Battery life prediction method based on importance entropy feature selection

Abstract

The invention relates to the field of battery life prediction, and discloses a battery life prediction method based on importance entropy feature selection, which comprises the following steps: constructing a battery life prediction model; the method comprises the following steps: collecting data between battery capacity attenuation and battery cycle life, preprocessing the data, and constructing an initial entropy feature set Q1 by using a multi-scale permutation entropy as an entropy feature; extracting an initial entropy feature set Q1, and selecting an optimal scale factor s and an embedding dimension m by analyzing the change trend of an MPE mean value under different scales; measuring the importance of the entropy features by adopting a Gini coefficient method, and sorting the entropy features through contribution values among the entropy features; and selecting entropy features which contribute to a prediction result to form an importance entropy feature set Q2 as the input of a battery life prediction model, and predicting the battery life. According to the method, the new entropy feature set is selected and constructed, and the importance entropy feature set Q2 is used as model input, so that the accuracy of the battery life prediction model is improved.

Description

Technical Field

[0001] The present invention relates to a battery life prediction method, and in particular to a battery life prediction method based on importance entropy value feature selection. Background Art

[0002] Lithium-ion batteries have the advantages of high energy storage density and long service life. As a complex electrochemical system, the capacity of lithium-ion batteries will gradually degrade during use. Therefore, the remaining useful life (RUL) prediction of lithium batteries, as a cutting-edge technology for lithium-ion battery fault diagnosis and health management, has become one of the important research directions in the field of lithium-ion batteries, and has gradually become a research hotspot in the health management and fault diagnosis of electronic systems. Timely prediction of battery service life helps to plan and manage batteries more reasonably, thereby reducing accident risks and ensuring the safe operation of equipment. RUL prediction methods can be roughly divided into two methods: model-driven and data-driven. Since the data-driven method is based on the historical data of battery operation, it performs regression prediction by mining the sequence features in the historical data. It does not require an accurate electrochemical model of the internal system of the battery, so it does not require specific battery prior knowledge and is easy to operate. Therefore, it has become the mainstream research method for the performance degradation analysis and RUL prediction of lithium battery nonlinear systems.

[0003] When modeling the battery from the perspective of physical quantities such as electrical impedance, uncertain factors such as impedance parameters, operating environment and operating conditions in actual applications make physical modeling difficult, and its nonlinear model physical model is difficult to accurately establish. When the traditional data-driven method predicts RUL, all the collected preliminary data is directly used for modeling. There is too much redundant information, which leads to problems such as poor generalization ability of the model when making actual predictions. Summary of the invention

[0004] In view of the shortcomings of the above problems, the present invention provides a battery life prediction method based on importance entropy value feature selection to solve the problem of low accuracy of battery life prediction in the prior art.

[0005] To solve the above problems, the present invention adopts the following solution: a battery life prediction method based on importance entropy feature selection, characterized by comprising: Step 1: Build a battery life prediction model; Step 2: Collect data on battery capacity attenuation and battery cycle life, pre-process the collected data on battery capacity attenuation and battery cycle life, and use multi-scale permutation entropy as entropy feature to construct the initial entropy feature set Q 1 ; Step 3: Extract the initial entropy feature set Q1 By analyzing the changing trend of the MPE mean at different scales, the optimal scale factor is selected s With embedding dimension m ; Step 4: Use the Gini coefficient method to measure the importance of entropy features, and sort the entropy features according to their contribution values; Step 5: Select the entropy features that contribute most to the prediction results to form the important entropy feature set Q 2 , as the input of the battery life prediction model, the battery life is predicted and analyzed.

[0006] Furthermore, the preprocessing of the collected data on battery capacity attenuation and battery cycle life includes: Eliminate unqualified battery data and data anomalies caused by test circuits and sensors in the battery data.

