A power battery fault sensitive feature selection method and a fault detection method

Abstract

The application relates to a power battery fault sensitive feature selection method and a fault detection method, and belongs to the technical field of batteries. The selection method divides a sliding window for a voltage sequence of all monomers in each battery pack and extracts multi-dimensional original features; distance features of each monomer and each dimension are constructed; a reference abnormal score sequence of each fault monomer is obtained according to all original features; the correlation coefficient of each distance feature and the reference abnormal score sequence is calculated to obtain a feature correlation ranking; an R scanning loss method is used to reconstruct an abnormal score sequence, and a first round of relevant feature subsets are screened out; feature sensitivity analysis is carried out on the first round of relevant feature subsets to obtain a sensitivity ranking; the final sensitive feature subset is obtained again by R scanning according to the sensitivity ranking. The detection method applies the final sensitive feature subset to online fault detection. The application accurately screens out a small amount of features most sensitive to faults under the premise of ensuring the accuracy of the reconstructed reference abnormal score.

Description

Technical Field

[0001] This invention belongs to the field of battery technology and relates to a method for selecting fault-sensitive features of power batteries and a fault detection method. Background Technology

[0002] With the rapid promotion and popularization of new energy vehicles, the power battery, as the core energy carrier of the entire vehicle, is a crucial component. Power battery packs typically consist of hundreds of cells connected in series and parallel. If any one of these cells fails, it can not only lead to battery performance degradation but, in severe cases, even cause safety accidents such as thermal runaway, threatening the lives and property of passengers. However, there are numerous features for detecting battery faults. How to select appropriate features from this vast pool and effectively detect battery faults is a significant challenge currently facing battery management systems (BMS).

[0003] Existing fault detection methods often fail to consider the relationship between fault type and feature type, resorting to either large feature inputs, feature dimensionality reduction, or single feature inputs. However, these methods have significant shortcomings in practical applications: First, the full feature set contains a large amount of redundant information unrelated to the fault, increasing computational burden and potentially introducing noise, thus reducing detection sensitivity and accuracy. Second, feature dimensionality reduction loses a significant amount of information, ignoring the ability of feature combinations to reconstruct fault representations, meaning the reduced features may not optimally restore the complete fault behavior pattern. Furthermore, a single feature is insufficient to identify the diverse and complex fault types, increasing the risk of false alarms. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a method for selecting fault-sensitive features of power batteries and a fault detection method, which can automatically select a small number of features that are most sensitive to faults for subsequent efficient and accurate fault detection.

[0005] To achieve the above objectives, the present invention provides the following technical solution: On the one hand, a method for selecting fault-sensitive features of power batteries is first provided, which includes: S1. Collect the power battery operation data of the BMS of new energy vehicles, including the single cell voltage operation data of multiple faulty battery packs; S2: Divide the voltage sequence of all cells in each battery pack into a sliding window, and extract multidimensional original features within each window; S3: Construct distance features: For each window and each feature dimension, calculate the median of each feature for all individuals within each window, then subtract the median from each individual feature value and take the absolute value to obtain the distance features; S4: Using all the original features as input to the isolated forest algorithm, obtain the baseline anomaly score sequence for each faulty unit; S5: Calculate the absolute value of the Pearson correlation coefficient between each distance feature and the baseline anomaly score sequence, and average it over multiple faulty individuals to obtain the feature correlation ranking. S6: The R-scan loss method is adopted. The number of candidate features R is traversed. The top-R distance features are used to reconstruct the abnormal score sequence each time. The root mean square error with the benchmark sequence is calculated and a dimension penalty term is added. The feature subset corresponding to R1 that minimizes the comprehensive loss is selected as the first round of relevant feature subset. S7: Within the first set of relevant features, perform feature sensitivity analysis to obtain a sensitivity ranking; S8: Sort by sensitivity and perform R scan again, select the feature subset corresponding to R2 that minimizes the loss as the final sensitive feature subset.

[0006] Furthermore, step S1 includes: S11: Collect power battery data for normal operation of new energy vehicles, including at least battery parameters such as time and voltage; S12: Collect all cell operating data in the faulty power battery pack, including the voltage of normal cells and abnormal cells; S13: Establish a power battery database based on the collected battery data.

