Energy storage battery state of health assessment method based on neural network model

Abstract

The application relates to the technical field of battery monitoring, in particular to a method for evaluating the health state of an energy storage battery based on a neural network model, which comprises the following steps: step 1: obtaining the cycle characteristics of the energy storage battery; step 2: performing regression classification on the cycle characteristics based on a support vector machine to generate an initial life interval; step 3: collecting original relaxation characteristics of the energy storage battery for several adjacent times, and based on the continuity factor of the original relaxation characteristics and historical relaxation characteristics; the sparse feature set comprises a pulse factor, a waveform factor, a peak value factor, a wavelet energy ratio, a wavelet energy entropy and a root mean square value; step 4: performing feature expansion on the sparse feature set to obtain three-dimensional information; and step 5: inputting the key feature matrix and the initial life interval into the neural network model to generate accurate life information of the energy storage battery. In the technical scheme, the relatively simple cycle characteristics are used to obtain the approximate life interval of the energy storage battery.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery monitoring, in particular to a method for evaluating the health state of energy storage batteries based on a neural network model. BACKGROUND

[0002] Energy storage batteries play an important role in the fields of renewable energy grid connection and grid peak regulation. It is crucial to accurately predict the life of energy storage batteries. Existing life prediction methods are usually based on extracting battery life-related features, such as cycle number, voltage / current data during charging / discharging process, internal resistance data, etc., and analyzing the correlation between these features for prediction. However, the prediction accuracy of such methods is limited, and they can only provide a rough estimate of the life. To improve the accuracy, some solutions further utilize the relaxation characteristics of the battery to extract more detailed information. However, this leads to an increase in feature dimension, which puts higher requirements on the prediction model: the model not only needs to handle more complex information, but also the output of the life prediction label (or the segmentation method of the detailed features) must be more accurate. This leads to a dramatic expansion of the information processing dimension of the model, which in turn causes the risk of model structure redundancy and overfitting during the training process.

[0003] In summary, the current energy storage battery life prediction system cannot accurately predict the service life. SUMMARY

[0004] The summary section of the present application is used to introduce the concepts in a brief form, which will be described in detail in the specific embodiments section. The summary section of the present application is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0005] Some embodiments of the present application propose a method for evaluating the health state of energy storage batteries based on a neural network model to solve the technical problems mentioned in the background section.

[0006] As a first aspect of the present application, some embodiments of the present application provide a method for evaluating the health state of energy storage batteries based on a neural network model, comprising the following steps:

[0007] Step 1: Obtain the cycle characteristics of the energy storage battery, and pre-process the cycle characteristics to generate initial life characteristics; the cycle characteristics include: cycle number, time when the charging voltage reaches the peak value, time when the discharging voltage reaches the peak value, time when the charging current reaches the peak value;

[0008] Step 2: Perform regression classification on the cycle characteristics based on a support vector machine to generate an initial life interval;

[0009] Step 3: Collect the original relaxation characteristics of the energy storage battery for several adjacent times, select the original relaxation characteristics with the highest correlation with the historical relaxation characteristics from the several original relaxation characteristics based on the continuity factor of the original relaxation characteristics and the historical relaxation characteristics, and extract a sparse feature set from the V-t curve corresponding to the original relaxation characteristics;

[0010] The sparse feature set includes a pulse factor, a waveform factor, a peak factor, a wavelet energy ratio, a wavelet energy entropy, and a root mean square value;

[0011] Step 4: Perform feature expansion on the sparse feature set to obtain three-dimensional information;

[0012] Step 5: Input the key feature matrix and the initial life interval into the neural network model to generate accurate life information of the energy storage battery.

[0013] In the technical solution provided in the present application, a relatively simple cycle feature is used to obtain a rough life interval of the energy storage battery, and a key feature matrix with higher dimensions is used to further refine the rough life interval of the energy storage battery. Therefore, the accuracy of life prediction can be effectively increased. In practice, this life prediction method can effectively increase the accuracy of life prediction, and in order to avoid model dimension explosion, the rough life interval is used to reduce the dimension of the neural network model output. Furthermore, only the latest V-t curve needs to be input by using the original relaxation characteristic analysis, thereby reducing the dimension of the neural network model input. In the case of synchronous reduction of the input dimension and the output dimension, the internal gradient explosion of the neural network model can be reduced and the overfitting phenomenon of the model can be reduced while ensuring the prediction accuracy of the neural network model.

[0014] Further, step 1 includes the following steps:

[0015] Step 11: Obtain the model of the energy storage battery, and obtain the charging voltage peak value, the discharging voltage peak value, and the charging current peak value of the energy storage battery according to the model of the energy storage battery;

[0016] Step 12: Obtain the historical charging and discharging data of the energy storage battery, and obtain the time when the charging voltage reaches the peak value, the time when the discharging voltage reaches the peak value, and the time when the charging current reaches the peak value during each charging and discharging according to the charging voltage peak value, the discharging voltage peak value, and the charging current peak value;

[0017] Step 13: Extract the cycle number from the historical charging and discharging data;

[0018] Step 14: Take the time when the charging voltage reaches the peak value, the time when the discharging voltage reaches the peak value, the time when the charging current reaches the peak value, and the cycle number as cycle characteristics.

[0019] According to the signal of the energy storage battery, the pre-defined charging voltage peak value, discharging voltage peak value and charging current peak value of the energy storage battery are obtained, so that the cycle characteristics of the energy storage battery can be accurately obtained, and the cycle characteristics of the same type of energy storage battery cannot be accurately collected due to the lack of standard charging voltage peak value, discharging voltage peak value and charging current peak value of the energy storage battery.

[0020] In the scheme, the cycle characteristics and the support vector machine are used to obtain the initial life interval. Although the accuracy requirement of the life collection is not high, the initial life interval must cover the actual life of the battery. The support vector machine classifies the collected cycle characteristics in a feature space. Therefore, when the cycle characteristics have noise, the accuracy of the obtained initial life interval is not high.

