CN116794516A

CN116794516A - Method for predicting residual service life of lithium ion battery based on algorithm fusion

Info

Publication number: CN116794516A
Application number: CN202310576576.1A
Authority: CN
Inventors: 汤易; 牟健慧; 高立喜; 王书豪
Original assignee: Yantai University
Current assignee: Yantai University
Priority date: 2023-05-22
Filing date: 2023-05-22
Publication date: 2023-09-22

Abstract

The invention belongs to the technical field of lithium ion battery remaining service life prediction, and relates to a lithium ion battery remaining service life prediction method based on algorithm fusion. According to the invention, a NASA (non-self-adaptive analysis) published 18650 type lithium ion battery test data set is adopted, a Pelson correlation coefficient method and a gray correlation analysis method are introduced to extract characteristic data with strong correlation with battery capacity indexes according to the change condition of various indexes of the battery during charge and discharge, and a health factor based on fusion of two time sequences is provided to extract the most simplified data for residual service life prediction and reveal the internal rule of battery service life attenuation. The feature data provided based on the preprocessing method is applied to a fusion prediction model based on CNN-BiGRU, so that the residual service life of the lithium ion battery is predicted. Compared with other two single model prediction methods, the method has the advantages of minimum prediction error and more accurate prediction result.

Description

Method for predicting residual service life of lithium ion battery based on algorithm fusion

Technical Field

The invention belongs to the technical field of lithium ion battery remaining service life prediction, and relates to a lithium ion battery remaining service life prediction method based on algorithm fusion.

Technical Field

Energy conservation and emission reduction are perpetual subjects of automobiles. The lithium ion battery has the advantages of good safety, low self-discharge rate, high energy density, long cycle life, no memory effect, strong economical efficiency and the like, and becomes the most competitive power supply variety of the new energy automobile. However, the performance of the lithium battery is reduced due to continuous charge and discharge during the use process, and particularly when the actual capacity is lower than 70% of the rated capacity, various performance indexes of the battery are seriously reduced, and the safety problem is more remarkable. Therefore, fast and accurate prediction of the remaining service life (Remaining Useful Life, RUL) of the lithium battery and realization of safe and reliable battery health management and fault early warning are one of hot spots and challenges facing the current industry, and have very important research significance.

Regarding the RUL prediction of lithium ion batteries, students at home and abroad mainly develop research work from the following three aspects: model-based methods, data-based methods, and fusion-based methods. The method based on model driving is divided into an electrochemical model, an equivalent circuit model, an empirical model driving method and the like, and the method requires researchers to have rich development experience. Research methods based on data driving benefit from the continuous progress of artificial intelligence (Artificial Intelligence, AI) and machine learning in recent years, and the requirements of the machine learning method on a large amount of data are met by combining the large amount of data generated in the battery production and manufacturing and experimental processes. The method based on the fusion of the two methods needs to fully develop the advantages of the two driving methods to improve the accuracy of the machine learning prediction model, and becomes the main research direction of domestic and foreign scholars at the present stage. Because the data research on the input model is less at present, the life of the lithium ion battery is gradually attenuated along with the continuous increase of the charge and discharge cycle times, potential safety hazards such as battery short circuit and explosion exist, and the existing prediction method is influenced by the battery experimental data noise and the defects of a single algorithm model, the phenomena such as gradient elimination and gradient explosion can occur, and the prediction method has low precision and inaccurate prediction results. Therefore, research and development of a machine learning prediction method based on algorithm fusion are urgently needed to improve the accuracy of lithium ion battery residual service life prediction.

Disclosure of Invention

The invention provides a method for predicting the residual service life of a lithium ion battery based on algorithm fusion. The data preprocessing adopts a multi-feature extraction and simulation mode, and characteristic parameters are efficiently mined by analyzing the internal rules of the acquired data, so that parameter simplification and reduction are realized, and the preprocessed data is used as a trusted input variable of a machine learning model. The RUL prediction adopts a CNN-BiGRU fused algorithm model, so that the defects of a single model are overcome, and the accuracy and the calculation efficiency of battery life prediction are improved.

The technical scheme of the invention is as follows:

a lithium ion battery residual service life prediction method based on algorithm fusion comprises the following steps:

step 1, data selection and extraction: and extracting effective information such as voltage, current and temperature in all charge and discharge cycles of the lithium ion batteries B0005, B0006, B0007 and B0018 by adopting a 18650 lithium ion battery detection data set with rated capacity of 2Ah disclosed by NASA. In the charging process, the batteries are charged in a constant current mode, and a constant voltage charging mode is maintained when the voltage reaches the maximum value until the current is reduced to 20mA, and the charging is finished; in the discharging process, the battery is discharged in a constant current manner until the battery is reduced to a certain value. And respectively extracting the terminal voltage, the charging output current, the charging temperature, the charging measurement voltage, the charging measurement current, the discharging battery terminal voltage, the discharging output current, the discharging temperature, the discharging battery load voltage and the discharging load current of the rechargeable battery in the constant-current constant-voltage charging stage, the constant-current discharging stage and the battery temperature change as battery performance indexes.

Step 2, data analysis: and selecting data of the lithium ion battery in a charging and discharging cycle process for at least 20 times, and performing visual analysis on the selected data. And analyzing the change conditions of the voltage, the current and the temperature of the lithium ion battery in the three complete charge and discharge cycle processes. The curves of the parameters relative to time are found by analysis to have more obvious points, such as maximum values, minimum values, inflection points, sudden increases or sudden decreases, and the like; and finding out the change rule of the characteristic points in three complete charge and discharge cycles and the corresponding time of the characteristic points through analysis.

Step 3, extracting features: and extracting the characteristics by a data analysis and a Pearson correlation coefficient method. Extracting characteristic point data representing each period of the battery, and requiring the data point to have high reliability and accuracy along with the aging process of the battery. And extracting the characteristic data and the corresponding time to represent the health state of the lithium ion battery, and verifying the correlation between the extracted characteristic and the battery capacity.