[0007] Furthermore, step S2 includes: S21: For each battery pack, let the total number of sampling points be... Set window length Points, sliding step size If a point, Then the number of windows “ " indicates that the quotient is rounded down; if Then the number of windows ; S22: Within each window, for the voltage sequence of each individual, use the TSFEL toolkit to extract multidimensional raw features, including time-domain, frequency-domain, and statistical features, for a total of [number missing]. Dimensional features; Note the first The first window The first monomer The original features are ,in, , , The total number of individuals, .

[0008] Furthermore, step S3 includes: S31: For each window In each feature dimension Above, calculate the median for each feature of all individuals within each window:

[0009] in, This indicates the median operation; S32: Based on feature dimension The median, defining each individual Distance characteristics:

[0010] Distance features Reflecting monomers In features The degree of deviation from the group center; normal individuals are consistent with the group, distance characteristics The smaller the value, the more likely the faulty individual is to deviate from the group, indicating a distance characteristic. The larger; S33: After calculating for each window, each individual, and each feature dimension, the distance feature tensor of all windows is obtained. .

[0011] Furthermore, step S4 includes: S41: For each faulty unit... Using the original features as input, the Isolation Forest algorithm is applied to all cells in each battery pack, including normal and faulty cells, to train the model and obtain the anomaly score for each cell in each window; the smaller the anomaly score, the more abnormal it is; the process of training the model to obtain the anomaly score for each cell is as follows (taking a faulty cell as an example). Faulty unit ( X The anomaly score (where the feature number is 0) is:

[0012] here This refers to outlier scores. It is the average path length of the faulty unit Z. When the number of subsamples is n The normalization factor for the average path length over time.

[0013] The average path length is calculated as follows:

[0014] here t It is the number of isolated trees in the forest. It is a faulty unit. z In the i The path length of each tree.

[0015] The normalization factor is calculated as follows:

[0016] here The harmonic number is calculated as follows:

[0017] here It is the natural logarithm. , is the Euler-Mascheroni constant.

[0018] S42: Note the first The faulty unit is in the first The abnormal score for each window is Then the baseline score sequence for the faulty unit is: ,in, , This represents the total number of faulty individual units.

[0019] Furthermore, step S5 includes: S51: For each distance feature Each faulty unit Calculate the Pearson correlation coefficient from the baseline score series and take the absolute value:

[0020] in, It is a faulty unit. In features The mean of the above, It is a faulty unit. The baseline score mean; S52: Calculation based on S51 The average value is taken from multiple faulty units:

[0021] S53: Press Sort the candidate features from largest to smallest to obtain their ranking.

[0022] Furthermore, step S6 includes: S61: Define the number of candidate features from arrive For each Select the top rankings The distance feature is denoted as the set. ; S62: Using Sets The features in the data are used to retrain the isolated forest model to obtain each faulty individual. In each window New abnormal scores That is, the reconstructed sequence; S63: Calculate the root mean square error between the reconstructed sequence and the reference sequence:

[0023] S64: Add a dimensional penalty term The overall loss is as follows:

[0024] in, The regularization coefficient is used. S65: Drawing Follow The change curve, take smallest The optimal number of features in the first round corresponds to the feature subset. This is the first feature subset.

[0025] Furthermore, step S7 includes: S71: In the first round of feature subsets Within the process, each feature is deleted one by one, the anomaly score sequence is reconstructed using the remaining features, and the comprehensive loss of the remaining features is calculated. Let the baseline loss corresponding to the first feature subset be . Define features The sensitivity index is:

[0026] in, The larger the value, the greater the increase in loss after removing the feature, and the more sensitive the feature is. S72: Sort the features from largest to smallest to obtain a sensitivity-ranked feature list. .

[0027] Furthermore, step S8 includes: S81: Feature list sorted by sensitivity Perform the R scan again to traverse... from arrive For each Before selection Based on the features, reconstruct the anomaly score sequence and recalculate the comprehensive loss. ; S82: Plot the loss curve and select the curve that minimizes the loss. As the final number of features, the corresponding feature subset This is the final subset of sensitive features.

[0028] On the other hand, a fault detection method based on the final sensitive feature subset selected by the aforementioned power battery fault sensitive feature selection method is also provided. This method uses the selected final sensitive feature subset for online fault detection, extracts the corresponding distance features from new battery data, inputs them into the trained isolated forest model, obtains real-time anomaly scores, and determines whether to alarm based on a preset threshold.