[0021] Further, step 14 includes the following steps:

[0022] Step 141: arranging the time when the charging voltage reaches the peak value in each cycle in sequence to obtain a charging voltage peak value sequence;

[0023] arranging the time when the discharging voltage reaches the peak value in each cycle in sequence to obtain a discharging voltage peak value sequence;

[0024] arranging the time when the charging current reaches the peak value in each cycle in sequence to obtain a charging current peak value sequence;

[0025] Step 142: sequentially performing noise reduction processing on the charging voltage peak value sequence, the discharging voltage peak value sequence and the charging current peak value sequence;

[0026] Step 143: taking the noise-reduced charging voltage peak value sequence, the discharging voltage peak value sequence, the charging current peak value sequence and the cycle number as cycle characteristics.

[0027] The scheme obtains the standardized charging voltage peak value, discharging voltage peak value and charging current peak value of the energy storage battery according to the type of the energy storage battery in advance, and uses the same as a reference to calculate the time when the voltage / current reaches the peak value in each charging and discharging process. This method ensures the consistency and accuracy of the cycle characteristics (especially the time when the peak value is reached) of the same type of battery, effectively eliminates the data deviation caused by the lack of unified standard reference value, and provides a more reliable and accurate initial feature input basis for subsequent preliminary estimation and fine prediction of the life interval.

[0028] In the preprocessing of the charging and discharging data of the energy storage battery (especially the peak time series), the traditional noise reduction method often cannot adapt to the local fluctuation trend of the data due to the use of fixed threshold or global statistics, resulting in inaccurate identification of abnormal values: either too sensitive to delete effective data points, or too insensitive to leave significant noise points. In addition, the static processing method also easily causes the accumulation effect of abnormal data, further distorts the data distribution characteristics, and affects the input quality of the subsequent life prediction model.

[0029] Further, the noise reduction method in step 142 is as follows:

[0030] Step 1421: Pre-set window length;

[0031] Step 1422: For each data point x in the window, calculate whether x satisfies x∈(u-2σ, u+2σ), where u represents the mean of the window, and σ represents the standard deviation of the window;

[0032] If x satisfies x∈(u-2σ, u+2σ), the data point x is normal data, and if x If x does not satisfy x∈(u-2σ, u+2σ), the data point x is noise data;

[0033] Step 1423: Slide the window to remove all noise data, to complete the noise reduction processing of all peak value sequences.

[0034] The scheme can adaptively follow the change trend of the data by using the dynamic threshold noise reduction method (steps 1421-1423) based on the sliding window and local statistics (mean u, standard deviation σ). The ±2σ criterion (covering about 95% of normally distributed data) is used to accurately identify and remove significant noise points (abnormal values) in the local window, effectively retaining data points reflecting the true state of the battery. At the same time, the sliding processing mechanism of the window avoids the accumulation of abnormal values in the fixed area, ensuring the smoothness and reliability of the peak time series after noise reduction, and providing a cleaner and more accurate data basis for subsequent feature extraction and life prediction.

[0035] Step 2 includes the following steps:

[0036] Step 21: Pre-acquire a plurality of initial life interval labels, the initial life intervals are continuous and do not overlap; collect a plurality of corresponding cycle features for each initial life interval label, and collect the collected cycle features and corresponding life interval labels as a sample; collect N samples to generate a training data set S;

[0037] ; x i represents the cycle feature of the i-th sample, i represents the index of the sample, y iKt represents the i-th sample life interval label, and all labels are extracted to obtain a label set Kt; ; the elements in the label set are different life interval labels;

[0038] Step 22: constructing a kernel function ;

[0039] ; j represents the index of the sample, x j represents the cycle feature of the j-th sample, represents a kernel function bandwidth parameter;

[0040] Step 23: training a binary SVM classifier for each category, defining an original optimization problem of the SVM classification and setting an optimization condition for each SVM classifier;

[0041] The original optimization problem is:

[0042] ;

[0043] w represents a weight vector, b represents an offset, C represents a penalty factor, represents a slack variable;

[0044] The constraint condition is:

[0045] ;

[0046] represents a high-dimensional mapping;

[0047] Step 24: converting the original problem into a dual problem by using a Lagrange multiplier method;

[0048] ;

[0049] The constraint condition is:

[0050] ;

[0051] represents a Lagrange multiplier;

[0052] Step 25: constructing a decision function according to a solution result;

[0053] ; m represents the index of the SVM classifier;

[0054] represents a temporary binary label of the m-th classifier;

[0055] represents a Lagrange multiplier of the m-th classifier;

[0056] offset of the mth classifier;

[0057] classification result of the mth classifier on the sample x;

[0058] Step 26: input the collected cycle characteristics into the trained SVM classifier to obtain the classification results of all SVM classifiers, and take the label with the maximum decision value as the output.

[0059] In this scheme, support vector machine (SVM) is used to replace neural network for initial life interval classification prediction (steps 21-26). The SVM model has higher transparency and interpretability advantage, which can clearly reveal the influence weight of different cycle characteristics (such as peak time, cycle number) on the classification decision boundary. This interpretability not only helps to deeply understand the relationship between features and life interval, but also guides feature optimization and model parameter tuning, thereby effectively improving the accuracy of initial life interval classification (i.e. "overall classification effect"). Compared with the neural network black box model, SVM significantly enhances the analyzability of the prediction process and the reliability of the results while ensuring the classification accuracy.

[0060] When performing fine life prediction based on battery discharge curve (V-t curve), the battery state is not only affected by life decay, but also easily disturbed by other instantaneous factors (such as environmental fluctuations, working condition changes, measurement noise, etc.), resulting in that the relaxation features (such as VC, VD, ReVD, tVD, tReVD) and their corresponding V-t curves directly extracted from the current or adjacent charge-discharge cycle contain significant noise and non-life-related fluctuation information. These noise information will seriously pollute the feature extraction process, making the subsequent clustering and analysis results based on these features deviate from the true battery aging state, and it is difficult to accurately represent the remaining life of the current battery.

