Battery health state evaluation method and device based on improved adaptive fuzzy neural network

Abstract

The invention discloses a battery health state evaluation method and device based on an improved adaptive fuzzy neural network, and belongs to the technical field of battery management. The method comprises the following steps: firstly, constructing a lithium battery capacity aging model, extracting a plurality of groups of health indexes from charging and discharging data, and carrying out median-quartile distance normalization processing; then, through a random forest-analytic hierarchy process, key features are screened out by integrating feature accuracy, correlation and extraction complexity; then, hyperparameters such as the number of membership functions and an initial standard deviation are determined through Bayesian optimization, a rule activation mechanism of a dynamic parameter optimization system driven by a working condition is introduced, model parameters are iteratively updated in combination with a recursive least square method and a gradient descent method, and high-precision estimation is achieved; and finally, constructing a full-life-cycle evaluation system covering four stages of normal, attention, abnormity and severity, and dynamically adjusting health indexes. According to the method, the dynamic parameters driven by the working condition characteristics are introduced, so that high-precision SOH estimation under the complex working condition is realized.

Description

Technical Field

[0001] This invention relates to the field of battery management technology, and in particular to a method and apparatus for evaluating battery health status based on an improved adaptive fuzzy neural network. Background Technology

[0002] Accurate estimation of battery State of Health (SOH) is a key technology for ensuring the safe operation of energy storage systems and electric vehicles, and also a guarantee for determining whether backup power can be used normally. Chinese patent CN112765772A discloses an SOH estimation method based on a traditional data-driven model. This method predicts SOH through feature extraction and machine learning algorithms, but it has the following drawbacks:

[0003] Traditional methods struggle to dynamically adapt to complex working conditions, resulting in insufficient estimation accuracy; furthermore, traditional data-driven models lack interpretability when utilizing conventional machine learning.

[0004] Regarding feature selection, only relevance is considered without taking into account the impact of feature importance on SOH estimation, and the efficiency of feature extraction complexity in practical engineering problems is also neglected. Therefore, the feature selection before SOH estimation in the existing technology still needs improvement.

[0005] Existing technologies mostly focus on SOH estimation itself, lacking a phased evaluation system for the health status of batteries throughout their entire life cycle. This results in the inability to dynamically adjust management strategies and limits the long-term tracking capability of batteries in complex scenarios. Summary of the Invention

[0006] Purpose of the invention: The purpose of this invention is to provide a battery health status evaluation method and device based on an improved adaptive fuzzy neural network. By constructing a lithium battery capacity aging model, combining RF-AHP for multi-criteria feature screening, using an improved ANFIS model for data-driven training, and constructing a phased full life cycle SOH evaluation system, the invention aims to improve the accuracy of battery SOH estimation, the interpretability of the model, and the efficiency of engineering applications.

[0007] Technical solution: A battery health status evaluation method based on an improved adaptive fuzzy neural network, comprising the following steps: S1, Based on the battery health indicators extracted from the multi-battery dataset, a lithium battery capacity aging model is constructed; S2, based on the lithium battery capacity aging model, constructs a hierarchical analysis model, takes battery health indicators as features, and uses the random forest-analysis method to perform multi-criteria feature screening to select key features; S3 uses key features as input and battery SOH estimates as output to train the improved dynamic ANFIS model, and saves the best improved dynamic ANFIS model for SOH estimation. S4. Based on the SOH value and degradation rate, a phased full life cycle battery health status evaluation system is constructed.

[0008] Furthermore, the method for extracting battery health indicators is as follows: using a multi-battery dataset, based on the charging and discharging data of lithium batteries, the original charging and discharging cycle structure containing information on time, voltage, current, temperature and capacity is structured into unified charging and discharging data information, and several sets of health indicators are extracted. Each health indicator is then standardized using the median-interquartile range scale.

[0009] The lithium battery capacity aging model uses an exponential decay model to fit the capacity degradation curve and describe the battery aging trend.

