Lithium ion battery degradation classification method based on BP neural network

Abstract

The invention relates to the technical field of lithium batteries, in particular to a lithium ion battery degradation classification method based on a BP neural network, and the method comprises the steps: obtaining the related data of an incremental capacity curve of a battery, constructing a data set, training a dynamic adaptive time sequence BP neural network model according to the data set, and enabling the model to be used for predicting the capacity of the battery; acquiring a charging data set, wherein the charging data set comprises constant-current charging stage data, constant-voltage charging stage data and trickle charging stage data; a charging BP neural network model is trained according to the charging data set, and the model comprises three sub-models and is used for obtaining three degradation correction terms; a degradation correction term is obtained and processed, and a comprehensive degradation correction term is obtained; and obtaining and processing the predicted battery capacity and the comprehensive degradation correction term to obtain a battery degradation index, and performing battery classification according to a preset battery degradation classification function. According to the invention, by introducing the predicted battery capacity and the comprehensive degradation correction term, the accuracy of estimating the degradation degree of the battery is improved.

Description

Technical Field

[0001] The invention relates to the technical field of lithium batteries, and in particular to a lithium ion battery degradation classification method based on BP neural network. Background Art

[0002] BP neural network is a computational model that simulates the interconnection of neurons in the human brain. It minimizes prediction errors by adjusting connection weights and biases, and is used to handle complex nonlinear relationship problems. BP neural network can complete tasks such as recognition, classification, and regression prediction by training large amounts of data. Currently, BP neural network is widely used in many fields, such as image recognition, natural language processing, medical diagnosis, and autonomous driving.

[0003] Lithium-ion battery is a rechargeable battery widely used in consumer electronics, electric vehicles and energy storage systems. Its working principle is based on the reciprocating movement of lithium ions between the positive and negative electrodes to achieve charging and discharging. Lithium-ion battery has the advantages of high energy density, light weight and long life, and has become the mainstream energy storage battery technology.

[0004] During the charge and discharge cycle, the capacity and performance of lithium-ion batteries gradually decrease due to factors such as electrolyte decomposition, lithium dendrite growth and electrode material aging, which is manifested in the decrease of battery charging efficiency, discharge capacity and service life. The BP neural network adjusts the weights through the back propagation algorithm, so that the model can continuously learn and optimize, which makes it highly adaptable when dealing with complex nonlinear problems and is suitable for multivariate and complex classification problems such as lithium-ion batteries.

[0005] In existing research, most of them rely on estimating the battery capacity that the battery can currently carry to judge the health or degree of degradation of the battery. The judgment and classification of the degree of degradation of lithium-ion batteries using a constant current and constant voltage charging strategy are not accurate enough. The performance of the battery during the charging process also reflects the degree of battery degradation, which needs further exploration and analysis.

[0006] Therefore, a lithium-ion battery degradation classification method based on BP neural network is proposed. Summary of the invention

[0007] The purpose of the present invention is to provide a method for classifying the degradation of lithium-ion batteries based on a BP neural network. By constructing a battery incremental capacity data set and training a dynamic adaptive time-series BP neural network model according to the data set to predict the battery capacity; by constructing a charging data set and training a charging BP neural network model according to the charging data set, the model includes three sub-models for obtaining three degradation correction terms; obtaining and processing the three degradation correction terms to obtain a comprehensive degradation correction term; obtaining a battery degradation index based on the predicted battery capacity and the comprehensive degradation correction term, and classifying the battery according to a preset battery degradation classification function.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A method for classifying the degradation of lithium-ion batteries based on a BP neural network, including:

[0010] Obtaining and processing relevant data of the incremental capacity curve of the battery to obtain a battery incremental capacity data set; training a dynamic adaptive time-series BP neural network model according to the battery incremental capacity data set;

[0011] The dynamic adaptive time-series BP neural network model obtains and processes the incremental capacity data of the battery to obtain a predicted battery capacity.

[0012] Further, the battery incremental capacity data set includes:

[0013] The battery incremental capacity data set includes peak features, valley features, peak slope features, curve area features, degradation features, and area change amounts on the incremental capacity curve of the battery.

[0014] Further, the structure of the dynamic adaptive time-series BP neural network model includes:

[0015] The dynamic adaptive time-series BP neural network model includes an input layer, a hidden layer, an association layer, a fusion layer, and an output layer; wherein the association layer includes several LSTM layers, and the fusion layer obtains the outputs and fusion weights of the LSTM layers and processes them to obtain fusion features.

[0016] Obtaining a first battery charging data set, the battery charging data set including constant current charging stage data, constant voltage charging stage data, and trickle charging stage data; obtaining the first battery charging data set and performing data augmentation to obtain a second battery charging data set.

[0017] Further, the first battery charging data set includes: obtaining charging data of a complete time series, the charging data including the battery terminal voltage V(t) and the battery terminal current I(t) in the time series;

[0018] Further, obtain the battery terminal voltage and the battery terminal current and process them according to a data partitioning method to obtain the constant current charging stage data, the constant voltage charging stage data, and the trickle charging stage data.

[0019] Further, the first battery charging data set further includes:

[0020] The constant current charging stage data includes: initial voltage, initial current, maximum charging voltage, voltage rise rate, constant current charging duration, constant current, current fluctuation data, initial battery capacity, cumulative charge in the constant current charging stage, remaining charge in the constant current charging stage, initial battery temperature, ambient temperature, and termination temperature of the constant current charging stage; the constant voltage charging stage data includes: constant voltage, voltage fluctuation data, initial current in the constant voltage charging stage, current decay rate, cumulative charge in the constant voltage charging stage, remaining charge in the constant voltage charging stage, initial temperature in the constant voltage stage, termination temperature in the constant voltage stage, the ambient temperature, constant voltage charging duration, and preset current threshold; the trickle charging stage data includes: the constant voltage, trickle charging stage voltage fluctuation data, minimum current, cumulative charge in the trickle charging stage, charging power, initial temperature in the trickle stage, termination temperature in the trickle stage, the ambient temperature, and trickle charging duration.