[0061] Further, step 3 includes the following steps:

[0062] Step 31: collect the original relaxation features of the adjacent several times of the energy storage battery, and the original relaxation features include VC, VD, ReVD, tVD and tReVD; wherein, VC represents the voltage at the beginning of the charging relaxation; VD represents the voltage drop amplitude at the initial stage of relaxation; ReVD represents the voltage value at the end of the relaxation process; tVD represents the time required for the voltage drop to reach a preset amplitude; tReVD represents the time required for the voltage drop to approach ReVD;

[0063] Step 32: perform continuity analysis on the collected original relaxation features and the historical relaxation features of the energy storage battery, and select the original relaxation features with the highest continuity with the historical relaxation features from the several original relaxation features;

[0064] Step 33: Extract the sparse feature set from the V-t curve corresponding to the original relaxation feature.

[0065] The present scheme effectively solves the noise interference problem by introducing a relaxation feature screening mechanism based on continuity analysis (steps 31-33). Specifically, the current collected several times of original relaxation features are analyzed for continuity with historical relaxation features, and the relaxation feature with the highest continuity with the historical decay trajectory and its corresponding V-t curve are selected. This method preferentially selects the data segment that best reflects the inherent aging trend of the battery and is least affected by instantaneous interference. Based on the high representation V-t curve after screening, the sparse feature set (such as pulse factor, waveform factor, etc.) is extracted, which significantly improves the correlation and reliability of the extracted features with the true remaining life of the battery, providing more accurate and pure input data for the subsequent refined life prediction model, thereby more accurately evaluating the remaining life of the current battery.

[0066] Further, step 32 includes the following steps:

[0067] Step 321: Obtain the initial life interval calculated by the energy storage battery when collecting the original relaxation feature;

[0068] Obtain the initial life interval corresponding to the historical relaxation feature;

[0069] Step 322: Arrange the VCs in the historical relaxation feature and the original relaxation feature in time sequence to generate a VC sequence;

[0070] Arrange the VDs in the historical relaxation feature in time sequence to generate a VD sequence;

[0071] Arrange the ReVDs in the historical relaxation feature in time sequence to generate a ReVD sequence;

[0072] Arrange the tVDs in the historical relaxation feature in time sequence to generate a tVD sequence;

[0073] Arrange the tReVDs in the historical relaxation feature in time sequence to generate a tReVD sequence;

[0074] Arrange the initial life interval corresponding to the historical relaxation feature in time sequence to generate a life sequence;

[0075] Step 323: Increase the VCs corresponding to the original relaxation feature of the adjacent several times in the VC sequence in turn, and add the initial life interval corresponding to the original relaxation feature of this time to the life sequence, and calculate the continuity coefficient of the VC sequence and the life sequence ; k represents the index of the original relaxation feature, represents the continuity coefficient calculated by adding the VC in the kth original relaxation feature to the VC sequence and the life sequence;

[0076] VD sequence is added to the VD sequence and the lifetime sequence to obtain continuity coefficient ; k represents the index of the original relaxation feature, represents the continuity coefficient obtained by adding the VD in the kth original relaxation feature to the VD sequence and the lifetime sequence;

[0077] ReVD sequence is added to the ReVD sequence and the lifetime sequence to obtain continuity coefficient ; k represents the index of the original relaxation feature, represents the continuity coefficient obtained by adding the ReVD in the kth original relaxation feature to the ReVD sequence and the lifetime sequence;

[0078] tVD sequence is added to the tVD sequence and the lifetime sequence to obtain continuity coefficient ; k represents the index of the original relaxation feature, represents the continuity coefficient obtained by adding the tVD in the kth original relaxation feature to the tVD sequence and the lifetime sequence;

[0079] tReVD sequence is added to the tReVD sequence and the lifetime sequence to obtain continuity coefficient ; k represents the index of the original relaxation feature, represents the continuity coefficient obtained by adding the tReVD in the kth original relaxation feature to the tReVD sequence and the lifetime sequence;

[0080] Step 324: Calculate the sum R k of continuity coefficients for each original relaxation feature k.

[0081]

[0082] Step 325: Take the original relaxation feature k with the highest sum R k of continuity coefficients as the original relaxation feature with the highest continuity with the historical relaxation feature.

[0083] ​The application innovatively proposes an automatic screening mechanism based on multi-parameter continuity coefficient calculation (steps 321-325). By sequentially adding the newly collected original relaxation characteristics (according to types VC, VD, etc.) to the corresponding historical characteristic sequence, and synchronously adding the initial life interval corresponding to the original relaxation characteristics to form a life sequence, the continuity coefficient (quantitative representation of the synchronization of the characteristic change trend and the life decay trend) between the characteristic sequence and the life sequence after each new characteristic is added is calculated. Finally, for each candidate original relaxation characteristic (corresponding to a group of k times of collection), the sum Rk of the continuity coefficients of all types of relaxation characteristics (VC, VD, ReVD, tVD, tReVD) of the candidate original relaxation characteristic is calculated, and the characteristic k with the highest Rk is selected. This method: accurately measures the matching degree of the new characteristic and the historical decay trajectory using the continuity coefficient (such as the correlation index), avoiding subjectivity. By accumulating the continuity coefficients of five key relaxation characteristics (VC, VD, ReVD, tVD, tReVD), the consistency of the new data segment with the life decay trend in multiple dimensions is comprehensively evaluated, and the screening result is more representative. Based on the criterion of maximizing Rk, the original relaxation characteristic and its V-t curve that best match the historical aging process and best reflect the current battery remaining life state (i.e., "most representative") are automatically and efficiently identified. Therefore, the mechanism ensures that the V-t curve used for extracting the sparse feature set (such as pulse factor, waveform factor, etc.) in the subsequent process has higher representation and reliability, provides a better input data basis for the refined life prediction model, and significantly improves the accuracy of the final life prediction.

[0084] Further, the number of collected historical relaxation characteristics is c, and the number of collected original relaxation characteristics is d; 5d

[0085] In the technical solution provided by the application, by controlling the number of historical relaxation characteristics and the number of original relaxation characteristics, the number of historical relaxation characteristics is prevented from being too large, which makes the change of the newly added original relaxation characteristics not obvious, and thus the original relaxation characteristics with low continuity cannot be accurately found.

