Battery residual life prediction method and system based on stack type self-coding

Abstract

The invention discloses a battery residual life prediction method and system based on stack type self-coding, and the method comprises the steps: inputting collected battery charging and discharging data into a preset stack type self-coding model, carrying out the extraction of linear features and nonlinear features, and obtaining a plurality of health factors of a battery; generating a corresponding weight coefficient for each health factor based on an actual prediction demand according to an excitation unit of the stack type self-encoding model; and performing feature fusion on the health factor through the weight coefficient, and predicting the performance degradation degree of the battery to obtain a residual life prediction result of the battery. According to the method, the battery performance degradation information is mined from the battery charging and discharging data, multi-dimensional features are fused, and the remaining life of the battery is accurately predicted.

Description

Technical Field

[0001] This application belongs to the technical field of intelligent battery energy management, and specifically relates to a method and system for predicting the remaining life of a battery based on stack encoding. Background Art

[0002] With the development of renewable energy and energy storage technologies, intelligent battery energy management systems have gradually become an important part of power systems and smart grids. Traditional battery energy management systems mostly adopt rule-based management methods to achieve battery energy management through fixed parameters and logical judgments. However, this method has problems of inflexible response and insufficient accuracy when dealing with complex battery operating states, environmental factor impacts, and battery state changes, and it is difficult to adapt to complex and changeable working conditions. In order to improve the performance of battery energy management systems, deep learning algorithms have gradually been applied in recent years to predict battery performance for more effective management of intelligent batteries.

[0003] Currently, most studies characterize the degree of battery performance degradation and the health state of batteries by defining capacity or internal resistance as health indicators (HIs), and use principal component analysis (PCA) to predict the remaining life of batteries by extracting feature vectors. However, various HIs exhibit different characteristics in local mutations of the lithium battery capacity degradation curve, conflicting with each other, which affects the prediction accuracy of the remaining life of lithium batteries. Moreover, PCA can only perform linear transformations, and when complex multi-dimensional data has a non-linear structure, PCA cannot comprehensively extract the essential features of the data, so PCA has poor adaptability and can only be used in specific scenarios. Summary of the Invention

[0004] This application proposes a method and system for predicting the remaining life of a battery based on stack encoding, which can accurately predict the remaining life of a battery by mining battery performance degradation information from battery charge and discharge data and fusing multi-dimensional features.

[0005] The first aspect of this application provides a method for predicting the remaining life of a battery based on stack encoding, and the method includes:

[0006] Input the collected battery charge and discharge data into a preset stack autoencoder model to extract linear and non-linear features, and obtain several health indicators of the battery;

[0007] According to the excitation unit of the stack autoencoder model, generate corresponding weight coefficients for each health indicator based on actual prediction requirements;

[0008] Perform feature fusion on the health indicators through the weight coefficients, predict the degree of battery performance degradation, and obtain the prediction result of the remaining life of the battery.

[0009] The above solution first performs preliminary feature extraction on the collected battery charge and discharge data, unifies the extraction of linear and non-linear features, enables the method to be applied to various scenarios, and makes the extracted feature data not limited to linear or non-linear structures; corresponding health factors are generated according to the extracted features, and each health factor can characterize the performance of the battery. According to the influence degree of each health factor on predicting the remaining life of the battery, weight assignment is performed for each health factor to highlight the importance of important health factors in the evolution process of the battery performance degradation degree, so as to increase the performance ability of important health factors in feature fusion and further improve the prediction accuracy. Moreover, the weight coefficient can also increase the discrimination speed of important health factors in the fusion process, improve the operation efficiency, and can more efficiently predict the current battery performance.

[0010] In a possible implementation method of the first aspect, the collected battery charge and discharge data is input into a preset stacked autoencoder model to extract linear and non-linear features, and several health factors of the battery are obtained. Specifically:

[0011] According to the preset battery degradation characteristics, the first threshold number of feature parameters are screened out from the battery charge and discharge data through the stacked autoencoder model;

[0012] The linear feature parameters and the non-linear feature parameters are normalized to generate the first threshold number of the health factors, and the health factors are integrated into a feature vector of the first threshold dimension.

[0013] In the above solution, several feature parameters are extracted from the battery charge and discharge data, and each feature parameter is related to the characterization of the battery performance to a certain extent. Since these feature parameters are linear and non-linear, in order to facilitate subsequent unified calculation, these feature parameters are normalized to obtain a multi-dimensional feature vector, which is convenient for subsequent feature fusion and extraction of deeper features.

[0014] In a possible implementation method of the first aspect, the feature parameters include: initial voltage drop value, battery discharge power, average battery temperature during discharge, constant current charging time, ratio of constant current charging time to constant voltage charging time, and discharge cut-off voltage.