[0010] Furthermore, the hierarchical analysis model includes a target layer, a criterion layer, and a scheme layer, wherein the target layer is used to obtain the best features for the SOH estimation effect;

[0011] The criteria layer serves as the selection criteria and includes an accuracy layer, a correlation layer, and a complexity layer. The accuracy layer calculates the contribution of each feature to the SOH estimation using the random forest algorithm. The correlation layer uses the Spearman correlation coefficient to evaluate the correlation between each feature and the SOH. The complexity layer scores the feature extraction complexity.

[0012] The solution layer uses the analytic hierarchy process (AHP) to determine the weights of each criterion, calculates the comprehensive score of the features, and selects the top four features as key features.

[0013] Furthermore, during the training of the improved dynamic ANFIS model, key hyperparameters of the ANFIS model are determined through Bayesian optimization, and dynamic parameters driven by working condition features are introduced. To achieve rule-adaptive activation, the consequent parameters of the fuzzy rules are optimized using the recursive least squares method, and the individual difference adjustment factor is updated using the gradient descent method to characterize the capacity deviation of the same batch of batteries under the same operating conditions.

[0014] Furthermore, let the set of hyperparameters of the ANFIS model be defined. for: , in, Let the initial standard deviation of the membership function be . This is the minimum value of the dynamic parameter. This represents the maximum value of the dynamic parameter. The number of membership functions controls the total number of fuzzy rules. n is the number of key features; Using the root mean square error of the three-fold cross-validation as the objective function, the corresponding hyperparameter combination is obtained iteratively. In the fuzzification layer, the key feature inputs are converted into fuzzy membership degrees using a Gaussian membership function. At the criterion level, the complexity of the operating condition is measured by the instantaneous rate of change of the input features; The emission intensity of the activation rules is standardized in the normalization layer; in the defuzzification layer, each activation rule corresponds to a linear output function, and the SOH estimate is obtained by weighted summation.

[0015] Furthermore, in step S4, the battery life cycle stages are divided according to the estimated SOH value: SOH > 90% is the normal state, 80% < SOH ≤ 90% is the attention state, 70% < SOH ≤ 80% is the abnormal state, and SOH ≤ 70% is the severe state. Among them, the SOH of the battery in the normal state and the attention state decays at a slower rate with the number of cycles, while the decay rate of the abnormal state and the severe state increases sharply.

[0016] Differentiated evaluation indicators are designed for each stage. Under normal conditions, the focus is on whether electrical quantities exhibit extreme abnormal values, and under special conditions, the focus is on the rate of change of electrical quantities under operating conditions. Under abnormal and severe conditions, the focus is on the attenuation rate and the rate of change of internal resistance.

[0017] A battery health status evaluation device, used to perform any of the above-mentioned battery health status evaluation methods, comprising: The data acquisition module is used to match battery types to determine corresponding key features, retrieve the key feature categories to be extracted from the database, and collect key feature data. The battery health state estimation module takes the features obtained by the data acquisition module as input and estimates the battery health state through an improved dynamic ANFIS model. The health assessment and communication module is used to implement the battery health assessment system, output a health score, perform corresponding data storage and alarm operations based on the health score, and back up the data to the cloud.

[0018] Compared with the prior art, the significant advantages of this invention are as follows: 1. This invention adopts the Random Forest (RF)-Analytic Hierarchy Process (AHP) feature selection method to optimize the feature input dimension, comprehensively reduce computational complexity, improve model training efficiency, and significantly improve the feasibility of engineering applications. 2. This invention adopts an improved dynamic ANFIS (Adaptive Network-based Fuzzy Inference System) model. By introducing dynamic parameters driven by working condition features, it achieves higher SOH estimation accuracy under complex working conditions and has stronger adaptability and interpretability compared to the traditional ANFIS model. 3. This invention adopts a phased, full life-cycle SOH evaluation system to comprehensively consider the SOH value, providing dynamic and accurate support for battery health management, and is applicable to various scenarios such as energy storage systems and electric vehicles. Attached Figure Description

