Battery failure prediction method, device, storage medium and program product

Abstract

The present disclosure provides a battery failure prediction method, device, storage medium and program product, which relate to the field of battery technology. The battery failure prediction method includes: obtaining cycle test data of the battery; calculating the cycle test data of the battery to obtain multiple dimensional curves, the multiple dimensional curves including the SOH curve; obtaining the number of first alarm points existing in the multiple dimensional curves, wherein, in each dimensional curve, each occurrence of a feature exceeding the feature threshold of the dimensional curve is recorded as the existence of a first alarm point; performing inflection point detection on the SOH curve to obtain the number of inflection points in the SOH curve; and determining whether the battery has failed based on the number of first alarm points and the number of inflection points. The present disclosure improves the accuracy of battery failure prediction.

Description

Technical Field

[0001] The present disclosure relates to the field of battery technology, and in particular to a battery failure prediction method, device, storage medium, and program product. Background Art

[0002] In recent years, with the rapid development of the new energy vehicle and energy storage industries, the performance stability and reliability of new energy batteries have received increasing attention. Battery safety is a key concern for consumers, directly impacting product usage and market acceptance. In practical applications, battery safety is effectively guaranteed through rigorous verification and design. However, during the R&D phase, battery design is still immature, and various failure modes may occur during the exploration process, leading to safety issues. These failure modes include lithium dendrites piercing the separator, lithium powder growth on the anode or separator, and lithium overhang (tab protrusion) loss. These can cause internal short circuits and capacity degradation during battery testing, seriously impacting battery life and safety. In related technologies, the accuracy of battery failure prediction is relatively low. Summary of the Invention

[0003] One of the purposes of the present disclosure is to improve the accuracy of battery failure prediction.

[0004] According to one aspect of the present disclosure, a battery failure prediction method is provided, comprising: obtaining cycle test data of a battery; calculating the cycle test data of the battery to obtain a plurality of dimensional curves, the plurality of dimensional curves including a battery state of health (SOH) curve; obtaining the number of first alarm points existing in the plurality of dimensional curves, wherein, in each dimensional curve, each occurrence of a feature exceeding a feature threshold of the dimensional curve is recorded as the existence of a first alarm point; performing inflection point detection on the SOH curve to obtain the number of inflection points in the SOH curve; and determining whether the battery has a failure condition based on the number of the first alarm points and the number of the inflection points.

[0005] In the technical solution of the embodiment of the present disclosure, multiple dimensional curves are obtained by calculating the cycle test data of the battery, and the first alarm point and its number are obtained by using the features in the multiple curves. The number of inflection points in the SOH curve is obtained by performing inflection point detection on the SOH curve, and then it is determined whether the battery has failed based on the number of first alarm points and the number of inflection points. In this way, the prediction of battery failure is achieved, and the safety failure of the battery is reflected by the first alarm point and its number, and the performance failure of the battery is reflected by the inflection point and its number. Therefore, the above method can predict battery failure more comprehensively, thereby improving the accuracy of battery failure prediction.

[0006] In some embodiments, determining whether a battery failure exists based on the number of the first alarm points and the number of the inflection points includes: performing a weighted sum calculation on the number of the first alarm points and the number of the inflection points to obtain a first weighted result; and determining that the battery failure exists when the first weighted result is greater than or equal to a first threshold. This achieves the purpose of combining the number of the first alarm points and the number of the inflection points through weighted summation to determine whether the battery failure exists, thereby improving the accuracy of battery failure prediction.

[0007] In some embodiments, the first weighted result C r1 for , where x is the number of first alarm points, α is the weight corresponding to the number of first alarm points, z is the number of inflection points, and δ is the weight corresponding to the number of inflection points. In this formula, the number of first alarm points is squared and then weighted together with the number of inflection points. This square operation increases the importance of the first alarm points, thereby improving the accuracy of battery failure prediction.

[0008] In some embodiments, δ is equal to the first threshold. Since the number of inflection points is 0 or 1, as long as an inflection point exists, the first weighted result can be greater than or equal to the first threshold, indicating that the battery has failed. In other words, the presence of a performance failure can be determined, which is more accurate than actual conditions and improves the accuracy of battery failure prediction.

[0009] In some embodiments, the battery failure prediction method further includes: extracting features from multiple dimensional curves of the battery to obtain extracted features; obtaining the battery's test conditions, process measurement values, and material code as design features of the battery, wherein the material code corresponds to the battery's recipe information; combining the extracted features with the design features to form a combined feature; and inputting the combined feature into a pre-trained prediction model, performing failure prediction on the battery using the prediction model to obtain the battery's failure mode. This implements a method for predicting battery failure using a prediction model, which can predict the specific mode of battery failure and further improve the accuracy of battery failure prediction.

[0010] In some embodiments, performing failure prediction on the battery using the prediction model to obtain the failure mode of the battery includes: performing failure prediction on the battery using the prediction model to obtain the number of failure modes of the battery as the number of second alarm points, wherein each predicted failure mode of the battery is recorded as a second alarm point. This achieves statistics on the number of battery failure modes.

[0011] In some embodiments, determining whether a battery failure exists based on the number of the first alarm points and the number of the inflection points includes determining whether a battery failure exists based on the number of the first alarm points, the number of the inflection points, and the number of the second alarm points. In this embodiment, the number of second alarm points predicted using the prediction model is combined with the number of the first alarm points and the number of inflection points described above to determine the battery failure, thereby further improving the accuracy of battery failure prediction.

[0012] In some embodiments, determining whether a battery failure exists based on the number of first alarm points, the number of inflection points, and the number of second alarm points includes: performing a weighted sum calculation on the number of first alarm points, the number of inflection points, and the number of second alarm points to obtain a second weighted result; and determining that the battery has failed if the second weighted result is greater than or equal to a second threshold. This achieves the purpose of determining whether a battery failure exists by combining the number of first alarm points, the number of inflection points, and the number of second alarm points through weighted summation, thereby improving the accuracy of battery failure prediction.

[0013] In some embodiments, the second weighted result C r2 for , where x is the number of first alarm points, α is the weight corresponding to the number of first alarm points, z is the number of inflection points, δ is the weight corresponding to the number of inflection points, y is the number of second alarm points, and β is the weight corresponding to the number of second alarm points. In this formula, the number of first alarm points is squared and then weighted together with the number of inflection points and the number of second alarm points. This square operation increases the importance of the first alarm point, thereby improving the accuracy of battery failure prediction.

[0014] In some embodiments, δ is equal to the second threshold. Since the number of inflection points is 0 or 1, as long as an inflection point exists, the second weighted result can be greater than or equal to the second threshold, indicating that the battery has failed. In other words, the presence of a performance failure can be used to determine that the battery has failed. This is more accurate and improves the accuracy of battery failure prediction.

[0015] In some embodiments, the battery failure prediction method further includes: training the prediction model before performing failure prediction on the battery using the prediction model. By training the prediction model, the trained prediction model can be used to easily predict the failure mode of the battery.

[0016] In some embodiments, training the prediction model includes: grading different failure modes of the battery based on the severity of the failure mode of the battery to obtain the level of each failure mode; calculating the cycle test data of multiple batteries used for training to obtain multiple dimensional curves of the multiple batteries; performing statistics on the cycle test data under each failure mode, and selecting a failure mode with a uniform data volume and failure mode level for modeling and prediction; obtaining extracted features of each dimensional curve of the battery with the selected failure mode based on the multiple dimensional curves of the battery with the selected failure mode; obtaining test conditions, process measurement values, and material codes of the battery with the selected failure mode as design features of the battery with the selected failure mode, wherein the material code corresponds to the recipe information of the battery with the selected failure mode; combining the extracted features with the design features to obtain combined features of the battery with the selected failure mode; and training the prediction model based on the combined features of the battery with the selected failure mode and the label value of the selected failure mode. This achieves the purpose of training the prediction model and improves the accuracy of predicting the failure mode of the battery.

[0017] In some embodiments, training the prediction model further includes: in the process of obtaining the multiple dimensional curves of the multiple batteries, filling the empty values ​​in each dimensional curve by linear interpolation, and smoothing the cycle test data of the multiple batteries by a median average algorithm with a predetermined window length. Here, by interpolating and smoothing the curves, the adverse effects of noisy data can be reduced.