[0007] Furthermore, step three includes coarsening the original battery life sequence composed of the collected battery life cycle data at different scales, constructing life sequences at multiple scales, and then calculating the permutation entropy of the coarse-grained sequences at different scales; specifically, for the original battery life sequence of length L To coarse-grained: In the formula, s and y s,j are the scale factor and multi-scale life sequence, respectively, and j is the granulation step length; For s,j Time reconstruction yields:

[0008] Where m represents the embedding dimension, represents the delay factor, t represents the t-th reconstruction component, and arranging (2) in ascending order, we can obtain: Reconstruct the elements

[0009] There are m! permutations in total. Assuming that each permutation occurs D times, the probability of the life sequence at each scale is , n is the number of features; Combining the above formulas, the MPE value is:

[0010] Among them, P s,l =1 / m!, H s,p Take the maximum value ln(m!) and set the MPE value H s,p Normalized to get h s,p =Hs,p / ln(m!).

[0011] Furthermore, step four includes: The Gini coefficient is used to measure the importance of entropy features, and the calculation formula is:

[0012] In the formula, n represents the number of features, G(F) represents the Gini coefficient value of the feature set F, and p i Indicates the proportion of the i-th feature point in the total entropy value feature; With feature f i As the variable feature of the branch, the metric of this feature in the branch is:

[0013] Where K is the total number of branches, Indicates the number of contained elements; Feature i The importance is: Combining equations (7) and (8), the weight of the i-th feature is normalized as follows: In the formula, n represents the number of features. Through the above formula (9), the importance of the entropy value feature is intuitively output, and all the entropy value features are sorted and output according to the important features.

[0014] Furthermore, step five includes: Filter out the top 10 entropy features in terms of importance to form the important entropy feature set Q 2 , and the importance entropy feature set Q 2 As the input of the battery life prediction model, the battery life prediction analysis is performed; According to the battery life prediction results, the entropy feature set Q is compared and analyzed. 1 And the selected importance entropy feature set Q 2 , the accuracy of battery life prediction.

[0015] Furthermore, the method also includes evaluating the battery life prediction model using mean square error, and the specific calculation formula is: in, and i It is i The predicted life and actual life of a battery tuple, n represents the number of features.

[0016] Compared with the prior art, the present invention uses the Gini coefficient method to measure the importance of entropy features, sorts entropy features by comparing the contribution values ​​between features, selects and constructs a new entropy feature set, and uses the importance entropy feature set Q 2 , as the input spatial sequence of the model, it reduces redundancy and improves the effectiveness and accuracy of the battery life prediction model. It has important economic and application value. It also uses actual operating battery attenuation data to build a case library, which provides powerful guidance for operators to find the accuracy of battery life prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of a battery life prediction method based on importance entropy value feature selection of the present invention; Figure 2 yes Figure 1 When calculating the entropy feature in the method shown, the relationship between the embedding dimension and the MPE mean; Figure 3 yes Figure 1 The original entropy feature set Q in the method shown 1 And the selected importance entropy feature set Q 2 ,Battery life prediction comparison experiment statistics chart. DETAILED DESCRIPTION

[0018] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, the present invention provides a battery life prediction method based on importance entropy feature selection, and the specific steps are as follows: Step 1: Build a battery life prediction model; Step 2: Collect data on battery capacity attenuation and battery cycle life, pre-process the collected data on battery capacity attenuation and battery cycle life, and use multi-scale permutation entropy as entropy feature to construct the initial entropy feature set Q 1 ; Specifically, first, the lithium battery is charged and discharged in a cycle, and data between battery capacity decay and battery cycle life are collected during the process, and the preliminary collection of data between battery capacity decay data and battery cycle life is completed. After that, the collected data between battery capacity decay and battery cycle life are preprocessed, including removing unqualified battery data (cycle times <500) and abnormal data caused by test circuits, sensors, etc., and removing or optimizing them. Then, based on the preprocessed data, multi-scale permutation entropy is used as the entropy feature to construct an initial entropy feature set Q 1 .

[0020] The battery pack is managed by the A123 system (APR18650M1A) and the temperature is set at 30°C. The rated capacity of each battery is 1.1Ah and the rated voltage is 3.3V. When the battery capacity decays to 80% of the rated capacity, it is considered as the cycle life of the battery.

