Insulator zero value detection robot control method

Abstract

The invention relates to the field of power equipment detection and intelligent robot control, and discloses an insulator zero value detection robot control method. The method comprises the following steps: deploying a rail-mounted robot on a power transmission line, and synchronously acquiring infrared thermal imaging, ultraviolet corona, capacitance coupling voltage and laser contour four-dimensional data; performing space-time alignment and multi-modal feature extraction on the multi-source information; generating an insulator unit-level health score through cross-modal attention fusion; and judging a zero-value fault according to the scoring threshold value and triggering a rechecking or confirming process. The system comprises a double-wheel clamping driving mechanism, a multi-source sensing array and a central controller. According to the invention, full-automatic, high-precision and high-safety detection in an electrified state is realized, and the recognition accuracy and the operation efficiency are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of power equipment testing and intelligent robot control, and specifically relates to a robot control method for detecting zero values ​​of insulators. Background Technology

[0002] With the continuous improvement of the intelligent operation and maintenance level of high-voltage transmission lines, the real-time monitoring of the performance status of insulators, as key components ensuring the safe operation of the power grid, has become one of the core tasks of power system maintenance. During long-term operation, insulators are susceptible to environmental corrosion, mechanical stress, and uneven electric field distribution, leading to the loss of insulation capacity in some insulators, resulting in "zero value" or "low value" defects. If such defects are not identified in time, they may cause flashover, breakdown, or even line breakage accidents, seriously threatening the stability of the power grid and the safety of personnel. Therefore, accurate and efficient detection of the electrical performance of insulators has become an important research direction in the field of intelligent inspection.

[0003] The insulator zero-value online diagnostic technology based on multimodal sensing fusion and digital twin collaboration aims to achieve dynamic assessment and autonomous decision-making control of insulator status in a non-contact, uninterrupted power-off manner. This technical approach relies on the joint sensing of physical features such as acoustic signatures and ultraviolet pulses, combined with real-time mapping and simulation of physical entities using virtual models, to improve the robustness and response speed of fault identification under complex electromagnetic and meteorological environments. However, existing detection methods still have significant limitations: traditional methods rely on manual tower climbing or contact measurement after power outages, which is not only inefficient but also poses high safety risks in high-voltage environments; while current mainstream single-sensor online detection solutions (such as those relying solely on infrared thermal imaging or ultraviolet imaging) are susceptible to interference from environmental noise, background radiation, or equipment aging, resulting in a high false positive rate and failing to meet the requirements of high-reliability operation and maintenance.

[0004] Existing technologies have not yet achieved effective integration in terms of multi-source information fusion mechanisms, dynamic environmental adaptability, and closed-loop control for detection and decision-making. Specifically, while acoustic signals can reflect abnormal vibrations caused by internal cracks or loosening of insulators, they are easily affected by wind noise and conductor corona interference; ultraviolet pulses can capture partial discharge phenomena, but they are not sensitive enough to weak discharge signals and are easily suppressed by sunlight background. If used independently, neither can accurately distinguish between zero-value insulators and normal aging states. Furthermore, the lack of deep coupling with a digital twin platform prevents the detection results from driving the robot to perform path replanning, sensor parameter self-adjustment, or re-inspection strategy generation, resulting in a fragmented "perception-judgment" state for the entire system. In critical scenarios such as ultra-high voltage and inter-regional power transmission, these deficiencies directly restrict the timeliness, accuracy, and operational safety of insulator defect identification, necessitating a control method for a zero-value insulator detection robot that integrates acoustic and ultraviolet multimodal perception and possesses real-time reasoning and autonomous control capabilities. Summary of the Invention

[0005] This invention provides a robot control method for zero-value detection of insulators, aiming to solve the technical problems of low efficiency and high operational risks caused by traditional detection methods that require machine shutdown or contact measurement, as well as the susceptibility to misjudgment by single sensors. The method achieves fully automatic, high-precision, and high-safety zero-value fault identification of transmission line insulator strings under energized operation by constructing a multi-source heterogeneous sensor fusion architecture and a non-contact dynamic detection mechanism.