[0021] Further, the data enhancement includes:

[0022] Perform preliminary filling processing on the missing data and standardize the data to obtain a standard data set; obtain the standard data set and perform clustering to obtain multiple clustering clusters; calculate the intra-class data similarity of the missing data within the clustering clusters, and obtain the corrected data of the missing data through weighting;

[0023] Correct all the missing data to obtain a second battery charging data set.

[0024] Train a charging BP neural network model according to the second battery charging data set;

[0025] The charging BP neural network model obtains the charging data of the battery and processes it to obtain a first deterioration correction term, a second deterioration correction term, and a third deterioration correction term.

[0026] Further, the charging BP neural network model includes:

[0027] The charging BP neural network model includes a constant current model, a constant voltage model, and a trickle model. Among them, the constant current model obtains the constant current charging stage data and processes it to output the first deterioration correction term, the constant voltage model obtains the constant voltage charging stage data and processes it to output the second deterioration correction term, and the trickle model obtains the trickle charging stage data and processes it to output the third deterioration correction term.

[0028] Furthermore, the charging BP neural network model further includes:

[0029] The constant current model, the constant voltage model, and the trickle model are all dynamic adaptive BP neural network models; the structure of the dynamic adaptive BP neural network model includes: an input layer, a first hidden layer, a second hidden layer, and an output layer.

[0030] Furthermore, the dynamic adaptability includes: introducing a dynamic adaptive learning rate during the model training stage to optimize the training process of the neural network model; the calculation formula for the dynamic adaptive learning rate is:

[0031]

[0032] where η(e) represents the learning rate of the e-th iteration, and η initial represents the initial learning rate; β(e) represents the learning rate decay factor;

[0033] Obtain the data saved in the loss window, where the loss window saves the loss values of the most recent G iterations; obtain the loss value and process it to obtain the average change rate of the loss value. The calculation formula for the average change rate of the loss value is:

[0034]

[0035] where ΔL represents the average change rate of the loss value, and L e-g+1 represents the loss value of the (e - g + 1)-th iteration, and L e-g represents the loss value of the (e - g)-th iteration;

[0036] If the average change rate of the loss value is less than the preset first change threshold L 1 , where L 1 < 0, then the calculation formula for the learning rate decay factor of the e-th iteration is:

[0037] β(e) = β 0 ×(1 + |ΔL|) -1 ;

[0038] where β(e) represents the learning rate decay factor of the e-th iteration, and β 0 represents the initial learning rate decay factor;

[0039] If the average change rate of the loss value belongs to [L 1 , L 2 , where L 2 represents the preset second change threshold, and L 2 > 0, then β(e) = β 0 +(1 + |ΔL|);

[0040] If the average change rate of the loss value is greater than the preset second change threshold, the calculation formula for the learning rate decay factor in the e-th iteration is:

[0041] β(e) = min(β max , β 0 ×(1 + |ΔL| 2 ));

[0042] where β max represents the maximum learning rate decay factor.

[0043] Obtain the first deterioration correction term, the second deterioration correction term, and the third deterioration correction term and process them to obtain a comprehensive deterioration correction term.

[0044] Obtain the predicted battery capacity and the comprehensive deterioration correction term and process them to obtain a battery deterioration index, and classify the battery according to a preset battery deterioration classification function.

[0045] Furthermore, the battery classification includes:

[0046] The calculation formula for the battery deterioration index is:

[0047]

[0048] where W represents the battery deterioration index, map() represents a mapping function, C i represents the predicted battery capacity during the i-th charging, C 0 represents the rated battery capacity, w represents the comprehensive deterioration correction term, and μ 1 represents the deterioration correction term coefficient;

[0049] Obtain the battery deterioration index and the preset battery deterioration classification function and process them to obtain a battery deterioration classification result.

[0050] Compared with the prior art, the beneficial effects of the present invention are:

[0051] 1. By using various data such as peak value, valley value, slope, curve area, and peak-valley distance, which have a high correlation with battery deterioration and capacity change, it can more comprehensively reflect battery characteristics and improve the accuracy of predicting battery capacity; through the multi-layer LSTM in the association layer, the model can extract and capture key features in the battery time series data, and then through the fusion layer, the outputs of each LSTM layer are integrated and weighted with weights to ensure the efficient integration of information and enhance the prediction ability of time series data; using a dynamic adaptive time series BP neural network model that combines the time series characteristics of LSTM and the non-linear fitting ability of the BP neural network can better capture the dynamic characteristics of the time series during battery deterioration and improve the prediction accuracy of battery capacity.

[0052] 2. By separately obtaining data in the constant current charging, constant voltage charging, and trickle charging stages, the characteristic differences in battery charging in each stage can be fully captured, the degradation information under different charging modes can be refined, the charging behavior of the battery can be more accurately reflected, and the accuracy of battery degradation classification can be improved; through data augmentation, the initial charging data set is expanded to ensure that there is a sufficient sample size during model training; missing data is corrected by methods such as clustering and weighted processing to ensure the integrity and accuracy of the data; by comprehensively processing the first degradation correction term, the second degradation correction term, and the third degradation correction term, the degradation process of the battery can be corrected and optimized in an all-round and multi-level manner, and a more reliable comprehensive degradation correction term can be obtained.

