A lithium battery state of charge estimation method based on shrinkage and spatial interaction network

Abstract

The present application relates to a kind of lithium battery state of charge estimation method based on shrinkage and spatial interaction network, belong to new energy technology field.The method includes the following steps: S1: using one or more of lithium battery current, voltage, temperature and sampling time sequence as model input, on the basis of residual network, soft threshold is inserted as nonlinear conversion layer into deep architecture, to eliminate unimportant features, realize the purpose of de-noising, obtain new feature matrix as the input of next module;S2: using the feature matrix X' of data de-noising processing after last module as data input, spatial interaction network learns the dependence between data, finally output the state of charge of lithium battery.The present application uses simple convolution and interaction network to extract time series data features while increasing the connection between each sub-sequence, and then improves the ability of model to learn the correlation of time series data before and after, to complete the accurate estimation of lithium battery state of charge.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of new energy, and relates to a lithium battery state of charge estimation method based on contraction and spatial interaction network. BACKGROUND

[0002] At present, electric vehicles are widely concerned by the academic and automotive industries due to their environmental friendliness and resource conservation, and develop rapidly. As one of the most important devices in electric vehicles, the health and safety of lithium batteries are crucial. As a key technology for guaranteeing the health state of lithium batteries, accurate lithium battery state of charge estimation is one of the important ways for the battery management system to manage the safety of the battery. In order to improve the accuracy of state of charge estimation, many studies have utilized the characteristics of the time series correlation of lithium battery characterization data, and proposed many methods capable of extracting and fully utilizing the data before and after the relationship and obtained many high-precision results.

[0003] There are three main categories of existing deep learning methods in the field of lithium battery state of charge estimation: various recurrent neural networks and variants; various encoder-decoder and variants; and time-domain convolutional networks. The best method at present is to use a time-domain convolutional network for prediction. This method can effectively capture local features using convolutional filters, and can finally integrate these local features into a whole to complete prediction. However, this method has two problems. First, the data collected by the sensor in the actual situation is noisy, which can cause overfitting or underfitting of the model, affecting the accuracy of the model. Second, the time series convolutional network only has the last layer with a global view of the entire backtracking window, while the effective receptive field of the intermediate layers (especially the layers close to the input) is limited, resulting in the loss of the relationship between the data before and after the feature extraction process. Therefore, the accurate lithium battery state of charge estimation task still faces major challenges. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a lithium battery state of charge estimation method based on contraction and spatial interaction network.

[0005] To achieve the above purpose, the present application provides the following technical scheme:

[0006] A lithium battery state of charge estimation method based on contraction and spatial interaction network, the method comprising the following steps:

[0007] S1: using one or more of the lithium battery current, voltage, temperature and sampling time series as the model input, inserting a soft threshold as a nonlinear conversion layer into the deep architecture based on the residual network to eliminate unimportant features and achieve the purpose of denoising, and obtaining a new feature matrix as the input of the next module;

[0008] S2: using the last module, the feature matrix X' after data denoising processing is taken as data input, the spatial interaction network learns the dependence between data, and finally the state of charge of the lithium battery is output.

[0009] Optionally, the S1 is specifically:

[0010] S11: the two-dimensional data first passes through three layers of ordinary two-dimensional convolution layers, wherein the channel number is set to 1, a new feature map x is obtained, the absolute value of the feature map x is subjected to a global average pooling operation to obtain a one-dimensional vector; the one-dimensional vector is propagated to a two-layer fully connected network to obtain a scaling parameter; then a Sigmoid function is applied at the end of the two-layer fully connected network to scale the scaling parameter to the range of (0, 1), which is represented as

[0011]

[0012] wherein z is the output of the two-layer fully connected network, and a is the corresponding scaling parameter; then the average value of |x| is multiplied by the scaling parameter a to obtain a threshold value; the threshold value is represented as follows:

[0013] τ=α·average|x i,j,c |(2)

[0014] wherein τ is the threshold value, i, j and c are the width, height and channel index of the feature map x respectively; after obtaining the threshold value τ, the feature map x and the threshold value τ are subjected to a soft thresholding operation, for x, the absolute value of x less than τ is directly set to zero, otherwise the corresponding scaling processing is performed, a new output y is obtained, and the operation process is as follows:

[0015]

[0016] The new output y and the output y1 of the first convolution layer are added to obtain the final output Y of the basic module:

[0017] Y=y+y1(4)

[0018] S12: the various characterization data of the lithium battery are taken as a set X as the input of the model, and a new feature matrix X' is obtained after the shrinkage network:

[0019] X'=Θ(X)(5)

[0020] wherein Θ is a nonlinear mapping relationship of the entire shrinkage network.

