Battery health state intelligent prediction method based on random quantization data enhancement

Abstract

A battery health state intelligent prediction method based on random quantization data enhancement comprises the steps that firstly, existing battery charging data are acquired through a BMS, linear interpolation downsampling operation is conducted on the data to unify the length of the data, and then an original training set and a test set are constructed; then, constructing a random quantizer, performing random quantization on the original training set data by using the random quantizer to generate enhanced data, and combining the obtained enhanced data with the original training set data to form an enhanced training set; thirdly, respectively inputting the data set before enhancement and the data set after enhancement into the SOH prediction model, respectively calculating and comparing the test loss average values of the data set before enhancement and the data set after enhancement, and if the test loss after enhancement is smaller than the loss before enhancement, verifying the validity of the battery health state intelligent prediction method based on random quantization data enhancement; the method can effectively solve the SOH estimation problem under the condition of limited battery data.

Description

Technical Field

[0001] The present invention belongs to the field of battery state estimation, and in particular relates to a battery health state intelligent prediction method based on random quantization data enhancement. Background Art

[0002] Whether the battery management system (BMS) can effectively intervene and manage the battery status to ensure the safe and reliable operation of the battery depends on the accurate estimation of the battery state of health (SOH). SOH is characterized by the capacity and internal resistance of the battery and is an internal parameter of the battery. Existing SOH estimation methods can generally be divided into: measurement-based methods, model-based methods, and data-driven methods ([1] Hu X, Xu L, Lin X, et al. Battery lifetime prognostics [J]. Joule, 2020, 4 (2): 310-346.). The measurement-based method relies on directly measuring the capacity and internal resistance of the battery. The test conditions are harsh and the test cycle is long, so it cannot be applied to online evaluation. The model-based method mainly estimates the SOH by establishing a physical model of the battery. It can better describe the internal characteristics of the battery, but due to the complexity and nonlinear behavior of the model, the accuracy of the result may decrease. The data-driven method mainly uses deep learning, support vector machine and other methods to estimate SOH. It requires a large amount of battery data for training to achieve accurate estimation of SOH. Compared with direct measurement and physical model methods, data-driven methods can achieve efficient and accurate estimation of SOH through big data analysis and training without detailed understanding of the internal mechanism of the battery. Therefore, the research on battery SOH using data-driven methods is increasing.

[0003] Although data-driven methods can be trained with big data and achieve high accuracy, in practical applications, the data of some battery models may be limited and cannot meet the requirements of training deep learning or other data-driven algorithms. Therefore, how to achieve accurate SOH estimation under the condition of limited battery data has become a key challenge. Summary of the invention

[0004] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a battery health status intelligent prediction method based on random quantization data enhancement, which increases the diversity and quantity of data sets through data enhancement technology, improves the generalization ability of the model, and can effectively improve the prediction accuracy of SOH when battery data is limited.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A battery health status intelligent prediction method based on random quantization data enhancement includes the following steps:

[0007] Step 1: Obtain the current battery charging data through the BMS, including voltage data (V n，k，1 , ..., V n，k,i , ..., V n，k,l ), current data (I n，k，1 , ..., I n，k，i , ..., I n，k，l ), and its corresponding time series is T n，k，1 , ..., T n，k，i , ..., T n,k，l , where i∈1,...,l represents the i-th data point of the k-th charging cycle of the n-th battery;

[0008] Step 2: Perform a linear interpolation downsampling operation on the existing battery charging data obtained in step 1, and downsample the voltage data, current data, and time series data of all batteries to the same data length l′;

[0009]

[0010]

[0011] Among them, i p ′ represents the location of the sampling point, V n，k，p , I n，k，p , T n，k，p Represents the value of the pth sampling point of the kth cycle of the nth battery after downsampling, V n，k，i , I n，k， i,T n，k，i Represents the value of the i-th value of the k-th cycle of the n-th battery before downsampling, where I is the value obtained by rounding down ip′, and I+1 is i p 'The value obtained by rounding up is