[0018] In some embodiments, the multiple dimensional curves further include at least one of a DC internal resistance curve, a charging time curve, a coulombic efficiency curve, a peak change curve of a single-cycle capacity increment curve, and a variance curve of the difference between single-cycle capacity curves. This provides multiple dimensional curve types.

[0019] In some embodiments, performing inflection point detection on the SOH curve includes: generating a predicted straight line based on a first discrete point corresponding to the first cycle test and a current discrete point corresponding to the current cycle test in the SOH curve, the predicted straight line being a line segment connecting the first discrete point and the current discrete point; obtaining the midpoint of the line segment, and based on the midpoint, obtaining a first portion of discrete points corresponding to the line segment on the left side of the midpoint and a second portion of discrete points corresponding to the line segment on the right side of the midpoint; calculating the absolute value of a first difference between the ordinate value of each discrete point in the first portion of discrete points and the ordinate value of a corresponding point on the predicted straight line corresponding to each discrete point, and calculating a first sum of the absolute values ​​of all first differences corresponding to the first portion of discrete points; calculating the absolute value of a second difference between the ordinate value of each discrete point in the second portion of discrete points and the ordinate value of a corresponding point on the predicted straight line corresponding to each discrete point, and calculating a second sum of the absolute values ​​of all second differences corresponding to the second portion of discrete points; when the second sum is greater than the sum of the first sum and a third threshold value, determining that the current discrete point is an inflection point. This achieves the purpose of detecting the inflection point of the SOH curve, and can detect the inflection point of the SOH curve more accurately.

[0020] According to another aspect of the present disclosure, a battery failure prediction device is provided, comprising: a memory; and a processor coupled to the memory, the processor being configured to execute the aforementioned battery failure prediction method based on instructions stored in the memory. This improves the accuracy of battery failure prediction.

[0021] According to another aspect of the present disclosure, a computer-readable storage medium is provided, on which computer instructions are stored. When the computer instructions are executed by a processor, the battery failure prediction method described above is implemented, thereby improving the accuracy of battery failure prediction.

[0022] According to another aspect of the present disclosure, a computer program product is provided, which includes a computer program or instructions, and when executed by a processor, the computer program or instructions implements the battery failure prediction method as described above, thereby improving the accuracy of battery failure prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the drawings without creative work.

[0024] Figure 1is a flowchart illustrating a battery failure prediction method according to some embodiments of the present disclosure;

[0025] Figure 2 is a schematic diagram illustrating inflection point detection on an SOH curve according to some embodiments of the present disclosure;

[0026] Figure 3 is a flowchart illustrating a method for predicting a failure mode of a battery using a prediction model according to some embodiments of the present disclosure;

[0027] Figure 4 is a flowchart illustrating a method for training a prediction model according to some embodiments of the present disclosure;

[0028] Figure 5 is a flowchart illustrating a battery failure prediction method according to other embodiments of the present disclosure;

[0029] Figure 6 is a schematic diagram illustrating sliding on an extracted feature using a predetermined window according to some embodiments of the present disclosure;

[0030] Figure 7 is a block diagram schematically illustrating a structure of a battery failure prediction device according to some embodiments of the present disclosure;

[0031] Figure 8 is a block diagram schematically illustrating a structure of a battery failure prediction device according to other embodiments of the present disclosure.

[0032] In the drawings, the drawings are not drawn to scale. DETAILED DESCRIPTION

[0033] The following detailed description of the embodiments of the present application is provided in conjunction with the accompanying drawings and examples. The following detailed description of the embodiments and the accompanying drawings are used to illustrate the principles of the present application, but are not intended to limit the scope of the present application, that is, the present application is not limited to the described embodiments.

[0034] Before describing the method of the embodiment of the present disclosure, the terms that may be involved in the embodiment of the present disclosure are explained as follows.

[0035] (1) Failure Mode: Failure modes refer to various types of failures that may occur during battery use. For example, failure modes include performance failure and safety failure. The method of the disclosed embodiment can predict safety failure, and the inflection point detection algorithm can be used to determine performance failure.

[0036] (2) Rule algorithm: In the embodiments of the present disclosure, the rule algorithm may refer to an empirical rule algorithm based on multiple electrochemical signals designed using a data analysis method.

[0037] (3) AI (Artificial Intelligence) algorithm: In the embodiments of the present disclosure, AI algorithm may refer to a machine learning algorithm, including a deep learning algorithm.

[0038] (4) Inflection point detection algorithm: In the embodiments of the present disclosure, the inflection point detection algorithm may refer to a binary cutting method designed based on the change trend of the capacity decay curve (ie, SOH curve) and empirical rules.

[0039] (5) Lithium dendrites: Dendritic lithium deposits formed on the surface of the negative electrode. The growth of lithium dendrites may cause internal short circuits in the battery, leading to safety issues such as thermal runaway.

[0040] (6) Binary classification model: The target variable of the model has only two categories, such as failure and normal.

[0041] (7) Multi-classification model: The target variable of the model has three or more categories, such as mild, normal, and severe.

[0042] (8) Electrochemical signals: The electrical signals generated by electrochemical reactions are the basis of battery testing, which may include current, voltage, capacity, power, energy, etc.

[0043] (9) Coulombic efficiency: refers to the ratio of the actual amount of charge released by the battery during the charging and discharging process to the input charge, reflecting the effectiveness of the battery's internal charge utilization.

[0044] (10) DC internal resistance: refers to the internal resistance of a battery under DC current conditions. It reflects the battery's resistance to current flow during the charge and discharge process. The DC internal resistance can be calculated by applying a DC current pulse and recording the voltage change before and after the pulse.

[0045] (11) Linear interpolation: A mathematical method used to estimate or construct a new data point between two known data points, whose value is calculated based on the values ​​of the two known points and their relative positions, through a linear relationship (i.e., a straight line).

[0046] (12) SOH: SOH is the abbreviation of State of Health, which means "battery health status" in Chinese. It is used to measure the degree of battery degradation and remaining service life, usually expressed as a percentage.

[0047] (13) Lithium ion deintercalation and intercalation: two key steps in the charging and discharging process of lithium-ion batteries. These two processes directly determine the performance and service life of the battery.

[0048] (14) Overall failure condition: Only determine whether the battery has safety failure, regardless of the specific type of failure.

[0049] (15) Specific failure conditions: Specific failure modes of the battery, such as lithium deposition on the tab, lithium deposition on the side, tab breakage, etc.

[0050] (16) Deep learning: An important branch of the field of artificial intelligence (AI) and machine learning (ML). Its core idea is to learn and extract features from large amounts of data by building and training deep neural network models to achieve automated processing and decision-making of complex tasks.

[0051] (17) Median filter: It is a nonlinear filtering technique that reduces noise by replacing each pixel value with the median value of its neighborhood.

[0052] (18) Extracted features: In the embodiments of the present disclosure, extracted features refer to features extracted from electrochemical data.

[0053] (19) Design features: In the embodiments of the present disclosure, design features refer to features extracted from data such as test conditions, processes, and recipes.

[0054] (20) Recall rate: This measures the proportion of positive examples that a model correctly identifies among all actual positive examples. It reflects the model's ability to identify positive examples, that is, the model's ability to "recall" positive examples. A higher recall rate indicates that the model is able to identify more actual positive examples.

[0055] (21) Accuracy: One of the most basic indicators for measuring the performance of a classification model, it represents the proportion of samples correctly predicted by the model to the total number of samples. The higher the accuracy, the better the overall prediction effect of the model.

[0056] (22) Grid search: A technique for hyperparameter optimization that improves model performance by traversing a predefined parameter grid to find the optimal parameter combination.

[0057] Figure 1 FIG. 1 is a flow chart illustrating a method for predicting battery failure according to some embodiments of the present disclosure. Figure 1 As shown, the method includes steps S102 to S110.

[0058] In step S102, cycle test data of the battery is obtained. For example, the cycle test data of the battery can be obtained through a cycle test.

[0059] In step S104 , the cycle test data of the battery is calculated to obtain a plurality of dimension curves, which include a SOH curve.