[0019] Figure 1 This is a flowchart of the battery health status evaluation method of the present invention; Figure 2 This is a graph showing the capacity decay of the same batch of batteries in an embodiment of the present invention. Figure 3 This is a schematic diagram of the feature selection process of the random forest-analytic hierarchy process in an embodiment of the present invention; Figure 4 The training curves for Bayesian optimization with the root mean square error of three-fold cross-validation as the objective function are shown. Figure 5 To improve the schematic diagram of the adaptive fuzzy neural network inference system; Figure 6 This is an error graph from the training process of the improved ANFIS model in this embodiment of the invention. Figure 7 The above are the predicted RMSE histograms for each battery under the improved ANFIS model in this embodiment of the invention. Figure 8 This is a histogram showing the distribution of prediction errors for each battery under the improved ANFIS model in this embodiment of the invention. Figure 9 This is a comparison chart of the estimated and actual SOH capacity under different operating conditions using the improved ANFIS model in this embodiment of the invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0021] like Figure 1 As shown, this invention provides a battery health status evaluation method based on an improved adaptive fuzzy neural network, comprising the following steps:

[0022] Step 1: Construct a lithium battery capacity aging model and a health indicator extraction mechanism; Common battery health indicators were extracted from a multi-battery dataset, and a model was established to show the relationship between SOH decay and cycle count, providing a data foundation for subsequent feature selection. The specific steps are as follows: Step 11: Based on common battery health indicators (HIs) from multi-battery datasets (such as NASA datasets), collect charge and discharge data of lithium batteries, including measured voltage, measured current, battery impedance and other parameters, and establish a nonlinear relationship model between capacity decay and cycle number.

[0023] In this embodiment, the lithium battery capacity aging model uses an exponential decay model to fit the capacity degradation curve and describe the battery aging trend, such as... Figure 2 As shown in the figure, B0005, B0006, B0007 and B0018 are the numbers of the measured batteries, which are derived from the NASA dataset.

[0024] Step 12: Based on the charge and discharge data, extract several sets of health indicators (HIs), including but not limited to constant current charging time, constant voltage charging time, and constant voltage rise time, to cover multi-dimensional information during the battery aging process.

[0025] In this implementation, 12 health indicators HI1 to HI12 are extracted, including isobaric rise time, constant current charging time, constant voltage charging time, isobaric drop time, charging current area, constant current area, constant voltage current area, IC (capacity increment curve) peak value, IC peak voltage, maximum discharge temperature, constant current time percentage, and maximum charging temperature, covering multi-dimensional information during the battery aging process. A health indicator matrix is ​​constructed from these 12 health indicators. and the health indicator matrix Perform robust normalization: (1) (2) in, Health Indicator Matrix The matrix obtained after normalization, Health Indicator Matrix The sample median; , Health indicator matrix The minimum and maximum values, It is the interquartile range.

[0026] Step 2: Construct a hierarchical analysis model, using battery health indicators as features, and perform multi-criteria feature screening based on RF-AHP to obtain key features; Specifically, the constructed hierarchical analysis model includes an objective layer, a criterion layer, and a final selected solution layer, such as... Figure 3 As shown.

[0027] The target layer is used to obtain the features that best perform on SOH estimation. The criteria layer is used to define the screening criteria, including the accuracy layer, the correlation layer, and the complexity layer. The accuracy layer calculates the contribution of each feature to the SOH estimation using the random forest algorithm. The correlation layer uses the Spearman correlation coefficient to evaluate the correlation between each feature and the SOH. The complexity layer scores the feature extraction complexity.

[0028] For the accuracy layer, RF is used to characterize the importance of features for the lithium battery capacity aging model: (3) In the formula, This represents the number of out-of-bag (OOB) samples. This represents the true value of the k-th OOB sample, i.e., the actual SOH value. This is the model's estimate for the k-th sample, i.e., the SOH estimate; The mean square error of the prediction. The feature matrix obtained by randomly permuting the i-th feature. The corresponding estimated value, Let be the mean square error after the i-th feature permutation.

[0029] The importance Imp(i) of the i-th feature is defined as: (4) in, Let be the mean squared error of the predicted i-th feature.

[0030] For each health metric, the random forest will directly output an importance value. Let i = 1, 2, 3, ..., 12. The importance of all health indicators is organized into a column vector matrix (a 12×1 matrix in this example) called the Imp matrix. Min-max normalization is performed to map to the same scale to obtain feature importance scores. : (5) In the formula, , These are the minimum and maximum values ​​of the Imp matrix, respectively.

[0031] Feature importance score It serves as the feature importance score after normalization of the accuracy layer.

