A battery performance evaluation method based on knowledge confidence and contribution degree

Abstract

The application discloses a battery performance evaluation method based on knowledge confidence and contribution degree, comprising the following steps: 1, collecting original feature vectors of batteries to be evaluated and dividing them into healthy and abnormal battery sample sets; 2, performing standardization processing on the features, and screening a combined feature subset based on filtering and bypassing feature selection methods; 3, performing self-updating on the original knowledge base based on the combined features to supplement new battery sample instances and eliminate low-value old instances; 4, calculating sample performance deviation based on the updated knowledge base, and combining model consistency, appearance defect and electrical performance stability factors to calculate evaluation scores, so as to determine the performance grade. The application can unify the appearance defect of the battery, the electrical performance stability, the model consistency and the knowledge base matching result into one evaluation framework, so that the evaluation result has both quantitative scores and clear grade output, and is convenient for direct use in production quality inspection, operation and maintenance, recycling and sorting and step utilization scenes.

Description

Technical Field

[0001] This invention belongs to the field of battery testing and performance evaluation technology, specifically a battery performance evaluation method based on knowledge confidence and contribution. Background Technology

[0002] With the rapid development of the new energy industry, energy storage systems, electric vehicles, and portable electronic devices, batteries, as the core component for energy storage and supply, directly affect the safety, stability, and lifespan of equipment operation. To ensure the reliability of batteries during production, sorting, recycling, secondary use, and operation and maintenance, it is usually necessary to inspect and evaluate the battery's appearance, electrical performance, and operational stability.

[0003] Existing battery performance evaluation methods mainly include manual inspection, electrical parameter testing, image recognition inspection, and classification evaluation based on historical sample data. Manual inspection typically relies on inspectors to observe and judge surface damage, leakage, markings, and sealing conditions. Electrical parameter testing usually analyzes the battery's operating status by collecting parameters such as voltage, current, capacity, state of charge, and temperature. Image recognition inspection uses features such as cracks, damage, leakage, and sealing defects in battery appearance images to assist in judging the battery's condition. Classification evaluation methods based on historical sample data typically use existing samples to build reference datasets or models to perform category matching or performance determination on the batteries under test.

[0004] However, existing battery performance evaluation methods still have certain shortcomings in practical applications. First, manual testing is greatly affected by the experience, fatigue level, and subjective judgment of the testers, making it difficult to guarantee the consistency and stability of the test results, and its efficiency is low in large-scale battery testing scenarios. Second, when relying solely on electrical parameters such as voltage, current, capacity, or temperature for evaluation, it is easy to overlook the impact of appearance defects such as battery surface damage, abnormal sealing, and leakage on battery safety and subsequent performance, resulting in incomplete evaluation results.

[0005] Furthermore, existing evaluation methods based on image or feature data typically collect a large number of raw features, but these features may suffer from redundancy, noise, or insufficient correlation. Using all features directly for performance evaluation not only increases computational complexity but may also reduce the accuracy and stability of the evaluation results. While some methods employ feature selection to reduce data dimensionality, their selection process often focuses on single statistical differences or single image features, making it difficult to consider the comprehensive relationship between battery sample category differences and local structural states.

[0006] Meanwhile, in existing battery evaluation methods based on historical samples or knowledge bases, the reference samples are usually relatively fixed. When the model, batch, usage environment, or aging state of the tested object changes, the original sample data cannot reflect the new battery state distribution in a timely manner, which can easily lead to the matching results deviating from the actual state. As test data continues to accumulate, there may also be data instances in the old samples that have low retrieval frequency, small classification contribution, or high repetition with new samples. If such samples are retained for a long time, it will not only increase the data storage and computational burden, but may also affect the reliability of subsequent evaluations.

[0007] Furthermore, existing battery performance evaluation results are mostly expressed as qualified, unqualified, or simply abnormal categories, resulting in a rather coarse evaluation dimension. In practical applications, battery performance status is not only related to the differences between healthy benchmark samples, but may also be affected by factors such as model consistency, surface damage, leakage status, state of charge fluctuations, voltage stability, and temperature changes. Without a comprehensive quantification of these factors, evaluation results are prone to being one-sided, with unclear grading boundaries and insufficient identification of abnormal states.

[0008] Therefore, how to improve the effectiveness of feature utilization, enhance the adaptability of reference sample data, and achieve a multi-factor comprehensive evaluation of battery performance status during battery performance evaluation has become an urgent problem to be solved in the field of battery testing and performance evaluation. Summary of the Invention

[0009] This invention addresses the shortcomings of existing technologies by proposing a battery performance evaluation method based on knowledge confidence and contribution. The aim is to provide a more comprehensive, stable, and accurate performance evaluation of battery samples from different batches and in different states across various application scenarios, including battery production testing, operation and maintenance, recycling and sorting, and secondary utilization. This will improve the match between battery performance assessment results and actual usage conditions, reduce the risk of misjudgment due to incomplete evaluation or outdated reference data, and provide a reliable basis for battery safety management, quality control, and subsequent usage decisions.

[0010] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The present invention provides a battery performance evaluation method based on knowledge confidence and contribution, characterized by the following steps: Step 1, collect the first One battery sample to be evaluated The original feature vector ,in, Indicates the first One battery sample to be evaluated The primitive characteristics, , Represents the total number of original features. ; This indicates the total number of battery samples to be evaluated; Based on the label of each battery sample to be evaluated, The battery samples to be evaluated were divided into a healthy battery sample set. and abnormal battery sample set ; Step 2, use equation (1) to... The first in primitive characteristics After standardization, the first... One battery sample to be evaluated Standardized Type of characteristics Thus, the first One battery sample to be evaluated Standardized feature vectors : (1) In equation (1), This indicates a positive number that prevents the denominator from being zero. Indicates the first primitive characteristics in The mean of the battery samples to be evaluated; Indicates the first primitive characteristics in One battery sample to be evaluated The standard deviation in; Step 3, based on the filtering feature selection method and the bypass feature selection method, for By performing combination filtering, the first... One battery sample to be evaluated Filtered combined feature vectors Thus, a subset of combined features is obtained. ; Step 3.1, use equation (2) to calculate the standardized first... Filtering discriminant value of a feature : (2) In equation (2), Represents the standardized first Features in healthy battery sample set The mean of the values; Represents the standardized first Features in the abnormal battery sample set The mean of the values; Represents the standardized first Features in healthy battery sample set The standard deviation in; Represents the standardized first Features in the abnormal battery sample set The standard deviation in; Step 3.2, calculate the standardized first step using equation (3). Correlation values ​​of various features : (3) In equation (3), Represents the standardized first The closure ratio of the battery sealing area corresponding to this feature; Represents the standardized first The continuity ratio of the battery sealing area corresponding to this feature; Represents the standardized first The grayscale transition between the edge-sealed region and the neighboring substrate corresponding to this feature; Step 3.3, use equation (4) to obtain the standardized first... Combination selection coefficient of features : (4) In equation (4), Represents the portfolio balance coefficient. ; Step 3.4, Set the combined selection threshold Select the one that satisfies Features constitute a combination of feature subsets ;in, Indicates the filtered first Combinatorial characteristics; This represents the total number of combined features after filtering. Based on the combined feature subset , obtained the One battery sample to be evaluated Filtered combined feature vectors ; Step 4, construct the original knowledge base ,include: Individual old battery sample instances, and based on combined feature subsets For the original knowledge base The knowledge base is updated by adding new battery sample instances and removing some old battery sample instances. Among them, the first Used battery sample example This includes: combined feature vector, true class label, timestamp, recent call frequency, number of matches, and classification contribution; ; Step 4.1, based on the original knowledge base Examples of old batteries of the same category in the original knowledge base The middle belongs to the first The mean of the combined feature vectors of the old battery sample instances of the category is calculated to obtain the first... Battery Sample Instance Center by Category ;in, Indicates the first The center of the battery sample instance of the category is in the first The central value of a combination of features; Step 4.2, with the first One battery sample to be evaluated As a new battery sample, Equation (5) is used for calculation. With the original knowledge base Middle Center for old battery sample examples Weighted difference between : (5) In equation (5), Indicates the first Selection coefficients for various combinations of features; Indicates the filtered first Combinatorial characteristics; Step 4.3, calculate using equation (6) Minimum difference and determine Recent Category : (6) In equation (6), Represents the original knowledge base The total number of categories; Indicates taking The category number corresponding to the minimum difference value; Step 4.4, Calculation Cluster confidence and time series stability index ; when , and At that time, Combined feature vectors The real category label, timestamp, number of matches, and classification contribution are used as new battery sample instances. Write it into the knowledge base; among which, This indicates the addition of a new judgment threshold. Indicates the confidence threshold. Indicates the timing stability threshold; otherwise, do not write. Step 4.5, use equation (7) to calculate the first [item] in the knowledge base. Used battery sample example Elimination Index : (7) In equation (7), Indicates the first Used battery sample example The recent call frequency normalized value; Indicates the first Used battery sample example Normalized value of the number of matches between the battery sample instance and any newly added battery sample instance; Indicates the first Used battery sample example The contribution of the category; Step 4.6, when At that time, the first Remove the old battery sample instance from the knowledge base; otherwise, retain the first one. Used battery sample instances; thus obtaining the updated knowledge base. ;in, Indicates the rejection threshold; Step 5, based on the updated knowledge base Using equation (8) for the first One battery sample to be evaluated Perform performance matching to obtain performance deviation : (8) In equation (8), This indicates the updated knowledge base. China Health Benchmark Center in the 19th The central value of a combination of features; Indicates the filtered first Combinatorial characteristics; Step 6, use equation (9) to calculate the first... One battery sample to be evaluated Assessment score : (9) In equation (9), for Model consistency factor; for Surface damage penalty factor, for The exudate penalty factor, for The charge state stability factor, for The normalization result; and ; Step 7, when At that time, the judgment of the first One battery sample to be evaluated For excellent performance; when At that time, the judgment of the first One battery sample to be evaluated For performance to be available; when At that time, the judgment of the first One battery sample to be evaluated This indicates a performance anomaly; among them, and This indicates two hierarchical thresholds, and .