Step 4, constructing health factors: and reconstructing an isobaric discharging time sequence and an isothermal elevating discharging time sequence according to correlation analysis aiming at the discharging voltage and the discharging temperature index of the lithium ion battery to represent the characteristic of the degradation of the lithium ion battery, and carrying out normalization treatment on the characteristic parameters. Fitting the normalized isobaric descending time sequence and isothermal ascending time sequence with a battery capacity sequence, analyzing the fitting degree, re-scrambling and sequencing the isobaric descending time sequence and the isothermal ascending time sequence according to an analysis result, and performing feature simulation and feature fusion to construct a health factor based on the fusion of the isobaric descending time sequence and the isothermal ascending time sequence of the discharge battery.

The method comprises the following specific steps:

4.1 judging the association degree by using a gray analysis method through the curve shape of each health factor, wherein the gray association analysis method comprises the following formula:

wherein y (k) is a battery capacity sequence, k is a sequence length, x _i (k) For the isobaric discharge time series or isothermal rise time series, i represents the number of sequences, ρ represents the resolution coefficient, where ρ=0.5. The association solving relational expression is as follows:

wherein r is _i In order for the degree of association to be the degree of association,0＜r _i < 1. The more the correlation value goes to 1, the greater the correlation of the constructed health factor with the battery capacity. The correlation between each characteristic and the battery capacity can be obtained by simply extracting and calculating the data.

Step 4.2, normalization: because the battery capacity sequence, the isobaric descending time sequence and the isothermal ascending time sequence are different in size but consistent in comparison standard, the battery capacity sequence, the isobaric descending time sequence and the isothermal ascending time sequence are normalized, and the data normalization is to scale the extracted data between smaller areas so as to eliminate the influence caused by the different sizes among the data. The invention adopts a minimum-maximum normalization method to enable the value range of the extracted characteristic to fall between [0,1], and the normalization method can improve the convergence rate of the model and improve the prediction precision of the model, as shown in figure 4. The min-max normalization method formula is:

In which x is _min To the minimum value of the extracted features, x _max Is the maximum of the extracted features.

And 4.3, carrying out visual processing on the health factor, the battery capacity sequence, the isobaric descending time sequence and the isothermal ascending time sequence, and finding that the constructed health factor is closer to the battery capacity sequence curve than the isobaric descending time sequence curve and the isothermal ascending time sequence curve in the whole charge-discharge cycle period of the lithium ion battery, so that the constructed health factor can represent the degradation condition of the battery capacity more than the original index.

And 5, evaluating the comprehensive effect of the health factors fused with the equivalent pressure drop time sequence and the isothermal rise time sequence by adopting root mean square error (Root Mean Squared Error, RMSE), average absolute error (Mean Absolute Error, MAE) and R square (R-squared) to obtain the battery capacity sequence.

In the fifth step, the specific operation is as follows:

the capacity root mean square error assessment mechanism for judging the deviation between the predicted value and the true value can be used for assessing the accuracy of the comprehensive effect of the health factor. The root mean square error is expressed as:

wherein y is _i Is a capacity sequence value; x is x _i Is a health factor sequence value; n is the number of cycles.

The average absolute error of the deviation of the predicted value and the true value is judged, so that the actual condition of the error can be better reflected.

And R square is used for measuring the fitting degree of the predicted value and the true value, and the predicted value and the true value have strong positive correlation when the correlation is closer to 1.

In the middle ofIs the average of the capacity sequences.

Step 6, preprocessing data: the constructed health factor is integrated with the extracted feature data.

In the step 6, the specific steps are as follows:

the extracted characteristic data of 6 dimensions of the voltage characteristic, the current characteristic and the temperature characteristic of the charge and discharge battery are classified according to the batteries B0005, B0006, B0007 and B0018 to form a data set to be processed.

Step 7, constructing a CNN-BiGRU fusion model:

dividing the data set to be processed generated in the sixth step into a training set and a testing set, selecting the battery capacity as a target output, constructing a CNN-BiGRU fusion model, and setting model super-parameters. Substituting test data of the test set into a CNN-BiGRU prediction model to obtain a residual service life prediction curve of the lithium ion battery.

The CNN-BiGRU fusion model comprises a characteristic data input layer, a CNN layer, a BiGRU layer and a characteristic data output layer.

The function of the characteristic data input layer is to analyze the data, determine the dimension and output of the characteristic variable, divide the training set and the testing set, normalize the data and input the characteristic data into the neural network model.

The data processed by the characteristic data input layer firstly enter the CNN layer to carry out two-round convolution calculation, then enter the full-connection layer to carry out time sequence expansion, and then the characteristic data is transmitted to the BiGRU layer.

The BiGRU layer further extracts local features of the feature data, learns and trains the extracted feature data, predicts the residual service life of the battery, and outputs a predicted result after inverse normalization.

The CNN-BiGRU neural network adopts two layers of convolution and bidirectional GRU for calculation. The convolution kernel size of the first layer of convolution setting is 3 multiplied by 1, the number of filters is 64, and each convolution kernel can perform dot product operation when the convolution operation is performed, so that characteristic data is learned, a filling mode is added at the same time, zero padding operation is performed on input characteristic data, a relu activation function is used after the convolution operation and is applied to the characteristic data obtained by the convolution operation, nonlinear processing is performed on the characteristic data, and therefore convergence of a neural network is enhanced. The second layer convolution sets a convolution kernel size of 3 x 1 and a number of filters of 32. The data obtained by the first layer convolution is used as input data of the second layer convolution, the first layer convolution operation is repeated, and the Padding processing and the relu activation function are performed. The BiGRU neural network collects data information in the forward direction and the reverse direction of the time sequence, has the capability of the hidden layers in two directions for memorizing and removing the information, increases the flexibility of a model, and reduces the accidental of characteristic data of the same time sequence. Two standard GRU neural networks forming a BiGRU neural network, one GRU neural network uses data of historical time series for forward training, and the other GRU neural network uses data of future time series for reverse training. And obtaining the final output of the BiGRU neural network under a certain time sequence by fusing the output characteristics obtained by forward training and the output characteristics obtained by reverse training. The BiGRU neural network can enable each hidden layer to extract historical data and future data in a specific step length, so that the BiGRU neural network can acquire more comprehensive characteristic data, and the prediction performance of the BiGRU neural network is greatly improved.

In the step 1, the voltage, current and temperature change conditions of the lithium ion battery in the 30 th, 60 th and 120 th charge-discharge cycle processes are extracted, and visual treatment is carried out to show the change trend.

