A multi-parameter matching method for lithium-ion batteries based on pre-training

Abstract

The present invention provides a multi-parameter matching method for lithium-ion batteries based on pre-training, which integrates different dynamic parameters through a unified data representation method, which is beneficial to improving the feature extraction of the matching model. The feature encoder of the multi-layer attention layer extracts the features of different dynamic parameters and fuses them with the similarity information of the features of different parameters to obtain global features, which is beneficial to improving the matching accuracy. The pre-training of the matching model does not require data labels, and can use a large amount of unlabeled process data to enhance the feature extraction ability of the model under unsupervised conditions, reducing the model's dependence on matching label data while ensuring the accuracy of the model matching.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium ion batteries, and in particular to a multi-parameter grouping method for lithium ion batteries based on pre-training. Background Art

[0002] Energy storage and conversion technology is crucial to the development of modern society. In the context of the global energy crisis and environmental pollution, how to utilize and develop energy efficiently, cleanly and safely has become an important research field. Batteries are one of the most popular electrical energy storage devices and have attracted widespread attention due to their low environmental pollution cost and high energy conversion efficiency. Among different types of batteries, lithium-ion batteries have the advantages of light weight, high energy density, no memory effect, and high single-cell voltage, making them widely used in portable electronic devices. However, due to the voltage upper limit of single lithium-ion batteries, when dealing with different application scenarios, single cells need to be combined into battery packs through series and parallel connection to provide the required power and capacity.

[0003] In the manufacturing process of lithium-ion batteries, changes in raw materials and production processes can lead to slight differences in the electrical characteristics of batteries in the same batch, that is, battery inconsistency. This inconsistency between batteries, such as capacity, internal resistance, voltage, etc., may cause overcharge and over-discharge of lithium-ion batteries in the battery pack, and further aggravate the working conditions of the batteries in different environments and conditions, resulting in a reduction in service life and threatening the safety of the battery pack.

[0004] In order to reduce the impact of cell inconsistency on battery pack performance and safety, usually after the single cell is manufactured, the cells are divided into different groups according to their electrical characteristics, and the battery packs are configured within the groups. Better screening methods can improve the initial consistency of the battery pack, which is the basis for long life and high performance, and can reduce the difficulty of cell balancing.

[0005] The electrical characteristic parameters of lithium-ion batteries are mainly divided into two categories, static parameters and dynamic parameters. Static parameters are scalars that describe a fixed characteristic of the battery, such as capacity, open circuit voltage, etc. Dynamic parameters are usually time series data collected during the battery charging and discharging process, such as charging voltage sequence, charging current sequence, etc. Existing lithium-ion battery matching methods are mainly divided into single-parameter-based matching methods, multi-parameter-based matching methods, and dynamic parameter-based matching methods according to the parameters used.

[0006] The matching method based on a single parameter is to match the lithium-ion battery by using an electrical characteristic parameter, such as capacity, voltage, internal resistance, etc., such as patent application CN113884906A (a power battery matching method based on normal distribution analysis. These methods are simple and fast and are widely used in the large-scale production of lithium-ion batteries. However, a single parameter cannot describe the overall performance of the battery, which leads to poor consistency within the battery pack.

[0007] The multi-parameter-based matching method takes multiple battery parameters into account to achieve better consistency performance, such as capacity-voltage, capacity-resistance, voltage-current, etc. For example, patent applications CN115764024A (a battery and its matching method), CN113937373A (a lithium battery matching method), CN115856688A (a battery and its internal resistance and capacity testing method and battery sorting method and matching method), CN114167287A (a battery pack sorting method, device and electronic device based on neural network). The multi-parameter-based method can obtain a more comprehensive description of the battery and accurately match it. However, most multi-parameter-based methods only consider the stable parameters of the battery and ignore the dynamic consistency performance of the battery.

[0008] The dynamic parameter-based matching method not only uses the stable parameters of the battery, but also uses the dynamic time series data in the testing and production process, such as the charging voltage curve, the discharge current curve, the capacity curve, etc., to evaluate the similarity of the electrochemical properties of lithium-ion batteries. For example, patent application CN115640702A (a battery matching method and system). The dynamic parameter-based method can achieve better consistency performance than the single parameter and multi-parameter methods. However, the dynamic parameter-based method has a higher model complexity, and the matching consistency performance depends on the quality and quantity of the labeled data.