[0011] The battery performance evaluation method based on knowledge confidence and contribution described in this invention is characterized in that step 6 uses equation (10) to calculate... Model consistency factor : (10) In equation (10), for The model code vector of the battery surface marking area. for Reference model template vector for the battery surface marking area; Indicates modulo, This indicates a positive number that prevents the denominator from being zero.

[0012] Furthermore, in step 6, the surface damage penalty factor is calculated using equation (11). : (11) In equation (11), for The normalized value of the crack length, for The percentage of damaged area for The dispersion of the damaged edge, , and The range of values ​​is .

[0013] Furthermore, in step 6, equation (12) is used to calculate... Exudate penalty factor : (12) In equation (12), Indicates the first One battery sample to be evaluated The color difference between the exudate area and the normal area is abnormal. for The ratio of the area of ​​the seepage zone, for The boundary diffusion rate.

[0014] Furthermore, in step 6, equations (13-1) and (13-2) are used to calculate... Charge state stability factor : (13-1) (13-2) In equations (13-1) and (13-2), for In a single sampling time window The state of charge within, for During the sampling time window Internal operating current, For rated capacity, for The initial state of charge, This represents the reference state of charge value corresponding to the healthy battery sample; Indicates the first One battery sample to be evaluated During the sampling time window Average operating voltage within; Indicates the first One battery sample to be evaluated During the sampling time window The average surface temperature inside; and These represent the reference voltage and reference surface temperature corresponding to the healthy battery sample, respectively. , and There are three stability weights, all of which are not less than 0. This indicates a positive number that prevents the denominator from being zero.

[0015] Furthermore, the battery sealing area closure ratio in step 3.2 Through the standardized first The ratio of the closed contour length of the battery sealing area corresponding to each feature to the theoretical contour length is obtained; Battery sealing area continuity ratio Through the standardized first The ratio of the length of the continuous effective pixel segment corresponding to each feature to the total edge length is obtained; Gray-scale transition amount Through the standardized first The grayscale difference between the edge-sealing region corresponding to each feature and the neighboring substrate is obtained by normalization.

[0016] Furthermore, in step 4.4, equation (14) is used to calculate... Cluster confidence ; (14) In equation (14), For indicator functions, when the first One new battery sample and If the category labels are the same, the indicator function is 1; otherwise, the indicator function is 0. express and The distance between them; In order to be with the first One battery sample to be evaluated Recent The newly added battery samples constitute The neighborhood set, Indicates the first One battery sample to be evaluated With neighborhood set The mean distance between each newly added battery sample in the sample. This indicates a positive number that prevents the denominator from being zero.

[0017] Furthermore, in step 4.4, equation (15) is used to calculate... Temporal stability index : (15) In equation (15), for In the The standardized first sampling time window Such characteristics for In the The standardized first sampling time window Such characteristics This represents the number of consecutive sampling time windows; For the first Selection coefficients for various combinations of features; For continuous Within each sampling time window The number of times the classification results remain consistent; It is the balance coefficient, and .

[0018] Furthermore, in step 4.5, equation (16) is used to calculate... Classification contribution : (16) In equation (16), For the first Used battery sample example Number of times it participates in the correct classification for Number of times it participates in the overall category This indicates a positive number that prevents the denominator from being zero.

[0019] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program supporting the processor in performing the method described therein, and the processor is configured to execute the program stored in the memory.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention overcomes the problem of inconsistent dimensions and large differences in numerical ranges among different types of features affecting the evaluation results by standardizing the original characteristics of the battery samples to be evaluated. It enables features such as voltage, current, temperature, appearance defects, and sealing condition to participate in screening, matching, and scoring under a unified scale, thereby improving the stability and objectivity of the evaluation process.

[0021] 2. This invention overcomes the problems of feature redundancy, noise interference, and insufficient single screening criteria by combining a filtering feature selection method with a sealing feature selection method. This allows the combined feature subset to reflect both the categorical differences between healthy and abnormal samples and the structural state of the battery sealing area, thereby improving the accuracy of performance status identification.

[0022] 3. This invention overcomes the problem that a fixed sample library is difficult to adapt to changes in battery batches, models, usage environments, and aging states by constructing and updating a knowledge base. It allows newly added battery sample instances that meet the criteria to be added to the knowledge base, avoiding the deviation of matching results from actual conditions due to long-term fixed reference data, thereby improving adaptability and reliability in long-term testing scenarios.

[0023] 4. This invention controls the writing of new samples by using cluster confidence and temporal stability indices, overcoming the problem of introducing low-confidence samples or randomly fluctuating samples during knowledge base updates. By judging the consistency of neighborhood categories, sample distance relationships, and state changes within continuous sampling time windows, it reduces the impact of erroneous collections and short-term fluctuations on the quality of the knowledge base.

[0024] 5. This invention overcomes the problems of sample redundancy, accumulation of low-value samples, and decreased matching efficiency after long-term use of the knowledge base by eliminating some old battery sample instances through a phasing index. Low-value instances are deleted based on recent call frequency, number of matches, and classification contribution, thereby reducing storage and computational burden and improving the representativeness and matching efficiency of the knowledge base.

[0025] 6. This invention calculates and grades an evaluation score by comprehensively considering performance deviation, model consistency, surface damage, leakage status, and charge state stability, overcoming the problems of single evaluation dimensions and insufficiently refined grading results. This allows the evaluation results to simultaneously reflect the degree of health deviation, identity matching, appearance safety, and operational stability, providing a technical basis for quality inspection, sorting and reuse, safety maintenance, and retirement decisions. Attached Figure Description

[0026] Figure 1 This is a flowchart of the steps of the present invention; Figure 2 This is a flowchart of the combined feature filtering process of the present invention; Figure 3 This is a flowchart of the knowledge base update process for this invention; Figure 4 This is a schematic diagram illustrating the battery performance scoring and grading method of the present invention; Figure 5 This is a schematic diagram of the feature regions of a battery sample image according to the present invention. Detailed Implementation

[0027] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments. Obviously, these embodiments are only used to illustrate the technical solution of the present invention and are not intended to limit the scope of protection of the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the relevant steps, parameters, or implementation methods without departing from the concept of the present invention should fall within the scope of protection of the present invention.

[0028] like Figure 1 As shown, a battery performance evaluation method based on knowledge confidence and contribution includes the following steps: Step 1, collect the first One battery sample to be evaluated The original feature vector ,in, Indicates the first One battery sample to be evaluated The primitive characteristics, , Represents the total number of original features. ; This indicates the total number of battery samples to be evaluated.

[0029] Based on the label of each battery sample to be evaluated, The battery samples to be evaluated were divided into a healthy battery sample set. and abnormal battery sample set ; In practice, This refers to one or a group of battery samples being tested. This is the multidimensional original feature vector formed by the sample at the detection time. The original features may include, but are not limited to: battery terminal voltage, operating current, estimated internal resistance, average surface temperature, temperature variance, mean brightness of the casing, variance of brightness of the casing, outline length of the sealing area, length of continuous pixel segments in the sealing area, grayscale transition amount in the sealing area, crack length, damaged area, dispersion of damaged edge, area of ​​suspected leakage area, diffusion rate of leakage boundary, model encoding vector of surface marking area, and state of charge related features, etc.

[0030] In the process of image feature extraction, the acquired battery surface image can first be processed by grayscale conversion, noise reduction, distortion correction and outer contour positioning. The battery body coordinate system is established according to the battery outer contour, and the surface marking area, sealing area, neighboring base area, damage candidate area and leakage candidate area are divided according to the position of the battery outer contour.

[0031] For crack length, edge detection or binarization segmentation can be used to extract crack candidate regions. Then, the crack candidate regions are skeletonized, and the ratio of the skeleton pixel length to the battery long side length or a preset reference length is used as the crack length normalization value.

[0032] The proportion of damaged area can be determined by grayscale thresholding, texture anomaly detection, or connected component analysis, and the ratio of the area of ​​the damaged candidate area to the area of ​​the effective area on the battery surface is used as the proportion of damaged area.

[0033] The dispersion of the damaged edge can be calculated based on the boundary length of the damaged area, the change in the direction of the boundary points, the change in the boundary curvature, or the dispersion of boundary pixels. The more irregular the edge, the greater the dispersion of the damaged edge.

[0034] For areas of seepage, it is possible to , or The color difference between candidate regions and normal regions in the color space is calculated, and the seepage penalty factor is determined by combining the area ratio of seepage regions and the boundary diffusion rate.

[0035] The model code vector for the surface marking area can be obtained through character recognition, template matching, QR code recognition, or preset code mapping. The more consistent the recognition result is with the reference model template, the larger the model consistency factor.

[0036] In order to The various features in the data can stably participate in subsequent calculations, and the battery sample can be located first during data collection. For example, in... Figure 5 In this process, the battery's main coordinate system can be determined by its outer contour. The long side of the battery is defined as the horizontal coordinate, and the short side as the vertical coordinate. Then, based on their position relative to the outer contour, the surface marking area, sealing area, adjacent base area, potential damage area, and potential leakage area are divided. The advantage of this approach is that the feature areas of batteries of the same or similar models are comparable, and feature distortion will not occur due to changes in battery placement angle or shooting position.