[0053] 3. By introducing a dynamic adaptive learning rate, while ensuring the rapid convergence of the model, the learning rate is flexibly adjusted, enhancing the stability and robustness of the model, avoiding overfitting and underfitting phenomena, significantly improving the training efficiency of the model, and ultimately enhancing the classification accuracy and generalization ability of the model; by introducing the predicted battery capacity and the comprehensive degradation correction term, the battery degradation index is obtained, comprehensively reflecting the current health state of the battery, providing a reliable basis for the degradation classification of the battery, and improving the accuracy of battery degradation assessment. Description of the Drawings

[0054] Figure 1 It is a flowchart of a lithium-ion battery degradation classification method based on a BP neural network provided in Embodiment 1 of the present invention;

[0055] Figure 2 It is a schematic diagram of the incremental capacity curve of the battery provided in Embodiment 1 of the present invention;

[0056] Figure 3 It is a schematic diagram of the structure of a dynamic adaptive time-series BP neural network model provided in Embodiment 1 of the present invention;

[0057] Figure 4 It is a schematic diagram of the structure of a dynamic adaptive BP neural network model provided in Embodiment 1 of the present invention. Detailed Embodiments

[0058] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0059] Embodiment 1

[0060] With the popularization of smart phones, the health status of the battery is crucial for the user experience. After hundreds of charging cycles, the mobile phone battery will gradually deteriorate, resulting in problems such as shortened battery life, slower charging speed, and even automatic shutdown. Therefore, in order to extend the life of the mobile phone battery and optimize the user experience, a certain mobile phone operating system adopts a lithium-ion battery degradation classification method based on the BP neural network provided by the present invention to monitor the degradation status of the battery in real time.

[0061] As Figure 1 shown in the flowchart of a lithium-ion battery degradation classification method based on the BP neural network provided by the present invention, first, relevant data of the incremental capacity curve of the battery is obtained and processed to obtain a battery incremental capacity data set; a dynamic adaptive time series BP neural network model is trained according to the battery incremental capacity data set; the dynamic adaptive time series BP neural network model obtains the incremental capacity data of the battery and processes it to obtain the predicted battery capacity.

[0062] Further, as Figure 2 shown in the schematic diagram of the incremental capacity curve of the battery, where Q represents the increment of the battery capacity, the abscissa represents the charging voltage at the battery terminal, and the ordinate represents the derivative of the battery capacity increment with respect to the voltage. The battery incremental capacity data set includes peak features, valley features, peak slope features, curve area features, degradation features, and area change amounts.

[0063] Further, the peak features include the first peak and its voltage and the second peak and its voltage; the valley features include the first valley and its voltage; the peak slope features include obtaining the coordinates of the first peak and the second peak, and the slope of the line connecting the two coordinates is the peak slope feature;

[0064] Further, the curve area features include: a perpendicular line is drawn from the first peak coordinate to the abscissa and a perpendicular line is drawn from the second peak coordinate to the abscissa, and the area enclosed by the two perpendicular lines and the incremental capacity curve represents the peak area; the total curve area is the area enclosed by the projection of the two ends of the incremental capacity curve on the abscissa and the curve; the spacing features include the distance between the peak and the valley and the distance between the first peak and the second peak;

[0065] Further, the degradation features include the change amounts of the first peak and the second peak in the battery incremental capacity curve compared with the first peak and the second peak in the previous charge, and the change amount of the valley in the battery incremental capacity curve compared with the valley in the previous charge; the area change amount includes the change amount of the total curve area compared with the total curve area of the battery incremental capacity curve in the previous charge;

[0066] Further, the fluctuation feature includes segmentation according to the curve slope. The part greater than the segmentation threshold indicates a larger volatility, and the length of the interval greater than the segmentation threshold is obtained. In this embodiment, a horizontal line is drawn based on the coordinates of the trough value, which intersects with the incremental capacity curve. The length of the curve part above the horizontal line represents the fluctuation feature.

[0067] By analyzing the first peak and the second peak and their corresponding voltages, the key features of the incremental capacity curve are refined, and the subtle deterioration trend of the battery during charging can be detected; the change in area can reflect the attenuation of the battery capacity, which is beneficial to improving the accuracy of battery capacity prediction; by comparing the peak, trough, and area features of the current charging cycle with the changes in the previous charge, the deterioration progress of the battery can be accurately tracked; by extracting various features such as the peak, trough, slope, and area of the incremental capacity curve, the deterioration process of the battery can be comprehensively captured.

[0068] Table 1. Training Results

[0069] Training cycle Training loss Validation loss Training accuracy Validation accuracy 1 0.076 0.086 87.05％ 85.34％ 2 0.064 0.072 89.56％ 87.02％ 3 0.053 0.064 91.02％ 88.53％ 4 0.047 0.052 93.19％ 91.26％ 5 0.032 0.046 94.54％ 92.49％ 6 0.025 0.034 96.02％ 94.28％ 7 0.018 0.027 96.59％ 94.89％ 8 0.014 0.022 97.04％ 95.12％ 9 0.012 0.017 97.62％ 95.67％ 10 0.009 0.015 98.13％ 96.14％

[0070] The dynamic adaptive time-series BP neural network model obtains and processes the battery incremental capacity data to obtain the predicted battery capacity C i .

[0071] Further, the dynamic adaptive time-series BP neural network model is as Figure 3 shown, including an input layer, a hidden layer, an association layer, a fusion layer, and an output layer; wherein the association layer includes multiple LSTM layers, and the fusion layer obtains the outputs and fusion weights of the LSTM layers and processes them to obtain fusion features.