[0015] In a possible implementation method of the first aspect, according to the excitation unit of the stacked autoencoder model, corresponding weight coefficients are generated for each of the health factors based on actual prediction requirements. Specifically:

[0016] According to actual prediction requirements, the influence value of each of the health factors on predicting the battery performance degradation degree is determined;

[0017] The excitation unit assigns the corresponding weight coefficient of each size to each of the health factors according to the influence value;

[0018] Among them, the excitation unit generates the weight coefficient by the gradient descent method based on the principle of minimum mean square error.

[0019] The above solution assigns different weights to each health factor based on the role of each health factor in predicting the degree of performance degradation of the battery, ensuring that the prediction result can meet the actual prediction requirements. Health factors with high weights can more significantly exhibit features during the feature fusion process, and health factors with low weights will not cause conflicts during feature fusion and affect the data prediction result.

[0020] In a possible implementation method of the first aspect, the weight coefficient is generated by the gradient descent method based on the principle of minimum mean square error. Specifically:

[0021] The weight coefficient is generated by the following formula:

[0022]

[0023] (w 1 , b 1 , b 2 ) ← arg min(Loss(w 1 , b 1 , b 2 ));

[0024]

[0025]

[0026] In the formula, w 1 is the matrix composed of the weight coefficients, b 1 , b 2 are the bias coefficients of the stacked autoencoder model, Loss is the loss function of the stacked autoencoder model, is the i-th output data of the stacked autoencoder model, x i is the i-th input data of the stacked autoencoder model, m is the number of health factors, Δw i,j is the change value of the j-th weight coefficient of the i-th dimensional feature vector, Δb 1i , Δb 2i are the change values of the bias coefficients of the i-th dimensional feature vector, is the transpose of w 1 .

[0027] In a possible implementation method of the first aspect, the health factors are subjected to feature fusion by the weight coefficients, and the degree of performance degradation of the battery is predicted to obtain a prediction result of the remaining life of the battery. Specifically:

[0028] Adjust the proportion when performing feature fusion on the health factors according to the magnitude of the weight coefficients, and perform feature fusion on the weight coefficients with the goal of evolving the degree of performance degradation of the battery to obtain the degree of performance degradation of the current battery;

[0029] Obtain the remaining life prediction result according to the degree of performance degradation.

[0030] In a possible implementation method of the first aspect, it further includes:

[0031] Set the number of times of feature extraction and the number of times of feature fusion of the stacked autoencoder model according to actual prediction requirements.

[0032] The above solution reduces the computational complexity by controlling the number of times of feature extraction and the number of times of feature fusion. At the same time, setting multiple times of feature extraction and feature fusion can also improve the accuracy of data prediction.

[0033] In a possible implementation method of the first aspect, the battery charge and discharge data collected is specifically:

[0034] In several charge and discharge cycles, collect the first state data of the battery before each charge and discharge cycle and the second state data after each charge and discharge cycle to obtain the battery charge and discharge data.

[0035] The second aspect of the present application provides a battery remaining life prediction system based on a stacked autoencoder. The system includes: a health factor extraction module, a weight coefficient assignment module, and a remaining life prediction module;

[0036] Among them, the health factor extraction module is used to input the collected battery charge and discharge data into a preset stacked autoencoder model to extract linear features and non-linear features to obtain several health factors of the battery;

[0037] The weight coefficient assignment module is used to generate corresponding weight coefficients for each of the health factors based on the excitation unit of the stacked autoencoder model according to actual prediction requirements;

[0038] The remaining life prediction module is used to perform feature fusion on the health factors by the weight coefficients, predict the degree of performance degradation of the battery, and obtain a prediction result of the remaining life of the battery.

[0039] A third aspect of the present application provides a terminal device, which includes: a terminal device including a processor and a memory, where the memory stores a computer program, and when the processor executes the computer program, the steps of a battery remaining life prediction method based on stacked auto-encoder according to any one of the embodiments of the present application are implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions of the present application, the accompanying drawings required for the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other accompanying drawings can be obtained based on these drawings without creative efforts.

[0041] Figure 1 is a specific flowchart of a battery remaining life prediction method based on stacked auto-encoder provided by an embodiment of the present application;

[0042] Figure 2 is a battery charge and discharge schematic diagram of a battery remaining life prediction method based on stacked auto-encoder provided by an embodiment of the present application;

[0043] Figure 3 is a battery aging analysis diagram of a battery remaining life prediction method based on stacked auto-encoder provided by an embodiment of the present application;

[0044] Figure 4 is a network architecture diagram of a battery remaining life prediction method based on stacked auto-encoder provided by an embodiment of the present application;

[0045] Figure 5 is a model structure diagram of a battery remaining life prediction method based on stacked auto-encoder provided by an embodiment of the present application;

[0046] Figure 6 is a structure diagram of a battery remaining life prediction system based on stacked auto-encoder provided by an embodiment of the present application;

[0047] Figure 7 is a structure diagram of a terminal device provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0048] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts belong to the scope of protection of the present application.