[0060] In some embodiments, the multiple dimensional curves may further include at least one of a DC internal resistance curve, a charging time curve, a Coulomb efficiency curve, a peak change curve of a single-cycle capacity increment curve, and a variance curve of the difference between single-cycle capacity curves. This provides multiple dimensional curve types.

[0061] In step S106, the number of first alarm points existing in the plurality of dimensional curves is obtained, wherein in each dimensional curve, each occurrence of a feature exceeding the feature threshold of the dimensional curve is recorded as the existence of a first alarm point.

[0062] Each curve has its own characteristics, such as its peak value or differential. A threshold can be set for each characteristic of each curve. In each curve, if there is one feature exceeding (greater than or equal to) the threshold, it is considered a first alarm point. If there are two features exceeding (greater than or equal to) the threshold, it is considered two first alarm points, and so on. The number of first alarm points across all curves is then counted, which is the number of first alarm points in the multi-dimensional curves.

[0063] For example, the characteristic threshold value for a DC internal resistance curve may range from 0.001 to 0.005 mΩ (milliohms). For example, the characteristic threshold value for a charge duration curve may range from 100 seconds to 500 seconds. For example, the characteristic threshold value for a SOH curve may range from 0.01 to 0.1. For example, the characteristic threshold value for a Coulombic efficiency curve may range from 0.01 to 0.05. For example, the characteristic threshold value for a peak change curve of a single-cycle capacity increment curve may range from 10 to 50. For example, the characteristic threshold value for a variance curve of differences between single-cycle capacity curves may range from 0.0001 to 0.0005.

[0064] It should be noted that the ranges of the characteristic thresholds of the above curves are only exemplary, and the scope of the present disclosure is not limited thereto. The characteristic thresholds of the above curves can be set according to actual conditions or actual needs.

[0065] The above steps S102 to S106 belong to the rule algorithm described above.

[0066] In step S108, an inflection point detection is performed on the SOH curve to obtain the number of inflection points in the SOH curve. This step S108 belongs to the inflection point detection algorithm described above.

[0067] Here, if an inflection point appears in the SOH curve, it means that the battery has experienced performance failure, and the number of inflection points is recorded as 1. Since the battery has already failed, the number of inflection points will not be increased subsequently. Therefore, if the battery has experienced performance failure, the number of inflection points in the SOH curve is 1; if the battery has not experienced performance failure, the number of inflection points in the SOH curve is 0. In other words, for a battery, the number of inflection points in its SOH curve is 0 or 1.

[0068] In step S110, it is determined whether the battery is in a failure state according to the number of first alarm points and the number of inflection points.

[0069] Here, the number of first alarm points can reflect the safety failure of the battery determined by the rule algorithm, and the number of inflection points can reflect the performance failure of the battery determined by the inflection point detection algorithm. Therefore, it is possible to determine whether the battery has a failure based on the number of first alarm points and the number of inflection points.

[0070] Thus, a battery failure prediction method according to some embodiments of the present disclosure is provided. The method includes: obtaining battery cycle test data; calculating the battery cycle test data to obtain multiple dimensional curves, the multiple dimensional curves including a battery state of health (SOH) curve; obtaining the number of first alarm points present in the multiple dimensional curves, wherein, in each dimensional curve, each occurrence of a feature exceeding a feature threshold of the dimensional curve is recorded as the presence of a first alarm point; performing inflection point detection on the SOH curve to obtain the number of inflection points in the SOH curve; and determining whether the battery has failed based on the number of first alarm points and the number of inflection points. That is to say, in this method, multiple dimensional curves are obtained by calculating the cycle test data of the battery, and the first alarm point and its number are obtained by using the characteristics of the multiple curves. The number of inflection points in the SOH curve is obtained by performing inflection point detection on the SOH curve, and then it is determined whether the battery has failed based on the number of first alarm points and the number of inflection points. In this way, the prediction of battery failure is achieved, and the first alarm point and its number are used to reflect the safety failure of the battery (rule algorithm), and the inflection point and its number are used to reflect the performance failure of the battery (inflection point detection algorithm). Therefore, the above method can predict battery failure more comprehensively, thereby improving the accuracy of battery failure prediction.

[0071] In some embodiments, step S110 includes: performing a weighted sum calculation on the number of first alarm points and the number of inflection points to obtain a first weighted result; and determining that the battery has failed if the first weighted result is greater than or equal to a first threshold. This achieves the purpose of combining the number of first alarm points and the number of inflection points through weighted summation to determine whether the battery has failed, thereby improving the accuracy of battery failure prediction.

[0072] For example, the first weighted result C r1 for

[0073] , (1)

[0074] Where x is the number of first alarm points, α is the weight corresponding to the number of first alarm points, z is the number of inflection points, and δ is the weight corresponding to the number of inflection points. Here, α and δ are pre-set weight coefficients that can be set according to actual conditions.

[0075] In the above formula, the number of first alarm points is squared and then weighted summed with the number of inflection points. This square operation can increase the importance of the first alarm point, thereby improving the accuracy of battery failure prediction.

[0076] For example, δ is equal to the first threshold. Since the number of inflection points is 0 or 1, as long as there is an inflection point, the first weighted result C in the above formula (1) can be made r1 If the value is greater than or equal to the first threshold, the battery is in a failure state. In other words, when there is a performance failure, it can be determined that the battery is in a failure state, which is closer to the actual situation and improves the accuracy of battery failure prediction.

[0077] In some embodiments, the first threshold value ranges from 0.001 to 0.01. It should be noted that the first threshold value can be designed according to actual conditions, and the scope of the present disclosure is not limited to the specific value of the first threshold value.

[0078] Figure 2 Schematic diagram showing the inflection point detection of the SOH curve according to some embodiments of the present disclosure. Figure 2 The process of detecting the inflection point of the SOH curve is described in detail.

[0079] First, a prediction line is generated based on the first discrete point corresponding to the first cycle test and the current discrete point corresponding to the current cycle test in the SOH curve. The prediction line is a line segment connecting the first discrete point and the current discrete point. For example, Figure 2 2 shows discrete points A to I on the SOH curve 201, where the first discrete point is A, the current discrete point is H, and the predicted line 202 is the line segment between the discrete points A and H. It should be noted that starting from the first cycle, a predicted line is generated for each subsequent cycle.

[0080] Next, the midpoint of the line segment is obtained, and based on the midpoint, a first portion of discrete points corresponding to the line segment on the left side of the midpoint and a second portion of discrete points corresponding to the line segment on the right side of the midpoint are obtained. Figure 2As shown, the midpoint 212 of the line segment is obtained, and the first portion of discrete points A to D corresponding to the line segment on the left side of the midpoint 212 and the second portion of discrete points E to H corresponding to the line segment on the right side of the midpoint are obtained.

[0081] Next, the absolute value of the first difference between the ordinate value of each discrete point in the first part of the discrete points and the ordinate value of the corresponding point on the predicted straight line corresponding to each discrete point is calculated, and the first sum of the absolute values ​​of all the first differences corresponding to the first part of the discrete points is calculated.

[0082] Here, as Figure 2 As shown, the discrete points on the SOH curve correspond to the corresponding points on the prediction line 202. The discrete points and their corresponding points have the same horizontal coordinate value, but the vertical coordinate values ​​may be different. Therefore, the absolute value of the difference between the vertical coordinate value of each discrete point in the first part of discrete points A to D and the vertical coordinate value of the corresponding point of each discrete point (i.e., the first difference) can be calculated, and the sum of the absolute values ​​of these differences, i.e., the first sum, can be calculated.

[0083] Next, the absolute value of the second difference between the ordinate value of each discrete point in the second part of discrete points and the ordinate value of the corresponding point on the predicted straight line corresponding to each discrete point is calculated, and the second sum of the absolute values ​​of all second differences corresponding to the second part of discrete points is calculated.

[0084] Similar to the above, if Figure 2 As shown, the absolute value of the difference (i.e., the second difference) between the ordinate value of each discrete point in the second part of discrete points E to H and the ordinate value of the corresponding point of each discrete point (the corresponding point on the predicted straight line) can be calculated, and the sum of the absolute values ​​of these differences, i.e., the second sum, can be calculated.

[0085] When the second sum is greater than the sum of the first sum and the third threshold, the current discrete point is determined to be an inflection point.