[0012] Step 3: Construct the voltage, current, and time sample data set after linear interpolation downsampling processing n = {O n，1 , O n，2 , O n，3 , …O n， k, ... O n，K}, where O n，k represents the kth cycle of the nth battery, O n，k =(V n，k , I n，k , T n，k ), that is, each sample contains three channels, and the three channels store voltage data V n，k and current data I n，k And the time data Tn,k , V n,k , I n，k , T n，k Each of them consists of l′ data points; and the original training set Train and the test set Test are divided, one battery data is selected as the test set Test, and the other battery data is used for the original training set Train;

[0013] Step 4: Construct a random quantizer, which dynamically quantizes the data in each channel, and randomly samples the quantized values ​​from randomly divided intervals;

[0014] Step 5: Determine the number of segmentation intervals m, use a random quantizer to randomly quantize the original training set Train, the voltage data, current data, and time data information in each sample are effectively retained, generate enhanced samples, and obtain the generated training set Train′, and merge the original training set Train and the generated training set Train′ generated after random quantization enhancement to obtain the enhanced training set Train aug = Train ∪ Train′;

[0015] Step 6: Construct a battery state of health (SOH) prediction model. The battery state of health (SOH) prediction model is based on the ResNet network architecture, including more than three residual blocks and a predictor. Each residual block contains two convolutional layers, and batch normalization (BatchNormalization) and ReLU activation functions are connected after the convolutional layers. Finally, the features extracted by the residual block are mapped into a scalar through the predictor of the two fully connected layers to output the battery state of health (SOH);

[0016] Step 7: Separate the original training set Train and the enhanced training set Train aug = Train ∪ Train ′ Input the battery health state (SOH) prediction model respectively, obtain the SOH prediction result, and calculate the loss between the prediction result and the actual result; the loss function uses the MSE loss function, the optimizer uses the Adam optimizer, and the pre-enhancement training loss Train is recorded respectively pre , Test loss before enhancement pre And the enhanced training loss Train pos , enhanced test loss Test pos ;

[0017] Step 8: Calculate the last e times of training loss before enhancement Train pre , Test loss before enhancement pre And the enhanced training loss Train pos , enhanced test loss Test posThe average value of ave(Train pre ) e 、ave(Test pre ) e 、ave(Train pos ) e 、ave(Test pos ) e , e represents the set number of times, if the following conditions are met:

[0018] ave(Test pre ) e >ave(Test pos ) e

[0019] This indicates that the intelligent prediction method for battery health status based on random quantization data enhancement is effective and can better solve the SOH estimation problem when battery data is limited. If it is not satisfied, reset the number of segmentation intervals m in step 6 and repeat steps 7 to 9 until ave(Test pre ) e >ave(Test pos ) e .

[0020] The step 4 is specifically as follows:

[0021] 4.1) Determine the number of random quantization intervals m and the quantization interval, obtain the minimum value min(x) and maximum value max(x) of each channel, and perform m uniformly distributed random sampling in the interval (min(x), max(x)) to obtain a j ′, and arrange them in order from small to large to obtain a 0 , a 1 , a 2 , ...a m-1 , get the quantization interval S j =[a j , a j+1 );

[0022] S j =[a j , a j+1 )

[0023] a 0 , a 1 , a 2 , ...a m-1 =sort(a 0 ′,a 1 ′,a 2 ′, ...a m-1 ′)

[0024] a j ′=U(min(x),max(x)),j=0,1,2,...,m-1

[0025] 4.2) Map all values ​​in different quantization intervals to a single value yj,y, representing a random sample uniformly distributed in the corresponding interval:

[0026] y j =U(a j , a j+1 )

[0027] 4.3) Quantize the scalar y obtained in different quantization intervals j Combine and restore the processed data to the shape of the original data.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] The present invention discloses a battery health state intelligent prediction method based on random quantization data enhancement. First, existing battery charging data is obtained through a BMS, and the data is linearly interpolated and downsampled to make them uniform to the same length, and an original training set and a test set are constructed; then a random quantizer is constructed, and the original training set is randomly quantized by the random quantizer to obtain a generated training set, and the generated training set is combined with the original training set to form an enhanced training set, and finally the enhanced data set and the pre-enhancement data set are input into the constructed SOH prediction model, and the test losses after and before enhancement are calculated respectively. When the loss after enhancement is less than the loss before enhancement, it means that the battery health state intelligent prediction method based on random quantization data enhancement is effective. The present invention effectively solves the accuracy problem of SOH estimation under the condition of limited battery data.

