An online fault diagnosis method for main transformer based on missing data repair

Abstract

The application discloses an online fault diagnosis method for a main transformer based on missing data repair. In order to overcome the problem of missing data of transformer oil chromatography online monitoring, the application comprises the following steps: S1: sequentially segmenting the collected oil chromatography data samples in different index dimensions, and constructing a tree model; S2: determining the data set of the end of the tree branch in the tree model corresponding to different fault types according to the association relationship between the fault types and the index dimensions; S3: judging whether the missing data in the data set affects the fault diagnosis, if yes, entering step S4, otherwise, directly entering step S5; S4: repairing the missing data according to the most adjacent tree branch node data obtained by Manhattan distance searching; and S5: matching historical data to diagnose the fault type. Compared with the existing historical data comparison, the missing data is repaired by searching the most adjacent node data, which is more accurate.

Description

Technical Field

[0001] This invention relates to the field of main transformer fault diagnosis, and more particularly to an online fault diagnosis method for main transformers based on missing data repair. Background Technology

[0002] The task of fault diagnosis is to predict, analyze, and judge potential or already occurring faults based on information obtained from condition monitoring, combined with known system structure and parameters, environmental conditions, or historical operating data of system equipment (including historical operating records, past fault records, and maintenance records). This involves determining the nature, category, degree, cause, and location of the fault, indicating the trend of fault occurrence and development and its consequences, proposing and implementing countermeasures for controlling the continued development of the fault and eliminating it, and ultimately restoring the equipment to a normal state. Different types and locations of faults in different parts of production units or equipment cause different changes in function, leading to different changes in the state and operating parameters of the unit as a whole and its individual parts. The purpose of fault diagnosis is to infer the type and location of the fault causing these changes from the changes in these states and parameters when a certain fault occurs in a certain part. Because the amount of process parameter data is very large, it is necessary to identify and extract characteristic information to facilitate fault diagnosis.

[0003] Data quality is a crucial factor affecting the accuracy of transformer fault diagnosis. However, various factors can lead to missing online monitoring data, and current fault diagnosis methods do not consider the related issues of data loss. For example, a fault diagnosis method and system for transformer oil chromatography based on multiple linear regression, disclosed in Chinese patent literature (publication number CN114152685A), includes the following steps: acquiring transformer characteristic gas information; obtaining transformer fault diagnosis results based on the characteristic gas information and a preset transformer oil chromatography fault diagnosis model; the transformer oil chromatography fault diagnosis model uses a multiple linear regression model to predict the components of the characteristic gases, and judges the transformer operating status based on the predicted characteristic gas quantities and preset threshold values. This solution enables real-time monitoring of the transformer operating status, timely detection of potential internal faults in the transformer, prevention of problems before they occur, and improvement of the reliability of power system operation. However, this solution does not consider the problem of missing online monitoring data for transformer oil chromatography. Summary of the Invention

[0004] This invention primarily addresses the problem of missing data in existing transformer oil chromatography online monitoring technologies; it provides a method for online fault diagnosis of main transformers based on missing data repair, obtaining the nearest neighbor data to repair the missing data.

[0005] The above-mentioned technical problems of the present invention are mainly solved by the following technical solutions:

[0006] A method for online fault diagnosis of main transformers based on missing data repair includes the following steps:

[0007] S1: The collected oil chromatographic data samples are sequentially segmented according to different index dimensions to construct a tree model;

[0008] S2: Based on the correlation between fault type and indicator dimension, determine the data set at the end of the branches in the tree model corresponding to different fault types;

[0009] S3: Determine whether missing data in the dataset affects fault diagnosis. If yes, proceed to step S4; otherwise, proceed directly to step S5.

[0010] S4: Based on the Manhattan distance, find the nearest neighboring tree node data and repair the missing data;

[0011] S5: Match historical data to diagnose fault types.