In the step 3, a formula method and a pearson correlation coefficient method are adopted to extract the characteristics. The pearson correlation coefficient is an evaluation index of the degree of correlation between different variables, the value of the pearson correlation coefficient is in the range from-1 to 1, and the pearson correlation coefficient is a linear correlation analysis method. The formula of the pearson correlation coefficient method is:

wherein Y represents battery capacity degradation data, and X represents relevant data of each extracted feature. The larger the absolute value of the correlation coefficient value, the stronger the correlation between the battery capacity degradation data and the extracted feature data. The correlation between each characteristic and the battery capacity can be obtained by simply extracting and calculating the data.

The invention has the beneficial effects that:

the invention analyzes 18650 type lithium ion battery detection data set disclosed by NASAPCoE research center in the United states, introduces a Pelson correlation coefficient method and a gray correlation analysis method to extract characteristic data with strong correlation with battery capacity indexes, and provides a health factor based on fusion of two time sequences to extract the most simplified data for residual service life prediction and reveal the internal rule of battery life attenuation. The feature data provided based on the preprocessing method is applied to a fusion prediction model based on CNN-BiGRU, so that the residual service life of the lithium ion battery is predicted. The method solves the problems of inaccurate feature selection, redundant feature data, high data preprocessing error, defects of the model and high model prediction error of the existing method, realizes parameter simplification and reduction, and improves the accuracy and calculation efficiency of battery life prediction.

Drawings

FIG. 1 is a schematic diagram of the overall process of the method of the present invention.

Fig. 2 is a B0005 battery data set architecture diagram.

Fig. 3 is a graph of a B0005 one-time charge-discharge experiment.

Fig. 4 (a) is a graph of the battery terminal voltage at the state of charge of a B0005 lithium ion battery.

Fig. 4 (B) is a graph of the battery output current at the state of charge of a B0005 lithium ion battery.

Fig. 4 (c) is a graph of battery temperature at the state of charge of a B0005 lithium ion battery.

Fig. 4 (d) is a graph of the measured voltage of the battery in the state of charge of the B0005 lithium ion battery.

Fig. 4 (e) is a graph of the charge measurement current at the state of charge of a B0005 lithium ion battery.

Fig. 5 (a) is a graph of the cell terminal voltage in the discharge state of a B0005 lithium ion battery.

Fig. 5 (B) is a graph of the battery output current in the discharge state of the B0005 lithium ion battery.

Fig. 5 (c) is a graph of the battery temperature in the discharge state of the B0005 lithium ion battery.

Fig. 5 (d) is a graph of the battery load voltage in the discharge state of a B0005 lithium ion battery.

Fig. 5 (e) is a graph of the load current in the discharge state of a B0005 lithium ion battery.

Fig. 6 (a) is a B0005 rechargeable battery voltage profile.

Fig. 6 (B) is a B0005 rechargeable battery current signature.

Fig. 6 (c) is a B0005 rechargeable battery temperature profile.

Fig. 6 (d) is a B0005 discharge cell voltage signature.

Fig. 6 (e) is a B0005 discharge cell current signature.

Fig. 6 (f) is a B0005 discharge cell temperature profile.

Fig. 6 (g) is a B0005 battery capacity fade curve.

Fig. 7 (a) is a graph of constructed time series versus battery capacity series trend.

FIG. 7 (b) is a graph showing a comparison of a sequence curve containing health factors.

Fig. 8 is a flow chart for predicting the remaining useful life of CNN.

Fig. 9 (a) is a residual life prediction diagram.

Fig. 9 (b) is a residual life prediction error map.

Fig. 10 is a diagram of a biglu neural network.

Fig. 11 (a) is a diagram of residual life prediction results of biglu iteration 100 times.

Fig. 11 (b) is a residual life prediction error map for biglu iteration 100 times.

Fig. 12 (a) is a graph of residual life prediction results for 160 biglu iterations.

Fig. 12 (b) is a residual life prediction error plot for biglu iteration 160 times.

Fig. 13 (a) is a diagram showing a residual lifetime prediction result of biglu.

Fig. 13 (b) is a residual life prediction error map of biglu.

FIG. 14 is a diagram of a CNN-BiGRU neural network.

FIG. 15 (a) is a graph showing the result of predicting the remaining useful life of CNN-BiGRU.

FIG. 15 (b) is a graph of residual life prediction error for CNN-BiGRU.

FIG. 16 (a) is a graph showing the predicted result of CNN-BiGRU after 30 th time.

FIG. 16 (b) is a graph showing the predicted result of CNN-BiGRU at time 60.

Detailed Description

The invention is described in detail below with reference to the drawings and the detailed description.

According to the method for predicting the residual service life of the lithium ion battery based on algorithm fusion, as shown in fig. 1, a 18650 lithium ion battery test data set with rated capacity of 2Ah disclosed by NASA is adopted, multi-feature extraction, simulation and fusion pretreatment method research is carried out on experimental data of a B0005 battery under a plurality of charge and discharge cycles, the pretreated data are divided into a training set and a testing set, the battery capacity is selected as a target output, and the target output is brought into a CNN-BiGRU fusion algorithm model to predict the residual service life of the lithium ion battery. The pearson correlation coefficient method is introduced to extract characteristic data with strong correlation with battery capacity indexes, and a health factor based on fusion of two time sequences is provided, and the feasibility of the health factor is judged by using error analysis methods such as root mean square error, average absolute error, R square and the like. The feature data extracted based on the data preprocessing method is brought into the fused CNN-BiGRU for training and testing, the residual service life of the lithium ion battery is predicted, the model is subjected to a segmentation test, and the model is compared with a classical single neural network model to verify the effectiveness and superiority of the model.

The method specifically comprises the following steps:

step 1, data selection and extraction, as shown in fig. 2 and fig. 3, specifically comprises the following steps:

step 1.1, as shown in fig. 2, a 18650 lithium ion battery detection data set with rated capacity of 2Ah disclosed by NASA is adopted to extract effective information such as voltage, current and temperature in all charge and discharge cycles of the lithium ion batteries B0005, B0006, B0007 and B0018.

