjected to dimensionality reduction processing to obtain dimensionality reduction voltage data, and then the dimensionality reduction voltage data is converted into polar coordinate data, and then the polar coordinate data is converted into a coded image using the Gram angle field, and finally the coded image is input into the ViT model to extract image features. On the other hand, the dimensionality reduction voltage data is input into the LSTM model to extract time series features, and then the image features and the time series features are fused, and the SVM is used to obtain the lithium battery category, thereby realizing the gradient screening of retired lithium batteries. In this process, there is no need for a complex manual extraction process, and the ViT model is used to extract rich image features and the LSTM model is used to extract time series features, thereby avoiding complex feature engineering. By fusing the information of the two modalities of image features and time series features, the limitations of the one-dimensional time series signal and the two-dimensional image are compensated, and a richer and more distinctive feature representation is provided, thereby improving the gradient screening accuracy of retired batteries, and only the charging voltage data is required to complete the screening, which is suitable for different retired battery classifications and has stronger adaptability and practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The present invention is further described in detail below with reference to the accompanying drawings.

[0014] Figure 1 It is a flow chart of the present invention.

[0015] Figure 2 Schematic diagram of the ViT model structure of the present invention.

[0016] Figure 3 It is a schematic diagram of the Encoder Block structure in the ViT model of the present invention.

[0017] Figure 4 Schematic diagram of the MLP Block structure in the ViT model of the present invention. DETAILED DESCRIPTION

[0018] like Figures 1 to 4 As shown, the gradient screening method for retired lithium batteries based on multimodal information fusion includes the following steps: Step S1, obtaining charging voltage data of retired lithium batteries during the charge and discharge cycle of retired lithium batteries; Step S2, performing dimension reduction processing on the charging voltage data to obtain dimension-reduced voltage data, converting the dimension-reduced voltage data into polar coordinate data, converting the polar coordinate data into a coded image using a Gram angle field, and then inputting the coded image into a ViT model to extract image features; The specific steps include: Step S21: preprocess the charging voltage data to obtain voltage time series data with a length of m. Q = q 1, q 2,…, q m , after compression ratio n = m / k The dimensionality reduction process obtains a length of k Dimensionality-reduced voltage time series data Q '= q' 1, q' 2,…, q' k , the dimension-reduced voltage time series data corresponds to k timestamps, where ; Step S22: In order to prevent the inner product from being biased towards the maximum value observation, a Min-Max scaler limited to [-1, 1] is used to scale the voltage time series after dimension reduction to [-1, 1]. Specifically, the formula is used. Scale the reduced-dimensional voltage time series data to the interval [-1,1] to obtain the scaled reduced-dimensional voltage time series data ,in, ; Step S23: Using formula The scaled voltage time series data Convert to corresponding polar coordinate data , that is, the scaled voltage value is encoded with the angle cosine to obtain the angle, the value is between [0,π], and the corresponding timestamp is encoded as the radius ρ, where, t i express The corresponding timestamp, N It means to divide the interval [-1,1] into N equal parts; Polar coordinate conversion puts the time series of Cartesian coordinate system into the unit circle of polar coordinate system. The advantages of this encoding are: the encoding process is a double mapping, given a time series of Cartesian coordinate system, the result in polar coordinate system is unique; at the same time, because the timestamp is converted to polar axis, the time correlation of the original time series is retained, and time series with different intervals will contain different angle limits after rescaling and conversion to polar coordinates, providing different information granularity for converting one-dimensional time series into two-dimensional image data.

[0019] Step S24: Use the Gram angle field to convert the polar coordinate data into a coded image. Expressed as ,in, is the vector angle in polar coordinates; Specifically, the Gram Matrix (GM) of a set of n vectors is a matrix defined by the dot product of each pair of vectors, expressed as , because the vector dot product contains the angle and direction information of the two vectors, the Gram matrix can express a certain relationship between the vectors in the group of vectors, so the features extracted according to GM also contain the vector relationship in the time series. However, the G value converted in this way is roughly Gaussian distributed with 0 as the center, which makes it difficult to distinguish valid data from Gaussian noise. Therefore, polar coordinate data with dual mapping is selected to be converted into a coded image. At the same time, in order to retain the information of the two vector angles and parse the individual and connection information from two perspectives, the encoding method used is the Gram trigonometric function difference field GADF, which is expressed as .