[0030] The present invention enhances the battery data by introducing a random quantization method, thereby effectively alleviating the problem of low SOH prediction accuracy caused by limited battery data, thereby significantly improving the prediction performance of battery SOH. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 The figure is a flow chart of a method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0033] Reference Figure 1 , a battery health status intelligent prediction method based on random quantization data enhancement, comprising the following steps:

[0034] Step 1: Obtain the current battery charging data through the BMS, including voltage data (V n，k，1, ..., V n，k,i , ..., V n，k,l ), current data (I n，k，1 , ..., I n，k，i , ..., I n，k，l ), and its corresponding time series is T n，k，1 , ..., T n，k，i , ..., T n,k，l , where I∈1,...,l represents the Ith data point of the kth charging cycle of the nth battery;

[0035] Step 2: Perform a linear interpolation downsampling operation on the existing battery charging data obtained in step 1, and downsample the voltage data, current data, and time series data of all batteries to the same data length l′;

[0036]

[0037] Among them, I p ′ represents the location of the sampling point, V n，k，p , I n，k，p , T n，k，p Represents the value of the pth sampling point of the kth cycle of the nth battery after downsampling, V n，k，i , I n，k，i , T n，k，i Represents the value of the i-th value of the k-th cycle of the n-th battery before downsampling, where i is I p 'The value obtained by rounding down, I+1 is i p 'The value obtained by rounding up is

[0038] Step 3: Construct the voltage, current, and time sample data set after linear interpolation downsampling processing n = {O n，1 , O n，2 , O n，3 , …O n，k , …O n，K}, where O n，k represents the kth cycle of the nth battery, O n，k =(V n，k , I n，k , T n，k ), that is, each sample contains three channels, and the three channels store voltage data V n，k and current data I n，k And the time data T n，k , V n,k , I n，k , T n，kEach of them consists of l′ data points; and the original training set Train and the test set Test are divided, one battery data is selected as the test set Test, and the other battery data is used for the original training set Train;

[0039] Step 4: Construct a random quantizer. The random quantizer dynamically quantizes the data in each channel, and the quantized values ​​are randomly sampled from randomly divided intervals. The specific steps are as follows:

[0040] 4.1) Determine the number of random quantization intervals m and the quantization interval, obtain the minimum value min(x) and maximum value max(x) of each channel, and perform m uniformly distributed random sampling in the interval (min(x), max(x)) to obtain a j ′, and arrange them in order from small to large to obtain a 0 , a 1 , a 2 , ...a m-1 , get the quantization interval S j =[a j , a j+1 );

[0041] S j =[a j , a j+1 )

[0042] a 0 , a 1 , a 2 , ...a m-1 =sort(a 0 ′,a 1 ′,a 2 ′, ...a m-1 ′)

[0043] a j ′=U(min(x),max(x)),j=0,1,2,...,m-1

[0044] 4.2) Map all values ​​in different quantization intervals to a single value yj,y, representing a random sample uniformly distributed in the corresponding interval;

[0045] y j =U(a j , a j+1 )

[0046] 4.3) Quantize the scalar y obtained in different quantization intervals j Combine and restore the processed data to the shape of the original data;

[0047] Step 5: Determine the number of segmentation intervals m, use a random quantizer to randomly quantize the original training set Train, the voltage data, current data, and time data information in each sample are effectively retained, generate enhanced samples, and obtain the generated training set Train′, and merge the original training set Train and the generated training set Train′ generated after random quantization enhancement to obtain the enhanced training set Train aug = Train ∪ Train′;