[0029] The beneficial effects of this invention are as follows: This invention first constructs a "distance feature"—the absolute value of each individual's deviation from the median of all individuals in the window across each feature dimension—to quantify the difference between an individual and the population, making fault features easier to capture. Then, using the isolated forest anomaly scores of all original feature inputs as a benchmark, the Pearson correlation coefficient between each distance feature and this benchmark is calculated, achieving the first round of correlation screening. Based on this, an innovative "R-scan loss" criterion is introduced. By traversing the number of features R, the difference between the anomaly scores reconstructed using Top-R features and the benchmark scores is examined, and a dimensionality penalty term is added. The subset of features that minimizes the overall loss is selected as the first-round result. Further, features within this subset are deleted one by one, and the increase in loss is evaluated to obtain a sensitivity ranking. A second round of R-scanning is then performed in sensitivity order to finally determine the few features most sensitive to faults and with the strongest reconstruction capabilities. Through these two-stage screening, this invention can automatically remove redundant features and retain the feature subset most representative of fault behavior, thereby significantly improving the accuracy and efficiency of early fault detection and providing reliable support for battery system safety early warning.

[0030] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0031] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of the method for selecting power battery fault-sensitive features in an embodiment of the present invention; Figure 2 The average absolute correlation coefficient ranking curves of 156 features of 3 faulty battery cells in this embodiment of the invention are shown. Figure 3 This is a distance feature bar chart showing the highest correlation among the top 25 faulty battery cells in this embodiment of the invention. Figure 4The curve showing the correlation loss of a faulty battery cell as a function of the number of features in an embodiment of the present invention is shown. Figure 5 This is a curve showing the overall sensitivity loss of a faulty battery cell as a function of the number of features in an embodiment of the present invention. Figure 6 This is a flowchart of a fault detection method based on the fault-sensitive features of a power battery according to an embodiment of the present invention; Figure 7 This is a heatmap showing the abnormal score of 20 battery cell fault detections in an embodiment of the present invention. Detailed Implementation

[0032] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0033] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0034] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0035] Please see Figures 1 to 7 This paper presents a method for selecting fault-sensitive features of power batteries and a fault detection method.

[0036] Example 1 This embodiment first provides a method for selecting fault-sensitive features of a power battery, such as... Figure 1As shown, it specifically includes the following steps: S1: Collects power battery operation data from the BMS of new energy vehicles, including individual cell voltage operation data of multiple faulty battery packs; S2: Divide the voltage sequence of all cells in each battery pack into a sliding window, and extract multidimensional original features within each window; S3: Construct distance features: For each window and each feature dimension, calculate the median of each feature for all individuals within each window, then subtract the median from each individual feature value and take the absolute value to obtain the distance features; S4: Using all the original features as input to the isolated forest algorithm, obtain the baseline anomaly score sequence for each faulty unit; S5: Calculate the absolute value of the Pearson correlation coefficient between each distance feature and the baseline anomaly score sequence, and average it over multiple faulty individuals to obtain the feature correlation ranking. S6: The R-scan loss method is adopted. The number of candidate features R is traversed. The top-R distance features are used to reconstruct the abnormal score sequence each time. The root mean square error with the benchmark sequence is calculated and a dimension penalty term is added. The feature subset corresponding to R1 that minimizes the comprehensive loss is selected as the first round of relevant feature subset. S7: Within the first set of relevant features, perform feature sensitivity analysis to obtain a sensitivity ranking; S8: Sort by sensitivity and perform R scan again, select the feature subset corresponding to R2 that minimizes the loss as the final usable sensitive feature subset.

[0037] As an optional embodiment, step S1 specifically includes S11-S13: S11: Collect power battery data for normal operation of new energy vehicles, including battery parameters such as time and voltage; S12: Collect all cell operating data in the faulty power battery pack, including the voltage of normal cells and abnormal cells; S13: Establish a power battery database based on the collected battery data.

[0038] As an optional embodiment, step S2 specifically includes S21-S22: S21: For each battery pack, let the total number of sampling points be... Set window length Points, sliding step size If a point, Then the number of windows “ " indicates that the quotient is rounded down; S22: Within each window, for the voltage sequence of each individual unit, use the TSFEL toolkit to extract multidimensional raw features, including time-domain, frequency-domain, and statistical features, for a total of 156 dimensions of features. Let the first... windows ( ) individual monomers ( , (the total number of individuals) The original features are .