[0086] Here, the second sum is greater than the sum of the first sum and the third threshold, indicating that the distances between the second part of discrete points and their corresponding points are relatively large, which reflects that the current discrete point is an inflection point. Figure 2 The inflection point vertical line 203 is shown in FIG.

[0087] In some embodiments, the third threshold value ranges from 0.01 to 0.1. Of course, the third threshold value can be adjusted according to needs, and the scope of the present disclosure is not limited thereto. The third threshold value can control the timing of inflection point detection.

[0088] In this way, the inflection point on the SOH curve is detected, which can detect the inflection point of the SOH curve more accurately, thereby improving the accuracy of battery failure prediction.

[0089] It should be noted that starting from the i-th cycle (i is a positive integer), the value of the third threshold can be determined based on actual conditions. Furthermore, the above algorithm can also be used to detect the inflection point of a constant-capacity test. Although the SOH curve does not have an inflection point in this case, it can be determined from the cutoff voltage and starting voltage curves. The principle of this method is the same or similar to the above process.

[0090] The inflection point detection algorithm of the above embodiment of the present disclosure is a bisection cut method. This algorithm uses the fitted SOH curve for detection and uses a point-slope formula to calculate any interpolation point on the line connecting two points. The calculation formula is as follows.

[0091] Given the coordinates of two points (x1, y1) and (x2, y2), the slope m can be calculated as Here, y1 and y2 are the SOH values ​​for each cycle, which are decimals ranging from 1 to 0, while x1 and x2 are the number of cycles (i.e., the number of cycles) corresponding to the battery test SOH, which are positive integers. m is the slope value calculated continuously between two points.

[0092] The formula for calculating the intercept, b, is b = y1 - m × x1. b is the intercept, which is the y-coordinate of the point where the line intersects the y-axis. x1 and y1 are the horizontal and vertical coordinates of the known point. m is the slope calculated from the above formula.

[0093] The formula for calculating the predicted value y is y = m × x + b. Here, the slope m and intercept b are used to calculate the y value corresponding to any x.

[0094] The following describes how to calculate the y value corresponding to the x value on the line connecting any two points based on these formulas:

[0095] The goal is to convert the point-slope form equation to the slope-intercept form equation y = mx + b, where m is the slope and b is the intercept.

[0096] 1) Expand the point-slope form of the equation

[0097] Start with the point-slope form equation: y-y1=m(x-x1).

[0098] Expand the expression on the right: y-y1=mx-mx1.

[0099] 2) Solve for y

[0100] Move y1 to the right side of the equation: y=mx-mx1+y1.

[0101] 3) Arrange it into oblique intercept form

[0102] Rearranging the terms gives the slope-intercept form of the equation: y=mx+(y1-mx1),

[0103] Here, y1-mx1 is the intercept b. Therefore, the slope-intercept form equation is: y=mx+b where: b=y1-mx1.

[0104] Figure 3 is a flow chart showing a method for predicting a battery failure mode using a prediction model according to some embodiments of the present disclosure. Here, the method for predicting a battery failure mode using a prediction model is the AI ​​algorithm described above. Figure 3 As shown, the method includes steps S302 to S308.

[0105] In step S302 , feature extraction is performed on multiple dimension curves of the battery to obtain extracted features.

[0106] For example, for each dimensional curve, the features of the curve may include: the historical peak value of the curve, the difference of the historical peak value of the curve, the current value of the curve and the difference of the current value of the curve, etc. These features are called extracted features. The extracted features can also be called extracted vector features.

[0107] In step S304, the battery test conditions, process measurement values, and material codes are obtained as design features of the battery, where the material code corresponds to the battery recipe information. This design feature can also be called a design vector feature.

[0108] In step S306 , the extracted features and the designed features are combined into combined features.

[0109] For example, the design feature is placed behind the extracted feature to form a combined feature, which can also be called a combined vector feature.

[0110] In step S308, the combined features are input into a pre-trained prediction model, and the battery failure is predicted by the prediction model to obtain the failure mode of the battery.

[0111] Thus, a method for predicting battery failure modes using a prediction model according to some embodiments of the present disclosure has been provided. This method can predict the specific failure mode of a battery, further improving the accuracy of battery failure prediction.

[0112] In some embodiments, step S308 may include: performing a failure prediction on the battery using a prediction model to obtain a number of failure modes that the battery may have, as the number of second alarm points, wherein each predicted failure mode of the battery is recorded as a second alarm point. In some cases, a battery may have multiple failure modes. In addition to obtaining the failure mode of the battery using the prediction model, the number of failure modes of the battery may also be counted.

[0113] In some embodiments, the aforementioned determination of whether a battery failure exists based on the number of first alarm points and the number of inflection points includes determining whether a battery failure exists based on the number of first alarm points, the number of inflection points, and the number of second alarm points. In other words, the number of second alarm points predicted by the prediction model is combined with the aforementioned number of first alarm points and the number of inflection points to determine the battery failure, thereby further improving the accuracy of battery failure prediction.

[0114] In some embodiments, determining whether a battery has failed based on the number of first alarm points, the number of inflection points, and the number of second alarm points includes: performing a weighted sum calculation on the number of first alarm points, the number of inflection points, and the number of second alarm points to obtain a second weighted result; and determining that the battery has failed if the second weighted result is greater than or equal to a second threshold. This achieves the purpose of determining whether a battery has failed by combining the number of first alarm points, the number of inflection points, and the number of second alarm points through weighted summation, thereby improving the accuracy of battery failure prediction.

[0115] For example, the second weighted result C r2 for

[0116] , (2)

[0117] Where x is the number of first alarm points, α is the weight corresponding to the number of first alarm points, z is the number of inflection points, δ is the weight corresponding to the number of inflection points, y is the number of second alarm points, and β is the weight corresponding to the number of second alarm points. α, β, and δ are preset weight coefficients that can be set according to actual conditions.

[0118] In the above formula, the number of first alarm points is squared and then weighted together with the number of inflection points and the number of second alarm points. This square operation can increase the importance of the first alarm point, thereby improving the accuracy of battery failure prediction.

[0119] For example, δ is equal to the second threshold. Since the number of inflection points is 0 or 1, as long as there is an inflection point, the second weighted result C in the above formula (2) can be made r2 If the value is greater than or equal to the second threshold, the battery is in a failure state. In other words, when there is a performance failure, it can be determined that the battery is in a failure state, which is closer to the actual situation and improves the accuracy of battery failure prediction.

[0120] In some embodiments, the second threshold is in the range of 0 < second threshold ≤ 1. It should be noted that the second threshold can be designed according to actual conditions, and the scope of the present disclosure is not limited to a specific value of the second threshold.

[0121] In some embodiments, the battery failure prediction method may further include: training the prediction model before predicting battery failure using the prediction model. By training the prediction model, the trained prediction model can be used to easily predict the failure mode of the battery.

[0122] Figure 4 is a flow chart illustrating a method for training a prediction model according to some embodiments of the present disclosure. Figure 4 As shown, the method includes steps S402 to S414.

[0123] In step S402 , different failure modes of the battery are graded based on the severity of the failure modes of the battery to obtain a level for each failure mode.

[0124] For example, based on the existing data, qualitative judgments can be made on different failure modes, and the severity of each failure mode can be given. For example, failure modes can be divided into three levels: no failure (i.e., normal), minor failure, and severe failure.

[0125] There are four possible classification situations: (1) None: indicating that the current failure mode has not occurred; (2) Minor: indicating that the failure mode has occurred but has not reached the severe threshold; (3) Severe: indicating that the failure mode has occurred and has exceeded the severe threshold; (4) Undisassembled / Empty: indicating that the current battery has not been observed and analyzed for this failure mode. Of course, data such as "Undisassembled / Empty" will not be used. Therefore, there are the above three levels: None, Minor, and Severe. For example, Table 1 shows exemplary classification label data.

[0126] Table 1. Example of classification label data

[0127]

[0128] That is, for each failure mode (ie, each type of failure mode), several batteries tested under this failure mode may have at least one level of no failure, slight failure, and severe failure.

[0129] In step S404 , calculations are performed on the cycle test data of the plurality of batteries used for training to obtain a plurality of dimensional curves of the plurality of batteries.