[0021] Step 3: Extract the initial entropy feature set Q 1 By analyzing the changing trend of the MPE mean at different scales, the optimal scale factor is selected s With embedding dimension m ; Specifically, Figure 2 As shown, the horizontal axis is the embedding dimension m The vertical axis is the MPE value. The original battery life sequence collected through the experiment is coarse-grained at different scales, and life sequences of multiple scales are constructed. Then, the permutation entropy of the coarse-grained sequences at different scales is calculated. Specifically, the original battery life sequence of length L is To coarse-grained:

[0022] In the formula, s and y s,j are the scale factor and multi-scale life sequence, respectively, and j is the granulation step length; For s,j Time reconstruction yields:

[0023] Where m represents the embedding dimension, represents the delay factor, t represents the t-th reconstruction component, and arranging (2) in ascending order, we can obtain: Reconstruct the elements There are m! permutations in total. Assuming that each permutation occurs D times, the probability of the life sequence at each scale is , n is the number of features; The MPE value in the above steps is calculated as:

[0024] When P s,l =1 / m!, the above formula takes the maximum value ln(m!), and the MPE value H s,p Normalized to get h s,p =H s,p / ln(m!).h s,p The larger it is and the closer it is to 1, the more non-stationary the battery life sequence is, and the farther it is from 1, the more regular the life sequence is.

[0025] When calculating MPE, the mean change trend under different embedding dimensions is used to select the most appropriate optimal scale factor. s With embedding dimension m When the scale factor is large, the extracted features will have information confusion, affecting the prediction effect of the battery life prediction model. When the scale factor is small, the extracted information is incomplete. Therefore, the present invention uses different embedding dimensions to m In this case, different scaling factors with MPE < 0.2 are selected.

[0026] Step 4: Use the Gini coefficient method to measure the importance of entropy features, and sort the entropy features according to their contribution values; Among them, the Gini coefficient is used to calculate the importance of entropy features. It takes the average value of the contribution of all features, determines the most important feature by the contribution value between features, and completes the sorting and screening of entropy features. The calculation formula is:

[0027] In the formula, n represents the number of features, G(F) represents the Gini coefficient value of the feature set F, and p i Indicates the proportion of the i-th feature point in the total entropy value feature, such as dimension m= 3, the entropy feature set contains 3 sub-features with different scale factors, and the feature accounts for p i =3 / 20=0.15.

[0028] With feature f i As the variable feature of the branch, the metric of this feature in the branch is:

[0029] Where K is the total number of branches, Indicates the number of contained elements. The larger the metric is, the more unbalanced the branch under this attribute is.

[0030] Feature i The importance is:

[0031] Combining equations (7) and (8), the weight of the i-th feature is normalized as follows: In the formula, n represents the number of features. Through the above formula (9), the weight of each input variable can be intuitively output as the proportion of each weight, that is, the importance of the entropy feature, so that all entropy features can be output as important features.

[0032] Step 5: Select the entropy features that contribute most to the prediction results as important features to form the important entropy feature set Q 2 Perform predictive analysis on battery life.

[0033] Specifically, the top 10 entropy features with the highest importance threshold are selected as the importance features to form the importance entropy feature set Q 2 As the input of the battery life prediction model, the battery life is predicted and analyzed, and the mean square error (MSE) is used to evaluate the battery life prediction model. The specific calculation formula is: in, and i It is i The predicted life and actual life of each battery tuple are calculated, and other battery attenuation fragments are passed through the same model in turn to output the results of battery life prediction.

[0034] According to the battery life prediction results, the entropy feature set Q is compared and analyzed. 1 And the selected importance entropy feature set Q 2 , the accuracy of battery life prediction.

[0035] Specifically, the feature set Q after entropy feature extraction 1 , too high a dimension will increase the training cost and even cause overfitting, so it is necessary to 1 Perform dimensionality reduction operations, use the Gini coefficient to calculate feature importance, and sort the features by contribution to form the importance entropy value feature set Q 2 , then, the feature set Q 1 And the importance entropy feature set Q 2 They are respectively used as the input of the battery life prediction model to compare and analyze the entropy feature set Q 1 And the selected importance entropy feature set Q 2 , the accuracy of battery life prediction.

[0036] like Figure 3As shown in the figure, the case library is constructed using the actual battery attenuation data, and the importance of entropy features is sorted and reselected using the Gini coefficient evaluation. 2 As the input of the battery life prediction model, a prediction experiment is carried out. Through comparative analysis, through the prediction comparison results corresponding to the four experimental objects in the case library, it can be clearly seen that the feature set selected by the importance entropy value feature has a better correlation with the data set. 2 The evaluation index (MSE) of the model is smaller than that of the entropy feature set Q 1 data.