[0039] As an optional embodiment, step S3 specifically includes S31-S33: S31: For each window In each feature dimension Above, calculate the median of this feature for all individuals within the window:

[0040] in, This indicates the median operation; S32: Based on the median of S31, define the distance feature:

[0041] This characteristic reflects monomers In features The degree to which an individual deviates from the group center is called the distance feature. Normal individuals are consistent with the group and have a small distance; faulty individuals deviate from the group and have a larger distance, making them easier for anomaly detection algorithms to detect. S33: After calculating for each window, each individual, and each feature dimension, the distance feature tensor of all windows is obtained. .

[0042] As an optional embodiment, step S4 specifically includes S41-S42: S41: Using all 156 original features of each faulty cell as input, apply the Isolation Forest algorithm to all cells (including normal and faulty cells) within each battery pack to train the model and obtain the anomaly score for each cell in each window. The smaller the anomaly score in the Isolation Forest, the more abnormal it is. The process of training the model to obtain the anomaly score for each cell is as follows (taking a faulty cell as an example). Faulty unit ( X The anomaly score (where the feature number is 0) is:

[0043] here This refers to outlier scores. It is the average path length of the faulty unit Z. When the number of subsamples is n The normalization factor for the average path length over time.

[0044] The average path length is calculated as follows:

[0045] here t It is the number of isolated trees in the forest. It is a faulty unit. z In the i The path length of each tree.

[0046] The normalization factor is calculated as follows:

[0047] here The harmonic number is calculated as follows:

[0048] here It is the natural logarithm. , is the Euler-Mascheroni constant.

[0049] S42: Note the first One faulty unit ( , (Total number of faulty units) in the 1st The abnormal score for each window is Then the baseline score sequence for the faulty unit is: .

[0050] As an optional embodiment, step S5 specifically includes S51-S53: S51: For each distance feature Each faulty unit Calculate the Pearson correlation coefficient from the baseline score series and take the absolute value:

[0051] in, It is a faulty unit. In features The mean of the above, It is a faulty unit. The baseline score mean; S52: Calculation based on S51 The average value is taken from multiple faulty units:

[0052] in, This represents the number of faulty units; in this example, the number of faulty units is 3. S53: Press Sort the candidate features from largest to smallest to obtain a ranking table. Figure 2 The correlation coefficient ranking curves of 156 features are displayed. Figure 3 The bar chart shows the top 25 most relevant distance features, denoted as Feature 1 to Feature 25.

[0053] As an optional embodiment, step S6 specifically includes S61-S65: S61: Define the number of candidate features from arrive ( This embodiment represents the total number of distance features. However, the scan limit is set to (To save on calculations). For each Select the top rankings The distance feature is denoted as the set. .

[0054] S62: Use these features to retrain the isolated forest model (model parameters are the same as in step S4) to obtain each faulty individual. In each window New abnormal scores That is, the reconstructed sequence.

[0055] S63: Calculate the root mean square error between the reconstructed sequence and the reference sequence:

[0056] S64: Add a dimensional penalty term The overall loss is as follows:

[0057] in, The regularization coefficient is 1 in this embodiment. This represents the current number of features.

[0058] S65: Drawing Follow The change curve, such as Figure 4 As shown, take smallest The optimal number of features in the first round corresponds to the feature subset. This is the first-round feature subset. In this embodiment... .

[0059] As an optional embodiment, step S7 specifically includes S71-S72: S71: In the first round of feature subsets Within the process, each feature is deleted one by one, the anomaly score sequence is reconstructed using the remaining features, and the comprehensive loss of the remaining features is calculated. Let the baseline loss corresponding to the first feature subset be . Define features The sensitivity index is:

[0060] in, The larger the value, the greater the increase in loss after removing the feature, and the more sensitive the feature is.

[0061] S72: Sort the features from largest to smallest to obtain a sensitivity-ranked feature list. .

[0062] As an optional embodiment, step S8 specifically includes S81-S82: S81: Feature list sorted by sensitivity Perform the R scan again to traverse... from arrive For each Before selection Based on the features, reconstruct the anomaly score sequence and calculate the comprehensive loss. (The formula is the same as step S6).

[0063] S82: Plot the loss curve as follows Figure 5 Take the option that minimizes the loss. As the final number of features, the corresponding feature subset This is the final set of usable sensitive features. In this embodiment... The three features selected were feature 4, feature 1, and feature 3.

