CN117007977B

CN117007977B - Energy storage battery health state diagnosis method

Info

Publication number: CN117007977B
Application number: CN202311280860.0A
Authority: CN
Inventors: 朱小宝; 邹雨欣; 邹炜康
Original assignee: Nanchang Hangkong University
Current assignee: Nanchang Hangkong University
Priority date: 2023-10-07
Filing date: 2023-10-07
Publication date: 2024-01-09
Anticipated expiration: 2043-10-07
Also published as: CN117007977A

Abstract

The invention provides a method for diagnosing the health state of an energy storage battery, which is used for collecting health degree data corresponding to various parameters of an overcharged and overdischarged battery monomer; extracting fluctuation amplitude characteristics and frequency distribution characteristics by a wavelet transformation improvement method; randomly sampling after constructing the label to form a plurality of mutually exclusive differential subsets; calculating Euclidean distance and calculating diagnosis evidence; constructing a prediction model of the health state of the battery; judging whether the battery has the risk of overcharge or overdischarge; fusing the prediction model and the diagnosis evidence obtained in the offline diagnosis through the D-S evidence theory to obtain a final diagnosis result; based on battery health data, a wavelet transformation improvement algorithm and a support vector machine are adopted to conduct feature extraction and model construction, meanwhile, a K-nearest neighbor algorithm and a D-S evidence theory are combined to conduct evaluation and diagnosis of battery health states, risks of battery overcharge or overdischarge can be effectively identified, early warning and treatment are timely conducted, and safety and reliability of the battery are guaranteed.

Description

Energy storage battery health state diagnosis method

Technical Field

The invention relates to the technical field of energy storage batteries, in particular to a method for diagnosing the health state of an energy storage battery.

Background

With the increase of charge and discharge times, the energy storage battery is at risk of capacity reduction or even failure, the energy efficiency of the energy storage power station is affected when the energy storage battery is light, and fire accidents can be caused when the energy storage power station is heavy. Because the energy storage power station is huge in scale, the fire extinguishing difficulty is extremely high after a fire disaster occurs, the direct loss and the indirect loss are extremely high, and the integral fluctuation of a power grid is easy to cause.

The battery can be damaged due to overcharge or overdischarge and the like, the safety can be caused by factors such as the health condition of the energy storage battery, unbalanced charge and the like, and the method for evaluating the state of the energy storage battery and analyzing the health condition is very important for improving the safety and reliability of the energy storage battery. Along with the development of machine learning and intelligent optimization technology, an intelligent monitoring service for identifying the health state of an energy storage battery is applied, and the establishment of an intelligent diagnosis model for the health state of the energy storage battery has become an important research point of an energy storage battery system.

The existing energy storage battery health diagnosis method has the problems of reduced training data capacity, large influence of sample set division and the like, and therefore the energy storage battery health state diagnosis method is provided.

Disclosure of Invention

In order to solve the problems, the invention provides a method for diagnosing the health state of an energy storage battery, so as to more exactly solve the problems of reduced training data capacity, larger influence of sample set division and the like in the existing method for diagnosing the health of the energy storage battery.

The invention is realized by the following technical scheme:

the invention provides a method for diagnosing the health state of an energy storage battery, which comprises the following steps:

s1: performing repeated overcharging and overdischarging on the energy storage batteries of the same type, and collecting health degree data corresponding to each parameter of the energy storage battery monomer one by one during each charging and discharging period;

s2: extracting fluctuation amplitude characteristics and frequency distribution characteristics carrying battery health state characteristics from each health degree data by a wavelet transformation improvement method according to the health degree data changing along with time;

s3: after a sample characteristic data set containing category labels is constructed according to the fluctuation amplitude characteristics and the frequency distribution characteristics, randomly sampling the sample characteristic data set to form a plurality of mutually exclusive differential subsets; calculating Euclidean distance between a sample to be detected and samples in each subset, and finding out k nearest neighbor samples and corresponding categories thereof; counting the number of samples of different types in each subset and calculating the diagnosis evidence of each subset;

s4: according to the fluctuation amplitude characteristics and the frequency distribution characteristics processed in the sample characteristic data set, constructing a prediction model of the battery health state through a support vector machine;

s5: inputting health degree data of all parameters of the energy storage battery monomer collected in real time, obtaining the health state of the battery through a prediction model, and judging whether the battery has the risk of overcharge or overdischarge; if the battery meets the judgment conditions of overcharge or overdischarge or the results of the health degree evaluation and the diagnosis evidence indicate that the battery state is abnormal, an early warning signal is sent out, and the battery is processed according to the diagnosis result; if the alarm is still continuous, removing the deteriorated battery monomer according to the alarm information, and performing off-line diagnosis to find out the failure reason;

s6: and fusing the prediction model with the diagnosis evidence obtained in the offline diagnosis through the D-S evidence theory to obtain a final diagnosis result.