[0072] Further, the input layer receives the features [F 1 , F 2 , F 3 , F 4 , F 5 , F 6 , F 7 of the input battery incremental capacity curve, where F 1 represents the peak feature, F 2 represents the trough feature, F 3 represents the peak slope feature, F 4 represents the curve area feature, F 5 represents the deterioration feature, F 6 represents the change in area, F 7 represents the fluctuation feature; the input layer processes the features and outputs the processing results to the hidden layer.

[0073] Further, the association layer contains multiple LSTM layers. In this embodiment, there are three LSTM layers;

[0074] Further, the association layer receives the output result of the hidden layer and transports the output result to three LSTM layers. The three LSTM layers process the received data respectively to obtain three output results. The fusion layer receives the data output by the association layer and performs fusion processing to obtain a fusion feature. The calculation formula of the fusion feature is as follows:

[0075]

[0076] where, F merge represents the fusion feature, F LSTM1 represents the output result of the first LSTM layer, and α 1 represents the weight coefficient of the output result of the first LSTM layer. F LSTM2 represents the output result of the second LSTM layer, and α 2 represents the weight coefficient of the output result of the second LSTM layer. F LSTM3 represents the output result of the third LSTM layer, and α 3 represents the weight coefficient of the output result of the third LSTM layer. represents the way of feature fusion.

[0077] Further, the dynamic adaptive time series BP neural network model receives the above data and processes it to obtain the predicted battery capacity C i .

[0078] As shown in Table 1, the training results of the dynamic adaptive time series BP neural network model are presented. Through analysis, it can be found that after complete training, the prediction accuracy of the model can reach more than 98%, and the accuracy on the test set can also reach more than 96%.

[0079] Through the design of the input layer, hidden layer, association layer and fusion layer, this model can process various features of the battery incremental capacity curve in multiple levels and all-round. Especially through the processing of time series features by multiple LSTM layers, the model can capture the dynamic behavior of the battery changing with time, further enhancing the prediction ability of the model and improving the accuracy of battery capacity prediction.

[0080] Obtain the first battery charging data set, where the battery charging data set includes constant current charging stage data, constant voltage charging stage data and trickle charging stage data; obtain the first battery charging data set and perform data augmentation to obtain the second battery charging data set.

[0081] Further, the first battery charging data set includes:

[0082] Obtain the complete time-series charging data, where the charging data includes the battery terminal voltage V(t) and the battery terminal current I(t) in the time series;

[0083] Obtain the battery terminal voltage and the battery terminal current and process them according to the data partitioning method to obtain the constant current charging stage data, the constant voltage charging stage data, and the trickle charging stage data.

[0084] Further, if during the charging stage, the current remains constant and the charging voltage gradually increases until the set maximum charging voltage is reached, then this stage belongs to the constant current charging stage; if when the battery terminal voltage reaches the set maximum voltage, the current gradually decreases and the current is greater than the preset current threshold; then this section of data belongs to the constant voltage charging stage; if the current is less than or equal to the preset current threshold, then this section of data belongs to the trickle charging stage.

[0085] Further, the steps of the specific data partitioning method are as follows:

[0086] The constant current charging stage segmentation includes:

[0087] Step S101: Initialize the time point t 0 As the charging starting point;

[0088] Step S102: During the time period [t 0 , t end , detect the voltage V(t) and current I(t) at each moment, where t end represents the charging end point;

[0089] Step S103: If the following conditions are met, there is a voltage derivative and V(t) < V max , then the stage [t 0 , t 1 belongs to the constant current charging stage, where t 1 represents the time point when the voltage reaches the set maximum charging voltage V max ;

[0090] The constant voltage charging stage includes:

[0091] Step S201: During the time period [t 1 , t end , detect the voltage V(t) and current I(t) at each moment;

[0092] Step S202: If the following conditions are met, there is and the current derivative and I(t) > I min , then [t 1 , t2 This stage belongs to the constant voltage charging stage, where t 2 represents the time point when the current drops to the preset current threshold I min .

[0093] The trickle charging stage includes:

[0094] Step S301: During the time period [t 2 , t end , detect the voltage V(t) and current I(t) at each moment;

[0095] Step S302: If the following conditions are met, there is and I(t) ≤ I min , then the stage [t 2 , t end belongs to the trickle charging stage.

[0096] By detecting the changes in current and voltage in the charging data of the complete time, the charging process is accurately divided into three stages: constant current, constant voltage, and trickle charging, and the data of each stage is extracted and analyzed. This segmentation method based on charging characteristics can accurately reflect the performance of the battery in different charging stages, making the data segmentation more accurate and helping to comprehensively analyze the charge and discharge characteristics of the battery.

[0097] Furthermore, the charging data further includes:

[0098] The data of the constant current charging stage includes: initial voltage, initial current, maximum charging voltage, voltage rise rate, constant current charging duration, constant current, current fluctuation data, initial battery capacity, cumulative charge in the constant current charging stage, remaining charge in the constant current charging stage, initial battery temperature, ambient temperature, and termination temperature of the constant current charging stage. Table 2 shows some of the data of the constant current charging stage;

[0099] Furthermore, the current fluctuation data refers to the data extracted after performing Fourier transform on the current data in the constant current charging stage;

[0100] Table 2. Some data of the constant current charging stage

[0101]

[0102] The data of the constant voltage charging stage includes: constant voltage, voltage fluctuation data, initial current in the constant voltage charging stage, current decay rate, cumulative charge in the constant voltage charging stage, remaining charge in the constant voltage charging stage, initial temperature of the constant voltage stage, termination temperature of the constant voltage stage, the ambient temperature, constant voltage charging duration, and preset current threshold. Table 3 shows some of the data of the constant voltage charging stage;

[0103] Further, the voltage fluctuation data refers to the data extracted after performing Fourier transform on the voltage data in the constant voltage charging stage;

[0104] Table 3. Partial data in the constant voltage charging stage

[0105]

[0106] The trickle charging stage data includes: the constant voltage, the voltage fluctuation data in the trickle charging stage, the minimum current, the cumulative charge in the trickle charging stage, the charging power, the initial temperature in the trickle stage, the termination temperature in the trickle stage, the ambient temperature, and the trickle charging duration. Table 4 shows partial data in the trickle charging stage.

