Lithium ion battery safety state estimation method based on working condition fragment characteristics and unsupervised clustering

Abstract

The invention discloses a lithium ion battery safety state estimation method based on working condition fragment features and unsupervised clustering, and the method comprises the steps: achieving the segmentation of different working condition data based on K-means based on an MIT-Stanford mixed working condition cycle data set; for the segmentation data, segment safety features and global safety features of the constant current charging stage, the constant voltage charging stage and the constant current discharging stage are extracted respectively; performing normalization, direction unification, dimension reduction and correlation screening processing on a data set formed by feature extraction in sequence, and combining to generate a plurality of feature data sets; aiming at a plurality of security feature sets, adopting a PSO algorithm to optimize FCM clustering center selection, introducing GG distance measurement to replace Euclidean distance, and establishing an SOS estimation method based on particle swarm optimization-fuzzy clustering; and independently clustering each feature set, giving a 0 / 1 score according to different clustering center results, and finally aggregating a good reputation rate as an SOS estimated value. According to the method, the dependence of SOS estimation on the parameter threshold is overcome by combining the working condition characteristics and the unsupervised clustering scoring mechanism, and a new method is provided for the actual use of SOS estimation.

Description

Technical Field

[0001] The present invention relates to the field of lithium-ion battery safety state estimation, and in particular to a lithium-ion battery safety state estimation method based on operating condition segment features and unsupervised clustering. Background Art

[0002] Lithium-ion batteries are widely used in the field of electric vehicles due to their high power density, high energy density, long life, and good cycle performance. However, lithium-ion batteries themselves have safety risks that cannot be ignored. As indicators such as energy density increase, the safety issues of lithium-ion batteries have become more acute. In recent years, while various types of high-energy-density lithium-ion batteries have been put on the market, lithium-ion battery-related safety accidents have occurred frequently at home and abroad. According to data released by the China New Energy Vehicle Evaluation Procedure (CEVE), since the beginning of 2019, there have been more than 40 electric vehicle safety accidents related to power batteries at home and abroad, which have caused serious loss of life and property. Therefore, how to achieve safe operation of lithium-ion batteries throughout their life cycle has become a major demand in the industry.

[0003] Extensive research has been conducted on lithium-ion battery safety, and the threshold method is relatively widely used in industrial scenarios. This method assesses battery safety by defining a safety threshold or a safety zone for a specific battery attribute and determining whether the battery's current attribute is within the threshold or the safety zone. This method simply categorizes a battery's safety status as safe or unsafe and is currently a common safety diagnostic method in BMSs. In addition to the threshold method, Cabrera formally introduced the concept of a safe state (SOS). The SOS is expressed similarly to other conditions such as the battery's state of health (SOH), quantifying battery safety using a numerical value between 0 and 1. The SOS uses the inverse of various abuse functions that affect battery safety performance as subfunctions, such as the battery's temperature, voltage, current, internal resistance, state of charge (SOC), SOH, and mechanical condition, to model and assign values ​​to the battery's SOS.

[0004] However, the research on the threshold method for lithium-ion battery safety is only based on the classification of various battery runaway phenomena, and a mathematical model that fully characterizes battery safety has not yet been proposed. The SOS abuse model requires the construction of abuse functions with different attributes to build a complete SOS model, which still faces certain difficulties in practical applications. The construction of abuse functions with multiple characterization quantities requires the safety thresholds of specific battery models under different attributes to be clarified in advance, which requires tedious testing for different models in actual use. The real-time sensor measurement of multiple attributes throughout the life cycle also creates an excessive data processing burden for the BMS.

[0005] In response to the above problems, a practical SOS estimation method is urgently needed to avoid a large number of mathematical modeling processes for SOS, so as to achieve real-time SOS estimation in different scenarios. The existing technology lacks the integration technology of safety features. Most methods need to list different safety features and discuss them independently. They cannot collaboratively explore the dynamic evolution laws between various features and between them and SOS, which limits the progress of SOS in actual references. At the same time, batteries face different working conditions and environments during actual operation. How to extract highly applicable safety features under complex conditions to ensure that the battery SOS can be evaluated in various environments. Therefore, the development of a lithium-ion battery safety state estimation method based on working condition fragment features and unsupervised clustering has become a key research direction to promote the intelligent upgrade of battery safety management. Summary of the Invention

[0006] The present invention aims to overcome the above-mentioned shortcomings of the prior art and proposes a lithium-ion battery safety state estimation method based on operating condition segment features and unsupervised clustering. It also proposes a systematic multi-source feature screening framework, combined with an improved unsupervised clustering model, to improve the accuracy and robustness of lithium-ion battery safety state SOS estimation.

[0007] In order to achieve the above object, the technical solution of the present invention is:

[0008] A method for estimating the safety state of a lithium-ion battery based on operating condition segment features and unsupervised clustering comprises the following steps:

[0009] S1: Based on the MIT-Stanford lithium-ion battery mixed operating cycle dataset, a K-means-based segmentation method is proposed to segment the data under different operating conditions according to the different voltage and current characteristics of the mixed operating conditions.

[0010] S2: Screen the extracted data for security features during constant current (CC) charging, constant voltage (CV) charging, and constant current discharging phases, and introduce global security features.

[0011] S3: Normalize, unify directions, reduce dimensions, and perform correlation screening on the extracted security features. Group the features based on the correlation analysis results to generate a feature dataset for clustering.