[0049] It should be understood that the step numbers used in the text are only for convenience of description and not for limiting the order of execution of the steps.

[0050] The first embodiment

[0051] Currently, most studies usually define the battery capacity or internal resistance as the health indicator (HI) to measure the battery performance. However, the characterization ability of a single HI for battery performance degradation varies, and it cannot accurately characterize the current battery performance in all scenarios. Given that the battery capacity is not easily obtained directly, how to accurately predict the remaining life of the battery through the charge and discharge data of the battery is the main technical problem to be solved in the embodiments of this application.

[0052] As Figure 1 shown, Figure 1 FIG. is a schematic flow chart of a method for predicting the remaining life of a battery based on a stacked autoencoder provided by an embodiment of this application. The method for predicting the remaining life of a battery based on a stacked autoencoder in this embodiment includes steps S1 to S3, which are described in detail as follows:

[0053] Step S1: Input the collected battery charge and discharge data into a preset stacked autoencoder model to extract linear features and non-linear features, and obtain several health factors of the battery.

[0054] In the embodiments of this application, the remaining life of a lithium-ion battery is predicted. Since the capacity of a lithium-ion battery is not easily obtained directly, potential degradation information is found through the analysis of the historical experimental data of the battery.

[0055] Exemplarily, the battery data in the current and previous states are directly collected from several charge and discharge cycles of the lithium-ion battery. Specifically: the first state data of the battery before each charge and discharge cycle and the second state data after each charge and discharge cycle are collected to obtain the battery charge and discharge data; these state data include current, voltage, temperature, internal resistance, etc. during charge and discharge.

[0056] Then, the obtained battery charge and discharge data are input into the trained stacked autoencoder model, and potential battery degradation information is found through data analysis to estimate the remaining life of the battery.

[0057] A lithium-ion battery is composed of multiple components, such as a carbonaceous anode that can embed lithium ions, a metal oxide cathode that can de-embed lithium ions, a lithium salt electrolyte, and a polymer separator that only allows lithium ions to pass through and separates the anode and cathode. As Figure 2 shown, it shows the operation process of a lithium-ion battery during charge and discharge. During the charging process, lithium ions move from the positive electrode through the separator to the negative electrode; during the discharging process, lithium ions move from the negative electrode through the separator to the positive electrode.

[0058] To accurately predict the remaining life of a lithium-ion battery, it is necessary to evolve the degree of battery degradation. The degradation of lithium-ion batteries mainly includes: loss of lithium inventory and loss of active materials. The loss of lithium inventory can be understood as the consumption of available lithium ions by side reactions. When a lithium-ion battery is charged for the first time, a solid electrolyte interface film, also known as the SEI film (Solid Electrolyte Interphase), is formed. The loss of active lithium is mainly caused by the growth of the SEI film, the degradation of the SEI film, the degradation of the electrolyte, and the formation of dendrites. External factors such as charge and discharge time, high temperature, and high voltage can all affect the SEI film. The loss of active materials can be understood as the loss of the space for storing lithium ions, which is specifically divided into the loss of positive active substances (Loss of positive active substances, LOP) and the loss of negative active substances (Loss of negative active substances, LON).

[0059] To better demonstrate the reasons for the degradation of lithium-ion batteries, Figure 3 a battery aging analysis diagram is provided, which gives different reasons for different degradation modes and analyzes the external and internal impacts respectively.

[0060] To accurately predict the degradation law and degree of the battery, the embodiment of the present application adopts a stacked autoencoder model based on the autoencoder algorithm and excitation unit.

[0061] After the battery charge and discharge data is input into the preset stacked autoencoder model, the model first extracts the data, selects several characteristic parameters that meet the preset battery degradation characteristics, and then normalizes and integrates these characteristic parameters into multi-dimensional data.

[0062] Exemplarily, six data, namely the initial voltage drop value, battery discharge power, average battery temperature during the discharge stage, constant current charging time, ratio of constant current charging time to constant voltage charging time, and discharge cut-off voltage, are respectively extracted from the battery charge and discharge data. Then, these 6 data are preprocessed and normalized to obtain the corresponding six health factors, and each health factor can characterize the battery performance degradation state to varying degrees. Then, these health factors are integrated into six-dimensional data to obtain a six-dimensional feature vector.

[0063] Step S2, according to the excitation unit of the stacked autoencoder model, generate corresponding weight coefficients for each of the health factors based on the actual prediction requirements.

[0064] In the embodiment of the present application, the excitation unit is used to generate corresponding weight coefficients for the health factors.