[0130] For example, the cycle test data of all batteries are calculated to obtain their DC internal resistance curve (for example, 30S DC internal resistance curve), charging time curve, SOH curve, Coulomb efficiency curve, peak change curve of single-cycle capacity increment curve (two curves of peak X value and Y value), and variance curve of the difference between single-cycle capacity curves.

[0131] For example, each curve (i.e., feature) is calculated as follows:

[0132] (1) 30S DC internal resistance: The DC internal resistance is calculated by comparing the voltage difference between the battery being fully charged and the battery being discharged for 30 seconds to the first discharge. The 30S DC internal resistance reflects the battery's performance and charge and discharge characteristics.

[0133] (2) Charging time: The time it takes to charge from 3.5V to 4.3V using constant current. This can intuitively show the charging rate and efficiency of the battery at different charging stages.

[0134] (3) SOH curve: The decay curve composed of the discharge capacity of each cycle can intuitively reflect the health status of the battery.

[0135] (4) Coulombic efficiency: The ratio of charge capacity to discharge capacity. This reflects the ratio between the energy stored in the battery and the energy actually released during discharge. It is usually expressed as the ratio of discharge capacity to charge capacity, but here it is the opposite, and is also called coulombic efficiency.

[0136] (5) Capacity increment curve: The relationship between dQ (capacity increment) and dV (voltage increment) is plotted against voltage. This curve reflects the rate of change in the system's ability to release charge under a given voltage change. The X-axis is voltage and the Y-axis is the ratio.

[0137] (6) Capacity difference curve: The standard deviation of the difference between the capacity-voltage curves for different cycle numbers. For example, different cycle numbers mean: 8th cycle - 2nd cycle, 9th cycle - 3rd cycle, etc.

[0138] In some embodiments, training the prediction model may further include: in the process of obtaining multiple dimensional curves for multiple batteries, filling in blank values ​​in each dimensional curve using a linear interpolation method, and smoothing the cycle test data of the multiple batteries using a median average algorithm with a predetermined window length. Interpolation and smoothing the curves can reduce the adverse effects of noisy data.

[0139] For example, linear interpolation can be used to fill null values ​​in each curve, followed by data smoothing using a median average algorithm with a window length of 5. The SOH curve can then be smoothed using an SG filter (Savitzky-Golay filter) for inflection point detection. Furthermore, for the charging duration curve, sampling intervals can be used to identify abnormal sampling points. Using these sampling points, subsequent data sampling can be checked. Once a curve with normal voltage changes is found, all previous sampling times can be deleted to ensure the current cycle's charging duration is normal.

[0140] For example, the smoothing process for each curve is as follows: the median function of Python's numpy library is used to calculate the median filter, and its formula is as follows:

[0141] . (3)

[0142] Here, a one-dimensional data sequence x[n], of length N, can be filtered using a median filter with a window size of 5, for example. Each element y[n] of the output sequence y[n] is the median of the input sequence x[n] within a window of length 5 centered at n.

[0143] Inflection point model: The SG filter fitting is done using the savgol_filter function from the scipy.signal library.

[0144] The savgol_filter function takes as input parameters x (the SOH curve), window (the window width), and polyorder (the order). For example, the window width is calculated as: window=min(len(x), 35), with polyorder set to 2. The min function in this formula selects the minimum between two numbers, and the len function represents the number of points in the SOH curve to be obtained, i.e., the number of loops. These input parameters significantly influence the smoothing effect.

[0145] In step S406 , statistics are collected on the cyclic test data under each failure mode, and failure modes with a good balance between the amount of cyclic test data and the level of the failure mode are selected for modeling and prediction.

[0146] For example, for a certain failure mode, the number of batteries with no failure, minor failure, and severe failure is 3, 3, and 3, respectively. This indicates that the amount of cycle test data for this failure mode is evenly distributed with the failure mode level. This failure mode is then selected for modeling and prediction. After the prediction model is trained, predictions can be made for this failure mode.

[0147] It should be noted that the above-mentioned "uniformity" includes but is not limited to absolute "uniformity," and a certain degree of error can be allowed. Therefore, "uniformity" here can be understood as "substantially uniform." For example, in a certain failure mode, the number of batteries with no failure, slight failure, and severe failure levels is 3, 4, and 3, respectively. This can be understood as the data volume of the cycle test data under this failure mode and the level of the failure mode are basically uniform. Such failure modes can also be used for modeling and prediction.

[0148] In step S408 , an extracted feature of each dimension curve of the battery with the selected failure mode is obtained based on the multiple dimension curves of the battery with the selected failure mode.

[0149] Electrochemical features are extracted from each curve and mathematically transformed. For example, using a median filter with a window of 5, every five cycles of data are processed, ultimately filtering each cycle of each battery. Therefore, the following extracted features for each cycle are obtained for each battery, as shown in Table 2:

[0150] Table 2 Exemplary extracted features

[0151]

[0152] Note: If the battery has 100 cycles, then Table 2 will have 100 rows. This is just the example filtering window size and data, i.e., not all columns are shown in this table. Curves A and B mentioned in this table represent the various curves extracted in the previous steps. The mathematical transformation of the curve features is explained as follows:

[0153] Historical peak value: The maximum value corresponding to the curve (before the current circle).

[0154] Historical peak difference: The maximum value of the difference corresponding to the curve (before the current circle).

[0155] Current value: The value corresponding to the curve of this circle.

[0156] Current value difference: the difference corresponding to the curve of this circle.

[0157] Difference: Starting from the second value, the difference between each value and the previous value (starting from the second difference value, the average of the current difference and the previous difference is used as the smoothed difference value to avoid excessive influence of a single point).

[0158] In step S410, the test conditions, process measurement values ​​and material codes of the battery with the selected failure mode are obtained as design features of the battery with the selected failure mode, wherein the material code corresponds to the recipe information of the battery with the selected failure mode.

[0159] For example, the obtained battery test conditions, design information and labels are shown in Table 3 below:

[0160] Table 3 Exemplary battery test conditions and design information and labels

[0161]

[0162] The fields are examples. For each battery, the design features do not change. In the above table, each battery has a record, which can be called the design vector feature of the battery, that is, the design feature.

[0163] In step S412 , the extracted features are combined with the design features to form a combined feature of the battery of the selected failure mode.

[0164] For example, the design features are placed after the extracted features to form a combined feature, also known as a combined vector feature. Combining the extracted features with the design features yields the complete input features required by the model. For example, Table 4 shows an example of a combined feature and label value.

[0165] Table 4 Example combination features and label values

[0166]

[0167] Among them, [a, b, c] represents a one-dimensional vector. In fact, all the values ​​in this vector can be spliced ​​after the previous column as separate columns like other features.

[0168] In step S414 , the prediction model is trained based on the combined features of the battery of the selected failure mode and the label value of the selected failure mode.

[0169] Thus, a method for training a prediction model according to some embodiments of the present disclosure is provided, which achieves the purpose of training the prediction model. By using the trained prediction model to predict the failure mode of the battery, the accuracy of predicting the failure mode of the battery can be improved.

[0170] Figure 5 FIG. 1 is a flow chart showing a method for predicting battery failure according to other embodiments of the present disclosure. Figure 5 As shown, the method includes steps S501 to S506.

[0171] In step S501, the existing data is labeled, classified, and feature extracted, and the curve is smoothed and cleaned.

[0172] For example, each failure mode after battery disassembly is roughly graded, and the failure mode is divided into three levels according to the severity: none, mild, and severe to give a classification label.

[0173] For all battery cycle data, calculations are performed to obtain the peak variation curves (X and Y values ​​of the peak) for the 30S DC internal resistance curve, charge time curve, SOH curve, Coulombic efficiency curve, and single-cycle capacity increment curve, as well as the variance curve of the difference between the single-cycle capacity curves. These curves can be used in AI models (i.e., predictive models) and rule-based models (i.e., rule-based algorithms). First, linear interpolation is used to fill in blank values ​​for each curve. The data is then smoothed using a median average algorithm with a predetermined window length (e.g., 5). An SG filter is then used to smooth the SOH curve for inflection point detection. Furthermore, for the charge time curve, sampling intervals can be used to identify abnormal sampling points. Using these sampling points, subsequent data sampling can be checked. Once the voltage variation conforms to a normal pattern is identified, all previous sampling times are deleted to ensure a normal charge time for the current cycle.