[0012] S4: PSO is used to optimize the cluster centers, and the GG distance metric is combined to replace the Euclidean distance. The scoring mechanism is used to assign points to the cluster center, and the SOS estimation of lithium-ion batteries is achieved by calculating the praise rate.

[0013] S5: Use clustering quality indicators, life correlation verification and robustness testing to output the evaluation results of the lithium-ion battery safety status SOS.

[0014] Furthermore, in step S1, the specific contents of the working condition segmentation method constructed are as follows:

[0015] The working condition segmentation method for mixed working condition data is divided into feature construction and segmentation parts under different working conditions.

[0016] S1-1: Based on the collected battery data, the static phase data is screened out and the average values ​​of current, voltage and SOC are calculated. At the same time, a parameterized global change point detection method (Findchangepts) is used to detect the mutation points of current, voltage and SOC.

[0017] Based on the preliminary segmentation results of Findchangepts for current, voltage, and SOC, we assigned a uniform value within the range [0.1, 1] to the segmented segments and multiplied them by the average values ​​of current, voltage, and SOC, respectively, to construct cluster features based on current, voltage, and SOC. We also added time as a feature and assigned it a higher weight to ensure the temporal consistency of the segmentation results. Therefore, the constructed feature matrix is:

[0018] features=[I mean *I abr SOC mean *SOC abr V mean *V abr time] (1)

[0019] Among them, I abr ,SOC abr ,V abr Represent the current, voltage and SOC partition assignment matrices respectively, I mean ,SOC mean ,V mean Represent the average values ​​of current, voltage and SOC respectively, and time represents time.

[0020] S1-2: Based on the preliminary segmentation results, set the number of K-means clusters:

[0021] k=min(max(C find )) (2)

[0022] Among them, C find Indicates the number of blocks divided after the initial segmentation of current, voltage and SOC.

[0023] K-means clustering is performed based on the number of clusters. Based on the clustering results, the charge-discharge and CC / CV characteristics of the same partitioned data are determined based on the partition current size and current change rate. Finally, the voltage change rate is used to verify and output the results.

[0024] Furthermore, in step S2, the specific contents of the constructed security feature screening method are as follows:

[0025] The safety features used for SOS assessment of lithium-ion batteries are divided into charge and discharge segment data and global variables.

[0026] S2-1: The charge and discharge segment data is segmented based on the operating condition to extract safety features. In the constant current charge (CC) segment (1C, SOC from 80% to the cut-off voltage), the time required for SOC to rise by 10% is extracted. cc ; Extract the residual square sum S between the actual current curve and the initial reference curve in the constant voltage charging (CV) segment SSE The mean absolute error E MAE , and its calculation formula is:

[0027]

[0028] Among them, I ref,i and I act,i are the reference and actual current values ​​respectively; N is the data length. ref and I act To have the same data length, align the two lengths through linear interpolation.

[0029] The incremental capacity (IC) curve is generated by the incremental capacity analysis (ICA) method in the constant current discharge segment. The core calculation formula is:

[0030]

[0031] Among them, Q k , Q k-1 Represent the capacity changes at time k and k-1, V B,k ,V B,k-1 They represent the battery terminal voltage values ​​at time k and k-1 respectively.

[0032] In order to ensure the integrity of the IC curve and avoid abnormal noise interference, the minimum sampling interval is set while performing linear interpolation on the collected voltage data to prevent dV B The noise step caused by too small a value is eliminated, and the Savitzky-Golay filtering method is introduced to reduce the noise of the IC curve.

[0033] S2-2: For the global variables, extract the parameter difference characteristics between the initial battery cycle and the cycles at different aging stages, and obtain the average temperature difference △avgT by segmenting the cycle nodes based on the MIT-Stanford dataset. i,j , capacity difference C ini ,i and internal resistance difference R ini,i, where ini represents the initial cycle, i and j represent the percentage position of the total cycle number (such as 10%, 20%).

[0034] This type of feature enhances the robustness to segmented feature noise by integrating the full-cycle temperature, capacity, and internal resistance evolution information, maintains reliability in the event of data anomalies, and jointly constructs a multi-dimensional safety status characterization system of time-voltage-temperature-capacity-internal resistance with the charge and discharge segment features.

[0035] Furthermore, in step S3, the steps for constructing a lithium-ion battery safety status dataset based on the extracted safety features are as follows:

[0036] S3-1: Normalize the extracted features and control their range within the range of [-1,1];

[0037] S3-2: The correlation between all features and capacity changes is analyzed using the Pearson correlation coefficient method (PCC). The calculation process is as follows.

[0038]

[0039] Among them, X is the data sequence of a single safety feature, Y is the capacity change sequence of the battery cell, and are the corresponding average values, and n is the sequence length. The features that are inversely proportional to the capacity change are reciprocated to ensure that all features are positively correlated with the capacity.

[0040] S3-3: Use principal component analysis (PCA) to reduce the dimension, sort and sum the information provided by the features, and select key features with a cumulative contribution rate of more than 90% to construct a new feature matrix.

[0041] S3-4: Based on the PCC analysis results between the filtered features, 3 features in each group (PCC ≤ 0.7) are selected for combination, and features are allowed to be reused to cover complete information. Finally, 107 feature groups are generated to provide input for PSO-GG clustering.