Step 1.2, as shown in fig. 3, in the charging process, the batteries are charged with constant current, and when the voltage reaches the maximum value, the constant voltage charging mode is maintained until the current is reduced to 20mA, and the charging is finished; in the discharging process, the battery is discharged in a constant current manner until the battery is reduced to a certain value. And respectively extracting the terminal voltage, the charging output current, the charging temperature, the charging measurement voltage, the charging measurement current, the discharging battery terminal voltage, the discharging output current, the discharging temperature, the discharging battery load voltage and the discharging load current of the rechargeable battery in the constant-current constant-voltage charging stage, the constant-current discharging stage and the battery temperature change as battery performance indexes.

Step 2, data analysis, as shown in fig. 4 and fig. 5, specifically the steps are as follows:

step 2.1, after the end voltage of the rechargeable battery rises to the highest point in the charging process, maintaining a constant voltage state, gradually reducing the time when the end voltage reaches the maximum value along with the increase of the charging times, maintaining the output current at 1.5A, gradually advancing the time when the current decreases along with the change of the charging time, shortening the time when the temperature of the battery reaches the highest point along with the increase of the charging and discharging times, gradually increasing the peak value of the temperature, gradually advancing the time when the measured voltage of the rechargeable battery reaches the maximum value along with the increase of the charging and discharging times, and enabling the measured charging current to be approximately the same as the output current curve of the rechargeable battery.

Step 2.2, in the discharging process, along with the increase of the cycle times, the time for the battery voltage to drop to the lowest point is gradually shortened, along with the increase of the charge and discharge times, the voltage valley value is slightly increased, the output current of the discharging battery is subjected to constant current discharging, the discharging time is gradually shortened, the temperature rising speed of the discharging battery is continuously accelerated along with the increase of the charge and discharge times, the battery temperature is increased, the time for the load voltage of the discharging battery to be reduced to 0V is gradually shortened, and the discharging load current curve is approximately the same as the discharging battery output current curve.

Step 3, feature extraction, as shown in fig. 6, specifically comprises the following steps:

in step 3.1, taking the B0005 battery as an example, during the charging process of the lithium ion battery, the terminal voltage of the rechargeable battery is gradually increased to 4.2V along with the charging time, and then the voltage is charged in a constant voltage charging mode. As can be seen from the voltage profile of a complete charge process, there are abrupt points in the geometric curve, i.e. when the battery charging terminal voltage reaches 4.2V. With the increase of the charge and discharge cycle times, the corresponding moments when the rechargeable battery reaches the abrupt change point are different, and the changed moments have close relation with the residual service life of the lithium ion battery. In all charging cycles, the time point corresponding to the point when the battery terminal voltage reaches the abrupt point is extracted and is taken as the characteristic extracted by the battery terminal voltage of the rechargeable battery, and the characteristic is marked as the characteristic of the battery terminal voltage of the B0005 rechargeable battery. By the above, the output current characteristic of the B0005 rechargeable battery, the temperature characteristic of the B0005 rechargeable battery, the measured voltage characteristic of the B0005 rechargeable battery and the measured current characteristic of the B0005 rechargeable battery are extracted.

In step 3.2, taking the B0005 battery as an example, in the discharging process of the lithium ion battery, the voltage of the end of the discharging battery gradually drops to the lowest point along with the discharging time, and then the voltage value is slightly and temporarily increased. It can be seen from the voltage change graph of a complete discharge process that there are abrupt points in the geometric curve, i.e., when the voltage at the end of the discharge cell drops to a minimum. With the increase of the charge-discharge cycle times, the corresponding moments when the discharge battery reaches the abrupt change point are different, and the changed moments have close relation with the residual service life of the lithium ion battery. In all discharging cycles, the point of time corresponding to the point of time when the battery terminal voltage reaches the abrupt point is extracted as the characteristic extracted from the discharging battery terminal voltage, and is denoted as the characteristic of the B0005 discharging battery terminal voltage. By this, the B0005 discharge battery output current characteristic, the B0005 discharge battery temperature characteristic, the B0005 discharge battery load voltage characteristic, and the B0005 discharge battery load current characteristic are extracted.

And 3.3, performing correlation analysis on the extracted characteristic data by using a pearson correlation coefficient method, wherein the pearson correlation coefficient method has the formula:

And 3.4, through correlation analysis and calculation, selecting the voltage characteristic of the B0005 rechargeable battery, the current characteristic of the B0005 rechargeable battery, the temperature characteristic of the B0005 rechargeable battery, the voltage characteristic of the B0005 discharging battery, the current characteristic of the B0005 discharging battery and the temperature characteristic of the B0005 discharging battery as battery indexes extracted by the characteristics, wherein the values of the Pearson correlation coefficients are all larger than 0.99 and are superior to other battery indexes.

Step 4, constructing health factors, as shown in fig. 7, specifically comprising the following steps:

in the step 4.1, in the charging process, the terminal voltage of the rechargeable battery is selected from a 0V-4.2V interval, and is marked as T ₁ The method comprises the steps of carrying out a first treatment on the surface of the The output current of the rechargeable battery is selected from 1.5A-0A interval, and is marked as T ₂ The method comprises the steps of carrying out a first treatment on the surface of the The temperature of the rechargeable battery is selected from 24V-28V interval, and is recorded as T ₃ The method comprises the steps of carrying out a first treatment on the surface of the The measured voltage of the rechargeable battery is selected from 4V-4.9V interval, and is recorded as T ₄ The method comprises the steps of carrying out a first treatment on the surface of the The charging measurement voltage is selected from 1.5A-0A interval and is marked as T ₅ . The extraction formula is as follows:

T ₁ ＝T _4.2V -T _0V

T ₂ ＝T _0A -T _1.5A

T ₃ ＝T _27℃ -T _24℃

T ₄ ＝T _4.9V -T _4V

T ₅ ＝T _0A -T _1.5A

In the discharging process, the end voltage of the discharging battery is selected from a 4.2V-2.7V interval, and is recorded as T ₆ The method comprises the steps of carrying out a first treatment on the surface of the The output current of the discharge battery is selected from the interval of-2A-0A and is marked as T ₇ The method comprises the steps of carrying out a first treatment on the surface of the The temperature of the discharge battery is selected from 24-38 ℃ interval, and is marked as T ₈ The method comprises the steps of carrying out a first treatment on the surface of the The load voltage of the discharge battery is selected from 4.2V-0V interval, and is recorded as T ₉ The method comprises the steps of carrying out a first treatment on the surface of the The discharge load current is selected from the interval of-2A-0A, and is marked as T ₁₀ . The extraction formula is as follows:

T ₆ ＝T _2.7V -T _4.2V

T ₇ ＝T _0A -T _-2A

T ₈ ＝T _38℃ -T _24℃

T ₉ ＝T _0V -T _4.2V

T ₁₀ ＝T _0A -T _-2A

wherein T is ₁ -T ₁₀ T is the time sequence extracted for each factor _0V 、T _2.7V 、T _4V 、T _4.2V And T _4.9V T is the time when the voltage reaches 0V, 2.7V, 4V, 4.2V and 4.9V respectively _-2A 、T _0A And T _1.5A The moments when the currents reach-2A, 0A and 1.5A, respectively，T _38℃ 、T _27℃ And T _24℃ The times at which the cell temperature reached 38 ℃, 27 ℃ and 24 ℃, respectively.

And 4.3, judging the association degree through the curve shape of each health factor by using a gray analysis method, wherein the gray association analysis method comprises the following formula:

wherein r is _i For the degree of association, 0 < r _i < 1. The more the correlation value goes to 1, the greater the correlation of the constructed health factor with the battery capacity. The correlation between each characteristic and the battery capacity can be obtained by simply extracting and calculating the data.

And 4.4, according to the correlation analysis of the constructed time sequence and the battery capacity sequence, the highest correlation of the discharge battery terminal voltage, the discharge battery temperature time sequence and the battery capacity sequence can be obtained, so that the two time sequences of the discharge battery terminal voltage and the discharge battery temperature are extracted.

And 4.5, extracting the isobaric dropping discharge time sequence as a health factor for representing the battery capacity. And (3) performing correlation analysis on an isobaric drop time sequence of the discharge battery voltage divided into three sections of 4.2V-3.9V, 3.9V-3V and 3V-2.7V and a battery capacity sequence, wherein the isobaric discharge time is expressed as follows:

T _4.2V-3.9V ＝T _3.9V -T _4.2V

T _3.9V-3V ＝T _3V -T _3.9V

T _3V-2.7V ＝T _2.7V -T _3V

t in _4.2V-3.9V 、T _3.9V-3V 、T _3V-2.7V Respectively, the time required for the voltage to decrease from 4.2V to 3.9V, the time required for the voltage to decrease from 3.9V to 3V, and the time required for the voltage to decrease from 3V to 2.7V; t (T) _4.2V 、T _3.9V 、T _3V 、T _2.7V The time points corresponding to the voltage reduction to 4.2V, 3.9V, 3V and 2.7V are respectively shown.

Will T ₆ 、T _4.2V-3.9V 、T _3.9V-3V 、T _3V-2.7V Correlation analysis with a battery capacity sequence shows that T is ₆ 、T _3.9V-3V The correlation with the battery capacity sequence was all 0.99.T (T) _4.2V-3.9V 、T _3V-2.7V The correlation with the battery capacity sequence is weak, and the interval section comprises the beginning and the end of the battery discharge, and the data noise can be caused by the influence of external environment factors, so T is taken _3.9V-3V The interval section is used as an index for representing the change of the battery capacity along with the charge-discharge cycle.

Step 4.6, dividing the temperature into three sections of 24-29 ℃, 29-33 ℃ and 33-38 ℃, and carrying out correlation analysis on the corresponding time sequence and battery capacity sequence, wherein the isothermal ascending and discharging time sequence formula is as follows:

T _24℃-29℃ ＝T _29℃ -T _24℃

T _29℃-33℃ ＝T _33℃ -T _29℃

T _33℃-38℃ ＝T _38℃ -T _33℃

t in _24℃-29℃ 、T _29℃-33℃ 、T _33℃-38℃ Respectively, the time required for the temperature to rise from 24 ℃ to 29 ℃, the time required for the temperature to rise from 29 ℃ to 33 ℃ and the time required for the temperature to rise from 33 ℃ to 38 ℃; t (T) _24℃ 、T _29℃ 、T _33℃ 、T _38℃ The temperature was raised to 24 ℃, 29 ℃, 33 ℃, 38 ℃ at the corresponding time points. Will T ₈ 、T _24℃-29℃ 、T _29℃-33℃ 、T _33℃-38℃ From the correlation analysis with the battery capacity sequence, it can be seen that T ₈ The highest correlation with the battery capacity sequence is achieved, so T is still chosen ₈ The interval section is used as an index for representing the change of the battery capacity along with the charge-discharge cycle.

Step 4.7, normalization: because the battery capacity sequence, the isobaric descending time sequence and the isothermal ascending time sequence are different in size but consistent in comparison standard, the battery capacity sequence, the isobaric descending time sequence and the isothermal ascending time sequence are normalized, and the data normalization is to scale the extracted data between smaller areas so as to eliminate the influence caused by the different sizes among the data. The invention adopts a minimum-maximum normalization method to enable the value range of the extracted characteristic to fall between [0,1], and the normalization method can improve the convergence rate of the model and improve the prediction precision of the model, as shown in figure 4. The min-max normalization method formula is:

Step 4.8, in two partial enlarged diagrams in fig. 7 (a), it can be seen that the equal temperature rise time series curves of the lithium battery in 34-41 times and 58-59 times are closer to the battery capacity series curves than the equal pressure drop time series curves, so that the equal pressure drop time series are adopted in the charge-discharge cycle times in 1-33 times interval, 42-57 times interval and 60-168 times interval, the isothermal temperature rise time series are adopted in the charge-discharge cycle times in 34-41 times interval and 58-59 times interval, and the health factor which can characterize the battery capacity in the whole cycle period is effectively extracted through the fusion of the two time series.

And 4.9, carrying out visual processing on the health factor, the battery capacity sequence, the isobaric descending time sequence and the isothermal ascending time sequence, wherein the constructed health factor is closer to the battery capacity sequence curve than the isobaric descending time sequence curve and the isothermal ascending time sequence curve in the whole charge and discharge cycle period of the lithium ion battery through two partial enlarged graphs of the graph (b) of fig. 7, so that the constructed health factor can represent the degradation condition of the battery capacity more than the original index.