[0020] Step S25, input the encoded image into the ViT model to obtain image features. Since the length of the compressed time series is consistent, the converted image size is consistent and the resolution is small. The ViT model is used to extract image features. ViT regards the image as a serialized input and uses the self-attention mechanism to process the pixel relationship in the image. First, the image is divided into a series of tiles, and each tile is converted into a vector representation as an input sequence; then these vectors are processed through a multi-layer Transformer encoder to capture the spatial dependencies between different positions in the image; finally, the image features are output through linear transformation. The ViT model structure is as follows Figure 2As shown. First, the input image is passed through the Patches Embedding layer to generate L tokens. Embedding includes convolution operation on the input image to divide it into L patches, and flattening the length and width of the image to obtain L tokens; then the classification token is added through the concat operation to obtain L+1 tokens, and then the position encoding information is added to each token. Finally, the Transformer encoder is passed and the encoded classification token is extracted as the output feature. The Transformer encoder contains several Encoder Blocks. The Encoder Block structure is as follows: Figure 3 As shown in the figure, it is divided into two parts. The first part inputs L+1 tokens into the Layer Norm layer, then passes through the Multi-Head Attention and Dropout layers, and finally adds the Dropout result to the initial input; the second part inputs the result of the first part into the LayerNorm layer, then passes through the MLP Block and Dropout layers, and finally adds the Dropout result to the result of the first part. Among them, the MLP Block structure is as follows Figure 4 shown.

[0021] Step S3, inputting the dimension-reduced voltage data obtained in step S2 into the LSTM model to extract time series features; The specific steps include: Step S31: compress the dimension-reduced voltage time series data, and scale the compressed data to the interval [-1,1] through a minimum-maximum scaler. X Input LSTM model. The purpose of this step is to speed up the convergence speed when training LSTM and prevent gradient explosion or gradient disappearance problems; Step S32, the LSTM core includes a cell state and three gates (forget gate, input gate, output gate) structure. According to the formula Calculate the forget gate, which determines the information of the previous moment to be retained at the current moment, where: f t is the output of the forget gate, its value range is between 0 and 1, σ is the sigmoid activation function, W f is the weight of the forget gate, b f is the bias of the forget gate, h t-1 is the hidden state of the previous moment, x t is the input at the current moment, x t ∈X ; Step S33: According to the formula Calculate the input gate and calculate the candidate memory unit according to the formula , the input gate determines the impact of the current input on the cell state, where i t is the output of the input gate, is the output of the candidate memory cell, W i and W C are the weight matrices of the input gate and candidate memory cells, respectively. b i and b C is bias; Step S34: According to the formula Calculate and update the cell state, which is updated through the forget gate and input gate. C t Current moment t The cell state, f t To decide what information should be forgotten, It is the current moment t Newly added cell status information; Step S35: According to the formula Calculate the output gate, which determines the effect of the cell state on the output, where o t is the output of the output gate, W o is the weight matrix of the output gate, b o is the bias of the output gate; Step S36: According to the formula Calculate the hidden state at the current moment and use the hidden state as a time series feature. The hidden state combines the output gate and the cell state at the current moment, where: h t is the hidden state at the current moment, tanh( C t ) is used to scale the cell state.

[0022] In the LSTM layer, each time step receives the current input and the cell state of the previous time step. Through the control of the forget gate, input gate, and output gate, LSTM can decide whether to forget or add information, and whether information is output to the hidden state. In this process, the hidden state gradually contains the local feature information of the sequence.

[0023] Step S4: after fusing the image features obtained in step S2 and the time series features obtained in step S3 to obtain fused features, SVM is used to obtain the lithium battery category according to the fused features, thereby realizing gradient screening of retired lithium batteries.

[0024] The specific steps include: Step S41: N' Retired lithium batteries include N' For the sample set of samples, i' samples, and use concat to fuse the image features and time series features to obtain the fusion features x i' ; Step S42: The fusion features of each sample x i' And the classification label y of the sample i' Form sample data and convert the sample data Input SVM for training; The optimization goal of SVM is to find a hyperplane , so that the interval between the sample point and the hyperplane is maximized, that is, find the maximum interval The hyperplane of the optimization problem is as follows: Indicates that the solution to the optimization objective is the support vector, that is, the set of sample points closest to the hyperplane, and all samples satisfy the constraints ,in, represents the norm of the hyperplane, w represents the normal vector of the hyperplane, which determines the direction of the hyperplane. b represents the bias, which controls the distance between the hyperplane and the origin, and C represents the regularization coefficient, which is used to control the error term The weight of , ; In order to deal with nonlinear classification problems, the optimization problem is converted into Lagrangian dual form, and the dual optimization problem of SVM is obtained , the constraints are transformed into, , the optimization objective of the dual optimization problem contains the inner product term , reflects the similarity of sample points in the feature space, so the kernel function is introduced Replacing the inner product term, the dual optimization problem is transformed into a new optimization problem, expressed as , the constraints are still ,in, is the Lagrange multiplier, and the sample point Related, , Kernel Function Select the radial basis kernel, that is, .

[0025] Step S43: After the training is completed, SVM learns a set of support vectors And the Lagrange multiplier corresponding to each element of the support vector , predict new samples x When, according to the formula Classify to get lithium battery classification, is the kernel function, that is , represents the new sample x With support vector x i' The similarity of b is the offset of the classification hyperplane, i '=1,2..., N '.

[0026] The process of gradient screening retired batteries according to the retired battery classification results is an existing technology.

[0027] The above description is only a preferred embodiment of the present invention, and therefore cannot be used to limit the scope of implementation of the present invention. That is, equivalent changes and modifications made according to the scope of the patent application of the present invention and the contents of the specification should still fall within the scope covered by the patent of the present invention.