[0042] Furthermore, in step S4, the battery SOS estimation process based on the PSO-GG algorithm is proposed as follows:

[0043] S4-1: The SOS estimation process based on particle swarm optimization-fuzzy clustering (PSO-GG) uses the PSO algorithm to optimize the cluster center of a single iteration of FCM, and combines the GG distance measurement module to replace the Euclidean distance to capture the non-spherical cluster structure. Finally, the safety status of lithium-ion batteries is quantitatively evaluated through the scoring mechanism and the calculation of the praise rate.

[0044] S4-2: The specific process of using the PSO-GG-based SOS estimation framework to perform safety status assessment is as follows:

[0045] Set the core parameters of PSO-GG clustering, set the number of clusters c = 2, the fuzzy index m = 2, the convergence threshold ε = 1e-8, and the maximum number of iterations to 200;

[0046] The 107 generated feature sets are input into the PSO-GG framework for clustering. Each feature set is clustered independently. The safety of the battery on this feature set is distinguished based on the size of the cluster center value optimized by PSO, and a safety score of 0 / 1 is assigned. Finally, the positive rate of all subsets is calculated as the SOS value.

[0047] In the clustering process, the PSO optimization module uses the GG distance objective function to dynamically search for cluster centers. The objective function expression is:

[0048]

[0049] Among them, fit represents the objective function of the PSO algorithm, u k-1 represents the membership matrix of the k-1th iteration in the FCM algorithm iteration process, x j Represents the jth sample point in the sample data set. Each particle in the PSO represents a complete cluster center. By calculating the fit corresponding to each particle and iteratively updating the position of the particle (i.e., the candidate center), we continuously search for a combination of centers that makes the fit smaller.

[0050] In the clustering process, the GG distance metric module replaces the traditional Euclidean distance with the following formula:

[0051]

[0052] Among them, d ij 2 represents the square of the distance from sample j to the center of cluster i, det(F i ) represents the covariance matrix F i The determinant, P i represents the prior probability of cluster i, F i is the covariance matrix.

[0053] Furthermore, in step S5, the safety status SOS of the lithium-ion battery is evaluated as follows:

[0054] S5-1: In order to evaluate the ability of the PSO-GG model to estimate SOS, the compactness Cmp, separation Spt and their ratio Cmp / Spt are used to evaluate the clustering quality. The definitions of compactness and separation are:

[0055]

[0056] Spt=min i≠k||c i -c k || 2 (10)

[0057] Among them, Cmp reflects the compactness of the data after data partitioning, that is, the quantitative similarity of data within a class. Spt is a function of the center matrix, which reflects the separability of the data after data partitioning, that is, the dissimilarity between classes as much as possible.

[0058] S5-2: To evaluate the reliability of the algorithm under actual noise interference, 30% random noise (the noise range is ±30% of the characteristic value) is applied to the safety characteristics of 30 random batteries. The SOS estimation is re-performed on the noise-imposed data, and the SOS estimation values ​​before and after the noise is applied are compared to verify the robustness of the proposed SOS estimation method.

[0059] The present invention is based on the MIT-Stanford mixed working condition cycle data set, and realizes the segmentation of different working condition data based on K-means; for the segmented data, the segmented safety features and global safety features of the constant current charging, constant voltage charging, and constant current discharging stages are extracted respectively; the data set composed of the extracted features is normalized, directional unified, dimensionally reduced, and correlation screened in turn, and several feature data sets are generated in combination; for several safety feature sets, the PSO algorithm is used to optimize the FCM clustering center selection, and the GG distance metric is introduced to replace the Euclidean distance to establish an SOS estimation method based on particle swarm optimization-fuzzy clustering (PSO-GG); each feature set is clustered independently, and a 0 / 1 score is assigned according to the results of different cluster centers, and the final aggregated praise rate is used as the SOS estimation value. The present invention overcomes the dependence of SOS estimation on parameter thresholds by combining working condition characteristics with an unsupervised clustering scoring mechanism, and provides a new method for the actual use of SOS estimation.

[0060] The beneficial effects of the present invention are:

[0061] 1. A safety state estimation method for lithium-ion batteries based on operating condition fragment features and unsupervised clustering is proposed. Safety features are collected for fragment information under different operating conditions, and the feature set is screened and split. The filtered feature set is used to achieve SOS scoring through unsupervised clustering. While taking into account the complementarity of multi-source data, the accuracy and robustness of SOS estimation are significantly improved.

[0062] 2. For mixed operating condition data, a feature combining voltage, current and SOC mutation fragments and data information is constructed, and the accurate segmentation of battery charging and discharging conditions is achieved based on the K-means algorithm. In response to changes in battery safety performance, a safety feature collection method under different operating conditions based on different aging periods is proposed.

[0063] 3. A multidimensional safety characterization system was constructed that integrates charging and discharging segment features and global variables. The feature correlation direction was unified through Pearson correlation analysis. PCA dimensionality reduction and low-correlation feature grouping strategies were used to generate independent safety feature subsets while retaining a high information contribution rate. This not only covers complete feature information but also reduces data redundancy, thereby improving the algorithm's adaptability to non-spherical feature clusters and noise robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a schematic flow diagram of the present invention.

[0065] Figure 2 It is a schematic diagram of the preliminary segmentation result of the charge and discharge data in the present invention.

[0066] Figures 3(a) and 3(b) are the working condition segmentation result diagrams based on K-means in the present invention, wherein Figure 3(a) is the working condition segmentation current result diagram, and Figure 3(b) is the working condition clustering result diagram based on SOC, voltage, and current characteristics.

[0067] Figure 4 This is a comparison chart of the IC curve filtering effect in the present invention.

[0068] Figure 5 This is a characteristic information contribution rate analysis diagram in the present invention.

[0069] Figure 6 This is a graph showing the PCC analysis results of the feature dataset in the present invention.

[0070] Figure 7 This is the flow chart of the PSO iterative process in the present invention

[0071] Figure 8 This is a relationship diagram between the battery SOS and cycle life in the present invention.

[0072] Figure 9 This is a comparison chart of the noise robustness test results in the present invention. Specific implementation methods

[0073] The present invention will be further described below with reference to the accompanying drawings.

[0074] Reference Figure 1 As shown, a lithium-ion battery safety state estimation method based on operating condition segment features and unsupervised clustering includes the following steps:

[0075] S1: Based on the MIT-Stanford lithium-ion battery mixed operating cycle dataset, a K-means-based segmentation method is proposed to segment the data under different operating conditions according to the different voltage and current characteristics of the mixed operating conditions.

[0076] S2: Screen the data of the constant current (CC) charging, constant voltage (CV) charging, and constant current discharging stages obtained by extraction, and introduce global safety features at the same time;

[0077] S3: Perform normalization, direction unification, dimensionality reduction, and correlation screening on the extracted safety features, and perform feature grouping according to the correlation analysis results to generate a feature data set for clustering;

[0078] S4: Optimize the clustering center using PSO, combine the GG distance metric to replace the Euclidean distance, assign scores through a scoring mechanism for the clustering result center, and achieve the SOS estimation of the lithium-ion battery through the calculation of the favorable rate;

[0079] S5: Use clustering quality indicators, life correlation verification, and robustness testing to output the evaluation results of the safety state SOS of the lithium-ion battery.

[0080] Furthermore, in the step S1, the specific content of the constructed working condition segmentation method is as follows:

[0081] The working condition segmentation method for hybrid working condition data is divided into the feature construction and segmentation parts under different working conditions.

[0082] S1-1: Based on the collected battery data, screen out the data in the static stage, and calculate the average values of current, voltage, and SOC; at the same time, use the parameterized global change point detection method (Findchangepts) to detect the change points of current, voltage, and SOC, and set the minimum segmentation length l min and the threshold β, and calculate the residual R(i,j) of all possible intervals [i,j]:

[0083]

[0084] where, μ i,j represents the mean logarithmic variance of the interval [i + 1,j]. The residual R(i,j) reflects the volatility of the data within the interval. The larger the standard deviation, the higher the residual.

[0085] Update the minimum total cost of the first j points in chronological order and the position of the last change point cp(j) that makes F(j) the smallest, that is, the backtracking pointer cp(j):

[0086]

[0087] where, F(0) = -β, cp(0) = []. Update F(j) according to equation (2), and take the i that makes the above formula the smallest as cp(j). In the recursive process, if there exists k < i satisfying: [[ID=...]]

[0088] F(k) + R(k,i) ≥ F(i) (3)

[0089] In subsequent calculations, k is pruned and no longer participates in the calculation of j>i, so that the recursive process does not need to retain the historical state of k, which greatly reduces the amount of calculation. Repeat the above process and retain the mutation point that satisfies F(j)-F(i)-R(i,j)>β.

[0090] Findchangepts' preliminary segmentation results for current, voltage, and SOC are as follows Figure 2 As shown, the segmented segments are uniformly assigned values ​​within [0.1, 1] and multiplied by the average values ​​of current, voltage, and SOC, respectively, to construct cluster features based on current, voltage, and SOC. Time is added as a feature and given a higher weight to ensure the temporal nature of the segmentation results. Therefore, the constructed feature matrix is:

[0091] features=[I mean *I abr SOC mean *SOC abr V mean *V abr time] (4)

[0092] Among them, I abr ,SOC abr ,V abr Represent the current, voltage and SOC partition assignment matrices respectively, I mean ,SOC mean ,V mean Represent the average values ​​of current, voltage and SOC respectively, and time represents time.

[0093] S1-2: Based on the preliminary segmentation results, set the number of K-means clusters:

[0094] k=min(max(C find )) (5)

[0095] Among them, C find Indicates the number of blocks divided after the initial segmentation of current, voltage and SOC.

[0096] K-means clustering is performed based on the number of clusters.

[0097] The basic process of the K-means algorithm is:

[0098] STEP 1: Based on the charge and discharge data, randomly select k objects as the initial cluster centers;

[0099] STEP 2: Determine the similarity metric and standardize the data objects;

[0100] STEP 3: Calculate the Euclidean distance from each sample data to the cluster center and perform cluster division;

[0101] STEP 4: Return to STEP 3 until the clustering stop condition (number of iterations, SSE threshold) is met.

[0102] STEP 5. Output clustering results;

[0103] Based on the clustering results, the charge and discharge and CC / CV characteristics of the same partition data are judged according to the partition current size and current change rate, and finally verified by the voltage change rate. The output working condition segmentation results are shown in Figure 3. Referring to Figure 3(a), the current results of the working condition segmentation are partitioned according to different characteristics, which well ensures the classification of similar working conditions; as shown in Figure 3(b), different constant current charge and discharge segments maintain continuity in current characteristics, while the constant voltage stage has a certain continuity in voltage characteristics. Further, in step S2, the specific contents of the constructed safety feature screening method are as follows:

[0104] The safety features used for SOS assessment of lithium-ion batteries are divided into charge and discharge segment data and global variables.

