Method, device and equipment for detecting interpretable anomaly of top drive drilling device and medium

By selecting feature parameters based on sensitivity and conflict indices and constructing a memory matrix using clustering algorithms, and combining this with multivariate state estimation methods for anomaly detection of top-driven drilling rigs, the problem of low detection accuracy and efficiency in existing technologies is solved, achieving efficient and accurate anomaly detection and tracing.

CN122065057APending Publication Date: 2026-05-19CHINA UNIV OF PETROLEUM (BEIJING)
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610084531.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-01-21
Publication Date
2026-05-19

AI Technical Summary

Technical Problem

Existing methods for detecting anomalies in top-drive drilling rigs do not adequately consider the complexity and correlation of operating conditions when selecting key monitoring variables, resulting in reduced accuracy and efficiency of detection results, and also have the problem of redundant memory matrix calculations.

Method used

The memory matrix is ​​constructed based on sensitivity and conflict indices. Feature parameters are selected and clustering algorithms are used to select sample data. The memory matrix is ​​constructed using the selected feature parameters and sample data. Anomaly detection is performed using multivariate state estimation. The source anomaly parameters are traced through Shapley value analysis.

Benefits of technology

It enables efficient and accurate detection and anomaly tracing of various mechanical components in the top-drive drilling unit, improving the accuracy and efficiency of detection, reducing computational redundancy, and enhancing the stability of the equipment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122065057A_ABST
    Figure CN122065057A_ABST
Patent Text Reader

Abstract

The invention discloses an interpretable anomaly detection method and device for a top drive drilling device, equipment and a medium, which are applied to the field of drilling engineering, and are used for acquiring operation data of a to-be-detected mechanical part in the top drive drilling device, and dividing the operation data into various working condition data sets based on working condition parameters, extracting feature data of each feature parameter from each working condition data set; determining a working condition sensitivity index and a conflict index based on a grey correlation degree coefficient and a correlation coefficient between the characteristic parameters, and optimizing memory matrix input parameters by utilizing sensitivity analysis and correlation analysis; a clustering algorithm is used for screening sample data to construct a spatial memory matrix, and a multivariate state estimation method and a sequential probability ratio test method are combined to carry out anomaly detection on the mechanical part to be detected; furthermore, interpretability analysis of the abnormal parameters is carried out, and the abnormal key parameters are visually revealed. The self-adaptive judgment of the running state of the equipment and the abnormal traceability of each mechanical part in the top drive well drilling device are realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of drilling engineering, and particularly to a method for detecting interpretable anomalies in a top-drive drilling rig, a device for detecting interpretable anomalies in a top-drive drilling rig, an electronic device, and a computer-readable storage medium. Background Technology

[0002] Top drive drilling rigs (or top drives) are critical equipment in oil drilling operations, and their operational stability directly affects the smooth progress of drilling operations. In the typical "trip-run-drill-connect single-joint" cycle, operating parameters such as load, speed, and torque need to be frequently switched according to process requirements, exhibiting high complexity and dynamic time-varying characteristics. This operational characteristic results in the top drive's series power transmission chain, consisting of the main motor, gearbox, spindle, back clamp, and rotary head, being in a long-term unsteady-state excitation environment. Even minor anomalies in any component can easily be rapidly transmitted and amplified along the structural path, leading to various typical faults such as bearing breakage, gearbox oil seal failure, coupling damage, and alignment cylinder failure. These faults not only cause top drive shutdowns and operational interruptions, severely impacting drilling efficiency, but can also lead to significant economic losses and even serious drilling safety accidents. Multivariate state estimation is a commonly used method for detecting equipment defects, and the memory matrix is ​​a core step in multivariate state estimation. Selecting appropriate state vectors to construct the memory matrix directly affects the accuracy and real-time performance of the multivariate state estimation model. Existing methods fail to adequately consider the complexity and correlation of operating conditions when selecting key monitoring variables and suffer from redundant memory matrix calculations, resulting in reduced accuracy and efficiency of anomaly detection results for various mechanical components in the top-drive drilling rig. Summary of the Invention

[0003] The purpose of this invention is to provide an interpretable anomaly detection method, device, electronic equipment, and computer-readable storage medium for top-drive drilling rigs, applicable to the drilling engineering field. This method selects feature parameters for constructing a memory matrix based on sensitivity and conflict indices, selects sample data for constructing the memory matrix based on a clustering algorithm, and constructs the memory matrix using the selected feature parameters and sample data, thereby achieving efficient and accurate detection of anomalies in various mechanical components of the top-drive drilling rig.

[0004] To address the aforementioned technical problems, this invention provides a method for interpretable anomaly detection in top-drive drilling rigs, comprising:

[0005] Obtain the operating dataset of the mechanical components to be tested in the top drive drilling unit, divide the dataset into various operating condition datasets based on the operating condition parameters, and extract the feature data of each feature parameter from each operating condition dataset.

[0006] Based on the feature data under each working condition, determine the gray correlation coefficient between each feature parameter, determine the working condition sensitivity index of each feature parameter based on the gray correlation coefficient, and determine the pre-screened feature parameters from the feature parameters based on the working condition sensitivity index.

[0007] Based on the characteristic data under each working condition, the correlation coefficient between each of the pre-screened characteristic parameters is determined, the conflict index of each of the pre-screened characteristic parameters is determined based on the correlation coefficient, and the target characteristic parameter is determined from the pre-screened characteristic parameters based on the conflict index.

[0008] Based on the feature data of the target feature parameters, a sample dataset for each working condition is constructed. A target sample dataset is obtained by filtering from the sample dataset based on a clustering algorithm. A memory matrix for each working condition is constructed based on the target sample dataset.

[0009] Based on the memory matrix, the residual vector of the observation vector to be detected is determined by the multivariate state estimation method, and the residual vector is tested by the sequential probability ratio test method to obtain the abnormal detection result of the mechanical component to be detected.

[0010] The residual vector determined by the multivariate state estimation method is subjected to interpretability analysis to obtain the Shapley value of each target feature parameter. Based on the Shapley value, the mean absolute Shapley value of the target feature parameter is determined. Based on the mean absolute Shapley value, the dominant anomaly parameter is determined from the target feature parameters to obtain the anomaly tracing result.

[0011] Optionally, based on the feature data under each working condition, a grey relational coefficient is determined between each feature parameter; based on the grey relational coefficient, a working condition sensitivity index is determined for each feature parameter; and based on the working condition sensitivity index, pre-screened feature parameters are determined from the feature parameters, including:

[0012] Based on the feature data under each working condition, the grey correlation coefficient between each feature parameter is determined, and the mean grey correlation coefficient of the feature parameter is determined based on the grey correlation coefficient between the feature parameter and other feature parameters.

[0013] The working condition sensitivity index of the feature parameter is determined based on the ratio of the mean of the grey relational coefficients of the feature parameter between different working conditions. The index deviation is determined based on the mean of the working condition sensitivity index. When the index deviation of the feature parameter is greater than or equal to the preset index deviation threshold, the feature parameter is determined as the pre-screened feature parameter.

[0014] Optionally, the correlation coefficient between each of the pre-screened feature parameters is determined based on the feature data under each working condition; the conflict index of each of the pre-screened feature parameters is determined based on the correlation coefficient; and the target feature parameter is determined from the pre-screened feature parameters based on the conflict index, including:

[0015] Based on the feature data of the pre-screened feature parameters under each working condition, the correlation coefficient between each pre-screened feature parameter is determined by correlation calculation, and the mean correlation coefficient of the pre-screened feature parameters is determined based on the correlation coefficient between the pre-screened feature parameters and other pre-screened feature parameters.

[0016] The result of subtracting the mean of the correlation coefficient from 1 is determined as the conflict index of the pre-screened feature parameters, and the target feature parameters are determined based on the conflict index of the pre-screened feature parameters under each working condition.

[0017] Optionally, the target sample dataset is obtained by filtering from the sample dataset based on a clustering algorithm, including:

[0018] Healthy sample data are determined from the sample data in the sample dataset based on the KNN algorithm;

[0019] Target sample data is determined from the health sample data based on the K-Means clustering algorithm;

[0020] Alternatively, the target sample data can be determined from the healthy sample data based on an equal-interval sampling algorithm;

[0021] Alternatively, the target sample data can be determined from the healthy sample data based on the Gaussian mixture clustering algorithm;

[0022] The target sample dataset is constructed based on the target sample data.