[0086] Further, the continuity coefficient is calculated according to the following formula:

[0087] ; wherein X i represents the value of the i-th element in the VC sequence, represents the average value of the VC sequence; represents the average value of the life sequence, represents the value of the i-th element in the life sequence;

[0088] The continuity coefficient , , , , The calculation manner is the same.

[0089] In the technical solution provided in the application, the continuity coefficient is the correlation between two number sequences, and it can be foreseen that if the correlation between two number sequences is high, the inertia of data is relatively strong, the continuity is high, and the credibility of the original relaxation feature inserted into the number sequence is high.

[0090] Further, step 4 comprises the following steps:

[0091] Step 41: acquiring a sparse feature set, aligning each feature in the sparse feature set with each other according to time, so as to expand the sparse feature set into a two-dimensional matrix, the rows in the two-dimensional matrix being the types of the sparse feature set, and the columns of the two-dimensional matrix being time;

[0092] Step 42: adding data values of the sparse feature set in the two-dimensional matrix to form a third dimension, so as to convert the two-dimensional matrix into three-dimensional information;

[0093] Step 43: performing data enhancement on the three-dimensional information to obtain a key feature matrix.

[0094] In the technical solution provided in the application, the mutually isolated sequences are converted into corresponding elements by aligning the sparse feature set according to time.

[0095] Further, the neural network model comprises: an input layer configured to receive the key feature matrix R;

[0096] a convolution layer configured to perform convolution operation on the key feature matrix R to generate a feature map;

[0097] a BN layer configured to perform standardization processing on the feature map;

[0098] an activation layer configured to activate the standardized feature;

[0099] a pooling layer configured to perform maximum pooling on the output of the activation layer;

[0100] a connection layer configured to connect the output of the pooling layer;

[0101] an output layer configured to output the accurate life information corresponding to the key feature.

[0102] In the technical solution provided in the application, the neural network model constructed can capture the key information in the key feature matrix R through convolution, so as to accurately identify the corresponding relationship between the key feature matrix R and the feature map.

[0103] The technical solution of the embodiment of the application has at least the following advantages and beneficial effects:

[0104] Precision improvement: Firstly, the preliminary life interval is estimated by the basic cycle features (step 1-2), and then the key feature matrix formed by the high-dimensional relaxation features (step 3-4) is introduced to refine the interval (step 5). This two-stage strategy of "rough screening + fine adjustment" effectively integrates information of different granularities, overcomes the limitations of insufficient prediction accuracy of single feature dimension, and significantly improves the accuracy of final life prediction.

[0105] Model complexity and overfitting control: Output dimension compression: The neural network model receives the preliminary life interval instead of the original, massive fine feature segmentation label, which significantly reduces the complexity of the model output layer.

[0106] Input dimension compression: Through relaxation feature screening based on continuity factors (step 3) and data augmentation (step 4), only the key feature matrix extracted from the recent, highly relevant V-t curve needs to be input, instead of the lengthy historical relaxation data sequence, effectively reducing the complexity of the model input layer.

[0107] Model stability enhancement: The synchronous reduction of input and output dimensions significantly alleviates the information processing pressure within the neural network model. This not only reduces the model structure redundancy, but also effectively suppresses the risk of gradient instability (gradient explosion / disappearance) and overfitting phenomenon during the training process caused by high-dimensional features and complex output space, improving the model's generalization ability and training stability. BRIEF DESCRIPTION OF DRAWINGS

[0108] Figure 1 Flowchart of the method for evaluating the state of health of energy storage batteries based on a neural network model.

[0109] Figure 2 Time curve of the charging voltage reaching the peak value.

[0110] Figure 3 Time curve of the discharging voltage reaching the peak value.

[0111] Figure 4 Cycle number curve of peak sequence denoising.

[0112] Figure 5 Frequency graph of the pulse factor, with frequency on the horizontal axis and pulse factor value on the vertical axis.

[0113] Figure 6 Frequency graph of the waveform factor, with frequency on the horizontal axis and waveform factor value on the vertical axis.

[0114] Figure 7 Frequency graph of the peak factor, with frequency on the horizontal axis and peak factor value on the vertical axis.

[0115] Figure 8The horizontal coordinate is frequency, and the vertical coordinate is the wavelet energy ratio value.

[0116] Figure 9 The horizontal coordinate is frequency, and the vertical coordinate is the wavelet energy entropy value. DETAILED DESCRIPTION

[0117] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described clearly and completely in combination with specific embodiments. The same reference signs in the drawings represent the same components. It should be noted that the described embodiments are part of the embodiments of the present application, not all. Based on the described embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0118] Compared with the embodiments shown in the drawings, the feasible implementation solutions within the scope of protection of the present application can have fewer components, other components not shown in the drawings, different components, differently arranged components or differently connected components, etc. In addition, two or more components in the drawings can be implemented in a single component, or a single component shown in the drawings can be implemented as a plurality of separate components.

[0119] Unless otherwise defined, the technical terms or scientific terms used herein should be understood as their usual meanings by those of ordinary skill in the art to which the present application belongs. The use of "first", "second" and similar words in the specification and claims of the present application does not represent any order, quantity or importance, but is only used to distinguish different components. Similarly, "one" or "a" and the like do not necessarily represent a quantity limitation. "Up", "down" and the like are only used to represent relative positional relationships, which may change accordingly when the absolute position of the described object changes.

[0120] Reference Figure 1 , Embodiment 1: Energy storage battery health state evaluation method based on neural network model, characterized in that, comprising:

[0121] Step 1: Obtain the cycle characteristics of the energy storage battery, and pre-process the cycle characteristics to generate initial life characteristics; the cycle characteristics include: cycle number, time when the charging voltage reaches the peak value, time when the discharging voltage reaches the peak value, time when the charging current reaches the peak value.

[0122] Step 1 includes the following steps:

[0123] Step 11: Obtain the model of the energy storage battery, and obtain the charging voltage peak value, discharging voltage peak value and charging current peak value of the energy storage battery according to the model of the energy storage battery.

[0124] Each model has different peak charging voltage, peak discharging voltage, and peak charging current, so these need to be preset according to the battery model.