[0037] The present invention collects battery data during the battery charging and discharging process, pre-processes it, and then extracts the entropy feature set Q 1 By analyzing the changing trend of the MPE mean at different scales, the optimal scale factor is selected s With embedding dimension m ; The Gini coefficient method is used to measure the importance of entropy features, and the entropy features are ranked by comparing the contribution values ​​between the features; finally, the features that contribute most to the prediction results are obtained through importance screening, and the important features are used to form the important entropy feature set Q 2 , reducing redundancy and improving the accuracy of the battery life prediction model, which has important economic and application value. It also uses actual operating battery attenuation data to build a case library, which provides powerful guidance for operators to find the accuracy of battery life prediction.

[0038] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A battery life prediction method based on importance entropy feature selection, characterized in that: include: Step 1: Build a battery life prediction model; Step 2: Collect data on battery capacity attenuation and battery cycle life, pre-process the collected data on battery capacity attenuation and battery cycle life, and use multi-scale permutation entropy as entropy feature to construct an initial entropy feature set Q1; Step 3: Extract the initial entropy feature set Q1, and select the optimal scale factor by analyzing the changing trend of the MPE mean at different scales s With embedding dimension m ; Step 4: Use the Gini coefficient method to measure the importance of entropy features, and sort the entropy features according to their contribution values; Step 5: Select the entropy features that contribute most to the prediction results to form the important entropy feature set Q2, which is used as the input of the battery life prediction model to perform prediction analysis on the battery life.

2. The method for predicting battery life based on importance entropy feature selection according to claim 1, characterized in that: The preprocessing of the collected data on battery capacity attenuation and battery cycle life includes: Eliminate unqualified battery data and abnormal data caused by test circuits and sensors from the battery data.

3. The method for predicting battery life based on importance entropy feature selection according to claim 1, characterized in that: The step three includes coarsening the original battery life sequence composed of the collected battery life cycle data at different scales, constructing life sequences at multiple scales, and calculating the permutation entropy of the coarse-grained sequences at different scales; specifically, for the original battery life sequence of length L To coarse-grained: In the formula, s and y s,j are the scale factor and multi-scale life sequence, respectively, and j is the granulation step length; For s,j Time reconstruction yields: Where m represents the embedding dimension, represents the delay factor, t represents the t-th reconstruction component, and arranging the above formula in ascending order, we can get: ； Reconstruct the elements There are m! permutations in total. Assuming that each permutation occurs D times, the probability of the life sequence at each scale is , n is the number of features; Combining the above formulas, we can get the MPE value as follows: Among them, P s,l =1 / m!, H s,p Take the maximum value ln(m!) and set the MPE value H s,p Normalized to get h s,p =H s,p / ln(m!).

4. The method for predicting battery life based on importance entropy feature selection according to claim 1, characterized in that: The fourth step comprises: The Gini coefficient is used to measure the importance of entropy features, and the calculation formula is: In the formula, n represents the number of features, G(F) represents the Gini coefficient value of the feature set F, and p i Indicates the proportion of the i-th feature point in the total entropy value feature; With feature f i As the variable feature of the branch, the metric of this feature in the branch is: Where K is the total number of branches, Indicates the number of contained elements; Feature i The importance is: Combining equations (7) and (8), the weight of the i-th feature is normalized as follows: In the formula, n represents the number of features. Through the above formula (9), the importance of the entropy value feature is intuitively output, and all the entropy value features are sorted and output according to the important features.

5. The method for predicting battery life based on importance entropy feature selection according to claim 1, characterized in that: The step five comprises: The top 10 entropy features in importance are selected to form an importance entropy feature set Q2, and the importance entropy feature set Q2 is used as the input of the battery life prediction model to perform prediction analysis on the battery life; According to the battery life prediction results, the entropy feature set Q1 and the screened importance entropy feature set Q2 are compared and analyzed to determine the accuracy of the battery life prediction.

6. The method for predicting battery life based on importance entropy feature selection according to claim 5, characterized in that: The method also includes evaluating the battery life prediction model using mean square error, and the specific calculation formula is: in, and i It is i The predicted life and actual life of a battery tuple, n represents the number of features.

Images