[0012] This application's solution segments data samples across different dimensions, resulting in better balance for each sample. It repairs missing data by finding the nearest neighboring nodes, a more accurate method compared to existing methods that rely on historical data comparison.

[0013] Preferably, the oil chromatography data sample includes the content of several gaseous components in soluble oils; each gaseous component is used as an indicator dimension. Using the type of gaseous component as the segmentation dimension ensures the balance of the tree model.

[0014] Preferably, the tree model construction process is as follows:

[0015] S101: Calculate the variance of each indicator dimension in the data sample respectively;

[0016] S102: Take the dimension with the largest variance as the splitting dimension, and arrange the data samples of each group in order of the size of the data in the splitting dimension; the larger the variance, the better the dispersion of the data in that dimension, and splitting in this direction can effectively ensure the balance of the tree.

[0017] S103: Divide a data queue into two data sub-queues using the data sample in the middle of the queue as the dividing line; S104: Determine whether the number of data samples in the data sub-queue is less than the number threshold or whether the number of divisions has reached the division threshold; if yes, complete the division of the branch node; if no, return to step S101 for each data sub-queue.

[0018] Each node in the tree-like model is a multi-dimensional binary tree, and all non-leaf nodes can be viewed as a spatially partitioned hyperplane. By performing non-overlapping hierarchical partitioning of the search space, an index structure suitable for efficient data retrieval is established. When this method is applied to the structured storage of transformer oil chromatography data, each oil chromatography sample contains multi-dimensional indicators and is stored in a node of the tree-like model.

[0019] Preferably, each indicator dimension of each branch node in the tree model is labeled with a segmentation coefficient; during the segmentation process to obtain each branch node, the segmentation coefficient corresponding to the segmentation dimension of each previous segmentation is incremented by one. Data is segmented based on variance; a larger variance indicates better data dispersion along that dimension, and segmenting in this direction effectively ensures the balance of the tree. This scheme is used for subsequent calculations of the correlation between fault types and dimension indicators.

[0020] Preferably, the fault types include two main categories: overheating faults and discharge faults.

[0021] The overheating faults mentioned include high-temperature overheating, medium-temperature overheating, and low-temperature overheating;

[0022] The discharge faults mentioned include partial discharge, high-energy discharge, and low-energy discharge.

[0023] Different fault types are associated with different dimensions.

[0024] Preferably, step S2 specifically includes the following process:

[0025] S201: Construct a table relating fault types to indicator dimensions; the table includes the primary and secondary indicator dimension features corresponding to each fault type.

[0026] S202: Assign dimensional weight values ​​to the primary and secondary indicator dimensional features of each fault type;

[0027] S203: Taking a specific fault type as the diagnostic target, calculate the correlation index of the data samples at the end of each branch in the tree model using the corresponding dimension weight values;

[0028] S204: The set of data samples whose correlation index is greater than the correlation threshold is taken as the diagnostic data sample set.

[0029] Select a set of data samples that are highly correlated with the type of fault to be diagnosed for the diagnosis of the corresponding fault type.

[0030] As a preferred method, the correlation index is calculated as follows:

[0031]

[0032] Among them, R yn The correlation index of the data sample at the end of the nth branch in the tree model corresponding to the yth type of fault;

[0033] ω yk Let y be the dimension weight value of the k-th indicator dimension corresponding to the y-th fault type;

[0034] K represents the total number of indicator dimensions;

[0035] f nk It represents the segmentation coefficient of the data sample at the end of the nth branch in the tree model in the kth index dimension.

[0036] The correlation index is calculated based on the number of times the segmentation dimension is used during segmentation.

[0037] Preferably, step S3 specifically includes the following process:

[0038] S301: Determine whether there are missing data in the diagnostic data sample set; if yes, proceed to step S302; otherwise, proceed to step S5.

[0039] S302: Query the relationship between fault type and indicator dimension. For the fault type to be diagnosed, if there is missing data for the main indicator dimension feature or the secondary indicator dimension feature, proceed to step S4; otherwise, proceed directly to step S5.