[0125] Step 12: Obtain historical charge and discharge data of the energy storage battery. Based on the peak charging voltage, peak discharging voltage, and peak charging current, obtain the time when the charging voltage reaches its peak value, the time when the discharging voltage reaches its peak value, and the time when the charging current reaches its peak value for each charge and discharge cycle.

[0126] like Figure 2 As shown, Figure 2 The time (F1) when the charging voltage reaches its peak value is shown. Figure 3 The time it takes for the discharge voltage to reach its peak value is shown. Figure 3 The peak value in the equation is the first defined minimum voltage value.

[0127] Historical charge and discharge data of a battery refers to the current and voltage information during the discharge and charging processes. Obtaining historical charge and discharge data can yield peak charging voltage, peak discharging voltage, and peak charging current.

[0128] Step 13: Extract the number of cycles from historical charge / discharge data;

[0129] Step 14: Use the time when the charging voltage reaches its peak value, the time when the discharging voltage reaches its peak value, the time when the charging current reaches its peak value, and the number of cycles as cycle characteristics for each charge and discharge cycle.

[0130] Step 14 includes the following steps:

[0131] Step 141: Arrange the times when the charging voltage reaches its peak value in each cycle to obtain the charging voltage peak value sequence;

[0132] Arrange the times when the discharge voltage reaches its peak value in each cycle sequentially to obtain the discharge voltage peak value sequence;

[0133] Arrange the times when the charging current reaches its peak value in each cycle sequentially to obtain the charging current peak value sequence;

[0134] Step 142: Perform noise reduction processing on the peak sequence of charging voltage, peak sequence of discharging voltage, and peak sequence of charging current in sequence.

[0135] The noise reduction method in step 142 is as follows:

[0136] Step 1421: Preset the window length;

[0137] Step 1422: For each data point x within the window, calculate whether x∈(u-2σ, u+2σ), where u represents the mean of the window and σ represents the standard deviation of the window;

[0138] If x satisfies x∈(u-2σ, u+2σ), the data point x is normal data, if x (u-2σ, u+2σ), the data point x is noise data.

[0139] Step 1423: sequentially sliding window to remove all noise data, to complete the noise reduction processing of all peak value sequences.

[0140] Step 143: taking the noise-reduced charging voltage peak value sequence, discharging voltage peak value sequence, charging current peak value sequence and cycle number as cycle characteristics.

[0141] As shown in Figure 4 , after preprocessing, some outliers can be removed. Figure 4

[0142] Step 2: performing regression classification on the cycle characteristics based on a support vector machine to generate an initial life interval.

[0143] The initial life interval is a larger life interval set in advance. For example, the accurate life information is set to 0-1000 weeks; the accurate life information needs to be accurate to weeks, and the error cannot exceed two weeks. Therefore, the span of the initial life interval can be set to 50, so that 0-1000 can be divided into 20 initial life intervals. The length of each initial life interval is 20 weeks.

[0144] Step 2 includes the following steps:

[0145] Step 21: pre-acquiring a plurality of initial life interval labels, the initial life intervals being continuous with each other and not overlapping with each other; collecting a plurality of corresponding cycle characteristics for each initial life interval label, taking the collected cycle characteristics and the corresponding life interval label as one sample; collecting N samples to generate a training data set S;

[0146] ; x i represents the cycle characteristics of the i-th sample, i represents the index of the sample, y i represents the life interval label of the i-th sample, and all labels are extracted to obtain a label set Kt; ; the elements in the label set are different life interval labels;

[0147] Step 22: constructing a kernel function ;

[0148] ; j represents the index of the sample, x j represents the cycle characteristics of the j-th sample, represents a kernel function bandwidth parameter; ​

[0149] Step 23: Train binary SVM classifiers for each class, define the primal optimization problem of SVM classification for each SVM classifier and set the optimization conditions;

[0150] The primal optimization problem is:

[0151] ;

[0152] w represents a weight vector, b represents an offset, C represents a penalty factor, represents a slack variable;

[0153] The constraint condition is:

[0154] ;

[0155] represents a high-dimensional mapping;

[0156] Step 24: Convert the primal problem into a dual problem using the Lagrange multiplier method;

[0157] ;

[0158] The constraint condition is:

[0159] ;

[0160] represents a Lagrange multiplier;

[0161] Step 25: Construct a decision function according to the solution;

[0162] ; m represents the index of the SVM classifier;

[0163] represents a temporary binary label of the mth classifier;

[0164] represents a Lagrange multiplier of the mth classifier;

[0165] represents an offset of the mth classifier;

[0166] represents the classification result of the mth classifier on the sample x;

[0167] Step 26: Input the collected cycle features into the trained SVM classifier to obtain the classification results of all SVM classifiers, and take the label with the maximum decision value as the output.

[0168] Thus, after the SVM classifier is classified by the scheme provided in step 2, an initial life interval can be obtained more accurately.

[0169] Step 3: Collecting original relaxation characteristics of the energy storage battery for several adjacent times, screening the original relaxation characteristics with the highest correlation with the historical relaxation characteristics from the several original relaxation characteristics based on a continuity factor of the original relaxation characteristics and the historical relaxation characteristics, and extracting a sparse feature set from a V-t curve corresponding to the original relaxation characteristics; the sparse feature set includes a pulse factor, a waveform factor, a peak value factor, a wavelet energy ratio, a wavelet energy entropy, and a root mean square value.

[0170] The "relaxation characteristic" refers to a process in which the terminal voltage (or current) of the energy storage battery slowly recovers to an equilibrium state (open circuit voltage) after experiencing a rapid charge-discharge pulse. This process reflects physical and chemical phenomena such as charge redistribution, ion diffusion, and interface reaction inside the device. During the working process of the energy storage battery, multiple relaxation characteristics can be collected, but these relaxation characteristics may not all accurately reflect the current state of the energy storage battery. Therefore, the essence of step 3 is to further screen the single life information that can reflect the current state of the energy storage battery. Specifically:

[0171] Step 3 includes the following steps:

[0172] Step 31: Collecting original relaxation characteristics of the energy storage battery for several adjacent times, the original relaxation characteristics including VC, VD, ReVD, tVD, and tReVD; wherein VC represents the voltage at the beginning of the charge relaxation; VD represents the voltage drop amplitude at the initial stage of relaxation; ReVD represents the voltage value at the end of the relaxation process; tVD represents the time required for the voltage drop to reach a preset amplitude; and tReVD represents the time required for the voltage drop to approach ReVD.