[0064] Example 2 This embodiment provides a method for fault detection based on features selected according to the power battery fault sensitive feature selection method in Embodiment 1. The selected sensitive feature subset is used for online fault detection. Corresponding distance features are extracted from new battery data, input into a trained isolated forest model, and a real-time anomaly score is obtained. An alarm is then triggered based on a preset threshold. Figure 6 As shown, the specific steps include the following: S1: The voltage data of all battery cells of the new energy vehicle is used as input (in this embodiment, the voltage data of 20 battery cells in a battery pack are selected). The final usable sensitive feature subset of S8 in embodiment 1 is extracted according to the time window. In this embodiment, it is feature 4, feature 1 and feature 3.

[0065] S2: Perform Z-Score standardization on the three selected features, that is, standardize the features to a mean of 0 and a variance of 1 to prevent the dimension of a certain feature from being too large and affecting the detection results; input the standardized features into the isolated forest algorithm in S41 of Example 1 to obtain the anomaly score.

[0066] As an optional embodiment, the Z-Score normalization process in step S2 has the following formula:

[0067] in, Features The standardized value Features The original value, The mean of all selected features. The standard deviation for all selected features.

[0068] S3: Compare the abnormal scores of all individual battery cells within a given window with a fixed threshold. (In this example, we take -0.2) We compare sequentially; if a battery's abnormal score in a certain window falls below a fixed threshold, it is determined to be a faulty cell; otherwise, it is a normal cell. The detection results are as follows: Figure 7 Anomalies with a score below -0.2 are indicated by a darker color, representing the fault window of the faulty unit. Figure 7 The faulty cell is cell number 16.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for selecting fault-sensitive features of a power battery, characterized in that: The method includes: S1. Collect the power battery operation data of the new energy vehicle battery management system, including the individual cell voltage operation data of multiple faulty battery packs; S2: Divide the voltage sequence of all cells in each battery pack into a sliding window, and extract multidimensional original features within each window; S3: Construct distance features: For each window and each feature dimension, calculate the median of each feature for all individuals within each window, then subtract the median from each individual feature value and take the absolute value to obtain the distance features; S4: Using all the original features as input to the isolated forest algorithm, obtain the baseline anomaly score sequence for each faulty unit; S5: Calculate the absolute value of the Pearson correlation coefficient between each distance feature and the baseline anomaly score sequence. For each distance feature, average the absolute values ​​of the Pearson correlation coefficient between the distance feature and the baseline anomaly score sequence for all faulty units to obtain the feature correlation and sort them from largest to smallest to obtain the candidate feature ranking. S6: The R-scan loss method is adopted. The number of candidate features R is traversed. The top-R distance features are used to reconstruct the anomaly score sequence each time. The root mean square error between the anomaly score sequence and the benchmark anomaly score sequence is calculated. A dimension penalty term is added. The feature subset corresponding to R1 that minimizes the comprehensive loss is selected as the first round of relevant feature subset. S7: In the first round of relevant feature subsets, perform feature sensitivity analysis to obtain sensitivity ranking. In the first round of feature subsets, delete each feature one by one, use the remaining features to reconstruct the abnormal score sequence, and calculate the comprehensive loss of the remaining features. The difference between the comprehensive loss of the remaining features and the benchmark loss corresponding to the first round of feature subsets is defined as the feature sensitivity index. Sensitivity ranking is performed based on the size of the sensitivity index. S8: Sort by sensitivity and perform R scan again, select the feature subset corresponding to R2 that minimizes the loss as the final sensitive feature subset.

2. The method for selecting fault-sensitive features of a power battery according to claim 1, characterized in that: Step S1 includes: S11: Collect power battery data for normal operation of new energy vehicles, including at least battery parameters such as time and voltage; S12: Collect all cell operating data in the faulty power battery pack, including the voltage of normal cells and abnormal cells; S13: Establish a power battery database based on the collected battery data.

3. The method for selecting fault-sensitive features of a power battery according to claim 1, characterized in that: Step S2 includes: S21: For each battery pack, let the total number of sampling points be... Set window length Points, sliding step size If a point, Then the number of windows " " indicates that the quotient is rounded down; if Then the number of windows ; S22: Within each window, for the voltage sequence of each individual, use the TSFEL toolkit to extract multidimensional raw features, including time-domain, frequency-domain, and statistical features, for a total of [number missing]. Dimensional features; Note the first The first window The first monomer The original features are ,in, , , The total number of individuals, .