[0006] As one embodiment of the present invention, the insulator zero-value detection robot control method includes the following steps: deploying a track-type robot with autonomous walking capability on the ground wire or conductor of the transmission line; synchronously collecting multi-dimensional physical field information of the insulator string through an infrared thermal imaging sensor, an ultraviolet corona detector, a distributed capacitance coupled voltage sensing array and a laser contour scanner; and performing timestamp alignment and spatial coordinate normalization processing on the above-mentioned multi-source sensing data based on a preset spatiotemporal synchronization benchmark. Spatial coordinate normalization is achieved through the collaboration of robot IMU data and laser geometric reference. First, a unified three-dimensional coordinate system with the insulator string axis as the reference is constructed. Then, the observation results of each sensor are mapped to this coordinate system through homogeneous coordinate transformation to ensure the spatial alignment accuracy of multi-source data.

[0007] A multimodal feature extraction network is used to analyze the raw signals of each sensing channel to generate thermal distribution feature vectors, corona discharge intensity sequences, voltage gradient distribution maps, and geometric deviation parameters. These four types of features are then input into a cross-modal attention fusion module to calculate the confidence weight of each feature channel at the insulator unit level and to synthesize a weighted comprehensive health status score. The comprehensive health status score is compared with a preset threshold range to determine whether there are zero-value insulators, and the fault location number and confidence level are output. Based on the determination result, the central controller generates path correction instructions and re-inspection scheduling strategies to drive the robot to perform local backtracking scans or cross-verification operations of adjacent insulators.

[0008] Furthermore, the track-type robot is equipped with a dual-wheel clamping drive mechanism, which includes an upper pressure wheel assembly, a lower load-bearing wheel assembly, and a servo motor drive unit. The upper pressure wheel assembly applies a constant preload through a spring loading mechanism to ensure that the robot maintains stable attachment on tracks with different inclination angles. The servo motor drive unit adopts a closed-loop vector control mode to adjust the speed difference between the left and right wheels in real time to achieve precise displacement control along the axial direction of the guide wire, with a positioning error not exceeding ±5 mm.

[0009] Furthermore, the infrared thermal imaging sensor is configured as a long-wave infrared focal plane array with a working wavelength of 8 to 14 micrometers, a spatial resolution of 640×480 pixels, and a frame rate of no less than 25 Hz. The sensor is mounted on the front gimbal of the robot and achieves automatic focusing and field of view coverage of the target insulator string through a two-dimensional servo mechanism, ensuring that the area of ​​each insulator skirt is completely imaged.

[0010] Furthermore, the ultraviolet corona detector adopts a solar-blind ultraviolet photomultiplier tube with a response wavelength range of 240 to 280 nanometers and a minimum detectable photon flux of 50 photons per second. The detector is installed on the same optical axis as a visible light auxiliary camera. The ultraviolet discharge signal is superimposed on the visible light image through an image registration algorithm to form a thermal map of the corona spatial distribution. Specifically, this is achieved through joint camera calibration, ultraviolet signal quantization, affine transformation registration, and semi-transparent color superposition. That is, the intrinsic and extrinsic parameters of the two cameras are first calibrated to establish a coordinate mapping, then the ultraviolet photon signal is quantized into a grayscale image and mapped to the visible light coordinate system, and finally converted into a thermal color mode superposition.

[0011] Furthermore, the distributed capacitively coupled voltage sensing array consists of nine miniature capacitive probes arranged laterally along the robot, with a spacing of 10 cm between each probe and the sensing surface of the probe facing the insulator string below. Each probe is connected to a 12-bit analog-to-digital converter via a high input impedance buffer amplifier, with a sampling frequency of 1000 Hz. By measuring the rate of change of potential difference between adjacent probes, the voltage distribution curve of the insulator string along the axial direction is inverted.

[0012] Furthermore, the laser profile scanner adopts the triangulation principle, emits a semiconductor laser beam with a wavelength of 650 nanometers, a scanning frequency of 2000 Hz, and a ranging accuracy of ±0.1 mm; the scanner performs continuous line scanning along the insulator string axis to obtain three-dimensional point cloud data of the outer diameter of the shed, the exposed length of the steel foot, and the depth of surface cracks of each insulator.

[0013] Furthermore, the multimodal feature extraction network includes four parallel sub-networks: the first sub-network is a one-dimensional convolutional neural network used to process the voltage gradient distribution map and extract the voltage drop inflection point position and amplitude attenuation rate; the second sub-network is a two-dimensional residual network used to analyze infrared thermal images and identify abnormal hot spots and their temperature gradient features; the third sub-network is a time-series pulse counting model used to count the ultraviolet photon event density per unit time and combine it with a spatial clustering algorithm to locate the discharge source; the fourth sub-network is a point cloud semantic segmentation network based on the PointNet architecture, which performs component-level segmentation on the laser scanning point cloud and outputs structural integrity indicators.

[0014] Furthermore, the cross-modal attention fusion module adopts a gated cross-attention mechanism, and its input is the feature tensor output by the above four sub-networks. The module first linearly projects each feature tensor to a unified embedding space, and then uses the voltage distribution feature as the query vector, and the other three types of features as key vectors and value vectors respectively. The correlation weight is calculated by the query vector and each key vector, and then the value vectors are weighted and summed to achieve fusion, and the attention weight is calculated. The final output is the weighted fusion insulator unit-level health state embedding vector with a dimension of 256.

[0015] Furthermore, the comprehensive health status score is obtained by mapping the health status embedding vector through a fully connected regression layer, and the value range is 0 to 100; the preset threshold range includes a first threshold of 70 and a second threshold of 90; when the score is below 70, it is determined to be a zero-value insulator; when the score is between 70 and 90, it is marked as a suspected fault and triggers the re-inspection process; when the score is above 90, it is determined to be normal.

[0016] Furthermore, the re-inspection scheduling strategy includes: controlling the robot to retreat to the position of the previous insulator and re-execute multi-source data acquisition; adjusting the integration time of the infrared and ultraviolet sensors to improve the signal-to-noise ratio of weak signals; starting the high-density scanning mode of the laser scanner and reducing the sampling point spacing to 0.5 mm; if the two detection results are consistent, the fault is confirmed; otherwise, continue to detect the next insulator and record contradictory data for background analysis.

[0017] Furthermore, the central controller has a built-in task state machine, whose states include initialization, cruise detection, partial re-inspection, fault confirmation, and return to standby; the state transition is driven by the detection results and environmental feedback; in the cruise detection state, the robot moves forward at a constant speed of 3 meters per minute; in the partial re-inspection state, the speed is reduced to 0.5 meters per minute, and a high-precision positioning mode is enabled.

[0018] Furthermore, the method also includes a communication assurance mechanism: the robot establishes a two-way data link with the ground base station through the power line carrier communication module; key detection data is compressed and encoded and transmitted in groups, each group containing the insulator number, four types of original sensor segments, feature vectors and preliminary judgment results; the ground base station receives the data and performs secondary verification, and can remotely issue intervention commands.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: By constructing a four-source heterogeneous sensing system encompassing infrared, ultraviolet, capacitively coupled voltage, and laser contour sensing, the system simultaneously acquires four-dimensional physical field information of insulators (thermal, optical, electrical, and morphological) under uninterrupted power supply and non-contact conditions. This fundamentally avoids the safety risks and efficiency bottlenecks associated with manual tower climbing or power outage detection. Employing a feature fusion strategy based on a cross-modal attention mechanism effectively suppresses misjudgments caused by environmental interference or device drift from a single sensor, significantly improving the accuracy and robustness of zero-value insulator identification. The introduction of closed-loop re-inspection scheduling and task state machine control logic enables the robot to make autonomous decisions and handle anomalies, achieving full automation from data acquisition to fault confirmation. The entire method supports stable operation on UHV AC or DC transmission lines, with a detection speed more than five times faster than traditional methods, a false alarm rate of less than 2%, and a false negative rate of less than 0.5%, providing reliable technical support for intelligent power grid operation and maintenance. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the cross-modal attention fusion module in this invention; Figure 3 This is a logical flowchart of the multi-source heterogeneous sensor data acquisition and preprocessing stage in this invention. Figure 4 This is a diagram of the parallel processing logic framework of the multimodal feature extraction network in this invention; Figure 5 This is a flowchart of the detection and re-inspection scheduling logic driven by the robot task state machine in this invention; Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the track-mounted robot and the ground base station in this invention. Detailed Implementation

[0021] Please refer to the attached document. Figures 1 to 6 This invention provides a robot control method for detecting zero-value insulators. The core of this method lies in constructing a closed-loop control system that enables fully automatic, non-contact, and high-precision identification of zero-value insulators during the energized operation of transmission lines. The method deploys an autonomous, track-mounted robot on the ground wire or conductor of the transmission line to simultaneously collect multi-dimensional physical field information of the insulator strings. Based on multi-source heterogeneous sensor fusion and cross-modal feature interaction mechanisms, it performs dynamic assessment of the insulator's health status and fault determination.

[0022] The method first executes the robot deployment step. The track-mounted robot is installed on the ground wire or conductor of the power transmission line, and its structure includes a dual-wheel clamping drive mechanism. This mechanism consists of an upper pressure wheel assembly, a lower load-bearing wheel assembly, and a servo motor drive unit. The upper pressure wheel assembly applies a constant preload through a spring loading mechanism to ensure stable attachment of the robot on lines with different inclination angles, preventing derailment due to wind loads or line vibrations. The servo motor drive unit adopts a closed-loop vector control mode, adjusting the speed difference between the left and right wheels in real time to achieve precise displacement control along the conductor axis, with a positioning error not exceeding ±5 mm. After startup, the robot enters an initialization state, completing self-testing, sensor calibration, and communication link establishment.

[0023] The robot then performs a multi-source sensor data synchronous acquisition step. It is equipped with four core sensors: an infrared thermal imaging sensor, an ultraviolet corona detector, a distributed capacitively coupled voltage sensing array, and a laser contour scanner. The infrared thermal imaging sensor is configured as a long-wave infrared focal plane array, operating in the 8-14 micrometer band, with a spatial resolution of 640×480 pixels and a frame rate of no less than 25 Hz. This sensor is mounted on the robot's front-end gimbal and uses a two-dimensional servo mechanism to achieve automatic focusing and field-of-view coverage of the target insulator string, ensuring that each insulator skirt area is completely imaged.

[0024] The ultraviolet corona detector employs a solar-blind ultraviolet photomultiplier tube with a response wavelength range of 240 to 280 nanometers and a minimum detectable photon flux of 50 photons per second. This detector is mounted coaxially with a visible light auxiliary camera. An image registration algorithm is used to superimpose the ultraviolet discharge signal onto a visible light image, forming a thermal map of the corona spatial distribution.

[0025] The specific implementation process is as follows: First, the ultraviolet camera and the visible light camera are jointly calibrated to obtain the intrinsic parameters (focal length, principal point coordinates) and extrinsic parameters (relative translation, rotation angle) of the two cameras, and establish a pixel coordinate mapping relationship; then, the discharge photon count signal collected by the ultraviolet detector is quantized into grayscale values ​​(the higher the discharge intensity, the larger the grayscale value), forming a single-channel ultraviolet grayscale image; subsequently, an affine transformation algorithm is used to map the ultraviolet grayscale image to the visible light image coordinate system according to the calibration parameters, and the coordinate deviation is corrected by bilinear interpolation; finally, the ultraviolet grayscale image is converted into a heat map color mode (the discharge intensity changes from low to high, corresponding to blue, green, yellow, and red gradients), and superimposed on the corresponding area of ​​the visible light image in a semi-transparent mode with 50% transparency, ultimately forming a heat map that intuitively reflects the spatial distribution of the corona.

[0026] The distributed capacitively coupled voltage sensing array consists of nine miniature capacitive probes arranged laterally along the robot, with a spacing of 10 cm between each probe and the sensing surface of the probe facing the insulator string below. Each probe is connected to a 12-bit analog-to-digital converter via a high input impedance buffer amplifier, with a sampling frequency of 1000 Hz. By measuring the rate of change of potential difference between adjacent probes, the voltage distribution curve of the insulator string along the axial direction is inverted.

[0027] The laser profile scanner employs the triangulation principle, emitting a semiconductor laser beam with a wavelength of 650 nanometers and a scanning frequency of 2000 Hz, achieving a ranging accuracy of ±0.1 mm. This scanner performs continuous line scanning along the insulator string axis, acquiring three-dimensional point cloud data of the outer diameter of the sheds, the exposed length of the steel feet, and the depth of surface cracks for each insulator. All four types of sensors operate synchronously under a unified clock source, ensuring strict time alignment of all raw data.

[0028] After data acquisition, the system performs spatiotemporal synchronization and coordinate normalization. Based on a preset spatiotemporal synchronization benchmark, the system timestamps infrared image frames, ultraviolet event streams, voltage sampling sequences, and laser point cloud trajectories. Time alignment uses a hardware trigger signal as a common reference point; all sensor data streams are interpolated and corrected with this reference point as their zero point, unifying data from different sampling frequencies onto the same time grid. The time resolution is set to 10 milliseconds. Spatial coordinate normalization, based on the robot's current position and attitude information, combined with the geometric benchmark provided by the laser scanner, maps the sensor observations to a unified three-dimensional coordinate system with the insulator string axis as the reference.

[0029] The specific mapping process is as follows: Basic data acquisition: The robot acquires its own position coordinates in real time through an inertial measurement unit (IMU). And attitude angles (roll angle, pitch angle, yaw angle); the laser scanner scans the hardware, steel cap and other feature points of the insulator string to obtain the three-dimensional coordinates of the feature points, and establishes the geometric reference by taking the line connecting the center points of the hardware at both ends of the insulator string as the initial axis; Unified coordinate system construction: With the insulator string axis as the Z-axis, the center point of one end fitting as the origin O, the horizontal direction perpendicular to the axis as the X-axis, and the vertical direction perpendicular to the axis as the Y-axis, a unified three-dimensional coordinate system O-XYZ is constructed. Coordinate transformation mapping: Sensor data mapping is achieved through a homogeneous coordinate transformation matrix. First, the original sensor observation coordinates are converted to world coordinates based on the robot's position and attitude information. Then, translation and rotation operations are performed using the laser geometric reference to map the world coordinates to the O-XYZ coordinate system. The transformation formula is as follows: ,in To unify the coordinates of the target point in the coordinate system, These are the original observation coordinates. Here is the rotation matrix corresponding to the attitude angle. Let be the translation vector of the robot relative to the origin. To unify the coordinates of the origin of the coordinate system in the world coordinate system.

[0030] Each insulator is assigned a unique number, and its physical location is determined by the axial distance measured by a laser scanner, with an error controlled within ±2 mm. The normalized data is sliced ​​by insulator unit to form a multi-channel input block with a single insulator as the processing granularity.

[0031] Next, the multimodal feature extraction step is executed. The system calls the multimodal feature extraction network, which consists of four parallel sub-networks. The first sub-network is a one-dimensional convolutional neural network used to process the voltage gradient distribution map. This sub-network receives a normalized 9-point voltage sequence, extracts local potential change features step by step through three one-dimensional convolutional layers, and finally outputs the voltage drop inflection point position and amplitude attenuation rate to form a one-dimensional voltage feature vector with a dimension of 64. The second sub-network is a two-dimensional residual network used to analyze infrared thermal images. This sub-network takes the infrared image block corresponding to a single insulator as input, extracts the spatial thermal distribution pattern through residual block stacking, identifies abnormal hot spots and their temperature gradient features, and outputs a two-dimensional thermal feature tensor, which is compressed into a 64-dimensional vector after global average pooling. The third sub-network is a time-series pulse counting model used to count the ultraviolet photon event density per unit time. This model performs sliding window counting on the photon arrival time series output by the ultraviolet detector, combines spatial clustering algorithm to aggregate discrete photon events into discharge power clusters, calculates the center coordinates and intensity weight of each cluster, and finally generates a 32-dimensional corona feature vector. The fourth sub-network is a point cloud semantic segmentation network, based on the PointNet architecture, which performs component-level segmentation on laser scanning point clouds.

[0032] The network receives a subset of point clouds corresponding to a single insulator, extracts point-level features through a shared multilayer perceptron, and then obtains a global shape descriptor through max pooling. It outputs structural integrity indicators, including skirt deformation, steel foot exposure ratio, and crack coverage, forming a 64-dimensional geometric feature vector. These four types of feature vectors together constitute the multimodal representation set of the insulator unit.

[0033] The cross-modal attention fusion step is then executed. The system calls the cross-modal attention fusion module, which employs a gated cross-attention mechanism. The input consists of feature vectors output from the four sub-networks mentioned above, with a total dimension of 224. The module first linearly projects each feature vector to a unified embedding space, resulting in a dimension of 64 for each vector, forming a query vector, a key vector, and a value vector. Using the voltage distribution feature as the query vector, and the other three types of features as the key and value vectors respectively, attention weights are calculated.

[0034] The key and value vectors of the other three types of features are assigned as follows: the thermal distribution feature vector is used as the first key vector and the first value vector, the corona discharge intensity sequence is used as the second key vector and the second value vector, and the geometric structure deviation parameter is used as the third key vector and the third value vector. The query vector (voltage distribution feature projection vector) is used to calculate the correlation weight with each key vector. Then, the weighted value vector is obtained by calculating the sum of the products of each correlation weight and the corresponding value vector, so as to achieve the targeted fusion of the three types of features with the voltage feature.

[0035] Attention weight Calculated using the scaled dot product formula:

[0036] in The voltage feature projection vector, For the first Key vectors of non-voltage characteristics, For the first Key vectors of non-voltage characteristics, The embedding dimension is 64. The calculated attention weights reflect the contribution of thermal, optical, and shape features to fault determination under the current voltage anomaly background. Subsequently, each value vector is weighted and summed according to its corresponding weight, and then concatenated with the query vector and nonlinearly fused through a gated linear unit. The final output is a 256-dimensional insulator unit-level health state embedding vector.

[0037] Based on this embedding vector, a comprehensive health status score generation step is performed. The system calls a fully connected regression layer, which consists of two perceptron layers with a hidden layer dimension of 128 and an output layer activation function of sigmoid. This layer maps the health status embedding vector to a single scalar score, with a value ranging from 0 to 100. This score comprehensively reflects the electrical performance, thermal stability, discharge activity, and structural integrity of the insulator. The preset threshold range includes a first threshold of 70 and a second threshold of 90. When the score is below 70, it is judged as a zero-value insulator; when the score is between 70 and 90, it is marked as a suspected fault; when the score is above 90, it is judged as normal.

[0038] After the judgment result is generated, the fault judgment and output steps are executed. The system iterates through the scores of all detected insulator units and classifies them according to the threshold range. For insulators judged to have a zero value, their location number, comprehensive score, and confidence level of the four types of original features are recorded; for suspected faults, a re-inspection flag is set. The judgment result, along with the original sensor segments, feature vectors, and location information, is packaged and prepared for subsequent control decisions.

[0039] Based on the judgment results, the path correction and re-inspection scheduling steps are executed. The central controller has a built-in task state machine, whose states include initialization, cruise detection, partial re-inspection, fault confirmation, and return to standby. In the cruise detection state, the robot moves forward at a constant speed of 3 meters per minute, continuously executing the above data acquisition and judgment process. Once a suspected fault or zero-value insulator is detected, the state machine immediately switches to the partial re-inspection state. In this state, the robot speed is reduced to 0.5 meters per minute, and high-precision positioning mode is activated.

[0040] The re-inspection scheduling strategy specifically includes: controlling the robot to retreat to the position of the previous insulator and re-execute multi-source data acquisition; adjusting the integration time of the infrared and ultraviolet sensors, extending the infrared integration time to 1.5 times the original value and the ultraviolet integration time to 2 times the original value to improve the signal-to-noise ratio of weak signals; and activating the high-density scanning mode of the laser scanner, reducing the sampling point spacing to 0.5 mm to improve geometric detail resolution. After the re-inspection is completed, feature extraction and score generation are performed again. If the two detection results are consistent, the fault is confirmed, the state machine enters the fault confirmation state, and the final judgment result is recorded; otherwise, the re-inspection flag is cleared, and the robot continues to detect the next insulator, marking the contradictory data as samples to be analyzed in the background.

[0041] Throughout the entire testing process, a communication safeguard mechanism is implemented. The robot establishes a two-way data link with the ground base station via a power line carrier communication module. Key testing data is compressed and encoded before being transmitted in packets. Each packet contains the insulator number, four types of original sensor segments, feature vectors, and preliminary judgment results. Compression employs a combination of lossless differential coding and Huffman coding to ensure that key features are not distorted. After receiving the data, the ground base station performs secondary verification, including data integrity verification and logical consistency checks. If an anomaly is detected, the base station can remotely issue intervention commands, such as mandatory re-inspection, suspension of the mission, or emergency return. The communication link uses frequency hopping spread spectrum technology, which has strong anti-electromagnetic interference capabilities and is suitable for ultra-high voltage AC or DC transmission line environments.

[0042] In actual operation on UHV transmission lines, the described method involves a robot traversing the ground wire at a speed of 3 meters per minute, with each insulator inspection taking approximately 8 seconds. Data is collected simultaneously by four types of sensors, and feature extraction and fusion are performed by a local edge computing unit, with a total latency of less than 500 milliseconds. The cross-modal attention mechanism effectively suppresses false alarms from a single sensor: for example, in strong sunlight, infrared images may show false hotspots, but if the voltage distribution is normal and there is no ultraviolet discharge, the attention weight automatically reduces the contribution of thermal features, avoiding false alarms. Conversely, in rainy or foggy weather when ultraviolet signals attenuate, if the voltage gradient drops significantly and the geometry remains intact, the system can still accurately identify zero-value faults. Experimental data shows that in a 5000-kilometer line test, the false alarm rate was less than 2%, the false negative rate was less than 0.5%, and the detection speed was more than 5 times faster than the traditional manual tower climbing method.

[0043] The mechanical structure of the track-mounted robot ensures stable operation on a 30-degree incline, while the dual-wheel gripping mechanism provides an adhesion force greater than 50 Newtons, capable of withstanding winds up to level 10. The servo motor drive unit employs incremental encoder feedback, combined with a Kalman filter to estimate position, achieving millimeter-level positioning accuracy. The sensor gimbal has two degrees of freedom: pitch and yaw, and can automatically plan the optimal observation angle before detection, ensuring unobstructed imaging of each insulator.

[0044] The distributed capacitively coupled voltage sensing array inverts the voltage distribution by measuring the rate of change of potential difference between probes; its mathematical model is based on the principle of capacitive voltage division. Let the... With the The potential difference between the probes is Then the voltage gradient of the corresponding section on the insulator string is ,in The probe spacing is 10 cm. The voltage gradient distribution spectrum is the sequence. Zero-value insulators manifest as a certain section The gradient is close to 0, while the gradient of adjacent segments increases abruptly. .

[0045] The point cloud data acquired by the laser contour scanner is processed by denoising, registration, and segmentation before being input into the PointNet network. This network extracts the local geometric features of each point using a shared MLP and then aggregates them into a global descriptor using a symmetric function. The structural integrity index is a weighted composite of three sub-indices: the skirt deformation degree is defined as the absolute value of the relative deviation between the measured outer diameter and the standard value; the exposed length of the steel foot is calculated by the axial extension of the metal portion in the point cloud; and the crack depth is obtained by fitting the abrupt change point of the laser reflection intensity. The three indices are normalized and then weighted and summed, with weights of 0.4, 0.3, and 0.3, respectively.

[0046] The state transitions of the task state machine are driven by both detection results and environmental feedback. For example, in a partial re-inspection state, if the re-inspection result is still suspected, a maximum of two re-inspection attempts are allowed; if the third result is still inconsistent, it is marked as a difficult sample and the process continues. After the fault confirmation state, the robot can choose to continue detection or return immediately, depending on the remaining battery power and task priority. The return path adopts the original route return strategy, with the speed increased to 4 meters per minute to save energy.

[0047] In summary, the insulator zero-value detection robot control method of the present invention achieves high-precision and high-robustness identification of zero-value insulators under uninterrupted power supply and non-contact conditions through multi-source heterogeneous sensor fusion, cross-modal attention mechanism and closed-loop re-inspection scheduling, which significantly improves the safety and efficiency of power grid operation and maintenance.

Claims

1. An insulator zero value detection robot control method, characterized by, The application relates to a track robot for detecting insulator strings on a power transmission line. The track robot is provided with a double-wheel clamping driving mechanism, which comprises an upper pressing wheel assembly, a lower supporting wheel assembly and a servo motor driving unit. The infrared thermal imaging sensor is configured as a long-wave infrared focal plane array, the working wave band of which is 8-14 micrometers, the spatial resolution is 640*480 pixels, and the frame rate is not less than 25 Hz. The ultraviolet corona detector adopts a sun-blind ultraviolet photomultiplier, the response wavelength range of which is 240-280 nanometers, and the minimum detectable photon flux is 50 photons per second. The ultraviolet corona detector and the visible light auxiliary camera are coaxially installed, the ultraviolet discharge signal is superimposed into the visible light image through an image registration algorithm, a corona spatial distribution thermogram is formed, and the thermogram is specifically realized through camera joint calibration, ultraviolet signal quantization, affine transformation registration and semi-transparent color superimposition, that is, the internal and external parameters of the two cameras are calibrated to establish coordinate mapping, the ultraviolet photon signal is quantized into a gray image and mapped into the visible light coordinate system, and finally the thermogram is converted into a thermal color mode and superimposed. ​ ​ ​ ​ 2. The insulator null detection robot control method of claim 1, wherein ​ 3. The insulator null detection robot control method of claim 1, wherein ​ 4. The insulator null detection robot control method of claim 1, wherein ​ 5. The insulator null detection robot control method of claim 1, wherein The distributed capacitive coupling voltage sensing array is composed of 9 micro-capacitive probes arranged transversely to the robot, each probe is spaced 10 cm apart, and the probe sensing surface faces downward to the insulator string; each probe is connected to a 12-bit analog-to-digital converter through a high-input-impedance buffer amplifier, and the sampling frequency is 1000 Hz; the voltage distribution curve along the axial direction of the insulator string is obtained by measuring the rate of change of the potential difference between adjacent probes.

6. The insulator null detection robot control method of claim 1, wherein The laser profile scanner adopts a triangulation ranging principle, emits a semiconductor laser beam with a wavelength of 650 nm, and has a scanning frequency of 2000 Hz and a ranging accuracy of ±0.1 mm; the scanner performs continuous line scanning along the axial direction of the insulator string to obtain three-dimensional point cloud data of the outer diameter of the shed, the exposed length of the steel foot and the surface crack depth of each insulator.

7. The insulator null detection robot control method of claim 1, wherein The multi-modal feature extraction network includes 4 parallel sub-networks: the first sub-network is a one-dimensional convolutional neural network for processing voltage gradient distribution maps to extract voltage drop inflection point positions and amplitude decay rates; the second sub-network is a two-dimensional residual network for analyzing infrared thermal images to identify abnormal hot spot regions and their temperature gradient features; the third sub-network is a time series pulse counting model for counting the density of ultraviolet photon events per unit time and locating the discharge source combined with a spatial clustering algorithm; and the fourth sub-network is a point cloud semantic segmentation network based on the PointNet architecture for component-level segmentation of laser scanning point clouds to output structural integrity indicators.

8. The insulator null detection robot control method of claim 7, wherein, The cross-modal attention fusion module adopts a gated cross-attention mechanism, and the input is the feature tensors output by the above-mentioned 4 sub-networks; the module first projects each feature tensor linearly to a unified embedding space, then takes the voltage distribution features as a query vector, the remaining three types of features as key vectors and value vectors respectively, calculates the correlation weight between the query vector and each key vector, and then weightedly sums each value vector to realize fusion and calculate the attention weight; finally, the output is a weightedly fused insulator unit-level health state embedding vector with a dimension of 256.

9. The insulator null detection robot control method of claim 8, wherein, The comprehensive health state score is obtained by mapping the health state embedding vector through a fully connected regression layer, and the value range is 0 to 100; the preset threshold interval includes a first threshold of 70 and a second threshold of 90; when the score is lower than 70, it is determined as a zero-value insulator; when the score is between 70 and 90, it is marked as a suspected fault, triggering a re-inspection process; when the score is higher than 90, it is determined as normal.

10. The insulator null detection robot control method of claim 9, wherein, The re-inspection scheduling strategy includes: controlling the robot to retreat to the position of the previous insulator and re-executing multi-source data acquisition; adjusting the integration time of the infrared and ultraviolet sensors to improve the signal-to-noise ratio of weak signals; starting the high-density scanning mode of the laser scanner, and the sampling point spacing is reduced to 0.5 mm; if the results of 2 detections are consistent, the fault is confirmed; otherwise, continue to detect the next insulator and record the contradictory data for background analysis.

Images