[0105] S2-1: The charge and discharge segment data is segmented based on the operating condition to extract safety features. In the constant current charge (CC) segment (1C, SOC from 80% to the cut-off voltage), the time required for SOC to rise by 10% is extracted. cc ; Extract the residual square sum S between the actual current curve and the initial reference curve in the constant voltage charging (CV) segment SSE The mean absolute error E MAE , and its calculation formula is:

[0106]

[0107] Among them, I ref,i and I act,i are the reference and actual current values ​​respectively; N is the data length. ref and I act To have the same data length, align the two lengths through linear interpolation.

[0108] The incremental capacity (IC) curve is generated by the incremental capacity analysis (ICA) method in the constant current discharge segment. The core calculation formula is:

[0109]

[0110] Among them, Q k , Q k-1 Represent the capacity changes at time k and k-1, V B,k ,V B,k-1They represent the battery terminal voltage values ​​at time k and k-1 respectively.

[0111] Since the IC curve is affected by the sampling time non-fixed and the voltage sampling process error during the actual data collection process, some important information is lost. Therefore, while performing linear interpolation on the collected voltage data, a minimum sampling interval is set to prevent dV B In order to extract relevant safety feature factors without being affected by noise, the Savitzky-Golay filtering method is introduced to reduce the noise of the IC curve. The IC curves before and after the processing are compared. Figure 4 shown.

[0112] S2-2: For the global variables, extract the parameter difference characteristics between the initial battery cycle and the cycles at different aging stages, and obtain the average temperature difference △avgT by segmenting the cycle nodes based on the MIT-Stanford dataset. i,j , capacity difference C ini ,i and internal resistance difference R ini,i , where ini represents the initial cycle, i and j represent the percentage position of the total cycle number (such as 10%, 20%).

[0113] This type of feature enhances the robustness to segmented feature noise by integrating the full-cycle temperature, capacity, and internal resistance evolution information, maintains reliability in the event of data anomalies, and jointly constructs a multi-dimensional safety status characterization system of time-voltage-temperature-capacity-internal resistance with the charge and discharge segment features.

[0114] To demonstrate the safety of battery cells throughout their entire cycle, all batteries are characterized by early, mid, and late cycle signatures. Early cycles are characterized by single cycles at 10% and 20% of the total number of cycles; mid-cycles are characterized by single cycles at 30% and 40% of the total number of cycles; and late cycles are characterized by single cycles at 70% and 80% of the total number of cycles.

[0115] Table 1 Security status characteristics based on different fragments

[0116]

[0117]

[0118] Furthermore, in step S3, the steps for constructing a lithium-ion battery safety status dataset based on the extracted safety features are as follows:

[0119] S3-1: Normalize the extracted features and control their range within the range of [-1, 1] to ensure that the cluster shape of each feature is as regular as possible during the SOS estimation process;

[0120] S3-2: The correlation between all features and capacity changes is analyzed using the Pearson correlation coefficient method (PCC). The calculation process is as follows.

[0121]

[0122] Among them, X is the data sequence of a single safety feature, Y is the capacity change sequence of the battery cell, and are the corresponding average values, and n is the sequence length. Regardless of whether the correlation is obvious or not, the features that are inversely proportional to the capacity correlation are reciprocated to ensure that all features are positively correlated with the capacity.

[0123] S3-3: Dimensionality reduction is performed using principal component analysis (PCA). The principle is as follows:

[0124] STEP 1: The matrix of the m principal components involved in the analysis is as follows:

[0125]

[0126] Where, X ij Represents the element in the i-th row and j-th column of the matrix. The m*n elements contained in the matrix indicate that the matrix is ​​composed of elements with m characteristics of n batteries.

[0127] STEP 2: Standardize the data in X:

[0128]

[0129] Where, X ij Represents the original data, represents normalized data, represents the mean of the jth column, s j Table Euclidean standard deviation of column j.

[0130] STEP 3: Use X ro =Q T X vs. X * Perform eigenvalue decomposition on the correlation coefficient matrix and rotate the data to project it onto the principal component axis;

[0131] STEP 4: Calculate the principal component contribution rate and cumulative contribution rate:

[0132]

[0133] Where C h Indicates the contribution rate of the hth principal component, C sum represents the cumulative contribution rate of the first k principal components, and λ represents the eigenvalue.

[0134] The information provided by different features is sorted and summed up. Figure 5 As shown in Figure 2, key features with cumulative contribution rates exceeding 90% are selected to construct a new feature matrix.

[0135] S3-4: PCC analysis results based on the selected features are as follows Figure 6 As shown in the figure, 3 features in each group (PCC ≤ 0.7 with each other) are selected for combination, and features are allowed to be reused to cover complete information. Finally, 107 feature groups are generated to provide input for PSO-GG clustering.