[0037] Healthy battery sample set and abnormal battery sample set These can be derived from manual labeling, historical quality inspection records, laboratory performance test results, or business rule judgments. For example, samples that meet the requirements for capacity retention, voltage stability, appearance defects, and safety risks can be labeled as healthy; samples with obvious bulging, abnormal sealing, leakage, severe damage, abnormal voltage fluctuations, or abnormal state of charge drift can be labeled as abnormal. This can be achieved by setting... and The subsequent step 3.1 can calculate the ability of each feature to distinguish between the two types of samples.

[0038] For example, in one parameter instantiation method, one can take... , The healthy battery sample set Includes 180 samples, a collection of abnormal battery samples. It includes 120 samples. This is not a limiting value; it can be 20, 50, or more, depending on the number of detection devices and the feature design. The key is that the original feature vectors for each sample are extracted according to the same rules. Then you can proceed to step 2.

[0039] Step 2, use equation (1) to... The first in primitive characteristics After standardization, the first... One battery sample to be evaluated Standardized Type of characteristics Thus, the first One battery sample to be evaluated Standardized feature vectors : (1) In equation (1), This prevents positive numbers with a denominator of zero. Can be taken , Or other sufficiently small positive numbers to avoid a zero denominator when a feature hardly changes in the sample. Indicates the first primitive characteristics in The mean of a sample of batteries to be evaluated can be obtained by... Indicates the first primitive characteristics in One battery sample to be evaluated The standard deviation in the data can be obtained by analyzing the standard deviation of the data. The first sample The standard deviation of the original features is obtained.

[0040] If the first The original characteristic is surface temperature, for a certain batch Average surface temperature of each sample The temperature was 28.0℃, and the standard deviation was [missing value]. At 2.0℃, a certain battery sample Surface temperature The temperature was 31.0℃. The standardized result of this feature is approximately This result indicates that the sample is approximately 1.5 standard deviations above the population mean for this characteristic.

[0041] In practical implementation, standardized parameters and It can be calculated within a single testing batch, or it can be calculated and solidified based on historical training samples, and periodically updated during production. For online testing, it is preferable to... and Storing parameters in a configuration file or database allows batteries of the same batch and model to use the same set of standardized parameters, thereby improving the consistency of results.

[0042] The purpose of standardization is to eliminate differences in the units and numerical ranges of different features. Taking raw features as an example, voltage might be measured in volts, temperature in degrees Celsius, crack length in millimeters or pixels, and area percentages are typically decimals between 0 and 1. Without standardization, features with large numerical ranges may unreasonably dominate distance calculations, difference calculations, and selection calculations.

[0043] Step 3, based on the filtering feature selection method and the bypass feature selection method, for By performing combination filtering, the first... One battery sample to be evaluated Filtered combined feature vectors Thus, a subset of combined features is obtained. The core idea is that simply selecting features based on statistical differences might lead to the selection of features that are irrelevant to the appearance structure but are occasionally easily distinguishable; similarly, selecting features based solely on the edge sealing structure might overlook electrical parameters or temperature features that are more sensitive to health / abnormality classifications. Therefore, this implementation method filters discriminant values. Related values ​​of sealing Combined, we obtain the combination selection coefficient. This ensures that the selected combined features have both category discrimination capabilities and take into account information from key structural areas such as battery sealing edges.

[0044] like Figure 2 As shown, we can first standardize each feature. Calculate the filter discriminant value separately Then, calculate the wrap-around correlation value based on the features related to the image's edge-sealing area. For non-image border features, corresponding mappings can also be assigned according to preset mapping relationships. , and Alternatively, its bypass correlation value can be set to a lower baseline value. Then, the combined selection coefficient is calculated. and combined with the selection threshold Compare.

[0045] Step 3.1, use equation (2) to calculate the standardized first... Filtering discriminant value of a feature The filter discriminant value It measures the separability of a certain feature between healthy and abnormal samples. (Molecular) The larger the sample size, the healthier the battery sample set. and abnormal battery sample set The more pronounced the difference in mean values ​​for this characteristic, the more significant the difference in the denominator. The smaller the value, the smaller the internal fluctuation of the two classes of samples on this feature. Therefore, The larger the value, the more suitable the characteristic is for distinguishing between health and abnormality. (2) In equation (2), Represents the standardized first Features in healthy battery sample set The mean of the values; Represents the standardized first Features in the abnormal battery sample set The mean of the values; Represents the standardized first Features in healthy battery sample set The standard deviation in; Represents the standardized first Features in the abnormal battery sample set The standard deviation in.

[0046] For example, if the mean of a certain edge grayscale transition feature is 0.20 in healthy samples and 0.85 in abnormal samples, with a standard deviation of 0.10 for healthy samples and 0.15 for abnormal samples, then... This indicates that the feature has strong discriminative power. Conversely, if the mean of a certain background brightness feature is similar in healthy and abnormal samples, but fluctuates significantly, then... The value is low, so this feature is not suitable as a key composite feature.

[0047] The advantage of using equation (2) is that it has low computational cost, strong interpretability, and does not rely on complex model training. For scenarios in the production site where the number of detection samples is constantly increasing and the importance of features needs to be quickly reassessed, the filter discriminant value can be updated in a short time.

[0048] Step 3.2, calculate the standardized first step using equation (3). Correlation values ​​of various features : (3) In equation (3), Represents the standardized first The closure ratio of the battery sealing area corresponding to this feature; Represents the standardized first The continuity ratio of the battery sealing area corresponding to this feature; Represents the standardized first The grayscale transition amount between the edge-sealing region and the neighboring substrate corresponding to this feature.

[0049] like Figure 5As shown, the sealing area, located near the outer contour of the battery, is a crucial region reflecting the battery's structural integrity, safety risks, and manufacturing defects. Inadequate sealing closure, localized sealing breaks, and abnormal grayscale transitions between the sealing edge and the adjacent substrate can all be associated with bulging, damage, leakage, encapsulation abnormalities, or mechanical damage. Therefore, a sealing correlation value is introduced during the feature selection process. This can make the screening results closer to the actual structure of the battery.

[0050] The battery sealing area closure ratio in step 3.2 Through the standardized first The ratio of the closed contour length of the battery sealing region corresponding to each feature to the theoretical contour length is obtained; the continuity ratio of the battery sealing region... Through the standardized first The grayscale transition amount is obtained by the ratio of the length of the continuous effective pixel segment corresponding to the feature to the total edge length. Through the standardized first The grayscale difference between the edge-sealing region corresponding to each feature and the neighboring substrate is obtained by normalization.

[0051] Specifically, the closure ratio of the battery sealing area The length of the detected closed contour can be calculated as "detected closed contour length / theoretical edge sealing contour length". The theoretical edge sealing contour length can be obtained from the standard external dimensions of the battery, the image calibration ratio, or the template contour; the detected closed contour length can be extracted after edge detection, contour tracking, or semantic segmentation. The closer it is to 1, the more complete the edge sealing is usually; if A significantly low reading may indicate edge banding breakage, obstruction, damage, or abnormal imaging.

[0052] Battery sealing area continuity ratio It can be calculated as "length of continuous effective pixel segment / total edge length". A continuous effective pixel segment can be understood as a sequence of pixels that has continuous brightness, texture or edge response within the edge area and meets the threshold condition. The higher the value, the more continuous the edge-sealing area; The lower the value, the more likely it is that there are breaks, cracks, or contamination in the edge banding.

[0053] Gray-scale transition amount It can be obtained by normalizing the difference between the average gray level of the edge sealing area and the average gray level of the neighboring base area, or by obtaining the mean edge gradient, local contrast, or absolute value of gray level difference. The denominator in equation (3) is... This setting is used to reduce the sealing correlation value when the grayscale transition is too large. The reason for this setting is that abnormally large grayscale transitions may originate from stains, reflections, damaged boundaries, or seepage boundaries, and their corresponding characteristics do not necessarily represent stable and reliable sealing structure characteristics.

[0054] For example, if a certain edge-sealing related feature corresponds to ,but If another feature ,but The former is more likely to be selected into the subset of combined features, while the latter may be suppressed due to poor edge continuity and large gray-level transitions.

[0055] Step 3.3, use equation (4) to obtain the standardized first... Combination selection coefficient of features : (4) In equation (4), Represents the portfolio balance coefficient. Equation (4) consists of three parts: This represents the weighted contribution of the filtering decision value; This represents the weighted contribution of the bypass correlation value; This indicates the interaction contribution between the two. The purpose of introducing interaction terms is to encourage features that possess both high discriminative power and high correlation with the sealing structure. If a feature only shows statistical differences but is unrelated to the key structural region, or is related to the sealing region but cannot distinguish between healthy and abnormal features, then the interaction term is low, and the overall priority of that feature will be limited.

[0056] In practice, It can be adjusted according to the detection target. If more attention is paid to overall classification accuracy, then [adjustment can be made]. 0.8 gives the filtering decision value a greater weight; if more emphasis is placed on identifying anomalies in the sealing structure, then... This gives greater weight to the correlation value between battery appearance defects and performance risks. For scenarios where battery appearance defects are highly correlated with performance risks, the preferred value is... This ensures that statistical judgment and edge sealing structure information remain relatively balanced.

[0057] For example, when hour, ;when hour, If the combined selection threshold If the former can enter the subset of combined features, the latter will not.

[0058] Step 3.4, Set the combined selection threshold Select the one that satisfies Features constitute a combination of feature subsets ;in, Indicates the filtered first Combinatorial characteristics; This represents the total number of combined features after filtering; in specific implementation, The settings can be adjusted based on experience, validation set results, or the number of target features. For example, when... At that time, it can be expected that the filter will be selected. Key combined features; if Too low a value may retain too many weakly correlated features, increasing the computational load of the subsequent knowledge base; if... If the threshold is too high, some features with auxiliary value may be excluded, leading to unstable evaluation results.

[0059] Based on combined feature subsets , obtained the One battery sample to be evaluated Filtered combined feature vectors .