[0107] Table 4. Partial data in the trickle charging stage

[0108]

[0109] Comprehensively extracting the data in the constant current charging stage, constant voltage charging stage, and trickle charging stage, and recording multiple key parameters including the initial voltage, current fluctuation, voltage fluctuation, and charging power, etc., can accurately characterize the characteristics of each stage in the charging process; refined data acquisition helps to deeply understand the performance of the battery in different charging stages and provides more detailed basic data for subsequent battery degradation assessment.

[0110] Further, the data enhancement includes:

[0111] Performing preliminary filling processing on the missing data and normalizing the data to obtain a standard data set; obtaining the standard data set and performing clustering to obtain multiple clustering clusters; calculating the within-class data similarity of the missing data within the clustering clusters, and obtaining the corrected data of the missing data through weighting;

[0112] Correcting all the missing data to obtain a second battery charging data set.

[0113] Further, first perform preliminary filling processing on the missing data so that the subsequent clustering algorithm can run normally. Common preliminary filling methods include mean filling, median filling, and K-nearest neighbor filling, etc.; after filling, perform normalization processing on the data so that the data is within the same scale range.

[0114] Further, for each data column, calculate the mean of all non-missing data and fill the missing values in this column with this mean. The formula is:

[0115]

[0116] where x miss represents the filling value, N represents the number of non-missing data, xn represents the nth non-missing data; after filling, all data vectors are standardized so that their mean is 0 and their standard deviation is 1; after completing the initial filling and standardization, a standard data set without missing data is obtained for subsequent clustering.

[0117] Furthermore, the data is divided into K clusters to minimize the distance from each sample to the center of the cluster to which it belongs, where the K value can be determined by the elbow method or the silhouette coefficient; in the kth cluster J k In the cluster, there are several samples. For the samples in the cluster, the similarity between the samples in the class is calculated, and the missing data in the class is corrected by weighted method, and finally the corrected data is obtained.

[0118] Furthermore, cosine similarity or Euclidean distance is first used to measure the similarity between samples within a class. In this embodiment, cosine similarity is used, and the calculation formula is:

[0119]

[0120] in, Indicates that in cluster J k The ath data vector in and the bth data vector The cosine similarity of

[0121] Furthermore, after calculating all the similarities, the weighted correction value of the missing data is calculated according to the similarities. The calculation formula is:

[0122]

[0123] Among them, x re represents the corrected data, λ 1 and λ 2 represents the weight coefficient, M represents the number of data vectors whose similarity is greater than the preset similarity threshold, ω m Represents the data x on the data column corresponding to the missing value in the data vector whose mth similarity is greater than the preset similarity threshold m After filling and correcting all missing data, the second battery charging data set is obtained.

[0124] Through the operation of preliminary data filling, the missing data can be quickly corrected, ensuring the integrity of the data set and ensuring that subsequent data processing such as standardization and clustering can proceed smoothly; by calculating the sample similarity within the cluster where the missing data is located and weighted correction based on similar samples, the missing values ​​can be completed more accurately, significantly improving the filling accuracy of missing data, reducing filling deviations, and thus improving the overall quality of the data set.

[0125] Train a charging BP neural network model based on the second battery charging dataset; the charging BP neural network model obtains and processes the charging data of the battery to obtain the first deterioration correction term w 1 、the second deterioration correction term w 2 and the third deterioration correction term w 3 .

[0126] Furthermore, the charging BP neural network model includes:

[0127] The charging BP neural network model includes a constant current model, a constant voltage model, and a trickle model. Among them, the constant current model obtains and processes the constant current charging stage data and outputs the first deterioration correction term. The constant voltage model obtains and processes the constant voltage charging stage data and outputs the second deterioration correction term. The trickle model obtains and processes the trickle charging stage data and outputs the third deterioration correction term.

[0128] The charging process in each stage has different effects on the deterioration of the battery. For example, the constant current stage may more reflect the initial health state of the battery, the constant voltage stage reflects the charging performance when the battery is close to saturation, and the trickle stage is used to maintain the balance of the battery terminal voltage. Therefore, by processing the data in the constant current, constant voltage, and trickle stages respectively and predicting the deterioration correction term in stages, it can ensure that the deterioration effects in each stage are accurately captured, generate the corresponding deterioration correction terms, and improve the accuracy and reliability of the deterioration assessment.

[0129] Furthermore, the charging BP neural network model also includes:

[0130] The constant current model, the constant voltage model, and the trickle model are all dynamic adaptive BP neural network models; the structure of the dynamic adaptive BP neural network model is as Figure 4 shown, including: an input layer, a first hidden layer, a second hidden layer, and an output layer.

[0131] Furthermore, the input layer of the dynamic adaptive BP neural network model receives and processes the charging data to obtain a first output result. The first hidden layer obtains and processes the first output result to obtain a second output result. The second hidden layer receives and processes the second output result to obtain a third output result. The output layer receives and processes the third output result to obtain an output result.