[0065] Optionally, the excitation unit adopted in the embodiments of the present application is SENet, which is short for Squeeze and Excitation Networks. The core idea of SENet is to enhance the model's recognition ability of important features by adaptively recalibrating the channel feature responses. In the embodiments of the present application, it is to improve the data performance ability of the stacked autoencoder model in the feature extraction process and further improve the prediction accuracy.

[0066] SENet obtains the normalized feature information, adaptively generates corresponding weight coefficients, and uses these weight coefficients to enhance the important features of the data and suppress the features that have less effect on the current task, thereby strengthening the performance ability of the important feature information in the extraction process. The feature information after weight rescaling is input into the feature fusion module of the stacked autoencoder model for feature fusion, making it easier for the feature fusion module to distinguish the feature information of each channel and improving the operation efficiency. Compared with a single deep neural network, a deep neural network containing an excitation unit can extract deeper feature expressions more efficiently.

[0067] Based on the actual prediction requirements, the model is trained so that the excitation unit can judge the influence of each health factor on the performance degradation degree of the predicted battery, and then generate corresponding weight coefficients.

[0068] Optionally, the embodiments of the present application select sigmoid as the activation function of the excitation unit.

[0069] Step S3, perform feature fusion on the health factors through the weight coefficients, predict the performance degradation degree of the battery, and obtain the predicted result of the remaining life of the battery.

[0070] In the embodiments of the present application, the feature vector containing the health factors with weights is input into the input layer of the feature fusion module. In the input layer, data dimensionality reduction is performed on the feature vector, and the proportion of the health factors during feature fusion is adjusted according to the magnitude of the weight coefficients. With the goal of evolving the performance degradation degree of the battery, feature fusion is performed on the weight coefficients to obtain the performance degradation degree of the current battery.

[0071] Optionally, the feature fusion module adopted in the embodiments of the present application is Autoencoder, abbreviated as AE. AE consists of an input layer, a hidden layer, and an output layer network, and is the basis for constructing a stacked autoencoder model. AE was initially used to solve the dimensionality reduction problem of neural networks in representation learning, and it performs unsupervised learning through the reconstruction error. The autoencoder algorithm includes two parts: encoding and decoding. The main role of the encoder is to extract features from the original input data through an encoding function to obtain the hidden layer. The decoder is the opposite of the encoder, and it reconstructs the data from the feature expression of the hidden layer through a decoding function.

[0072] To improve the data processing accuracy, the stacked autoencoder model can include multiple feature fusion modules, and the deep neural network composed of these multiple feature fusion modules is called SAE. Figure 4 The network architecture diagrams of AE and SAE are provided. As shown in the figure, the left side is the schematic diagram of the AE network structure, and the right side is the schematic diagram of the SAE network structure. The difference between AE and SAE is that SAE stacks multiple AEs, takes the output of the previous AE as the input of the next AE for feature extraction and fusion. SAE can represent both linear transformation and nonlinear transformation, and can more completely extract the degradation feature information of the battery. Therefore, in the process of extracting health factors, it is not necessary to consider whether the extracted data is linear or nonlinear, and these data can be uniformly processed in SAE.

[0073] An activation unit is added before each feature fusion module to assign weights to the data. Therefore, each feature fusion module will perform feature fusion according to the weight coefficients. Among them, each feature fusion module sequentially includes an input layer, a hidden layer, and an output layer.

[0074] Specifically, the input layer passes the n-dimensional feature vector of the health factor with weights to the hidden layer, performs data fusion through decoding to obtain an m-dimensional feature vector, and passes the m-dimensional feature vector to the output layer to generate a new n-dimensional feature vector. Calculate the difference between the n-dimensional feature vector of the input layer and the n-dimensional feature vector of the output layer through the loss function to obtain the error of the model, and continuously train the stacked autoencoder model with the goal of minimizing the error to obtain a high-precision prediction result.

[0075] Therefore, SAE composed of multiple feature fusion modules extracts the deep feature expressions of data through layer-by-layer encoding training. When dealing with complex and non-linear high-dimensional data, it has stronger feature extraction ability than AE with only a single hidden layer, can reduce the dimension of the input data, and adaptively and deeply extract features.

[0076] The loss function, the specific expression is:

[0077]

[0078] In the formula, w 1 is the matrix composed of the weight coefficients, b 1 , b 2 are the bias coefficients of the stacked autoencoder model, Loss is the loss function of the stacked autoencoder model, is the i-th output data of the stacked autoencoder model, x i is the i-th input data of the stacked autoencoder model, and m is the number of health factors.

[0079] Among them, the first feature fusion module can be expressed as:

[0080] H(H) = f(w 1 H + b 1 );

[0081]

[0082] In the formula, H is, f is the feature vector obtained by the hidden layer in the first feature fusion module, w 1 is the weight coefficient, b 1 , b 2 is the bias coefficient, is the transpose of w 1 .

