Lithium battery health state estimation method based on data driving

Abstract

The invention belongs to the technical field of lithium battery state estimation, and particularly relates to a lithium battery health state estimation method based on data driving, which comprises the following steps: firstly, collecting lithium battery operation data; carrying out characteristic quantity extraction and data preprocessing on the collected data; carrying out dimension reduction on the preprocessed data by utilizing a neighbor component analysis method, and dividing the data subjected to dimension reduction into a training set and a test set; according to the method, chaotic mapping, adaptive weight, a reverse learning strategy and a Cauchy variation disturbance strategy are introduced to improve a sparrow search algorithm to obtain an improved sparrow search algorithm, and the improved sparrow search algorithm ISSA is utilized to optimize parameters of the DESN, so that an ISSA-DESN lithium battery state estimation model is obtained, and lithium battery health estimation is realized. According to the method, data dimension reduction is performed through the neighbor component analysis method, and the depth echo state network parameters are optimized through the improved sparrow search algorithm, so that the accuracy of the model and the prediction precision of lithium battery health can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of lithium battery state estimation, and specifically relates to a data-driven method for estimating the state of health of lithium batteries. Background Art

[0002] In the context of the shortage of traditional energy and the accelerating transformation of the global energy structure, lithium batteries have been widely used in electric vehicles and microgrid energy storage power stations, and their economy and safety have gradually become the focus of attention. However, due to the relatively easy occurrence of thermal runaway in lithium batteries, it will affect the safe and stable operation of electric vehicles and microgrid energy storage power stations. Therefore, accurately predicting the state of health of lithium batteries is of crucial significance for ensuring the normal use and safe operation of lithium batteries.

[0003] Currently, the methods for predicting the state of health of lithium batteries include model-based prediction methods and data-driven prediction methods. The model-based prediction methods mainly include the equivalent circuit method and the electrochemical model method. They usually consider factors such as the materials of lithium batteries, impedance changes, and internal materials to build models, and can achieve relatively high prediction accuracy. However, the modeling process of the physical and chemical dynamic changes of lithium batteries is difficult. The data-driven prediction methods usually use neural networks for prediction, which do not require an accurate analysis of the degradation process of lithium batteries, and only require a large amount of lithium battery data to predict the state of health of lithium batteries.

[0004] The state of health (SOH) of lithium batteries is an important parameter for the normal and safe operation of lithium batteries. When using neural networks to predict the state of health of lithium batteries, there are problems such as being easily trapped in local optima, slow convergence speed, and overfitting. Therefore, a method needs to be proposed to improve these technical problems, so as to accurately predict the state of health of lithium batteries. Summary of the Invention

[0005] In order to solve the problems that the prediction model based on neural networks is easily trapped in local optima, slow convergence speed, overfitting, etc. when predicting the state of health of lithium batteries, the present invention proposes a data-driven method for estimating the state of health of lithium batteries, which can accurately predict the state of health of lithium batteries and ensure the safe and stable operation of lithium batteries.

[0006] The object of the present invention can be achieved by the following technical solutions: A data-driven method for estimating the state of health of lithium batteries, comprising the following steps:

[0007] Step S1: Collect lithium battery operation data;

[0008] Step S2: Extract feature quantities from the collected lithium battery operation data;

[0009] Step S3: Perform data preprocessing on the extracted feature quantities;

[0010] Step S4: Use the nearest neighbor component analysis method to screen the preprocessed feature quantities and eliminate redundant data;

[0011] Step S5: Use the screened feature quantities to train the deep state echo network DESN, and at the same time use the improved sparrow search algorithm ISSA to optimize the parameters of the deep state echo network DESN, so as to obtain the ISSA-DESN lithium battery state estimation model;

[0012] Step S6: Input the screened feature quantities of the lithium battery to be predicted into the ISSA-DESN lithium battery state estimation model, and output the health prediction result of the lithium battery to be predicted.

[0013] For the above data-driven lithium battery health state estimation method, the lithium battery operation data collected in step S1 includes voltage and current data during the constant current charging and discharging stages of the battery.

[0014] For the above data-driven lithium battery health state estimation method, the feature quantities extracted in step S2 include 6 dimensions: root mean amplitude , absolute average value , root mean square value , kurtosis index , skewness index and peak index .

[0015] For the above data-driven lithium battery health state estimation method, the data preprocessing method in step S3 includes removing outliers and maximum-minimum normalization operations.