[0060] In one specific implementation, it can be done by first pressing Sort from largest to smallest, then combine Truncate it. For example, if And a total of 12 features satisfy ,but For each battery sample Only the standardized values ​​corresponding to these 12 features are retained. This reduces the feature dimensions while retaining key features with high explanatory value.

[0061] The technical benefit of this step is that it reduces redundant input and improves the efficiency and stability of subsequent knowledge base matching. Because... Only by combining feature subsets The features in the model are well-defined, so the difference calculations in subsequent steps 4 and 5 will not be diluted by a large number of irrelevant features.

[0062] For ease of description, combine feature subsets This represents the set of features retained after filtering by combination selection coefficients. It can represent the filtered feature names or feature index set, or in specific calculations, the set of feature values ​​of the corresponding sample on the filtered feature.

[0063] Specifically, for the first One battery sample to be evaluated Based on combined feature subsets The resulting combined feature vector It can be represented as ,in for In the The standardized eigenvalues ​​on the combined features. In equations (5) and (8), Used to indicate the current sample in the [number]th [year]. The eigenvalues ​​of a combination of features, and their relationship with They have the same meaning.

[0064] Step 4, construct the original knowledge base ,include: Individual old battery sample instances, and based on combined feature subsets For the original knowledge base The knowledge base is updated by adding new battery sample instances and removing some old battery sample instances. Among them, the first Used battery sample example This includes: combined feature vector, true class label, timestamp, recent call frequency, number of matches, and classification contribution; .

[0065] Original knowledge base This can be understood as a collection of historical battery sample experiences. Each old battery sample instance... It not only stores the combined feature vectors, but also the true class labels, timestamps, recent call frequency, number of matches, and classification contribution. The reason for this design is that knowledge base updates must consider not only whether samples are similar, but also whether samples are recent, frequently called, whether they contribute to correct classification, and whether they have been replaced by new samples.

[0066] For example, The following data structure can be used to store it: ,in To combine feature vectors, For real category labels, For the timestamp of the sample being written or most recently updated, This is a normalized value based on recent call frequency. This is the normalized value for the number of matches. This data structure represents the contribution to the classification. It can be stored in a relational database, key-value database, vector database, or local file system.

[0067] like Figure 3 As shown, knowledge base updates involve two directions: first, adding representative battery sample instances; and second, removing low-value, old battery sample instances. Through this two-way update, Compared to It has better adaptability and can be updated according to changes in battery batch, model, aging stage and operating conditions.

[0068] Step 4.1, based on the original knowledge base Examples of old batteries of the same category in the original knowledge base The middle belongs to the first The mean of the combined feature vectors of the old battery sample instances of the category is calculated to obtain the first... Battery Sample Instance Center by Category ;in, Indicates the first The center of the battery sample instance of the category is in the first The central value of a combination of features. Categories can include health baseline, minor anomaly, and severe anomaly categories, or, depending on business needs, can be set as excellent performance, usable performance, and anomaly performance categories. For each category, the category center is obtained by averaging the combined feature vectors of all old battery sample instances within that category dimension by dimension. . This represents the typical feature distribution of samples of this category in the knowledge base.

[0069] For example, when the first The category is the health benchmark class, and this class is in When it contains 100 old battery sample instances, these 100 samples can be analyzed on the 1st... We obtain the average of the values ​​of the combined features. For all Performing this operation on a combination of features yields... In subsequent step 4.2, new battery samples are added. Will be with various categories of centers Calculate the weighted difference.

[0070] Using category centers instead of comparing each sample with all previous samples can significantly reduce computational cost. The number of old samples is large, and performing full matching directly will increase the response time. Calculating the category center first can quickly determine the overall difference between the new sample and each category, and then decide whether to perform further neighborhood matching or instance-level retrieval.

[0071] Step 4.2, with the first One battery sample to be evaluated As a new battery sample, Equation (5) is used for calculation. With the original knowledge base Middle Center for old battery sample examples Weighted difference between : (5) In equation (5), Indicates the first Selection coefficients for various combinations of features; Indicates the filtered first Combinatorial characteristics. Using combinational selection coefficients. As weights, more important combined features contribute more to the difference. If a feature is determined to have high discriminative power and bypass correlation in step 3, then deviations of that feature will have a more significant impact. Conversely, even slight fluctuations in low-importance features will not excessively affect the degree of difference.

[0072] denominator This is used to normalize the scale of the class centers, preventing the absolute difference from having an unbalanced impact on the results when the value of a certain class center is large. Adding 1 can prevent the denominator from being zero when the class center is 0.

[0073] For example, if The values ​​of the newly added samples on the three combined features are as follows: A certain category center for ,but It can be obtained by adding each item together. If Smaller, indicating Proximity to the center of that category; if Larger, indicating If it differs significantly from this category, it may be a novel or anomalous sample.

[0074] Step 4.3, calculate using equation (6) Minimum difference and determine Recent Category : (6) In equation (6), Represents the original knowledge base The total number of categories; Indicates taking The category number corresponding to the minimum difference is found from all category centers using equation (6). The closest category. The smaller, the more... The more it resembles an existing category in the knowledge base; The larger, the more The results show significant differences from existing categories. This information can be used not only for temporary classification but also as a basis for determining whether to add new knowledge base samples.

[0075] It should be noted that the most recent category This is not necessarily equivalent to the final performance level. The final performance level also requires the calculation of performance deviation in step 5, the comprehensive scoring in step 6, and the threshold determination in step 7. This setup avoids the knowledge base's nearest neighbor relationship solely determining the final evaluation result, thus improving the robustness of the evaluation.

[0076] Step 4.4, Calculation Cluster confidence and time series stability index ; when , and At that time, Combined feature vectors The real category label, timestamp, number of matches, and classification contribution are used as new battery sample instances. Write it into the knowledge base; among which, This indicates the addition of a new judgment threshold. Indicates the confidence threshold. This indicates the timing stability threshold; otherwise, it is not written.

[0077] in, This indicates that the newly added sample differs sufficiently from existing category centers and has the potential to be included in the knowledge base. If the value is too small, it indicates that the sample is highly similar to existing samples in the knowledge base, and writing it may result in duplicate storage, so it is not advisable to add it.

[0078] This indicates that the sample has a high clustering confidence level within its neighborhood. That is, Several newly added samples in the vicinity maintain consistency with their class labels, and samples that are closer in distance contribute more. If the classes within the neighborhood are mixed, then... A low value indicates that the sample may be located at the category boundary or belong to noise samples, and should not be directly written into the knowledge base.

[0079] This indicates that the sample exhibits stable performance within the continuous sampling time window. If If the features fluctuate drastically or the classification results change frequently in a short period of time, it may be due to acquisition noise, transient conditions or temporary occlusion, and it is not suitable to write it into the knowledge base as stable knowledge.

[0080] In practice, formula (14) is used to calculate... Cluster confidence ; (14) In equation (14), For indicator functions, when the first One new battery sample and If the category labels are the same, the indicator function is 1; otherwise, the indicator function is 0. express and The distance between them; In order to be with the first One battery sample to be evaluated Recent The newly added battery samples constitute The neighborhood set, Indicates the first One battery sample to be evaluated With neighborhood set The mean distance between each newly added battery sample in the sample. This indicates a positive number that prevents the denominator from being zero.

[0081] Equation (14) can be understood as the neighborhood confidence score of "category consistency × distance decay". If the neighborhood set... Most samples and If they are of the same category and are close in distance, then High; if a large number of samples in the neighborhood set are If the categories are inconsistent, or if the categories are the same but far apart, then Lower.

[0082] For example, you can take If 4 out of 5 neighboring samples are... If the categories are the same and the distances are all small, then It may reach above 0.60; if only one of the five neighboring samples matches the same class, then Typically lower. The validation set can be set to 0.50, 0.55, or 0.60.

[0083] Calculate using equation (15) Temporal stability index : (15) In equation (15), for In the The standardized first sampling time window Such characteristics for In the The standardized first sampling time window Such characteristics This represents the number of consecutive sampling time windows; For the first Selection coefficients for various combinations of features; For continuous Within each sampling time window The number of times the classification results remain consistent; It is the balance coefficient, and .

[0084] The first term of equation (15) describes the continuous stability of the feature sequence. The smaller the change in combined features between adjacent sampling time windows, the closer the exponential term is to 1; the larger the change, the smaller the exponential term. The second term of equation (15) describes the stability of the classification result. If continuous... If the classification results remain largely consistent across multiple sampling time windows, then... Relatively high.

[0085] In practice, Values ​​of 3, 5, 10, etc., can be used. If the testing station continuously acquires 5 frames of images and 5 sets of electrical parameters for the same battery within a short period of time, then... ; This indicates the number of times the classification results remain consistent across four adjacent time window comparisons. A value of 0.5 can be used, indicating that feature stability and classification stability are equally important; or a value of 0.7 can be used, indicating that more emphasis is placed on the continuous stability of features.

[0086] For example, take If the combined features change little in adjacent time windows, resulting in an index of 0.90, and three out of four adjacent classifications remain consistent, then... .like If the sample meets the time stability requirement, then the sample satisfies the time stability requirement.

[0087] when , and When all three conditions are met, it means It possesses sufficient differences from existing categories, along with reliable category attribution and temporal stability, and therefore can be used as... Write it to the knowledge base. If any condition is not met, then... Used only for this evaluation, not included in the final report. This prevents noisy, duplicate, or unstable samples from contaminating the knowledge base.

[0088] Step 4.5, use equation (7) to calculate the first [item] in the knowledge base. Used battery sample example Elimination Index The elimination index Used to determine whether old samples still have value. If A lower value indicates that the old sample has rarely been called recently; if A lower value indicates that the old sample rarely forms a valid match with the new sample; if A lower value indicates that the old sample did not contribute sufficiently to the correct classification. An increase in the value indicates that the old sample is more suitable for elimination.