[0083] The first feature fusion module transmits the output n-dimensional feature vector after fusion to the excitation unit, and updates the weight coefficient of the n-dimensional feature vector according to the principle of minimum mean square error by the gradient descent method. The specific expression is:

[0084] (w 1 , b 1 , b 2 ) ← arg min(Loss(w 1 , b 1 , b 2 ));

[0085]

[0086] In the formula, w 1 is the weight coefficient, b 1 , b 2 is the bias coefficient, Loss is the loss function, Δw i,j is the change value of the j-th weight coefficient of the i-th dimensional feature vector, Δb 1i , Δb 2i is the change value of the bias coefficient of the i-th dimensional feature vector.

[0087] According to the updated weight coefficient, feature fusion of the n-dimensional feature vector is performed in the second feature fusion module to achieve deeper feature extraction of the data and improve the prediction accuracy.

[0088] Although setting multiple feature fusion modules in the stacked autoencoder model can improve the feature extraction and fusion effect, it will increase the computational complexity of the model. Therefore, in the actual operation process, it is also necessary to adjust the number of feature fusion modules according to the actual prediction requirements to set the feature extraction times and feature fusion times of the stacked autoencoder model.

[0089] Implementing the embodiments of the present application has the following beneficial effects:

[0090] In the embodiments of the present application, the collected battery charge and discharge data is first subjected to preliminary feature extraction, and linear and non-linear features are extracted uniformly, so that the method can be applied to various scenarios, and the extracted feature data is not limited to linear or non-linear structures; corresponding health factors are generated according to the extracted features, and each health factor can characterize the performance of the battery. According to the influence degree of each health factor on predicting the remaining life of the battery, weight values are assigned to each health factor to highlight the importance of important health factors in the process of evolving the degree of battery performance degradation, so as to increase the performance of important health factors in feature fusion and further improve the prediction accuracy. Moreover, the weight coefficient can also increase the discrimination speed of important health factors in the fusion process, improve the operation efficiency, and can more efficiently predict the current battery performance.

[0091] Second Embodiment

[0092] Furthermore, in order to implement the battery remaining life prediction system based on stacked autoencoders corresponding to the above method embodiments to achieve the corresponding functions and technical effects, Figure 6 A structural diagram of a battery remaining life prediction system based on stacked autoencoders is provided. For the sake of illustration, only the parts related to this embodiment are shown. The battery remaining life prediction system based on stacked autoencoders provided by the embodiments of the present application includes:

[0093] A health factor extraction module 201, configured to input the collected battery charge and discharge data into a preset stacked autoencoder model to extract linear features and non-linear features, and obtain a plurality of health factors of the battery.

[0094] In the embodiments of the present application, according to the preset battery degradation characteristics, the stacked autoencoder model is used to screen out the first threshold number of feature parameters from the battery charge and discharge data; the linear feature parameters and the non-linear feature parameters are normalized to generate the first threshold number of the health factors, and the health factors are integrated into a feature vector of the first threshold dimension.

[0095] A weight coefficient assignment module 202, configured to generate corresponding weight coefficients for each of the health factors based on the excitation unit of the stacked autoencoder model according to the actual prediction requirements.

[0096] In the embodiments of the present application, according to the actual prediction requirements, the influence value of each of the health factors on predicting the degree of battery performance degradation is determined.

[0097] The excitation unit assigns the corresponding weight coefficients of corresponding magnitudes to each of the health factors through the influence value.

[0098] Among them, the excitation unit generates the weight coefficient by the gradient descent method based on the principle of minimum mean square error.

[0099] The remaining life prediction module 203 is configured to perform feature fusion on the health factors through the weight coefficient, predict the degree of performance degradation of the battery, and obtain the remaining life prediction result of the battery.

[0100] In the embodiment of the present application, the proportion of feature fusion of the health factors is adjusted according to the magnitude of the weight coefficient, and the weight coefficient is subjected to feature fusion with the goal of evolving the degree of performance degradation of the battery to obtain the degree of performance degradation of the current battery;

[0101] According to the degree of performance degradation, the remaining life prediction result is obtained.

[0102] In some embodiments, the health factor extraction module 201 further includes:

[0103] In the embodiment of the present application, the remaining life of the lithium-ion battery is predicted. Since the capacity of the lithium-ion battery is not directly available, potential degradation information is found through the analysis of the historical experimental data of the battery.

[0104] Exemplarily, the battery data in the current and previous states are directly collected from several charge and discharge cycles of the lithium-ion battery. Specifically, the first state data of the battery before each charge and discharge cycle and the second state data after each charge and discharge cycle are collected to obtain the charge and discharge data of the battery; these state data include current, voltage, temperature, internal resistance, etc. during charge and discharge.

[0105] Then, the obtained battery charge and discharge data are input into the trained stacked autoencoder model, and potential battery degradation information is found through data analysis to estimate the remaining life of the battery.