[0009] The existing technology mainly has the following problems:

[0010] 1. Most battery matching methods only use one data characteristic curve or dynamic data to match the battery, but do not use the multi-parameter data in the battery production process, that is, multiple dynamic parameter data between different steps and processes.

[0011] 2. The feature extraction of the dynamic parameter-based matching method is relatively independent and does not consider the correlation between different source data, resulting in poor consistency performance of the final battery pack.

[0012] 3. Most battery matching methods rely on a lot of assumptions and the quantity and quality of manually labeled data. However, the consistency of different types of lithium-ion batteries is different, and it is difficult to establish the same assumptions and a large number of high-quality labeled data sets for each type of lithium-ion battery. Summary of the invention

[0013] The present invention aims to provide a multi-parameter matching method for lithium-ion batteries based on pre-training, which uses multiple electrical characteristic parameters of lithium-ion batteries, such as voltage curve, current curve, capacity curve, etc., to guide battery matching and improve matching accuracy; combined with the pre-training method, it reduces the dependence of the matching model on the quality and quantity of label data, reduces the cost of manual labeling, and improves the applicability of the model.

[0014] To achieve the above object, the present invention provides a multi-parameter matching method for lithium-ion batteries based on pre-training, and the specific steps are as follows:

[0015] S1. Pre-training: Obtain multi-parameter data of lithium-ion batteries without matching labels, construct a pre-training data set, and use the pre-training data set in combination with the fragment predictor to pre-train the feature encoder of the matching model to obtain a pre-trained complete feature encoder;

[0016] S2. Fine-tuning training: obtain multi-parameter data of lithium-ion batteries with matching labels, construct a matching training set, combine the matching classifier and the pre-trained feature encoder, perform fine-tuning training on the model, and obtain a fully trained multi-parameter matching model for lithium-ion batteries;

[0017] S3. Battery grouping: obtaining multi-parameter data of the battery to be grouped, and classifying the battery to be grouped according to the multi-parameter data of the battery to be grouped and the trained multi-parameter grouping model of lithium-ion batteries, and the classification result is the group to which the battery to be grouped belongs.

[0018] As one of the preferred technical solutions, in step S1, the multi-parameter data of the lithium-ion battery is derived from the process data generated by the battery testing process in the lithium-ion battery production line.

[0019] As one of the further preferred technical solutions, the process of the lithium-ion battery testing process is: formation battery→constant current charging→standing→constant current constant voltage charging #1→standing→constant current discharge→standing→constant current constant voltage charging #2→test completed.

[0020] As one of the further preferred technical solutions, by collecting the test process data corresponding to each battery, multi-parameter data of the lithium-ion battery is obtained, which includes five dynamic parameter curves related to the characteristics of the lithium-ion battery, namely, the constant current charging voltage curve, the constant current and constant voltage charging #1 voltage curve, the constant current discharge voltage curve, the constant current and constant voltage charging #2 voltage curve and the constant current and constant voltage charging #2 current curve.

[0021] As one of the further preferred technical solutions, different dynamic parameters in the multi-parameter data of lithium-ion batteries are integrated to obtain an overall data representation form; for the input multi-parameter data of lithium-ion batteries X = [x0, x1, ..., x q ], where X is the input lithium-ion battery multi-parameter data, q is an integer greater than 1, indicating the number of dynamic parameters, is the i-th dynamic parameter, 0≤i≤q, is time series data, R is the real number domain, T i is the timing length of the dynamic parameter, is the data point of the ii-th dynamic parameter. First, all dynamic parameters in X are normalized, as shown in formula (1);

[0022]

[0023] in are the maximum and minimum values ​​of the ith dynamic parameter, respectively. is the normalized dynamic parameter Time series data of

[0024] For the normalized multi-parameter data of lithium-ion batteries Each dynamic parameter Fragmentation Decompose into n fixed length d in Sub-series data is the vector dimension d in A real vector, All data fragments of the same dynamic parameter are combined to form the data fragment representation of the i-th dynamic parameter. is a matrix of dimension n×d in A real matrix with n rows and d in List.