[0089] (7) In equation (7), Indicates the first Used battery sample example The recent call frequency normalized value; Indicates the first Used battery sample example Normalized value of the number of matches between the battery sample instance and any newly added battery sample instance; Indicates the first Used battery sample example The classification contribution is calculated using equation (16). Classification contribution : (16) In equation (16), For the first Used battery sample example Number of times it participates in the correct classification for Number of times it participates in the overall category This indicates a positive number that prevents the denominator from being zero.

[0090] Category contribution The higher the value, the more frequently the old sample participated in the correct classification during the historical matching process; The lower the value, the more likely the old sample is to offer limited help in classification, or even interfere with classification. This is calculated using equation (16). This allows for data-driven decision-making in knowledge base elimination, rather than simply deleting items based on chronological order.

[0091] For example, an old sample Participating in general classification recently The correct classification was performed the following times. Next, then This indicates that its contribution is relatively high; another old sample participated in the total classification 80 times, but only participated correctly 20 times, then... This indicates that the sample may not match the current data distribution.

[0092] If an old sample ,but If the threshold is removed If so, then the old sample meets the removal criteria. If another old sample ,but It should obviously be retained.

[0093] Step 4.6, when At that time, the first Remove the old battery sample instance from the knowledge base; otherwise, retain the first one. Used battery sample instances; thus obtaining the updated knowledge base. ;in, Indicates the rejection threshold; The settings can be adjusted based on the knowledge base storage capacity, computation latency, and sample update frequency. For example, to maintain a high level of knowledge base freshness, the settings can be appropriately reduced. If you wish to retain more historical samples to cover more aging states, you can appropriately increase the [specific value]. .

[0094] knowledge base It consists of both the old battery sample instances that have been retained and the new battery sample instances that meet the new conditions. This will be used as the basis for performance matching in step 5. Because Low-contribution old samples have been removed and stable new samples have been absorbed, thus enabling better adaptation to the distribution of the current testing batch and the current battery model.

[0095] Step 5, based on the updated knowledge base Using equation (8) for the first One battery sample to be evaluated Perform performance matching to obtain performance deviation This performance deviation Used to measure the battery sample to be evaluated The degree of deviation from the health baseline class center. (Compared to step 4.2) Unlike other steps, step 5 always involves selecting a health benchmark center. As a reference object, therefore It directly reflects the distance between the sample and the health status.

[0096] (8) In equation (8), This indicates the updated knowledge base. China Health Benchmark Center in the 19th The central value of a combination of features; Indicates the filtered first Combinatorial characteristics; denominator This scale is used to simultaneously consider sample feature values ​​and health baseline center values, avoiding excessive absolute difference effects when certain feature amplitudes are large. Still used as weights, causing the deviation of key combined features from the pair The impact is greater.

[0097] For example, if In terms of combined characteristics such as edge sealing continuity, seepage color difference, and temperature stability, it is close to the healthy benchmark class center. Smaller; if the seal continuity decreases, the suspected leakage area increases, or the temperature is significantly higher than the healthy baseline, then Increase. The subsequent step 6 will... Normalization And it is used in the exponential decay term in the scoring function.

[0098] The benefit of this step is that it transforms the knowledge base update results into quantifiable performance deviation metrics. Since the health benchmark class center comes from the updated knowledge base... Therefore, the health reference can be gradually adjusted with sample batches and equipment environment, reducing misjudgments caused by a fixed threshold remaining unchanged for a long time.

[0099] Step 6, use equation (9) to calculate the first... One battery sample to be evaluated Assessment score : (9) In equation (9), for Model consistency factor; for Surface damage penalty factor, for The exudate penalty factor, for The charge state stability factor, for The normalization result; and ; Equation (9) integrates the four types of evaluation information into an assessment score ranging from 0 to 100. The first part, 100, serves as the score benchmark; This indicates that the greater the performance deviation, the lower the score becomes exponentially. This indicates that both surface damage and leakage will result in penalties; Higher scores indicate higher model consistency and state of charge stability.

[0100] The advantage of using multiplicative fusion is that the final score will drop significantly if any key risk factor is severely abnormal. For example, even if the knowledge base deviation is small, the penalty will be reduced if the surface is severely damaged or leaks. Even if the appearance is good, if the state of charge is unstable or the model designation is inconsistent, the final score will not be too high.

[0101] It can be by It is obtained through linear normalization, quantile normalization, or max-min normalization. For example, in a certain batch, it can be taken as... And truncate the results to It can also be based on historical healthy samples and abnormal samples. The distribution is set with normalization parameters.

[0102] Step 6 involves using equation (10) to calculate... Model consistency factor : (10) In equation (10), for The model code vector of the battery surface marking area. for Reference model template vector for the battery surface marking area; Indicates modulo, This indicates a positive number that prevents the denominator from being zero.

[0103] like Figure 5 As shown, the marking area on the battery surface can be located on the front or side of the battery casing, and is used to read the model, batch, barcode, QR code, production date or specification markings. It can be obtained through character recognition, barcode recognition, or image encoding models; It can be a standard template from the model to be tested.

[0104] Equation (10) is essentially used for calculation and Cosine similarity between them. The closer the value is to 1, the more consistent the surface markings of the tested battery are with the reference model template; The lower the value, the more likely there is a mismatch between the model and the label, wear and tear on the label, contamination of the label, mixing of materials, or failure to identify the material.

[0105] In practical implementation, if the object to be tested must belong to a specific model, then... As a strong constraint. If If the value is below the preset model consistency limit, such as 0.80, then even if other performance indicators are good, manual verification can be prompted to prevent different models of batteries from being mistakenly mixed into the same sorting process.

[0106] Step 6 involves calculating the surface damage penalty factor using equation (11). The surface damage penalty factor It is used to quantify the impact of appearance defects such as shell cracks, dents, scratches, peeling, and edge damage. Reflects the length of cracks or scratches. Reflects the area of ​​the damaged region. It reflects the degree of irregularity of the damage boundary. The more discrete and burr-like the edge, the more complex the damage or the higher the risk.

[0107] (11) In equation (11), for The normalized value of the crack length, for The percentage of damaged area for The dispersion of the damaged edge, , and The range of values ​​is Equation (11) adopts The form makes the penalty less severe for minor damage, but as As the penalty increases, it rises rapidly and gradually approaches 1. This setting highlights the risk of severely damaged samples while avoiding the excessive influence of minor surface noise on the score.

[0108] For example, if ,but This indicates minor damage; if ,but This indicates significant damage and a correspondingly low assessment score.

[0109] Step 6 involves using equation (12) to calculate... Exudate penalty factor : (12) In equation (12), Indicates the first One battery sample to be evaluated The color difference between the exudate area and the normal area is abnormal. for The ratio of the area of ​​the seepage zone, for The boundary diffusion rate.

[0110] like Figure 5 As shown, suspected leakage areas typically exhibit localized color abnormalities, abnormal reflections, boundary smudging, liquid stain diffusion, or increased color difference between the area and the normal shell area. Used to measure the color difference between suspected effusion areas and normal areas; Used to measure the percentage of area suspected of seepage; It is used to measure the extent to which the boundaries of the region have spread outward.

[0111] Equation (12) averages the area and boundary diffusion, then multiplies it by the constant of color difference. The amplified term of the decision. If The area is relatively small; even if there are a few suspected areas, it could be due to lighting or material texture. It won't be too high; if A larger area indicates a significant difference between the suspected effusion area and the normal area; in this case, the larger the area and the greater the diffusion, the better. The higher.

[0112] For example, if ,but This indicates a low risk of leakage; if ,but This indicates that the penalty for leakage is significant.

[0113] Step 6 involves calculating using equations (13-1) and (13-2). Charge state stability factor : (13-1) (13-2) In equations (13-1) and (13-2), for In a single sampling time window The state of charge within, for During the sampling time window Internal operating current, For rated capacity, for The initial state of charge, This represents the reference state of charge value corresponding to the healthy battery sample; Indicates the first One battery sample to be evaluated During the sampling time window Average operating voltage within; Indicates the first One battery sample to be evaluated During the sampling time window The average surface temperature inside; and These represent the reference voltage and reference surface temperature corresponding to the healthy battery sample, respectively. , and There are three stability weights, all of which are not less than 0. This indicates a positive number that prevents the denominator from being zero.

[0114] Equation (13-1) can be used to estimate the sampling time window through Coulomb integration. Internal state of charge When the discharge current is positive, the integral term will decrease. When the charging current direction is defined with opposite signs, the sign can be adjusted according to the specific current direction convention. The rated capacity is used to convert the current integral into a change in state of charge.

[0115] Equation (13-2) will Deviation, voltage deviation, and surface temperature deviation are all included in the stability evaluation. If of , and If all values ​​are close to the healthy reference range, then the negative terms in the index are relatively small. If it is close to 1; if one or more of them deviate significantly, then decline.

[0116] , and This is used to adjust the importance of the three types of deviations. For example, in energy storage battery scenarios, temperature anomalies can directly affect safety risks, and the importance of these deviations can be appropriately increased. In capacity consistency screening scenarios, the efficiency can be appropriately increased. In voltage platform stability screening scenarios, the appropriate increase can be made. .

[0117] For example, take ,like of relatively Deviation of 5%, average voltage relative Deviation of 2%, average surface temperature relative If it deviates by 3%, then This indicates that the state of charge is relatively stable.

[0118] The following is a non-restrictive scoring example to illustrate how equation (9) works. Suppose there is a battery sample to be evaluated. The normalized result of the performance deviation is obtained through step 5. The model consistency factor is obtained through equation (10). The surface damage penalty factor is obtained through equation (11). The percolation penalty factor is obtained by equation (12). The charge state stability factor is obtained through equation (13-2). .