[0106] The lithium-ion battery is composed of multiple components, such as a carbonaceous anode that can embed lithium ions, a metal oxide cathode that can de-embed lithium ions, a lithium salt electrolyte, and a polymer separator that only allows lithium ions to pass through and separates the anode and cathode.

[0107] To accurately predict the remaining life of a lithium-ion battery, it is necessary to evolve the degree of battery degradation. The degradation of lithium-ion batteries is mainly divided into: loss of lithium inventory and loss of active materials. The loss of lithium inventory can be understood as the consumption of available lithium ions by side reactions. When a lithium-ion battery is charged for the first time, a solid electrolyte interface film, also known as the SEI film (Solid Electrolyte Interphase), is formed. The loss of active lithium is mainly caused by the growth of the SEI film, the degradation of the SEI film, the degradation of the electrolyte, and the formation of dendrites. External factors such as charge and discharge time, high temperature, and high voltage can all affect the SEI film. The loss of active materials can be understood as the loss of the space for storing lithium ions, which is specifically divided into the loss of positive active substances (Loss of positive active substances, LOP) and the loss of negative active substances (Loss of negative active substances, LON).

[0108] In order to accurately predict the degradation law and degree of the battery, the embodiments of this application adopt a stacked autoencoder model based on the autoencoder algorithm and excitation unit.

[0109] After the battery charge and discharge data is input into the preset stacked autoencoder model, the model first extracts the data, selects several characteristic parameters that meet the preset battery degradation characteristics, and then normalizes and integrates these characteristic parameters into multi-dimensional data.

[0110] Exemplarily, six data, namely the initial voltage drop value, battery discharge power, average battery temperature during the discharge stage, constant current charging time, ratio of constant current charging time to constant voltage charging time, and discharge cut-off voltage, are respectively extracted from the battery charge and discharge data. Then, these 6 data are preprocessed and normalized to obtain the corresponding six health factors, and each health factor can characterize the battery performance degradation state to varying degrees. Then, these health factors are integrated into six-dimensional data to obtain a six-dimensional feature vector.

[0111] In some embodiments, the weight coefficient assignment module 202 further includes:

[0112] Generate corresponding weight coefficients for the health factors through the excitation unit.

[0113] Optionally, the excitation unit adopted in the embodiments of this application is SENet, which is short for Squeeze and Excitation Networks. The core idea of SENet is to adaptively recalibrate the channel feature response to enhance the model's ability to identify important features. In the embodiments of this application, it is to improve the data performance ability of the stacked autoencoder model during the feature extraction process and further improve the prediction accuracy.

[0114] SENet obtains the normalized feature information, adaptively generates the corresponding weight coefficients, and uses the weight coefficients to enhance the important features of the data and suppress the features that have little effect on the current task, thereby strengthening the performance ability of the important feature information in the extraction process. The feature information after weight rescaling is input into the feature fusion module of the stacked autoencoder model for feature fusion, making it easier for the feature fusion module to distinguish the feature information of each channel and improving the operation efficiency. Compared with a single deep neural network, a deep neural network containing an excitation unit can extract deeper feature expressions more efficiently.

[0115] Based on the actual prediction requirements, the model is trained so that the excitation unit can judge the influence of each health factor on the performance degradation degree of the predicted battery, and then generate the corresponding weight coefficients.

[0116] Optionally, the sigmoid function is selected as the activation function of the excitation unit in the embodiments of the present application.

[0117] In some embodiments, the remaining useful life prediction module 203 further includes:

[0118] The feature vector containing the health factors with weights is input into the input layer of the feature fusion module. In the input layer, the data of the feature vector is dimensionally reduced, and the proportion of the health factors during feature fusion is adjusted according to the magnitude of the weight coefficients. With the goal of evolving the performance degradation degree of the battery, the weight coefficients are subjected to feature fusion to obtain the performance degradation degree of the current battery.

[0119] Optionally, the feature fusion module adopted in the embodiments of the present application is Autoencoder, abbreviated as AE. AE consists of an input layer, a hidden layer, and an output layer network, and is the basis for constructing a stacked autoencoder model. AE was initially used to solve the dimensionality reduction problem of neural networks in representation learning, and it performs unsupervised learning through the reconstruction error. The autoencoder algorithm includes two parts: encoding and decoding. The main role of the encoder is to extract features from the original input data through the encoding function to obtain the hidden layer. The decoder is the opposite of the encoder, and it reconstructs the data of the feature expression of the hidden layer through the decoding function.

[0120] To improve the data processing accuracy, the stacked autoencoder model can include multiple feature fusion modules. The deep neural network composed of these multiple feature fusion modules is called SAE. The difference between AE and SAE is that SAE stacks multiple AEs, takes the output of the previous AE as the input of the next AE for feature extraction and fusion. SAE can represent both linear transformations and non-linear transformations, and can extract the degradation feature information of the battery more completely. Therefore, when extracting the health factor, it is not necessary to consider whether the extracted data is linear or non-linear, and SAE can uniformly process these data.