[0048] Step 6: Construct a battery state of health (SOH) prediction model. The battery state of health (SOH) prediction model is based on the ResNet network architecture, including five residual blocks and a predictor. Each residual block contains two convolutional layers, and batch normalization (BatchNormalization) and ReLU activation functions are connected after the convolutional layer to ensure the training stability and nonlinear expression ability of the model; finally, through the predictor of two fully connected layers, the features extracted by the residual block are mapped to a scalar for outputting the battery state of health (SOH);

[0049] Step 7: Separate the original training set Train and the enhanced training set Train aug = Train ∪ Train ′ Input the battery health state (SOH) prediction model respectively, obtain the SOH prediction result, and calculate the loss between the prediction result and the actual result; the loss function uses the MSE loss function, the optimizer uses the Adam optimizer, and the pre-enhancement training loss Train is recorded respectively pre , Test loss before enhancement pre And the enhanced training loss Train pos , enhanced test loss Test pos ;

[0050] Step 8: Calculate the last e times of training loss before enhancement Train pre , Test loss before enhancement pre And the enhanced training loss Train pos , enhanced test loss Test pos The average value of ave(Train pre ) e 、ave(Test pre ) e 、ave(Train pos ) e 、ave(Test pos ) e , e represents the set number of times, if the following conditions are met:

[0051] ave(Testpre ) e >ave(Test pos ) e

[0052] This indicates that the intelligent prediction method for battery health status based on random quantization data enhancement is effective and can better solve the SOH estimation problem when battery data is limited. If it is not satisfied, reset the number of segmentation intervals m in step 6 and repeat steps 7 to 9 until ave(Test pre ) e >ave(Test pos ) e .

[0053] The embodiment is based on a Tesla car battery data set from a laboratory to verify the effectiveness of the method of the present invention. The data set is verified using a Tesla battery data set from a laboratory. The nominal capacity of the battery is 2.7Ah, the nominal voltage is 3.6V, the charging cut-off voltage and the discharging cut-off voltage are 4.2V and 2.0V respectively. The entire experiment is carried out at room temperature; 5 batteries in a certain pack group are selected for k-fold cross validation, and one of the battery data is selected as the test set, and the other battery data is used as the original training set. Repeat 5 groups of experiments; the mean square error (MSE) indicator is used to measure the effect of the data enhancement of the present invention. The experimental results are shown in Table 1:

[0054] Table 1

[0055]

[0056] The results of converting it into the form of SOH absolute error are shown in Table 2:

[0057] Table 2

[0058]

[0059]

[0060] It can be found from Tables 1 and 2 that the test data set loss generated by the data set enhanced by the random quantization method is smaller than the loss before enhancement, indicating that the method proposed in the present invention can effectively solve the SOH prediction problem when battery data is limited.

[0061] The present invention is suitable for SOH evaluation when battery data is limited. When battery data is limited, the present invention can enhance the existing data through random quantization, which can improve the accuracy of model prediction of battery SOH to a certain extent.