[0021] Optionally, the S2 is specifically:

[0022] S21: the basic module decomposes the input feature F into F even and F odd two sub-features through splitting and interaction learning operations;

[0023] The splitting process down-samples the original sequence F into two sub-sequences F even and F odd by separating the odd and even elements; features are extracted from F even and F odd using different convolution kernels; interactive learning is adopted to realize mutual mapping transformation and information exchange;

[0024] The interactive learning is specifically:

[0025] Firstly, F even and F odd are projected into hidden states with two different one-dimensional convolution modules Φ and Ψ, and are converted into exponential form, and interact with F even and F odd in the form of element-related products, that is:

[0026]

[0027] Secondly, two scaled features F and F are further projected into two other hidden states using two other one-dimensional convolution modules ρ and η, and the final output is two updated sub-features F' even and F' odd by adding or subtracting F , that is:

[0028]

[0029] S22: The basic modules are arranged in the form of a binary tree to form an overall module, a total of L layers, and the number of basic modules in each layer is 2 l The input of a single module is a time series X', k = 1 or the feature vector output by the previous overall module k is the number of overall modules in the final structure of the next section, and τ is the dimension of the data, capturing short-term and long-term temporal dependencies in the time series; after passing through the L basic modules, a merging operation is performed to rearrange the elements in all sub-features and connect them into a new sequence representation, which is then added to the original time series through the remaining connections to generate a new sequence; finally, a simple fully connected network is used to decode the enhanced sequence representation to obtain a new feature vector

[0030] S23: Stack K overall modules to obtain the final structure of the network to obtain the prediction accuracy; wherein the input X k of each module is the output of the previous module The output of the last module is the final prediction result; in each overall module, an intermediate supervision is introduced, the output of each module is compared with the input of the module to obtain an error, and the error is sent to the total error for back propagation;

[0031] The error of each module is:

[0032]

[0033] The total error of the final structure is:

[0034]

[0035] The beneficial effects of the present application are:

[0036] From the perspective of the model, in order to solve the problem of noise interference on the model and the limited effective acceptance domain of the intermediate layer of the model, the noise data can be denoised by soft thresholding, the data with little relevance to the current task is reduced or zeroed, and the simple convolution and interaction network are used to extract the time sequence data features while increasing the connection between each sub-sequence, thereby improving the ability of the model to learn the correlation between the time sequence data, and completing the accurate estimation of the state of charge of the lithium battery.

[0037] Other advantages, objects, and features of the present application will be apparent from the following specification, and in some aspects, will be apparent to those skilled in the art from the following description, or can be learned by practice of the present application. The objects and other advantages of the present application can be achieved and obtained by the following description. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to make the objects, technical solutions and advantages of the present application clearer, the preferred detailed description of the present application will be combined with the drawings as follows, wherein:

[0039] Figure 1 The schematic diagram of the present application is shown in the figure. DETAILED DESCRIPTION

[0040] The embodiments of the present application are described below by specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the disclosure of the present specification. The present application can also be implemented or applied by different specific embodiments, and the details in the specification can be modified or changed based on different views and applications without departing from the spirit of the present application. It should be noted that the diagrams provided in the following examples only illustrate the basic concept of the present application in a schematic manner, and the following examples and features in the examples can be combined with each other without conflict.

[0041] Among them, the drawings are only used for illustrative description, and the representation is only a schematic diagram, not a physical diagram, and cannot be understood as a limitation on the present application; in order to better illustrate the embodiments of the present application, some components of the drawings will be omitted, enlarged or reduced, and do not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings can be omitted.

[0042] The same or similar reference numerals in the drawings of the embodiments of the present application correspond to the same or similar components; in the description of the present application, it should be understood that if the terms "upper", "lower", "left", "right", "front", "back" and the like indicate the orientation or positional relationship shown in the drawings, only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore the positional relationship described in the drawings is only used for illustrative description, and cannot be understood as a limitation on the present application, for those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0043] In view of the demand for high-precision prediction of lithium battery state of charge estimation, a lithium battery state of charge estimation method based on shrinkage and spatial interaction network is studied, which mainly includes two modules: a shrinkage network module with soft threshold and a spatial interaction network module with simple convolution. Specifically, the shrinkage network module with soft threshold is used to denoise the noise data and obtain a new time series feature matrix; the spatial interaction network module with simple convolution is responsible for learning the forward and backward dependency between the time series feature matrices, and outputs the final estimation value. The lithium battery state of charge estimation based on shrinkage and spatial interaction network is as shown in Figure 1 .