[0132] Furthermore, establish a training set and a test machine using the constant current charging stage data in the corrected second battery charging dataset. Train the constant current model according to the constant current charging stage data, and calculate the error between the network output and the true value using the mean square error loss function. The calculation formula of the loss function is:

[0133]

[0134] where \(L\) represents the loss function, \(d\) y represents the true value of the \(y\)-th sample, which represents the true degree of deterioration of the battery in this embodiment, represents the predicted value of the \(y\)-th sample predicted by the model, which represents the first deterioration correction term of the battery in this embodiment, and \(Y\) represents the total number of samples.

[0135] Further, similar to the above method, the constant voltage model and the trickle charging model are trained using the constant voltage charging stage data and the trickle charging stage data respectively. The trained constant voltage model and the trained trickle charging model receive data and output the second deterioration correction term and the third deterioration correction term respectively.

[0136] The constant current model, the constant voltage model, and the trickle charging model all adopt a dynamic adaptive BP neural network model, and are independently trained for the three charging stages of constant current, constant voltage, and trickle charging respectively, and predict the deterioration correction term, which helps to reduce the overfitting risk generated when the global model processes too much data, and can capture the deterioration characteristics of the battery in different stages more carefully, thereby improving the prediction accuracy of the deterioration correction term.

[0137] Further, the dynamic adaptation includes:

[0138] Introduce a dynamic adaptive learning rate, use a larger learning rate at the initial stage of training to quickly adjust the weights, and automatically reduce the learning rate when the error approaches the minimum value; the calculation formula of the dynamic adaptive learning rate is:

[0139]

[0140] where \(\eta(e)\) represents the learning rate of the \(e\)-th iteration, \(\eta\) initial represents the initial learning rate; \(\beta(e)\) represents the learning rate decay factor;

[0141] Obtain the data saved in the loss window, and the loss window saves the loss values of the last \(G\) iterations; obtain the loss value and process it to obtain the average change rate of the loss value, and the calculation formula of the average change rate of the loss value is:

[0142]

[0143] where \(\Delta L\) represents the average change rate of the loss value, \(L\) e-g+1 represents the loss value of the \((e - g + 1)\)-th iteration, \(L\) e-g represents the loss value of the \((e - g)\)-th iteration;

[0144] If the average change rate of the loss value is less than the preset first change threshold \(L\) 1 , where \(L\) 1 <0, then the calculation formula of the learning rate decay factor of the \(e\)-th iteration is:

[0145] β(e) = β 0 ×(1 + |ΔL|) -1 ;

[0146] where β(e) represents the learning rate decay factor for the e-th iteration, and β 0 represents the initial decay factor;

[0147] If the average change rate of the loss value belongs to [L 1 , L 2 , where L 2 represents the preset second change threshold and L 2 > 0, then β(e) = β 0 + (1 + |ΔL|);

[0148] If the average change rate of the loss value is greater than the second preset change threshold, the calculation formula for the learning rate decay factor of the e-th iteration is:

[0149] β(e) = min(β max , β 0 ×(1 + |ΔL 2 ));

[0150] where β max represents the maximum learning rate decay factor.

[0151] Furthermore, at the end of each training cycle, re-evaluate the learning effect. If the model converges too slowly or the learning rate drops too fast, the decay factor can be adjusted.

[0152] The dynamic adaptive learning rate strategy can significantly improve the training efficiency and accuracy of the model by dynamically adjusting the learning rate and decay factor according to the change of the loss value, effectively shortening the training time, reducing manual intervention, preventing overfitting and local optimum problems, and improving the self-adaptability and generalization ability of the model.

[0153] Obtain the first deterioration correction term, the second deterioration correction term, and the third deterioration correction term and process them to obtain the comprehensive deterioration correction term;

[0154] Furthermore, the calculation formula for the comprehensive deterioration correction term is:

[0155] w = θ 1 ·w 1 + θ 2 ·w 2 + θ 3 ·w 3 ;

[0156] where w represents the comprehensive deterioration correction term, and θ 1 , θ 2 and θ 3Respectively represent the weight coefficients of the first deterioration correction term, the second deterioration correction term, and the third deterioration correction term.

[0157] Obtain the predicted battery capacity and the comprehensive deterioration correction term and process them to obtain a battery deterioration index, and classify the battery according to a preset battery deterioration classification function.

[0158] Furthermore, the formula for the battery deterioration index in the battery deterioration classification is as follows:

[0159]

[0160] Among them, W represents the battery deterioration index, map() represents a mapping function, and C i represents the predicted battery capacity during the i-th charge, and μ 0 represents the weight coefficient of the predicted battery capacity, and C 0 represents the rated battery capacity, w represents the comprehensive deterioration correction term, and μ 1 represents the deterioration correction term coefficient;

[0161] Obtain the battery deterioration index and the preset battery deterioration classification function and process them to obtain a battery deterioration classification result.

[0162] Furthermore, in this embodiment, the mapping function maps the battery deterioration index to the range of [0, 100], and the formula of the preset battery deterioration classification function is:

[0163]

[0164] Among them, CD represents the degree of battery deterioration. In this embodiment, after the mobile phone battery is charged 275 times, the battery deterioration index of the battery is 86, and the battery belongs to mild deterioration. Through the battery health monitoring interface of the mobile phone, users can view the health status of the battery in real time, including the current deterioration index, battery capacity, and battery life estimate.