[0040] Determine whether the missing data will affect subsequent diagnosis. If it will, repair the data; otherwise, ignore the missing data.

[0041] Preferably, step S4 includes the following steps:

[0042] S401: Identify the branch ends in a tree model where there are missing data for either the primary or secondary indicator dimension features.

[0043] S402: Obtain all data samples from the previous level tree node;

[0044] S403: Determine whether there are missing data of the same indicator dimension in the data sub-queues belonging to the same parent tree node; if so, return to step S402; otherwise, select the data sub-queue as the repair queue and proceed to step S404.

[0045] S404: Calculate the Manhattan distance between the dataset to be repaired and the repair queue, and determine whether it is less than the distance threshold; if so, take the average value of the data in the corresponding indicator dimension in the repair queue as the repair value; otherwise, return to step S402 until tracing back to the top level.

[0046] By using Manhattan distance to find the nearest data sample, the data can be repaired and its accuracy improved.

[0047] Preferably, the Manhattan distance calculation process is as follows:

[0048]

[0049]

[0050] Where, ρ pq Let be the correlation coefficient between the gas content of the p-th indicator dimension and the gas content of the q-th indicator dimension; the value range is [-1, 1], where -1 indicates a strong negative correlation between the two indicator dimensions, 1 indicates a strong positive correlation, and 0 indicates no correlation.

[0051] N is the number of data samples;

[0052] x ip For the data set x to be repaired i The p-th indicator dimension;

[0053] x iq For the data set x to be repaired i The q-th indicator dimension;

[0054] μ p Let be the mean of the p-th indicator dimension;

[0055] μ q Let be the mean of the q-th indicator dimension;

[0056] x jq To repair queue x j The q-th indicator dimension;

[0057] d ij|p For the data set x to be repaired i With repair queue x j Manhattan distance for the p-th index dimension.

[0058] By assigning spatial distance weights to gas indicators using correlation coefficients, the correlation between gas components can be reflected. Gas indicators with a higher correlation to missing gases have lower weights, making it easier for test samples to approach training samples that are similar to strongly correlated gas indicators.

[0059] As a preferred option, the correlation index of the data sub-queues in other tree nodes at the same level as the previous tree node is calculated; if the correlation index of the data sub-queue is less than the correlation index of the data set to be repaired, then the data sub-queue is determined to be the repair queue.

[0060] Data repair is best performed by selecting data subqueues of the same level, resulting in more accurate repair results.

[0061] As a preferred approach, a multi-class SVM model is used to diagnose faults in the repair queue. The model determines whether the diagnosed fault type matches the fault type required for the data set to be repaired. If so, the data in the repair queue is used for repair; otherwise, the repair queue is removed. This narrows down the scope of the repair queue.

[0062] Preferably, the fault diagnosis process is as follows:

[0063] The data of each indicator dimension in the diagnostic data sample is matched with the historical fault data to calculate the matching coefficient. If the matching coefficient is greater than the matching threshold, it is determined that a fault of the corresponding fault type has occurred; otherwise, it is determined that a fault of the corresponding fault type has not occurred.

[0064] By matching historical data, it can be determined whether the same type of failure has occurred.

[0065] Preferably, the matching coefficient calculation process is as follows:

[0066]

[0067] Among them, C y Let y be the matching coefficient for the y-th fault type;

[0068] ω yk Let y be the dimension weight value of the k-th indicator dimension corresponding to the y-th fault type;

[0069] D ks This refers to the data for the k-th indicator dimension in the diagnostic data sample.

[0070] D kh This refers to the data for the k-th indicator dimension in the historical fault data.

[0071] The similarity to historical fault data is calculated to diagnose whether a fault has occurred.

[0072] The beneficial effects of this invention are:

[0073] 1. By segmenting the data samples according to different dimensions, each data sample can have better balance.