[0044] 1) Shrinkage network module with soft threshold

[0045] In this study, the current, voltage, temperature and sampling time series of lithium battery are used as model input, or one or more of them can be freely combined as input, and the network inserts a soft threshold as a nonlinear conversion layer into a deep architecture based on a residual network to eliminate unimportant features and achieve the purpose of denoising, obtaining a new feature matrix as the input of the next module.

[0046] The content of the algorithm is as follows:

[0047] ① Basic module of shrinkage network

[0048] The two-dimensional data first passes through three layers of ordinary two-dimensional convolutional layers, where the channel number is set to 1, to obtain a new feature map x, and the absolute value of the feature map x is globally averaged to obtain a one-dimensional vector. Then, the one-dimensional vector is propagated to the second layer of the fully connected network to obtain a scaling parameter. Then, the Sigmoid function is applied at the end of the second layer of the fully connected network to scale the scaling parameter to the range of (0, 1), which can be expressed as

[0049]

[0050] where z is the output of the second layer of the fully connected network, and a is the corresponding scaling parameter. Then, the average value of |x| is multiplied by the scaling parameter a to obtain the threshold value. The reason for this processing is that the model needs a soft threshold value, and the threshold value is a positive number, and the number cannot be too large. If the size of the number is greater than the maximum value in the feature map, it will cause all the data to be zero after soft thresholding. In summary, the threshold value used in the basic module is expressed as follows:

[0051] τ=α·average|x i,j,c |(2)

[0052] where τ is the threshold value, i, j and c are the width, height and channel index of the feature map x respectively. After obtaining the threshold value τ, the feature map x and the threshold value τ are subjected to soft thresholding operation. For x, the absolute value of x is directly set to zero if it is less than τ, otherwise it is subjected to corresponding scaling processing to obtain a new output y, and the operation process is as follows:

[0053]

[0054] Finally, the new output y is added to the output y1 of the first layer of the convolutional layer to obtain the final output Y of the basic module:

[0055] Y=y+y1(4)

[0056] ②Overall structure of the shrinkage network

[0057] The overall structure of the shrinkage network studied is stacked with many basic modules, so that the noise-related characteristics can be gradually reduced. Another advantage of the shrinkage network studied is that the threshold value is automatically learned in the deep architecture, rather than manually set by an expert. The various characterization data of the lithium battery are regarded as a set X as the input of the model, and finally a new feature matrix X' is obtained after passing through the shrinkage network:

[0058] X'=Θ(X)(5)

[0059] where Θ is the nonlinear mapping relationship of the entire shrinkage network.

[0060] 2) Spatial interaction network module with simple convolution

[0061] The data input used in this module is the feature matrix X' after the data has been denoised in the previous module. Although the feature matrix has been transformed, it still has the dependencies between time-series data. In this module, the spatial interaction network will fully learn the dependencies between data and finally accurately output the state of charge of the lithium battery.

[0062] The algorithm is as follows:

[0063] ① Basic Modules of Spatial Interaction Network

[0064] The basic module decomposes the input feature F into Fi through splitting and interactive learning operations. even and F odd Two sub-features.

[0065] The splitting process downsamples the original sequence F into two subsequences F by separating the odd and even elements. even and F odd Although the amount of data per sequence is reduced, most of the information from the original sequence is preserved. Next, different convolution kernels are used to extract data from F... even and F odd Features are extracted from the kernels. Since the kernels are independent, the features extracted from them will contain different but valuable temporal relationships, thus enhancing the representational power. However, there is potential information loss in the previous downsampling. This study proposes an interactive learning strategy that uses mutual mapping transformations to achieve information exchange. The interactive learning process is described below:

[0066] First, F even and F odd They are projected onto hidden states with two different one-dimensional convolutional moduli Φ and Ψ, respectively, and transformed into an exponential form of e, and then compared with F. even and F odd The product interaction related to the elements, namely:

[0067]

[0068] Secondly, using two additional one-dimensional convolutional moduli ρ and η, the two scaled features are... and Further projection onto the other two hidden states, through and The final output of addition or subtraction is two updated sub-features F. e ' ven and F o ' dd ,Right now:

[0069]