[0016] For the above data-driven lithium battery health state estimation method, the specific steps of using the nearest neighbor component analysis method to screen the preprocessed feature quantities in step S4 include:

[0017] (1) The probability of selecting from the preprocessed feature set as the reference point of is: P[Ref( p m )= p m , n |S]= k[ d w p m , p m , n ] ∑ n =1 N k [ d w p m , p m , n ] , where represents the nth data of the -dimensional feature quantity , is selected from the feature set as the reference point of for , , represents the exponential function with the natural constant \(e\) as the base, is the kernel width; is the distance from to any data in

[0018] (2) Use the leave-one-out method to predict the feature set in the output value of, set the training set not to contain the sample ( , ), represents the \(i\)-th data of the dimensional feature quantity ; represents the label of the \(i\)-th data of the dimensional feature quantity , then is selected as the probability of the reference point is: P in = P[Ref( p m ,i )= p m ,n | S -i ]= k[ d w p m ,i , p m , n ] ∑ n = 1, n ≠ i N k [ d w ( p m, i , p m, n )] , , where represents the \(i\)-th data of the dimensional feature quantity ; represents the \(n\)-th data of the dimensional feature quantity ; is the feature dimension, is the total number of feature dimensions;

[0019] Use the prediction value and the error expectation as the loss function: , where is the expectation operation, and on this basis, introduce a regularization term as the minimization objective function: , where is the vector composed of the feature weights ; is the regularization parameter;

[0020] According to the given kernel width and the regularization parameter , find the appropriate vector to minimize the objective function: , the final solution The size of the element reflects the characteristic weights of each characteristic quantity. Finally, the optimal characteristic quantity is selected according to the characteristic weights; that is, the characteristic quantity corresponding to the optimal characteristic weight is selected from the characteristic quantities of 6 dimensions. The finally dimension-reduced characteristic quantity is , representing the th data of the dimension-reduced characteristic quantity.

[0021] The above method for estimating the state of health of a lithium battery based on data driving, step S5 specifically includes the following steps:

[0022] Step S5-1: Randomly divide the characteristic quantities screened by the neighborhood component analysis method into a test set and a training set;

[0023] Step S5-2: Use the training set data to train the deep state echo network DESN;

[0024] Step S5-3: Use the improved sparrow search algorithm ISSA to optimize the parameters of the deep state echo network DESN, and then obtain the ISSA-DESN lithium battery state estimation model;

[0025] Step S5-4: Use the test set data to test the constructed ISSA-DESN lithium battery state estimation model.

[0026] In the above method for estimating the state of health of a lithium battery based on data driving, in step S5, a chaotic mapping, an adaptive weight, a reverse learning strategy, and a Cauchy mutation perturbation strategy are introduced to improve the sparrow search algorithm, and an improved sparrow search algorithm is obtained.

[0027] Compared with the existing lithium battery state estimation methods, the advantages of the present invention are as follows:

[0028] A method for estimating the state of health of a lithium battery based on data driving is proposed. The high-dimensional characteristic quantity data is converted into low-dimensional characteristic quantity data by the neighborhood component analysis method, redundant data is eliminated, and most of the useful information is obtained from the original lithium battery operation data; a chaotic mapping, an adaptive weight, a reverse learning strategy, and a Cauchy mutation perturbation strategy are introduced to improve the sparrow search algorithm, and an improved sparrow search algorithm is obtained, which can effectively improve the local optimization and global search capabilities of the sparrow search algorithm;

[0029] The parameters of the deep state echo network DESN are optimized by using the improved sparrow search algorithm, which better solves the problems of slow convergence speed and overfitting of the deep state echo network DESN. Finally, the ISSA-DESN lithium battery state estimation model is established, which has higher health prediction accuracy than other ordinary models. Description of the Drawings

[0030] Figure 1This is the flowchart of the data-driven method for estimating the health state of lithium batteries according to the present invention.

[0031] Figure 2 This is the battery parameter diagram.

[0032] Figure 3 This is the flowchart of the improved sparrow search algorithm. Detailed implementation manners

[0033] The present invention will be described in detail below in conjunction with the accompanying drawings and specific implementation cases.

[0034] The present invention is a data-driven method for estimating the health state of lithium batteries. Figure 1 This is the flowchart of this method, including the following steps:

[0035] Step S1: Collect the operation data of the lithium battery, including the voltage and current data during the constant current charge and discharge stages of the battery.

[0036] During the constant current process, the lithium battery is charged at 1.5 A. When the voltage reaches 4.2 V, the charging mode is switched to constant voltage charging. When the current drops to 20 mA, the charging process terminates. The discharge runs at a constant current of 2 A until the lithium battery voltage drops to the minimum cut-off voltage. The lithium battery parameters are as Figure 2 shown.

[0037] Step S2: Extract the characteristic quantities from the collected operation data of the lithium battery.

[0038] The extracted characteristic quantities include 6 dimensions: root mean square amplitude , absolute average value , root mean square value , kurtosis index , skewness index , peak index . The calculation formulas for each characteristic quantity are as follows:

[0039]

[0040]

[0041]

[0042]

[0043]

[0044]

[0045] In the formula, is the collected operation data of the lithium battery, is the total number of the operation data of the lithium battery.