[0074] 2. Missing data is repaired by finding the nearest neighboring node data, which is more accurate than the existing method of obtaining data by comparing historical data.

[0075] 3. The correlation index is calculated based on the number of times the segmentation dimension is used during the segmentation, which makes the selection of data samples more accurate.

[0076] 4. Ignore missing data that will not affect fault diagnosis, thus improving diagnosis speed.

[0077] 5. Calculate the similarity with historical fault data to diagnose whether a fault has occurred. Attached Figure Description

[0078] Figure 1 is a flowchart of the online fault diagnosis method for main transformers according to the present invention. Detailed Implementation

[0079] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0080] Example:

[0081] This embodiment presents an online fault diagnosis method for main transformers based on missing data repair, such as... Figure 1 As shown, an online fault diagnosis method for main transformers based on missing data repair includes the following steps:

[0082] S1: The collected oil chromatographic data samples are sequentially segmented according to different index dimensions to construct a tree model.

[0083] The oil chromatographic data sample includes the content of several soluble gaseous components in the oil; each gaseous component is used as an indicator dimension. In this embodiment, five dissolved gas indicators in the oil are included: H2, CH4, C2H6, C2H4, and C2H2.

[0084] Using the types of gas components as the segmentation dimension ensures the balance of the tree model.

[0085] The process of building a tree model is as follows:

[0086] S101: Calculate the variance of each indicator dimension in the data sample respectively.

[0087] S102: Take the index dimension with the largest variance as the splitting dimension, and arrange the data samples of each group in order of the size of the data in the splitting dimension.

[0088] A larger variance indicates better dispersion of data in that dimension, and splitting in this direction can effectively ensure the balance of the tree.

[0089] S103: Divide a data queue into two data sub-queues using the data sample in the middle of the queue as the dividing line.

[0090] In the tree model, each index dimension of each branch node is labeled with a segmentation coefficient.

[0091] During the process of obtaining each branch node through segmentation, the segmentation coefficient corresponding to the segmentation dimension of each segmentation is incremented by one. For example, the segmentation coefficient corresponding to the branch node obtained after segmenting once with the five dimensions H2, CH4, C2H6, C2H4 and C2H2 is (1, 1, 1, 1, 1), that is, the segmentation coefficient of each dimension is 1.

[0092] S104: Determine whether the number of data samples in the data sub-queue is less than the quantity threshold or whether the number of splits has reached the split threshold; if yes, complete the split of the branch node; if no, return to step S101 for each data sub-queue.

[0093] Each node in the tree-like model is a multi-dimensional binary tree, and all non-leaf nodes can be viewed as a spatially partitioned hyperplane. By performing non-overlapping hierarchical partitioning of the search space, an index structure suitable for efficient data retrieval is established. When this method is applied to the structured storage of transformer oil chromatography data, each oil chromatography sample contains multi-dimensional indicators and is stored in a node of the tree-like model.

[0094] S2: Based on the correlation between fault type and indicator dimension, determine the data set at the end of the branches in the tree model corresponding to different fault types.

[0095] The types of faults are divided into two main categories: overheating faults and discharge faults.

[0096] Overheating faults include high temperature overheating (above 700℃), medium temperature overheating (300-700℃), and low temperature overheating (below 300℃).

[0097] Discharge faults include partial discharge, high-energy discharge, and low-energy discharge.

[0098] Step S2 specifically includes the following processes:

[0099] S201: Construct a table relating fault types to indicator dimensions; the table includes the primary and secondary indicator dimensions for each fault type.

[0100] S202: Assign dimensional weight values ​​to the primary and secondary indicator dimensional features of each fault type.

[0101] S203: Using a specific fault type as the diagnostic target, calculate the correlation index of the data samples at the end of each branch in the tree model using the corresponding dimension weight values.