[0174] Abnormal sampling points are mainly determined based on test conditions and experience.

[0175] For example, on a machine with a 30-second test sampling rate, the cycle data has a sampling point every 30 seconds, meaning a piece of electrochemical data including voltage, current, capacity, and other dimensions. Based on this interval (30 seconds) and the charge and discharge voltage pattern (i.e., monotonically increasing voltage during constant-current charging and monotonically decreasing voltage during constant-voltage discharge), we filter out significantly abnormal sampling points. For example, if the interval caused by a test interruption exceeds one hour, this data is removed to avoid significant abnormal increases in charging time.

[0176] In addition, the SOH curve fitted by the SG filter is used for inflection point detection. This is because the SG filter may mask the failure signal during failure prediction. However, inflection point detection considers the drop in the SOH curve, so this effect can be ignored.

[0177] In step S502 , a failure mode with a better distribution is selected as a prediction target, and a deep learning algorithm is used to fill the data set.

[0178] For example, statistically analyze the distribution of each failure mode category and select the failure mode with the most balanced overall number and level as the target failure mode for modeling. If the selected target data volume is still insufficient or unevenly distributed, optionally generate new data using a data augmentation algorithm to fill the dataset and ensure that the data volume and distribution meet the requirements. Alternatively, consider classifying minor and normal failures as one category and abnormal failures as another.

[0179] For example, after obtaining the labels for each failure mode organized in the previous step, you can quickly obtain the number of different levels of each failure mode and the total number of empty / unopened items, thereby obtaining the corresponding classification distribution and total data volume. Among these failure modes, choose a failure mode with a "no failure (i.e., normal): minor failure: major failure" ratio close to 1:1:1 and a large total volume to build a multi-classification model, or choose a failure mode with a "minor failure + no failure: major failure" ratio close to 1:1 to build a binary classification model.

[0180] After selecting the target failure mode, if the distribution and quantity do not meet the requirements, a supplementary dataset can be generated using a deep learning-based data augmentation algorithm.

[0181] For example, the method for generating a supplementary data set may include: calculating the number of normal batteries to be expanded; obtaining normal battery weighted cycle data; using deep learning technology to train a first model, which can convert a curve into a latent space vector (encoding) and can convert the latent space vector back to the original curve (decoding); training a second model to learn the hidden laws within the first model and, by applying specific conditions, generate latent space vectors that meet the conditions, so that the second model can generate new curve data that meets the conditions through the decoding part of the first model; and using the trained first and second models to generate new normal battery data. In this way, a supplementary data set can be generated. The data augmentation algorithm here is implemented based on the TimeVAE algorithm.

[0182] It should be noted that the above-mentioned specific conditions refer to batteries generated under different conditions, such as rate, temperature, etc. For example, if batteries are generated at 25 degrees and 45 degrees, 25 and 45 are the conditions.

[0183] In step S503, information such as test conditions, formula, and process of the battery is obtained as design features.

[0184] For example, the test temperature, recipe information, and process measurement values ​​for each battery are obtained. The recipe information can use the material code as input to facilitate model differentiation. The process measurement features are filtered using the Pearson correlation coefficient to reduce the number of input features and ease the model fitting difficulty. This results in a one-dimensional vector (design features) for each battery, where each column represents a model input feature. For reference, see Table 3.

[0185] In step S504, the crop prediction labels are classified to different degrees after quantization, the extracted features and the designed features are combined, a sliding window is used to generate a data set and a prediction model is trained.

[0186] For example, based on the N one-dimensional vectors in step S503 (N is the number of batteries with data for the current failure mode), the target failure mode severity classification is added as a label value to the end of the corresponding design feature. Then, a sliding window is used to slide the various dimensional curves of each cell (i.e., battery) to generate a time series data set (i.e., extracted features). For example, with a sliding window size of 2, the window slides one point at a time. To avoid abnormal charging and discharging data in the early stages, in actual use, the sliding will start from the data of the mth cycle (e.g., m=11), that is, the data must have m points or more to be used. The input features of the resulting time series data set are in the form of: {number of records, number of extracted features, sequence length}, where the number of records is the total number of records generated for all batteries after the sliding window operation, the extracted features include the historical peak value, historical peak difference, current value, and current value difference of the current position of each curve, and the sequence length is the size of the sliding window. Secondly, the vector obtained at the beginning of step S504 is filled into the rightmost side of the corresponding battery data in the time series data set to supplement the design information and target prediction label (the design and label can be spliced ​​before sliding, or the design features and labels can be spliced ​​after sliding as described in the present disclosure). The shape of the input feature is: {number of records, number of extracted features + design features, sequence length}. The embodiment of the present disclosure may not use all the time series data, because the failure of a battery does not mean that it has failed in every previous cycle. Therefore, the present disclosure may only use the last sequence and label before the inflection point as the data set. If it can be determined in which cycles the failure occurs, all the time series data can be used. In some cases, the time series data can be converted into a piece of data. The shape of the input feature is: {number of batteries, number of extracted features × sequence length + number of design features}, which is called a "model data set". This conversion step is not required if time series is used.

[0187] For example, in step 1, a sliding window (e.g., window length is 2) is used to slide over the extracted features of all cycles. The sliding process is as follows Figure 6 shown. Figure 6 In the figure, 11, 12, 13, 14, and 15 represent the cycle number of the battery, that is, the cycle number. Here, the battery cycle starts from the 1st cycle, but the prediction starts from the 11th cycle.

[0188] In step 2, all data within the window are concatenated into a one-dimensional vector.

[0189] In step 3, the design information vector is expanded and appended to the end of the one-dimensional vector in step 2.

[0190] In step 4, the label is appended to the end of the vector in step 3. This will result in a table such as that shown in Table 4.

[0191] In step S505 , a rule algorithm is designed to mine potential failure signals, a threshold value is given for each curve based on the current data set, and effective extraction features are screened from all curves.

[0192] Here, the rule algorithm (i.e., rule model) refers to the classification of current data as normal or abnormal based on defined rules. For example, the judgment rule could be a differential judgment of previously extracted curves for each dimension. The current threshold setting is based on the curve differentials between data in the dataset that do not contain any failure modes and other data that do contain failure modes. When a dimension exceeds the threshold more than the number of times, a failure signal is detected. When a failure signal occurs in any dimension, the rule algorithm will deem an alarm to be present. In some cases, the final alarm can be determined by priority rules rather than solely based on the alarm of the rule model. This alarm indicates whether the battery has experienced any of the considered failure modes. For specific failure modes, the previously constructed predictive model (AI algorithm model) is used for judgment.

[0193] The AI ​​algorithm models mentioned above include a variety of models, such as random forests, support vector machines, multi-layer perceptrons, XGBoost (eXtreme Gradient Boosting), and Tabular Prior-Data Fitted Networks (TabPFN). This method uses a dataset to train these models, then selects the optimal model as the final prediction model by comparing metrics such as recall and precision. Grid search techniques are used during model construction to traverse the space of possible hyperparameters to obtain the optimal performance of the current model. The dataset used by the model is the "model dataset" in step S504.

[0194] The calculation formulas for recall and precision are as follows:

[0195] , (4)

[0196] . (5)

[0197] The goal of grid search optimization is to find a set of parameters that optimizes the average performance of the model during cross-validation. The optimization goals of the disclosed embodiment are recall and precision.

[0198] Before optimization, a parameter grid must be defined, listing all possible parameter combinations. For example, for a support vector machine, parameters that may need to be optimized include the regularization parameter C and the kernel function parameters. Cross-validation is then performed to evaluate the effectiveness of each parameter combination, and the optimal parameter combination is selected based on precision and recall. The optimization process is shown below:

[0199] . (6)

[0200] The explanation of the formula is as follows:

[0201] True Positives (TP): The number of samples that the model predicts to be positive and are actually positive.

[0202] True Negatives (TN): The number of samples that the model predicts to be negative and are actually negative.

[0203] False Positives (FP): The number of samples that the model predicts are positive but are actually negative.

[0204] False Negatives (FN): The number of samples that the model predicts as negative but are actually positive.