[0032] For the correlation layer, the Spearman correlation coefficient is used to characterize the correlation between each feature and SOH. The absolute value is used to obtain the score of feature i in the relevance layer. : (6) In the formula, n is the sample size of feature i. is the difference between the feature rank and the SOH rank of the k-th sample.

[0033] For the complexity layer, based on practical engineering needs, the original complexity score of the i-th feature is manually defined. And form a 12×1 complexity score matrix. A higher score indicates that the feature is easier to obtain, and thus the complexity layer score is obtained after min-max normalization. for: (7) In the formula, For the first The original complexity score of each feature. , These are the complexity score matrices. The minimum and maximum values.

[0034] A judgment matrix is ​​constructed using the AHP (Analytic Hierarchy Process) method. The judgment matrix is ​​a 3×3 square matrix with elements... Defined as the "importance ratio" between the i-th feature and the j-th feature, the judgment matrix is ​​filled in based on expert experience. =1 / The judgment matrix here can be simplified to a 1×3 weight matrix. This is pre-defined based on specific project requirements. Criterion weights are set; here, a higher weight is adopted to achieve the fastest possible SOH estimation according to actual project requirements. Values ​​were selected, and the top four health indicators were chosen as key features: (8) In the formula, For the first The overall score of each feature , , These represent the weights of each layer of the criteria layer under actual engineering requirements.

[0035] In this implementation, the four health indicators ultimately selected for the two batteries B0005 and B0006 in the NASA dataset were: IC peak value, constant voltage current area, IC peak voltage, and constant current area.

[0036] Step 3: Use four health indicators of batteries B0005 and B0006 from the NASA dataset to train the improved dynamic ANFIS model. Specifically, the key hyperparameters of the improved dynamic ANFIS model are determined through Bayesian optimization, and dynamic parameters driven by operating condition characteristics are introduced. To achieve rule-based adaptive activation, the model combines recursive least squares (RLS) and gradient descent to optimize the antecedent and consequent parameters, constructing an improved dynamic ANFIS model that balances accuracy, efficiency, and adaptability to operating conditions, and finally outputs the battery SOH estimate.

[0037] Step 31: Optimize the ANFIS model using Bayesian methods to select the optimal hyperparameter combination; let the set of hyperparameters for the ANFIS model be: (9) in, Let the initial standard deviation of the membership function be . For dynamic parameters The minimum value, For dynamic parameters The maximum value; The number of membership functions controls the total number of fuzzy rules. , where n=4 is the number of key features.

[0038] Using the root mean square error of three-fold cross-validation as the objective function, i.e., minimizing the SOH estimation error of the model in the cross-validation of the training set, the corresponding hyperparameter combination is obtained iteratively, such as... Figure 4 As shown.

[0039] Step 32: In the fuzzification layer, the key feature inputs are converted into fuzzy membership values ​​using a Gaussian membership function. (10) In the formula, Input features membership function, , Let be the center and radius of the membership function of the q-th input, respectively. Figure 5 In network structure for The membership function center and membership function radius are expressed as: and .

[0040] Step 33: At the criterion layer, the complexity of the working condition is measured by the instantaneous rate of change of the input features;

[0041] Based on an improved adaptive fuzzy neural network inference system, such as Figure 5 As shown, the definition Let the emission intensity be the k-th rule, then: (11) Introducing dynamic parameters driven by operating condition characteristics Based on operating condition characteristics such as current ratio changes and temperature fluctuations, the activation threshold of the rules is dynamically adjusted, activating only rules whose cumulative intensity is greater than the product of the dynamic parameter and the total emission intensity. This introduces dynamic parameters driven by operating condition characteristics. : (12) Among them, the hyperparameter optimization obtained and Calculate the baseline activation coefficients as boundaries. ; Define the operating condition complexity factor The degree of fluctuation in the current operating condition is characterized by calculation based on the rate of change of key features, as shown in the following formula: (13) In the formula, tanh is the hyperbolic tangent function; This is the operating condition sensitivity coefficient, configured according to actual engineering needs; The normalized operating condition complexity is obtained by normalizing the ratio of the rate of change of key features to the rate of change of SOH through moving average.