[0102] The calculation process for the correlation index is as follows:

[0103]

[0104] Among them, R ynThe correlation index of the data sample at the end of the nth branch in the tree model corresponding to the yth type of fault;

[0105] ω yk Let y be the dimension weight value of the k-th indicator dimension corresponding to the y-th fault type;

[0106] K represents the total number of indicator dimensions;

[0107] f nk It represents the segmentation coefficient of the data sample at the end of the nth branch in the tree model in the kth index dimension.

[0108] The correlation index is calculated based on the number of times the segmentation dimension is used during segmentation.

[0109] For example, for a branch node with a segmentation coefficient of (1, 1, 1, 1, 1), CH4 corresponding to the diagnosed fault is the primary indicator dimension feature, and C2H2 and C2H4 are secondary indicator dimension features. Then, CH4 is assigned a weight of 3, C2H2 and C2H4 are assigned a weight of 2, and other dimension indicators are assigned a weight of 0.5. The calculated correlation index result is 8.

[0110] S204: The set of data samples whose correlation index is greater than the correlation threshold is taken as the diagnostic data sample set.

[0111] Select a set of data samples that are highly correlated with the type of fault to be diagnosed for the diagnosis of the corresponding fault type.

[0112] S3: Determine whether missing data in the dataset affects fault diagnosis. If yes, proceed to step S4; otherwise, proceed directly to step S5.

[0113] Step S3 specifically includes the following processes:

[0114] S301: Determine whether there is missing data in the diagnostic data sample set; if yes, proceed to step S302 for judgment; otherwise, proceed to step S5.

[0115] S302: Query the relationship between fault type and indicator dimension. For the fault type to be diagnosed, if there is missing data for the main indicator dimension feature or the secondary indicator dimension feature, proceed to step S4; otherwise, proceed directly to step S5.

[0116] Determine whether the missing data will affect subsequent diagnosis. If it will, repair the data; otherwise, ignore the missing data.

[0117] S4: Based on the Manhattan distance, find the nearest neighboring tree node data and repair the missing data.

[0118] Step S4 includes the following steps:

[0119] S401: Identify the branch ends in a tree model where there is missing data for either the primary or secondary indicator dimension features.

[0120] S402: Retrieve all data samples from the previous level tree node.

[0121] Calculate the association index of data subqueues in other tree nodes at the same level as the previous tree node;

[0122] If the correlation index of a data subqueue differs from the correlation index of the data set to be repaired by less than the nominal threshold, then the data subqueue is determined to be a repair queue.

[0123] Data repair is best performed by selecting data subqueues of the same level, resulting in more accurate repair results.

[0124] S403: Determine whether there are missing data of the same indicator dimension in the data sub-queues belonging to the same parent-level tree node; if so, return to step S402; otherwise, select the data sub-queue as the repair queue and proceed to step S404.

[0125] A multi-class SVM model is used to diagnose faults in the repair queue. The model determines whether the diagnosed fault type matches the fault type required for the data set to be repaired. If so, the data in the repair queue is used for repair; otherwise, the repair queue is removed. This narrows down the scope of the repair queue.

[0126] S404: Calculate the Manhattan distance between the dataset to be repaired and the repair queue, and determine whether it is less than the distance threshold; if so, take the average value of the data in the corresponding indicator dimension in the repair queue as the repair value; otherwise, return to step S402 until tracing back to the top level.

[0127] By using Manhattan distance to find the nearest data sample, the data can be repaired and its accuracy improved.

[0128] The Manhattan distance calculation process is as follows:

[0129]

[0130]

[0131] Where, ρ pq Let be the correlation coefficient between the gas content of the p-th indicator dimension and the gas content of the q-th indicator dimension; the value range is [-1, 1], where -1 indicates a strong negative correlation between the two indicator dimensions, 1 indicates a strong positive correlation, and 0 indicates no correlation.

[0132] N is the number of data samples;

[0133] x ipFor the data set x to be repaired i The p-th indicator dimension;

[0134] x iq For the data set x to be repaired i The q-th indicator dimension;

[0135] μ p Let be the mean of the p-th indicator dimension;

[0136] μ q Let be the mean of the q-th indicator dimension;

[0137] x jq To repair queue x j The q-th indicator dimension;

[0138] d ij|p For the data set x to be repaired i With repair queue x j Manhattan distance for the p-th index dimension.

[0139] By assigning spatial distance weights to gas indicators using correlation coefficients, the correlation between gas components can be reflected. Gas indicators with a higher correlation to missing gases have lower weights, making it easier for test samples to approach training samples that are similar to strongly correlated gas indicators.

[0140] S5: Match historical data to diagnose fault types.

[0141] The fault diagnosis process is as follows:

[0142] The data of each indicator dimension in the diagnostic data sample are matched with historical fault data to calculate the matching coefficient. If the matching coefficient is greater than the matching threshold, it is determined that a fault of the corresponding fault type has occurred; otherwise, it is determined that a fault of the corresponding fault type has not occurred.

[0143] By matching historical data, it can be determined whether the same type of failure has occurred.

[0144] Preferably, the matching coefficient calculation process is as follows:

[0145]

[0146] Among them, C y Let y be the matching coefficient for the y-th fault type;

[0147] ω yk Let y be the dimension weight value of the k-th indicator dimension corresponding to the y-th fault type;

[0148] D ks This refers to the data for the k-th indicator dimension in the diagnostic data sample.

[0149] D kh This refers to the data for the k-th indicator dimension in the historical fault data.

[0150] The similarity to historical fault data is calculated to diagnose whether a fault has occurred.

[0151] This embodiment's solution segments data samples across different dimensions, resulting in better balance for each sample. It repairs missing data by finding the nearest neighboring node, a more accurate method compared to existing methods that rely on historical data comparison.

[0152] It should be understood that the embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

Claims

1. A method for on-line fault diagnosis of a main transformer based on missing data repair, characterized in that, The method comprises the following steps: S1: sequentially segmenting the collected oil chromatographic data samples in different index dimensions to construct a tree model; S2: determining the data set of the end of the tree branch in the tree model corresponding to different fault types according to the association relationship between the fault types and the index dimensions; S3: judging whether the missing data in the data set affects fault diagnosis, if yes, entering step S4, otherwise, directly entering step S5; S4: repairing the missing data according to the most adjacent tree branch node data obtained by Manhattan distance searching; S5: matching historical data to diagnose the fault type.

2. The method according to claim 1, wherein, The oil chromatographic data sample includes the content of gas components in several soluble oils; and each gas component is taken as an index dimension.

3. The method according to claim 1 or 2, characterized in that, The tree model construction process comprises the following steps: S101: calculating the variance of each index dimension in the data sample; S102: taking the index dimension with the largest variance value as the segmentation dimension, and arranging the data samples in each group according to the size of the data in the segmentation dimension; S103: taking the data sample in the middle of the queue as the segmentation boundary, and dividing one data queue into two data sub-queues; S104: judging whether the number of data samples in the data sub-queue is less than the number threshold or the segmentation times reach the segmentation threshold; If yes, the segmentation of the tree branch node is completed; if no, returning to step S101 for each data sub-queue.

4. The method according to claim 3, wherein, Each index dimension of each tree branch node in the tree model is marked with a segmentation coefficient; and the segmentation coefficient corresponding to the segmentation dimension of each segmentation is added by one in the process of segmenting to obtain each tree branch node.

5. The method of claim 1, wherein the method is characterized by, The fault types include two categories of overheat fault and discharge fault; The overheat fault includes high-temperature overheat, medium-temperature overheat and low-temperature overheat; The discharge fault includes partial discharge, high-energy discharge and low-energy discharge.