[0205] In the disclosed embodiments, positive examples represent failures, while negative examples represent normal results. A high recall rate reduces false negatives, meaning the battery failure warning prediction results have a lower false negative rate. Accuracy is generally used to measure the overall prediction effectiveness of a model, but accuracy alone is insufficient to assess a model's false negative rate.

[0206] Cross-validation involves dividing a dataset into multiple subsets, training the model multiple times, and calculating the average recall and precision for each parameter combination within each model, thereby achieving more stable and reliable performance evaluation results. The optimal parameter combination represents the best performance of the model. By comparing the best performance of each model, the best performing models are ultimately selected as failure prediction models.

[0207] In step S506, the rule algorithm is used to predict the overall failure situation, the AI ​​algorithm is used to predict the specific failure situation, and the capacity inflection point detection algorithm is used to determine whether the test is stopped due to performance failure, and output performance failure, safety failure and / or specific failure.

[0208] During testing, an inflection point detection algorithm can be used to determine whether the current cycle has reached an inflection point. If an inflection point is detected, testing can be stopped early. There are multiple inflection point detection algorithms, and this method uses a binary cut method for judgment. When the weighted sum of the inflection point detection algorithm, rule-based algorithm, and failure mode prediction algorithm exceeds the threshold, the relevant personnel responsible for cell testing must be notified and a decision must be made. At this point, the tester can choose to stop the test or arrange for a disassembly analysis to ensure timely and accurate results.

[0209] Here, failure prediction is performed using a combination of AI models, rule-based models, and inflection point models. The alarm point refers to the number of cycles after various algorithms detect a failure signal. Each algorithm generates an alarm point each time a failure signal is detected.

[0210] For example, the timing of alarm point generation for each algorithm is as follows: (1) The rule algorithm has the judgment of multi-dimensional signals. Therefore, when the number of alarm points in one dimension exceeds the first predetermined number (for example, 5 times) or the number of alarm points in the predetermined number (for example, 3) or more dimensions exists, each alarm will increase the warning coefficient αx 2 (2) When an alarm occurs in the failure mode prediction algorithm (specific failure), if it occurs for more than a second predetermined number of times (for example, 3 times), it will be considered that there is a possibility of specific failure, and each alarm will increase the warning coefficient βy; when an alarm occurs in the inflection point detection algorithm, it will be considered that there is a possibility of performance failure, and the warning coefficient δz will be increased.

[0211] judge Is it greater than or equal to the threshold (i.e., the second threshold mentioned above), where Indicates the cumulative sum. δ can be the same as the threshold to ensure that even if the battery is not dead, it will be warned when it reaches the inflection point. α and β should be adjusted according to the actual situation. Note: αx 2 This is because the signals captured by the rule-based algorithm are more obvious anomalies. If such signals occur, the tester needs to be notified as soon as possible. x is the number of alarm points in the rule-based algorithm (i.e., the number of first alarm points), y is the number of specific failure alarm points (i.e., the number of second alarm points), and z is the number of alarm points in the inflection point detection algorithm (i.e., the number of inflection points).

[0212] When an early warning occurs, the tester can be informed to urge the battery test to be stopped as soon as possible to reduce safety risks.

[0213] Thus, a battery failure prediction method according to some other embodiments of the present disclosure is provided. This method uses three algorithms (or models) in combination to perform battery failure prediction, which can improve the accuracy of battery failure prediction.

[0214] In the AI ​​algorithm, existing data is labeled, classified, feature extracted, and abnormal sampling values ​​are processed to obtain information such as battery test conditions, formula, and process as design features. The quantified classifications of different degrees are used as prediction labels, and the extracted features and design features are merged. A sliding window is used to generate a data set and train the AI ​​model.

[0215] In the rule algorithm, feature extraction is performed, abnormal sampling values ​​are processed, and thresholds are selected based on the data set. Each extracted feature has a threshold.

[0216] In the inflection point detection algorithm, the SOH curve is extracted and smoothed, and the threshold is selected according to the data set.

[0217] In some embodiments, a battery failure prediction method uses existing data to qualitatively determine different failure modes, assigning a severity level to each failure mode (e.g., none, mild, severe). The data distribution for each failure mode is then statistically analyzed, and failure modes with relatively uniform data distribution are selected for modeling. The battery's process measurements and material codes are then incorporated into the data set as design features. The prediction is based on multi-cycle, time-series data, providing more comprehensive potential failure characteristics.

[0218] In addition, an inflection point detection algorithm is used to check in real time whether the battery has failed. If the data is severely uneven, two methods can be used based on the label (severity): (1) Generate new similar data to expand the dataset through a data augmentation algorithm; (2) Build a binary classification model by classifying mild and normal as one category and severe as another. This can solve the problem of uneven data distribution to a certain extent, but it depends on whether there is a difference between severe and mild, and whether this can be determined in the disassembly photos.

[0219] Output failure prediction results, which are comprehensively judged by inflection point detection, rule algorithms, AI algorithms combined with the capabilities and rules of different models.

[0220] When dealing with obviously abnormal sampling points, missing values ​​can be interpolated to improve data quality and enhance model robustness. For example, abnormal sampling points can be found by testing the sampling interval, and sampling values ​​that do not conform to normal patterns can be deleted; missing data can be filled by linear interpolation.

[0221] In addition, abnormal curves can be processed. For example, a median average algorithm with a predetermined window length (e.g., a window length of 5) can be used to smooth the data to address abnormal values. The smoothed SOH curve can then be fitted using an SG filter for inflection point detection.

[0222] A designed rule-based algorithm is used to make rough judgments on the characteristic dimensions of multiple electrochemical signals, uncovering potential failure factors in these signals and improving the timeliness, accuracy, and interpretability of early warnings. The rule-based algorithm can be used simultaneously with the inflection point detection algorithm and the machine learning algorithm for prediction. Therefore, the final prediction result depends on all three algorithms.

[0223] In the method disclosed herein, after evaluating the data distribution and total amount of each failure mode, feasible failure modes are selected and modeled separately to predict whether the corresponding failure mode exists in the battery; qualitative judgments are made on the possible failure modes in the data, and a rough classification of the failure conditions is given (for example, none, mild, severe), so as to obtain certain quantitative labels to establish a multi-classification model. Alternatively, a two-classification model is established by taking mild and normal as one category and severe as another. Methods of considering two different dimensions, labeling and data enhancement, are used to improve the problem of small data volume as much as possible. Test condition characterization, process measurement, recipe and other design information features are quantized, and the features are used as input to the model after feature screening to provide more sufficient information for the model and improve the accuracy of the model.

[0224] The aforementioned inflection point detection algorithm is used to promptly identify inflection points and terminate the test. This timely termination minimizes the coupling of performance failures with safety failures, which can make it difficult to trace the cause of the failure.

[0225] Data quality is crucial for small sample datasets. Considering measurement errors during battery testing, different R&D personnel setting different test conditions, varying measurement accuracy across different machines, and other external interventions (e.g., power outages, test personnel requesting a test postponement, or stopping and restarting a test), these noisy data can interfere with model judgment. Data cleaning, smoothing, and interpolation can mitigate these effects.

[0226] Because models focus on specific failure modes, they may overlook obvious signals from other failure modes. Using rule-based algorithms to assist in judgment can reduce these omissions. Furthermore, because rule-based algorithms are based on electrochemical data, they are highly interpretable and can assist battery developers in uncovering failure mechanisms.

[0227] Figure 7 Schematically illustrates a block diagram of a battery failure prediction device according to some embodiments of the present disclosure. The battery failure prediction device includes a memory 710 and a processor 720.

[0228] The memory 710 can be a disk, flash memory or any other non-volatile storage medium. Figure 1 、 Figure 3 、 Figure 4 and Figure 5 At least one of the corresponding embodiments corresponds to the instructions.

[0229] The processor 720 is coupled to the memory 710 and can be implemented as one or more integrated circuits, such as a microprocessor or a microcontroller. The processor 720 is used to execute instructions stored in the memory, thereby improving the accuracy of battery failure prediction.

[0230] In one embodiment, it is also possible to Figure 8 As shown, battery failure prediction device 700 includes memory 710 and processor 720. Processor 720 is coupled to memory 710 via BUS 730. Battery failure prediction device 700 can also be connected to an external storage device 750 via a storage interface 740 to access external data, and can also be connected to a network or another computer system (not shown) via a network interface 760, which will not be described in detail here.