[0173] Step 32: performing continuity analysis on the collected original relaxation characteristics and historical relaxation characteristics of the energy storage battery, and screening the original relaxation characteristics with the highest continuity with the historical relaxation characteristics from the several original relaxation characteristics.

[0174] The number of collected historical relaxation characteristics is c, and the number of collected original relaxation characteristics is d; 5d < c < 10d.

[0175] Step 32 includes the following steps:

[0176] Step 321: obtaining an initial life interval calculated by the energy storage battery when collecting the original relaxation characteristics;

[0177] obtaining an initial life interval corresponding to the historical relaxation characteristics;

[0178] Step 322: arrange the VCs in the historical relaxation feature and the original relaxation feature in time sequence to generate a VC sequence;

[0179] arrange the VDs in the historical relaxation feature in time sequence to generate a VD sequence;

[0180] arrange the ReVDs in the historical relaxation feature in time sequence to generate a ReVD sequence;

[0181] arrange the tVDs in the historical relaxation feature in time sequence to generate a tVD sequence;

[0182] arrange the tReVDs in the historical relaxation feature in time sequence to generate a tReVD sequence;

[0183] arrange the initial life intervals corresponding to the historical relaxation features in time sequence to generate a life sequence;

[0184] Step 323: sequentially add the VCs corresponding to adjacent original relaxation features in the VC sequence, add the initial life interval corresponding to the original relaxation feature to the life sequence, and obtain the continuity coefficient calculated from the VC sequence and the life sequence ; k represents the index of the original relaxation feature, represents the continuity coefficient calculated from the VCs in the kth original relaxation feature added to the VC sequence and the life sequence;

[0185] sequentially add the VDs corresponding to adjacent original relaxation features in the VD sequence, add the initial life interval corresponding to the original relaxation feature to the life sequence, and obtain the continuity coefficient calculated from the VD sequence and the life sequence ; k represents the index of the original relaxation feature, represents the continuity coefficient calculated from the VDs in the kth original relaxation feature added to the VD sequence and the life sequence;

[0186] sequentially add the ReVDs corresponding to adjacent original relaxation features in the ReVD sequence, add the initial life interval corresponding to the original relaxation feature to the life sequence, and obtain the continuity coefficient calculated from the ReVD sequence and the life sequence ; k represents the index of the original relaxation feature, represents the continuity coefficient calculated from the ReVDs in the kth original relaxation feature added to the ReVD sequence and the life sequence;

[0187] sequentially add the tVDs corresponding to adjacent original relaxation features in the tVD sequence, add the initial life interval corresponding to the original relaxation feature to the life sequence, and obtain the continuity coefficient calculated from the tVD sequence and the life sequence ; k represents the index of the original relaxation feature, represents the continuity coefficient calculated by adding the tVD in the kth original relaxation feature to the tVD sequence and the life sequence;

[0188] In the tReVD sequence, the tReVD corresponding to the adjacent original relaxation features is sequentially increased, the life sequence adds the initial life interval corresponding to the original relaxation feature, and the continuity coefficient calculated by the VC sequence and the life sequence is ; k represents the index of the original relaxation feature, represents the continuity coefficient calculated by adding the tReVD in the kth original relaxation feature to the tReVD sequence and the life sequence.

[0189] For example, in the historical relaxation feature, the VD sequence is: 1, 2, 3, 4. The life sequence (the number of life intervals) is 4, 5, 6, 7, and the VD of a certain original relaxation feature generated at present is 5, and the life interval calculated by step 2 is 6.

[0190] Then the VD sequence: 1, 2, 3, 4, 5 and the life sequence 4, 5, 6, 7, 6 are placed together for continuity calculation. The calculation method is as follows:

[0191] The continuity coefficient is calculated as follows:

[0192] ; wherein X i represents the value of the i th element in the VC sequence, represents the average value of the VC sequence; represents the average value of the life sequence, represents the value of the i th element in the life sequence;

[0193] The continuity coefficient , , , , is calculated in the same way.

[0194] Step 324: calculate the sum R k of the continuity coefficients for each original relaxation feature k;

[0195] ;

[0196] Step 325: the original relaxation feature k with the highest sum R k of the continuity coefficients is taken as the original relaxation feature with the highest continuity with the historical relaxation feature.

[0197] In combination with the foregoing, for each of the several original relaxation features generated most recently, an addition can be made to the historical relaxation features for continuity analysis, so as to select the closest one of the original relaxation features.

[0198] Step 33: Extracting a sparse feature set from the V-t curve corresponding to the original relaxation feature. The sparse feature set includes pulse factor, waveform factor, peak factor, wavelet energy ratio, wavelet energy entropy, and root mean square value.

[0199] This V-t curve is the relationship between current and voltage during the relaxation process, and each element in the sparse feature set can be obtained by wavelet transform, fuzzy entropy calculation, etc. from the graph. In this scheme, the pulse factor, waveform factor, peak factor, wavelet energy ratio, wavelet energy entropy, and root mean square value are several important features extracted from the V-t curve. In practice, energy entropy and fuzzy entropy can be further increased, or some features can be reduced. The key of the present application is to find an accurate V-t curve.

[0200] As shown in Figures 5-9 , it can be seen that the feature graphs of each element in the sparse feature set.

[0201] Step 4: Expanding the sparse feature set to obtain three-dimensional information;

[0202] Step 4 includes the following steps:

[0203] Step 41: Obtain the sparse feature set, and align each feature in the sparse feature set with each other according to time, so as to expand the sparse feature set into a two-dimensional matrix. The rows in the two-dimensional matrix are the types of the sparse feature set, and the columns of the two-dimensional matrix are time.

[0204] Step 42: Adding data values of the sparse feature set in the two-dimensional matrix to form a third dimension, so as to convert the two-dimensional matrix into three-dimensional information.