4. The method for selecting fault-sensitive features of a power battery according to claim 3, characterized in that: Step S3 includes: S31: For each window In each feature dimension Above, calculate the median of this feature for all individuals within the window: in, This indicates the median operation; S32: Based on feature dimension The median, defining each individual Distance characteristics: Distance features Reflecting monomers In features The degree of deviation from the group center; normal individuals are consistent with the group, distance characteristics The smaller the value, the more likely the faulty individual is to deviate from the group, indicating a distance characteristic. The larger; S33: After calculating for each window, each individual, and each feature dimension, the distance feature tensor of all windows is obtained. .

5. The method for selecting fault-sensitive features of a power battery according to claim 4, characterized in that: Step S4 includes: S41: For each faulty unit... Using the original features as input, the Isolation Forest algorithm is applied to all cells within each battery pack, including both normal and faulty cells, to train the model and obtain anomaly scores for each cell in each window. Lower anomaly scores indicate greater abnormality. The process of training the model to obtain the anomaly score for each cell is as follows: Assume a faulty unit , X If the feature number is denoted as , then its anomaly score is: In the formula, This refers to outlier scores. It is the average path length of the faulty unit Z. When the number of subsamples is n The normalization factor for the average path length over time; The average path length is calculated as follows: In the formula, t It is the number of isolated trees in the forest. It is a faulty unit. z In the i The path length of each tree; The normalization factor is calculated as follows: In the formula, For harmonic numbers, , It is the natural logarithm. It is the Euler-Mascheroni constant; S42: Note the first The faulty unit is in the first The abnormal score for each window is Then the baseline anomaly score sequence for the faulty unit is: ,in, , This represents the total number of faulty individual units.

6. The method for selecting fault-sensitive features of a power battery according to claim 5, characterized in that: Step S5 includes: S51: For each distance feature Each faulty unit The Pearson correlation coefficient was calculated from the baseline abnormality score sequence, and the absolute value was taken: in, It is a faulty unit. In features The mean of the above, It is a faulty unit. The baseline score mean; S52: Calculation based on S51 The average value is taken from multiple faulty units: S53: Press Sort the candidate features from largest to smallest to obtain their ranking.

7. The method for selecting fault-sensitive features of a power battery according to claim 6, characterized in that: Step S6 includes: S61: Define the number of candidate features from arrive For each Select the top rankings The distance feature is denoted as set. ; S62: Using Sets The features in the data are used to retrain the isolated forest model to obtain each faulty individual. In each window New abnormal scores That is, to reconstruct the abnormal score sequence; S63: Calculate the root mean square error between the reconstructed outlier score sequence and the baseline outlier score sequence: S64: Add a dimensional penalty term The overall loss is as follows: in, The regularization coefficient is used. S65: Drawing Follow The change curve, take smallest The optimal number of features in the first round corresponds to the feature subset. This is the first feature subset.

8. The method for selecting fault-sensitive features of a power battery according to claim 7, characterized in that: Step S7 includes: S71: In the first round of feature subsets Within the process, each feature is deleted one by one, the anomaly score sequence is reconstructed using the remaining features, and the comprehensive loss of the remaining features is calculated. Let the baseline loss corresponding to the first feature subset be . Define features The sensitivity index is: in, The larger the value, the greater the increase in loss after removing the feature, and the more sensitive the feature is. S72: Sort the features from largest to smallest to obtain a sensitivity-ranked feature list. .

9. The method for selecting fault-sensitive features of a power battery according to claim 8, characterized in that: Step S8 includes: S81: Feature list sorted by sensitivity Perform the R scan again to traverse... from arrive For each Before selection Based on the features, reconstruct the anomaly score sequence and recalculate the comprehensive loss. ; S82: Plot the loss curve and select the curve that minimizes the loss. As the final number of features, the corresponding feature subset This is the final subset of sensitive features.

10. A fault detection method based on the final sensitive feature subset selected by the power battery fault sensitive feature selection method according to any one of claims 1-9, characterized in that: The selected subset of sensitive features is used for online fault detection. The corresponding distance features are extracted from the new battery data, input into the trained isolated forest model, and real-time anomaly scores are obtained. An alarm is then triggered based on a preset threshold.

Images