[0136] Furthermore, in step S4, the battery SOS estimation process based on the PSO-GG algorithm is proposed as follows:

[0137] S4-1: The SOS estimation process based on particle swarm optimization-fuzzy clustering (PSO-GG) uses the PSO algorithm to optimize the cluster center of a single iteration of FCM, and combines the GG distance measurement module to replace the Euclidean distance to capture the non-spherical cluster structure. Finally, the safety status of lithium-ion batteries is quantitatively evaluated through the scoring mechanism and the calculation of the praise rate.

[0138] S4-2: The specific process of using the PSO-GG-based SOS estimation framework to perform security status assessment is as follows:

[0139] Set the core parameters of PSO-GG clustering, set the number of clusters c = 2, the fuzzy index m = 2, the convergence threshold ε = 1e-8, and the maximum number of iterations to 200;

[0140] The 107 generated feature sets are input into the PSO-GG framework for clustering. Each feature set is clustered independently. The safety of the battery on this feature set is distinguished based on the size of the cluster center value optimized by PSO, and a safety score of 0 / 1 is assigned. Finally, the positive rate of all subsets is calculated as the SOS value.

[0141] In the clustering process, the PSO optimization module uses the GG distance objective function to dynamically search for cluster centers. The objective function expression is:

[0142]

[0143] Among them, fit represents the objective function of the PSO algorithm, u k-1 represents the membership matrix of the k-1th iteration in the FCM algorithm iteration process, x j Represents the jth sample point in the sample data set. Each particle in the PSO represents a complete cluster center. By calculating the fit corresponding to each particle and iteratively updating the position of the particle (i.e., the candidate center), we continuously search for a combination of centers that makes the fit smaller.

[0144] In the clustering process, the GG distance metric module replaces the traditional Euclidean distance with the following formula:

[0145]

[0146] Among them, d ij 2 represents the square of the distance from sample j to the center of cluster i, det(F i ) represents the covariance matrix F i The determinant, P i represents the prior probability of cluster i, F i is the covariance matrix.

[0147] S4-3: The PSO iteration process is shown in Figure 7. After the PSO iteration is completed, the global optimal particle is returned, that is, the cluster center with the minimum objective function fit is used as the cluster center of the FCM in the kth iteration. This cluster center is re-introduced into the FCM to update the membership degree and calculate the clustering result of the FCM.

[0148] S4-3: In the process of estimating SOS, the data set constructed by different battery cell characteristics is obviously non-spherical, so the Euclidean distance is difficult to capture the true structure of the data, which greatly affects the results of FCM. Therefore, GG clustering is introduced, which improves the clustering performance by introducing a distance measure based on fuzzy maximum likelihood estimation. GG first introduces the covariance matrix F i :

[0149]

[0150] Then, through the covariance matrix F i Calculate the prior probability P for each cluster i :

[0151]

[0152] Finally, the exponential distance measure is used to calculate the distance from each data point to each cluster, and GG clustering is combined with PSO-FCM. GG replaces the Euclidean distance calculation method in FCM, thereby achieving dual optimization of cluster center and distance measure of PSO-GG.

[0153] S4-5: Since all features are guaranteed to be positively correlated with capacity, classification is performed based on the cluster center obtained by PSO optimization. In the feature set consisting of n features, if the battery belongs to a cluster with a larger cluster center mean, it is defined as a battery cell with better safety, and a value of 1 is assigned to it in this judgment.

[0154] According to the clustering method, the generated k feature matrices are clustered respectively to obtain k ratings. The positive rating rate of the battery is calculated based on the k rating results:

[0155]

[0156] Among them SOS cell Indicates the praise rate of the battery cell, that is, the corresponding SOS; Y cell It represents the scoring result of the battery cell in the i-th clustering; k represents the number of feature sets, that is, the number of clustering times.

[0157] This method avoids the problem of difficulty in determining the threshold value of different parameters in the process of establishing the SOS mathematical model. The battery is automatically classified through an unsupervised learning framework and the SOS value of the battery is finally determined.

[0158] In step S5, the safety status SOS of the lithium-ion battery is evaluated as follows:

[0159] S5-1: In order to evaluate the superiority of the PSO-GG model in clustering, the compactness (Cmp), separation (Spt) and the ratio of the two (Cmp / Spt) are used to evaluate the clustering quality. The definitions of compactness and separation are:

[0160]

[0161] Spt=min i≠k ||c i -c k || 2 (20)

[0162] Among them, Cmp reflects the compactness of the data after data partitioning, that is, the quantitative similarity of data within a class. Spt is a function of the center matrix, which reflects the separability of the data after data partitioning, that is, the dissimilarity between classes as much as possible.

[0163] To verify the advancedness of the proposed method, the clustering results of traditional FCM, GG, and PSO-GG are compared, and the evaluation index results are shown in Table 1. All three methods perform complete clustering and take the average value of all evaluation indicators. The initial class centroid matrix given by the two algorithms is automatically generated by the algorithm.

[0164] Table 1 Comparison of three evaluation index values ​​of FCM, GG and PSO-GG clustering

[0165]

[0166] The results show that the Cmp values ​​for the PSO-GG clustering are all smaller than those for the FCM and GG clustering methods, indicating that PSO-GG has slightly better compactness. The Spt values ​​for PSO-GG and GG are larger than those for the FCM clustering methods, indicating that the separation between PSO-GG and GG is better using the fuzzy maximum likelihood distance measure. The Cmp / Spt values ​​for PSO-GG are all smaller than those for FCM and GG clustering methods, indicating that PSO-GG has better quality in terms of both compactness and separation.