[0042] The set of rules that are activated is as follows: (14) in, The set of rules that are activated. The total number of fuzzy rules.

[0043] Step 34: The emission intensity of the activation rules is standardized in the normalization layer to eliminate the influence of dimensions; in the defuzzification layer, each activation rule corresponds to a linear output function. The weighted summation yields the estimated SOH value: (15) In the formula, Let x1 be the linear parameter of the consequent of the k-th rule and the weight vector; x2, x3, and x4 are the four key features obtained after RF-AHP: IC peak value, constant voltage current area, IC peak voltage, and constant current area.

[0044] Step 35: Optimize the consequent parameters of the fuzzy rules using the recursive least squares method, and update the individual difference adjustment factor serving as the bias term using the gradient descent method to characterize the capacity deviation of the same batch of batteries under the same working conditions, thereby improving the generalization ability of the model for the same model but different individual batteries.

[0045] Train with the battery data of B0005 and B0006 in the NASA dataset, and perform generalization detection using the data of B0007 and B0018 batteries, as Figure 6 、 Figure 7 、 Figure 8 and Figure 9 . The verification results show that the improved ANFIS model has good generalization ability, as shown in Table 1.

[0046] Table 1 is a comparison chart of the verification results of the improved ANFIS, standard ANFIS, LSTM, and GPR

[0047] As can be seen from Table 1, the global RMSE is approximately 0.0293, the low SOH interval is approximately 0.0310, the MAE is approximately 0.0238, the MAPE is approximately 2.97%, and the training time is less than 30 seconds, which is much faster than other data-driven models, and it has strong interpretability and better comprehensive performance than other methods.

[0048] Step 4: Construct a phased full-life cycle battery SOH evaluation system;

[0049] Specifically, divide the battery life cycle stages according to the SOH estimated value and the attenuation rate.

[0050] Step 41: Define four stages according to the SOH estimated value, normal state (SOH > 90%), attention state (80% < SOH ≤ 90%), abnormal state (70% < SOH ≤ 80%), severe state (SOH ≤ 70%). Among them, the SOH attenuation rate of common batteries in the normal state and attention state is relatively slow, while the attenuation rate in the abnormal state and severe state increases sharply.

[0051] Step 42: Design differentiated evaluation indicators for each stage. In the normal state, pay attention to whether there are extreme abnormal values in the electrical quantities to characterize circuit faults; in the attention state, pay attention to the change rate of the electrical quantities under working conditions, such as current and voltage; in the abnormal state and severe state, pay attention to the attenuation rate and the internal resistance change rate. In addition, configure a battery internal pressure detection device and perform DEMS (Differential Electrochemical Mass Spectrometry) tests in a timely manner to prevent potential safety hazards due to battery load.

[0052] Step 43: Taking into account parameters such as SOH estimate, temperature change rate, internal resistance change rate, current and voltage amplitude, an alarm mechanism or cloud backup will be triggered if an abnormal value is detected.

[0053] A battery health status evaluation device, used to perform the above-described battery health status evaluation method, includes: The data acquisition module is used to match battery types to determine corresponding key features, retrieve the key feature categories to be extracted from the database, and collect key feature data. The battery health status estimation module is used to input the features obtained by the data acquisition module into the data-driven model (i.e., the improved dynamic ANFIS model) based on the improved adaptive fuzzy neural network inference system for this battery type, so as to realize the battery health status estimation. The model running time is less than 1 second. The health assessment and communication module is used to implement the battery health assessment system, output a health score, and perform corresponding data storage and alarm operations based on the health score. The data is also backed up to the cloud for retrospective verification of the model's accuracy.

Claims

1. A battery state of health evaluation method based on an improved adaptive fuzzy neural network, characterized in that, The method comprises the steps of: S1, constructing a lithium battery capacity aging model based on battery health indicators extracted from a multi-battery data set; S2, constructing an analytic hierarchy process model based on the lithium battery capacity aging model, taking the battery health indicators as features, and performing multi-criteria feature screening based on the random forest-analytic hierarchy process method to screen key features; S3, training the improved dynamic ANFIS model by taking the key features as input and the battery SOH estimation value as output, and saving the best improved dynamic ANFIS model for SOH estimation; S4, constructing a staged full-life-cycle battery health state evaluation system according to the SOH estimation value and the decay rate.