Further, the step of extracting fluctuation amplitude characteristics and frequency distribution characteristics carrying battery health status characteristics from each health degree data through a wavelet transformation improvement method according to the health degree data changing along with time comprises the following steps of;

performing Fourier transform on the health degree data changing along with time to adaptively divide the Fourier spectrum of the health degree data into N continuous intervals; obtaining an initial selected frequency band boundary point through a classical scale space method, calculating parameter values of each frequency band and solving a parameter mean value; determining whether the frequency band parameter value is larger than the average value or not, and then determining the frequency band demarcation point again;

the determining whether the frequency band parameter value is greater than the average value includes: if yes, the frequency band is reserved, and if not, the adjacent frequency bands are combined.

Further, the step of constructing a prediction model of the battery health state through a support vector machine according to the fluctuation amplitude characteristic and the frequency distribution characteristic processed in the sample characteristic data set comprises the following steps of;

and taking the extracted fluctuation amplitude characteristic and frequency distribution characteristic as input, taking the health state of the battery as output, and then using an SVM to establish a mapping relation between the fluctuation amplitude characteristic and the frequency distribution characteristic and train a prediction model of the health state of the battery.

Further, after the prediction model predicts, the accuracy and precision of the modeled model are verified by adopting a root mean square deviation algorithm, a mean absolute percentage error algorithm and a mean absolute value error algorithm as evaluation criteria.

Further, the root mean square difference algorithm is：Wherein->Is true value +.>N is the number of samples for the predicted value.

Further, the average absolute percentage error algorithm is as follows:。

further, the average absolute value error algorithm is as follows:。

further, the step of fusing the prediction model and the diagnosis evidence obtained in the offline diagnosis through the D-S evidence theory to obtain a final diagnosis result includes:

according to the D-S evidence theory, representing evidence of each source as a credibility allocation function, and obtaining the credibility value of each result through the credibility allocation function;

fusing the prediction model with diagnostic evidence obtained in offline diagnosis through a Dempster rule;

and taking the result with the maximum credibility value as a final diagnosis result.

The invention has the beneficial effects that: according to the battery health degree data, a wavelet transformation improvement algorithm and a support vector machine are adopted to conduct feature extraction and model construction, meanwhile, a K-nearest neighbor algorithm and a D-S evidence theory are combined to conduct evaluation and diagnosis of the battery health state, risks of overcharge or overdischarge of the battery can be effectively identified, early warning and treatment are timely conducted, and safety and reliability of battery use are guaranteed.

Drawings

Fig. 1 is a flow chart of a method for diagnosing a state of health of an energy storage battery according to the present invention.

The implementation, functional features and advantages of the present application will be further described with reference to the accompanying drawings in conjunction with the embodiments.

Detailed Description

In order to more clearly and completely describe the technical scheme of the invention, the invention is further described below with reference to the accompanying drawings.

Referring to fig. 1, the present invention provides a method for diagnosing a health status of an energy storage battery, which comprises the following steps:

specifically, the multiple overcharging and overdischarging experiments are performed on the same type of energy storage battery, so as to simulate various states possibly encountered by the battery in the actual use process, including a normal state, a slightly degraded state, a severely degraded state, and an extremely degraded state, and the like, under these different states, parameters including voltage, current, temperature, internal resistance, and the like of the battery are different, where the voltage and current can reflect the charge and discharge states of the battery, and the temperature and internal resistance can reveal the chemical reaction conditions occurring inside the battery, and in this way, we can collect parameter data of the battery under different health states, so as to provide basis for subsequent data analysis and model establishment, and during each charge and discharge period we can use existing professional equipment to collect various parameters of the energy storage battery monomer, including but not limited to: the parameters such as voltage, current, temperature and internal resistance can directly reflect the working state of the battery and are closely related to the health degree of the battery, meanwhile, the health state of the battery corresponding to each sampling point, namely health degree data, is required to be recorded, and the health degree of the battery is usually measured by the performance parameters such as the capacity and the internal resistance of the battery; the collected parameter data with category labels (namely health degree) form a sample characteristic data set; each sample data contains a set of battery parameters and corresponding health, which will be used to train a predictive model so that the model can predict the health of the battery from the input battery parameters.