[0025] In different data segments, F i Insert specific data fragments between It is called parameter interval marker, which is a sequence of length d in , a special data segment with all zero values, plays the role of identifying the boundaries of different segment data representations in the unified data representation; the separated data representation of the input lithium ion multi-parameter data X is obtained Separated data representation F in is a matrix of the real field R, with dimensions q(n+1)-1 rows and d inColumn; q is the number of dynamic parameters, each dynamic parameter is divided into n data segments, and then a parameter interval identifier is inserted between each data segment. All the above data segments, a total of q(n+1)-1 data segments, are connected in rows to obtain a separated data representation.

[0026] As one of the preferred technical solutions, in step S1, for the input lithium-ion battery multi-parameter data X, a special segment data is also used. That is, the fragment replacement mark replaces the data fragment in the separated data representation according to the probability p to obtain the pre-trained separated data representation F i ' n and the replaced data fragment Where i is the i-th dynamic parameter, j is the j-th replaced data segment, 0≤j≤n; To separate data representations, data segments are randomly replaced with segment replacement identifiers;

[0027] The separated data is input into the feature encoder and processed to obtain the global feature F global ; The global feature F global The feature data corresponding to the position of the fragment replacement mark in the fragment is extracted and input into the fragment predictor in sequence; the fragment predictor predicts the original data fragment according to the input features, and obtains the fragment prediction results in sequence Predicting results by calculating fragments With the replaced fragment data The mean square error (MSE) between them is shown in formula (2), and the model parameters of the feature encoder and the fragment predictor are updated according to the error; when the error of the model pre-training is less than ε, the model pre-training is terminated and a pre-trained feature encoder is obtained;

[0028]

[0029] Where N is the number of replaced data segments.

[0030] As one of the preferred technical solutions, in step S1, the structure of the feature encoder is composed of L attention layers, and the attention layer is composed of attention calculation, layer normalization and feedforward neural network; for the input feature with the lth attention layer First, map it to different spaces, as shown in equations (3)(4)(5);

[0031]

[0032]

[0033]

[0034] in is the input feature of the lth attention layer; They are the query matrix, key matrix, and value matrix corresponding to the input features respectively; is the linear mapping operator of the corresponding matrix; then, the attention weight is calculated by calculating the query matrix Q l With the key matrix K l The similarity of the pair value matrix V l The features in are weighted, so that the similarity information, that is, the relationship information, is integrated into the features, as shown in formula (6);

[0035]

[0036] Among them, Softmax(.) is the softmax function used to normalize the attention weights, M l is the attention calculation result of the lth attention layer, d k is the dimension of the eigenvector of the value matrix; then, M l Through residual connection and Fusion Input feedforward neural network. The feedforward neural network in the attention layer is a two-layer neural network, as shown in formula (7);

[0037]

[0038] in is the output of the network, Relu(.) is the Relu activation function, W1, W2 are the weights of the feedforward neural network, b1, b2 are the weight biases of the feedforward neural network; Through residual connection and The sum is finally obtained to get the output of the lth attention layer

[0039] As one of the preferred technical solutions, in step S1, the structure of the fragment predictor consists of two linear layers, the linear layers are connected by an activation function, and the number of hidden units in the linear layer is h; the task of the fragment predictor is to obtain the global feature F in the pre-training stage. global The features of the corresponding positions of the fragment replacement identifiers are used to predict the original data fragments, and the fragment prediction results are obtained in turn. And by calculating the fragment prediction results With the replaced fragment data The mean square error (MSE) between them is used to update the feature encoder.

[0040] As one of the preferred technical solutions, the specific method of step S2 is as follows: the separated data is input into a pre-trained feature encoder, and the global feature is obtained after processing; then the global feature is input into the matching classifier, and the matching classifier predicts the predicted matching label corresponding to the battery data according to the input global feature By calculating the matching group annotation y and predicting the matching group annotation The cross entropy error (CE) between them, C is the number of battery grouping groups (for example, there are three groups A, B, and C, then C = 3), as shown in formula (8), and the model parameters of the feature encoder and the grouping classifier are updated according to the error, and a well-trained lithium-ion battery multi-parameter grouping model is obtained by fine-tuning;

[0041]

[0042] Among them, p t is the predicted probability of the corresponding group label, y t The value marked for the group (for example, the corresponding number for group A is 0, the corresponding number for group B is 1, and the corresponding number for group C is 2), 0≤t≤C.

[0043] For example, the original labeled data is letters such as "A". Before entering the grouping model, the grouped labeled data needs to be converted into a one-hot encoded input, that is, a vector, for example, C=3, labeled group A->[1,0,0].