[0046] Step S3: Perform data preprocessing on the extracted feature quantities;

[0047] Since there may be outliers in the original lithium battery operation data, there may also be outliers in the extracted feature quantities. Therefore, outlier removal operations are performed before model establishment.

[0048] Select the average of ten data on each side of the outlier in the feature quantity to replace the outlier. The outlier removal formula is as follows:

[0049]

[0050] In the formula: is the dimensional feature quantity of the th abnormal replacement value; is the feature dimension, = 1, 2, 3, 4, 5, 6, is the first data on the left side of the outlier, is the first data on the right side of the outlier, is the tenth data on the right side of the outlier, is the tenth data on the left side of the outlier.

[0051] Since the collected lithium battery operation data has a large dispersion, it is also necessary to perform maximum-minimum normalization processing on each dimensional feature quantity.

[0052] Step S4: Use the Nearest Neighbor Component Analysis (NCA) to screen the preprocessed feature quantities and eliminate redundant data;

[0053] The present invention considers the information redundancy between feature quantities and uses NCA to reduce the dimension of feature quantities to eliminate redundant data. NCA is a distance metric learning algorithm related to nearest neighbors on the original data set, which can transform high-dimensional data into low-dimensional data and obtain most of the useful information from the original data. The NCA screening steps are as follows:

[0054] Let the preprocessed feature set = {( , )}, where is the dimensional feature quantity of the th data, is the dimensional feature quantity of the th data label. Take any data from the feature set and calculate the distance to its neighboring samples:

[0055]

[0056] In the above formula, represents the dimensional feature quantity of the th data, is the feature dimension, is the total number of feature dimensions, is the feature weight corresponding to the th dimensional feature quantity. Assume that the probability of selecting as the reference point satisfies:

[0057] P[Ref( p m )= p m , n |S] ∝ k[ d w p m , p m, n ]

[0058] In the above formula, is selected from the feature set and is used as the reference point; is the kernel function or similarity function; The smaller is, the larger is. The expression of

[0059]

[0060] In the above formula, represents the exponential function with the natural constant e as the base; is the kernel width, which is used to control the probability of each data being selected;

[0061] The probability of selecting from the feature set and using it as the reference point is:

[0062] P[Ref( p m )= p m , n |S]= k[ d w p m , p m , n ] ∑ n =1 N k [ d w p m , p m , n ]

[0063] In the above formula, the distance is affected by the feature weight and thus indirectly affects the probability of being selected as the reference point;

[0064] To adjust the feature weights in the model , the leave-one-out method is used to predict the feature set in . At this time, the training set does not contain the sample ( , ), then is selected as the probability of the reference point is:

[0065] P in = P[Ref( p m ,i )= p m ,n | S -i ]= k[ d w p m ,i , p m , n ] ∑ n = 1, n ≠ i N k [ d w ( p m, i , p m, n )]

[0066] Taking the expected error of the predicted value as the loss function:

[0067]

[0068] In the above formula, is the expectation operation. On this basis, a regularization term is introduced as the minimization objective function:

[0069]

[0070] In the above formula, is the vector composed of feature weights ; the regularization parameter can make some of the feature weights in become 0. According to the given kernel width and the regularization parameter , find the appropriate vector

[0071] to minimize the objective function:

[0072] The size of the elements of the final solution reflects the feature weights of each feature quantity. The larger , the smaller , indicating that the current feature quantity has a higher correlation with the lithium battery state estimation model, and vice versa. Finally, select the optimal feature quantity according to the feature weights; that is, select the feature quantity corresponding to the optimal feature weight from the feature quantities of 6 dimensions. The finally dimension-reduced feature quantity is , representing the th data of the dimensionality-reduced feature quantity.

[0073] Step S5: Train the Deep State Echo Network (DESN) using the filtered eigenvalue, and at the same time optimize the parameters of the Deep State Echo Network (DESN) using the Improved Sparrow Search Algorithm (ISSA), thereby obtaining the ISSA-DESN lithium battery state estimation model;

[0074] Step S5-1: Randomly divide the feature quantity after NCA screening into a test set and a training set;

[0075] Step S5-2: Use the training set data to train the Deep State Echo Network (DESN). Assume there are input samples and output neurons. At time , the input can be expressed as = , and the output can be expressed as = ;

[0076] Step S5-3: Use the Improved Sparrow Search Algorithm (ISSA) to optimize the parameters of the Deep State Echo Network (DESN), thereby obtaining the ISSA-DESN lithium battery state estimation model;

[0077] The Sparrow Search Algorithm is an intelligent optimization algorithm, which is derived from the foraging and anti-predation behaviors of sparrow populations in nature, and has stronger optimization ability, faster convergence speed and better robustness. The specific steps for optimizing the parameters of the Deep State Echo Network (DESN) using the Improved Sparrow Search Algorithm are as follows:

[0078] Step S5-3-1: Introduce chaotic mapping, adaptive weight, reverse learning strategy and Cauchy mutation perturbation strategy to improve the Sparrow Search Algorithm, obtaining the Improved Sparrow Search Algorithm;

[0079] Step S5-3-2: Use the Improved Sparrow Search Algorithm to optimize the parameters of the Deep State Echo Network (DESN), thereby establishing the ISSA-DESN lithium battery state estimation model;

[0080] Step S5-4: Use the test set data to test the constructed ISSA-DESN lithium battery state estimation model.