[0231] In this embodiment, the data instructions are stored in the memory and then processed by the processor, thereby improving the accuracy of battery failure prediction.

[0232] In another embodiment, the present disclosure further provides a computer-readable storage medium (eg, a non-transitory computer-readable storage medium) having computer program instructions stored thereon. When the instructions are executed by a processor, the computer-readable storage medium implements the following operations: Figure 1 、 Figure 3 、 Figure 4 and Figure 5 The present invention relates to a method comprising at least one of the steps of the method in the corresponding embodiment. Those skilled in the art will appreciate that the embodiments of the present disclosure may be provided as methods, apparatuses, or computer program products. Therefore, the present disclosure may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present disclosure may take the form of a computer program product implemented on one or more computer-usable non-transitory storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0233] The present disclosure is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present disclosure. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0234] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0235] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0236] In some embodiments of the present disclosure, a computer program product is further provided. The computer program product includes a computer program or instructions. When the computer program or instructions are executed by a processor, the battery failure prediction method as described above is implemented.

[0237] In some embodiments of the present disclosure, a computer program is further provided, comprising: instructions, which, when executed by a processor, cause the processor to execute the battery failure prediction method as described above.

[0238] Although the present application has been described with reference to preferred embodiments, various modifications may be made thereto and components may be substituted with equivalents without departing from the scope of the present application. In particular, the various technical features described in the various embodiments may be combined in any manner as long as there are no structural conflicts. The present application is not limited to the specific embodiments disclosed herein, but encompasses all technical solutions within the scope of the claims.

Claims

1. A battery failure prediction method, characterized in that: include: Obtain battery cycle test data; Calculating the cycle test data of the battery to obtain a plurality of dimensional curves, wherein the plurality of dimensional curves include a battery state of health (SOH) curve; Obtaining the number of first alarm points in the plurality of dimension curves, wherein in each dimension curve, each occurrence of a feature exceeding a feature threshold of the dimension curve is recorded as the presence of a first alarm point; performing inflection point detection on the SOH curve to obtain the number of inflection points in the SOH curve; and It is determined whether the battery has failed according to the number of the first alarm points and the number of the inflection points.

2. The battery failure prediction method according to claim 1, wherein: Determining whether the battery has failed according to the number of the first alarm points and the number of the inflection points includes: Performing a weighted sum calculation on the number of the first alarm points and the number of the inflection points to obtain a first weighted result; and When the first weighted result is greater than or equal to a first threshold, it is determined that the battery has failed.

3. The battery failure prediction method according to claim 2, characterized in that: The first weighted result C r1 for ， Wherein, x is the number of the first alarm points, α is the weight corresponding to the number of the first alarm points, z is the number of the inflection points, and δ is the weight corresponding to the number of the inflection points.

4. The battery failure prediction method according to claim 3, characterized in that: The δ is equal to the first threshold.

5. The battery failure prediction method according to claim 1, characterized in that: Also includes: performing feature extraction on multiple dimensional curves of the battery to obtain extracted features; Obtaining test conditions, process measurement values, and material codes of the battery as design features of the battery, wherein the material codes correspond to formula information of the battery; combining the extracted features and the designed features into a combined feature; and The combined features are input into a pre-trained prediction model, and failure prediction of the battery is performed using the prediction model to obtain a failure mode of the battery.

6. The battery failure prediction method according to claim 5, characterized in that: Performing failure prediction on the battery using the prediction model to obtain a failure mode of the battery includes: The battery is subjected to failure prediction using the prediction model to obtain the number of failure modes of the battery as the number of second alarm points, wherein each predicted failure mode of the battery is recorded as a second alarm point.

7. The battery failure prediction method according to claim 6, characterized in that: Determining whether the battery has failed according to the number of the first alarm points and the number of the inflection points includes: Whether the battery is in failure state is determined according to the number of the first alarm points, the number of the inflection points, and the number of the second alarm points.

8. The battery failure prediction method according to claim 7, wherein: Determining whether a battery failure occurs according to the number of the first alarm points, the number of the inflection points, and the number of the second alarm points includes: performing a weighted sum calculation on the number of the first alarm points, the number of the inflection points, and the number of the second alarm points to obtain a second weighted result; and When the second weighted result is greater than or equal to a second threshold, it is determined that the battery has failed.

9. The battery failure prediction method according to claim 8, characterized in that: The second weighted result C r2 for ， Among them, x is the number of the first alarm points, α is the weight corresponding to the number of the first alarm points, z is the number of the inflection points, δ is the weight corresponding to the number of the inflection points, y is the number of the second alarm points, and β is the weight corresponding to the number of the second alarm points.

10. The battery failure prediction method according to claim 9, characterized in that: The δ is equal to the second threshold.

11. The battery failure prediction method according to claim 5, characterized in that: Also includes: Before performing failure prediction on the battery using the prediction model, the prediction model is trained.

12. The battery failure prediction method according to claim 11, characterized in that: Training the prediction model includes: Classifying different failure modes of the battery based on the severity of the failure mode of the battery to obtain a level for each failure mode; performing calculations on cycle test data of a plurality of batteries for training to obtain a plurality of dimension curves of the plurality of batteries; Collect statistics on the cyclic test data under each failure mode, and select failure modes with a uniform ratio between the amount of cyclic test data and the level of the failure mode for modeling and prediction; Obtaining an extracted feature of each dimensional curve of the battery with the selected failure mode based on a plurality of dimensional curves of the battery with the selected failure mode; Obtaining test conditions, process measurement values, and material codes of the battery with the selected failure mode as design features of the battery with the selected failure mode, wherein the material code corresponds to recipe information of the battery with the selected failure mode; combining the extracted features with the design features into a combined feature of the battery for the selected failure mode; and The prediction model is trained based on the combined features of the battery of the selected failure mode and the label value of the selected failure mode.

13. The battery failure prediction method according to claim 12, characterized in that: Training the prediction model further includes: In the process of obtaining multiple dimensional curves of the multiple batteries, the null values ​​in each dimensional curve are filled by a linear interpolation method, and the cycle test data of the multiple batteries are smoothed by a median average algorithm with a predetermined window length.

14. The battery failure prediction method according to claim 1, characterized in that: The multiple dimensional curves further include: at least one of a DC internal resistance curve, a charging time curve, a Coulomb efficiency curve, a peak change curve of a single-cycle capacity increment curve, and a variance curve of a difference between single-cycle capacity curves.

15. The battery failure prediction method according to any one of claims 1 to 14, characterized in that: Detecting the inflection point of the SOH curve includes: generating a prediction straight line based on a first discrete point corresponding to the first cycle test and a current discrete point corresponding to the current cycle test in the SOH curve, the prediction straight line being a line segment connecting the first discrete point and the current discrete point; Obtaining a midpoint of the line segment, and based on the midpoint, obtaining a first portion of discrete points corresponding to the line segment on the left side of the midpoint and a second portion of discrete points corresponding to the line segment on the right side of the midpoint; Calculating an absolute value of a first difference between a ordinate value of each discrete point in the first portion of discrete points and a ordinate value of a corresponding point on the predicted straight line corresponding to the discrete point, and calculating a first sum of the absolute values ​​of all first differences corresponding to the first portion of discrete points; Calculating an absolute value of a second difference between the ordinate value of each discrete point in the second portion of discrete points and the ordinate value of a corresponding point on the predicted straight line corresponding to the discrete point, and calculating a second sum of the absolute values ​​of all second differences corresponding to the second portion of discrete points; When the second sum is greater than the sum of the first sum and a third threshold, the current discrete point is determined to be an inflection point.

16. A battery failure prediction device, characterized in that: include: Memory; as well as A processor coupled to the memory, wherein the processor is configured to execute the battery failure prediction method according to any one of claims 1 to 15 based on instructions stored in the memory.

17. A computer-readable storage medium, characterized in that Computer instructions are stored thereon, and when the computer instructions are executed by a processor, the battery failure prediction method according to any one of claims 1 to 15 is implemented.

18. A computer program product, characterized in that The computer program product includes a computer program or instructions, and when the computer program or instructions are executed by a processor, the battery failure prediction method according to any one of claims 1 to 15 is implemented.

Images