[0167] S5-2: After verifying the reliability of the proposed safety feature dataset, 124 batteries were estimated using the proposed SOS estimation method based on PSO-GG. The estimated results were renumbered according to the battery's cycle life, and a comparison between the score and the final life was plotted. Figure 8 It can be seen that there is a certain correlation between the battery safety score and the cycle life. Although the exact safety cannot be quantitatively predicted, the relative safety status of the battery cell can basically be judged based on the results.

[0168] Re-estimate the SOS of the data after adding noise, and the estimated results are as follows Figure 9 As shown in Figure 2, it can be seen that the 30% error has a very limited impact on the error of the SOS estimation, and the SOS result after fitting is almost the same as the SOS estimation without noise, which shows that the SOS estimation method proposed in this paper has a certain degree of robustness.

[0169] In summary, this method proposes a lithium-ion battery safety state estimation method based on multi-source features and unsupervised clustering. It uses battery charge and discharge segment features and global variables as inputs to the SOS estimation model. It leverages the advantages of K-means operating condition segmentation and the PSO-GG clustering algorithm in feature extraction and dynamic distance measurement, respectively, to achieve adaptive modeling of complex multi-source safety features. By integrating multi-dimensional safety information through a feature grouping scoring and praise rate fusion mechanism, an unsupervised clustering-driven SOS quantitative assessment process is constructed. This effectively overcomes the SOS estimation's dependence on parameter thresholds and provides a new approach for practical SOS estimation.

[0170] In this specification, the schematic descriptions of the present invention do not necessarily refer to the same embodiment or example. Those skilled in the art may combine and combine different embodiments or examples described in this specification. In addition, the contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The scope of protection of the present invention should not be considered as limited to the specific forms described in the implementation cases. The scope of protection of the present invention also includes equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A lithium-ion battery safety state estimation method based on operating condition segment features and unsupervised clustering, characterized in that: The following steps are involved: S1: Based on the MIT-Stanford lithium-ion battery mixed operating cycle dataset, a K-means-based segmentation method is proposed to segment the data under different operating conditions according to the different voltage and current characteristics of the mixed operating conditions. S2: Screen the extracted data for security features during constant current (CC) charging, constant voltage (CV) charging, and constant current discharging phases, and introduce global security features. S3: Normalize, unify directions, reduce dimensions, and perform correlation screening on the extracted security features. Group the features based on the correlation analysis results to generate a feature dataset for clustering. S4: PSO is used to optimize the cluster centers, and the GG distance metric is combined to replace the Euclidean distance. The scoring mechanism is used to assign points to the cluster center, and the SOS estimation of lithium-ion batteries is achieved by calculating the praise rate. S5: Use clustering quality indicators, life correlation verification and robustness testing to output the evaluation results of the lithium-ion battery safety status SOS.

2. The method for estimating the safety state of a lithium-ion battery based on operating condition segment features and unsupervised clustering according to claim 1, wherein: In step S1, the segmentation method based on K-means is constructed as follows: The working condition segmentation method for mixed working condition data is divided into feature construction and segmentation under different working conditions; S1-1: Based on the collected battery data, the static phase data is screened out and the average values ​​of current, voltage and SOC are calculated. At the same time, a parameterized global change point detection method (Findchangepts) is used to detect the mutation points of current, voltage and SOC. Based on the preliminary segmentation results of Findchangepts for current, voltage, and SOC, the segmented segments are uniformly assigned values ​​within [0.1, 1] and multiplied by the average values ​​of current, voltage, and SOC, respectively, to construct clustering features based on current, voltage, and SOC. Time is added as a feature and given a higher weight to ensure the temporal nature of the segmentation results. Therefore, the constructed feature matrix is: features=[I mean *I abr SOC mean *SOC abr V mean *V abr time] (1) Among them, I abr ,SOC abr ,V abr Represent the current, voltage and SOC partition assignment matrices respectively, I mean ,SOC mean ,V mean Represent the average values ​​of current, voltage and SOC respectively, and time represents time; S1-2: Based on the preliminary segmentation results, set the number of K-means clusters: k=min(max(C find )) (2) Among them, C find Indicates the number of blocks divided after the initial segmentation of current, voltage and SOC; K-means clustering is performed according to the number of clusters. Based on the clustering results, the charge-discharge and CC / CV characteristics of the same partition data are judged according to the partition current size and current change rate. Finally, the voltage change rate is used to verify and output the results.