[0119] Substituting the above values ​​into equation (9), we can obtain .like If the sample performs well, then it can be judged to have excellent performance.

[0120] For example, if another battery sample to be evaluated... ,but .like If the sample is deemed abnormal, it should be classified as having performance anomalies. This example illustrates that this method can reduce obviously abnormal samples to an abnormal level through a multi-factor comprehensive penalty.

[0121] Step 7, when At that time, the judgment of the first One battery sample to be evaluated For excellent performance; when At that time, the judgment of the first One battery sample to be evaluated For performance to be available; when At that time, the judgment of the first One battery sample to be evaluated This indicates a performance anomaly; among them, and This indicates two hierarchical thresholds, and .

[0122] like Figure 4 As shown, and The evaluation score from 0 to 100 is divided into three intervals. The excellent performance interval corresponds to direct storage, direct shipment, priority utilization, or continued service; the usable performance interval corresponds to downgraded use, manual review, restricted use, or entry into the observation queue; and the abnormal performance interval corresponds to rejection, repair, isolation, safe disposal, or further experimental testing.

[0123] In one specific implementation, it can be taken as follows: For high-security applications, it can improve... and For example, take For initial screening scenarios, one can take... The threshold can be adjusted based on the actual operational costs of misjudgment, the risk of missed judgment, and the detection objectives.

[0124] The advantage of this step is that the output format is simple and intuitive, making it easy to interface with production line sorting equipment, warehousing systems, or maintenance management systems. The system can write "excellent performance," "available performance," and "abnormal performance" as status codes into the database, or it can output them in the form of labels, colors, audible and visual alarms, or mechanical sorting actions.

[0125] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0126] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0127] To further illustrate the complete implementation process of this method, a series of embodiments are given below. This embodiment is only used to explain the implementation of this technical solution and is not intended to limit the scope of protection.

[0128] This embodiment uses the production quality inspection and recycling sorting scenario of square lithium-ion batteries as an example to illustrate a battery performance evaluation method based on knowledge confidence and contribution. This embodiment is only used to explain the specific implementation process of the present invention and is not intended to limit the scope of protection of the present invention.

[0129] In this embodiment, the detection system includes an image acquisition module, an electrical parameter acquisition module, a temperature acquisition module, a data processing module, and a knowledge base management module. The image acquisition module acquires images of the battery surface; the electrical parameter acquisition module acquires data related to terminal voltage, operating current, estimated internal resistance, and state of charge; the temperature acquisition module acquires the battery surface temperature; the data processing module performs feature standardization, combined feature screening, performance deviation calculation, and scoring; and the knowledge base management module stores historical sample instances and updates the knowledge base based on the value of new and old samples.

[0130] In this embodiment, samples are divided into historically labeled samples, original knowledge base samples, currently evaluated samples, and newly added samples after review. Historically labeled samples are used to form healthy battery sample sets and abnormal battery sample sets, and are used to calculate feature standardization parameters and feature selection parameters; original knowledge base samples are used to construct the original knowledge base and serve as reference data for subsequent category center, health benchmark center, and performance deviation calculation; currently evaluated samples are used to perform performance evaluation and grade determination, and their true category labels are not pre-assumed during initial testing; newly added samples after review refer to samples whose true category labels have been obtained after manual review, experimental testing, or confirmation by business rules, and these samples can participate in knowledge base updates as new samples.

[0131] In this embodiment, the battery sample to be evaluated can be a battery sample from the current testing batch, or a historical testing sample used to establish evaluation parameters, update the knowledge base, or verify evaluation results. For battery samples that have completed testing, manual review, or business rule confirmation, their labels can be health labels or abnormal labels; for battery samples that have not yet completed review during the online testing process, initial labels can be generated based on electrical parameter testing results, image defect detection results, or existing classification models, and then the true category labels can be formed after subsequent manual review, capacity testing, internal resistance testing, or safety testing.

[0132] Therefore, in this invention, "the healthy battery sample set is divided according to the label" and abnormal battery sample set This can be understood as: forming a sample set for feature selection, parameter calculation, and knowledge base update based on labeled samples, rule-determined samples, or samples that have been verified; after obtaining the evaluation results, the currently detected online samples can be verified and confirmed as new battery sample instances to participate in the knowledge base update.

[0133] In this embodiment, 300 historically labeled battery samples are selected as training samples, including 180 healthy samples and 120 abnormal samples. Based on the historical sample labels, the 180 healthy samples form a healthy battery sample set. The 120 abnormal samples were used to form an abnormal battery sample set. For each battery sample, 18 raw features are collected to form the raw feature vector:

[0134] The 18 original features are as follows: This is the terminal voltage, used to represent the terminal voltage during battery testing; This represents the operating current, used to indicate the average operating current within the sampling time window. This is an estimated value of internal resistance, obtained through pulse testing or equivalent estimation; The average surface temperature is used to represent the average surface temperature within the sampling time window. Temperature variance is used to characterize the degree of temperature fluctuation on the battery surface. This is the average brightness value of the casing, used to represent the average image brightness value of the battery casing area; Variance of the casing brightness is used to represent the fluctuation of image brightness in the battery casing area; This represents the length of the sealed edge closed profile, used to indicate the detected length of the sealed edge closed profile. This represents the length of a continuous pixel segment in the edge sealing area, used to indicate the length of a continuous effective pixel segment in the edge sealing region. This is the grayscale transition amount for edge sealing, used to represent the grayscale difference between the edge sealing area and the neighboring substrate; Crack length, used to represent the length or equivalent length of the crack region skeleton; This represents the percentage of damaged area, used to indicate the ratio of the damaged area to the battery surface area. This is the dispersion of the damage edge, used to represent the degree of irregularity of the damage boundary; This is the area ratio of the suspected leakage area to the battery surface area. This is the seepage boundary diffusion rate, used to indicate the degree of diffusion at the boundary of a suspected seepage area; The similarity features of the model code vector are identified from the marking area on the battery surface. The estimated state of charge is obtained from the current integral and the initial state of charge. This is a voltage stability characteristic used to represent voltage fluctuations within a sampling time window.

[0135] During image acquisition, the battery's outer contour is first located, establishing a coordinate system for the battery body. The battery's long side is used as the horizontal coordinate, and the short side as the vertical coordinate. Then, based on the relative position of the battery's outer contour, the system is divided into surface identification region, edge sealing region, neighboring base region, potential damage region, and potential leakage region. The edge sealing region is used to extract the length of the closed edge contour, the length of continuous pixel segments, and the grayscale transition amount. The surface identification region is used to extract the model code vector. The potential damage region is used to extract crack length, damage area ratio, and damage edge dispersion. The potential leakage region is used to extract the area ratio of suspected leakage areas and the leakage boundary diffusion rate.

[0136] To avoid the impact of differences in the dimensions and numerical ranges of different features on the evaluation results, this embodiment uses historically labeled samples to calculate standardized parameters, rather than using the current sample to be evaluated. Specifically, for the first... The original features were used to calculate the mean based on 300 historically labeled samples. and standard deviation and will and Stored in the standardized parameter table. For the current sample to be evaluated. The primitive characteristics Standardize to obtain :

[0137] in, To prevent positive numbers with a denominator of zero, this embodiment takes... For example, for a given sample to be evaluated... Its average surface temperature is If the mean surface temperature of the historically labeled samples is... The standard deviation is The normalized result of the average surface temperature of this sample is:

[0138] Following the above method, all 18 original features were standardized to obtain the current sample to be evaluated. Standardized feature vectors .

[0139] This embodiment uses both the filtering discriminant value and the edge-sealing structure correlation value to determine the combination selection coefficient, thereby filtering the combination feature subset. First, based on the historical labeled sample set... and Calculate the filtering discriminant value for each standardized feature between healthy and abnormal samples. The larger the filtering discriminant value, the better the feature can distinguish between healthy and abnormal samples. Secondly, the edge-sealing structure correlation value is calculated for the edge-sealing related image features. The correlation value of the edge sealing structure is determined by the closure ratio of the edge sealing area. Continuity of the sealing area and grayscale transition amount The calculations yielded the following results. For electrical performance characteristics that do not directly correspond to the edge-sealing area, their edge-sealing structure association value can be set to 0, or assigned a lower baseline value according to a preset mapping rule. In this embodiment, to maintain calculation consistency, the edge-sealing structure association value for non-edge-sealed image features is set to 0.

[0140] Specifically, in this embodiment, the closure ratio of the edge sealing area The ratio of the detected closed edge profile length to the theoretical edge profile length is obtained; the continuity ratio of the edge sealing area is... The grayscale transition amount is obtained by the ratio of the length of the continuous effective pixel segment to the total edge length. It is obtained by normalizing the grayscale difference between the edge-sealed area and the neighboring base area. Then, it is based on the filtering discriminant value. Related values ​​of edge banding structure Calculate the combination selection coefficient In this embodiment, the combined balance coefficient Setting it to 0.7 emphasizes the ability to distinguish between healthy and abnormal samples while preserving the association information of the edge-sealing structure. A combined selection threshold is then set. Select the one that satisfies Features constitute a combination of feature subsets .

[0141] In this embodiment, 12 combined features were obtained after screening, specifically: correspond That is, the terminal voltage; correspond That is, the estimated internal resistance value; correspond , i.e., average surface temperature; correspond That is, temperature variance; correspond That is, the length of the closed edge banding profile; correspond That is, the length of the continuous pixel segment of the edge sealing; correspond That is, the grayscale transition amount of the edge sealing; correspond That is, the crack length; correspond That is, the percentage of the damaged area; correspond , that is, the area ratio of the suspected seepage area; correspond , i.e., the permeation boundary diffusion rate; correspond This refers to voltage stability characteristics.