[0205] Step 4 is actually to expand multiple curves onto a two-dimensional plane to form a two-dimensional image information matrix with channel information.

[0206] Step 5: Inputting the key feature matrix and the initial life interval into the neural network model to generate accurate life information of the energy storage battery.

[0207] After the processing of steps 3 and 4, the three-dimensional information with the highest correlation to the accurate life of the battery is extracted. The three-dimensional information is used as input, and the accurate life information is used as output to create a sufficient number of data samples. Then the neural network model is trained using the data samples, and then the trained neural network model is used to predict the accurate life information.

[0208] The neural network model can be arbitrarily selected according to the computing resources, and the neural network model provided in the application is as follows:

[0209] Further, the neural network model comprises:

[0210] The input layer is used for receiving the key feature matrix R.

[0211] The convolution layer is used for extracting the key feature matrix R to perform a convolution operation to generate a feature map.

[0212] The BN layer is used for performing standardization processing on the feature map.

[0213] The activation layer is used for activating the standardized feature.

[0214] The pooling layer performs maximum pooling on the output of the activation layer.

[0215] The connection layer connects the output of the pooling layer.

[0216] The output layer outputs the accurate life information corresponding to the key feature.

[0217] The neural network model in the scheme mainly performs standardization processing on the feature, and the standardization processing is mainly standardization of the data. The standardization formula is:

[0218] ;

[0219] Q represents the feature map of the i th data sample, is the mean value of each channel, is the variance of each channel.

[0220] In practice, other neural network model structures can also be selected for training to obtain accurate accurate life information.

[0221] The above is only a preferred embodiment of the application and is not used to limit the application. For those skilled in the art, the application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. A method for evaluating the state of health of an energy storage battery based on a neural network model, characterized by, The application relates to a method for predicting the life of a storage battery, and belongs to the technical field of battery life prediction. The method comprises the following steps: Step 1: obtaining cycle characteristics of the storage battery, and preprocessing the cycle characteristics to generate initial life characteristics; The cycle characteristics comprise cycle times, time when the charging voltage reaches a peak value, time when the discharging voltage reaches a peak value, and time when the charging current reaches a peak value; Step 2: performing regression classification on the cycle characteristics based on a support vector machine to generate initial life intervals; Step 3: collecting original relaxation characteristics of the storage battery in adjacent times, screening, from the original relaxation characteristics, an original relaxation characteristic with the highest correlation with historical relaxation characteristics based on a continuity factor of the original relaxation characteristics and the historical relaxation characteristics, and extracting a sparse feature set from a V-t curve corresponding to the original relaxation characteristic with the highest correlation; The sparse feature set comprises a pulse factor, a waveform factor, a peak value factor, a wavelet energy ratio, a wavelet energy entropy and a root mean square value; Step 3 comprises the following steps: Step 31: collecting original relaxation characteristics of the storage battery in adjacent times, wherein the original relaxation characteristics comprise VC, VD, ReVD, tVD and tReVD; wherein VC represents a voltage at a charging relaxation starting moment, VD represents a voltage drop amplitude in an initial relaxation stage, ReVD represents a voltage value at an end of a relaxation process, tVD represents a time required for voltage drop to reach a preset amplitude, and tReVD represents a time required for voltage drop to approach ReVD; Step 32: performing continuity analysis on the collected original relaxation characteristics and historical relaxation characteristics of the storage battery, and screening, from the original relaxation characteristics, an original relaxation characteristic with the highest continuity with the historical relaxation characteristics; Step 33: extracting a sparse feature set from a V-t curve corresponding to the original relaxation characteristic with the highest continuity; Step 32 comprises the following steps: Step 321: obtaining an initial life interval calculated by the storage battery when the original relaxation characteristics are collected; obtaining an initial life interval corresponding to the historical relaxation characteristics; Step 322: arranging the historical relaxation characteristic VC in a time sequence to generate a VC sequence; arranging the historical relaxation characteristic VD in a time sequence to generate a VD sequence; arranging the historical relaxation characteristic ReVD in a time sequence to generate a ReVD sequence; arranging the historical relaxation characteristic tVD in a time sequence to generate a tVD sequence; arranging the historical relaxation characteristic tReVD in a time sequence to generate a tReVD sequence; Step 323: sequentially increase the VCs corresponding to the adjacent several times of the original relaxation characteristics in the VC sequence, add the initial lifetime interval corresponding to the original relaxation characteristics to the lifetime sequence, and calculate the continuity coefficient of the VC sequence and the lifetime sequence to obtain ; k represents the index of the original relaxation characteristics, represents the continuity coefficient obtained by adding the VC in the kth original relaxation characteristics to the VC sequence and calculating with the lifetime sequence. In the VD sequence, the VD corresponding to the original relaxation feature of the adjacent several times is sequentially increased, the initial life interval corresponding to the original relaxation feature is added to the life sequence, and the continuity coefficient is obtained by calculating the VD sequence and the life sequence ; k represents the index of the original relaxation feature, indicates that the VD in the kth original relaxation feature is added to the continuity coefficient obtained by calculating the VD sequence and the life sequence. In the ReVD sequence, the ReVD corresponding to the original relaxation feature of each adjacent order is sequentially added, the initial life interval corresponding to the original relaxation feature is added to the life sequence, and the continuity coefficient is calculated by the ReVD sequence and the life sequence to obtain ; k represents the index of the original relaxation feature, indicates that the ReVD in the kth original relaxation feature is added to the ReVD sequence, and the continuity coefficient is calculated by the life sequence. In the tVD sequence, the tVD corresponding to the original relaxation feature of the adjacent several times is sequentially increased, the initial life interval corresponding to the original relaxation feature of the time is added to the life sequence, and the continuity coefficient is obtained by calculating the tVD sequence and the life sequence ; k represents the index of the original relaxation feature, indicates that the tVD in the kth original relaxation feature is added to the tVD sequence, and the continuity coefficient is obtained by calculating the tVD sequence and the life sequence. In the tReVD sequence, the tReVD corresponding to the original relaxation feature of a number of adjacent times is sequentially increased, the initial life interval corresponding to the original relaxation feature is added to the life sequence, and the continuity coefficient is calculated by the tReVD sequence and the life sequence to obtain ; k represents the index of the original relaxation feature, indicates the continuity coefficient obtained by adding the tReVD in the kth original relaxation feature to the tReVD sequence and the life sequence. Step 324: Calculate the sum of continuity coefficients R for each original relaxation feature k k ; ； Step 325: Summing the continuity coefficients R k the highest original relaxation feature k as the original relaxation feature with the highest continuity with the historical relaxation features; arranging the initial life interval corresponding to the historical relaxation characteristics in a time sequence to generate a life sequence; Step 4: expanding the sparse feature set to obtain three-dimensional information; Step 4 comprises the following steps: Step 41: obtaining the sparse feature set, aligning each feature in the sparse feature set in a time sequence, expanding the sparse feature set into a two-dimensional matrix, and taking rows in the two-dimensional matrix as the types of the sparse feature set and taking columns in the two-dimensional matrix as time; Step 42: adding data values of the sparse feature set in the two-dimensional matrix to form a third dimension, and converting the two-dimensional matrix into three-dimensional information; Step 43: performing data enhancement on the three-dimensional information to obtain a key feature matrix. Step 5: input the key feature matrix and the initial life interval into the neural network model to generate accurate life information of the energy storage battery.