[0121] An excitation unit is added before each feature fusion module to assign weights to the data. Therefore, each feature fusion module will perform feature fusion according to the weight coefficients. Among them, each feature fusion module sequentially includes an input layer, a hidden layer, and an output layer.

[0122] Specifically, the input layer transfers the n-dimensional feature vector of the health factor with weights to the hidden layer, performs data fusion through decoding to obtain an m-dimensional feature vector, and transfers the m-dimensional feature vector to the output layer to generate a new n-dimensional feature vector. Calculate the difference between the n-dimensional feature vector of the input layer and the n-dimensional feature vector of the output layer through the loss function to obtain the error of the model, and continuously train the stacked autoencoder model with the goal of minimizing the error to obtain a high-precision prediction result.

[0123] Therefore, SAE composed of multiple feature fusion modules extracts the deep feature expressions of data through layer-by-layer encoding training. When dealing with complex and non-linear high-dimensional data, it has stronger feature extraction ability than AE with only a single hidden layer, can reduce the dimension of the input data, and adaptively and deeply extract features.

[0124] The specific expression of the loss function is:

[0125]

[0126] In the formula, w 1 is the matrix composed of the weight coefficients, b 1 , b 2 are the bias coefficients of the stacked autoencoder model, Loss is the loss function of the stacked autoencoder model, is the i-th output data of the stacked autoencoder model, x i is the i-th input data of the stacked autoencoder model, and m is the number of health factors.

[0127] Among them, the first feature fusion module can be expressed as:

[0128] H(H) = f(w 1 H + b1 )

[0129]

[0130] In the formula, H is, f is the feature vector obtained from the hidden layer in the first feature fusion module, w 1 is the weight coefficient, b 1 , b 2 is the bias coefficient, is the transpose of w 1 .

[0131] The first feature fusion module transmits the fused n-dimensional feature vector output to the excitation unit, and updates the weight coefficients of the n-dimensional feature vector according to the principle of minimum mean square error by the gradient descent method. The specific expression is:

[0132] (w 1 , b 1 , b 2 ) ← arg min(Loss(w 1 , b 1 , b 2 ));

[0133]

[0134] In the formula, w 1 is the matrix composed of the weight coefficients, b 1 , b 2 are the bias coefficients, Loss is the loss function, Δw i,j is the change value of the j-th weight coefficient of the i-th dimensional feature vector, Δb 1i , Δb 2i are the change values of the bias coefficients of the i-th dimensional feature vector.

[0135] According to the updated weight coefficients, feature fusion is performed on the n-dimensional feature vector in the second feature fusion module to achieve deeper feature extraction of the data and improve the prediction accuracy.

[0136] Although setting multiple feature fusion modules in the stacked autoencoder model can improve the feature extraction and fusion effects, it will increase the computational complexity of the model. Therefore, in the actual operation process, it is also necessary to adjust the number of feature fusion modules according to the actual prediction requirements to set the number of feature extraction times and feature fusion times of the stacked autoencoder model.

[0137] Implementing the embodiments of the present application has the following beneficial effects:

[0138] In the embodiments of the present application, the collected battery charge and discharge data is first subjected to preliminary feature extraction, unifying the extraction of linear and non-linear features, so that the method can be applied to various scenarios, and the extracted feature data is not limited to linear or non-linear structures; corresponding health factors are generated according to the extracted features, and each health factor can characterize the performance of the battery. According to the influence of each health factor on predicting the remaining life of the battery, weight assignment is performed for each health factor, highlighting the importance of important health factors in the process of evolving the degree of battery performance degradation, so as to increase the performance ability of important health factors in feature fusion and further improve the prediction accuracy. Moreover, the weight coefficient can also increase the discrimination speed of important health factors in the fusion process, improve the operation efficiency, and can more efficiently predict the current battery performance.

[0139] Further, Figure 7 is a structural diagram of a terminal device provided by an embodiment of the present application. As Figure 7 shown, the terminal device 3 of this embodiment includes: at least one processor 30 (only one is shown in Figure 7 ), a memory 31, and a computer program 32 stored in the memory 31 and executable on the at least one processor. When the processor 30 executes the computer program 32, the steps of a method for predicting the remaining life of a battery based on stacked autoencoders described in any one of the embodiments of the present application can be implemented.

[0140] The terminal device 3 can be a computing device such as a desktop computer, a cloud server, and a laptop computer. The computing device may include, but is not limited to, the processor 30 and the memory 31. Figure 7 This is only an example of the terminal device 3 and does not constitute a limitation on the terminal device 3. It may include more or fewer components than shown in the figure.