[0142] Therefore, the current sample to be evaluated The combined feature vector is:

[0143] Through the above processing, the combined feature subset includes both electrical performance characteristics such as voltage, internal resistance, and temperature, and appearance safety characteristics such as edge sealing, damage, and leakage.

[0144] In this embodiment, the original knowledge base Includes 240 used battery sample instances. Each used battery sample instance This includes the combined feature vector, true class label, timestamp, recent call frequency, number of matches, and classification contribution, specifically represented as follows:

[0145] in, This is a combined feature vector of old battery sample instances. For real category labels, The timestamp for writing the sample to the knowledge base. This is a normalized value based on recent call frequency. This is a normalized value representing the number of matches between the new sample instance and the new sample instance. Contribution to the classification.

[0146] In this embodiment, the samples in the knowledge base are divided into three categories: categories For the health baseline class, the sample size is 120; category The category is minor anomaly, with 80 samples; [Category] This is a severe exception class, with 40 samples. Recent call frequency. The number of times each sample was invoked within the last 30 days was counted, and then normalized by dividing by the maximum number of times a single sample was invoked within the last 30 days. (Normalized match count value) The number of matches between the old sample and the new candidate sample is counted, and then normalized by dividing by the maximum number of matches for old samples in the current knowledge base. (Classification Contribution) Calculated based on the ratio of the number of times the old sample participated in the correct classification to the total number of classifications:

[0147] in, This indicates the number of times the old sample participated in the correct classification. This indicates the total number of times the old sample participated in the classification.

[0148] Based on the original knowledge base Calculate the category centers for the healthy baseline class, the minor abnormality class, and the severe abnormality class, respectively. For the first... The category center is obtained by averaging the combined feature vectors of all old battery sample instances within that category across all dimensions.

[0149] in, Indicates the first Category in The central value of a combination of features. For example, in this embodiment, the central value of the health benchmark class. The partial feature center values ​​are: The corresponding center value is 0.02. The corresponding center value is -0.10. The corresponding center value is -0.15. The corresponding center value is -0.08. The corresponding center value is 0.12. The corresponding center value is 0.10. The corresponding center value is -0.05. The corresponding center value is -0.30. The corresponding center value is -0.28. The corresponding center value is -0.35. The corresponding center value is -0.32. The corresponding center value is -0.12. The above center value is used for subsequent performance deviation calculations.

[0150] Based on the current sample to be evaluated For example, the original detection data collected by the detection system is as follows: terminal voltage is 3.71V, estimated internal resistance is 42mΩ, average surface temperature is 31.0℃, temperature variance is 1.8, sealing edge closed contour length is 1860 pixels, sealing edge continuous pixel segment length is 1710 pixels, sealing edge grayscale transition is 0.36, crack length normalized value is 0.08, damaged area ratio is 0.04, suspected seepage area ratio is 0.02, seepage boundary diffusion rate is 0.03, and voltage stability feature is 0.12. After standardization and combined feature screening, the following data is obtained: Combined feature vectors :

[0151] In this embodiment, the current sample to be evaluated During the online testing phase, it is only used for scoring and rating, and is not directly written to the knowledge base. If the true category label is subsequently confirmed through manual verification or experimental testing, then... As newly added candidate samples after review, a knowledge base write decision is executed. Calculate the weighted difference between each category center and the original knowledge base. In this embodiment, With health benchmarks The weighted difference between them was 0.42, compared to the minor anomaly class. The weighted difference between them was 0.28, compared to the severe anomaly class. The weighted difference between them is 0.73. Therefore, it can be concluded that... The minimum degree of difference is: Its most recent category is the minor anomaly category. In this embodiment, a new judgment threshold is added. Set the confidence threshold to 0.25. Set the threshold to 0.55 for time series stability. Take 0.70. Further calculations. Cluster confidence and temporal stability indices. Let's assume... The five most recent new candidate samples constitute the neighborhood set. Four of the samples were related to If the review categories are consistent and the distances are close, then the calculation is as follows: .

[0152] right Data is collected continuously for 5 sampling time windows. If the combined features fluctuate little within adjacent time windows, and 3 out of 4 adjacent classification results are consistent, then the following can be calculated: ; because , ,and ,therefore The conditions for adding a new entry have been met. The knowledge base management module will... The combined feature vector, verified true category label, timestamp, initial matching count, and initial classification contribution are written into the knowledge base as new battery sample instances. The initial matching count for a new sample can be set to 0, and the initial classification contribution can be set to 1, or set to preset initial values ​​according to the verification results. In this embodiment, the initial matching count is set to 0, and the initial classification contribution is set to 1.

[0153] After adding new samples, the knowledge base management module updates the original knowledge base. The obsolescence index is calculated using old samples. An example of an old battery sample is used. For example, its normalized call frequency value in the last 30 days is The normalized value of the number of matches between it and the newly added samples is Its classification contribution is Then its elimination index is: ; Substituting the values, we get: ; The calculation yielded: .

[0154] In this embodiment, the threshold for elimination is... Take 0.60. Because... Therefore, old battery sample examples Remove it from the knowledge base. Then use an old battery sample as an example. For example, its recent call frequency normalization value is 0.75, its match count normalization value is 0.60, and its classification contribution is 0.85. Therefore:

[0155] because Therefore, old battery sample instances are retained. After adding new samples and removing old samples, the knowledge base management module recalculates the category centers to obtain the updated knowledge base. .

[0156] For the current sample to be evaluated Based on the updated knowledge base The health benchmark class center is used to calculate its performance deviation. In this embodiment, the calculation is as follows: The performance deviation is: ; To facilitate scoring, Normalize to the [0,1] interval. Let the reference maximum value of performance deviation in historical samples be 1.20, then: ;therefore, The normalized performance deviation is 0.30.

[0157] right The battery surface marking area is identified to obtain the model code vector. .Will With reference model template vector Cosine similarity calculation is performed to obtain the model consistency factor. In this embodiment, the calculation is as follows: This result indicates that the current sample to be evaluated... The model identifier is basically consistent with the reference model template. To prevent the model consistency factor from exceeding the scoring requirements, this embodiment limits the model consistency factor to the range of [0,1]. When the cosine similarity calculation result is less than 0, it is set to 0; when the calculation result is greater than 1, it is set to 1.

[0158] Based on the image detection results, The normalized value of the crack length is The percentage of damaged area is The dispersion of the damaged edge is The surface damage penalty factor is... for:

[0159] Substituting the values, we get:

[0160] Further results were obtained:

[0161] illustrate There is slight surface damage, but the degree of damage is not high.

[0162] Based on the detection results of the suspected seepage area The ratio of the area of ​​the seepage zone is Boundary diffusion rate The color difference between the exudate area and the normal area is abnormal. Then the exudate penalty factor for:

[0163] Substituting the values, we get:

[0164] The calculation yielded: ;illustrate There is virtually no risk of significant leakage.

[0165] During the sampling time window Inside, to By integrating the current, we can obtain its state of charge. .set up The initial state of charge is 0.82, the rated capacity is 50Ah, and the change in state of charge obtained by current integration within the sampling time window is 0.02. Therefore:

[0166] Let the reference state of charge for the healthy battery sample be... It is 0.82. Average operating voltage within the sampling time window It is 3.71V, the healthy reference voltage. It is 3.75V; average surface temperature The healthy reference surface temperature is 31.0℃. The temperature is 28.0℃. In this embodiment, the stability weight is:

[0167] Substituting into the formula for calculating the state-of-charge stability factor, we get:

[0168] The calculation yielded: ;illustrate Its state of charge, voltage, and temperature are generally stable.

[0169] Based on the above calculation results, , , , , Substitute the above factors into the scoring formula:

[0170] Substituting the specific values, we get:

[0171] The calculation yielded: ; In this embodiment, a first grading threshold is set. Second-level threshold .because Therefore, the current sample to be evaluated The sample was determined to be "performance-available." The system also output the main influencing factors for this sample, including: moderate performance deviation, minor surface damage, low risk of leakage, high model consistency, and high state-of-charge stability. This result can be used in subsequent sorting processes. It is classified as a battery that can be downgraded or requires further review.

[0172] To further illustrate the process of identifying anomalous samples using this method, another sample currently to be evaluated is selected. Conduct testing. Image inspection revealed poor edge sealing continuity, with large damaged areas and areas suspected of leakage; electrical parameter testing showed high internal resistance and significant surface temperature fluctuations. Calculations showed... The normalization performance deviation is Model consistency factor is The surface damage penalty factor is The exudate penalty factor is The charge state stability factor is Substitute into the scoring formula:

[0173] The calculation yielded: ;because Therefore, The sample was determined to be "performance abnormal." The main causes of the abnormality output by the system were: high performance deviation, high surface damage penalty, high leakage penalty, and decreased state of charge stability. For this sample, the system can further trigger isolation, manual re-inspection, or safety handling procedures.

[0174] In this embodiment, the knowledge base management module is maintained in batches or on a periodic basis. After each batch of battery testing is completed, the system counts the number of times each old battery sample instance is called in this batch of testing, counts the number of times each old battery sample instance participates in the correct classification, updates the recent call frequency normalization value, the matching number normalization value, and the classification contribution, performs a new write judgment on newly added samples with verified and confirmed labels, calculates the elimination index for old samples, deletes or logically archives old samples whose elimination index exceeds the threshold, recalculates the category centers of the healthy baseline class, the slightly abnormal class, and the seriously abnormal class, and uses the updated category centers for the performance deviation calculation of the next batch of samples to be evaluated.

[0175] The removal operation can be performed through physical deletion, logical deletion, or archiving. This embodiment uses logical deletion, marking eliminated samples as invalid samples, preventing them from participating in category center calculations and performance matching, but retaining their historical records for quality traceability.