3. The method for estimating the safety state of a lithium-ion battery based on operating condition segment features and unsupervised clustering according to claim 2, wherein: In step S2, the specific contents of the constructed security feature screening method are as follows: Safety features for SOS assessment of lithium-ion batteries, divided into charge and discharge segment data and global variables; S2-1: The charge and discharge segment data is segmented based on the operating condition to extract safety features. In the constant current charge (CC) segment (1C, SOC from 80% to the cut-off voltage), the time required for SOC to rise by 10% is extracted. cc ; Extract the residual square sum S between the actual current curve and the initial reference curve in the constant voltage charging (CV) segment SSE The mean absolute error E MAE , and its calculation formula is: Among them, I ref,i and I act,i are the reference and actual current values ​​respectively; N is the data length; in order to ensure I ref and I act To have the same data length, align the two lengths by linear interpolation; The incremental capacity (IC) curve is generated by the incremental capacity analysis (ICA) method in the constant current discharge segment. The core calculation formula is: Among them, Q k , Q k-1 Represent the capacity changes at time k and k-1, V B,k ,V B,k-1 They represent the battery terminal voltage values ​​at time k and k-1 respectively; In order to ensure the integrity of the IC curve and avoid abnormal noise interference, the minimum sampling interval is set while performing linear interpolation on the collected voltage data to prevent dV B The noise step caused by too small a value is eliminated, and the Savitzky-Golay filtering method is introduced to reduce the noise of the IC curve; S2-2: For the global variables, extract the parameter difference characteristics between the initial battery cycle and the cycles at different aging stages, and obtain the average temperature difference △avgT by segmenting the cycle nodes based on the MIT-Stanford dataset. i,j , capacity difference C ini ,i and internal resistance difference R ini,i , where ini represents the initial cycle, i and j represent the percentage position of the total number of cycles (e.g., 10%, 20%); This type of feature enhances the robustness to segmented feature noise by integrating the full-cycle temperature, capacity, and internal resistance evolution information, maintains reliability in the event of data anomalies, and jointly constructs a multi-dimensional safety status characterization system of time-voltage-temperature-capacity-internal resistance with the charge and discharge segment features.

4. The method for estimating the safety state of a lithium-ion battery based on operating condition segment features and unsupervised clustering according to claim 3, wherein: In step S3, the steps for constructing a lithium-ion battery safety status dataset based on the extracted safety features are as follows: S3-1: Normalize the extracted features and control their range within the range of [-1,1]; S3-2: The correlation between all features and capacity changes was analyzed using the Pearson correlation coefficient (PCC) method. The calculation process is as follows; Among them, X is the data sequence of a single safety feature, Y is the capacity change sequence of the battery cell, and are the corresponding average values, and n is the sequence length; the features that are inversely proportional to the capacity change are reciprocated to ensure that all features are positively correlated with the capacity; S3-3: Use principal component analysis (PCA) to reduce dimensionality, sort and sum the information provided by the features, and select key features with a cumulative contribution rate exceeding 90% to construct a new feature matrix; S3-4: Based on the PCC analysis results between the filtered features, 3 features in each group (PCC ≤ 0.7) are selected for combination, and features are allowed to be reused to cover complete information. Finally, 107 feature groups are generated to provide input for PSO-GG clustering.

5. The method for estimating the safety status of a lithium-ion battery based on multi-source features and unsupervised clustering according to claim 4, wherein: In step S4, the battery SOS estimation process based on the PSO-GG algorithm is proposed as follows: S4-1: The SOS estimation process based on particle swarm optimization-fuzzy clustering (PSO-GG) uses the PSO algorithm to optimize the cluster centers of a single FCM iteration, and combines the GG distance metric module to replace the Euclidean distance to capture non-spherical cluster structures. Finally, a scoring mechanism and praise rate calculation are used to achieve quantitative evaluation of the safety status of lithium-ion batteries. S4-2: The specific process of using the PSO-GG-based SOS estimation framework to perform safety status assessment is as follows: Set the core parameters of PSO-GG clustering, set the number of clusters c = 2, the fuzzy index m = 2, the convergence threshold ε = 1e-8, and the maximum number of iterations to 200; The 107 generated feature sets are input into the PSO-GG framework for clustering. Each feature set is clustered independently. The safety of the battery on this feature set is distinguished based on the size of the cluster center value optimized by PSO, and a safety score of 0 / 1 is assigned. Finally, the positive rate of all subsets is calculated as the SOS value. In the clustering process, the PSO optimization module uses the GG distance objective function to dynamically search for cluster centers. The objective function expression is: Among them, fit represents the objective function of the PSO algorithm, u k-1 represents the membership matrix of the k-1th iteration in the FCM algorithm iteration process, x j Represents the jth sample point in the sample data set; each particle in the PSO represents a complete cluster center. By calculating the fit corresponding to each particle and iteratively updating the position of the particle (i.e., the candidate center), we continuously search for a center combination that makes the fit smaller. In the clustering process, the GG distance metric module replaces the traditional Euclidean distance with the following formula: Among them, d ij 2 represents the square of the distance from sample j to the center of cluster i, det(F i ) represents the covariance matrix F i The determinant, P i represents the prior probability of cluster i, F i is the covariance matrix.

6. The method for estimating the safety status of a lithium-ion battery based on multi-source features and unsupervised clustering according to claim 5, wherein: In step S5, the safety status SOS of the lithium-ion battery is evaluated as follows: S5-1: In order to evaluate the estimation ability of the PSO-GG model for SOS, the compactness Cmp, separation Spt and their ratio Cmp / Spt are used to evaluate the clustering quality; the definitions of compactness and separation are: Spt=min i≠k ||c i -c k || 2 (10) Among them, Cmp reflects the compactness of the data after data partitioning, that is, the quantitative similarity of data within the class; Spt is a function of the center matrix, which reflects the separability of the data after data partitioning, that is, the classes are as dissimilar as possible; S5-2: To evaluate the reliability of the algorithm under actual noise interference, 30% random noise (the noise range is ±30% of the characteristic value) is applied to the safety characteristics of 30 random batteries. The SOS estimation is re-performed on the noise-imposed data, and the SOS estimation values ​​before and after the noise is applied are compared to verify the robustness of the proposed SOS estimation method.

Images