[0165] The deterioration correction term coefficient enables the model to adjust the deterioration evaluation result according to the actual charging situation, enhancing the flexibility of the evaluation; by introducing the predicted battery capacity, rated capacity, and comprehensive deterioration correction term to calculate the battery deterioration index, it can comprehensively reflect the current health status of the battery; combined with the deterioration classification function, it can achieve accurate classification of the degree of battery deterioration.

[0166] The present invention realizes the accurate prediction of the battery capacity after multiple charges of the battery by constructing a battery incremental capacity dataset and training a dynamic adaptive time series BP neural network model. On the other hand, considering that most existing lithium-ion batteries adopt a constant current and constant voltage charging strategy, and the performance of the battery during charging also reflects the degree of battery degradation, a charging dataset is constructed and a charging BP neural network model is trained. The model includes three sub-models for obtaining three degradation correction terms. The degradation correction terms are obtained and processed to obtain a comprehensive degradation correction term. The battery degradation index is obtained by predicting the battery capacity and the comprehensive degradation correction term, which is used to judge the degree of battery degradation, and the battery degradation is classified according to a preset battery degradation classification function, improving the accuracy of estimating the degree of battery degradation.

[0167] Embodiment 2

[0168] On the second-hand digital device recycling platform, the degree of battery degradation is one of the important indicators for evaluating the health status and value of the device. Especially for mobile devices such as tablet computers, the battery status directly affects its battery life and user experience. A recycling platform received a batch of second-hand tablet computers and used a lithium-ion battery degradation classification method based on BP neural network provided by the present invention to evaluate the battery degradation of the tablet computers. The method provided by the present invention can analyze the degree of battery degradation of each tablet computer device, providing an important reference basis for the subsequent pricing, repair or refurbishment of second-hand tablet computers.

[0169] First, the platform obtains the basic data of the tablet computer battery through a professional detection system. These data include the number of charging cycles of the battery, the current remaining capacity, the rated capacity, voltage, current, etc. After completing the acquisition of the basic data, the system extracts its key features by analyzing the incremental capacity curve of the battery. The incremental capacity curve can reflect the capacity change of the battery at different voltages. By analyzing information such as its peak value, valley value, and peak slope, the system can capture the degradation trend of the battery.

[0170] Through the dynamic adaptive time series BP neural network model, the data of the incremental capacity curve can be further processed to predict the battery capacity of each tablet computer. The model processes the time series characteristics of the battery through multiple LSTM layers, and can capture the dynamic changes of the battery in different charging cycles, thereby improving the prediction accuracy of the degree of battery degradation.

[0171] Furthermore, in the training stage, the calculation formula of the loss function of the dynamic adaptive time series BP neural network model is:

[0172]

[0173] where L DA-LSTM-BPNN represents the loss function of the model, H represents the number of samples, d hdenotes the true value of the h-th sample, representing the true battery capacity in this embodiment. denotes the predicted value of the h-th sample predicted by the model, representing the predicted battery capacity in this embodiment.

[0174] After obtaining the first battery charging dataset, to ensure data integrity, it is necessary to fill in and standardize the missing data. Common methods for handling missing data include mean filling, median filling, and K-nearest neighbor filling; after filling, all data vectors are standardized so that the data is distributed within the same scale range. Next, clustering analysis is performed based on the similarity of the battery data, and the data in each cluster is corrected to obtain a more accurate battery degradation assessment. In this embodiment, the Euclidean distance is used to measure the similarity between samples within a class, and the calculation formula for the similarity is:

[0175]

[0176] where denotes the a'-th data vector k in the k'-th cluster J and the b'-th data vector , U represents the length of the data vector, denotes the u-th data component of the a'-th data vector k in the k'-th cluster J , denotes the u-th data component of the b'-th data vector k in the k'-th cluster J .

[0177] During the battery charging process, the battery management system divides the charging process into a constant current stage, a constant voltage stage, and a trickle charging stage. Each stage has a different impact on battery degradation, so it is necessary to process these data separately. Based on the charging data of these stages, the charging BP neural network model can calculate the degradation correction terms respectively, so as to more accurately evaluate the health status of the battery.

[0178] After completing the battery data processing, according to the calculated comprehensive degradation correction term and the predicted battery capacity, the battery degradation index can be obtained. The battery degradation index reflects the actual attenuation of the battery and can help the platform classify the battery degradation of the tablet computer. The degradation state of the battery can usually be divided into four levels: healthy, slightly degraded, moderately degraded, and severely degraded.

[0179] For the recycling platform, the requirement for the degree of battery degradation of second-hand digital products is very high. As shown in Table 5, the relevant data such as the number of charging times, the corresponding battery degradation index, and the degree of degradation of this batch of tablet computers are presented. It can be seen from the table that as the number of charging times increases, the battery degradation index gradually decreases. Tablet computers with fewer charging times usually have good battery health, while those with more charging times have more serious battery degradation. The battery degradation classification method provided by the present invention provides a basis for the recycling platform to process this batch of second-hand tablet computers in terms of batteries.

