Intelligent Status Monitoring System for Generator Sets Based on Measurement While Drilling

By combining multidimensional matrix profile analysis and robust random forest algorithm, the problem of multi-parameter anomaly identification in the condition monitoring of downhole generator sets was solved, and the accurate location and differentiation of the anomaly start and end points, duration and fault type were achieved, thus improving the accuracy and reliability of monitoring.

CN123109633APending Publication Date: 2026-11-17HUBEI HONGHUA LONG TECH MASCH & ELECTRICITY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611305480.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-26
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing downhole generator unit status monitoring methods are unable to identify abnormal states caused by the joint changes of multiple parameters, cannot accurately locate the start and end points, duration, and degree of contribution of the abnormality, and are difficult to distinguish between instantaneous fluctuations and continuous faults. The fixed historical sample judgment model is difficult to update.

Method used

Multidimensional matrix profile analysis and robust random forest algorithm are used to perform correlation analysis on the voltage, current, temperature, vibration and speed data of generator set. Multidimensional matrix profile analysis is used to identify the start and end positions and duration of abnormal subsequences, and robust random forest algorithm is used to generate comprehensive anomaly scores and key isolation dimensions to realize fault category identification and status monitoring.

Benefits of technology

It improves the accuracy of anomaly location and parameter identification of downhole generator sets, can distinguish between instantaneous fluctuations and continuous faults, and generates monitoring results that include fault status and fault type, thus reducing the risk of misjudgment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123109633A_ABST
    Figure CN123109633A_ABST
Patent Text Reader

Abstract

This invention discloses an intelligent condition monitoring system for generator sets based on measurement-while-drilling (MWD) downhole, comprising the following steps: collecting voltage, current, temperature, vibration, and speed data of the generator set; performing time alignment and segmentation processing on the collected data; identifying abnormal data segments and the degree of influence of each monitoring parameter on the abnormality; establishing a reference sample based on normal operating data; filtering and comprehensively judging abnormal data; and updating the reference sample using data that meets normal conditions. When the duration of the abnormality reaches a set time and the degree of abnormality exceeds a limit, it is determined that the generator set has experienced a continuous fault; the fault category is determined based on the main abnormal parameters; and a condition monitoring result including the fault status, fault category, and abnormal location is generated. This invention relates to the field of downhole equipment condition monitoring technology.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of downhole equipment condition monitoring technology, and in particular to an intelligent condition monitoring system for generator sets based on measurement while drilling. Background Technology

[0002] As measurement-while-drilling (MWD) equipment evolves towards long-term, continuous downhole operation, downhole generator sets need to continuously supply power to the equipment under conditions of vibration, temperature changes, and load fluctuations. Existing generator set condition monitoring methods typically collect voltage, current, temperature, vibration, and speed data, and use single-parameter threshold comparisons, trend analysis, or classification models based on historical samples to determine the operating status. Some monitoring methods also extract statistical features through sliding windows and provide anomaly alerts for data exceeding preset limits to assist in identifying electrical anomalies, abnormal temperature rises, vibration anomalies, and abnormal speed anomalies.

[0003] Multiple monitoring parameters of underground generator sets exhibit correlated changes and response time differences. Existing technologies struggle to identify abnormal states resulting from the combined changes of multiple parameters when judging each parameter individually. Directly comparing different parameters at fixed time points can easily amplify data discrepancies due to asynchronous parameter responses. Current anomaly detection methods typically output anomaly scores at a specific sampling time or within a data window, making it difficult to simultaneously determine the start and end points of the anomaly, its duration, and the degree of contribution of each monitoring parameter. Furthermore, it is challenging to distinguish between instantaneous fluctuations and persistent faults. Judgment models built upon fixed historical samples struggle to update reference samples based on subsequent normal operating data, potentially leading to inconsistencies between judgment boundaries and the actual operating state of the generator set. Furthermore, determining the fault type solely based on a single anomaly score makes it difficult to combine key anomaly parameters and the model isolation process to identify single or compound fault categories.

[0004] Therefore, how to provide an intelligent condition monitoring system for generator sets based on measurement while drilling is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent condition monitoring system for generator sets based on measurement while drilling. This invention fully utilizes multidimensional matrix profile analysis and robust random forest algorithm to perform correlation analysis, anomaly screening, continuous fault determination, and fault category identification on the voltage, current, temperature, vibration, and speed data of the generator set. It can identify multi-parameter anomalies with time lag relationships, distinguish between instantaneous fluctuations and continuous faults, and generate generator set condition monitoring results that include fault status and fault category.

[0006] The intelligent condition monitoring system for generator sets based on downhole measurement-while-drilling according to an embodiment of the present invention includes the following steps: The data acquisition and processing module collects voltage, current, temperature, vibration and speed data of the generator set, performs alignment and windowing processing, and generates multi-dimensional state subsequences. The matrix profile module performs multidimensional matrix profile analysis on multidimensional state subsequences to obtain matrix profile distance, start and end positions of anomalous subsequences, duration length, and anomalous contribution vector. The feature reconstruction module constructs a state feature vector from the matrix contour distance, the start and end positions and duration of the abnormal subsequence, and the abnormal contribution vector, and establishes a reference sample set based on the state feature vector pre-calibrated as the normal operating state. The anomaly screening module determines the screening limit based on the matrix contour distance in the reference sample set, and identifies state feature vectors whose matrix contour distance exceeds the screening limit as candidate anomaly feature vectors. The forest module inputs the state feature vector into the robust random forest algorithm, normalizes each component of the anomaly contribution vector into the cutting probability of the corresponding feature dimension, outputs the comprehensive anomaly score and key isolation dimension, and determines the anomaly score boundary based on the comprehensive anomaly score. The update module writes the state feature vectors that do not exceed the screening limit in terms of matrix contour distance and comprehensive anomaly score not exceeding the anomaly score limit into the reference sample set; The judgment module determines a continuous fault when the duration of the candidate abnormal feature vector reaches the preset duration and the comprehensive abnormal score exceeds the abnormal score limit. It determines the fault category based on the abnormal contribution vector and the key isolation dimension, and generates generator set status monitoring results.

[0007] Optionally, the acquisition and processing module includes: The generator set collects voltage, current, temperature, vibration, and speed data. The voltage data is the voltage measurement value corresponding to each sampling time of the generator set. The current data is the current measurement value corresponding to each sampling time of the generator set. The temperature data is the temperature measurement value corresponding to each sampling time of the generator set. The vibration data is the vibration measurement value corresponding to each sampling time of the generator set. The speed data is the speed measurement value corresponding to each sampling time of the generator set. Each of the voltage, current, temperature, vibration, and rotational speed data is assigned a sampling time identifier, and a unified sampling time sequence is determined based on each sampling time identifier. Voltage data, current data, temperature data, vibration data, and rotational speed data are mapped to a unified sampling time sequence according to their corresponding sampling time identifiers. For sampling times in the unified sampling time sequence that lack corresponding measurement values, interpolation is performed using the same type of measurement values ​​corresponding to adjacent sampling times to generate time-aligned multidimensional state data. According to the set window length, voltage data, current data, temperature data, vibration data and rotation speed data corresponding to continuous and uniform sampling times are extracted from the time-aligned multidimensional state data. The extraction range is moved sequentially along the uniform sampling time sequence according to the set window movement step size. The data in each extraction range are arranged according to the sampling time to generate a multidimensional state subsequence.

[0008] Optionally, the matrix contour module includes: According to the set subsequence length, voltage data, current data, temperature data, vibration data and rotation speed data are synchronously extracted from each sampling position of the multidimensional state subsequence to generate a local multidimensional state subsequence. The data of each monitoring dimension are then normalized to generate a normalized local multidimensional state subsequence. Each normalized local multidimensional state subsequence is used as a query subsequence in turn, and the remaining normalized local multidimensional state subsequences that do not overlap with the query subsequences in sampling position are used as comparison subsequences to generate a set of comparison subsequences. For each monitoring dimension of each comparison subsequence in the query subsequence and comparison subsequence sets, the comparison subsequences are moved sequentially according to the unified sampling time within a set time lag range. The effective overlapping data segments corresponding to the time positions in the query subsequence and the moved comparison subsequence are extracted, and the one-dimensional distance is calculated. The minimum value is selected from the single-dimensional distances corresponding to each monitoring dimension under each time delay as the time delay matching distance. The time delay that generates the time delay matching distance is determined as the target time delay, and the time delay matching distance vector and the target time delay vector are generated. The time delay constraint multidimensional distance is calculated based on the difference between each target time delay and the median of the target time delay in the target time delay vector, and the time delay matching distance vector. The minimum value is selected from all time-delay constraint multidimensional distances corresponding to each query subsequence as the first matrix contour distance. The position of the comparison subsequence that produces the minimum value is determined as the first matching position. The first matrix contour distances are arranged according to the sampling positions of the query subsequences to generate the first matrix contour distance sequence. One of the monitoring dimensions, voltage, current, temperature, vibration and rotation speed, is removed in turn. After each removal, the corresponding set of comparison subsequences is retraced based on the remaining monitoring dimensions. The time delay matching distance vector and the target time delay vector are regenerated, and the time delay constraint multidimensional distance is recalculated. The smallest time delay constraint multidimensional distance obtained by recalculation is determined as the counterfactual profile distance corresponding to the removed monitoring dimension. Calculate the difference between the first matrix contour distance and the distances of each counterfactual contour, set the differences that are less than zero to zero, and normalize the differences after setting them to zero to generate an initial anomaly contribution vector. The corresponding monitoring dimensions are arranged from largest to smallest according to the initial abnormal contribution. Starting from the largest abnormal contribution, they are added sequentially. When the cumulative abnormal contribution first reaches the set ratio, the accumulated monitoring dimensions are determined as the set of contribution dimensions. Extract the data corresponding to the contribution dimension set from the normalized local multidimensional state subsequence, generate the contribution-focused state subsequence, and regenerate the query subsequence, comparison subsequence set, time delay matching distance vector, target time delay vector and time delay constraint multidimensional distance based on the contribution-focused state subsequence. Select the minimum value from the time delay constraint multidimensional distances corresponding to each query subsequence to generate the second matrix contour distance sequence. The first and second contour distance boundaries are determined based on the first and second matrix contour distance sequences, respectively. Sampling positions where both the first and second matrix contour distances continuously exceed the corresponding contour distance boundaries are merged into anomaly subsequences. The start and end positions and duration of the anomaly subsequences are obtained based on the first and last sampling positions and the corresponding unified sampling time. The components corresponding to the contribution dimension set in the initial anomaly contribution vector are renormalized to generate anomaly contribution vector.

[0009] Optionally, the feature reconstruction module includes: Obtain the matrix contour distance, start position, end position, duration, and anomaly contribution vector corresponding to each abnormal subsequence, and aggregate the matrix contour distance, start position, end position, duration, and anomaly contribution vector belonging to the same abnormal subsequence; Following the monitoring dimensions of voltage, current, temperature, vibration, and rotational speed, the components in the anomaly contribution vector are read sequentially to obtain the voltage anomaly contribution, current anomaly contribution, temperature anomaly contribution, vibration anomaly contribution, and rotational speed anomaly contribution. According to the order of matrix contour distance, starting position, ending position, duration, voltage anomaly contribution, current anomaly contribution, temperature anomaly contribution, vibration anomaly contribution, and rotational speed anomaly contribution, the data corresponding to the same anomaly subsequence are arranged sequentially to generate the state feature vector corresponding to the anomaly subsequence. Extract the state feature vectors that are pre-labeled as normal operating states from the state feature vectors, and arrange them in ascending order of their starting positions; Each pre-labeled state feature vector of the sorted system is used as a reference sample to establish a reference sample set.

[0010] Optionally, the anomaly filtering module includes: Read the state feature vectors of each normal operating state in the reference sample set in sequence, extract the matrix contour distance from the state feature vectors of each normal operating state, and arrange the extracted matrix contour distances according to the arrangement order of the state feature vectors of each normal operating state in the reference sample set to generate a reference matrix contour distance set. Calculate the median of all matrix contour distances in the reference matrix contour distance set, calculate the absolute difference between each matrix contour distance and the median, and calculate the median of all absolute differences to generate the median absolute deviation of the matrix contour distance. The median of the matrix contour distance and the absolute deviation of the median matrix contour distance are used to determine the screening limits; The state feature vectors that are not written into the reference sample set are used to form a set of state feature vectors to be screened. The matrix contour distance is extracted from each state feature vector in the set of state feature vectors to be screened in turn, and the extracted matrix contour distance is established with the state feature vector to generate a matrix contour distance record to be screened. The matrix contour distances in each matrix contour distance record to be screened are compared with the screening limit. The matrix contour distance records to be screened that are greater than the screening limit are extracted. The state feature vectors to which each matrix contour distance record to be screened belongs are obtained according to the correspondence. The obtained state feature vectors are determined as candidate anomaly feature vectors. The candidate anomaly feature vectors are arranged in ascending order of their starting positions to generate a set of candidate anomaly feature vectors.

[0011] Optionally, the determination of the cut probability driven by abnormal contribution in the forest module includes: Read each candidate anomaly feature vector in the candidate anomaly feature vector set in sequence, extract the anomaly contribution vector from each candidate anomaly feature vector, and read each component in the anomaly contribution vector in the order of voltage, current, temperature, vibration and rotation speed to generate an anomaly contribution component sequence corresponding to the candidate anomaly feature vector. The abnormal contribution values ​​of voltage, current, temperature, vibration, and rotational speed in the abnormal contribution component sequence are summed to generate the total abnormal contribution value of the corresponding candidate abnormal feature vectors. Divide each abnormal contribution in the abnormal contribution component sequence by the total abnormal contribution value to obtain the segmentation probability of the corresponding monitoring dimension. When the total abnormal contribution is zero, the cutting probability corresponding to voltage, current, temperature, vibration and rotation speed is set to 0.2 respectively; When the total value of abnormal contribution is greater than zero, the normalized cut probability is used. Based on the determined cutting probabilities arranged in the order of voltage, current, temperature, vibration, and rotational speed, a cutting probability vector corresponding to the corresponding candidate abnormal feature vector is generated. The cutting probability vectors are then arranged in the order of their arrangement in the candidate abnormal feature vector set to generate a cutting probability vector set.

[0012] Optionally, the generation of the comprehensive anomaly score and key isolation dimensions includes: The candidate anomaly feature vector and its corresponding cutting probability vector are input into the robust random forest algorithm. The number of cutting nodes that the candidate anomaly feature vector passes through in each random tree, the monitoring dimension used for each cutting node, and the cutting probability corresponding to the monitoring dimension are recorded to generate an isolation depth set and an isolation path record. Read each isolation depth in the isolation depth set in sequence, generate the corresponding single-tree anomaly score according to each isolation depth, and form a single-tree anomaly score set according to the arrangement order of the random trees. Calculate the average value of all single-tree anomaly scores in the single-tree anomaly score set to generate the comprehensive anomaly score of the corresponding candidate anomaly feature vector. The isolation path records are read sequentially, and the monitoring dimensions used for each cutting node are classified according to voltage, current, temperature, vibration and rotation speed. The cutting probability corresponding to the same monitoring dimension is accumulated to generate voltage isolation contribution, current isolation contribution, temperature isolation contribution, vibration isolation contribution and rotation speed isolation contribution. The monitoring dimension corresponding to the isolation contribution with the largest value is determined as the key isolation dimension. The state feature vectors of each normal operating state in the reference sample set are used to generate corresponding cutting probability vectors. These vectors are then input into the same robust random forest algorithm. Based on the method of generating isolation depth set and comprehensive anomaly score, comprehensive anomaly scores corresponding to the state feature vectors of each normal operating state are generated to form a reference comprehensive anomaly score set. Calculate the median and median absolute deviation of the reference comprehensive abnormal score set, and determine the abnormal score limit by summing the median with the product of the preset abnormal score coefficient and the median absolute deviation.

[0013] Optionally, the update module includes: Read the feature vectors of each state in sequence, extract the matrix contour distance from each feature vector, and obtain the comprehensive anomaly score; Compare the matrix contour distance corresponding to each state feature vector with the screening limit, and extract the state feature vectors whose matrix contour distance is less than or equal to the screening limit from each state feature vector. The comprehensive anomaly score corresponding to each extracted state feature vector is compared with the anomaly score limit, and state feature vectors with comprehensive anomaly scores less than or equal to the anomaly score limit are extracted again from each extracted state feature vector. The extracted state feature vectors are then sequentially written into the reference sample set. State feature vectors whose matrix contour distance is greater than the screening limit or whose comprehensive anomaly score is greater than the anomaly score limit are not written into the reference sample set.

[0014] Optionally, the determination of persistent fault in the determination module includes: Read each candidate anomaly feature vector sequentially, extract the duration length from each candidate anomaly feature vector, and read the comprehensive anomaly score; The duration length corresponding to each candidate abnormal feature vector is compared with the preset duration length. When the duration length is greater than or equal to the preset duration length, the comprehensive abnormal score of the corresponding candidate abnormal feature vector is read. The overall anomaly score is compared with the anomaly score limit. When the overall anomaly score is greater than the anomaly score limit, the corresponding candidate anomaly feature vector is determined to be a persistent fault, and the start position, end position, duration, anomaly contribution vector, overall anomaly score and key isolation dimension in the candidate anomaly feature vector are retained. When the duration of a candidate abnormal feature vector is less than the preset duration, or when the duration is greater than or equal to the preset duration and the overall abnormal score is less than or equal to the abnormal score limit, the corresponding candidate abnormal feature vector will not be judged as a persistent fault.

[0015] Optionally, the generation of fault category and generator set status monitoring results in the determination module includes: Read the candidate anomaly feature vectors that are determined to be persistent faults in sequence, extract the anomaly contribution vector from each candidate anomaly feature vector, and read the key isolation dimension. Following the order of voltage, current, temperature, vibration, and rotational speed, each abnormal contribution degree in the abnormal contribution degree vector is read sequentially, the values ​​of each abnormal contribution degree are compared, and the monitoring dimension corresponding to the abnormal contribution degree with the largest value is determined as the main abnormal dimension. When the anomaly contribution of multiple monitoring dimensions is at its maximum value, the monitoring dimension that is the same as the key isolation dimension is determined as the main anomaly dimension. If they are all different from the key isolation dimension, the monitoring dimension that comes first is determined in the order of voltage, current, temperature, vibration and rotation speed. The main anomaly dimension is compared with the key isolation dimension. When the two are the same, the fault corresponding to the monitoring dimension is determined as the fault category. When the two are different, the fault corresponding to the main anomaly dimension and the fault corresponding to the key isolation dimension are jointly determined as the fault category. Among them, voltage, current, temperature, vibration and speed correspond to voltage fault, current fault, temperature fault, vibration fault and speed fault, respectively. The results of persistent fault determination, identified fault category, anomaly contribution vector, key isolation dimension, and comprehensive anomaly score are correlated with candidate anomaly feature vectors to generate generator set condition monitoring results.

[0016] The beneficial effects of this invention are: This invention performs time alignment, windowing, and multidimensional matrix profile analysis on the voltage, current, temperature, vibration, and speed data of generator sets. By combining time-delay matching distance, time-delay constrained multidimensional distance, and counterfactual profile distance, it can identify multi-parameter correlated anomalies when the response times of different monitoring parameters are inconsistent. It also obtains the start and end positions, duration, and anomaly contribution vector of the anomaly subsequence, thereby reducing the impact of asynchronous parameter responses on the anomaly identification results and improving the accuracy of anomaly location and anomaly parameter identification in downhole generator sets.

[0017] This invention converts the anomaly contribution vector into the cutting probability of each monitoring dimension in a robust random forest algorithm. Combined with the isolation depth and isolation path of each random tree, it generates a comprehensive anomaly score and key isolation dimensions, enabling monitoring dimensions with high anomaly contributions to participate in fault status judgment. At the same time, it determines the screening threshold and anomaly score threshold based on the state feature vector of the normal operating state, and only writes the state feature vectors whose matrix contour distance and comprehensive anomaly score do not exceed the corresponding thresholds into the reference sample set. This allows the reference sample to be updated with the normal operating data, reducing misjudgments caused by inconsistencies between the fixed reference sample and the actual operating state.

[0018] This invention compares the duration of candidate anomaly feature vectors with a preset duration and compares the comprehensive anomaly score with anomaly score limits, determining a persistent fault only when both conditions are met simultaneously, thus distinguishing between short-term fluctuations and persistent anomalies. Furthermore, it combines the main anomaly dimension and key isolation dimension determined by the anomaly contribution vector to determine a single fault category or a compound fault category, generating generator set status monitoring results that include persistent fault determination results, fault category, anomaly contribution vector, key isolation dimension, and comprehensive anomaly score, thereby achieving continuous monitoring and fault classification of the operating status of downhole generator sets. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the intelligent status monitoring system for generator sets based on measurement while drilling proposed in this invention; Figure 2 This is a flowchart of the multidimensional matrix contour analysis and anomaly subsequence determination in this invention; Figure 3This is a graph showing the first matrix contour distance and the second matrix contour distance in an embodiment of the present invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figures 1-3 The intelligent condition monitoring system for generator sets based on measurement-while-drilling downhole includes the following steps: The data acquisition and processing module collects voltage, current, temperature, vibration and speed data of the generator set, performs alignment and windowing processing, and generates multi-dimensional state subsequences. The matrix profile module performs multidimensional matrix profile analysis on multidimensional state subsequences to obtain matrix profile distance, start and end positions of anomalous subsequences, duration length, and anomalous contribution vector. The feature reconstruction module constructs a state feature vector from the matrix contour distance, the start and end positions and duration of the abnormal subsequence, and the abnormal contribution vector, and establishes a reference sample set based on the state feature vector pre-calibrated as the normal operating state. The anomaly screening module determines the screening limit based on the matrix contour distance in the reference sample set, and identifies state feature vectors whose matrix contour distance exceeds the screening limit as candidate anomaly feature vectors. The forest module inputs the state feature vector into the robust random forest algorithm, normalizes each component of the anomaly contribution vector into the cutting probability of the corresponding feature dimension, outputs the comprehensive anomaly score and key isolation dimension, and determines the anomaly score boundary based on the comprehensive anomaly score. The update module writes the state feature vectors that do not exceed the screening limit in terms of matrix contour distance and comprehensive anomaly score not exceeding the anomaly score limit into the reference sample set; The judgment module determines a continuous fault when the duration of the candidate abnormal feature vector reaches the preset duration and the comprehensive abnormal score exceeds the abnormal score limit. It determines the fault category based on the abnormal contribution vector and the key isolation dimension, and generates generator set status monitoring results.

[0022] In this embodiment, the acquisition and processing module includes: The generator set collects voltage, current, temperature, vibration, and speed data. The voltage data is the voltage measurement value corresponding to each sampling time of the generator set. The current data is the current measurement value corresponding to each sampling time of the generator set. The temperature data is the temperature measurement value corresponding to each sampling time of the generator set. The vibration data is the vibration measurement value corresponding to each sampling time of the generator set. The speed data is the speed measurement value corresponding to each sampling time of the generator set. Each of the voltage, current, temperature, vibration, and rotational speed data is assigned a sampling time identifier, and a unified sampling time sequence is determined based on each sampling time identifier. Specifically, the sampling time corresponding to each measured value in voltage data, current data, temperature data, vibration data, and rotational speed data is obtained, and the sampling time is written into the corresponding measured value as a sampling time identifier; the earliest sampling time in each sampling time identifier is taken as the starting point and the latest sampling time is taken as the ending point, and the sampling time is generated sequentially according to the set time interval, and the generated sampling time is arranged in chronological order to form a unified sampling time sequence; Voltage data, current data, temperature data, vibration data, and rotational speed data are mapped to a unified sampling time sequence according to their corresponding sampling time identifiers. For sampling times in the unified sampling time sequence that lack corresponding measurement values, interpolation is performed using the same type of measurement values ​​corresponding to adjacent sampling times to generate time-aligned multidimensional state data. Specifically, for sampling moments in the unified sampling time sequence that are not mapped to corresponding measurement values, the closest similar measurement values ​​before and after the sampling moment are determined respectively. Based on the time position of the sampling moment with missing measurement values ​​between adjacent sampling moments, linear interpolation is performed on the adjacent similar measurement values, and the interpolation result is used as the measurement value corresponding to the sampling moment. According to the set window length, voltage data, current data, temperature data, vibration data and rotation speed data corresponding to continuous and uniform sampling times are extracted from the time-aligned multidimensional state data. The extraction range is moved sequentially along the uniform sampling time sequence according to the set window movement step size. The data in each extraction range are arranged according to the sampling time to generate a multidimensional state subsequence. In Example 1, voltage, current, temperature, vibration, and rotational speed data are all recorded by the downhole data acquisition processor along with their corresponding sampling times, and arranged chronologically. A unified sampling time sequence is established with the earliest sampling time as the starting point and the latest sampling time as the ending point, at 10ms intervals. If no similar measurement value is obtained at the unified sampling time t, the closest similar measurement values ​​before and after that time are determined. and Their sampling times are respectively and ,according to The interpolation results are calculated and written to the corresponding unified sampling time. After time alignment, the data corresponding to 200 consecutive unified sampling times are used as a window with a window length of 2 seconds. The window movement step is 50 unified sampling times, moving 0.5 seconds along the unified sampling time sequence each time. The voltage, current, temperature, vibration and rotation speed data within the same window are arranged according to the sampling time to generate a multidimensional state subsequence.

[0023] In this embodiment, the matrix contour module includes: According to the set subsequence length, voltage data, current data, temperature data, vibration data and rotation speed data are synchronously extracted from each sampling position of the multidimensional state subsequence to generate a local multidimensional state subsequence. The data of each monitoring dimension are then normalized to generate a normalized local multidimensional state subsequence. Specifically, the subsequence length is defined as the number of consecutive uniform sampling moments extracted from each multidimensional state subsequence. This number is less than the total number of uniform sampling moments contained in the multidimensional state subsequence and remains unchanged throughout the extraction process. For example, if the multidimensional state subsequence contains 200 uniform sampling moments, the subsequence length is set to 50 uniform sampling moments, and each monitoring dimension is extracted synchronously according to this length. Each normalized local multidimensional state subsequence is used as a query subsequence in turn, and the remaining normalized local multidimensional state subsequences that do not overlap with the query subsequences in sampling position are used as comparison subsequences to generate a set of comparison subsequences. Specifically, each normalized local multidimensional state subsequence has a corresponding start sampling position and end sampling position; after determining one of them as the query subsequence, the start sampling positions and end sampling positions of the remaining normalized local multidimensional state subsequences are compared one by one, and the normalized local multidimensional state subsequence whose end sampling position is earlier than the start sampling position of the query subsequence, or whose start sampling position is later than the end sampling position of the query subsequence, is determined as the comparison subsequence; For each monitoring dimension of each comparison subsequence in the query subsequence and comparison subsequence sets, within a set time lag range, the comparison subsequences are moved sequentially according to a unified sampling time. Effective overlapping data segments corresponding to the time positions in the query subsequence and the moved comparison subsequences are extracted, and the one-dimensional distance is calculated. ; in, Indicates the first The query subsequence and the first The comparison subsequence at the _th ... Time lag of each monitoring dimension One-dimensional distance below, Indicates the sequence number of the subsequence being queried. Indicates the first The comparison subsequence index in the set of comparison subsequences corresponding to each query subsequence. Indicates the sequence number of the monitoring dimension participating in the current distance calculation; This indicates the number of sampling moments the comparison subsequence has shifted relative to the query subsequence. A positive value indicates a shift backward, a negative value indicates a shift forward, and a zero value indicates no shift. Indicates time delay The query subsequence and the comparison subsequence contain the number of sampling points in the effective overlapping data segment. Indicates the sampling point number within the valid overlapping data segment. Indicates the first The query subsequence at the _th ... Within the monitoring dimension, the first Normalized data corresponding to each sampling point Indicates the first The comparison subsequence at the _th ... Within the monitoring dimension and the first Each sampling point is based on time delay The corresponding normalized data; This formula is derived from the Euclidean distance calculation method and improved by incorporating the time-delay shift relationship between the query subsequence and the comparison subsequence. Under each set time delay, the formula calculates the sum of squared differences between corresponding sampled data in the query subsequence and the shifted comparison subsequence. It then averages these differences by taking the square root of the average of the corresponding sampled data to obtain the single-dimensional distance for the corresponding monitoring dimension. Compared to the conventional distance calculation method that directly compares data from the same sampling location, this formula adds a time-delay variable and only calculates for sampled data that can form a corresponding relationship after shifting. This allows voltage, current, temperature, vibration, and rotational speed data to achieve morphological matching under different target time delays, avoiding inconsistencies in response time across different monitoring dimensions that could lead to excessively large distances, thereby improving the accuracy of matching results for similar state subsequences. Specifically, the set time delay range refers to the range of sampling times that the comparison subsequence is allowed to move relative to the query subsequence along the uniform sampling time sequence. Centered on zero time delay, the maximum forward and backward movement is the same and less than the set subsequence length. For example, when the set subsequence length is 50 uniform sampling times, the set time delay range is determined to be from -5 to 5 uniform sampling times, and the one-dimensional distance corresponding to each time delay is calculated sequentially. After shifting the comparison subsequence by the corresponding number of sampling times, delete the data in both the query subsequence and the comparison subsequence that cannot form a corresponding sampling time relationship, and retain the data that can correspond in the same sampling order after the shift. For example, when a comparison subsequence of length 50 is shifted backward by 5 sampling times, extract the first 45 data points of the query subsequence and the last 45 data points of the comparison subsequence to form an effective overlapping data segment; The minimum value is selected from the single-dimensional distances corresponding to each monitoring dimension under each time delay as the time delay matching distance. The time delay that generates the time delay matching distance is determined as the target time delay, and the time delay matching distance vector and the target time delay vector are generated. Specifically, for voltage, current, temperature, vibration, and rotational speed data, the one-dimensional distances obtained from the same query subsequence and the same comparison subsequence within a set time delay range are arranged according to time delay. The one-dimensional distance with the smallest value is selected as the time delay matching distance for the corresponding data, and the time delay corresponding to this one-dimensional distance is taken as the target time delay. If multiple identical minimum values ​​exist, the time delay with the smallest absolute value is selected. The time delay matching distances and target time delays are arranged in the order of voltage, current, temperature, vibration, and rotational speed, generating time delay matching distance vectors and target time delay vectors respectively. The time delay constraint multidimensional distance is calculated based on the difference between each target time delay and the median target time delay in the target time delay vector, and the time delay matching distance vector. ; in, Indicates the first The query subsequence and the first Time-delay constrained multidimensional distance between comparison subsequences This indicates the number of monitoring dimensions involved in the current time-delay constrained multidimensional distance calculation, during the first matrix profile analysis. For the fifth time, during the second matrix contour analysis... The number of monitoring dimensions included in the contribution dimension set; This indicates the sequence number of the monitoring dimension participating in the current calculation. Indicates the first The query subsequence and the first The comparison subsequence at the _th ... Time-delay matching distance across monitoring dimensions This indicates the target time delay that generates the time delay matching distance. Indicates the first The query subsequence and the first The median time delay of the target time delay vector corresponds to each comparison subsequence. This indicates that the median value is taken from all the monitoring dimensions involved in the current calculation. Indicates the first The absolute difference between the target time delay and the median time delay for each monitoring dimension This indicates the maximum number of sampling time shifts included within the set time delay range. This indicates a pre-defined positive number to prevent the denominator from being zero; It represents exponential operations with the natural constant as the base; This formula is composed of the average time-delay matching distance of each monitoring dimension and a cross-dimensional time-delay consistency constraint term. The average time-delay matching distance represents the overall difference between the query subsequence and the comparison subsequence across each monitoring dimension. The cross-dimensional time-delay consistency constraint term is formed by exponentially transforming the median of the absolute difference between each target time delay and its median. When the target time delays of each monitoring dimension are close, this constraint term approaches zero; when the target time delays are dispersed, this constraint term increases. Compared to conventional multidimensional matrix contour distance, which only sums or averages the distances of each dimension, this formula incorporates the response sequence relationship between different monitoring dimensions into the distance calculation. This allows comparison subsequences with only local morphological similarities but inconsistent cross-dimensional response relationships to obtain larger time-delay constrained multidimensional distances, thereby improving the ability to distinguish between abnormal and normal fluctuations in multi-parameter correlations of generator sets. The minimum value is selected from all time-delay constraint multidimensional distances corresponding to each query subsequence as the first matrix contour distance. The position of the comparison subsequence that produces the minimum value is determined as the first matching position. The first matrix contour distances are arranged according to the sampling positions of the query subsequences to generate the first matrix contour distance sequence. Specifically, for each query subsequence, it is compared with the time-delay constraint multidimensional distances corresponding to each comparison subsequence in the comparison subsequence set. The time-delay constraint multidimensional distance with the smallest value is selected as the first matrix contour distance, and the starting sampling position of the comparison subsequence that produces the minimum value is determined as the first matching position. When there are multiple identical minimum values, the comparison subsequence with the earliest starting sampling position is selected. One of the monitoring dimensions, voltage, current, temperature, vibration and rotation speed, is removed in turn. After each removal, the corresponding set of comparison subsequences is retraced based on the remaining monitoring dimensions. The time delay matching distance vector and the target time delay vector are regenerated, and the time delay constraint multidimensional distance is recalculated. The smallest time delay constraint multidimensional distance obtained by recalculation is determined as the counterfactual profile distance corresponding to the removed monitoring dimension. Specifically, voltage, current, temperature, vibration, and rotational speed are used as the data to be eliminated. Only one type of data is eliminated each time, while the other four types of data are retained. For the same query subsequence, the time delay matching distance vector, the target time delay vector, and the time delay constraint multidimensional distance between it and each comparison subsequence in the comparison subsequence set are recalculated, and the minimum time delay constraint multidimensional distance is used as the counterfactual contour distance corresponding to the eliminated data. Calculate the difference between the first matrix contour distance and the distances of each counterfactual contour, set the differences that are less than zero to zero, and normalize the differences after setting them to zero to generate an initial anomaly contribution vector. The corresponding monitoring dimensions are arranged from largest to smallest according to the initial abnormal contribution. Starting from the largest abnormal contribution, they are added sequentially. When the cumulative abnormal contribution first reaches the set ratio, the accumulated monitoring dimensions are determined as the set of contribution dimensions. Specifically, the sum of all components in the initial anomaly contribution vector is 1. The corresponding monitoring dimensions are arranged in descending order of their component values, and the sum is accumulated item by item starting from the first value in the sort. When the accumulated value first exceeds or equals a set proportion, the accumulation stops, and all monitoring dimensions from the first value in the sort to the current monitoring dimension are determined as the contribution dimension set. For example, when the set proportion is 0.8, all monitoring dimensions whose accumulated anomaly contribution first reaches 0.8 are selected. Extract the data corresponding to the contribution dimension set from the normalized local multidimensional state subsequence, generate the contribution-focused state subsequence, and regenerate the query subsequence, comparison subsequence set, time delay matching distance vector, target time delay vector and time delay constraint multidimensional distance based on the contribution-focused state subsequence. Select the minimum value from the time delay constraint multidimensional distances corresponding to each query subsequence to generate the second matrix contour distance sequence. Specifically, for each query subsequence corresponding to the contribution dimension set, the time-delay constrained multidimensional distance between it and each comparison subsequence in the comparison subsequence set is calculated, and the minimum value of the obtained distance is selected as the second matrix contour distance of the query subsequence; the second matrix contour distances are arranged in order according to the starting sampling position of each query subsequence to generate the second matrix contour distance sequence. The first and second contour distance boundaries are determined based on the first and second matrix contour distance sequences, respectively. Sampling positions where both the first and second matrix contour distances continuously exceed the corresponding contour distance boundaries are merged into anomaly subsequences. The start and end positions and duration of the anomaly subsequences are obtained based on the first and last sampling positions and the corresponding unified sampling time. The components corresponding to the contribution dimension set in the initial anomaly contribution vector are re-normalized to generate anomaly contribution vector. In Example 1, for each local multidimensional state subsequence, the median and median absolute deviation of voltage, current, temperature, vibration, and rotational speed data are calculated respectively. The difference between each sampled data and the corresponding median is divided by the sum of the corresponding median absolute deviation and a preset positive number to generate normalized data. The preset positive number is taken as... The subsequence length was set to 50 uniform sampling times, the time lag range was set to -5 to 5 uniform sampling times, and the contribution ratio was set to 0.8. The median and median absolute deviation were calculated for the first matrix contour distance sequence and the second matrix contour distance sequence, respectively. The sum of the corresponding median and three times the median absolute deviation was determined as the boundary of the first contour distance and the boundary of the second contour distance. After calculating the difference between the first matrix contour distance and each counterfactual contour distance, the negative difference value was set to zero, and the initial anomaly contribution was generated by dividing each non-negative difference value by the sum of all non-negative difference values. When the sum of all non-negative difference values ​​was zero, the initial anomaly contribution values ​​corresponding to voltage, current, temperature, vibration, and rotational speed were all set to 0.2, and the five monitoring dimensions were determined as the contribution dimension set. The query subsequences whose first matrix contour distance and second matrix contour distance both exceed the corresponding contour distance boundary and whose starting sampling positions are consecutive are merged. The starting sampling position of the first query subsequence is taken as the starting position of the abnormal subsequence, and the ending sampling position of the last query subsequence is taken as the ending position of the abnormal subsequence. The duration is determined according to the difference between the unified sampling time corresponding to the starting position and the ending position.

[0024] In this embodiment, the feature reconstruction module includes: Obtain the matrix contour distance, start position, end position, duration, and anomaly contribution vector corresponding to each abnormal subsequence, and aggregate the matrix contour distance, start position, end position, duration, and anomaly contribution vector belonging to the same abnormal subsequence; Following the monitoring dimensions of voltage, current, temperature, vibration, and rotational speed, the components in the anomaly contribution vector are read sequentially to obtain the voltage anomaly contribution, current anomaly contribution, temperature anomaly contribution, vibration anomaly contribution, and rotational speed anomaly contribution. According to the order of matrix contour distance, starting position, ending position, duration, voltage anomaly contribution, current anomaly contribution, temperature anomaly contribution, vibration anomaly contribution, and rotational speed anomaly contribution, the data corresponding to the same anomaly subsequence are arranged sequentially to generate the state feature vector corresponding to the anomaly subsequence. Extract the state feature vectors that are pre-labeled as normal operating states from the state feature vectors, arrange them in ascending order of their starting positions, and keep the order of the data in each state feature vector unchanged. Specifically, the state feature vector pre-labeled as the normal operating state refers to matching the sampling period when the generator set is confirmed to be in the normal operating state with the start and end positions of the corresponding abnormal subsequence of each state feature vector before establishing the reference sample set; when the start position to the end position are all within the sampling period, the corresponding state feature vector is configured with a normal operating state label; Each pre-labeled state feature vector of the sorted state is used as a reference sample to establish a reference sample set. In Example 1, the start and end positions of the same abnormal subsequence are used as the association basis. The matrix contour distance, duration, and abnormal contribution vector corresponding to the abnormal subsequence are written into the same data record. A state feature vector is generated in a fixed order of matrix contour distance, start position, end position, duration, voltage abnormal contribution, current abnormal contribution, temperature abnormal contribution, vibration abnormal contribution, and rotational speed abnormal contribution. For example, a certain state feature vector can be represented as [2.36, 120, 168, 4.8, 0.12, 0.08, 0.18, 0.47, 0.15], where the start and end positions are the sampling positions in the unified sampling time sequence, and the duration is the difference between the corresponding start and end sampling times. The sampling period when the generator set is confirmed to be in normal operation is matched with the start and end positions of the corresponding state feature vectors. The state feature vectors whose start and end positions are both within the sampling period are marked as normal operation states. The state feature vectors of each normal operation state are written into the same set in ascending order of the start position to form a reference sample set. Each item in the reference sample set corresponds to a state feature vector of a normal operation state.

[0025] In this embodiment, the anomaly filtering module includes: Read the state feature vectors of each normal operating state in the reference sample set in sequence, extract the matrix contour distance from the state feature vectors of each normal operating state, and arrange the extracted matrix contour distances according to the arrangement order of the state feature vectors of each normal operating state in the reference sample set to generate a reference matrix contour distance set. Specifically, the state feature vectors of the normal operation state in the reference sample set occupy the first to Nth items in order of writing; when reading the i-th state feature vector, the matrix contour distance that ranks first in the state feature vector is extracted and written into the i-th item of the reference matrix contour distance set, so that each matrix contour distance in the reference matrix contour distance set corresponds to the state feature vector in the reference sample set. Calculate the median of all matrix contour distances in the reference matrix contour distance set, calculate the absolute difference between each matrix contour distance and the median, and calculate the median of all absolute differences to generate the median absolute deviation of the matrix contour distance. Specifically, all matrix contour distances in the reference matrix contour distance set are arranged in ascending order of value; when the number is odd, the middle value is taken, and when the number is even, the average of the two middle values ​​is taken as the median. The absolute value of the difference between each matrix contour distance and the median is calculated and arranged in ascending order of value. The median of the absolute values ​​is determined in the same way and is taken as the median absolute deviation of the matrix contour distance. The median of the matrix contour distance and the absolute deviation of the median matrix contour distance are used to determine the screening threshold: ; in, As a screening boundary, This is the median of all matrix contour distances in the reference matrix contour distance set. The absolute deviation of the median distance from the matrix contour. The preset screening coefficient is greater than zero; The state feature vectors that are not written into the reference sample set are used to form a set of state feature vectors to be screened. The matrix contour distance is extracted from each state feature vector in the set of state feature vectors to be screened in turn, and the extracted matrix contour distance is established with the state feature vector to generate a matrix contour distance record to be screened. The matrix contour distances in each matrix contour distance record to be screened are compared with the screening limit. The matrix contour distance records to be screened that are greater than the screening limit are extracted. The state feature vectors to which each matrix contour distance record to be screened belongs are obtained according to the correspondence. The obtained state feature vectors are determined as candidate anomaly feature vectors. The candidate anomaly feature vectors are arranged in ascending order of their starting positions to generate a set of candidate anomaly feature vectors. In Example 1, the preset screening coefficient is 3. The sum of the median of the reference matrix contour distance set and the absolute deviation of three times the median of the matrix contour distance is determined as the screening limit. When the absolute deviation of the matrix contour distance is zero, the median is used as the screening limit. State feature vectors not written into the reference sample set are read sequentially, and their matrix contour distances are compared with the screening limit. Only when the matrix contour distance is greater than the screening limit is the corresponding state feature vector identified as a candidate anomaly feature vector; otherwise, no candidate anomaly processing is performed. For example, when the median is 1.20 and the absolute deviation of the median is 0.15, the screening limit is 1.65, and the state feature vector with a matrix contour distance of 1.82 is identified as a candidate anomaly feature vector.

[0026] In this embodiment, the determination of the cutting probability driven by abnormal contribution in the forest module includes: Read each candidate anomaly feature vector in the candidate anomaly feature vector set in sequence, extract the anomaly contribution vector from each candidate anomaly feature vector, and read each component in the anomaly contribution vector in the order of voltage, current, temperature, vibration and rotation speed to generate an anomaly contribution component sequence corresponding to the candidate anomaly feature vector. The abnormal contribution values ​​of voltage, current, temperature, vibration, and rotational speed in the abnormal contribution component sequence are summed to generate the total abnormal contribution value of the corresponding candidate abnormal feature vectors. Divide each abnormal contribution in the abnormal contribution component sequence by the total abnormal contribution value to obtain the segmentation probability of the corresponding monitoring dimension: ; in, For the first Among the candidate abnormal feature vectors, the th... The probability of segmentation for each monitoring dimension. For the first Anomaly contribution of each monitoring dimension For the first The abnormal contribution of each monitoring dimension; When the total abnormal contribution is zero, the cutting probability corresponding to voltage, current, temperature, vibration and rotation speed is set to 0.2 respectively; When the total value of abnormal contribution is greater than zero, the normalized cut probability is used. The cutting probabilities are determined by arranging them in the order of voltage, current, temperature, vibration and rotation speed. A cutting probability vector corresponding to the corresponding candidate abnormal feature vector is generated. The cutting probability vectors are then arranged in the order of the candidate abnormal feature vector set to generate a cutting probability vector set. In Example 1, the anomaly contribution vectors in the candidate anomaly feature vectors are read in a fixed order of voltage, current, temperature, vibration, and rotational speed. The five anomaly contribution values ​​are summed to obtain the total anomaly contribution value. When the total anomaly contribution value is greater than zero, each anomaly contribution value is divided by the total anomaly contribution value, and the five values ​​obtained are used as the cutting probabilities of the corresponding monitoring dimensions, with the sum of the five cutting probabilities being 1. For example, when the anomaly contribution vector is [0.12, 0.08, 0.18, 0.47, 0.15], the total anomaly contribution value is 1, and the corresponding cutting probability vector is [0.12, 0.08, 0.18, 0.47, 0.15]. When all five anomaly contribution values ​​are zero, the corresponding cutting probabilities are all set to 0.2. After generating the cutting probability vector, its one-to-one correspondence with the candidate anomaly feature vectors is maintained, and it is written into the cutting probability vector set according to the arrangement order in the candidate anomaly feature vector set.

[0027] In this embodiment, the generation of the comprehensive anomaly score and key isolation dimensions includes: The candidate anomaly feature vector and its corresponding cutting probability vector are input into the robust random forest algorithm, so that each candidate anomaly feature vector passes through the cutting nodes in each random tree in turn and reaches the leaf node. The number of cutting nodes passed by the candidate anomaly feature vector in each random tree, the monitoring dimension adopted by each cutting node and the cutting probability corresponding to the monitoring dimension are recorded to generate the isolation depth set and isolation path record. Read each isolation depth in the isolation depth set sequentially, generate a corresponding single-tree anomaly score for each isolation depth, and form a single-tree anomaly score set according to the random tree arrangement. Calculate the average of all single-tree anomaly scores in the single-tree anomaly score set to generate a comprehensive anomaly score for the corresponding candidate anomaly feature vector. ; in, For the first The combined anomaly score of the candidate anomaly feature vectors The number of random trees, For the first The candidate abnormal feature vectors at the th ... Isolation depth in a random tree; The isolation path records are read sequentially, and the monitoring dimensions used for each cutting node are classified according to voltage, current, temperature, vibration and rotation speed. The cutting probability corresponding to the same monitoring dimension is accumulated to generate voltage isolation contribution, current isolation contribution, temperature isolation contribution, vibration isolation contribution and rotation speed isolation contribution. The monitoring dimension corresponding to the isolation contribution with the largest value is determined as the key isolation dimension. For the same candidate anomaly feature vector, the isolation contributions of voltage, current, temperature, vibration, and rotational speed are initially set to zero. All isolation path records are read sequentially. For each cutting node read, the cutting probability corresponding to that node is accumulated into the isolation contribution of the adopted monitoring dimension. After all records are read, the five isolation contributions are compared, and the monitoring dimension corresponding to the maximum value is determined as the key isolation dimension. If there are identical maximum values, the monitoring dimension is determined in the order of voltage, current, temperature, vibration, and rotational speed. The state feature vectors of each normal operating state in the reference sample set are used to generate corresponding cutting probability vectors. These vectors are then input into the same robust random forest algorithm. Based on the method of generating isolation depth set and comprehensive anomaly score, comprehensive anomaly scores corresponding to the state feature vectors of each normal operating state are generated to form a reference comprehensive anomaly score set. Calculate the median and median absolute deviation of the reference comprehensive abnormal score set, and determine the abnormal score limit by the sum of the median and the product of the preset abnormal score coefficient and the median absolute deviation. The preset outlier coefficient is a fixed value greater than zero that is set before calculating the outlier limit. It is used to determine the multiple of the median absolute deviation within the outlier limit and remains unchanged during the calculation of the same reference aggregate outlier set. In Example 1, the preset outlier coefficient is 3, that is, the outlier limit is the sum of the median of the aggregate outlier and three times the median absolute deviation. In Example 1, for each state feature vector, the number of cutting nodes it passes through from the root node to the leaf node in each random tree is recorded, and this number is used as the isolation depth. The reciprocal of the isolation depth plus one is used as the corresponding single-tree anomaly score, and the average of all single-tree anomaly scores is calculated as the comprehensive anomaly score. For example, when the isolation depths of a certain state feature vector in three random trees are 2, 4, and 5 respectively, the single-tree anomaly scores are 0.333, 0.200, and 0.167 respectively, and the comprehensive anomaly score is 0.233. At the same time, the monitoring dimension and its cutting probability used for each cutting node in each isolation path are recorded, and the cutting probabilities corresponding to voltage, current, temperature, vibration, and rotation speed are accumulated respectively. The monitoring dimension with the largest accumulated value is determined as the key isolation dimension. The same method is used to calculate the comprehensive anomaly score for the state feature vectors of each normal operating state in the reference sample set, forming a reference comprehensive anomaly score set; for example, when the set is [0.10, 0.12, 0.13, 0.15, 0.16], the median is 0.13, the median absolute deviation is 0.02, and when the preset anomaly score coefficient is 3, the anomaly score limit is 0.19.

[0028] In this embodiment, the update module includes: Read each state feature vector sequentially, extract the matrix contour distance from each state feature vector, and obtain the comprehensive anomaly score, so that the matrix contour distance and the comprehensive anomaly score are respectively associated with the corresponding state feature vector; Compare the matrix contour distance corresponding to each state feature vector with the screening limit, and extract the state feature vectors whose matrix contour distance is less than or equal to the screening limit from each state feature vector. The comprehensive anomaly score corresponding to each extracted state feature vector is compared with the anomaly score limit, and state feature vectors with comprehensive anomaly scores less than or equal to the anomaly score limit are extracted again from each extracted state feature vector. Each extracted state feature vector is sequentially written into the reference sample set, so that each newly added item in the reference sample set corresponds to a state feature vector, and the correspondence between the matrix contour distance, start position, end position, duration length and abnormal contribution vector in the state feature vector remains unchanged. State feature vectors whose matrix contour distance is greater than the screening limit or whose comprehensive anomaly score is greater than the anomaly score limit are not written into the reference sample set. In Example 1, following the one-to-one correspondence between state feature vectors and comprehensive anomaly scores, the matrix contour distance and its corresponding comprehensive anomaly score in each state feature vector are read sequentially. First, it is determined whether the matrix contour distance is less than or equal to the screening threshold, and then it is determined whether the comprehensive anomaly score is less than or equal to the anomaly score threshold. Only when both determinations are true simultaneously is the state feature vector completely written into the reference sample set, while maintaining the arrangement of the matrix contour distance, start position, end position, duration, and anomaly contribution vector. For example, when the screening threshold is 1.65 and the anomaly score threshold is 0.19, the state feature vector with a matrix contour distance of 1.52 and a comprehensive anomaly score of 0.17 is written into the reference sample set; if any value exceeds the corresponding threshold, it is not written.

[0029] In this embodiment, the determination of persistent fault in the determination module includes: Each candidate anomaly feature vector is read sequentially, the duration length is extracted from each candidate anomaly feature vector, and the comprehensive anomaly score is read, so that the duration length, the comprehensive anomaly score and the corresponding candidate anomaly feature vector are kept correlated. The duration length corresponding to each candidate abnormal feature vector is compared with the preset duration length. When the duration length is greater than or equal to the preset duration length, the comprehensive abnormal score of the corresponding candidate abnormal feature vector is read. Specifically, the preset duration length refers to the minimum abnormal duration time pre-set before the continuous fault determination, and remains unchanged during the same determination process. The duration length in the candidate abnormal feature vector is the time difference between the start and end sampling times of the corresponding abnormal subsequence. The preset duration length is set to 5 seconds, and the duration determination condition is met when the duration length reaches 5 seconds. The overall anomaly score is compared with the anomaly score limit. When the overall anomaly score is greater than the anomaly score limit, the corresponding candidate anomaly feature vector is determined to be a persistent fault, and the start position, end position, duration, anomaly contribution vector, overall anomaly score and key isolation dimension in the candidate anomaly feature vector are retained. When the duration of a candidate abnormal feature vector is less than the preset duration, or when the duration is greater than or equal to the preset duration and the comprehensive abnormal score is less than or equal to the abnormal score limit, the corresponding candidate abnormal feature vector will not be judged as a persistent fault. In Example 1, according to the correspondence between candidate anomaly feature vectors and comprehensive anomaly scores, the duration and corresponding comprehensive anomaly score of each candidate anomaly feature vector are read sequentially. First, the duration is compared with a preset duration, and then the comprehensive anomaly score is compared with anomaly score limits. Only when the duration is greater than or equal to the preset duration and the comprehensive anomaly score is greater than the anomaly score limit is the corresponding candidate anomaly feature vector determined to represent a persistent fault, and its start position, end position, duration, anomaly contribution vector, comprehensive anomaly score, and key isolation dimension are retained. For example, when the preset duration is 5 seconds and the anomaly score limit is 0.19, a candidate anomaly feature vector with a duration of 6 seconds and a comprehensive anomaly score of 0.27 is determined to be a persistent fault; if the duration is less than 5 seconds or the comprehensive anomaly score is not greater than 0.19, it is not determined to be a persistent fault.

[0030] In this embodiment, the determination module further includes: Read the candidate anomaly feature vectors that are determined to be persistent faults in sequence, extract the anomaly contribution vector from each candidate anomaly feature vector, and read the key isolation dimension to keep the anomaly contribution vector, the key isolation dimension and the corresponding candidate anomaly feature vectors associated. Following the order of voltage, current, temperature, vibration, and rotational speed, each abnormal contribution degree in the abnormal contribution degree vector is read sequentially, the values ​​of each abnormal contribution degree are compared, and the monitoring dimension corresponding to the abnormal contribution degree with the largest value is determined as the main abnormal dimension. When the anomaly contribution of multiple monitoring dimensions is at its maximum value, the monitoring dimension that is the same as the key isolation dimension is determined as the main anomaly dimension. If they are all different from the key isolation dimension, the monitoring dimension that comes first is determined in the order of voltage, current, temperature, vibration and rotation speed. Specifically, the five components in the anomaly contribution vector are compared one by one in the order of voltage, current, temperature, vibration and rotation speed. The monitoring dimension corresponding to the largest component is determined as the main anomaly dimension. When there are multiple identical largest components, the monitoring dimension that is the same as the key isolation dimension is selected first. If there is no identical dimension, the monitoring dimension that appears first in the order is selected. The main anomaly dimension is compared with the key isolation dimension. When the two are the same, the fault corresponding to the monitoring dimension is determined as the fault category. When the two are different, the fault corresponding to the main anomaly dimension and the fault corresponding to the key isolation dimension are jointly determined as the fault category. Among them, voltage, current, temperature, vibration and speed correspond to voltage fault, current fault, temperature fault, vibration fault and speed fault, respectively. Specifically, a consistency comparison is performed between the primary anomaly dimension and the key isolation dimension; when the two are the same, a single fault category is determined according to this monitoring dimension, for example, when both are vibration dimensions, it is determined as a vibration fault; when the two are different, the two monitoring dimensions are respectively mapped to corresponding faults and combined to form a composite fault category, for example, when the primary anomaly dimension is temperature and the key isolation dimension is rotational speed, it is determined as a temperature-rotational speed composite fault; The determination result of the persistent fault, the determined fault category, the anomaly contribution vector, the key isolation dimension and the comprehensive anomaly score are associated with the candidate anomaly feature vector to generate a generator set state monitoring result; In Embodiment 1, the anomaly contribution vector is read in the order of voltage, current, temperature, vibration and rotational speed, and the monitoring dimension corresponding to the maximum anomaly contribution is determined as the primary anomaly dimension; when there are multiple identical maximum values, the monitoring dimension that is the same as the key isolation dimension is preferentially selected, and when there is no identical dimension, the monitoring dimension arranged earlier is selected. The primary anomaly dimension is compared with the key isolation dimension, when the two are the same, a corresponding single fault category is generated, and when the two are different, a composite fault category is generated in the order of the primary anomaly dimension and the key isolation dimension. For example, when the anomaly contribution vector is [0.10, 0.12, 0.18, 0.45, 0.15] and the key isolation dimension is vibration, the primary anomaly dimension is vibration, and the fault category is a vibration fault; when the primary anomaly dimension is temperature and the key isolation dimension is rotational speed, the fault category is a temperature-rotational speed composite fault. The persistent fault determination result, the fault category, the anomaly contribution vector, the key isolation dimension and the comprehensive anomaly score are associated with the corresponding candidate anomaly feature vector to generate the generator set state monitoring result.

[0031] Embodiment 2:

[0032] To verify the feasibility of the present invention in implementation, the present invention is applied to continuous operation state monitoring of the generator set supporting measurement while drilling equipment. During operation, the generator set is jointly affected by load changes, mechanical vibration, temperature rise and rotational speed fluctuation, and there are correlated changes among voltage, current, temperature, vibration and rotational speed. For example, when a mechanical component becomes loose, vibration data usually changes first, and then rotational speed, current and temperature data generate responses in sequence; when the load increases for a short time, current and voltage may have instantaneous fluctuations, but no persistent fault is formed in the generator set. If monitoring is performed only by means of a single parameter exceeding a fixed limit, short-term impacts and normal load regulation are easily identified as faults, and it is also difficult to determine whether the anomaly mainly comes from voltage, current, temperature, vibration or rotational speed.

[0033] In this embodiment, voltage, current, temperature, vibration, and speed data of the generator set are collected. The unit for voltage data is V, the unit for current data is A, the unit for temperature data is °C, and the vibration data is expressed as the effective value of vibration acceleration in m / s². 2 The unit of rotational speed data is r / min. The sampling interval for all types of data was set to 0.1s, and continuous data collection was carried out for 120 minutes, resulting in 72,000 uniform sampling moments. Each uniform sampling moment corresponds to one voltage measurement, one current measurement, one temperature measurement, one vibration measurement, and one rotational speed measurement, for a total of 360,000 measurement values.

[0034] During the data acquisition process, sampling time identifiers were assigned to each measurement value, and a unified sampling time sequence was generated at 0.1-second intervals, starting from the earliest sampling time and ending at the latest sampling time. Due to variations in data transmission intervals, among the 72,000 unified sampling times, 186 sampling times lacked voltage measurements, 142 lacked current measurements, 97 lacked temperature measurements, 213 lacked vibration measurements, and 128 lacked rotational speed measurements. For sampling times lacking measurements, the closest similar measurements before and after that sampling time were read, and linear interpolation was performed according to the time position of the sampling time lacking measurements between adjacent sampling times to generate time-aligned multidimensional state data.

[0035] A window length of 200 consecutive uniform sampling times was used, and a window movement step of 20 uniform sampling times was used to perform windowing processing on the time-aligned multidimensional state data. Each multidimensional state subsequence corresponds to 20 seconds of running data, with a 2-second shift between adjacent multidimensional state subsequences. This generated 3591 multidimensional state subsequences. During normal operation, the voltage mainly ranged from 47.6V to 48.5V, the current mainly ranged from 9.2A to 13.8A, the temperature mainly ranged from 51.4℃ to 63.7℃, and the effective value of vibration acceleration mainly ranged from 2.0 m / s². 2 Up to 4.1 m / s 2 The rotational speeds are mainly distributed between 2968 r / min and 3036 r / min.

[0036] From each multidimensional state subsequence, voltage, current, temperature, vibration, and rotational speed data are simultaneously extracted according to a set analysis length of 50 consecutive uniform sampling times to generate local multidimensional state subsequences. Each local multidimensional state subsequence corresponds to 5 seconds of operational data. Normalization is performed on each type of data to eliminate differences in numerical range and units for voltage, current, temperature, vibration, and rotational speed. Each normalized local multidimensional state subsequence is used as a query subsequence, and the remaining normalized local multidimensional state subsequences without overlapping sampling positions are used as comparison subsequences.

[0037] The comparison subsequence is allowed to move forward or backward by no more than 5 uniform sampling times relative to the query subsequence, with a corresponding maximum time delay of 0.5 s. For each movement position, data that can correspond one-to-one in the same sampling order are extracted from both the query and comparison subsequences. The one-dimensional distances for voltage, current, temperature, vibration, and rotational speed are calculated for each data type at different movement positions. The minimum value is selected from these one-dimensional distances to form a time delay matching distance vector and a target time delay vector. This allows different monitoring parameters to be matched within a range not exceeding 0.5 s.

[0038] In a normal load increase data segment, the current data is delayed by 0.2 seconds relative to the voltage data, and the temperature data is delayed by 0.4 seconds relative to the current data. Without time-delay matching, the multidimensional distance of this data segment is 1.47. After time-delay matching, the target time delays for voltage, current, temperature, vibration, and rotational speed are 0, 2, 4, 1, and 1 unified sampling times, respectively, resulting in a time-delay constrained multidimensional distance of 0.92. This data segment does not exceed the contour distance boundary and is therefore not identified as an anomaly. This avoids large direct distances caused by asynchronous responses of various monitoring parameters.

[0039] For each query subsequence, the minimum value among the time-delay constrained multidimensional distances corresponding to it and all comparison subsequences is selected as the first matrix contour distance, and a first matrix contour distance sequence is generated according to the sampling position of the query subsequence. Subsequently, one type of data from voltage, current, temperature, vibration, and rotational speed is removed respectively, and the counterfactual contour distance is recalculated using the remaining four types of data each time. The differences less than 0 obtained by subtracting each counterfactual contour distance from the first matrix contour distance are set to 0, and the remaining differences are normalized to generate an initial anomaly contribution vector.

[0040] In a segment of mechanical vibration anomaly data, the first matrix profile distance is 2.36. After removing voltage, current, temperature, vibration, and rotational speed, the resulting counterfactual profile distances are 2.21, 2.18, 2.05, 1.42, and 1.98, respectively. Normalization is performed based on the difference between the first matrix profile distance and the counterfactual profile distance, resulting in an initial anomaly contribution vector of [0.08, 0.10, 0.16, 0.48, 0.18]. Arranging the anomaly contributions from largest to smallest and accumulating them sequentially, when the contribution ratio is set to 0.8, the cumulative value of the corresponding components for vibration, rotational speed, and temperature reaches 0.82. Therefore, vibration, rotational speed, and temperature are determined as contribution data types.

[0041] Vibration, rotational speed, and temperature data are extracted from the normalized local multidimensional state subsequence. The time-delay matching distance vector, target time-delay vector, and time-delay constraint multidimensional distance are recalculated to generate a second matrix contour distance sequence. In normal operating data, the median of the first matrix contour distance sequence is 0.86, the median absolute deviation is 0.11, the preset coefficient is 3, and the first contour distance boundary is 1.19. The median of the second matrix contour distance sequence is 0.73, the median absolute deviation is 0.09, and the second contour distance boundary is 1.00. Sampling positions in both matrix contour distance sequences that continuously exceed the corresponding contour distance boundaries are merged to obtain an abnormal subsequence.

[0042] In the aforementioned abnormal mechanical vibration data, the first matrix profile distance continuously exceeded 1.19, and the second matrix profile distance continuously exceeded 1.00. The starting position corresponding to the continuous abnormal positions was 32640, and the ending position was 32718. Based on a sampling interval of 0.1s, the duration of this abnormal subsequence was 7.9s. During the abnormal period, the voltage decreased from 48.1V to 47.3V, the current increased from 11.6A to 14.9A, the temperature increased from 58.7℃ to 62.4℃, and the effective value of the vibration acceleration decreased from 3.2m / s². 2 Rising to 10.8 m / s 2 The rotational speed decreased from 3012 r / min to 2946 r / min. The initial abnormal contribution degree corresponding to the contribution data type was renormalized to obtain the abnormal contribution degree vector [0, 0, 0.20, 0.59, 0.21].

[0043] The state feature vector is constructed by arranging the matrix contour distance, start position, end position, duration, and anomaly contribution vector corresponding to the abnormal subsequence in a fixed order. In this embodiment, the state feature vector is represented as [2.36, 32640, 32718, 7.9, 0, 0, 0.20, 0.59, 0.21]. State feature vectors are extracted from sampling periods confirmed to be in normal operating condition to establish a reference sample set. The initial reference sample set contains 180 state feature vectors for normal operating conditions.

[0044] Matrix contour distances were extracted from 180 state feature vectors in the reference sample set. The median of the resulting reference matrix contour distance set was 1.08, and the median absolute deviation was 0.14. With a preset screening coefficient of 3, the calculated screening threshold was 1.50. The matrix contour distance corresponding to the aforementioned mechanical vibration anomaly was 2.36, which is greater than the screening threshold of 1.50. Therefore, the corresponding state feature vector was identified as a candidate anomaly feature vector.

[0045] The anomaly contribution vector in the candidate anomaly feature vector is converted into cutting probabilities corresponding to voltage, current, temperature, vibration, and rotational speed. Since the sum of all components of the anomaly contribution vector is 1, the cutting probability vector is [0, 0, 0.20, 0.59, 0.21]. In the robust random forest algorithm, 100 random trees are set up, and the candidate anomaly feature vector passes through the cutting nodes of each random tree sequentially according to the stated cutting probabilities. The isolation depth of the candidate anomaly feature vector in each random tree and the monitoring dimension used at each cutting node are recorded.

[0046] The isolation depths of the candidate anomaly feature vector across the 100 random trees are primarily distributed between 2 and 5. Specifically, there are 31 random trees with an isolation depth of 2, 38 with an isolation depth of 3, 22 with an isolation depth of 4, and 9 with an isolation depth of 5. The anomaly score for each tree is calculated by adding 1 to the reciprocal of each isolation depth, and then the average of the 100 individual tree anomaly scores is calculated, resulting in a comprehensive anomaly score of 0.274.

[0047] The same method was used to calculate the comprehensive anomaly score for the state feature vectors of the normal operating conditions in the reference sample set. The median of the resulting reference comprehensive anomaly score set was 0.167, and the median absolute deviation was 0.018. With a preset anomaly score coefficient of 3, the calculated anomaly score threshold was 0.221. The comprehensive anomaly score of 0.274 corresponding to the aforementioned mechanical vibration anomaly is greater than the anomaly score threshold of 0.221.

[0048] In the isolation paths of 100 random trees, the cutting probabilities of cutting nodes corresponding to voltage, current, temperature, vibration, and rotational speed are accumulated respectively. The resulting isolation contributions are 0, 0, 18.6, 54.7, and 20.3, respectively. Among them, the isolation contribution corresponding to the vibration dimension is the largest, so vibration is identified as the key isolation dimension. The preset duration is set to 5 seconds. The duration corresponding to the above candidate abnormal feature vector is 7.9 seconds, which is greater than the preset duration of 5 seconds, and the comprehensive abnormal score of 0.274 is greater than the abnormal score threshold of 0.221. Therefore, this candidate abnormal feature vector is judged as a continuous fault. The abnormal contribution degree corresponding to vibration in the abnormal contribution degree vector is the maximum value of 0.59, so the main abnormal dimension is vibration. Since both the main abnormal dimension and the key isolation dimension are vibration, the fault category is determined to be vibration fault. The final generator set condition monitoring results include the continuous fault determination result, vibration fault category, starting position 32640, ending position 32718, duration 7.9s, comprehensive anomaly score 0.274, anomaly contribution vector [0, 0, 0.20, 0.59, 0.21], and key isolation dimension vibration.

[0049] In another segment of anomalous data, the duration of the anomalous subsequence was 8.6 s, the anomalous contribution vector was [0.06, 0.09, 0.43, 0.18, 0.24], and the overall anomalous score was 0.258. The anomalous contribution score corresponding to temperature was the highest at 0.43, therefore temperature was the primary anomalous dimension; the key isolation dimension generated by the robust random forest algorithm was rotational speed. Since the primary anomalous dimension and the key isolation dimension were different, the fault category was determined to be a temperature-rotational speed composite fault. During this anomalous period, the temperature increased from 60.2℃ to 70.6℃, the rotational speed decreased from 3004 r / min to 2918 r / min, and the current increased from 12.1A to 15.3A, consistent with the state characteristics of both temperature and rotational speed being anomalous.

[0050] The continuous operation data also showed four instances of short-term vibration increases caused by mechanical shocks, with the effective values ​​of the four abnormal vibration accelerations reaching 8.7 m / s². 2 9.3 m / s 2 10.1 m / s 2 and 8.9m / s 2 The corresponding durations are 1.2s, 1.8s, 2.1s, and 1.5s, respectively. The matrix contour distances of the four state feature vectors all exceed the screening threshold, and the comprehensive anomaly scores are 0.241, 0.253, 0.266, and 0.248, respectively, all exceeding the anomaly score threshold. However, the durations are all less than the preset duration of 5s, so none of them are judged as continuous faults, thus distinguishing short-term mechanical impacts from continuous faults.

[0051] During the reference sample update process, a total of 67 newly generated state feature vectors participated in the update judgment. Of these, 42 state feature vectors had a matrix contour distance less than or equal to the screening threshold of 1.50 and a comprehensive anomaly score less than or equal to the anomaly score threshold of 0.221, and were therefore added to the reference sample set; 11 state feature vectors had a matrix contour distance greater than 1.50 and were not added to the reference sample set; 14 state feature vectors had a matrix contour distance less than or equal to 1.50, but a comprehensive anomaly score greater than 0.221, and were also not added to the reference sample set. After the update, the reference sample set increased from 180 state feature vectors to 222 state feature vectors.

[0052] When comparing the same operating data, using methods with fixed limits for vibration, temperature, current, voltage, and speed alone generated 18 anomaly alerts, of which 9 were continuous faults and 9 were anomalies caused by load regulation or short-term impacts. Using the system of this embodiment, 9 candidate anomalies were identified, of which 5 met the duration and comprehensive anomaly score conditions and were judged as continuous faults, while 4 were not judged as continuous faults because their duration was less than 5 seconds. Upon verification, the 5 continuous faults included 3 vibration faults, 1 temperature fault, and 1 temperature-speed composite fault. The fault categories were consistent with the actual changes in the corresponding voltage, current, temperature, vibration, and speed data. Without updating the reference sample set, 7 state feature vectors in subsequent normal load change data exceeded the original judgment conditions. After using the updated reference sample set, the matrix contour distance and comprehensive anomaly score of 5 state feature vectors were within the updated normal range, and only 2 state feature vectors entered the subsequent anomaly judgment. Therefore, in the operating data described in this embodiment, updating the reference sample allows the judgment criteria to adapt to changes in the normal operating state of the generator set.

[0053] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A generator set intelligent condition monitoring system based on downhole measurement-while-drilling, characterized in that, include: The data acquisition and processing module collects voltage, current, temperature, vibration and speed data of the generator set, performs alignment and windowing processing, and generates multi-dimensional state subsequences. The matrix profile module performs multidimensional matrix profile analysis on multidimensional state subsequences to obtain matrix profile distance, start and end positions of anomalous subsequences, duration length, and anomalous contribution vector. The feature reconstruction module constructs a state feature vector from the matrix contour distance, the start and end positions and duration of the abnormal subsequence, and the abnormal contribution vector, and establishes a reference sample set based on the state feature vector pre-calibrated as the normal operating state. The anomaly screening module determines the screening limit based on the matrix contour distance in the reference sample set, and identifies state feature vectors whose matrix contour distance exceeds the screening limit as candidate anomaly feature vectors. The forest module inputs the state feature vector into the robust random forest algorithm, normalizes each component of the anomaly contribution vector into the cutting probability of the corresponding feature dimension, outputs the comprehensive anomaly score and key isolation dimension, and determines the anomaly score boundary based on the comprehensive anomaly score. The update module writes the state feature vectors that do not exceed the screening limit in terms of matrix contour distance and comprehensive anomaly score not exceeding the anomaly score limit into the reference sample set; The judgment module determines a continuous fault when the duration of the candidate abnormal feature vector reaches the preset duration and the comprehensive abnormal score exceeds the abnormal score limit. It determines the fault category based on the abnormal contribution vector and the key isolation dimension, and generates generator set status monitoring results.

2. The intelligent status monitoring system for generator sets based on downhole measurement-while-drilling as described in claim 1, characterized in that, The acquisition and processing module includes: The generator set collects voltage, current, temperature, vibration, and speed data. The voltage data is the voltage measurement value corresponding to each sampling time of the generator set. The current data is the current measurement value corresponding to each sampling time of the generator set. The temperature data is the temperature measurement value corresponding to each sampling time of the generator set. The vibration data is the vibration measurement value corresponding to each sampling time of the generator set. The speed data is the speed measurement value corresponding to each sampling time of the generator set. Configure sampling time identifiers for voltage data, current data, temperature data, vibration data, and rotational speed data respectively, and determine a unified sampling time sequence based on each sampling time identifier; Voltage data, current data, temperature data, vibration data, and rotational speed data are mapped to a unified sampling time sequence according to their corresponding sampling time identifiers. For sampling times in the unified sampling time sequence that lack corresponding measurement values, interpolation is performed using the same type of measurement values ​​corresponding to adjacent sampling times to generate time-aligned multidimensional state data. According to the set window length, voltage data, current data, temperature data, vibration data and rotation speed data corresponding to continuous and uniform sampling times are extracted from the time-aligned multidimensional state data. The extraction range is moved sequentially along the uniform sampling time sequence according to the set window movement step size. The data in each extraction range are arranged according to the sampling time to generate a multidimensional state subsequence.

3. The intelligent status monitoring system for generator sets based on downhole measurement-while-drilling as described in claim 2, characterized in that, The matrix contour module includes: According to the set subsequence length, voltage data, current data, temperature data, vibration data and rotation speed data are synchronously extracted from each sampling position of the multidimensional state subsequence to generate a local multidimensional state subsequence. The data of each monitoring dimension are then normalized to generate a normalized local multidimensional state subsequence. Each normalized local multidimensional state subsequence is used as a query subsequence in turn, and the remaining normalized local multidimensional state subsequences that do not overlap with the query subsequences in sampling position are used as comparison subsequences to generate a set of comparison subsequences. For each monitoring dimension of each comparison subsequence in the query subsequence and comparison subsequence sets, the comparison subsequences are moved sequentially according to the unified sampling time within a set time lag range. The effective overlapping data segments corresponding to the time positions in the query subsequence and the moved comparison subsequence are extracted, and the one-dimensional distance is calculated. The minimum value is selected from the single-dimensional distances corresponding to each monitoring dimension under each time delay as the time delay matching distance. The time delay that generates the time delay matching distance is determined as the target time delay, and the time delay matching distance vector and the target time delay vector are generated. The time delay constraint multidimensional distance is calculated based on the difference between each target time delay and the median of the target time delay in the target time delay vector, and the time delay matching distance vector. The minimum value is selected from all time-delay constraint multidimensional distances corresponding to each query subsequence as the first matrix contour distance. The position of the comparison subsequence that produces the minimum value is determined as the first matching position. The first matrix contour distances are arranged according to the sampling positions of the query subsequences to generate the first matrix contour distance sequence. One of the monitoring dimensions, voltage, current, temperature, vibration and rotation speed, is removed in turn. After each removal, the corresponding set of comparison subsequences is retraced based on the remaining monitoring dimensions. The time delay matching distance vector and the target time delay vector are regenerated, and the time delay constraint multidimensional distance is recalculated. The smallest time delay constraint multidimensional distance obtained by recalculation is determined as the counterfactual profile distance corresponding to the removed monitoring dimension. Calculate the difference between the first matrix contour distance and the distances of each counterfactual contour, set the differences that are less than zero to zero, and normalize the differences after setting them to zero to generate an initial anomaly contribution vector. The corresponding monitoring dimensions are arranged from largest to smallest according to the initial abnormal contribution. Starting from the largest abnormal contribution, they are added sequentially. When the cumulative abnormal contribution first reaches the set ratio, the accumulated monitoring dimensions are determined as the set of contribution dimensions. Extract the data corresponding to the contribution dimension set from the normalized local multidimensional state subsequence, generate the contribution-focused state subsequence, and regenerate the query subsequence, comparison subsequence set, time delay matching distance vector, target time delay vector and time delay constraint multidimensional distance based on the contribution-focused state subsequence. Select the minimum value from the time delay constraint multidimensional distances corresponding to each query subsequence to generate the second matrix contour distance sequence. The first and second contour distance boundaries are determined based on the first and second matrix contour distance sequences, respectively. Sampling positions where both the first and second matrix contour distances continuously exceed the corresponding contour distance boundaries are merged into anomaly subsequences. The start and end positions and duration of the anomaly subsequences are obtained based on the first and last sampling positions and the corresponding unified sampling time. The components corresponding to the contribution dimension set in the initial anomaly contribution vector are renormalized to generate anomaly contribution vector.

4. The intelligent condition monitoring system for generator sets based on downhole measurement-while-drilling as described in claim 3, characterized in that, The feature reconstruction module includes: Obtain the matrix contour distance, start position, end position, duration, and anomaly contribution vector corresponding to each abnormal subsequence, and aggregate the matrix contour distance, start position, end position, duration, and anomaly contribution vector belonging to the same abnormal subsequence; Following the monitoring dimensions of voltage, current, temperature, vibration, and rotational speed, the components in the anomaly contribution vector are read sequentially to obtain the voltage anomaly contribution, current anomaly contribution, temperature anomaly contribution, vibration anomaly contribution, and rotational speed anomaly contribution. According to the order of matrix contour distance, starting position, ending position, duration, voltage anomaly contribution, current anomaly contribution, temperature anomaly contribution, vibration anomaly contribution, and rotational speed anomaly contribution, the data corresponding to the same anomaly subsequence are arranged sequentially to generate the state feature vector corresponding to the anomaly subsequence. Extract the state feature vectors that are pre-labeled as normal operating states from the state feature vectors, and arrange them in ascending order of their starting positions; Each pre-labeled state feature vector of the sorted system is used as a reference sample to establish a reference sample set.

5. The intelligent status monitoring system for generator sets based on downhole measurement-while-drilling as described in claim 1, characterized in that, The anomaly filtering module includes: Read the state feature vectors of each normal operating state in the reference sample set in sequence, extract the matrix contour distance from the state feature vectors of each normal operating state, and arrange the extracted matrix contour distances according to the arrangement order of the state feature vectors of each normal operating state in the reference sample set to generate a reference matrix contour distance set. Calculate the median of all matrix contour distances in the reference matrix contour distance set, calculate the absolute difference between each matrix contour distance and the median, and calculate the median of all absolute differences to generate the median absolute deviation of the matrix contour distance. The median of the matrix contour distance and the absolute deviation of the median matrix contour distance are used to determine the screening threshold; The state feature vectors that are not written into the reference sample set are used to form a set of state feature vectors to be screened. The matrix contour distance is extracted from each state feature vector in the set of state feature vectors to be screened in turn, and the extracted matrix contour distance is established with the state feature vector to generate a matrix contour distance record to be screened. The matrix contour distances in each matrix contour distance record to be screened are compared with the screening limit. The matrix contour distance records to be screened that are greater than the screening limit are extracted. The state feature vectors to which each matrix contour distance record to be screened belongs are obtained according to the correspondence. The obtained state feature vectors are determined as candidate anomaly feature vectors. The candidate anomaly feature vectors are arranged in ascending order of their starting positions to generate a set of candidate anomaly feature vectors.

6. The intelligent condition monitoring system for generator sets based on downhole measurement-while-drilling as described in claim 5, characterized in that, The determination of the cut probability driven by the abnormal contribution in the forest module includes: Read each candidate anomaly feature vector in the candidate anomaly feature vector set in sequence, extract the anomaly contribution vector from each candidate anomaly feature vector, and read each component in the anomaly contribution vector in the order of voltage, current, temperature, vibration and rotation speed to generate an anomaly contribution component sequence corresponding to the candidate anomaly feature vector. The abnormal contribution values ​​of voltage, current, temperature, vibration, and rotational speed in the abnormal contribution component sequence are summed to generate the total abnormal contribution value of the corresponding candidate abnormal feature vectors. Divide each abnormal contribution in the abnormal contribution component sequence by the total abnormal contribution value to obtain the segmentation probability of the corresponding monitoring dimension. When the total abnormal contribution is zero, the cutting probability corresponding to voltage, current, temperature, vibration and rotation speed is set to 0.2 respectively; When the total value of abnormal contribution is greater than zero, the normalized cut probability is used. Based on the determined cutting probabilities arranged in the order of voltage, current, temperature, vibration, and rotational speed, a cutting probability vector corresponding to the corresponding candidate abnormal feature vector is generated. The cutting probability vectors are then arranged in the order of their arrangement in the candidate abnormal feature vector set to generate a cutting probability vector set.

7. The intelligent condition monitoring system for generator sets based on downhole measurement-while-drilling as described in claim 6, characterized in that, The generation of the comprehensive anomaly score and key isolation dimensions includes: The candidate anomaly feature vector and its corresponding cutting probability vector are input into the robust random forest algorithm. The number of cutting nodes that the candidate anomaly feature vector passes through in each random tree, the monitoring dimension used for each cutting node, and the cutting probability corresponding to the monitoring dimension are recorded to generate an isolation depth set and an isolation path record. Read each isolation depth in the isolation depth set in sequence, generate the corresponding single-tree anomaly score according to each isolation depth, and form a single-tree anomaly score set according to the arrangement order of the random trees. Calculate the average value of all single-tree anomaly scores in the single-tree anomaly score set to generate the comprehensive anomaly score of the corresponding candidate anomaly feature vector. The isolation path records are read sequentially, and the monitoring dimensions used for each cutting node are classified according to voltage, current, temperature, vibration and rotation speed. The cutting probability corresponding to the same monitoring dimension is accumulated to generate voltage isolation contribution, current isolation contribution, temperature isolation contribution, vibration isolation contribution and rotation speed isolation contribution. The monitoring dimension corresponding to the isolation contribution with the largest value is determined as the key isolation dimension. The state feature vectors of each normal operating state in the reference sample set are used to generate corresponding cutting probability vectors. These vectors are then input into the same robust random forest algorithm. Based on the method of generating isolation depth set and comprehensive anomaly score, comprehensive anomaly scores corresponding to the state feature vectors of each normal operating state are generated to form a reference comprehensive anomaly score set. Calculate the median and median absolute deviation of the reference comprehensive abnormal score set, and determine the abnormal score limit by summing the median with the product of the preset abnormal score coefficient and the median absolute deviation.

8. The intelligent condition monitoring system for generator sets based on downhole measurement-while-drilling as described in claim 7, characterized in that, The update module includes: Read the feature vectors of each state in sequence, extract the matrix contour distance from each feature vector, and obtain the comprehensive anomaly score; Compare the matrix contour distance corresponding to each state feature vector with the screening limit, and extract the state feature vectors whose matrix contour distance is less than or equal to the screening limit from each state feature vector. The comprehensive anomaly score corresponding to each extracted state feature vector is compared with the anomaly score limit, and state feature vectors with comprehensive anomaly scores less than or equal to the anomaly score limit are extracted again from each extracted state feature vector. The extracted state feature vectors are then sequentially written into the reference sample set. State feature vectors whose matrix contour distance is greater than the screening limit or whose comprehensive anomaly score is greater than the anomaly score limit are not written into the reference sample set.

9. The intelligent condition monitoring system for generator sets based on downhole measurement-while-drilling as described in claim 1, characterized in that, The determination of persistent faults in the determination module includes: Read each candidate anomaly feature vector sequentially, extract the duration length from each candidate anomaly feature vector, and read the comprehensive anomaly score; The duration length corresponding to each candidate abnormal feature vector is compared with the preset duration length. When the duration length is greater than or equal to the preset duration length, the comprehensive abnormal score of the corresponding candidate abnormal feature vector is read. The overall anomaly score is compared with the anomaly score limit. When the overall anomaly score is greater than the anomaly score limit, the corresponding candidate anomaly feature vector is determined to be a persistent fault, and the start position, end position, duration, anomaly contribution vector, overall anomaly score and key isolation dimension in the candidate anomaly feature vector are retained. When the duration of a candidate abnormal feature vector is less than the preset duration, or when the duration is greater than or equal to the preset duration and the overall abnormal score is less than or equal to the abnormal score limit, the corresponding candidate abnormal feature vector will not be judged as a persistent fault.

10. The intelligent condition monitoring system for generator sets based on downhole measurement-while-drilling as described in claim 9, characterized in that, The generation of fault categories and generator set status monitoring results in the determination module includes: Read the candidate anomaly feature vectors that are determined to be persistent faults in sequence, extract the anomaly contribution vector from each candidate anomaly feature vector, and read the key isolation dimension. Following the order of voltage, current, temperature, vibration, and rotational speed, each abnormal contribution degree in the abnormal contribution degree vector is read sequentially, the values ​​of each abnormal contribution degree are compared, and the monitoring dimension corresponding to the abnormal contribution degree with the largest value is determined as the main abnormal dimension. When the anomaly contribution of multiple monitoring dimensions is at its maximum value, the monitoring dimension that is the same as the key isolation dimension is determined as the main anomaly dimension. If they are all different from the key isolation dimension, the monitoring dimension that comes first is determined in the order of voltage, current, temperature, vibration and rotation speed. The main anomaly dimension is compared with the key isolation dimension. When the two are the same, the fault corresponding to the monitoring dimension is determined as the fault category. When the two are different, the fault corresponding to the main anomaly dimension and the fault corresponding to the key isolation dimension are jointly determined as the fault category. Among them, voltage, current, temperature, vibration and speed correspond to voltage fault, current fault, temperature fault, vibration fault and speed fault, respectively. The results of persistent fault determination, identified fault category, anomaly contribution vector, key isolation dimension, and comprehensive anomaly score are correlated with candidate anomaly feature vectors to generate generator set condition monitoring results.