Claims

1. A battery health status intelligent prediction method based on random quantization data enhancement, characterized in that: The following steps are involved: Step 1: Obtain the current battery charging data through the BMS, including voltage data (V n,k,1 ,…,V n,k,i ,…,V n,k,l ), current data (I n,k,1 ,…,I n,k,i ,…,I n,k,l ), and its corresponding time series is T n,k,1 ,…,T n,k,i ,…,T n,k,l , where i∈1,…,l represents the i-th data point of the K-th charging cycle of the n-th battery; Step 2: Perform a linear interpolation downsampling operation on the existing battery charging data obtained in step 1, and downsample the voltage data, current data, and time series data of all batteries to the same data length l'; Among them, i p ' represents the location of the sampling point, V n,k,p ,I n,k,p ,T n,k,p Represents the value of the pth sampling point of the kth cycle of the nth battery after downsampling, V n,k,i ,I n,k,i ,T n,k,i Represents the value of the i-th value of the k-th cycle of the n-th battery before downsampling, where i is i p 'Round down to get the value, i+1 is i p 'The value obtained by rounding up is Step 3: Construct the voltage, current, and time sample data set after linear interpolation downsampling processing n = {O n,1 ,O n,2 ,O n,3 ,···O n,k ,···O n,K }, where O n,k represents the kth cycle of the nth battery, O n,k =(V n,k ,I n,k ,T n,k ), that is, each sample contains three channels, and the three channels store voltage data V n,k and current data I n,k And the time data T n,k , V n,k ,I n,k ,T n,k Each of them consists of l' data points; and the original training set Train and the test set Test are divided, one battery data is selected as the test set Test, and the other battery data is used for the original training set Train; Step 4: Construct a random quantizer, which dynamically quantizes the data in each channel, and randomly samples the quantized values ​​from randomly divided intervals; Step 5: Determine the number of segmentation intervals m, use a random quantizer to randomly quantize the original training set Train, the voltage data, current data, and time data information in each sample are effectively retained, generate enhanced samples, and obtain the generated training set Train', and merge the original training set Train and the generated training set Train' generated after random quantization enhancement to obtain the enhanced training set Train aug = Train ∪ Train'; Step 6: Construct a battery state of health (SOH) prediction model. The battery state of health (SOH) prediction model is based on the ResNet network architecture, including more than three residual blocks and a predictor. Each residual block contains two convolutional layers, and batch normalization (Batch Normalization) and ReLU activation functions are connected after the convolutional layers. Finally, the features extracted by the residual block are mapped to a scalar through the predictor of the two fully connected layers to output the battery state of health (SOH); Step 7: Separate the original training set Train and the enhanced training set Train aug = Train ∪ Train' Input the battery health status (SOH) prediction model respectively, obtain the SOH prediction result, and calculate the loss between the prediction result and the actual result; the loss function uses the MSE loss function, the optimizer uses the Adam optimizer, and the pre-enhancement training loss Train is recorded respectively pre , Test loss before enhancement pre And the enhanced training loss Train pos , enhanced test loss Test pos ; Step 8: Calculate the last e times of training loss before enhancement Train pre , Test loss before enhancement pre And the enhanced training loss Train pos , enhanced test loss Test pos The average value of ave(Train pre ) e 、ave(Test pre ) e 、ave(Train pos ) e 、ave(Test pos ) e , e represents the set number of times, if the following conditions are met: ave(Test pre ) e >ave(Test pos ) e This indicates that the intelligent prediction method for battery health status based on random quantization data enhancement is effective and can better solve the SOH estimation problem when battery data is limited. If it is not satisfied, reset the number of segmentation intervals m in step 6 and repeat steps 7 to 9 until ave(Test pre ) e >ave(Test pos ) e .

2. The battery health status intelligent prediction method according to claim 1, characterized in that: The step 4 is specifically as follows: 4.1) Determine the number of random quantization intervals m and the quantization interval, obtain the minimum value min(x) and maximum value max(x) of each channel, and perform m uniformly distributed random sampling in the interval (min(x), max(x)) to obtain a j ', and arrange them in ascending order to obtain a0, a1, a2, ... a m-1 , get the quantization interval S j =[a j ,a j+1 ); S j =[a j ,a j+1 ) <h2 style=";text-align:left;direction:ltr">a0,a1,a2,…a<h2 style=";text-align:left;direction:ltr"> m-1 <h2 style=";text-align:left;direction:ltr"> =sort(a0',a1',a2',…a<h2 style=";text-align:left;direction:ltr"> m-1 <h2 style=";text-align:left;direction:ltr"> ') a j '=U(min(x),max(x)),j=0,1,2,…,m-1 4.2) Map all values ​​in different quantization intervals to a single value y j ,y j represents a uniformly distributed random sample in the corresponding interval: y j =U(a j ,a j+1 ) 4.3) Quantize the scalar y obtained in different quantization intervals j Combine and restore the processed data to the shape of the original data.

Images