[0180] Table 5. Relevant data of tablet computers

[0181] Tablet PC number Number of charging times Battery deterioration index Degree of deterioration A01 51 0.99 Healthy A02 102 0.97 Healthy A03 149 0.95 Healthy A04 174 0.94 Slight deterioration A05 223 0.92 Slight deterioration A06 278 0.91 Slight deterioration A07 320 0.90 Slight deterioration A08 365 0.89 Moderate deterioration

[0182] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A lithium-ion battery degradation classification method based on BP neural network, characterized in that: include: Acquire and process relevant data of the incremental capacity curve of the battery to obtain a battery incremental capacity data set; train a dynamic adaptive time-series BP neural network model according to the battery incremental capacity data set; The dynamic adaptive time series BP neural network model acquires and processes the incremental capacity data to obtain the predicted battery capacity; Acquire a first battery charging data set, wherein the first battery charging data set includes constant current charging stage data, constant voltage charging stage data, and trickle charging stage data; Acquire the first battery charging data set and perform data enhancement to obtain a second battery charging data set; Training a charging BP neural network model according to the second battery charging data set; The charging BP neural network model acquires and processes charging data to obtain a first degradation correction term, a second degradation correction term, and a third degradation correction term; Acquire the first degradation correction term, the second degradation correction term, and the third degradation correction term, and process them to obtain a comprehensive degradation correction term; The predicted battery capacity and the comprehensive degradation correction item are obtained and processed to obtain a battery degradation index, and the battery is classified according to a preset battery degradation classification function.

2. A lithium-ion battery degradation classification method based on BP neural network according to claim 1, characterized in that: The battery incremental capacity data set includes: The battery incremental capacity data set includes peak characteristics, valley characteristics, peak slope characteristics, curve area characteristics, degradation characteristics and area change on the incremental capacity curve of the battery.

3. The lithium-ion battery degradation classification method based on BP neural network according to claim 1 is characterized in that: The dynamic adaptive time series BP neural network model includes: The structure of the dynamic adaptive time series BP neural network model includes an input layer, a hidden layer, an association layer, a fusion layer and an output layer; wherein the association layer includes an LSTM layer, and the fusion layer obtains and processes the output and fusion weight of the LSTM layer to obtain a fusion feature.

4. The lithium-ion battery degradation classification method based on BP neural network according to claim 1 is characterized in that: The first battery charging data set includes: Acquire complete time series charging data, wherein the complete time series charging data includes a battery terminal voltage V(t) and a battery terminal current I(t) in the time series; The battery terminal voltage and the battery terminal current are acquired and processed according to a data division method to obtain the constant current charging stage data, the constant voltage charging stage data and the trickle charging stage data.

5. The lithium-ion battery degradation classification method based on BP neural network according to claim 1 is characterized in that: The first battery charging data set also includes: The constant current charging stage data includes: initial voltage, initial current, maximum charging voltage, voltage rise rate, constant current charging time, constant current, current fluctuation data, initial battery capacity, cumulative charge amount in the constant current charging stage, remaining power in the constant current charging stage, initial battery temperature, ambient temperature and termination temperature of the constant current charging stage; the constant voltage charging stage data includes: constant voltage, voltage fluctuation data, initial current in the constant voltage charging stage, current decay rate, cumulative charge amount in the constant voltage charging stage, remaining power in the constant voltage charging stage, initial temperature in the constant voltage stage, termination temperature in the constant voltage stage, the ambient temperature, constant voltage charging time and preset current threshold; the trickle charging stage data includes: constant voltage, voltage fluctuation data in the trickle charging stage, minimum current, cumulative charge amount in the trickle charging stage, initial temperature in the trickle charging stage, termination temperature in the trickle charging stage, the ambient temperature and trickle charging time.

6. The lithium-ion battery degradation classification method based on BP neural network according to claim 1, characterized in that: The data enhancement includes: Perform preliminary filling processing on the missing data and standardize the data to obtain a standard data set; obtain the standard data set and cluster it to obtain multiple clusters; calculate the intra-class data similarity of the missing data in the clusters, and obtain the corrected data of the missing data by weighting; All missing data are corrected to obtain a second battery charging data set.

7. The lithium-ion battery degradation classification method based on BP neural network according to claim 1, characterized in that: The charging BP neural network model includes: The charging BP neural network model includes a constant current model, a constant voltage model and a trickle model, wherein the constant current model obtains and processes the constant current charging stage data, and outputs the first degradation correction item, the constant voltage model obtains and processes the constant voltage charging stage data, and outputs the second degradation correction item, and the trickle model obtains and processes the trickle charging stage data, and outputs the third degradation correction item.

8. The lithium-ion battery degradation classification method based on BP neural network according to claim 7 is characterized in that: The charging BP neural network model also includes: The constant current model, the constant pressure model and the trickle model are all dynamic adaptive BP neural network models. The structure of the dynamic adaptive BP neural network model includes: an input layer, a first hidden layer, a second hidden layer and an output layer.

9. The lithium-ion battery degradation classification method based on BP neural network according to claim 1, characterized in that: The dynamic adaptation includes: A dynamic adaptive learning rate is introduced in the model training stage to optimize the training process of the neural network model. The calculation formula of the dynamic adaptive learning rate is: Among them, η(e) represents the learning rate of the e-th iteration, η initial represents the initial learning rate; β(e) represents the learning rate attenuation factor; Obtain the data saved in the loss window, wherein the loss window saves the loss values ​​of the most recent G iterations; obtain and process the loss values ​​to obtain an average change rate of the loss values, and the calculation formula for the average change rate of the loss values ​​is: Wherein, ΔL represents the average change rate of the loss value, L e-g+1 represents the loss value of the e-g+1th iteration, L e-g Represents the loss value of the eg-th iteration; The average change rate of the loss value is obtained and processed to obtain the learning rate attenuation factor.

10. The lithium-ion battery degradation classification method based on BP neural network according to claim 1, characterized in that: The battery categories include: The calculation formula of the battery degradation index is: Wherein, W represents the battery degradation index, map() represents the mapping function, C i represents the predicted battery capacity at the i-th charge, C0 represents the rated capacity of the battery, w represents the comprehensive degradation correction term, and μ1 represents the degradation correction term coefficient; The battery degradation index and the preset battery degradation classification function are obtained and processed to obtain a battery degradation classification result.

Images