[0044] The present invention has the following beneficial effects:

[0045] The present invention integrates different dynamic parameters through a unified data representation method, which is conducive to improving the feature extraction of the matching model. The feature encoder of the multi-layer attention layer extracts the features of different dynamic parameters and fuses them with the similarity information of different parameter features to obtain global features, which is conducive to improving the matching accuracy. The pre-training of the matching model does not require data labels, and a large amount of unlabeled process data can be used to enhance the feature extraction ability of the model under unsupervised conditions, reducing the model's dependence on matching label data while ensuring the accuracy of the model matching.

[0046] The specific advantages of the present invention are as follows:

[0047] 1. Multiple dynamic parameter curves of lithium-ion batteries (such as voltage curve, current curve, capacity curve, etc.) are used to guide battery matching, with high matching accuracy and good effect, which further improves the consistency of battery packs compared with traditional methods.

[0048] 2. Use pre-training to train the model, reduce the dependence of the matching model on the quality and quantity of matching label data, reduce the cost of manual matching annotation, and improve the applicability of the matching model.

[0049] 3. Using an integrated data representation method to represent multi-parameter data of lithium-ion batteries is conducive to feature extraction. Multiple dynamic characteristic parameters are integrated into the same data representation, and on this basis, global features with correlation information of different dynamic parameter characteristics are extracted for grouping, thereby improving the accuracy of grouping.

[0050] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0052] Figure 1 It is the lithium-ion battery testing process flow;

[0053] Figure 2 This is an example of multi-parameter data for a lithium-ion battery;

[0054] Figure 3 This is an example of normalized multi-parameter data for a lithium-ion battery;

[0055] Figure 4 It is the lithium ion multi-parameter data representation process of the present invention;

[0056] Figure 5 It is the overall framework of the multi-parameter matching method for lithium-ion batteries based on pre-training;

[0057] Figure 6 It is the main step of the multi-parameter matching method of lithium-ion batteries based on pre-training;

[0058] Figure 7 This is the structure diagram of the model group in this paper during the model pre-training stage;

[0059] Figure 8 is the feature encoder structure;

[0060] Fig. 9 It is the attention layer structure;

[0061] Fig.10 is the fragment predictor structure;

[0062] Fig.11 This is a structural diagram of the matching model of the present invention in the fine-tuning stage;

[0063] Fig.12 is the matching classifier structure;

[0064] Fig.13 It is a structural diagram of the battery assembly model of the present invention at the battery assembly stage. DETAILED DESCRIPTION

[0065] The embodiments of the present invention are described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0066] Example:

[0067] A multi-parameter matching method for lithium-ion batteries based on pre-training. The method uses multiple electrical characteristic parameters of lithium-ion batteries, such as voltage curve, current curve, capacity curve, etc., to guide battery matching and improve matching accuracy. Combined with the pre-training method, the matching model is less dependent on the quality and quantity of label data, the cost of manual labeling is reduced, and the applicability of the model is improved.

[0068] 4.1 Data Source

[0069] The data used in the present invention is multi-parameter data of lithium-ion batteries, and the data source is process data generated by the battery testing process in the lithium-ion battery production line. Figure 1 By collecting the test process data corresponding to each battery, multi-parameter data of lithium-ion batteries can be obtained, such as Figure 2 As shown, there are five dynamic parameter curves related to the characteristics of lithium-ion batteries, namely, constant current charging voltage curve, constant current constant voltage charging #1 voltage curve, constant current discharge voltage curve, constant current constant voltage charging #2 voltage curve and constant current constant voltage charging #2 current curve. These curves contain information related to the static and dynamic consistency of the battery, especially the consistency performance under dynamic conditions, which plays an important role in guiding battery grouping, reducing the inconsistency of batteries within the group, and improving the performance and life of the battery pack.

[0070] 4.2 Data Representation

[0071] Most traditional battery matching methods based on dynamic parameters only use one battery dynamic parameter for matching, which cannot fully describe the consistency of the battery under dynamic working conditions. Secondly, some current data-driven multiple dynamic parameter matching methods extract the features of different dynamic parameters separately in terms of dynamic parameter feature extraction, and then perform simple fusion, such as. The multi-parameter data features of lithium-ion batteries obtained by this method do not contain information about the relationship between different parameters, while the relationship information between different dynamic parameters can fully reflect the performance of the battery under different working conditions, which is conducive to improving the consistency of the battery pack and reducing the complexity of the feature extraction module.