[0141] The specific embodiments described above further elaborate on the purpose, technical solutions, and beneficial effects of the present application. It should be understood that the above are only specific embodiments of the present application and are not used to limit the protection scope of the present application. In particular, for those skilled in the art, any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included within the protection scope of the present application.

Claims

1. A method for predicting remaining battery life based on stacked autoencoder, characterized in that: include: The collected battery charging and discharging data is input into the preset stacked autoencoder model to extract linear and nonlinear features to obtain several health factors of the battery; According to the excitation unit of the stacked autoencoder model, a corresponding weight coefficient is generated for each of the health factors based on actual prediction requirements; The health factors are feature-fused by using the weight coefficients to predict the degree of battery performance degradation, thereby obtaining a prediction result of the remaining life of the battery.

2. The method for predicting remaining battery life based on stacked autoencoder according to claim 1, characterized in that: The collected battery charge and discharge data is input into a preset stacked autoencoder model to extract linear and nonlinear features to obtain several health factors of the battery, specifically: According to the preset battery degradation characteristics, a first threshold number of characteristic parameters are screened out from the battery charge and discharge data by using the stacked autoencoder model; The linear characteristic parameters and the nonlinear characteristic parameters are normalized to generate a first threshold number of health factors, and the health factors are integrated into a characteristic vector of a first threshold dimension.

3. The method for predicting remaining battery life based on stacked autoencoder according to claim 2, characterized in that: The characteristic parameters include: initial voltage drop value, battery discharge power, average battery temperature during the discharge phase, constant current charging time, ratio of constant current charging time to constant voltage charging time, and discharge cut-off voltage.

4. The method for predicting remaining battery life based on stacked autoencoder according to claim 1, characterized in that: The excitation unit according to the stacked autoencoder model generates a corresponding weight coefficient for each health factor based on actual prediction requirements, specifically: Determine the influence value of each health factor on the predicted battery performance degradation degree according to actual prediction requirements; The incentive unit assigns the weight coefficient of corresponding size to each health factor through the influence value; The excitation unit generates the weight coefficient by using a gradient descent method and a minimum mean square error principle.

5. The method for predicting remaining battery life based on stacked autoencoder according to claim 4, characterized in that: The weight coefficient is generated by the gradient descent method according to the minimum mean square error principle, specifically: The weight coefficient is generated by the following formula: (w1, b1, b2)←arg min(Loss(w1, b1, b2)); Wherein, w1 is the matrix composed of the weight coefficients, b1 and b2 are the bias coefficients of the stacked autoencoder model, and Loss is the loss function of the stacked autoencoder model. is the i-th output data of the stacked autoencoder model, x i is the i-th input data of the stacked autoencoder model, m is the number of health factors, Δw i,j is the change value of the jth weight coefficient of the i-th eigenvector, Δb 1i , Δb 2i is the change in the bias coefficient of the i-th dimension eigenvector, is the transpose of w1.

6. The method for predicting remaining battery life based on stacked autoencoder according to claim 1, characterized in that: The health factor is feature-fused by the weight coefficient to predict the performance degradation of the battery, and the remaining life prediction result of the battery is obtained, specifically: The weight of the health factor during feature fusion is adjusted according to the size of the weight coefficient, and the weight coefficient is subjected to feature fusion with the goal of evolving the performance degradation degree of the battery to obtain the performance degradation degree of the current battery; The remaining life prediction result is obtained according to the performance degradation degree.

7. The method for predicting remaining battery life based on stacked autoencoder according to any one of claims 1 to 6, characterized in that: Also includes: According to actual prediction requirements, the number of feature extractions and feature fusions of the stacked autoencoder model is set.

8. The method for predicting remaining battery life based on stacked autoencoder according to claim 1, characterized in that: The collected battery charging and discharging data are specifically: In several charge and discharge cycles, the first state data of the battery before each charge and discharge cycle and the second state data after each charge and discharge cycle are collected to obtain the battery charge and discharge data.

9. A battery remaining life prediction system based on stacked autoencoder, characterized in that: include: Health factor extraction module, weight coefficient assignment module and remaining life prediction module; Among them, the health factor extraction module is used to input the collected battery charging and discharging data into the preset stacked autoencoder model to extract linear and nonlinear features to obtain several health factors of the battery; The weight coefficient assignment module is used to generate a corresponding weight coefficient for each of the health factors based on the actual prediction requirements according to the excitation unit of the stacked autoencoder model; The remaining life prediction module is used to perform feature fusion on the health factor through the weight coefficient, predict the performance degradation degree of the battery, and obtain the remaining life prediction result of the battery.

10. A terminal device, characterized in that: It comprises a processor and a memory, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of a method for predicting remaining battery life based on stacked autoencoding according to any one of claims 1 to 8 are implemented.

Images