6. The method of claim 4, wherein the method is characterized by, The step S2 specifically comprises the following process: S201: constructing a fault type and index dimension association table; the association table includes the main index dimension characteristics and the secondary index dimension characteristics corresponding to each fault type; S202: assigning a dimension weight value to the main index dimension characteristics and the secondary index dimension characteristics of each fault type; S203: taking a specific fault type as a diagnosis target, and calculating the association index of the data sample of each tree branch end in the tree model according to the corresponding dimension weight value; S204: taking the data sample set with the association index greater than the association threshold as the diagnosis data sample set.

7. The method according to claim 6, wherein, The calculation process of the association index is as follows: wherein R yn a correlation index of a data sample at the end of the nth branch in the tree model corresponding to the yth fault type; ω yk is a dimension weight value of the kth index dimension corresponding to the yth fault type; K is the total number of index dimensions; f nk is the splitting coefficient of the data sample at the end of the nth branch of the tree model in the kth index dimension.

8. The method according to claim 6, characterized in that, The step S3 specifically comprises the following process: S301: judging whether there is data missing in the diagnosis data sample set; If yes, entering step S302; otherwise, entering step S5; S302: querying the association relationship between the fault types and the index dimensions, and if there is data missing of the main index dimension characteristics or the secondary index dimension characteristics for the fault type to be diagnosed, entering step S4; otherwise, directly entering step S5. The step S4 comprises the following steps:

9. The method according to claim 8, wherein, S401: determining the end of the tree branch in the tree model with missing data of the main index dimension characteristics or the secondary index dimension characteristics; S402: obtaining all data samples in the upper tree branch node; ​ S403: judging whether there is data missing of the same index dimension in the data sub-queue belonging to the same upper level tree branch node; if yes, returning to step S402; otherwise, selecting the data sub-queue as a repair queue, and entering step S404; S404: calculating the Manhattan distance between the to-be-repaired data set and the repair queue, and judging whether it is less than the distance threshold; if yes, taking the average value of the data in the corresponding index dimension in the repair queue as the repair value; otherwise, returning to step S402 until the uppermost level is traced back.

10. The method according to claim 9, wherein, The Manhattan distance calculation process is: wherein, p pq is the correlation coefficient of the gas content of the pth index dimension and the gas content of the qth index dimension; the value range is [-1, 1], -1 indicates strong negative correlation between the two index dimensions, 1 indicates strong positive correlation, and 0 indicates no correlation; N is the number of data samples; x ip for the data set x to be repaired i the p-th index dimension in the data set x x iq for the data set x to be repaired i the qth index dimension in the data set μ p is the mean value for the p-th index dimension; μ q is the mean for the qth index dimension; x jq to repair the queue x j the qth index dimension d ij|p For the data set x to be repaired i With the repair queue x j Manhattan distance with respect to the p-th index dimension.

11. The method of claim 9, wherein the method is characterized by, calculating the association index of the data sub-queue in other tree branch nodes at the same level as the upper level tree branch node; if the association index of the data sub-queue and the association index of the to-be-repaired data set only differ by less than a rated threshold, the data sub-queue is determined as the repair queue.

12. The method according to claim 10 or 11, characterized in that, Using a multi-classification SVM model to diagnose the fault of the repair queue, judging whether the diagnosed fault type is the same as the fault type required to be diagnosed by the to-be-repaired data set, if yes, using the repair queue to repair the data; otherwise, eliminating the repair queue.

13. The method of claim 1, wherein the method is characterized by, The fault diagnosis process is: matching the data of each index dimension in the diagnostic data sample with the historical fault data, calculating the matching coefficient; if the matching coefficient is greater than the matching threshold, it is judged that the corresponding fault type fault occurs; otherwise, it is judged that the corresponding fault type fault does not occur.

14. The method according to claim 13, wherein, The matching coefficient calculation process is: wherein C y is the matching coefficient for the yth fault type; ω yk is a dimension weight value of the kth index dimension corresponding to the yth fault type; D ks diagnostic data sample for the kth indicator dimension; D kh is the data of the kth indicator dimension in the historical failure data.

Images