[0072] Therefore, the first step of the present invention is to integrate the different dynamic parameters in the lithium-ion battery multi-parameter data to obtain an overall data representation form; for the input lithium-ion battery multi-parameter data X = [x0, x1, ..., x q], where X is the input lithium-ion battery multi-parameter data, q is an integer greater than 1, indicating the number of dynamic parameters, is the ith dynamic parameter, is time series data, R is the real number domain, T i is the timing length of the dynamic parameter, is the data point of the ith dynamic parameter. First, all dynamic parameters in X are normalized, as shown in formula (1).

[0073]

[0074] in are the maximum and minimum values ​​of the i-th dynamic parameter respectively. is the normalized dynamic parameter An example of normalized multi-parameter data of lithium-ion batteries is shown in Figure 2. Figure 3 As shown, through data normalization, different dynamic parameters are mapped to expressions under the same dimension, which is beneficial for subsequent processing.

[0075] For the normalized multi-parameter data of lithium-ion batteries Each dynamic parameter Fragmentation Decompose into n fixed length d in Sub-series data is the vector dimension d in A real vector, All data fragments of the same dynamic parameter are combined to form the data fragment representation of the i-th dynamic parameter. is a matrix of dimension n×d in A real matrix with n rows and d in List;

[0076] Finally, in order to extract the relationship and data features of different dynamic data under a unified data representation, the matching model requires boundary information between different fragment data representations, so that the model can distinguish data fragments with different dynamic parameters under a unified data representation, thereby extracting overall features and relationship features. i Insert specific data fragments between It is called parameter interval marker, which is a sequence of length d in , a special data segment with all zero values, plays the role of identifying the boundaries of different data segments in the unified data representation. The debugging process of all the above data is as follows Figure 4 Finally, the separated data representation of the input lithium ion multi-parameter data X is obtained This representation integrates all dynamic parameters in the input lithium-ion multi-parameter data X into a unified data representation and uses parameter interval markers to separate the data of different dynamic parameters, so that the matching model can not only extract the data features of a single dynamic parameter, but also extract the relationship features between different dynamic parameters, which is beneficial to improving the accuracy of lithium-ion battery matching.

[0077] 4.3 Model matching steps

[0078] The overall framework of the present invention is as follows Figure 5 As shown, the structure of the matching model of the present invention is different in different steps, and the module represented by the dotted box is only used in the pre-training step.

[0079] The operation of the present invention specifically comprises three steps: Figure 6 shown.

[0080] 4.3.1 Model pre-training stage

[0081] like Figure 7 As shown in FIG. 1 , a large amount of multi-parameter data of lithium-ion batteries without matching labels is first obtained to construct a pre-training data set for the matching model. The pre-training data set is used in combination with the fragment predictor to pre-train the feature encoder of the matching model to obtain a pre-trained feature encoder. Specifically, for the input multi-parameter data X of lithium-ion batteries, in addition to using the above data representation method to process the data, another special fragment data is also used. That is, the segment replacement mark replaces the data segment in the separated data representation according to the probability p to obtain the pre-trained separated data representation F′ in and the replaced data fragment Among them, i is the i-th dynamic parameter, and j is the j-th replaced data segment. The separated data is input into the feature encoder, and the global feature F is obtained after processing. global . The global feature F global The feature data corresponding to the position of the fragment replacement mark in the fragment is extracted and input into the fragment predictor in sequence. The fragment predictor predicts the original data fragment according to the input features and obtains the fragment prediction results in sequence. Predicting results by calculating fragments With the replaced fragment data The mean square error (MSE) between them is shown in formula (2), and the model parameters of the feature encoder and the fragment predictor are updated according to the error. When the error of the model pre-training is less than ε, the model pre-training is terminated and a pre-trained feature encoder is obtained.

[0082]

[0083] Where N is the number of replaced data segments.

[0084] The model structure of the feature encoder and segment prediction in the above process is as follows Figure 8 and Fig.10 The structure of the feature encoder consists of L attention layers, where the structure of each attention layer is as follows Fig. 9 As shown in Figure 1, the attention layer consists of attention calculation, layer normalization and feedforward neural network. For the input feature with the lth attention layer First, map it to different spaces, as shown in equations (3)(4)(5).