[0070] ② Spatial Interaction Network Overall Module

[0071] After the structure of the basic module is determined, the basic modules are arranged in a binary tree manner to form an overall module, a total of L layers, and the number of basic modules in each layer is 2 l The input of a single module is the time series X(k=1) or the feature vector of the output of the previous overall module (k is the number of overall modules in the final structure of the next section, and τ is the dimension of the data), the information of the previous layers will gradually accumulate, that is, the features of the deeper layers will contain additional more detailed time information transmitted by the shallower layers. In this way, we can capture short-term and long-term temporal dependencies in the time series. After passing through the L layers of basic modules, a merging operation is performed to rearrange the elements in all sub-features and connect them into a new sequence representation, which is then added to the original time series through the remaining connections to generate a new sequence with enhanced predictability. Finally, a simple fully connected network is used to decode the enhanced sequence representation to obtain a new feature vector

[0072] ③Spatial interaction network final structure and error function

[0073] The study stacks K overall modules to obtain the final structure of the network, which can obtain better prediction accuracy. The input X k of each module is the output of the previous module The output of the last module is the final prediction result. An intermediate supervision is introduced in each overall module, and the output of each module is compared with the input of the module to obtain the error, which is sent to the total error for backpropagation.

[0074] The error of each module is:

[0075]

[0076] The total error of the final structure is:

[0077]

[0078] Finally, it should be pointed out that the above examples are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and all should be covered in the scope of the claims of the present application.

Claims

1. A method for lithium battery state-of-charge estimation based on shrinkage and spatial interaction networks, characterized by: The method comprises the following steps: S1: using one or more of lithium battery current, voltage, temperature and sampling time sequence as model input, inserting soft threshold as a nonlinear conversion layer into deep architecture based on residual network to eliminate unimportant features, achieving the purpose of denoising, and obtaining a new feature matrix as the input of the next module; S2: using the feature matrix denoised by the denoising module in S1 as data input, a spatial interaction network learns the dependency between data, and finally outputs the state of charge of the lithium battery; specifically: S21: the base module decomposes the input feature into and two sub-features through splitting and interactive learning operations; The splitting process separates the original sequence into two sub-sequences by separating the odd and even elements Down-sampling into two sub-sequences and Extract features from and using different convolution kernels; and mutual mapping transformation is performed by using interactive learning to realize information exchange; The interactive learning is specifically: First, and are projected to hidden states with two different one-dimensional convolutional modules and are converted to exponential form and interact with and and element-wise multiplication interaction, i.e.: （6） Secondly, use another two one-dimensional convolution modules and to project the two scaled features and to another two hidden states, through and , add or subtract the final output is two updated sub-features and , namely: （7） S22: The basic modules are combined into integral modules according to the arrangement of binary tree, and the total number of integral modules is layers, and the number of basic modules in each layer is , is the number of integral modules in the final structure of the next subsection, and the input of a single module is a time series, or the feature vector output by the previous integral module , , is the dimension of the data, capturing short-term and long-term temporal dependencies in the time series; after passing through the basic modules of the layer, a merging operation is performed to rearrange the elements in all sub-features and connect them into a new sequence representation, which is then added to the original time series through the remaining connections to generate a new sequence; finally, a simple fully connected network is used to decode the enhanced sequence representation to obtain a new feature vector ; S23: Will The final structure of the network is obtained by stacking the overall modules, which yields the prediction accuracy; the input of each module is... It is the output of the previous module. The output of the last module is the final prediction result. An intermediate supervisor is introduced in each overall module. The output of each module is compared with the input of that module to obtain the error, and the error is fed into the total error for backpropagation. The error of each module is: （8） The total error of the final structure is: （9）。 2. The method of claim 1, wherein: The S1 is specifically: S11: The two-dimensional data first goes through three ordinary two-dimensional convolutional layers, where the number of channels is set to 1, to obtain a new feature map. , for feature mapping The absolute value of the vector is used for global average pooling to obtain a one-dimensional vector; this one-dimensional vector is then propagated into a two-layer fully connected network to obtain a scaling parameter; finally, it is applied at the end of the two-layer fully connected network. The function that scales the scaling parameter to the range (0,1) is represented as: (1) wherein is the output of the two-layer fully connected network, are the respective scaling parameters; the scaling parameters are then multiplied by the average value of to obtain the threshold value; the threshold value used is expressed as follows: (2) wherein, is a threshold value, , and are the width, height and channel index of the feature map respectively; the threshold value is obtained after and the threshold value are soft-thresholded, for the absolute value of is less than directly set to zero, otherwise corresponding scaling is performed, to obtain a new output , the operation process is as follows: (3) add the output of the first layer of convolutional layers to the output of the second layer of convolutional layers to obtain the final output of the base module : (4) S12: Take each characterization data of the lithium battery as a set X as the input of the model, and obtain a new feature matrix after shrinking the network : （5） wherein is a non-linear mapping relationship of the entire shrink network.

Images