And 5, evaluating the comprehensive effect of the health factors fused with the equivalent pressure drop time sequence and the isothermal rise time sequence by adopting root mean square error (Root Mean Squared Error, RMSE), average absolute error (Mean Absolute Error, MAE) and R square (R-squared) to obtain the battery capacity sequence. The method comprises the following specific steps:

And 5.1, calculating the root mean square error. The root mean square error is the deviation between the judgment predicted value and the true value, and can be used as the accuracy assessment of the comprehensive effect.

Wherein y is _i Is a capacity sequence value; x is x _i Is a health factor sequence value; n is the number of cycles. Analysis of health factors: rmse=0.0021, calculated values of the isobaric falling time sequence and the isothermal rising time sequence are 0.0063 and 0.0454 respectively, the error value of the health factor and the battery capacity is lower, and the accuracy is higher.

And 5.2, calculating an average absolute error. The average absolute error is the average of absolute values of the predicted value and the true value deviation, and can better reflect the actual condition of the error.

Analysis of health factor in formula: mae=0.0017, calculated values of the isobaric falling time sequence and the isothermal rising time sequence are respectively 0.0059 and 0.0362, the error value of the health factor and the battery capacity is lower, and the accuracy is higher.

And 5.3, calculating R square. The R square is used for measuring the fitting degree of the predicted value and the true value, and the predicted value and the true value have strong positive correlation when the correlation is closer to 1.

Where y is the average of the capacity sequences. Analysis of health factors: r is R ² =0.9998, when the pressure drops are equal The calculated values of the inter-sequence and the isothermal rise time sequence are respectively 0.9980 and 0.8949, the error value of the health factor and the battery capacity is lower, and the accuracy is higher. The closer the RMSE and MAE are to 0, the better the fit of the health factor sequence curve to the degradation curve of battery capacity, R ² The closer to 1, the more strongly the health factor sequence curve is correlated with the degradation curve of battery capacity. The analysis result shows that compared with the index before the data preprocessing, the error value of the health factor and the battery capacity is lower, the accuracy is higher, and the data preprocessing work can be completed by adopting the method.

And 6, preprocessing the charge-discharge cycle data of the lithium ion battery to extract characteristic data of 6 dimensions of charge-discharge battery voltage characteristics, charge-discharge battery current characteristics and charge-discharge battery temperature characteristics and constructed health factors, and respectively putting the extracted and constructed data into a file according to the classification of the batteries B0005, B0006, B0007 and B0018. Therefore, 7 dimensions of characteristic data are obtained through data preprocessing, the data preprocessing is completed, and the obtained characteristic data are used in a machine learning model for predicting the residual service life of the lithium ion battery.

Step 7, research on a residual service life prediction method of a lithium ion battery based on a convolutional neural network (Convolutional Neural Networks, CNN), as shown in fig. 8 and 9, specifically comprises the following steps:

and 7.1, establishing a CNN prediction model. The method comprises the steps of dividing the whole CNN model into three parts, wherein the first part is a data input layer, uniformly normalizing characteristic data obtained by data preprocessing, eliminating the influence of dimension, and dividing the input data into a training set and a testing set. The second part is a hidden layer, and the CNN model structure is in this part. The third part is an output prediction part for predicting the test data.

And 7.2, predicting the residual service life of the lithium ion battery based on the CNN. The test uses MATLAB to construct a lithium ion battery residual service life model and predict the residual service life, and the version of MATLAB software is 2021a. Programming by MATLAB software and running on an experimental platform, wherein the basic information of the used experimental platform is as follows: the 64-bit operating system is based on a processor (system type) of x64, a 2.90GHz processor, 8.00GB of memory and 512GB of used operating system versions are Windows10 family Chinese version and solid state disk. Battery data of B0006, B0007 and B0018 are used as training sets, battery data of B0005 is used as a test set, and battery capacity is selected as output. The CNN prediction model constructed by the invention has the root mean square error of 3.57% and the average absolute error of 3.43% by evaluating the CNN prediction model through the root mean square error and the average absolute error.

And 7.3, verifying the validity of the CNN. The invention adopts the same data, respectively takes the test data of the 30 th, 60 th and 90 th test sets to substitute the CNN prediction model and the BP neural network prediction model, and compares the prediction results. The result shows that the error of the CNN prediction model for predicting the test data after the 30 th time, the 60 th time and the 90 th time of the test set is less than 5%, and the prediction effect of the CNN prediction model is better than that of the BP neural network prediction model. It can be seen that the convolutional neural network is effective in predicting the remaining service life of the lithium ion battery.

Step 8, predicting performance research of residual service life of the lithium ion battery based on a bi-directional gating cycle unit (BidirectionalGatedRecurrentUnit, biGRU), as shown in fig. 10 to 13, specifically comprises the following steps:

and 8.1, establishing a BiGRU prediction model. The biglu neural network is composed of two unidirectional and opposite-direction GRUs, and in each time sequence, the feature input is simultaneously provided to the two opposite-direction GRUs, and the output state is commonly influenced by the two GRUs.

And 8.2, testing the first group of super parameters. The super-parameters of the BiGRU model are set to be 100 for maximum iteration times, the batch size is 10, the initial learning rate is 0.0001, the discarding rate is 0.2, and the learning rate is reduced by 10% after 50 times of training. And writing a corresponding program to obtain a predicted residual service life result of the lithium ion battery. The root mean square error RMSE of the biglu neural network model is 3.70% and the mean absolute error MAE is 3.11%.

And 8.3, testing the second group of super parameters. The super-parameters of the BiGRU model are set to be 160 for maximum iteration times, the batch size is 30, the initial learning rate is 0.001, the discarding rate is 0.3, and the learning rate is reduced by 10% after 130 times of training. And writing a corresponding program to obtain a predicted residual service life result of the lithium ion battery. The root mean square error RMSE of the biglu neural network model was 2.60% and the mean absolute error MAE was 2.52%.