[0085]

[0086]

[0087]

[0088] in is the input feature of the lth attention layer. They are the query matrix, key matrix, and value matrix corresponding to the input features respectively. is the linear mapping operator of the corresponding matrix. Then, the attention weight is calculated by calculating the query matrix Q l With the key matrix K l The similarity of the pair value matrix V l The features in are weighted to fuse the similarity information, that is, the relationship information, into the features, as shown in formula (6).

[0089]

[0090] Among them, Softmax(.) is the softmax function used to normalize the attention weights, M l is the attention calculation result of the lth attention layer, d k is the dimension of the eigenvector of the value matrix. Then, M l Through residual connection and Fusion Input feedforward neural network. The feedforward neural network in the attention layer is a two-layer neural network, as shown in formula (7).

[0091]

[0092] in is the output of the network, Relu(.) is the Relu activation function, W1, W2 are the weights of the feedforward neural network, and b1, b2 are the weight biases of the feedforward neural network. Through residual connection and The sum is finally obtained to get the output of the lth attention layer

[0093] The structure of the fragment predictor is as follows Fig.10 As shown in the figure, it consists of two linear layers, which are connected by an activation function, and the number of hidden units in the linear layer is h. The task of the fragment predictor is to use the global feature F global The features of the corresponding positions of the fragment replacement identifiers are used to predict the original data fragments, and the fragment prediction results are obtained in turn. And by calculating the fragment prediction results With the replaced fragment data The feature encoder is updated by the mean square error (MSE) between the two, so that the feature encoder can adapt to the structure and characteristics of lithium-ion battery multi-parameter data under the condition of only lithium-ion battery multi-parameter data without matching annotations, which is beneficial to the feature extraction of subsequent matching tasks.

[0094] 4.3.2 Model fine-tuning stage

[0095] like Fig.11 As shown in the figure, firstly, a small amount of lithium-ion battery multi-parameter data with matching labels is obtained, and a matching model fine-tuning dataset is constructed. The fine-tuning dataset is used to combine the matching classifier and the pre-trained feature encoder to perform model fine-tuning training to obtain a fully trained lithium-ion battery multi-parameter matching model. Specifically, for the input lithium-ion battery multi-parameter data X and its corresponding matching annotation y k In addition to using the above data representation to process the data to obtain the separated data representation. The separated data is input into the pre-trained feature encoder, and the global features are obtained after processing. The global features are then input into the grouping classifier, and the grouping classifier predicts the predicted grouping label corresponding to the battery data based on the input global features. By calculating the matching group annotation y and predicting the matching group annotation The cross entropy error (CE) between them is , k is the number of battery grouping groups (for example, there are three groups A, B, and C, then k = 3), as shown in formula (8), and the model parameters of the feature encoder and the grouping classifier are updated according to the error, and a well-trained lithium-ion battery multi-parameter grouping model is obtained by fine-tuning.

[0096]

[0097] Among them, p k is the predicted probability of the corresponding group label, y k The value of the group annotation. For example, the original annotation data is the letter "A". Before entering the grouping model, the group annotation data needs to be converted into a one-hot encoding input, that is, a vector, for example, C = 3, annotating group A->[1,0,0].

[0098] like Fig.12As shown in the figure, the structure of the pairing classifier in the above process is similar to that of the fragment predictor in the pre-training stage, both of which are composed of several linear layers connected by activation functions. Unlike the fragment predictor, the input of the pairing classifier is the global feature and the output is the pairing annotation information, so the number of its hidden units h is different from that of the fragment predictor.

[0099] 4.3.3 Battery assembly stage

[0100] First, the multi-parameter data of the lithium-ion battery to be matched is obtained, and the batteries to be matched are classified according to the multi-parameter data of the batteries to be matched and the well-trained multi-parameter matching model of lithium-ion batteries. The matching result is the group to which the batteries to be matched belong, such as Fig.13 shown.

[0101] In summary, the present invention integrates different dynamic parameters through a unified data representation method, which is conducive to improving the feature extraction of the matching model. Secondly, the feature encoder of the multi-layer attention layer extracts the features of different dynamic parameters and fuses them with the similarity information of the features of different parameters to obtain global features, which is conducive to improving the matching accuracy. Finally, the pre-training of the matching model does not require data labels, and a large amount of unlabeled process data can be used to enhance the feature extraction ability of the model under unsupervised conditions, reducing the model's dependence on matching label data while ensuring the accuracy of the model matching.