[0023] Optionally, the residual vector of the observation vector to be detected is determined based on the memory matrix using a multivariate state estimation method, including:

[0024] The target memory matrix is ​​determined from the memory matrix under each working condition based on the working condition data corresponding to the observation vector to be detected.

[0025] The weight coefficients of each data vector are obtained by inputting the observation vector to be detected and the data vector in the target memory matrix into the weight coefficient formula constructed based on the nearest neighbor weight method.

[0026] A weight vector is constructed based on the weight coefficients, an estimation vector is determined based on the product of the target memory matrix and the weight vector, and the residual vector is determined based on the estimation vector and the observation vector to be detected.

[0027] Optionally, an interpretability analysis is performed on the residual vector determined by the multivariate state estimation method to obtain the Shapley value of each of the target feature parameters, and the mean absolute Shapley value of the target feature parameters is determined based on the Shapley values, including:

[0028] The KernelSHAP algorithm is invoked to determine the Shapley value of each target feature parameter based on the residual vector determined by the multivariate state estimation method;

[0029] The mean absolute Shapley value is determined based on the Shapley value of the target feature parameter.

[0030] Optionally, the method further includes:

[0031] The error contribution rate of each of the target feature parameters is determined based on the residual vector;

[0032] The cumulative error contribution rate of each target characteristic parameter is determined based on the error contribution rate of each target characteristic parameter during the entire monitoring period;

[0033] Anomaly parameter localization analysis is performed based on the cumulative error contribution rate of the target feature parameters.

[0034] To address the aforementioned technical problems, this invention provides an interpretable anomaly detection device for top-drive drilling rigs, comprising:

[0035] The first module is used to acquire the operating data set of the mechanical components to be tested in the top drive drilling device, divide the operating data into various operating data sets based on the operating condition parameters, and extract the feature data of each feature parameter from each operating data set.

[0036] The second module is used to determine the grey relational coefficient between each feature parameter based on the feature data under each working condition, determine the working condition sensitivity index of each feature parameter based on the grey relational coefficient, and determine the pre-screened feature parameters from the feature parameters based on the working condition sensitivity index.

[0037] The third module is used to determine the correlation coefficient between each of the pre-screened feature parameters based on the feature data under each working condition, determine the conflict index of each of the pre-screened feature parameters based on the correlation coefficient, and determine the target feature parameter from the pre-screened feature parameters based on the conflict index.

[0038] The fourth module is used to construct a sample dataset for each working condition based on the feature data of the target feature parameters, to filter the target sample dataset from the sample dataset using a clustering algorithm, and to construct a memory matrix for each working condition based on the target sample dataset.

[0039] The fifth module is used to determine the residual vector of the observation vector to be detected by the multivariate state estimation method based on the memory matrix, and to test the residual vector by the sequential probability ratio test method to obtain the abnormal detection result of the mechanical component to be detected.

[0040] The sixth module is used to perform interpretability analysis on the residual vector determined by the multivariate state estimation method to obtain the Shapley value of each target feature parameter, determine the mean absolute Shapley value of the target feature parameter based on the Shapley value, and determine the dominant anomaly parameter from the target feature parameter based on the mean absolute Shapley value to obtain the anomaly tracing result.

[0041] To solve the above-mentioned technical problems, the present invention provides an electronic device, comprising:

[0042] Memory, used to store computer programs;

[0043] A processor is used to implement the above-described top-drive drilling device interpretable anomaly detection method when executing the computer program.

[0044] To address the aforementioned technical problems, the present invention provides a computer-readable storage medium storing computer-executable instructions. When these computer-executable instructions are executed by a processor, they implement the aforementioned interpretable anomaly detection method for top-drive drilling apparatus.

[0045] As can be seen, this invention acquires the operating data of the mechanical components to be tested in the top-drive drilling rig, divides the operating data into various operating condition datasets based on operating condition parameters, extracts feature data of each feature parameter from each operating condition dataset, determines the grey relational coefficient between each feature parameter based on the feature data under each operating condition, determines the operating condition sensitivity index of each feature parameter based on the grey relational coefficient, and determines pre-screened feature parameters from the feature parameters based on the operating condition sensitivity index; determines the correlation coefficient between each pre-screened feature parameter based on the feature data under each operating condition, determines the conflict index of each pre-screened feature parameter based on the correlation coefficient, and determines the target feature parameter from the pre-screened feature parameters based on the conflict index; and finally, based on the feature data of the target feature parameter and the operating condition parameters... The system constructs sample datasets for each working condition based on the operating condition data, and selects target sample datasets from the sample datasets using a clustering algorithm. Based on the target sample datasets, it constructs memory matrices for each working condition, determines the residual vectors of the observation vectors to be detected based on the memory matrices, and tests the residual vectors using the sequential probability ratio test to obtain the anomaly detection results for the mechanical parts to be detected. It then performs interpretability analysis on the residual vectors determined by the multivariate state estimation method to obtain the Shapley values ​​of each target feature parameter, determines the mean absolute Shapley value of the target feature parameter based on the Shapley value, and determines the dominant anomaly parameter from the target feature parameters based on the mean absolute Shapley value to obtain the anomaly tracing results.

[0046] This invention selects feature parameters for constructing a memory matrix based on sensitivity and conflict indices, selects sample data for constructing the memory matrix based on a clustering algorithm, and constructs the memory matrix using the selected feature parameters and sample data. This enables efficient and accurate detection of various mechanical components in a top-drive drilling rig, and provides interpretability analysis of the multivariate state estimation method to trace the source of abnormal parameters in various mechanical components of the top-drive drilling rig. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0048] Figure 1 A flowchart illustrating an interpretable anomaly detection method for a top-drive drilling apparatus provided in an embodiment of the present invention;

[0049] Figure 2 This is an example diagram of a dataset partitioning provided in an embodiment of the present invention;

[0050] Figure 3 An example diagram of the working condition sensitivity index of characteristic parameters under different working conditions provided in an embodiment of the present invention;

[0051] Figure 4 An example diagram illustrating the conflict index of characteristic parameters under different working conditions, provided in an embodiment of the present invention;

[0052] Figure 5 This is a comparison chart of the marginal contribution rate of the average absolute Shapley value of the target feature parameter in the Bearing1_1 dataset provided in an embodiment of the present invention;

[0053] Figure 6 This is a comparison chart of the marginal contribution rate of the average absolute Shapley value of the target feature parameter in the Bearing2_1 dataset provided in an embodiment of the present invention;

[0054] Figure 7 This is a comparison chart of the marginal contribution rate of the average absolute Shapley value of the target feature parameter in the Bearing2_3 dataset provided in an embodiment of the present invention;

[0055] Figure 8 This is a comparison chart of the marginal contribution rate of the mean absolute Shapley value of the target feature parameters in the Bearing3_1 dataset provided in an embodiment of the present invention;

[0056] Figure 9 This is a feature Shapley value distribution map of the Bearing1_1 dataset provided in an embodiment of the present invention;

[0057] Figure 10 This is a cumulative error contribution rate distribution diagram of the Bearing1_1 dataset provided in an embodiment of the present invention;

[0058] Figure 11 This is an interpretability comparison chart of Shapley value and cumulative error contribution rate under normal conditions, provided by an embodiment of the present invention.

[0059] Figure 12 This is an interpretability comparison chart of marginal contribution index and cumulative error contribution rate index under abnormal conditions provided by an embodiment of the present invention.

[0060] Figure 13 The diagram below illustrates the structure of an anomaly detection device for a top-drive drilling apparatus provided in this embodiment of the invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] The following combination Figure 1 , Figure 1 A flowchart of an interpretable anomaly detection method for a top-drive drilling rig provided in an embodiment of the present invention is provided. The method may include:

[0063] S101: Acquire the operating data of the mechanical components to be tested in the top drive drilling unit, divide the operating data into various operating condition datasets based on the operating condition parameters, and extract the feature data of each feature parameter from each operating condition dataset.