Claims

1. A gradient screening method for retired lithium batteries based on multimodal information fusion, characterized in that: The steps include: Step S1, obtaining charging voltage data of retired lithium batteries during the charge and discharge cycle of retired lithium batteries; Step S2, performing dimension reduction processing on the charging voltage data to obtain dimension-reduced voltage data, converting the dimension-reduced voltage data into polar coordinate data, converting the polar coordinate data into a coded image using a Gram angle field, and then inputting the coded image into a ViT model to extract image features; Step S3, inputting the dimension-reduced voltage data obtained in step S2 into the LSTM model to extract time series features; Step S4: after fusing the image features obtained in step S2 and the time series features obtained in step S3 to obtain fused features, SVM is used to obtain the lithium battery category according to the fused features, thereby realizing gradient screening of retired lithium batteries.

2. According to claim 1, a gradient screening method for retired lithium batteries based on multimodal information fusion is characterized in that: In step S2, the polar coordinate data acquisition process includes the following steps: Step S21: preprocess the charging voltage data to obtain voltage time series data with a length of m. Q = q 1, q 2,…, q m , after compression ratio n = m / k The dimensionality reduction process obtains a length of k Dimensionality-reduced voltage time series data Q '= q' 1, q' 2,…, q' k , the dimension-reduced voltage time series data corresponds to k timestamps, where ; Step S22: Using formula Scale the reduced-dimensional voltage time series data to the interval [-1,1] to obtain the scaled reduced-dimensional voltage time series data ,in, ; Step S23: Using formula The scaled voltage time series data Convert to corresponding polar coordinate data ,in, t i express The corresponding timestamp, N It means to divide the interval [-1,1] into N Equal portions.

3. The method for gradient screening of retired lithium batteries based on multimodal information fusion according to claim 2 is characterized in that: In step S2, the polar coordinate data is converted into a coded image using the Gram angle field. Expressed as ,in, is the vector angle in polar coordinates.

4. The method for gradient screening of retired lithium batteries based on multimodal information fusion according to claim 3 is characterized in that: In step S3, the dimension-reduced voltage time series data is compressed, and the compressed data is scaled to the interval [-1,1] by a minimum-maximum scaler. X Enter the LSTM model.

5. The method for gradient screening of retired lithium batteries based on multimodal information fusion according to claim 4 is characterized in that: In step S3, extracting the time series features includes the following steps: Step S31: According to the formula Calculate the forget gate, where f t is the output of the forget gate, its value range is between 0 and 1, σ is the sigmoid activation function, W f is the weight of the forget gate, b f is the bias of the forget gate, h t-1 is the hidden state of the previous moment, x t is the input at the current moment, x t ∈ X ; Step S32: According to the formula Calculate the input gate and calculate the candidate memory unit according to the formula ,in, i t is the output of the input gate, is the output of the candidate memory cell, W i and W C are the weight matrices of the input gate and candidate memory cells, respectively. b i and b C is bias; Step S33: According to the formula Calculate and update the cell state, where C t Current moment t The cell state, f t To decide what information should be forgotten. It is the current moment t Newly added cell status information; Step S34: According to the formula Compute the output gate, where o t is the output of the output gate, W o is the weight matrix of the output gate, b o is the bias of the output gate; Step S35: According to the formula Calculate the hidden state at the current moment and use the hidden state as the time series feature, where h t is the hidden state at the current moment, tanh( C t ) is used to scale the cell state.

6. The method for gradient screening of retired lithium batteries based on multimodal information fusion according to claim 5 is characterized in that: The step S4 of obtaining the lithium battery classification specifically includes the following steps: Step S41: N' Retired lithium batteries include N' For the sample set of samples, i' samples, and use concat to fuse the image features and time series features to obtain the fusion features x i' ; Step S42: The fusion features of each sample x i' And the classification label y of the sample i' Form sample data and convert the sample data Input SVM for training; Step S43: After the training is completed, SVM learns a set of support vectors And the Lagrange multiplier corresponding to each element of the support vector , predict new samples x When, according to the formula Classify to get lithium battery classification, is the kernel function, which represents the new sample x With support vector x i' The similarity of b is the offset of the classification hyperplane, i '=1,2..., N '.

7. The method for gradient screening of retired lithium batteries based on multimodal information fusion according to claim 6 is characterized in that: In step S42, the optimization goal of SVM is to find the hyperplane that maximizes the interval with the samples in the sample data. The optimization problem is as follows: Indicates that all samples satisfy the constraints ,in, represents the norm of the hyperplane, w represents the normal vector of the hyperplane, b represents the bias, and C represents the regularization coefficient, which is used to control the error term The weight of , ; In order to deal with nonlinear classification problems, the optimization problem is converted into Lagrangian dual form and the kernel function is introduced. The new optimization problem is expressed as , the constraints are ,in, is the Lagrange multiplier, , kernel function Select the radial basis kernel.

Images