[0081] Step S6: Input the filtered feature quantity of the lithium battery to be predicted into the ISSA-DESN lithium battery state estimation model, and output the health prediction result of the lithium battery to be predicted.

[0082] The present invention establishes an efficient lithium battery state estimation model, which can accurately predict the health of lithium batteries, and is of great significance for the safe and reliable operation of lithium batteries.

Claims

1. A data-driven method for estimating the state of health of a lithium battery, characterized in that Including the following steps: Step S1: Collect the operation data of the lithium battery; Step S2: Extract the characteristic quantities from the collected operation data of the lithium battery; Step S3: Perform data preprocessing on the extracted characteristic quantities; Step S4: Use the Nearest Prototype Component Analysis (NPCA) method to screen the preprocessed characteristic quantities and eliminate redundant data; Step S5: Use the screened characteristic quantities to train the Deep State Echo Network (DESN), and at the same time use the Improved Sparrow Search Algorithm (ISSA) to optimize the parameters of the DESN, so as to obtain the ISSA-DESN lithium battery state estimation model; Step S6: Input the screened characteristic quantities of the lithium battery to be predicted into the ISSA-DESN lithium battery state estimation model, and output the health prediction result of the lithium battery to be predicted.

2. The method for estimating the health state of a lithium battery based on data driving according to claim 1, characterized in that, The operation data of the lithium battery collected in Step S1 includes the voltage and current data during the constant current charge and discharge stages of the battery.

3. A data-driven method for estimating the state of health of a lithium battery according to claim 1, characterized in that, The feature quantities extracted in step S2 include six dimensions: root mean square amplitude , absolute average value , root mean square value , kurtosis index , skewness index and peak value index .

4. A data-driven method for estimating the state of health of a lithium battery according to claim 1 or 2 or 3, characterized in that The data preprocessing method in Step S3 includes removing outliers and maximum-minimum normalization operations.

5. A data-driven method for estimating the state of health of a lithium battery according to claim 3, characterized in that, The specific steps of using the Nearest Prototype Component Analysis (NPCA) method to screen the preprocessed characteristic quantities in Step S4 include: (1)Select from the preprocessed feature set and serve as The probability of the reference point is: , where represents the n-th data of the dimensional feature quantity , is selected from the feature set and serves as The reference point of , represents the exponential function with the natural constant e as the base, is the kernel width; is and The distance to any data in (2)Using the leave-one-out method to predict the feature set in the output value, set the training set not to contain the sample( , ), denote the i-th data of the dimensional feature quantity; denote the label of the i-th data of the dimensional feature quantity, then is selected as the probability of the reference point is: , , where is the feature dimension,[[]] is the total number of feature dimensions; using the error expectation and as the loss function: , where is the expectation operation, and on this basis, a regularization term is introduced as the minimization objective function: , where , is the vector composed of feature weights ; is the regularization parameter; According to the given kernel width and the regularization parameter , find a suitable vector to minimize the objective function: , the final solution The magnitude of the elements reflects the characteristic weights of each feature quantity. Finally, select the optimal feature quantity according to the characteristic weights; that is, select the feature quantity corresponding to the optimal characteristic weight from the feature quantities of 6 dimensions. The finally dimension-reduced feature quantity is , representing the th data of the dimension-reduced feature quantity.

6. The method for estimating the state of health of a lithium battery based on data driving according to claim 5, wherein Step S5 specifically includes the following steps: Step S5-1: Randomly divide the characteristic quantities screened by the Nearest Prototype Component Analysis (NPCA) method into a test set and a training set; Step S5-2: Use the training set data to train the Deep State Echo Network (DESN); Step S5-3: Use the Improved Sparrow Search Algorithm (ISSA) to optimize the parameters of the Deep State Echo Network (DESN), so as to obtain the ISSA-DESN lithium battery state estimation model; Step S5-4: Use the test set data to test the constructed ISSA-DESN lithium battery state estimation model.

7. A data-driven lithium battery state of health estimation method according to claim 1 or 6, characterized in that In Step S5, the sparrow search algorithm is improved by introducing a chaotic map, an adaptive weight, a reverse learning strategy, and a Cauchy mutation perturbation strategy to obtain the Improved Sparrow Search Algorithm (ISSA).

Images