2. The method of claim 1, wherein the neural network model is trained based on a plurality of training data, and each of the plurality of training data comprises a plurality of features and a corresponding label. Step 1 includes the following steps: Step 11: obtain the model of the energy storage battery, and obtain the charging voltage peak value, discharging voltage peak value and charging current peak value of the energy storage battery according to the model of the energy storage battery; Step 12: obtain the historical charging and discharging data of the energy storage battery, and obtain the time when the charging voltage reaches the peak value, the time when the discharging voltage reaches the peak value and the time when the charging current reaches the peak value during each charging and discharging according to the charging voltage peak value, the discharging voltage peak value and the charging current peak value; Step 13: extract the cycle number from the historical charging and discharging data; Step 14: take the time when the charging voltage reaches the peak value, the time when the discharging voltage reaches the peak value, the time when the charging current reaches the peak value and the cycle number as the cycle feature.

3. The method for assessing the health status of energy storage batteries based on a neural network model according to claim 2, characterized in that, Step 14 includes the following steps: Step 141: arrange the time when the charging voltage reaches the peak value in each cycle in sequence to obtain a charging voltage peak value sequence; arrange the time when the discharging voltage reaches the peak value in each cycle in sequence to obtain a discharging voltage peak value sequence; arrange the time when the charging current reaches the peak value in each cycle in sequence to obtain a charging current peak value sequence; Step 142: sequentially perform noise reduction processing on the charging voltage peak value sequence, the discharging voltage peak value sequence and the charging current peak value sequence; Step 143: take the noise-reduced charging voltage peak value sequence, the discharging voltage peak value sequence and the charging current peak value sequence and the cycle number as the cycle feature.

4. The method for assessing the health status of energy storage batteries based on a neural network model according to claim 3, characterized in that, The noise reduction method in step 142 is as follows: Step 1421: pre-set a window length; Step 1422: for each data point x in the window, calculate whether x ∈ (u-2σ, u+2σ) is satisfied, wherein u represents the mean value of the window and σ represents the standard deviation of the window; If x satisfies x e (u - 2σ, u + 2σ), the data point x is normal data, and if x (u - 2σ, u + 2σ), the data point x is noise data. Step 1423: sequentially slide the window to remove all noise data to complete the noise reduction processing of all peak value sequences.

5. The method for assessing the health status of energy storage batteries based on a neural network model according to claim 3, characterized in that, Step 2 includes the following steps: Step 21: pre-acquire a plurality of initial life interval labels, the initial life intervals are continuous with each other and do not overlap with each other; collect a plurality of corresponding cycle features for each initial life interval label, take the collected cycle features and the corresponding life interval label as one sample; collect N samples to generate a training data set S; ; x i represents the cycle feature of the i-th sample, i represents the index of the sample, y i represents the life interval label of the i-th sample, and all labels are extracted to obtain a label set Kt; ; the elements in the label set are different life interval labels; Step 22: Constructing the kernel function ; ; j denotes the index of the sample, x j denotes the cyclic feature of the jth sample, denotes the kernel function bandwidth parameter; Step 23: train a binary SVM classifier for each category, and define an original optimization problem of the SVM classification and set optimization conditions for each SVM classifier; The original optimization problem is: ； w denotes a weight vector, b denotes an offset, C denotes a penalty factor, denotes a slack variable; The constraint condition is: ； represents a high-dimensional mapping; Step 24: use the Lagrange multiplier method to convert the original problem into a dual problem; ； The constraint condition is: ； denotes the Lagrange multiplier; Step 25: construct a decision function according to the solution; ; m denotes the index of the SVM classifier; a temporary binary label representing the mth classifier; λm represents the Lagrange multiplier of the mth classifier; represents an offset of the mth classifier; represents the classification result of the mth classifier on the sample x; Step 26: input the collected cycle features into the trained SVM classifier to obtain the classification results of all SVM classifiers, and take the label with the maximum decision value as the output.

6. The method of claim 1, wherein the neural network model is trained using a plurality of training data sets, each of the training data sets including a plurality of input data and a corresponding output data. Continuity coefficient The calculation formula is as follows: ; where X i denotes the value of the i-th element of the VC sequence, denotes the average value of the VC sequence; denotes the average value of the lifetime sequence, denotes the value of the i-th element of the lifetime sequence.

7. The method of claim 1, wherein the neural network model is trained using a dataset comprising a plurality of energy storage battery health states and a plurality of corresponding features. The neural network model includes: The input layer is used to receive the key feature matrix R; The convolution layer is used to extract the key feature matrix R for convolution operation to generate a feature map; The BN layer is used to perform standardization processing on the feature map; The activation layer is used to activate the standardized feature. The pooling layer performs maximum pooling on the output of the activation layer; The connection layer connects the output of the pooling layer; The output layer outputs the accurate life information corresponding to the key features.

Images