[0102] Test example

[0103] In order to verify the effectiveness of the present invention, dynamic parameter data of the test process of 43,361 18650 lithium-ion batteries were collected. The original data set is divided into a pre-training set, a training set, a validation set and a test set after interpolation alignment. The specific details are shown in Table 1, wherein the pre-training set data has no matching labels and is used for pre-training matching models. The training set and the verification machine are lithium-ion multi-parameter data sets containing matching labels, which are used to fine-tune the matching model and optimize the model hyperparameters. The test set is a battery data set to be matched, which contains the corresponding matching label data, and is used to test the performance of the matching method of the present invention and calculate the matching accuracy.

[0104] Table 1. Dataset information

[0105]

[0106] In order to compare the performance of the present invention and different model methods in the multi-parameter matching task of lithium-ion batteries, the present invention uses Macro F1, Macro Precision, Macro Recall, Top-1 accuracy and Top-3 accuracy as evaluation indicators to measure the performance and performance of the model in the multi-parameter matching task of lithium-ion batteries. The model complexity is measured by the average running time of a single sample. The method of the invention is compared with a multi-layer perceptron (MLP), a fully convolutional network (FCN), a deep residual network (ResNet), a multi-scale convolutional neural network (MCNN) and a multi-parameter-based lithium-ion battery matching method (DDDMO). The above methods are all trained in a supervised manner using data with matching labels. The experimental results are shown in Table 2, where the best results are represented in bold.

[0107] Table 2. Comparison results of different matching methods

[0108]

[0109] The experimental results show that the method of the present invention has achieved the best results in most evaluation indicators. The Top-1 accuracy of the method of the present invention is 5.53% higher than that of the DDDOM method. The Top-1 accuracy and Top-3 accuracy of the DDDOM method are roughly the same, indicating that the method is less effective when dealing with difficult samples. Other models performed poorly in the lithium ion multi-parameter matching method task, and only the method of the present invention and the DDDOM method exceeded 90% in the Top-1 accuracy. Secondly, the best running time result comes from the MLP method. The model complexity of the method of the present invention is relatively high, and the average running time of a single sample reaches 0.5468ms. In general, the method of the present invention can complete the lithium ion multi-parameter matching task with higher accuracy and improve the consistency and matching quality of the battery pack.

[0110] In summary, the present invention proposes a multi-parameter matching method for lithium-ion batteries based on pre-training. The method uses multiple dynamic parameter data in the lithium-ion battery testing process to guide the battery matching task. Combined with the pre-trained model training strategy, the model's dependence on matching annotation data is reduced. At the same time, a unified data representation method is used to improve the matching feature extraction performance, thereby improving the accuracy of battery matching, which is beneficial to improving the life, performance and safety of the battery pack.

[0111] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A multi-parameter matching method for lithium-ion batteries based on pre-training, characterized in that: The specific steps are as follows: S1. Pre-training: Obtain multi-parameter data of lithium-ion batteries without matching labels, construct a pre-training data set, use the pre-training data set, combined with the fragment predictor, pre-train the feature encoder of the matching model, and obtain a pre-trained complete feature encoder; The multi-parameter data of lithium-ion batteries are derived from the process data generated by the battery testing process in the lithium-ion battery production line; by collecting the test process data corresponding to each battery, the multi-parameter data of lithium-ion batteries are obtained, which includes five dynamic parameter curves related to the characteristics of lithium-ion batteries, namely, the constant current charging voltage curve, the constant current constant voltage charging #1 voltage curve, the constant current discharge voltage curve, the constant current constant voltage charging #2 voltage curve and the constant current constant voltage charging #2 current curve; Integrate different dynamic parameters in multi-parameter data of lithium-ion batteries to obtain an overall data representation; For the input lithium-ion battery multi-parameter data ,in is the input lithium-ion battery multi-parameter data, is an integer greater than 1, indicating the number of dynamic parameters. For the dynamic parameters, , is time series data, is the field of real numbers, is the time length of the dynamic parameter, For the data points of dynamic parameters; first All dynamic parameters in are normalized, and the formula is shown in formula (1); （1） in , Respectively The maximum and minimum values ​​of the dynamic parameters, is the normalized dynamic parameter Time series data of For the normalized multi-parameter data of lithium-ion batteries , each dynamic parameter Fragmentation Decompose into Fixed length Sub-series data , is the vector dimension A real vector, All data fragments with the same dynamic parameter are combined to form the first The data segment of the dynamic parameters , is the matrix dimension The real matrix of OK, List; Represented in different data segments Insert specific data fragments between , called parameter interval marker, which is a sequence of length , a special data segment with all zero values, which serves to identify the boundaries of different data segments in a unified data representation; the input lithium ion multi-parameter data is obtained Delimited data representation ; S2. Fine-tuning training: obtain multi-parameter data of lithium-ion batteries with matching labels, construct a matching training set, combine the matching classifier and the pre-trained feature encoder, perform fine-tuning training on the model, and obtain a fully trained multi-parameter matching model for lithium-ion batteries; S3. Battery grouping: obtaining multi-parameter data of the battery to be grouped, and classifying the battery to be grouped according to the multi-parameter data of the battery to be grouped and the trained multi-parameter grouping model for lithium-ion batteries, and the classification result is the group to which the battery to be grouped belongs.

2. The method according to claim 1, characterized in that: In step S1, the process of the lithium-ion battery test procedure is: formation battery→constant current charging→standing→constant current constant voltage charging #1→standing→constant current discharge→standing→constant current constant voltage charging #2→test completed.

3. The method according to claim 1, characterized in that In step S1, for the input lithium-ion battery multi-parameter data , and also uses a special fragment data , that is, the fragment replacement mark, according to the probability Replace the data fragments in the separated data representation to obtain the pre-trained separated data representation and the replaced data fragment ,in, For the dynamic parameters, For the The separated data is input into the feature encoder, and the global feature is obtained after processing. ; The global features The feature data corresponding to the position of the fragment replacement mark in the fragment is extracted and input into the fragment predictor in sequence; the fragment predictor predicts the original data fragment according to the input features, and obtains the fragment prediction results in sequence ; Predict results by calculating fragments With the replaced fragment data The mean square error between , as shown in formula (2), and the model parameters of the feature encoder and the segment predictor are updated according to the error; when the error of the model pre-training is less than , model pre-training is terminated, and a fully pre-trained feature encoder is obtained; （2） in is the number of replaced data fragments.

4. The method according to claim 1, characterized in that: In step S1, the structure of the feature encoder is composed of attention layers, which consist of attention calculation, layer normalization and feedforward neural network; for Input features of the attention layer , first map it to different spaces, as shown in equations (3), (4), and (5): （3） （4） （5） in For the Input features of the attention layer; , , They are the query matrix, key matrix, and value matrix corresponding to the input features respectively; , , is the linear mapping operator of the corresponding matrix; Then, the attention weights are calculated by calculating the query matrix With key matrix The similarity of the pair matrix The features in are weighted, so that the similarity information, that is, the relationship information, is integrated into the features, as shown in formula (6); （6） in is the softmax function used to normalize the attention weights, For the The attention calculation result of the attention layer, d k is the dimension of the eigenvectors of the value matrix; then, Through residual connection and Fusion , input feedforward neural network; the feedforward neural network in the attention layer is a two-layer neural network, as shown in formula (7); （7） in is the output of the network, is the Relu activation function, , is the weight of the feedforward neural network, , is the weight bias of the forward neural network; Through residual connection and Finally, we get Attention layer output .

5. The method according to claim 1, characterized in that In step S1, the structure of the fragment predictor consists of two linear layers connected by an activation function, and the number of hidden units in the linear layer is ; The task of the segment predictor is to use the global features in the pre-training stage The features of the corresponding positions of the fragment replacement identifiers are used to predict the original data fragments, and the fragment prediction results are obtained in turn. , and predict the result by calculating the fragment With the replaced fragment data The mean square error between them is used to update the feature encoder.

6. The method according to claim 1, characterized in that The specific method of step S2 is as follows: the separated data is input into a pre-trained feature encoder, and the global features are obtained after processing; the global features are then input into a matching classifier, and the matching classifier predicts the predicted matching label corresponding to the battery data according to the input global features. ; By calculating the matching group Predicted matching annotation The cross entropy error between is the number of battery grouping groups, as shown in formula (8), and the model parameters of the feature encoder and the grouping classifier are updated according to the error, and a fully trained lithium-ion battery multi-parameter grouping model is obtained by fine-tuning; （8） in, is the predicted probability of the corresponding group label, The value marked for the group.

Images