[0064] As a critical piece of equipment in oil drilling operations, the stability of the top drive drilling rig directly affects the smooth progress of drilling operations. Its operating parameters, such as load, speed, and torque, need to be frequently switched according to process requirements, exhibiting high complexity and dynamic time-varying characteristics. The top drive series power transmission chain, consisting of the main motor, gearbox, spindle, back clamp, and rotary head, operates under a non-steady-state excitation environment for extended periods. Even minor anomalies in any component can easily propagate and amplify rapidly along the structural path, leading to various typical faults such as bearing breakage, gearbox oil seal failure, coupling damage, and alignment cylinder failure. These faults not only cause top drive shutdowns and operational interruptions, severely impacting drilling efficiency, but can also lead to significant economic losses and even serious drilling safety accidents. Therefore, this embodiment can perform anomaly detection on various mechanical components in the top drive drilling rig.

[0065] This embodiment does not limit the specific method for anomaly detection of the mechanical components to be inspected. Multivariate state estimation is a commonly used method for detecting equipment defects. The memory matrix is ​​a core step in multivariate state estimation. Selecting appropriate state vectors to construct the memory matrix directly affects the accuracy and real-time performance of the multivariate state estimation model. Existing methods do not adequately consider the complexity and correlation of key monitoring variables and suffer from redundancy in memory matrix calculations, leading to reduced accuracy and efficiency in anomaly detection results for various mechanical components in top-drive drilling rigs.

[0066] For any mechanical component to be tested, this embodiment can screen the feature parameters for constructing the memory matrix based on sensitivity and conflict indices, and screen the sample data for constructing the memory matrix based on clustering algorithms. The memory matrix is ​​constructed using the screened feature parameters and sample data, thereby achieving efficient and accurate detection of each mechanical component in the top-drive drilling device.

[0067] In the industrial field, in order to fully characterize the abnormal states of typical equipment such as circulating pumps, wind turbine generators, and steam turbines, low-frequency signals such as internal temperature and current of the equipment, as well as high-frequency vibration signals obtained by external sensors, can be collected. Based on the historical health status data and measured data of the above monitoring data, data support can be provided for the subsequent construction of abnormal detection models of the equipment, thereby achieving accurate identification of abnormal signal samples.

[0068] This embodiment can acquire the operating data of the mechanical component to be tested in the top drive drilling device. The operating data set of the mechanical component to be tested is the monitoring data collected during the operation of the mechanical component to be tested. This embodiment does not limit the specific type of monitoring data. The monitoring data can be feature data related to the identification of abnormality of mechanical component, such as temperature, current, vibration and other signals. It is not limited here and can be set according to actual application.

[0069] This embodiment can perform anomaly detection on various mechanical components. The mechanical component requiring anomaly detection is the component to be detected. In this embodiment, the type of monitoring data for each mechanical component to be detected can be determined based on the characteristics of the mechanical component. For example, when the mechanical component to be detected is a bearing, the monitoring data used to construct the memory matrix of the bearing can be vibration signals.

[0070] To further explore the key parameters characterizing the equipment's multi-condition operating status, operating conditions are first divided by selecting appropriate parameters. These selected parameters possess the characteristic of directly reflecting the operating status of the top drive equipment. The equipment's preset operating condition space can be defined as G. p Among them, G p ∈(G1,G2,G3, ……, G w-1 G w ), where w is the total number of operating conditions.

[0071] This embodiment can divide the operating data of the mechanical component to be tested into various operating condition datasets based on operating condition parameters. For example, the operating condition of a bearing is generally distinguished by operating condition parameters such as speed s1 and load s2. In this embodiment, the operating data of the bearing can be divided into various operating condition datasets based on operating condition parameters such as speed and load. Each operating condition dataset contains the monitoring data collected by the bearing under the current operating condition.

[0072] This embodiment can extract feature data of each feature parameter from the dataset of various working conditions of the mechanical component to be tested. This embodiment can construct a memory matrix of the mechanical component to be tested under various working conditions based on the feature data of each feature parameter. This embodiment does not limit the specific type of feature parameter and can be set according to actual application.

[0073] Taking the top drive bearing as an example, a publicly available top drive bearing operation dataset is shown in Table 1.

[0074] Table 1: Summary Table of Rolling Bearing Dataset Details

[0075]

[0076] In this example, operating conditions can be divided based on rotational speed and load, and multi-domain features can be extracted from the one-dimensional raw signals collected under different operating conditions. The specific feature information is shown in Table 2.

[0077] Table 2: Summary Table of Rolling Bearing Dataset Details

[0078]

[0079] This example demonstrates how to extract feature data of types t1-t31 from monitoring data. The above process uses a bearing as an example; for any device under test, the method for acquiring its feature data can be set based on the device's characteristics.

[0080] This embodiment can determine the Fault Start Point (FSP) from the sampling points of the feature data of the component under test. The FSP can then be used to divide the feature data into normal data and fault data. This embodiment can determine the training set from the normal data and use the remaining feature data as the test set, which includes both normal and abnormal data that were not used in the training. This embodiment does not limit the specific method for dividing normal and fault data; generally, the 3-sigma method can be used to determine the FSP of the operational data. This embodiment can divide normal and abnormal data based on the FSP, determine the training set of the mechanical component under test from the normal data, and determine the test set of the mechanical component under test from the remaining normal and abnormal data.

[0081] For example, for the top drive bearing operation dataset, considering that the actual dataset Bearing1_1 has a small number of samples, the bearing dataset is divided into 4 groups using a sliding window. After expansion, the fault start point of the dataset is 300; the fault start point of the Bearing2_1 dataset is 453; the fault start point of the Bearing2_3 dataset is 313; and the fault start point of the Bearing3_1 dataset is 2385.

[0082] Four groups of bearings with fault initiation point labels were selected as the research objects. The training and test sets for each group of bearings were divided as follows: the first 200 test files of the Bearing1_1 dataset were selected as the training set of normal data; the first 300 test files of the Bearing2_1 dataset were selected as the training set of normal data; the first 200 test files of the Bearing2_3 dataset were selected as the training set of normal data; and the first 2000 test files of the Bearing3_1 dataset were selected as the training set of normal data. The test set consisted of the data corresponding to the four groups of bearings that were not used in the training. The specific fault initiation point locations and data partitioning methods are as follows: Figure 2 As shown.

[0083] S102: Determine the grey relational coefficient between each feature parameter based on the feature data under each working condition, determine the working condition sensitivity index of each feature parameter based on the grey relational coefficient, and determine the pre-screened feature parameters from the feature parameters based on the working condition sensitivity index.

[0084] This embodiment can optimize the characteristic parameters of the mechanical component under test from two aspects: the sensitivity of characteristic parameters to operating conditions and the correlation between characteristic parameters in multiple operating conditions. This embodiment can calculate the operating condition sensitivity index of characteristic parameters through correlation degree calculation, quantify the correlation between parameters and equipment operating conditions, and jointly evaluate the degree of correlation between characteristic parameters under multiple operating conditions using conflict index.

[0085] In this embodiment, the feature data used to calculate the operating condition sensitivity index and conflict index is the normal feature data in the training set.

[0086] This embodiment first determines the grey relational coefficient between each feature parameter based on the feature data under each working condition, then determines the working condition sensitivity index of each feature parameter based on the grey relational coefficient, and finally determines the pre-screened feature parameters from the feature parameters based on the working condition sensitivity index.

[0087] Specifically, this embodiment can determine the gray correlation coefficient between each feature parameter based on the feature data under each working condition, and determine the mean gray correlation coefficient of the feature parameter based on the gray correlation coefficient between the feature parameter and other feature parameters.

[0088] The working condition sensitivity index of the feature parameter is determined by the ratio of the mean of the grey relational coefficient of the feature parameter between different working conditions. The deviation of the index is determined by the mean of the working condition sensitivity index. When the deviation of the feature parameter is greater than or equal to the preset deviation threshold, the feature parameter is determined as a pre-screened feature parameter.

[0089] For example, we can first select G. p Feature data of a certain feature type in the working condition dataset is used as the original reference sequence. , where m is the number of sampling points for the feature data in the training set. Further, select the corresponding G... p Feature data of different feature types under different operating conditions are used as the original comparison sequence. .

[0090] The sequence data is made dimensionless, and the units of the feature data are eliminated by normalization methods, thus transforming it into a comparable time series.

[0091] This embodiment can obtain G. p Grey relational coefficient of the i-th feature parameter under the working condition The grey relational coefficient is calculated at each node of a time series to determine the degree of correlation between the comparison series and the reference series. The calculation method is shown in the following formula:

[0092] ;

[0093] In the formula, This is the original reference sequence. For G p The original comparison sequence of the i-th feature parameter under the working condition, where m is the number of sampling points of the feature data in the training set, and ρ is the resolution coefficient, which can be taken as 0.5 according to convention.

[0094] Furthermore, the correlation between the i-th feature parameter and other feature parameters can be calculated based on the grey relational coefficient. The grey relational coefficient only represents the correlation between data points. Its discrete distribution makes direct comparison and analysis difficult. Therefore, the average value method is used for centralized processing to reflect the correlation between each monitoring parameter and the reference sequence. The mean of the grey relational coefficients between each characteristic parameter and the remaining characteristic parameters is calculated and defined as the correlation degree. The calculation method is shown in the following formula:

[0095] ;

[0096] In the formula, For G p The correlation degree of the i-th feature parameter under the operating condition. For G p The grey relational coefficient of the β-th feature parameter under the operating condition. For G p The grey relational coefficient of the i-th feature parameter under the working condition, where m is the number of sampling points of the feature data in the training set, and n is the total number of feature parameters.

[0097] For different working conditions The characteristic data of the next characteristic parameter are used to perform the above calculations to obtain the correlation degree of each characteristic parameter under each working condition.

[0098] This embodiment can determine the condition sensitivity index of the feature parameters based on the ratio of the mean gray relational coefficients of the feature parameters between different operating conditions, as shown in the following formula:

[0099] ;

[0100] In the formula, S i Let i be the operating condition sensitivity index for the i-th characteristic parameter. To be with G p Different working conditions, For G p The correlation degree of the i-th feature parameter under the operating condition. for The correlation degree of the i-th feature parameter under the working condition.

[0101] When the characteristic parameters can effectively track changes in the operating conditions of rotating equipment, the correlation between the characteristic parameters under different operating conditions varies significantly. In this case, the parameter operating condition sensitivity index S... i Deviation from 1, conversely, when the parameter cannot well characterize the changes in the operating conditions of rotating equipment, the operating condition sensitivity index S... i Close to 1.

[0102] This embodiment determines the deviation of an indicator based on the mean of the operating condition sensitivity indicators. When the deviation of a feature parameter is greater than or equal to a preset deviation threshold, the feature parameter is determined as a pre-screening feature parameter for the mechanical component to be tested. This embodiment can use the absolute value of the difference between operating condition sensitivity indicators as the deviation of the indicator; this embodiment does not limit the value of the preset deviation threshold and can set it based on actual application.

[0103] Specifically, taking bearing operating conditions 1, 2, and 3 as examples, we can calculate the sensitivity indices of each feature parameter between the Bearing1_1 dataset and the Bearing3_1 dataset, and calculate the sensitivity indices of each feature parameter between the Bearing2_1 dataset and the Bearing3_1 dataset, such as... Figure 3 As shown, in this embodiment, t5, t6, t7, t10, t11, t18, t19, t20, t23, t24, and t25 can be selected from t1-t31 as pre-screening feature parameters.

[0104] S103: Determine the correlation coefficient between each pre-screened feature parameter based on the feature data under each working condition, determine the conflict index of each pre-screened feature parameter based on the correlation coefficient, and determine the target feature parameter from the pre-screened feature parameters based on the conflict index.

[0105] Parameters with high correlation to equipment status tend to have low conflict. To quantify the conflict between parameters under different operating conditions, correlation analysis is introduced for calculation. This embodiment can determine the correlation coefficient between each pre-screened feature parameter based on the feature data under each operating condition, determine the conflict index of each pre-screened feature parameter based on the correlation coefficient, and determine the target feature parameter of the mechanical component to be tested from the pre-screened feature parameters based on the conflict index.

[0106] Specifically, this embodiment library calculates the correlation coefficient between each pre-screened feature parameter based on the feature data of the pre-screened feature parameters under each working condition, and determines the mean correlation coefficient of the pre-screened feature parameters based on the correlation coefficient between the pre-screened feature parameters and other pre-screened feature parameters.

[0107] The result of subtracting the mean of the correlation coefficient from 1 is determined as the conflict index of the pre-screened feature parameters. The target feature parameters are then determined based on the conflict index of the pre-screened feature parameters under various working conditions.

[0108] For example, we can first assume that G is pre-selected based on the feature parameters. p The parameter sequence under the operating condition is as follows :

[0109] ;

[0110] In the formula, x i Let x be the feature data sequence under the i-th feature parameter, where i can take a maximum of p, and p is the number of pre-selected feature parameters. ij Let m be the feature data of the i-th feature parameter collected at the j-th sampling point, where m is the number of feature data sampling points in the training set.

[0111] Calculate G p The correlation coefficient between characteristic parameter i and characteristic parameter µ under the operating condition:

[0112] ;

[0113] In the formula, For G p The correlation coefficient between parameter i and parameter µ under operating conditions, x ij For the feature data of the i-th feature parameter collected at the j-th sampling point, x µj For the feature data of the µ-th feature parameter collected at the j-th sampling point, Let be the mean of the feature data under the i-th feature parameter. Let be the mean of the feature data under the µ-th feature parameter.

[0114] Furthermore, a conflict index can be calculated between the characteristic parameter i and the remaining characteristic parameters µ:

[0115] ;

[0116] In the formula, This is an index of the conflict between the i-th feature parameter and the other feature parameters. For G p The correlation coefficient between parameter i and parameter µ under the operating condition. For G p The correlation coefficient of parameter i under the working condition, and m is the number of sampling points of feature data in the training set.

[0117] In this embodiment, A smaller relative value indicates less conflict between parameters, a closer relationship between parameters, and a stronger correlation with equipment status; conversely, a larger value indicates a weaker correlation with equipment status. This embodiment does not limit the specific method of feature selection based on conflict indices; it can be set according to actual application. Generally, a conflict threshold can be set. When the conflict index of a feature parameter is greater than the conflict threshold under all operating conditions, the feature parameter can be removed.

[0118] Taking bearing operating conditions 1, 2, and 3 as examples, characteristic data under different operating conditions (Bearing1_1, Bearing2_1, and Bearing3_1) are selected to calculate the corresponding conflict indexes, such as... Figure 4 As shown, Figure 4 The energy ratio characteristic t25 of the second frequency band decomposition of the wavelet packet shows outstanding conflict index under all three operating conditions, indicating a weak correlation with other parameters. In this case, t25 can be removed, and t5, t6, t7, t10, t11, t18, t19, t20, t23, and t24 can be retained as target parameters.

[0119] S104: Construct sample datasets for each working condition based on feature data of target feature parameters and working condition data of working condition parameters. Filter the sample datasets from the sample datasets using a clustering algorithm to obtain the target sample dataset. Construct the memory matrix for each working condition based on the target sample dataset.

[0120] This embodiment can construct the final set of feature parameters of the mechanical component to be tested based on the target feature parameters and operating condition parameters. For example, the set of feature parameters of a bearing can be T=[t5, t6, t7, t10, t11, t18, t19, t20, t23, t24, s1, s2].

[0121] Further, based on the feature data of the target feature parameters and the working condition data of the working condition parameters in the training set, a sample dataset of the mechanical component to be tested under various working conditions is constructed.

[0122] In this embodiment, the working condition data is used to distinguish the working conditions to which the feature data belongs. The feature data of the target feature parameter is divided into working conditions by the working condition data of the working condition parameter, and a sample dataset containing the feature data of the target feature parameter under each working condition is constructed.

[0123] Considering the wide range of operating conditions and drastic fluctuations in operational monitoring parameters of top drive equipment, using a single process memory matrix to cover all operating condition data would result in poor modeling performance. This embodiment uses key monitoring parameters to classify operating conditions. It can then use a clustering algorithm to filter target sample datasets from the sample dataset, effectively eliminating abnormal sample data caused by interference from complex environmental noise in the equipment's field-acquired signals, thus reducing computational time complexity.

[0124] This embodiment first normalizes the training and test set data using a normalization method. Then, it uses the KNN (K-Nearest Neighbor) algorithm to determine healthy sample data from the sample data in the sample dataset. Next, it uses the K-Means clustering algorithm (a type of mean clustering algorithm) to determine target sample data from the healthy sample data. Alternatively, this embodiment can use an equal-interval sampling algorithm to determine target sample data from the healthy sample data, or a Gaussian mixture clustering algorithm.

[0125] Specifically, the operating condition space to which the sample dataset belongs is determined based on the operating condition parameters. For any operating condition G... p The sample dataset below can be L q×m =[X(t1)X(t2)…X(t m )],X(t j ) is the state vector composed of q feature data of feature types collected at the j-th sampling point, where j is at most m, m is the number of sampling points of feature data in the training set, and q is the number of target feature parameters.

[0126] Furthermore, calculate G p The mean vector is obtained by averaging the feature data of each target feature parameter under the operating conditions. =[V L,1 V L,2 V L,3 ,…,V L,q ] T , where V L,i For G p The average value of the feature data of the i-th target feature parameter under the working condition, where i is at most q, and T is the matrix transpose.

[0127] This embodiment can calculate L. q×m Each state vector and mean vector The distance d between jd j The distance between the state vector and the mean vector of the j-th sampling point is used to obtain the working condition G. p Distance vector =[d1,d2,d3,…,d m ] T .

[0128] This embodiment does not limit d. j The calculation method is generally based on Euclidean distance, as shown in the following formula:

[0129] ;

[0130] In the formula, d j Let X(t) be the Euclidean distance between the state vector and the mean vector of the j-th sampling point. j Let be the state vector at the j-th sampling point. Let x be the mean vector. i (t j V represents the feature data of the i-th target feature parameter at the j-th sampling point. L,i is the average value of the feature data of the i-th target feature parameter.

[0131] First, healthy sample data can be identified from the sample data in the sample dataset based on the KNN algorithm, and the distance vector can be used to... After sorting the elements in ascending order, the top 95% of the data with the smallest distance values ​​are selected as the healthy distances, resulting in a total of ρ1 distance elements. The state vectors corresponding to r distance elements, i.e., the healthy sample data, are then extracted from the sample dataset to construct the health matrix. , The matrix is ​​constructed from r healthy sample data, where each healthy sample data contains feature data of q target feature parameters.

[0132] Furthermore, this embodiment can determine the target sample data from the healthy sample data based on the K-Means clustering algorithm, and randomly select p healthy sample data from r as cluster centers c. i The cluster center set C = [c1, c2, c3, ..., c3] is constructed. p ].

[0133] Calculate the Euclidean distance from each healthy sample data point to the cluster center. Based on the Euclidean distance, assign the healthy sample data to the nearest cluster center. Update the cluster centers according to the new clustering results until the cluster centers no longer change, forming the final p cluster centers as the target sample data.

[0134] This embodiment can construct a memory matrix for each operating condition based on the target sample dataset, and equipment anomaly detection can be performed based on this memory matrix.

[0135] This embodiment can construct G based on p target sample data. p Memory matrix under operating conditions , For G p The memory matrix is ​​constructed from p target sample data under the working condition, and each target sample data contains feature data of q target feature parameters.

[0136] This embodiment can perform a clustering process to filter target sample data for each working condition's sample dataset, thereby constructing a memory matrix for each working condition. This embodiment does not limit the specific clustering method for each working condition; the number of cluster centers can be set based on the actual application. Generally, for the training sets in the Bearing1_1, Bearing2_1, and Bearing3_1 datasets for bearings, the number of cluster centers can be uniformly set to 100.

[0137] This embodiment can construct a spatial memory matrix based on the memory matrix under various working conditions. , , ,…, Anomaly detection is performed on the observation vector to be detected based on the spatial memory matrix. In this embodiment, the observation vector to be detected is the feature data of the target feature parameters that need to be detected, which are collected during the operation of the device.

[0138] S105: Based on the memory matrix, the residual vector of the observation vector to be detected is determined by the multivariate state estimation method, and the anomaly detection result of the mechanical part to be detected is obtained by testing the residual vector based on the sequential probability ratio test method.

[0139] This embodiment can combine the weighted nearest neighbor MSET (Multivariate State Estimation Technique) and SPRT (Sequential Probability Ratio Test) to detect anomalies in the observed vector. Specifically, the target memory matrix can be determined from the memory matrices under each working condition based on the working condition data corresponding to the observed vector. The observed vector to be detected and the data vectors in the target memory matrix are input into the weight coefficient formula constructed based on the nearest neighbor weight method to obtain the weight coefficients of each data vector. A weight vector is constructed based on the weight coefficients. The estimated vector is determined based on the product of the target memory matrix and the weight vector. The residual vector is determined based on the estimated vector and the observed vector to be detected. The anomaly detection result is obtained by testing the residual vector based on the sequential probability ratio test.

[0140] For example, if for a certain observation vector to be detected, its operating condition is determined to be G based on its operating condition data. p At this point, G in the spatial memory matrix can be called. p Target memory matrix under operating conditions Anomaly detection is performed on the observed vector to be detected, and the target memory matrix is ​​used. The structure can be:

[0141] ;

[0142] In the formula, For G p The target memory matrix under the working condition, X(t) j Let x be the j-th target sample data, where j can take a maximum value of p, and x i (t j ) represents the feature data of the i-th target feature parameter at the j-th sampling point, where i is at most q.

[0143] For a given moment, the real-time monitoring input of the observation vector to be detected, X obs Calculate the corresponding estimated vector:

[0144] ;

[0145] In the formula, X est Let W be the estimated vector of the observed vector to be detected, and W be the weight vector. k The weight coefficients are the weights of the k-th target sample data, where k can take a maximum of p. The weight coefficients represent X. est The similarity measure with each target sample data, T is the matrix transpose.

[0146] To further improve the model's estimation accuracy and robustness to outliers, the nearest neighbor weighting method is used instead of the traditional linear least squares method for weight calculation. The formula for calculating the weight coefficients is shown below:

[0147] ;

[0148] In the formula, w k Let X(t) be the weight coefficient of the k-th target sample data. j Let X(t) be the j-th target sample data. k Let X be the k-th target sample data. obs Let be the observation vector to be detected. The above calculation method can assign higher weights to parameters of samples closer to the memory matrix, thereby achieving a nonlinear weighted estimation of the observation vector to be detected.

[0149] The residual vector e = {e1, e2, ..., e} can be calculated by estimating the vector and the observed vector to be detected. q}, e qThe residual value is used to estimate the feature data of the q-th target feature parameter in the vector and the observation vector to be detected. The residual value globally follows a normal distribution.

[0150] Furthermore, anomaly detection results are obtained by testing the residual vector using the sequential probability ratio test. SPRT is a multi-stage detection method that can track the slow growth trend of faults and obtain detection results when the relevant indicator thresholds reach predetermined standards. Moreover, accurate judgments can still be made while ensuring that the error rates of both types of test parameters are set sufficiently low.

[0151] Suppose that event H0 represents the residual distribution under normal health conditions, and H0 ~ (µ0, σ0) 2 ), µ0 and σ0 2 Let H1 and H2 be the mean and variance of the residual distribution under normal healthy conditions, respectively; let H1 represent the residual distribution under abnormal conditions, H1 ~ (µ1, σ1) 2 ), µ1 and σ1 2 Let be the mean and variance of the residual distribution under abnormal conditions, respectively.

[0152] In practical applications, the mean of the deviation will increase or decrease with the occurrence of faults; that is, the mean of the residual in the abnormal state, µ1, will become µ0 + Δµ, where Δµ can be 3σ0. When using SPRT for anomaly detection, the data dispersion is set to be uniform, i.e., σ0 2 =σ1 2 The corresponding joint probability density distribution can be defined as follows:

[0153] ;

[0154] ;

[0155] In the formula, Let H0 be the joint probability density of p samples under the H0 distribution. Let be the joint probability density of p samples under the H1 distribution.

[0156] Likelihood ratio L e This is the ratio of the two joint probability densities, as shown in the following formula:

[0157] ;

[0158] In SPRT, it is also necessary to define values ​​for the false alarm rate α and the false negative rate β for two types of errors, both of which can be set to 0.05. Then, the anomaly detection thresholds A and B can be calculated using the following formula:

[0159] ;

[0160] ;

[0161] In this embodiment, the likelihood ratio can be calculated based on the residual vector between the to-be-detected observation vector and the estimated vector. If the likelihood ratio L e < - A, then the hypothesis H0 is accepted and the observation is stopped, and it is determined that the detection result is normal; if the likelihood ratio L e > - B, then the hypothesis H1 is accepted and the observation is stopped, and it is determined that the detection result is abnormal; if the likelihood ratio B ≤L e ≤A, then the detection of the observation vector continues.

[0162] In this embodiment, for any to-be-detected mechanical component, the memory matrix of the to-be-detected mechanical component under each working condition can be determined in the above manner, and the SPRT method is combined with the memory matrix to perform abnormal monitoring of the mechanical component.

[0163] In this embodiment, the nearest neighbor weight method is used to replace the traditional linear least squares method for the weight coefficient in MSET, improving the estimation accuracy of the MSET model and the robustness to outliers. The weighted nearest neighbor MSET model is trained using the data in the healthy period, which can effectively express the parameter relationship under healthy conditions. Combining the SPRT method to construct a multi-parameter abnormal detection model can adaptively detect the abnormal changes of multi-working condition data. Compared with the manually and empirically set residual thresholds, it can intuitively and effectively reduce the false alarm rate and missed alarm rate in abnormal detection.

[0164] S10⑥: Perform interpretability analysis on the residual vector determined by the multi-variable state estimation method to obtain the Shapley value of each target feature parameter, determine the mean absolute Shapley value of the target feature parameter based on the Shapley value, and determine the dominant abnormal parameter from the target feature parameters based on the mean absolute Shapley value to obtain the abnormal source tracing result.

[0165] This example further introduces the KernelSHAP method. By performing interpretability analysis on the abnormal detection results of the multi-variable state estimation model, it realizes quantifying the contribution degree of the dominant abnormal parameter, providing targeted maintenance decisions for on-site staff. KernelSHAP is a model interpretation method based on the Shapley value in game theory, used to quantify the contribution of each feature to a specific prediction result.

[0166] This embodiment uses KernelSHAP to calculate the mean absolute Shapley value of different target feature parameters for the residual value during the monitoring period of the device, further quickly locating the main parameters causing the abnormality, and then performing abnormal location according to the fault reasons represented by the corresponding parameter changes, showing strong interpretability and providing a reliable theoretical basis for subsequent fault diagnosis.

[0167] This embodiment can calculate the average absolute Shapley value of the target characteristic parameters. Based on the average absolute Shapley value, the interpretability of abnormal parameters can be analyzed from multiple dimensions such as operating status, different working conditions, and fault type, providing a reliable theoretical basis for subsequent fault diagnosis of the mechanical components under test in the top drive.

[0168] This embodiment does not limit the specific method for calculating the average absolute Shapley value of the target feature parameters. Generally, the residual vector can be determined based on the observation vector to be detected and the estimated vector; the KernelSHAP algorithm is called to determine the Shapley value of each target feature parameter based on the residual vector determined by the multivariate state estimation method; the average absolute Shapley value is determined based on the Shapley value of the target feature parameters; and anomaly parameter localization analysis is performed based on the average absolute Shapley value of the target feature parameters.

[0169] Specifically, the prediction error of the device at each sampling point time for each target feature parameter is defined as:

[0170] ;

[0171] In the formula, E(i,j) is the prediction error of the i-th target feature parameter at the j-th sampling point, and X... obs (i,j) represents the observed value of the i-th target feature parameter at the j-th sampling point, X est (i,j) represents the estimated value of the i-th target feature parameter at the j-th sampling point.

[0172] To measure the relative impact of each monitoring parameter on the overall error at the same time point, the KernelSHAP method is used to perform interpretability analysis on the error of multivariate state estimation. First, the total error score function of the system at time j is defined. The sum of prediction errors for all features at that moment:

[0173] ;

[0174] Based on KernelSHAP, a linear interpretive model is constructed to approximate the predicted output of the original model. Its mathematical expression is as follows:

[0175] ;

[0176] In the formula, The average error of all samples, That is, the Shapley value of the i-th target feature parameter at the j-th sampling point, where p is the total number of target feature parameters.

[0177] It reflects the abnormal parameter situation at a single moment, and can realize the calculation of Shapley values ​​between different target feature parameters. It can also construct the error contribution interpretability matrix Er by combining the Shapley values ​​of all time periods.

[0178] ;

[0179] In the formula, E r Let E be the Shapley value matrix. r (i,j) represents the Shapley value of the i-th target feature parameter at the j-th sampling point, where i is at most q and j is at most p.

[0180] The Er matrix characterizes the trajectory of each monitoring parameter's anomalous contribution throughout the entire monitoring period, providing data support for the subsequent interpretability and localization of anomalous parameters. Building upon this, the mean absolute Shapley value is further introduced to quantify the overall anomalousness of the parameter throughout the entire monitoring period.

[0181] ;

[0182] In the formula, E cr (i) is the average absolute Shapley value of the i-th target feature parameter.

[0183] In this implementation, anomaly parameter localization analysis was performed on four sets of faulty bearing datasets. A comparison chart of the mean absolute Shapley values ​​of errors under different conditions was plotted, with specific trends shown below. Figure 5 As shown. From Figure 5 The results show a significant difference in the mean absolute Shapley value between normal and abnormal states. The error value for the normal state, serving as a benchmark, is relatively small, with the main contributing features being the centroid frequency (t18), mean square frequency (t19), root mean square frequency (t20), and the energy ratio of the first frequency band of the wavelet packet (t24). The vibration signal of a normally operating bearing exhibits low and stable energy, with spectral energy concentrated in the low-frequency band and no significant fault characteristic frequencies. Specifically, t18, t19, and t20 characterize the concentration of energy distribution in the frequency domain, reflecting that the spectral energy is mainly concentrated in the low-frequency band; t24 directly quantifies the proportion of energy in the low-frequency band.

[0184] When the operating state shifts from normal to abnormal, the signal energy distribution undergoes significant reconstruction. Energy migrates and diffuses from the low-frequency main band to the fault characteristic frequency and high-frequency resonant band, resulting in a decrease in spectral energy concentration. This redistribution process causes a general attenuation of the average absolute Shapley values ​​of the t18, t19, t20, and t24 features, which play a dominant role under normal conditions. However, the outer ring fault mode exhibits a unique response. Despite the overall decrease in frequency domain concentration, the error contribution of the t24 feature increases in the abnormal state. This phenomenon stems from the structural characteristics of the outer ring being fixed to the base. The fault impact energy forms a direct force transmission path through the rigid connection. Rigid conduction has less attenuation of low-frequency components, allowing the fundamental frequency component and adjacent harmonics of the impact energy to be preserved more completely, resulting in a relative concentration of impact energy in the low-frequency band. This mechanism makes t24 an important indicator of outer ring faults, namely, the low-frequency impact energy contribution highlighted by the structural conduction advantage against the background of overall spectral energy dispersion.

[0185] Comparative analysis of outer ring fault data for Bearing1_1 under high load and low speed conditions and Bearing3_1 under low load and high speed conditions, such as... Figure 5 and Figure 8 As shown, the differences in operating conditions reveal the significant impact of operating conditions on the mean absolute Shapley value of multi-domain characteristics of outer ring faults. Under abnormal conditions, high load and low speed significantly increase the mean absolute Shapley value of both time-domain statistical characteristics and frequency-domain average amplitude characteristics. High load increases the stress and friction on the bearing contact surface, which directly manifests as an overall increase in signal amplitude in the time domain; at the same time, lower speed means a longer energy accumulation time between each impact or vibration event, making the energy of a single event stronger, thus making it more prominent in time-domain statistical characteristics and frequency-domain average amplitude reflecting the overall energy level.

[0186] Conversely, under low load and high speed conditions, the high-frequency excitation of the equipment is more likely to couple and resonate with the fault characteristic frequency, causing energy to be highly concentrated near the fault characteristic frequency and its harmonics, significantly enhancing the contribution of the relevant frequency domain energy characteristics t18, t19, t20, and t24. Therefore, the synergistic effect of load and speed deeply affects the intensity of fault excitation, energy distribution pattern, and its representation in multi-domain features. High load strengthens the time domain and overall amplitude response, while high speed is more conducive to stimulating and highlighting the energy concentration characteristics in the frequency domain. In bearing fault diagnosis applications, it is necessary to fully consider the operating conditions to optimize feature extraction and diagnostic strategies. That is, under low load and high speed conditions, the analysis of frequency domain energy distribution characteristics should be emphasized, while under high load and low speed conditions, more attention should be paid to the monitoring of time domain statistical characteristics and overall energy amplitude.

[0187] Figure 6 Bearing2_1 inner ring fault and Figure 7The Bearing2_3 cage failure data reveals the differences in characteristic mean absolute Shapley values ​​for different failure modes under the same operating conditions. Under normal conditions, the low-frequency impact, frequency shift, and modulation components of vibration energy distribution in inner ring failures result in a higher root-mean-square frequency characteristic contribution and a lower center-of-gravity frequency; cage failures, on the other hand, exhibit the opposite characteristic trend due to the high concentration of energy at low-frequency natural frequencies and their harmonics. Under abnormal conditions, the vibration propagation coupling effect caused by the physical proximity of the two leads to a convergence in the trend of characteristic mean absolute Shapley values, weakening the distinguishability of failure modes.

[0188] In summary, the trend differences in the average absolute Shapley value from multiple perspectives, such as operating status, different working conditions, and fault types, can intuitively identify abnormal parameters and provide a reliable decision-making basis for fault location of rotating equipment.

[0189] Furthermore, in addition to the mean absolute Shapley value of the target feature parameters, this embodiment can perform anomaly parameter localization analysis based on the cumulative error contribution rate index of the target feature parameters. Specifically, the error vector is determined based on the observed vector to be detected and the estimated vector, and the error contribution of each target feature parameter is determined based on the error vector; the cumulative error contribution rate index is determined based on the error contribution index of the target feature parameters, and anomaly parameter localization analysis is performed based on the cumulative error contribution rate index. The cumulative error contribution rate index is a method for measuring the magnitude of each participant's contribution to the overall error, which can achieve rapid localization.

[0190] To further explore the interpretability of multivariate parameters, this embodiment can introduce the cumulative error contribution rate index, which is a calculation method for quickly locating the dominant parameters of anomalies.

[0191] This embodiment selects the Bearing1_1 dataset for analysis of the average absolute Shapley value and cumulative error contribution rate. First, the Shapley values ​​of the target feature parameters are calculated, and the corresponding heatmaps are plotted, such as... Figure 9 As shown, the contribution of target feature parameters to the error of all running samples is analyzed from a global perspective. A two-dimensional matrix of the cumulative error contribution rate index is constructed and the corresponding heatmap is plotted, as shown. Figure 10 As shown; at the same time, from Figure 9 and Figure 10 The comparison shows that the distribution trends of the indicators presented by the two heatmaps are highly consistent.

[0192] Figure 11 This is a chart comparing the interpretability of the marginal contribution rate of the average Shapley value and the cumulative error contribution rate under normal conditions, as shown in the figure. Figure 11As shown, under normal conditions, the key features identified by the Marginal Contribution Rate and Cumulative Contribution Rate are t19, t18, t20, t24, and t23. Figure 12 As shown, under anomalous conditions, the feature importance ranking of the cumulative contribution rate indicator is completely consistent with the marginal contribution rate. This result provides dual validation for the accurate localization of anomalous parameters and also confirms the feasibility of the proposed method in the interpretability analysis of anomalous parameters.

[0193] Based on the above embodiments, the present invention filters the feature parameters for constructing the memory matrix based on sensitivity and conflict indices, filters the sample data for constructing the memory matrix based on clustering algorithms, and constructs the memory matrix using the filtered feature parameters and sample data. This enables efficient and accurate detection of anomalies in the top-drive drilling rig, and provides interpretability analysis for the multivariate state estimation method, enabling the tracing of abnormal parameters of each mechanical component in the top-drive drilling rig.

[0194] The following combination Figure 13 , Figure 13 A structural block diagram of an interpretable anomaly detection device for a top-drive drilling rig provided in an embodiment of the present invention is shown. The device may include:

[0195] The first module 100 is used to acquire the operating data of the component to be tested in the top drive drilling device, divide the operating data into various operating condition datasets based on the operating condition parameters, and extract the feature data of each feature parameter from each operating condition dataset.

[0196] The second module 200 is used to determine the gray relational coefficient between each feature parameter based on the feature data under each working condition, determine the working condition sensitivity index of each feature parameter based on the gray relational coefficient, and determine the pre-screened feature parameters from the feature parameters based on the working condition sensitivity index.

[0197] The third module 300 is used to determine the correlation coefficient between each pre-screened feature parameter based on the feature data under each working condition, determine the conflict index of each pre-screened feature parameter based on the correlation coefficient, and determine the target feature parameter from the pre-screened feature parameters based on the conflict index.

[0198] The fourth module 400 is used to construct sample datasets for each working condition based on the feature data of the target feature parameters, to filter the target sample dataset from the sample dataset based on the clustering algorithm, and to construct the memory matrix for each working condition based on the target sample dataset.

[0199] The fifth module 500 is used to determine the residual vector of the observation vector to be detected by multivariate state estimation based on the memory matrix, and to test the residual vector based on the sequential probability ratio test to obtain the abnormal detection result of the mechanical component to be monitored.

[0200] The sixth module 600 is used to perform interpretability analysis on the residual vector determined by the multivariate state estimation method to obtain the Shapley value of each target feature parameter, determine the mean absolute Shapley value of the target feature parameter based on the Shapley value, determine the dominant anomaly parameter from the target feature parameter based on the mean absolute Shapley value, and obtain the anomaly tracing result.

[0201] Based on the above embodiments, the present invention filters the feature parameters for constructing the memory matrix based on sensitivity and conflict indices, filters the sample data for constructing the memory matrix based on clustering algorithms, and constructs the memory matrix using the filtered feature parameters and sample data. This enables efficient and accurate detection of anomalies in the top-drive drilling rig, and provides interpretability analysis for the multivariate state estimation method, enabling the tracing of abnormal parameters of each mechanical component in the top-drive drilling rig.

[0202] Based on the above embodiments, the second module 200 may include:

[0203] The first unit is used to determine the grey relational coefficient between each feature parameter based on the feature data under each working condition, and to determine the mean grey relational coefficient of the feature parameter based on the grey relational coefficient between the feature parameter and other feature parameters.

[0204] The second unit is used to determine the working condition sensitivity index of the feature parameter based on the ratio of the mean of the gray relational coefficient of the feature parameter between different working conditions, and to determine the index deviation based on the mean of the working condition sensitivity index. When the index deviation of the feature parameter is greater than or equal to the preset index deviation threshold, the feature parameter is determined as a pre-screened feature parameter.

[0205] Based on the above embodiments, the third module 300 may include:

[0206] The third unit is used to perform correlation calculations based on the characteristic data of pre-screened characteristic parameters under various working conditions to determine the correlation coefficient between each pre-screened characteristic parameter, and to determine the mean correlation coefficient of the pre-screened characteristic parameters based on the correlation coefficient between the pre-screened characteristic parameters and other pre-screened characteristic parameters.

[0207] The fourth unit is used to determine the conflict index of the pre-screened feature parameters by subtracting the mean of the correlation coefficient from 1, and to determine the target feature parameters based on the conflict index of the pre-screened feature parameters under various working conditions.

[0208] Based on the above embodiments, the fourth module 400 may include:

[0209] The fifth unit is used to determine healthy sample data from the sample data in the sample dataset based on the KNN algorithm;

[0210] The sixth unit is used to determine target sample data from healthy sample data based on the K-Means clustering algorithm;

[0211] The seventh unit is used to determine target sample data from healthy sample data based on an equal-interval sampling algorithm;

[0212] Unit 8 is used to determine target sample data from healthy sample data based on Gaussian mixture clustering algorithm;

[0213] Unit 9 is used to construct the target sample dataset based on the target sample data.

[0214] Based on the above embodiments, the fifth module 500 may include:

[0215] The tenth unit is used to determine the target memory matrix from the memory matrix under each working condition based on the working condition data corresponding to the observation vector to be detected;

[0216] The eleventh unit is used to input the observation vector to be detected and the data vector in the target memory matrix into the weight coefficient formula constructed based on the nearest neighbor weight method to obtain the weight coefficient of each data vector;

[0217] Unit 12 is used to construct a weight vector based on the weight coefficients, determine the estimation vector based on the product of the target memory matrix and the weight vector, and determine the residual vector based on the estimation vector and the observation vector to be detected.

[0218] Based on the above embodiments, the sixth module 600 may include:

[0219] Unit 13 is used to call the KernelSHAP algorithm to determine the Shapley values ​​of each target feature parameter based on the residual vector determined by the multivariate state estimation method.

[0220] Unit 14 is used to determine the average absolute Shapley value based on the Shapley values ​​of the target feature parameters.

[0221] Based on the above embodiments, the device may further include:

[0222] The seventh module is used to determine the error contribution rate of each target feature parameter based on the residual vector;

[0223] The eighth module is used to determine the cumulative error contribution rate of each target characteristic parameter based on the error contribution rate of each target characteristic parameter throughout the entire monitoring period;

[0224] The ninth module is used for anomaly parameter localization analysis based on the cumulative error contribution rate of target feature parameters.

[0225] Based on the above embodiments, the present invention also provides an electronic device, which may include a memory and a processor. The memory stores a computer program, and when the processor calls the computer program in the memory, it can implement the steps provided in the above embodiments. Of course, the device may also include various necessary network interfaces, a power supply, and other components.

[0226] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by an execution terminal or processor, can implement the method provided in the embodiments of the present invention; the storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0227] In this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, without necessarily requiring or implying any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A method for interpretable anomaly detection in a top-drive drilling rig, characterized in that, include: The operating data of the mechanical components to be tested in the top drive drilling unit is acquired, and the operating data is divided into various operating condition datasets based on the operating condition parameters. Feature data of each feature parameter is extracted from each of the operating condition datasets. Based on the feature data under each working condition, determine the gray correlation coefficient between each feature parameter, determine the working condition sensitivity index of each feature parameter based on the gray correlation coefficient, and determine the pre-screened feature parameters from the feature parameters based on the working condition sensitivity index. Based on the characteristic data under each working condition, the correlation coefficient between each of the pre-screened characteristic parameters is determined, the conflict index of each of the pre-screened characteristic parameters is determined based on the correlation coefficient, and the target characteristic parameter is determined from the pre-screened characteristic parameters based on the conflict index. Based on the feature data of the target feature parameters, a sample dataset for each working condition is constructed. A target sample dataset is obtained by filtering from the sample dataset based on a clustering algorithm. A memory matrix for each working condition is constructed based on the target sample dataset. Based on the memory matrix, the residual vector of the observation vector to be detected is determined by the multivariate state estimation method, and the residual vector is tested by the sequential probability ratio test method to obtain the abnormal detection result of the mechanical component to be detected. The residual vector determined by the multivariate state estimation method is subjected to interpretability analysis to obtain the Shapley value of each target feature parameter. Based on the Shapley value, the mean absolute Shapley value of the target feature parameter is determined. Based on the mean absolute Shapley value, the dominant anomaly parameter is determined from the target feature parameters to obtain the anomaly tracing result.

2. The interpretable anomaly detection method for the top-drive drilling device according to claim 1, characterized in that, Based on the feature data under each operating condition, a grey relational coefficient is determined among the feature parameters; based on the grey relational coefficient, an operating condition sensitivity index is determined for each feature parameter; and based on the operating condition sensitivity index, pre-screened feature parameters are determined from the feature parameters, including: Based on the feature data under each working condition, the grey correlation coefficient between each feature parameter is determined, and the mean grey correlation coefficient of the feature parameter is determined based on the grey correlation coefficient between the feature parameter and other feature parameters. The working condition sensitivity index of the feature parameter is determined based on the ratio of the mean of the grey relational coefficients of the feature parameter between different working conditions. The index deviation is determined based on the mean of the working condition sensitivity index. When the index deviation of the feature parameter is greater than or equal to the preset index deviation threshold, the feature parameter is determined as the pre-screened feature parameter.

3. The interpretable anomaly detection method for the top-drive drilling rig according to claim 1, characterized in that, Based on the characteristic data under each working condition, the correlation coefficient between each of the pre-screened characteristic parameters is determined; based on the correlation coefficient, the conflict index of each of the pre-screened characteristic parameters is determined; and based on the conflict index, the target characteristic parameter is determined from the pre-screened characteristic parameters, including: Based on the feature data of the pre-screened feature parameters under each working condition, the correlation coefficient between each pre-screened feature parameter is determined by correlation calculation, and the mean correlation coefficient of the pre-screened feature parameters is determined based on the correlation coefficient between the pre-screened feature parameters and other pre-screened feature parameters. The result of subtracting the mean of the correlation coefficient from 1 is determined as the conflict index of the pre-screened feature parameters, and the target feature parameters are determined based on the conflict index of the pre-screened feature parameters under each working condition.

4. The interpretable anomaly detection method for the top-drive drilling rig according to claim 1, characterized in that, The target sample dataset is obtained by filtering from the sample dataset based on the clustering algorithm, including: Healthy sample data are determined from the sample data in the sample dataset based on the KNN algorithm; Target sample data is determined from the health sample data based on the K-Means clustering algorithm; Alternatively, the target sample data can be determined from the healthy sample data based on an equal-interval sampling algorithm; Alternatively, the target sample data can be determined from the healthy sample data based on the Gaussian mixture clustering algorithm; The target sample dataset is constructed based on the target sample data.

5. The interpretable anomaly detection method for the top-drive drilling rig according to claim 1, characterized in that, Based on the memory matrix, the residual vector of the observation vector to be detected is determined using a multivariate state estimation method, including: The target memory matrix is ​​determined from the memory matrix under each working condition based on the working condition data corresponding to the observation vector to be detected. The weight coefficients of each data vector are obtained by inputting the observation vector to be detected and the data vector in the target memory matrix into the weight coefficient formula constructed based on the nearest neighbor weight method. A weight vector is constructed based on the weight coefficients, an estimation vector is determined based on the product of the target memory matrix and the weight vector, and the residual vector is determined based on the estimation vector and the observation vector to be detected.

6. The interpretable anomaly detection method for the top-drive drilling rig according to claim 1, characterized in that, The interpretability analysis of the residual vector determined by the multivariate state estimation method is performed to obtain the Shapley value of each target feature parameter. Based on the Shapley value, the mean absolute Shapley value of the target feature parameter is determined, including: The KernelSHAP algorithm is invoked to determine the Shapley value of each target feature parameter based on the residual vector determined by the multivariate state estimation method; The mean absolute Shapley value is determined based on the Shapley value of the target feature parameter.

7. The interpretable anomaly detection method for the top-drive drilling rig according to claim 1, characterized in that, Also includes: The error contribution rate of each of the target feature parameters is determined based on the residual vector; The cumulative error contribution rate of each target characteristic parameter is determined based on the error contribution rate of each target characteristic parameter during the entire monitoring period; Anomaly parameter localization analysis is performed based on the cumulative error contribution rate of the target feature parameters.

8. A top-drive drilling rig with an interpretable anomaly detection device, characterized in that, include: The first module is used to acquire the operating data of the mechanical components to be tested in the top drive drilling device, divide the operating data into various operating condition datasets based on the operating condition parameters, and extract the feature data of each feature parameter from each of the operating condition datasets. The second module is used to determine the grey relational coefficient between each feature parameter based on the feature data under each working condition, determine the working condition sensitivity index of each feature parameter based on the grey relational coefficient, and determine the pre-screened feature parameters from the feature parameters based on the working condition sensitivity index. The third module is used to determine the correlation coefficient between each of the pre-screened feature parameters based on the feature data under each working condition, determine the conflict index of each of the pre-screened feature parameters based on the correlation coefficient, and determine the target feature parameter from the pre-screened feature parameters based on the conflict index. The fourth module is used to construct a sample dataset for each working condition based on the feature data of the target feature parameters, to filter the target sample dataset from the sample dataset using a clustering algorithm, and to construct a memory matrix for each working condition based on the target sample dataset. The fifth module is used to determine the residual vector of the observation vector to be detected by the multivariate state estimation method based on the memory matrix, and to test the residual vector by the sequential probability ratio test method to obtain the abnormal detection result of the mechanical component to be detected. The sixth module is used to perform interpretability analysis on the residual vector determined by the multivariate state estimation method to obtain the Shapley value of each target feature parameter, determine the mean absolute Shapley value of the target feature parameter based on the Shapley value, and determine the dominant anomaly parameter from the target feature parameter based on the mean absolute Shapley value to obtain the anomaly tracing result.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the interpretable anomaly detection method for a top-drive drilling apparatus as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, implement the interpretable anomaly detection method for a top-drive drilling apparatus as described in any one of claims 1 to 7.