[0176] Through the above implementation process, this embodiment can utilize historically labeled samples to calculate standardized parameters, avoiding instability in standardized results due to insufficient current sample quantity or batch changes; it can filter combined features by combining the filtering discriminant value and the sealing structure correlation value, so that the screening results can reflect the category differences between healthy and abnormal samples, while retaining safety-related information such as sealing structure, damage, and leakage; it can adjust the knowledge base according to changes in battery batch, model status, and aging status by adding new samples and removing low-value old samples, reducing matching deviations caused by long-term use of a fixed knowledge base; it can also calculate the evaluation score by combining performance deviation, model consistency factor, surface damage penalty factor, leakage penalty factor, and state of charge stability factor, so that battery performance level determination simultaneously considers health deviation, model matching, appearance safety, and operational stability.

Claims

1. A battery performance evaluation method based on knowledge confidence and contribution, characterized in that, Includes the following steps: Step 1, collect the first One battery sample to be evaluated The original feature vector ,in, Indicates the first One battery sample to be evaluated The primitive characteristics, , Represents the total number of original features. ; This indicates the total number of battery samples to be evaluated; Based on the label of each battery sample to be evaluated, The battery samples to be evaluated were divided into a healthy battery sample set. and abnormal battery sample set ; Step 2, use equation (1) to... The first in primitive characteristics After standardization, the first... One battery sample to be evaluated Standardized Type of characteristics Thus, the first One battery sample to be evaluated Standardized feature vectors : (1) In equation (1), This indicates a positive number that prevents the denominator from being zero. Indicates the first primitive characteristics in The mean of the battery samples to be evaluated; Indicates the first primitive characteristics in One battery sample to be evaluated The standard deviation in; Step 3, based on the filtering feature selection method and the bypass feature selection method, for By performing combination filtering, the first... One battery sample to be evaluated Filtered combined feature vectors Thus, a subset of combined features is obtained. ; Step 3.1, use equation (2) to calculate the standardized first... Filtering discriminant value of a feature : (2) In equation (2), Represents the standardized first Features in healthy battery sample set The mean of the values; Represents the standardized first Features in the abnormal battery sample set The mean of the values; Represents the standardized first Features in healthy battery sample set The standard deviation in; Represents the standardized first Features in the abnormal battery sample set The standard deviation in; Step 3.2, calculate the standardized first step using equation (3). Correlation values ​​of various features : (3) In equation (3), Represents the standardized first The closure ratio of the battery sealing area corresponding to this feature; Represents the standardized first The continuity ratio of the battery sealing area corresponding to this feature; Represents the standardized first The grayscale transition between the edge-sealed region and the neighboring substrate corresponding to this feature; Step 3.3, use equation (4) to obtain the standardized first... Combination selection coefficient of features : (4) In equation (4), Represents the portfolio balance coefficient. ; Step 3.4, Set the combined selection threshold Select the one that satisfies Features constitute a combination of feature subsets ;in, Indicates the filtered first Combinatorial characteristics; This represents the total number of combined features after filtering. Based on the combined feature subset , obtained the One battery sample to be evaluated Filtered combined feature vectors ; Step 4, construct the original knowledge base ,include: Individual old battery sample instances, and based on combined feature subsets For the original knowledge base The knowledge base is updated by adding new battery sample instances and removing some old battery sample instances. Among them, the first Used battery sample example This includes: combined feature vector, true class label, timestamp, recent call frequency, number of matches, and classification contribution; ; Step 4.1, based on the original knowledge base Examples of old batteries of the same category in the original knowledge base The middle belongs to the first The mean of the combined feature vectors of the old battery sample instances of the category is calculated to obtain the first... Battery Sample Instance Center by Category ;in, Indicates the first The center of the battery sample instance of the category is in the first The central value of a combination of features; Step 4.2, with the first One battery sample to be evaluated As a new battery sample, Equation (5) is used for calculation. With the original knowledge base Middle Center for old battery sample examples Weighted difference between : (5) In equation (5), Indicates the first Selection coefficients for various combinations of features; Indicates the filtered first Combinatorial characteristics; Step 4.3, calculate using equation (6) Minimum difference and determine Recent Category : (6) In equation (6), Represents the original knowledge base The total number of categories; Indicates taking The category number corresponding to the minimum difference value; Step 4.4, Calculation Cluster confidence and time series stability index ; when , and At that time, Combined feature vectors The real category label, timestamp, number of matches, and classification contribution are used as new battery sample instances. Write it into the knowledge base; among which, This indicates the addition of a new judgment threshold. Indicates the confidence threshold. Indicates the timing stability threshold; otherwise, do not write. Step 4.5, use equation (7) to calculate the first [item] in the knowledge base. Used battery sample example Elimination Index : (7) In equation (7), Indicates the first Used battery sample example The recent call frequency normalized value; Indicates the first Used battery sample example Normalized value of the number of matches between the battery sample instance and any newly added battery sample instance; Indicates the first Used battery sample example The contribution of the category; Step 4.6, when At that time, the first Remove the old battery sample instance from the knowledge base; otherwise, retain the first one. Used battery sample instances; thus obtaining the updated knowledge base. ;in, Indicates the rejection threshold; Step 5, based on the updated knowledge base Using equation (8) for the first One battery sample to be evaluated Perform performance matching to obtain performance deviation : (8) In equation (8), This indicates the updated knowledge base. China Health Benchmark Center in the 19th The central value of a combination of features; Indicates the filtered first Combinatorial characteristics; Step 6, use equation (9) to calculate the first... One battery sample to be evaluated Assessment score : (9) In equation (9), for Model consistency factor; for Surface damage penalty factor, for The exudate penalty factor, for The charge state stability factor, for The normalization result; and ; Step 7, when At that time, the judgment of the first One battery sample to be evaluated For excellent performance; when At that time, the judgment of the first One battery sample to be evaluated For performance to be available; when At that time, the judgment of the first One battery sample to be evaluated This indicates a performance anomaly; among them, and This indicates two hierarchical thresholds, and .

2. The battery performance evaluation method based on knowledge confidence and contribution as described in claim 1, characterized in that, In step 6, equation (10) is used to calculate... Model consistency factor : (10) In equation (10), for The model code vector of the battery surface marking area. for Reference model template vector for the battery surface marking area; Indicates modulo, This indicates a positive number that prevents the denominator from being zero.

3. The battery performance evaluation method based on knowledge confidence and contribution as described in claim 1, characterized in that, In step 6, the surface damage penalty factor is calculated using equation (11). : (11) In equation (11), for The normalized value of the crack length, for The percentage of damaged area for The dispersion of the damaged edge, , and The range of values ​​is .

4. The battery performance evaluation method based on knowledge confidence and contribution as described in claim 1, characterized in that, In step 6, equation (12) is used to calculate... Exudate penalty factor : (12) In equation (12), Indicates the first One battery sample to be evaluated The color difference between the exudate area and the normal area is abnormal. for The ratio of the area of ​​the seepage zone, for The boundary diffusion rate.

5. The battery performance evaluation method based on knowledge confidence and contribution as described in claim 1, characterized in that, Step 6 involves calculating using equations (13-1) and (13-2). Charge state stability factor : (13-1) (13-2) In equations (13-1) and (13-2), for In a single sampling time window The state of charge within, for During the sampling time window Internal operating current, For rated capacity, for The initial state of charge, This represents the reference state of charge value corresponding to the healthy battery sample; Indicates the first One battery sample to be evaluated During the sampling time window Average operating voltage within; Indicates the first One battery sample to be evaluated During the sampling time window The average surface temperature inside; and These represent the reference voltage and reference surface temperature corresponding to the healthy battery sample, respectively. , and There are three stability weights, all of which are not less than 0. This indicates a positive number that prevents the denominator from being zero.

6. The battery performance evaluation method based on knowledge confidence and contribution as described in claim 1, characterized in that, The battery sealing area closure ratio in step 3.2 Through the standardized first The ratio of the closed contour length of the battery sealing area corresponding to each feature to the theoretical contour length is obtained; Battery sealing area continuity ratio Through the standardized first The ratio of the length of the continuous effective pixel segment corresponding to each feature to the total edge length is obtained; Gray-scale transition amount Through the standardized first The grayscale difference between the edge-sealing region corresponding to each feature and the neighboring substrate is obtained by normalization.

7. The battery performance evaluation method based on knowledge confidence and contribution as described in claim 1, characterized in that, In step 4.4, equation (14) is used for calculation. Cluster confidence ; (14) In equation (14), For indicator functions, when the first One new battery sample and If the category labels are the same, the indicator function is 1; otherwise, the indicator function is 0. express and The distance between them; In order to be with the first One battery sample to be evaluated Recent The newly added battery samples constitute The neighborhood set, Indicates the first One battery sample to be evaluated With neighborhood set The mean distance between each newly added battery sample in the sample. This indicates a positive number that prevents the denominator from being zero.

8. The battery performance evaluation method based on knowledge confidence and contribution as described in claim 1, characterized in that, In step 4.4, equation (15) is used for calculation. Temporal stability index : (15) In equation (15), for In the The standardized first sampling time window Such characteristics for In the The standardized first sampling time window Such characteristics This represents the number of consecutive sampling time windows; For the first Selection coefficients for various combinations of features; For continuous Within each sampling time window The number of times the classification results remain consistent; It is the balance coefficient, and .

9. The battery performance evaluation method based on knowledge confidence and contribution as described in claim 1, characterized in that, In step 4.5, equation (16) is used for calculation. Classification contribution : (16) In equation (16), For the first Used battery sample example Number of times it participates in the correct classification for Number of times it participates in the overall category This indicates a positive number that prevents the denominator from being zero.

10. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports a processor in executing the method of any one of claims 1-9, the processor being configured to execute the program stored in the memory.

Images