2. The battery state of health evaluation method based on the improved self-adaptive fuzzy neural network according to claim 1, characterized in that, The extraction method of the battery health indicators is as follows: using a multi-battery data set, the original charge-discharge cycle information containing time, voltage, current, temperature and capacity is structured into unified charge data and discharge data information according to the charge-discharge data of the lithium battery, and several groups of health indicators are extracted, and the health indicators are standardized by median-quartile range scale; The lithium battery capacity aging model adopts an exponential decay model to fit the capacity degradation curve and describe the battery aging trend.

3. The battery state of health evaluation method based on the improved self-adaptive fuzzy neural network according to claim 1, characterized in that, The analytic hierarchy process model includes a target layer, a criterion layer and a scheme layer, wherein the target layer is used to obtain the best features for SOH estimation effect; The criterion layer serves as a screening criterion and includes an accuracy layer, a correlation layer and a complexity layer; the accuracy layer calculates the contribution of each feature to SOH estimation by a random forest algorithm, the correlation layer evaluates the correlation between each feature and SOH by a Spearman correlation coefficient, and the complexity layer scores in combination with the complexity of feature extraction; The scheme layer determines the weight of each criterion by the analytic hierarchy process method, calculates the comprehensive score of the features, and selects the top four features as key features.

4. The battery state of health evaluation method based on the improved self-adaptive fuzzy neural network according to claim 1, characterized in that, During the training process of the improved dynamic ANFIS model, the key hyperparameters of the ANFIS model are determined by Bayesian optimization, and the dynamic parameters driven by the working condition characteristics are introduced The rule adaptive activation is realized, the recursive least squares method is used to optimize the consequent parameters of the fuzzy rule, and the gradient descent method is used to update the individual difference adjustment factor, which represents the capacity deviation of the same batch of batteries under the same working condition.

5. The battery state of health evaluation method based on the improved self-adaptive fuzzy neural network according to claim 4, characterized in that, Setting the hyperparameter set of the ANFIS model is: ， wherein, is the initial standard deviation of the membership function, is the minimum value of the dynamic parameter, is the maximum value of the dynamic parameter; is the number of membership functions, controlling the total number of fuzzy rules n is the number of key features; The root mean square error of three-fold cross-validation is taken as the objective function, and the corresponding hyperparameter combination is obtained by iteration; In the fuzzification layer, the Gaussian membership function is used to convert the key feature input into fuzzy membership; In the criterion layer, the instantaneous change rate of the input feature is used to measure the complexity of the working condition; In the normalization layer, the emission intensity of the activation rule is standardized; in the defuzzification layer, each activation rule corresponds to a linear output function, and the SOH estimation value is obtained by weighted summation.

6. The battery state of health evaluation method based on improved self-adaptive fuzzy neural network according to claim 1, wherein, In step S4, the battery life cycle stage is divided according to the SOH estimation value: SOH > 90% is the normal state, 80% < SOH ≤ 90% is the attention state, 70% < SOH ≤ 80% is the abnormal state, and SOH ≤ 70% is the serious state, wherein the SOH of the normal state and the attention state decays slowly with the cycle number, and the decay rate of the abnormal state and the serious state increases sharply; Differential evaluation indexes are designed for each stage: for the normal state, whether the electrical quantity appears an extreme abnormal value is concerned; for the attention state, the change rate of the working condition electrical quantity is concerned; for the abnormal state and the serious state, the decay rate and the internal resistance change rate are concerned.

7. A battery state of health evaluation device for executing the battery state of health evaluation method according to any one of claims 1 to 6, characterized by The method comprises the steps of: The data acquisition module is used to match the battery type to determine the corresponding key features, retrieve the key feature categories to be extracted in the database, and acquire the key feature data; A battery health state estimation module, which takes the features obtained by the data acquisition module as input, estimates the battery health state through an improved dynamic ANFIS model; A health degree evaluation and communication module, which is used to realize a battery health degree evaluation system and output a health degree score; And according to the health degree score, corresponding data storage, alarm operation and data backup to the cloud are performed.

Images