S2: extracting fluctuation amplitude characteristics and frequency distribution characteristics carrying battery health state characteristics from each health degree data by a wavelet transformation improvement method according to the health degree data changing along with time; wavelet transformation is a mathematical tool that can be used for signal processing and feature extraction. Its main advantage is that it can take into account both the time (or space) information and the frequency information of the signal at multiple scales, thus being particularly suitable for processing non-stationary signals, such as battery parameter data; here we use a wavelet transform improvement algorithm to process the collected health data. First, we consider each piece of health data as a signal, and then wavelet transform this signal. Wavelet transformation breaks down a signal into a series of wavelets, each wavelet corresponding to a characteristic of the signal over a particular time period and frequency range; wave amplitude characteristics this refers to the amplitude of each wavelet obtained after wavelet transformation reflecting the intensity or energy of the signal over different time periods and frequency ranges. In battery health diagnostics, the fluctuation amplitude feature may help us find abnormal fluctuations in battery parameters, which is often an important indicator of battery health changes; the frequency distribution characteristic is the frequency of each wavelet obtained after wavelet transformation, and reflects the components or structures of the signal in different frequency ranges. In the diagnosis of the battery health state, the frequency distribution characteristic can help us analyze the change rule of the battery parameters, so as to reveal the potential rule of the battery health state; performing Fourier transform on the health degree data changing along with time to adaptively divide the Fourier spectrum of the health degree data into N continuous intervals; obtaining an initial selected frequency band boundary point through a classical scale space method, calculating parameter values of each frequency band and solving a parameter mean value; determining whether the frequency band parameter value is larger than the average value or not, and then determining the frequency band demarcation point again; the determining whether the band parameter value is greater than the mean value includes: if yes, reserving the frequency band, otherwise, merging adjacent frequency bands; in performing wavelet transforms, frequency band division is a critical step. Since this step determines the frequency resolution of the signal, i.e. the information about which specific frequencies in the signal can be observed. However, the traditional method based on the maximum point of the Fourier spectrum needs to rely on known prior information, which is often difficult to obtain under the actual condition; the scale-space approach may adaptively select the band-splitting points such that known a priori information is no longer required. However, classical scale-space methods screen out too many band-splitting points, which may lead to band-splitting phenomena, i.e. some bands may be excessively subdivided, affecting the subsequent analysis. In order to solve this problem, the present application performs fourier transform on the health degree data that changes with time, and adaptively divides the fourier spectrum of the health degree data into N consecutive sections. Fourier transform is a method of analyzing a signal in the frequency domain, which can exhibit the intensity of individual frequency components in the signal. And obtaining the boundary points of the initially selected frequency bands through a classical scale space method, calculating the parameter values of each frequency band and solving the parameter average value. The purpose of this step is to obtain a rough frequency band division. And judging whether the parameter value of each frequency band is larger than the average value. If so, then the band is reserved; otherwise, the band is merged with the adjacent band. The purpose of this step is to reduce the number of bands and avoid band breakage. The band split point is redetermined. After the above steps are completed, we obtain a new frequency band division, which is finer than the original division, and can better reflect the frequency information in the signal.

S3: after a sample characteristic data set containing category labels is constructed according to fluctuation amplitude characteristics and frequency distribution characteristics, randomly sampling the sample characteristic data set to form a plurality of mutually exclusive differential subsets; calculating Euclidean distance between a sample to be detected and samples in each subset, and finding out k nearest neighbor samples and corresponding categories thereof; counting the number of samples of different types in each subset and calculating the diagnosis evidence of each subset; in particular, in experiments of performing multiple overcharging and overdischarging on the battery, we can obtain a large amount of battery parameter data, and also record the health status of the battery at each experimental stage, i.e. the category label. For example, if the battery is normal, it may be marked as "normal"; if the battery is in a degraded state, it may be marked as "degraded". The specific label setting is determined according to actual requirements. Then we store these battery parameters (e.g. voltage, current, temperature, etc.) as features, along with the corresponding class labels, to form a sample. All samples are organized to obtain a sample characteristic data set A containing category labels; to ensure that the model learns and generalizes correctly, we will typically divide the entire data set into multiple subsets, such as training and testing sets. Wherein the training set is used to train the model and the test set is used to evaluate the performance of the model. The partitioning method can help us avoid model overfitting and can better evaluate the performance of the model on unknown data; then two n-dimensional feature vectors x= (X1, X2,..xn) and y= (Y1, Y2,., yn) are set by the euclidean distance method, the euclidean distance between X and Y being defined as: d (X, Y) =sqrt [ (X1-Y1) and + (X2-Y2) are + (xn-yn) for + (X1-Y2) for + (X2-Y2) for + (X1-Y2) for X, where sqrt represents the square root. The distance between the sample to be measured and each sample in each subset is calculated, the similarity degree of the two samples in the feature space is reflected, and for each sample to be measured, the k training samples closest to the sample to be measured are found out, and the samples are considered to be the most similar to the sample to be measured. We then record class labels for the k samples, which will be used to predict the class of the sample to be tested; counting the number of different class samples in each subset and calculating the diagnostic evidence of each subset: for the k samples found in the previous step, we count the number of samples for each category. From this quantitative information we can then calculate the diagnostic evidence for each category, i.e. the likelihood that the sample to be tested belongs to each category. Finally, we select the class with the highest probability as the predicted class of the sample to be tested.