And 8.4, testing a third group of super parameters. The super-parameters of the BiGRU model are set to be 160 for maximum iteration times, the batch size is 15, the initial learning rate is 0.001, the discarding rate is 0.3, and the learning rate is reduced by 10% after 80 times of training. And writing a corresponding program to obtain a predicted residual service life result of the lithium ion battery. The root mean square error RMSE of the biglu neural network model was 2.46% and the mean absolute error MAE was 2.33%. The result shows that the overall error of the third group of super-parameter settings is smaller, and the BiGRU super-parameter settings are finally determined.

Step 9, researching a residual service life prediction method of the lithium ion battery based on CNN-BiGRU fusion, as shown in fig. 14-16, specifically comprising the following steps:

and 9.1, constructing a CNN-BiGRU fusion prediction model. The CNN-BiGRU neural network structure mainly comprises four parts, namely a characteristic data input layer, a CNN layer, a BiGRU layer and a characteristic data output layer.

And 9.2, verifying the limitation of the super-parameter setting of the model. The super-parameters of the CNN-BiGRU model are set to be 160 for maximum iteration times, 15 for batch size, 0.001 for initial learning rate and 0.3 for discarding rate, the learning rate is reduced by 10% after 80 times of training, and the gradient descent algorithm adopts self-adaptive moment estimation. And writing a corresponding program to obtain a predicted residual service life result of the lithium ion battery. The root mean square error of the CNN-BiGRU prediction model constructed by the invention is 1.56%, and the average absolute error is 1.31%. Compared with a CNN model and a BiGRU model, the error of the CNN-BiGRU model is minimum, and the prediction accuracy is highest.

And 9.3, verifying the effectiveness and superiority of the CNN-BiGRU. In order to further evaluate the accuracy of the CNN-BiGRU prediction model constructed by the invention on the residual service life prediction result of the lithium ion battery, test data after 30 th and 60 th times of a test set are respectively substituted into the CNN-BiGRU prediction model, the CNN prediction model and the BiGRU prediction model, and the prediction results are compared. The error results of the CNN-BiGRU prediction model, the CNN prediction model and the BiGRU prediction model after 30 times are respectively as follows: 1.77%, 4.13% and 2.18%. The error results of the CNN-BiGRU prediction model, the CNN prediction model and the BiGRU prediction model after the 60 th time are respectively as follows: 1.85%, 4.23% and 2.12%. From the multiple predictions, it can be seen that the CNN-BiGRU prediction model has the lowest root mean square error compared to the other comparison models. Therefore, the CNN-BiGRU method constructed by the invention can be judged to improve the accuracy of predicting the residual service life of the lithium ion battery.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and not for limiting the same; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some or all of the technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit of the invention.

Claims

1. The method for predicting the residual service life of the lithium ion battery based on algorithm fusion is characterized by comprising the following steps of:

step 1, data selection and extraction: extracting effective information of voltage, current and temperature in all charge and discharge cycles of the lithium ion batteries B0005, B0006, B0007 and B0018 by adopting a lithium ion battery detection data set disclosed by NASA;

step 2, data analysis: selecting data of the lithium ion battery in a charging and discharging cycle process for at least 20 times, and performing visual analysis on the selected data; analyzing the change conditions of the voltage, the current and the temperature of the lithium ion battery in the three complete charge-discharge cycle processes; the curves of the parameters relative to time are found by analysis to have more obvious points;

Step 3, extracting features: extracting features by data analysis and a Pearson correlation coefficient method; extracting characteristic point data representing each period of the battery, and requiring the data point to have high reliability and accuracy along with the aging process of the battery; extracting the characteristic data and the corresponding moments thereof to represent the health state of the lithium ion battery, and verifying the correlation between the extracted characteristics and the battery capacity;

step 4, constructing health factors: reconstructing an isobaric discharging time sequence and an isothermal elevating discharging time sequence according to correlation analysis aiming at discharge voltage and discharge temperature indexes of the lithium ion battery to represent parameter characteristics of degradation of the lithium ion battery, and carrying out normalization treatment on characteristic parameters; fitting the normalized isobaric descending time sequence and isothermal ascending time sequence with a battery capacity sequence, analyzing the fitting degree, re-scrambling and sequencing the isobaric descending time sequence and the isothermal ascending time sequence according to an analysis result, and performing feature simulation and feature fusion to construct a health factor based on the fusion of the isobaric descending time sequence and the isothermal ascending time sequence of a discharge battery;

step 5, evaluating the comprehensive effect of health factors fused by the equivalent pressure drop time sequence and the isothermal rise time sequence by adopting root mean square error, average absolute error and R square (R-squared) and the battery capacity sequence;

Step 6, preprocessing data: integrating the constructed health factors with the extracted characteristic data;

classifying the extracted characteristic data of 6 dimensions of the voltage characteristic, the current characteristic and the temperature characteristic of the charge and discharge battery according to batteries B0005, B0006, B0007 and B0018, and then forming a data set to be processed;

step 7, constructing a CNN-BiGRU fusion model, and processing data:

dividing the data set to be processed generated in the sixth step into a training set and a testing set, selecting battery capacity as target output, constructing a CNN-BiGRU fusion model, and setting model super-parameters; substituting test data of the test set into a CNN-BiGRU prediction model to obtain a residual service life prediction curve of the lithium ion battery;

the CNN-BiGRU fusion model comprises a characteristic data input layer, a CNN layer, a BiGRU layer and a characteristic data output layer;

the characteristic data input layer is used for analyzing the data, determining the dimension and output of the characteristic variable, dividing the training set and the testing set, normalizing the data, and inputting the characteristic data into the neural network model;

the data processed by the characteristic data input layer firstly enter a CNN layer to carry out two-round convolution calculation, then enter a full-connection layer to carry out time sequence expansion, and then the characteristic data are transmitted to a BiGRU layer;

The BiGRU layer further extracts local features of the feature data, learns and trains the extracted feature data, predicts the residual service life of the battery, and reversely normalizes the predicted result and outputs the normalized result;

the CNN-BiGRU fusion model adopts two-layer convolution and bidirectional GRU for calculation; the convolution kernel size of the first layer of convolution setting is 3 multiplied by 1, the number of filters is 64, and each convolution kernel can perform dot product operation when the convolution operation is performed, so that characteristic data is learned, a filling mode is added at the same time, zero padding operation is performed on input characteristic data, a relu activation function is used after the convolution operation and is applied to the characteristic data obtained by the convolution operation, nonlinear processing is performed on the characteristic data, and the convergence of a neural network is enhanced; the convolution kernel of the second layer of convolution is 3×1, and the number of filters is 32; the data obtained by the first layer convolution is used as the input data of the second layer convolution, the first layer convolution operation is repeated, and the Padding processing and the relu activation function are carried out; the BiGRU neural network collects data information in the forward direction and the reverse direction of the time sequence, has the capability of the hidden layers in two directions for memorizing and removing the information, increases the flexibility of a model by the structure, and reduces the accidental of characteristic data of the same time sequence; two standard GRU neural networks forming a BiGRU neural network, wherein one GRU neural network adopts data of a historical time sequence for forward training, and the other GRU neural network adopts data of a future time sequence for reverse training; and obtaining the final output of the BiGRU neural network under a certain time sequence by fusing the output characteristics obtained by forward training and the output characteristics obtained by reverse training.

2. The method for predicting the remaining service life of a lithium ion battery based on algorithm fusion as claimed in claim 1, wherein in the fifth step, the specific operation is as follows:

a capacity root mean square error assessment mechanism for judging the deviation between the predicted value and the true value is adopted to be used for evaluating the accuracy of the comprehensive effect of the health factor; the root mean square error is expressed as:

wherein y is _i Is a capacity sequence value; x is x _i Is a health factor sequence value; n is the number of cycles;

the average absolute error of the deviation of the predicted value and the true value is judged, so that the actual condition of the error can be better reflected;

the R square measuring the fitting degree of the predicted value and the true value is adopted, and the predicted value and the true value have strong positive correlation when the correlation is closer to 1;

in the middle ofIs the average of the capacity sequences.

3. The method for predicting the remaining service life of the lithium ion battery based on algorithm fusion according to claim 1 or 2, wherein in the step 1, the specific operation is as follows:

in the charging process, the batteries are charged in a constant current mode, and a constant voltage charging mode is maintained when the voltage reaches the maximum value until the current is reduced to 20mA, and the charging is finished; in the discharging process, the battery is discharged in a constant current manner until the battery is reduced to a certain value; and respectively extracting the terminal voltage, the charging output current, the charging temperature, the charging measurement voltage, the charging measurement current, the discharging battery terminal voltage, the discharging output current, the discharging temperature, the discharging battery load voltage and the discharging load current of the rechargeable battery in the constant-current constant-voltage charging stage, the constant-current discharging stage and the battery temperature change as battery performance indexes.

4. The method for predicting the remaining service life of the lithium ion battery based on algorithm fusion according to claim 3, wherein in the step 1, the voltage, current and temperature change conditions of the lithium ion battery in the 30 th, 60 th and 120 th charge-discharge cycle processes are extracted, and the change trend is indicated by visual processing.

5. The method for predicting the remaining service life of the lithium ion battery based on algorithm fusion according to claim 1, 2 or 4, wherein in the step 3, a formula method and a pearson correlation coefficient method are adopted to extract the characteristics; the pearson correlation coefficient is an evaluation index of the degree of correlation between different variables, the value of the pearson correlation coefficient is in the interval of-1 to 1, and the pearson correlation coefficient is a linear correlation analysis method; the formula of the pearson correlation coefficient method is:

wherein Y represents battery capacity degradation data, and X represents relevant data of each extracted feature.

6. The method for predicting the residual service life of the lithium ion battery based on algorithm fusion as claimed in claim 3, wherein in the step 3, a formula method and a pearson correlation coefficient method are adopted to extract the characteristics; the pearson correlation coefficient is an evaluation index of the degree of correlation between different variables, the value of the pearson correlation coefficient is in the interval of-1 to 1, and the pearson correlation coefficient is a linear correlation analysis method; the formula of the pearson correlation coefficient method is:

7. The method for predicting the remaining service life of a lithium ion battery based on algorithm fusion according to claim 1, 2, 4 or 6, wherein in the step 4, the specific steps are as follows:

wherein y (k) is a battery capacity sequence, k is a sequence length, x _i (k) For the isobaric discharge time sequence or the isothermal rise time sequence, i represents the number of sequences, ρ represents the resolution coefficient, and the association degree solving relational expression is as follows:

wherein r is _i For the degree of association, 0 < r _i < 1; the more the correlation value tends to 1, the greater the correlation between the constructed health factor and the battery capacity; the correlation between each characteristic and the battery capacity can be obtained by simply extracting and calculating the data;

step 4.2, normalization: because the battery capacity sequence, the isobaric descending time sequence and the isothermal ascending time sequence are different in size but consistent in comparison standard, the battery capacity sequence, the isobaric descending time sequence and the isothermal ascending time sequence are normalized, and the data normalization is to scale the extracted data in a smaller area so as to eliminate the influence caused by the different sizes among the data; the invention adopts a minimum-maximum normalization method to enable the value range of the extracted characteristic to fall between [0,1], the normalization method can improve the convergence rate of the model and the precision of model prediction, and the formula of the minimum-maximum normalization method is as follows:

In which x is _min To the minimum value of the extracted features, x _max Is the maximum of the extracted features;

8. The method for predicting the remaining service life of a lithium ion battery based on algorithm fusion according to claim 3, wherein in the step 4, the specific steps are as follows:

wherein r is _i For the degree of association, 0 < r _i < 1; correlation valueThe more towards 1, the greater the correlation of the constructed health factor with battery capacity; the correlation between each characteristic and the battery capacity can be obtained by simply extracting and calculating the data;

9. The method for predicting the remaining service life of a lithium ion battery based on algorithm fusion according to claim 5, wherein in the step 4, the specific steps are as follows:

step 4.2, normalization: because the battery capacity sequence, the isobaric descending time sequence and the isothermal ascending time sequence are different in size but consistent in comparison standard, the battery capacity sequence, the isobaric descending time sequence and the isothermal ascending time sequence are normalized, and the data normalization is to scale the extracted data in a smaller area so as to eliminate the influence caused by the different sizes among the data; the invention adopts a minimum-maximum normalization method to enable the value range of the extracted characteristic to fall between [0,1], and the normalization method can improve the convergence rate of the model and improve the prediction precision of the model, as shown in figure 4; the min-max normalization method formula is: