A method for detecting the mechanical state of an isolating switch and a split detector

Abstract

The present invention relates to the technical field of disconnector defect detection, and discloses a disconnector mechanical state detection method and split detector. The method performs kernel K-means clustering on the operating torque-rotation angle vector; constructs a reward function that defines reinforcement learning using the silhouette coefficient S, the Davidson-Botting index (DBI), and the fault density index; optimizes the kernel parameter γ and the number of clusters K of kernel K-means clustering through deep reinforcement learning to obtain the optimal clustering result; establishes a mapping relationship between clusters and fault types based on the optimal clustering result, enabling diagnosis from data to fault type; and finally, uses an ELM-Tree model to quantify the fault severity and select treatment measures. The present invention can significantly improve the accuracy of disconnector fault classification and perform quantitative analysis of the fault severity.

Description

Technical Field

[0001] The present invention relates to the technical field of disconnector defect detection, and in particular to a disconnector mechanical state detection method and a split-type detector. Background Art

[0002] Among numerous types of power equipment, disconnectors rank first in both application volume and failure rate among high-voltage power transmission and transformation equipment. Disconnectors are often used in conjunction with circuit breakers to create a clear disconnection and ensure reliable power outages. When closed, they act like conductors, so their dynamic stability and current-carrying capacity are also required. Currently, the failure rate of outdoor disconnectors remains high. The main causes and symptoms of failure can be summarized as follows:

[0003] The connecting rod bearings and other transmission mechanisms of disconnectors are exposed to outdoor environments for extended periods of time, experiencing varying weather conditions. The accumulation of particulate matter such as dust and dirt in the air can lead to oxidation and corrosion, and the lubricant applied during installation can easily dry out. The effects of severe weather conditions such as acid rain, salt spray, and industrial exhaust are even more severe. These damages to the transmission mechanisms are relatively minor and typically manifest themselves as shaking, noise, and incomplete movement during operation.

[0004] Isolating switches are usually installed and debugged on-site. The reference standards have a certain range, and how to debug them to the standard state depends largely on manual experience. There is a lack of effective detection equipment for key parameters such as contact pressure. Therefore, defects such as insufficient clamping force caused by improper debugging are often difficult to detect. Problems such as contact heating only appear after operation, causing serious problems such as contact oxidation and melting.

[0005] Isolating switches do not have the ability to interrupt current, their electrical performance is relatively simple, they are easy to manufacture, and their overall price and profit are not high. Therefore, related R&D investment is insufficient and performance improvement is very slow.

[0006] Disconnectors are widely distributed and numerous, so timely maintenance of operating disconnectors is costly in terms of both manpower and material resources. As electricity loads continue to grow, the proportion of accidents caused by disrepair of disconnectors is increasing, significantly impacting the power grid. If disconnector hazards are not discovered promptly, they can potentially cause widespread power outages.

[0007] Although the proportion of disconnector mechanical failures to the total number of disconnector failures is very high, the research on disconnector mechanical failure diagnosis technology is still incomplete.

[0008] There are three major bottlenecks in the existing disconnector mechanical fault diagnosis technology: first, the traditional decision tree algorithm does not consider the redundancy between attributes when selecting features, resulting in high model complexity; second, K-means clustering is sensitive to noise and has difficulty processing abnormal status data; third, the lack of multi-source sensor data fusion makes it difficult to fully reflect the equipment status. Summary of the Invention

[0009] In order to solve the defects of the prior art, the present invention provides a method for detecting the mechanical state of an isolating switch and a split detector.

[0010] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0011] A method for detecting the mechanical state of an isolating switch comprises the following steps:

[0012] Step 1: Perform data preprocessing and feature normalization on the collected operating torque and rotation angle data to obtain the operating torque-rotation angle vector that matches the operating torque and rotation angle;

[0013] Step 2: Select the kernel function and calculate the kernel matrix for the obtained operating torque-rotation angle vector;

[0014] Step 3: Kernel K-means clustering initialization;

[0015] Step 4: Kernel K-means clustering iterative optimization, by iteratively adjusting the cluster center, minimizes the intra-cluster differences in the high-dimensional space and achieves effective separation of fault samples;

[0016] Step 5: Construct a reward function that defines reinforcement learning using the silhouette coefficient S, the Davidson-Botting index DBI, and the fault density index. Optimize the kernel parameter γ and the number of clusters K of kernel K-means clustering through deep reinforcement learning to obtain the optimal clustering result.

[0017] Step 6: Establish a mapping relationship between clusters and fault types based on the optimal clustering results to achieve diagnosis from data to fault types;

[0018] Step 7: Use the ELM-Tree model (extreme learning machine decision tree model) to quantify the fault severity and select treatment measures.

[0019] Furthermore, the silhouette coefficient S is used to measure the cohesion and separation of sample clusters; the Davidson-Botting index DBI is used to evaluate the separation between clusters and the compactness within clusters; and the fault density index is used to quantify the compactness of the distribution of fault samples within a cluster, which is defined as:

[0020] ;

[0021] in, is the density of the fault sample distribution within the kth cluster, is the kth cluster center, Cluster The number of samples; is the kth cluster The average distance from the inner sample to the cluster center; is the u-th sample;

[0022] The reward function R for deep reinforcement learning is constructed by combining the silhouette coefficient S, the Davidson-Botting index DBI, and the fault density index:

[0023] ;

[0024] Among them, α, β, and η are the weight coefficients of the silhouette coefficient S, the Davidson-Botting index DBI, and the fault density index, respectively.

[0025] Further preferably, in step 1, data preprocessing includes filling missing data by linear interpolation, and ensuring that the operating torque and rotation angle data correspond to the same time point by synchronous acquisition clock or timestamp matching, thereby achieving timing alignment.

[0026] Further preferably, in step 1, the process of normalizing the j-th dimension feature is:

[0027] ;

[0028] in, is the i-th sample of the j-th dimension feature, is the i-th sample of the j-th dimension feature after standardization, is the mean of the j-th dimension feature, is the standard deviation of the j-th dimension feature, N is the total number of samples, and D is the feature dimension.

[0029] Further preferably, in step 4, kernel K-means clustering assigns samples to the nearest cluster center in the high-dimensional space by minimizing the intra-cluster squared error (SSE).

[0030] The present invention also provides a split-type detector for the mechanical state of an isolating switch, comprising a detector body, a data acquisition terminal, and a data analyzer; the data acquisition terminal interacts with the data analyzer via a wireless communication module, and the data acquisition terminal acquires the operating torque and rotation angle of the isolating switch measured by the detector body; the data analyzer is used for data management and defect diagnosis; the data analyzer comprises an analysis and diagnosis module, a display module, and a storage module; the acquired operating torque and rotation angle data are analyzed by the analysis and diagnosis module, and the analysis and diagnosis module comprises:

[0031] A data preprocessing and feature standardization unit is used to perform data preprocessing and feature standardization on the collected operating torque and rotation angle data to obtain an operating torque-rotation angle vector that matches the operating torque and rotation angle;

[0032] The kernel K-means clustering unit is used to select the kernel function and calculate the kernel matrix for the obtained operating torque-rotation angle vector. It performs kernel K-means clustering initialization and iterative optimization, iteratively adjusting cluster centers to minimize intra-cluster differences in high-dimensional space and effectively separate fault samples. The validity of kernel K-means clustering is verified using the silhouette coefficient and Davidson-Botting index, and a mapping relationship between clusters and fault types is established to achieve diagnosis from data to fault types.

[0033] The fault quantification and treatment measure unit uses the ELM-Tree model to quantify the fault degree and select treatment measures.

[0034] Further preferably, the detector body comprises:

[0035] A stepper motor used to drive the disconnector to open or close;

[0036] A torque detection module for measuring the operating torque required for the disconnector operation process;

[0037] A rotation angle detection module is used to measure the rotation angle of the disconnector.

[0038] Further preferably, the data acquisition terminal is connected to the rotation angle detection module and the torque detection module to collect and store data on the operating torque and rotation angle of the isolation switch.

[0039] Further preferably, an FPGA chip is deployed in the detector body to realize the following parallel calculations:

[0040] Kernel function calculation unit: parallel calculation of Gaussian kernel values;

[0041] Matrix block operations: Decompose the kernel matrix block and the inter-anchor kernel matrix into sub-blocks, and use pipeline technology to accelerate matrix inversion and multiplication.

[0042] The present invention achieves the following beneficial effects: kernel K-means clustering significantly improves the accuracy of disconnector fault classification through nonlinear mapping and optimized initialization. A reward function defining reinforcement learning is constructed using the silhouette coefficient S, the Davidson-Botting index (DBI), and the fault density index. Deep reinforcement learning is used to optimize the kernel parameter γ and the number of clusters K in kernel K-means clustering, resulting in optimal clustering results and ensuring reliable clustering results. Parameter tuning and engineering adaptation techniques (such as the Nyström approximation) make the detection method of the present invention applicable to large-scale field data. The kernel K-means clustering results provide highly discriminative input features for subsequent quantitative analysis using the ELM-Tree model, forming a complete diagnostic closed loop from fault detection to maintenance guidance, providing an efficient solution for safe power grid operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is the operating torque limit protection action curve of the disconnector.

[0044] Figure 2 This is the operating torque-rotation angle curve of the disconnector opening process.

[0045] Figure 3 This is the operating torque-rotation angle curve of the disconnector when it is stuck.

[0046] Figure 4 The present invention is a flow chart of a method for detecting the mechanical state of an isolating switch.

[0047] Figure 5 A visualization diagram of kernel K-means clustering.

[0048] Figure 6 This is a schematic diagram of a split-type detector for the mechanical status of an isolating switch. DETAILED DESCRIPTION

[0049] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0050] Example 1

[0051] During the operation of the isolating switch, the operating torque and rotation angle data can be measured to draw an operating torque-rotation angle curve, such as Figure 1 It is the operating torque limit protection action curve of the disconnector; Figure 2 This is the operating torque-rotation angle curve of the disconnector opening process. Figure 1 and Figure 2In the figure, X is the horizontal coordinate value and Y is the vertical coordinate value. It has high repeatability and reliability. The operating torque-rotation angle curve can accurately reflect the change of operating torque during the operation of the disconnector. In other words, the mechanical state of the disconnector can be analyzed and judged based on the operating torque-rotation angle curve. Figure 3 The operating torque-rotation angle curve of the disconnector shows obvious changes when it is stuck.

[0052] For this purpose, refer to Figure 4 This embodiment provides a method for detecting the mechanical state of an isolating switch, comprising the following steps:

[0053] Step 1: Perform data preprocessing and feature normalization on the collected operating torque and rotation angle data to obtain the operating torque-rotation angle vector that matches the operating torque and rotation angle;

[0054] Step 2: Select the kernel function and calculate the kernel matrix for the obtained operating torque-rotation angle vector;

[0055] Step 3: Kernel K-means clustering initialization;

[0056] Step 4: Kernel K-means clustering iterative optimization, by iteratively adjusting the cluster center, minimizes the intra-cluster differences in the high-dimensional space and achieves effective separation of fault samples;

[0057] Step 5: Construct a reward function that defines reinforcement learning using the silhouette coefficient S, the Davidson-Botting index DBI, and the fault density index. Optimize the kernel parameter γ and the number of clusters K of kernel K-means clustering through deep reinforcement learning to obtain the optimal clustering result.

[0058] Step 6: Establish a mapping relationship between clusters and fault types based on the optimal clustering results to achieve diagnosis from data to fault types;

[0059] Step 7: Use the ELM-Tree model (extreme learning machine decision tree model) to quantify the fault severity and select treatment measures.

[0060] Specifically, in step 1, data preprocessing includes filling missing data through linear interpolation, and ensuring that the operating torque and rotation angle data correspond to the same time point by synchronizing the acquisition clock or timestamp matching, thereby achieving timing alignment.

[0061] Among them, the process of standardizing the j-th dimension feature is:

[0062] ;

[0063] in, is the i-th sample of the j-th dimension feature, is the i-th sample of the j-th dimension feature after standardization, is the mean of the j-th dimension feature, is the standard deviation of the j-th dimension feature, N is the total number of samples, and D is the feature dimension. Here, when j=1, it corresponds to the operating torque data, and j=2 corresponds to the rotation angle data.

[0064] For example, the average operating torque of a disconnector is 120 N·m (standard deviation 15 N·m), and the rotation angle data is the meshing angle, which is 85° (standard deviation 5°). After standardization, the i-th operating torque value =135 is converted to (135-120) / 15=1, the i-th meshing angle =80 is converted to (80-85) / 5=−1. This process makes the weights of the two types of features equal in the clustering, preventing the operating torque value from dominating the classification result due to its large absolute value.

[0065] Fault modes of disconnectors (such as jamming and three-phase synch) often manifest as nonlinear variations in the operating torque-rotation angle curve. Traditional Euclidean distance cannot effectively capture these complex relationships. Kernel functions, however, implicitly map the data to a high-dimensional space, making them linearly separable. Due to their local sensitivity, Gaussian kernels (RBF kernels) are particularly well-suited for processing sudden changes in disconnector operation signals (such as a sudden increase in torque caused by jamming).

[0066] In step 2, the Gaussian kernel function is defined as:

[0067] ;

[0068] Among them, γ controls the width of the kernel function and determines the locality of the high-dimensional space. is the u-th sample, which is the standardized operating torque-rotation angle vector. is the vth sample, u,v={1,2,…,N}; is the Gaussian kernel value of the u-th sample and the v-th sample;

[0069] Kernel matrix construction: Calculate the Gaussian kernel values ​​of all sample pairs to obtain the N×N kernel matrix K:

[0070] ;

[0071] Assume that the operating torque-rotation angle vectors of the two samples are =[1.2,-0.5] and =[0.8,-0.3], taking γ=0.5, the Gaussian kernel value is:

[0072] =exp(-0.5×((1.2-0.8) 2 +(-0.5+0.3)2 ))=exp(-0.5×0.2)≈0.9048;

[0073] A high kernel value indicates that two samples are close in high-dimensional space and may belong to the same fault type (such as normal state).

[0074] In step 3, the cluster centers are reasonably selected through kernel K-means clustering initialization to avoid the clustering deviation caused by traditional random initialization and ensure coverage of typical fault modes such as normal state, jamming, and three-phase out-of-phase.

[0075] The number of clusters K is preset according to the historical fault type, and then the square of the distance from the sample to the nearest cluster center is used as the kernel distance to select K cluster centers.

[0076] sample Squared distance to the nearest cluster center:

[0077] ;

[0078] in, It is a high-dimensional mapping, which is indirectly expressed by the kernel matrix in actual calculation. is the kth cluster center.

[0079] Assuming that three clusters need to be initialized (normal state, jammed state, and three-phase out-of-sync), kernel K-means clustering first randomly selects a normal state sample as the first cluster center. It then calculates the kernel distance of all samples to this cluster center and selects the sample with the farthest distance (likely jammed state) as the second cluster center. The third cluster center (e.g., three-phase out-of-sync) is then probabilistically selected. This process ensures that the initial cluster centers represent different fault modes, avoiding the classification bias often associated with traditional methods due to the clustering of initial centers.

[0080] In step 4, kernel K-means clustering assigns samples to the nearest cluster center in high-dimensional space by minimizing the squared error (SSE) within the cluster. Because disconnector fault data can exhibit complex distributions (for example, stuck operating torque values ​​are concentrated in a certain range), the kernel K-means clustering method enhances inter-cluster separation through nonlinear mapping.

[0081] Objective function: Minimize the intra-cluster squared error:

[0082] ;

[0083] in, is the kth cluster;

[0084] For each sample , calculate its relationship with all cluster centers The kernel distance is assigned to the cluster with the closest distance:

[0085] ;

[0086] in, For samples To cluster The square distance, Cluster The number of samples, Cluster The first samples;

[0087] A sample The row vector of the kernel matrix is ​​[1.0, 0.9, 0.2, …, 0.7]. If cluster C1 contains 3 samples and its kernel value mean is 0.85, the distance is calculated as:

[0088] d 2 ( ,C1)=1.0-2×0.85+0.85 2 ≈0.0225, if this distance is smaller than the distance of other clusters, the sample It is classified into cluster C1, which may correspond to the normal state.

[0089] In step 5, the clustering results need to be verified by evaluation indicators. For example, a high silhouette coefficient indicates that similar fault samples are closely clustered and different clusters are clearly separated; a low Davidson-Borgin index reflects significant differences between clusters.

[0090] Traditional clustering evaluation relies on a single metric (such as the silhouette coefficient), which suffers from a single evaluation dimension and a tendency to fall into local optima. This paper proposes a reward function that combines the silhouette coefficient with the Davidson-Botting Index (DBI), supplemented by the Fault Density Index (FDI), to construct an optimization objective for deep reinforcement learning (DQN).

[0091] The silhouette coefficient S is used to measure the cohesion and separation of sample clusters and is defined as follows:

[0092] ;

[0093] in, is the uth sample The average distance to other samples in the same cluster, is the uth sample The average distance to the nearest neighbor cluster samples, is the total number of samples.

[0094] The Davidson-Boulding Index (DBI) is used to evaluate the inter-cluster separation and intra-cluster compactness, and the formula is:

[0095] ;

[0096] in, is the kth cluster The average distance from the inner sample to the cluster center, is the jth cluster The average distance from the inner sample to the cluster center, is the Euclidean distance between the center of the kth cluster and the center of the jth cluster, and K is the number of clusters.

[0097] The fault density index is used to quantify the compactness of the distribution of fault samples within a cluster and is defined as:

[0098] ;

[0099] in, is the density of the fault sample distribution within the kth cluster;

[0100] The reward function R for deep reinforcement learning is constructed by combining the silhouette coefficient S, the Davidson-Botting index DBI, and the fault density index:

[0101] ;

[0102] Among them, α, β, and η are the weight coefficients of the silhouette coefficient S, the Davidson-Botting index DBI, and the fault density index, respectively, which are calibrated through experiments (α=0.6, β=0.3, η=0.1).

[0103] ‌Actual working condition verification‌:

[0104] Cluster 1 (normal state): The average operating torque is between 100 and 130 N·m, the engagement angle deviation is less than 2°, and the stop angle is within the standard range.

[0105] Cluster 2 (stuck): The mean operating torque exceeds 150 N·m, the angle hysteresis exceeds 5°, and the silhouette coefficient S>0.7.

[0106] Cluster 3 (three phases out of phase): The meshing angles differ by more than 10° among phases A, B, and C, and the DBI value is significantly lower than that of other clusters.

[0107] Operation process:

[0108] 1. Label historical data: Add samples of known fault types to the training set and mark the cluster to which they belong.

[0109] 2. Create a mapping table: Count the proportion of each fault type in each cluster. For example, in cluster 2, 90% of the faults are jams.

[0110] 3. Verify new data: For new detection samples, determine the fault type based on the cluster to which it belongs and the mapping table.

[0111] The selection of the kernel parameter γ and the number of clusters K directly impacts clustering performance. Excessively large γ can lead to overfitting (e.g., mistaking noise for fault patterns), while too small a γ may fail to distinguish similar faults. Practical applications require a balance between computational complexity and diagnostic accuracy. Deep reinforcement learning is used to optimize the kernel parameter γ and the number of clusters K, testing within the ranges of γ ∈ {0.1, 1, 10} and K ∈ {3, 4, 5}. Data is divided into training and validation sets to avoid parameter overfitting.

[0112] The clustering process can be accelerated using the Nyström approximation

[0113] ;

[0114] in, A matrix constructed for a representative subset of samples selected from the original data, for The transpose of .

[0115] This paper uses an improved Nyström sampling algorithm. The traditional Nyström approximation requires random sampling of 𝑚 anchor points. This paper proposes importance-weighted sampling:

[0116] Calculate the similarity matrix between samples and select the m samples with the highest weights as anchor points.

[0117] The kernel matrix is ​​approximated as:

[0118] ;

[0119] in, is the kernel matrix block of the anchor point and the full sample, represents the transpose of the kernel matrix block C, Indicates that the kernel matrix block dimension is N×m, is the kernel matrix between anchor points, Indicates that the dimension of the kernel matrix between anchor points is m×m.

[0120] Compared with the traditional Nyström sampling algorithm, the improved Nyström sampling algorithm selects the samples with the highest weight for kernel matrix calculation, which shortens the kernel matrix calculation time and improves the clustering accuracy.

[0121] An approximate kernel matrix was constructed by selecting representative samples. For example, a substation monitors 500 disconnectors, generating 100,000 data points daily. Using the Nyström approximation, kernel matrix calculation time was reduced from 10 hours to 2 hours, while maintaining a clustering accuracy of >90%.

[0122] In step 6, based on the best clustering result and combined with historical fault data, a mapping relationship between clusters and fault types is established.

[0123] In step 7, the cluster features are input into the ELM-Tree model, and the ELM-Tree model outputs the fault severity level.

[0124] For example, clustering results show that the meshing angles of phases A, B, and C in a cluster are 82°, 88°, and 79°, respectively, exceeding the threshold. ELM-Tree model analysis shows a phase deviation of +6° (phase B), providing guidance for adjusting the connecting rod length.

[0125] The ELM-Tree model includes:

[0126] Decision tree framework: The fault type is the root node and each child node corresponds to a quantitative analysis model of the fault.

[0127] Node ELM Design:

[0128] Input layer: fault feature vector;

[0129] Hidden layer: randomly generate weights W and bias b, and use ReLU as the activation function.

[0130] Output layer: Analytically solve the output weight β and predict the target variable (such as wear depth and remaining life).

[0131] Training process:

[0132] Data partitioning: For each fault type, the training set and test set are divided into 7:3 ratios.

[0133] ‌ELM Node Training‌:

[0134] Randomly generate input weights W and bias b.

[0135] Calculate the hidden layer output matrix H:

[0136] ;

[0137] in, Represents the dimension of the hidden layer output matrix, L is the number of hidden layer nodes, n is the number of input samples, is the first input sample, is the first input sample, is the activation function;

[0138] Analytically solve the output weight β:

[0139] Tree splitting condition: If the current node prediction error (RMSE) is higher than the threshold , split features are selected by sorting by feature importance.

[0140] The test data set is 3000 sets of historical operation data of a 220kV disconnector, including 4 types of faults. The traditional K-means clustering + SVM method and the kernel K-means clustering + ELM-Tree method of the present invention are compared using the test data set. The results are shown in Table 1. The performance of the method of the present invention is greatly improved. The clustering results are shown in Table 1. Figure 5 As shown, it includes normal state, jamming, incomplete movement and three-phase asynchrony.

[0141] Table 1

[0142]

[0143] The method of this invention achieves high-precision fault classification through kernel K-means clustering and then quantifies the fault severity using the ELM-Tree model, forming a "clustering-regression" collaborative diagnosis framework. Its core value lies in:

[0144] Engineering Practicality: Adapts to the complex, nonlinear failure modes of disconnectors under multiple operating conditions without relying on large amounts of labeled data.

[0145] Computational efficiency: The ELM-Tree model's fast training and incremental update features support online monitoring and real-time early warning.

[0146] Comprehensive diagnosis: It not only determines the fault type but also outputs quantitative parameters (such as wear rate and remaining life), providing a dual basis for maintenance decisions.

[0147] This method has been successfully applied to the intelligent transformation project of the disconnector at a converter station, reducing the false alarm rate by 42% and extending the maintenance cycle by 30%, verifying its engineering effectiveness.

[0148] Example 2

[0149] like Figure 6 As shown, a split-type detector for the mechanical status of an isolating switch includes a detector body, a data acquisition terminal and a data analyzer; the data acquisition terminal exchanges data with the data analyzer through a wireless communication module, and the wireless communication module is preferably a LoRa module; the data acquisition terminal collects the operating torque and rotation angle of the isolating switch, and the data analyzer is used for data management and defect diagnosis.

[0150] The detector body comprises:

[0151] A stepper motor used to drive the disconnector to open or close. Within a certain torque range, when the disconnector operating torque changes, the stepper motor's speed remains unchanged. Its rotation speed is controlled by the number of pulses and frequency input by the driver.

[0152] A torque detection module is used to measure the operating torque required for the disconnector to operate. The torque detection module can use a torque sensor, which is connected to the operating mechanism of the disconnector inside the operating mechanism box through a coupling and an interface rod;

[0153] A rotation angle detection module is used to measure the rotation angle of the disconnect switch. This module uses an angle encoder to measure the rotation angle of the torque sensor. This angle is converted to the reduction ratio of the reducer inside the operating mechanism box to determine the rotation angle of the disconnect switch. Specifically, the module includes a large gear fixed to the motor, an absolute angle encoder, a small gear mounted on the output shaft of the angle encoder, a universal joint, a magnetic switch, and an RS485 serial-to-USB data transmission cable.

[0154] It also includes a control box and a power module. The control box includes a control panel, a motion controller that can control the start and stop of the stepper motor, and a travel switch that controls the opening and closing of the isolating switch to stop the stepper motor.

[0155] The power module supplies power to the control box, stepper motor and other modules.

[0156] The data acquisition terminal is connected to the rotation angle detection module and the torque detection module to collect and store data on the operating torque and rotation angle of the disconnector, and send the data to the data analyzer.

[0157] The data analyzer includes an analysis and diagnosis module, a display module, a storage module, and a battery module for powering the data analyzer. The analysis and diagnosis module includes:

[0158] A data preprocessing and feature standardization unit is used to perform data preprocessing and feature standardization on the collected operating torque and rotation angle data to obtain an operating torque-rotation angle vector that matches the operating torque and rotation angle;

[0159] The kernel K-means clustering unit is used to select the kernel function and calculate the kernel matrix for the obtained operating torque-rotation angle vector. It performs kernel K-means clustering initialization and iterative optimization, iteratively adjusting cluster centers to minimize intra-cluster differences in high-dimensional space and effectively separate fault samples. The validity of kernel K-means clustering is verified using the silhouette coefficient and Davidson-Botting index, and a mapping relationship between clusters and fault types is established to achieve diagnosis from data to fault types.

[0160] The fault quantification and treatment measure unit uses the ELM-Tree model to quantify the fault degree and select treatment measures.

[0161] In addition, an FPGA chip is deployed in the detector body to implement the following parallel computing:

[0162] Kernel function calculation unit: parallel calculation of Gaussian kernel values ;

[0163] Matrix block operation: decompose the core matrix block C and the inter-anchor core matrix W into sub-blocks, and use pipeline technology to accelerate matrix inversion and multiplication.

[0164] Experiments show that the calculation time of the improved algorithm on the FPGA chip is :

[0165] ;

[0166] Where P is the number of parallel computing units on the FPGA chip, O is the computational time, m is the number of plot points, and N is the total number of samples. When P = 128 and m = 100, the original algorithm's hour-level computation time can be shortened to 3.2 minutes.

[0167] During the operation, adjust the isolating switch to the open position, turn on the stepper motor, and the stepper motor rotates at a constant speed, so that the isolating switch rotates slowly and evenly from the open position to the closing position. The operating torque-rotation angle curve of the isolating switch closing process is drawn based on the collected operating torque-rotation angle. Before the isolating switch is tested, the basic information of the isolating switch (model, voltage level, historical data, etc.) should be entered first, the isolating switch information should be established, the power supply of each part should be ensured to be normal, the connection status of the wireless communication module (LoRa) should be checked, and the communication between the detection data acquisition terminal and the data analyzer should be ensured to be normal.

[0168] During the data acquisition process, the stepper motor drives the isolating switch to rotate at a constant speed. The torque sensor and angle encoder work synchronously to collect the operating torque and rotation angle respectively. The operating torque and rotation angle data are aligned according to the timestamp to eliminate timing errors.

[0169] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for detecting the mechanical state of an isolating switch, characterized in that: The steps include: Step 1: Perform data preprocessing and feature normalization on the collected operating torque and rotation angle data to obtain the operating torque-rotation angle vector that matches the operating torque and rotation angle; Step 2: Select the kernel function and calculate the kernel matrix for the obtained operating torque-rotation angle vector; Step 3: Kernel K-means clustering initialization; Step 4: Kernel K-means clustering iterative optimization, by iteratively adjusting the cluster center, minimizes the intra-cluster differences in the high-dimensional space and achieves effective separation of fault samples; Step 5: Construct a reward function that defines reinforcement learning using the silhouette coefficient S, the Davidson-Botting index DBI, and the fault density index. Optimize the kernel parameter γ and the number of clusters K of kernel K-means clustering through deep reinforcement learning to obtain the optimal clustering result. Step 6: Establish a mapping relationship between clusters and fault types based on the optimal clustering results to achieve diagnosis from data to fault types; Step 7: Use the ELM-Tree model to quantify the fault severity and select treatment measures.

2. The method for detecting the mechanical state of an isolating switch according to claim 1, wherein: The silhouette coefficient S is used to measure the cohesion and separation of sample clusters; the Davidson-Botting index (DBI) is used to evaluate the separation between clusters and the compactness within clusters; the fault density index is used to quantify the compactness of the distribution of fault samples within a cluster and is defined as: ； in, is the density of the fault sample distribution within the kth cluster, is the kth cluster center, Cluster The number of samples; is the kth cluster The average distance from the inner sample to the cluster center; is the u-th sample; The reward function R for deep reinforcement learning is constructed by combining the silhouette coefficient S, the Davidson-Botting index DBI, and the fault density index: ； Among them, α, β, and η are the weight coefficients of the silhouette coefficient S, the Davidson-Botting index DBI, and the fault density index, respectively.

3. The method for detecting the mechanical state of an isolating switch according to claim 1, characterized in that: In step 1, data preprocessing includes filling missing data through linear interpolation and synchronizing acquisition clocks or timestamp matching to ensure that the operating torque and rotation angle data correspond to the same time point, thereby achieving timing alignment.

4. The method for detecting the mechanical state of an isolating switch according to claim 1, characterized in that: The feature normalization mentioned in step 1 refers to normalizing the j-th dimension feature. The process is: ； in, is the i-th sample of the j-th dimension feature, is the i-th sample of the j-th dimension feature after standardization, is the mean of the j-th dimension feature, is the standard deviation of the j-th dimension feature, N is the total number of samples, and D is the feature dimension.

5. The method for detecting the mechanical state of an isolating switch according to claim 1, characterized in that: In step 4, kernel K-means clustering assigns samples to the nearest cluster center in the high-dimensional space by minimizing the intra-cluster squared error.

6. The method for detecting the mechanical state of an isolating switch according to claim 1, characterized in that: In step 1, the clustering process is accelerated using the Nyström approximation.

7. A split-type detector for the mechanical state of an isolating switch, comprising a detector body, a data acquisition terminal, and a data analyzer; the data acquisition terminal interacts with the data analyzer via a wireless communication module, the data acquisition terminal acquires the operating torque and rotation angle of the isolating switch measured by the detector body, and the data analyzer is used for data management and defect diagnosis; the data analyzer comprises an analysis and diagnosis module, a display module, and a storage module, and the acquired operating torque and rotation angle data are analyzed by the analysis and diagnosis module, characterized in that: The analysis and diagnosis module includes: A data preprocessing and feature standardization unit is used to perform data preprocessing and feature standardization on the collected operating torque and rotation angle data to obtain an operating torque-rotation angle vector that matches the operating torque and rotation angle; The kernel K-means clustering unit is used to select the kernel function and calculate the kernel matrix for the obtained operating torque-rotation angle vector. It performs kernel K-means clustering initialization and iterative optimization, iteratively adjusting cluster centers to minimize intra-cluster differences in high-dimensional space and effectively separate fault samples. The validity of kernel K-means clustering is verified using the silhouette coefficient and Davidson-Botting index, and a mapping relationship between clusters and fault types is established to achieve diagnosis from data to fault types. The fault quantification and treatment measure unit uses the ELM-Tree model to quantify the fault degree and select treatment measures.

8. The split-type detector for the mechanical state of the disconnector according to claim 7, characterized in that: The detector body comprises: A stepper motor used to drive the disconnector to open or close; A torque detection module for measuring the operating torque required for the disconnector operation process; A rotation angle detection module is used to measure the rotation angle of the disconnector.

9. The split-type detector for the mechanical state of the disconnector according to claim 8, characterized in that: The data acquisition terminal is connected to the rotation angle detection module and the torque detection module to collect and store data on the operating torque and rotation angle of the disconnector.

10. The split-type detector for the mechanical state of the disconnector according to claim 9, characterized in that: An FPGA chip is deployed in the detector body to implement the following parallel computing: Kernel function calculation unit: parallel calculation of Gaussian kernel values; Matrix block operations: Decompose the kernel matrix block and the inter-anchor kernel matrix into sub-blocks, and use pipeline technology to accelerate matrix inversion and multiplication.

Images