S4: according to the fluctuation amplitude characteristics and the frequency distribution characteristics processed in the sample characteristic data set, constructing a prediction model of the battery health state through a support vector machine; the support vector machine (Support Vector Machine, SVM for short) is a powerful classification and regression tool, is particularly suitable for processing data of a high-dimensional feature space, and achieves the balance between accuracy and computational complexity by means of a loss function and a penalty factor; the loss function is used to measure the difference between the model predicted and actual values. For SVM, the usual loss function is"Hinge Loss" (Hinge Loss). Such a loss function would impose zero loss on samples located on the side of the correct classification boundary, while for samples located on the side of the incorrect classification boundary a linearly increasing loss would be imparted. The penalty factor, also called regularization parameter, is a superparameter used to control the complexity of the model. It is mainly used to balance the fit and complexity of the model to prevent over-fitting or under-fitting. In SVM, the penalty factor determines the "cost" of classifying the errors. If the penalty factor is large, the model will strive to avoid classification errors, possibly resulting in a model that is too complex and easy to overfit; conversely, if the penalty factor is small, the model may accept some sort of error, possibly making the model too simple and easy to under fit; specifically, based on the collected sample feature data set (including the fluctuation amplitude feature and the frequency distribution feature), we can construct a multidimensional feature space. Each sample point has a position in this space whose coordinates are determined by the eigenvalues and its class labels determine which class the point belongs to. The goal of the SVM is to find an optimal hyperplane in this feature space so that it can separate the different classes of samples as correctly as possible and so that the distance (i.e., spacing) of the classes of samples from this hyperplane is maximized. In practice, we will typically use kernel functions to map the original feature space to a higher dimensional feature space in order to find a better hyperplane. This is known as core skill. When the model training is completed, we can use it to predict the state of health of the new battery sample. Specifically, firstly, calculating fluctuation amplitude characteristics and frequency distribution characteristics of a new sample, inputting the characteristics into an SVM model, and finally obtaining output, wherein the finally obtained output is that the predicted health state takes the extracted fluctuation amplitude characteristics and frequency distribution characteristics as input, the health state of a battery as output, and then using the SVM to establish a mapping relation between the fluctuation amplitude characteristics and the frequency distribution characteristics and train a prediction model of the health state of the battery; after the prediction model predicts, modeling is carried out by adopting a root mean square deviation algorithm, an average absolute percentage error algorithm and an average absolute value error algorithm as evaluation criteriaVerifying the accuracy and precision of the model; root mean square difference (Root Mean Square Error, RMSE): it is the square root of the average of the sum of the squares of the observed and true deviations, and can measure the standard deviation of the prediction error. The smaller the value, the higher the accuracy of model prediction is explained; the root mean square difference algorithm is:wherein->Is true value +.>N is the number of samples for the predicted value; the mean absolute percentage error algorithm is: />The method comprises the steps of carrying out a first treatment on the surface of the Absolute percent error (Mean Absolute Percentage Error, MAPE): it is the average of the absolute error over the true value of all individual observations. This is a percentage reflecting the magnitude of the prediction error relative to the true value. The smaller the value, the higher the accuracy of model prediction is explained; the average absolute value error algorithm is:the method comprises the steps of carrying out a first treatment on the surface of the Mean absolute error (Mean Absolute Error, MAE): it is the average of the absolute errors of all individual observations. MAE is less sensitive to outliers than RMSE, and thus can better reflect the prediction accuracy of the model in most cases. The smaller the value, the higher the accuracy of model prediction.

S5: inputting health degree data of all parameters of the energy storage battery monomer collected in real time, obtaining the health state of the battery through a prediction model, and judging whether the battery has the risk of overcharge or overdischarge; if the battery meets the judgment conditions of overcharge or overdischarge or the results of the health degree evaluation and the diagnosis evidence indicate that the battery state is abnormal, an early warning signal is sent out, and the battery is processed according to the diagnosis result; if the alarm is still continuous, removing the deteriorated battery monomer according to the alarm information, and performing off-line diagnosis to find out the failure reason; specifically, inputting health degree data of all parameters of the energy storage battery monomer collected in real time, and obtaining the health state of the battery through a prediction model: then, the real-time data are input into a pre-trained support vector machine model, and the model predicts the current state of health of the battery according to the input data. Judging whether the battery has the risk of overcharge or overdischarge: then, we need to analyze the prediction result of the model to determine whether there is a risk of overcharging or overdischarging the battery. If the battery meets the judgment conditions of overcharge or overdischarge or the results of the health degree evaluation and the diagnosis evidence indicate that the battery state is abnormal, an early warning signal is sent out, and the battery is processed according to the diagnosis result: if the battery is judged to be at risk of overcharging or overdischarging, or the state of health of the battery is assessed as abnormal, we need to immediately send out an early warning signal and take appropriate processing measures such as reducing the charge current, increasing the discharge current, etc. If the alarm is still continuous, removing the deteriorated battery cell according to the alarm information, and performing off-line diagnosis to find out the failure reason: if the problem of the battery is not solved even if the treatment measures are taken, then we need to remove the problem battery from the system and then perform an offline diagnosis to find the specific cause of the battery failure. Thus, the problem that the battery affects the operation of the whole system can be avoided, and the failure mechanism of the battery can be better understood so as to improve future prediction and processing strategies; through the series of steps, the health state of the battery can be monitored in real time, early warning can be sent out in time when the battery is in a problem, and appropriate measures are taken to protect the safety of the battery and a user.

S6: fusing the prediction model and the diagnosis evidence obtained in the offline diagnosis through the D-S evidence theory to obtain a final diagnosis result; according to the D-S evidence theory, representing evidence of each source as a credibility allocation function, and obtaining the credibility value of each result through the credibility allocation function; fusing the prediction model with diagnostic evidence obtained in offline diagnosis through a Dempster rule; taking the result with the maximum credibility value as a final diagnosis result; D-S evidence Theory (Dempster-Shafer Theory) is a Theory for merging incomplete and uncertain information based on a probabilistic framework. D-S evidence theory is particularly useful for processing evidence provided by multiple sources that may be conflicting or incomplete. In this step, we use D-S evidence theory to integrate diagnostic results from two sources: one source is the predicted result obtained by the SVM model, and the other source is the result obtained by the offline diagnosis. In the D-S evidence theory, evidence for each source is represented as a confidence score function that gives confidence values for each possible outcome. These confidence values reflect our degree of trust in the individual results. When we have evidence from multiple sources, the evidence can be fused using the Dempster rule. The main idea of this rule is: if both sources support a result, the final confidence value for this result should be higher than the confidence value given by either source alone. Finally, we select the result with the highest confidence value as the final diagnostic result. If there are multiple results with equal confidence values, we need to rely on other information or take further checks to determine the final result. In this way, the D-S evidence theory can help us integrate evidence from different sources, and improve the accuracy and reliability of diagnosis.

Of course, there are many other embodiments of the present invention, and based on this embodiment, those of ordinary skill in the art will obtain other embodiments without any inventive effort, which fall within the scope of the present invention.

Claims

1. A method for diagnosing the state of health of an energy storage battery, the method comprising:

s6: fusing the prediction model and the diagnosis evidence obtained in the offline diagnosis through the D-S evidence theory to obtain a final diagnosis result;

the step of extracting fluctuation amplitude characteristics and frequency distribution characteristics carrying battery health state characteristics from each health degree data through a wavelet transformation improvement method according to the health degree data changing along with time comprises the following steps;

2. The method for diagnosing the health state of the energy storage battery according to claim 1, wherein the step of constructing a prediction model of the health state of the battery by a support vector machine according to the fluctuation amplitude characteristic and the frequency distribution characteristic processed in the sample characteristic data set comprises the steps of;

3. The method according to claim 2, wherein the model is used to verify the accuracy and precision of the model by using a root mean square error algorithm, a mean absolute percentage error algorithm and a mean absolute value error algorithm as evaluation criteria after the prediction model predicts.

4. A method of diagnosing a state of health of an energy storage battery according to claim 3, wherein the root mean square deviation algorithm is:wherein->Is true value +.>N is the number of samples for the predicted value.

5. The method for diagnosing a state of health of an energy storage battery according to claim 4, wherein the mean absolute percentage error algorithm is:。

6. the method of claim 5, wherein the mean absolute error algorithm is:。

7. the method for diagnosing a state of health of an energy storage battery according to claim 1, wherein the step of fusing the prediction model and the diagnostic evidence obtained in the offline diagnosis by the D-S evidence theory to obtain a final diagnostic result comprises: