An overhead insulated conductor surface defect visual inspection system based on industrial visual intelligence

Abstract

The application relates to the technical field of image processing and power conductor detection, and discloses an overhead insulated conductor surface defect visual detection system based on industrial visual intelligence. The system comprises a multi-dimensional optical information acquisition unit, a self-adaptive environment sensing and enhancing unit, a multi-scale feature cooperative extraction unit, a heterogeneous defect fusion diagnosis unit and an edge execution and cloud synchronization unit. The system acquires images and depth information, carries out pretreatment through brightness distribution modeling and fuzzy compensation, extracts multi-scale features based on a deep neural network and an attention mechanism, and carries out intelligent diagnosis in combination with a knowledge base. The application solves the problems of poor defect detection robustness, low micro-defect precision and insufficient real-time performance in a complex environment, improves the recognition rate of hidden defects, reduces false positives and false negatives, and provides reliable support for power inspection.

Description

Technical Field

[0001] This invention relates to the fields of image processing and power conductor inspection technology, specifically to a visual inspection system for surface defects of overhead insulated conductors based on industrial visual intelligence. Background Technology

[0002] In the process of intelligent operation and maintenance of power systems, the safety monitoring and reliability assessment of power grid infrastructure have always been key links in ensuring a stable energy supply. Overhead insulated conductors, as an important physical carrier of the distribution network, have their surface integrity directly related to the insulation performance and service life of the lines. With the rapid development of industrial vision intelligence technology, using non-contact optical sensing and image processing methods to replace traditional manual inspections has become the core path for digital transformation in the power inspection field. This field covers visual information acquisition, transmission, and intelligent diagnosis in large-scale power grid environments, aiming to improve the perception accuracy of physical defects in transmission equipment through advanced algorithm models.

[0003] Visual inspection of surface defects in overhead insulated conductors is a specific direction for improving the efficiency of power distribution network operation and maintenance. This technology integrates high-definition imaging equipment and edge computing units to monitor the operating status of conductors in real time and uses computer vision algorithms to extract conductor textures from complex spatial backgrounds. Its basic principle lies in building a feature recognition model to automatically identify and warn of abnormal phenomena such as mechanical damage, aging cracking, and foreign object suspension that may occur in the conductor insulation layer, thereby reducing the risk of short circuits or power outages caused by conductor faults.

[0004] Current technologies still face significant challenges in practical applications. Traditional image processing methods often exhibit poor robustness in dynamic conditions such as complex lighting interference, background occlusion, and random cable swaying in the field, leading to inaccurate target extraction. Furthermore, the defects in insulated conductors exhibit significant diversity and scale variations. Existing detection algorithms have limited ability to capture features of minute cracks or hidden abrasions, easily resulting in missed detections or false alarms. In addition, there is a contradiction between the real-time processing requirements of massive amounts of high-resolution images and the limited computing power of inspection equipment, making it difficult for the system to achieve low-latency response while ensuring high detection accuracy, and lacking the ability to efficiently fuse and analyze heterogeneous defect data. Therefore, overcoming environmental noise interference and improving the automatic identification efficiency of multi-scale complex defects has become a pressing technical challenge in the field of intelligent inspection of overhead insulated conductors. Summary of the Invention

[0005] The purpose of this invention is to provide a visual inspection system for surface defects of overhead insulated conductors based on industrial visual intelligence, so as to solve the problems mentioned in the background art, such as poor robustness of surface defect detection of overhead insulated conductors, low accuracy of identification of small defects, and insufficient real-time performance of high-resolution image processing.

[0006] The technical solution of the present invention includes: The multi-dimensional optical information acquisition unit is used to simultaneously acquire serialized image data of overhead insulated conductors and corresponding spatial depth information through an integrated high-definition camera array and structured light source module during dynamic inspection.

[0007] The adaptive environment perception and enhancement unit is used to receive raw data transmitted by the multi-dimensional optical information acquisition unit. By establishing a brightness distribution model and a motion blur compensation algorithm, it performs consistent characterization processing on images affected by variable outdoor lighting and conductor sway.

[0008] A multi-scale feature collaborative extraction unit is used to construct a feature evolution framework based on deep neural networks. It extracts the texture topology, color variation features, and geometric contour deformation of the conductor surface under different receptive fields, and uses an attention weighting mechanism to accurately locate significant defect areas.

[0009] The heterogeneous defect fusion diagnostic unit is used to deeply fuse the feature vectors output by the multi-scale feature collaborative extraction unit. Combined with the preset defect knowledge base and morphological classification criteria, it identifies the type and severity of insulation layer damage, aging, cracks, and external foreign object suspension, and generates a quantitative diagnostic report.

[0010] The edge execution and cloud synchronization unit is used to realize local real-time reasoning of detection logic through the embedded computing platform deployed on the inspection terminal, and to synchronize key defect samples and diagnostic results to the remote monitoring cloud through the wireless transmission network for long-term status trend analysis.

[0011] Furthermore, the multi-dimensional optical information acquisition unit integrates three sets of high-speed industrial cameras positioned at 120-degree angles to each other, achieving comprehensive coverage of the entire circumference of the overhead insulated conductor. The structured light source module employs pulse width modulation technology for precise synchronization with the camera's shutter timing, ensuring high-contrast surface reflection images are obtained in both strong and low-light conditions. The spatial depth information is acquired in real-time via a time-of-flight sensor to correct for imaging scale deviations caused by fluctuations in the distance between the conductor and the camera.

[0012] In one embodiment of the present invention, the adaptive environment perception and enhancement unit first models the image background using a Gaussian mixture model in its processing flow. Then, it removes complex vegetation, buildings, and sky interference using background subtraction, retaining only the main area of ​​the conductor. To address uneven illumination, the unit calculates the local histogram statistical characteristics of the image and applies adaptive histogram equalization technology to enhance details in dark areas. For motion blur caused by conductor swaying, the unit utilizes instantaneous acceleration parameters fed back from the accelerometer to construct a Wiener filter restoration function, performing deconvolution repair on the degraded image to restore its edge sharpness.

[0013] Furthermore, the multi-scale feature collaborative extraction unit employs a hierarchical, progressive convolutional operation structure. In the bottom feature layer, the system uses a small convolutional kernel to capture minute cracks and minor abrasions on the surface of the insulated wire. In the middle feature layer, the system increases the receptive field through dilated convolution to perceive the distribution patterns of large-area insulation peeling or aging cracks. In the high feature layer, the system extracts semantic information about the overall abnormal orientation of the wire and the presence of hanging foreign objects. The attention weighting mechanism dynamically adjusts channel weights by calculating the contribution of each pixel in the feature map to defect classification, prioritizing algorithm resources in high-probability regions.

[0014] In one embodiment of the present invention, the heterogeneous defect fusion diagnostic unit constructs a multi-branch classification network to nonlinearly map the texture statistics, color space deviation, and geometric topological parameters of an image. The system has a pre-stored standard defect sample library of over 100,000 examples, including carbonization trajectories caused by discharge, aging and pulverization caused by ultraviolet radiation, and circumferential and longitudinal cracks caused by mechanical stress. The diagnostic process determines the specific attributes of the defect by calculating the Mahalanobis distance between the feature to be tested and the center points of each category in the sample library. For cases with multiple defects coexisting, the system uses evidence theory for conflict resolution, outputting a logically consistent comprehensive judgment conclusion.

[0015] Furthermore, the edge execution and cloud synchronization unit employs hardware acceleration technology. Through a tensor processing unit, it quantizes and prunes the trained neural network model, compressing the model parameter size to less than 15% of the original size while ensuring an accuracy loss of less than 1%. The system retains only critical inference logic locally and can process no less than 60 frames of 1080-pixel high-definition images per second. When a major security vulnerability is detected, the system immediately pushes a high-definition thumbnail containing geographic coordinates, timestamps, and defect levels to the backend via a 5G mobile communication network or a dedicated power grid wireless frequency band. It then initiates defect evolution prediction based on a long short-term memory network on the cloud server to assess the probability of failure of the defect within the next month.

[0016] As one embodiment of the present invention, the system also includes a closed-loop self-learning module. This module is used to collect false positives and false negatives reported by the edge terminals, incrementally optimize the model in the cloud using a semi-supervised learning algorithm, and periodically push the updated weight parameters to each inspection terminal to achieve continuous evolution of the system's detection performance.

[0017] Furthermore, the high-speed industrial camera in the multi-dimensional optical information acquisition unit has an autofocus function. Based on the real-time distance data provided by the time-of-flight sensor, it drives the lens group to move rapidly through a micro stepper motor to ensure that the imaging focal plane is always locked on the outer circumferential surface of the conductor, and its focusing response time is less than 50 milliseconds.

[0018] In one embodiment of the present invention, the adaptive environment perception and enhancement unit further includes a color deviation repair subunit. This subunit calculates the gain compensation coefficients for the red, green, and blue channels by detecting a standard white balance reference area in the image, in order to eliminate color deviations caused by different insulating wire materials or different color temperature light sources, thereby improving the accuracy of the heterogeneous defect fusion diagnostic unit in identifying color-sensitive defects.

[0019] Furthermore, the multi-scale feature collaborative extraction unit introduces a feature pyramid architecture during the feature extraction process. This architecture fuses high-level features with rich semantic information with low-level features with high-resolution detail information pixel by pixel through a top-down path and lateral connections. This structure ensures that the system can obtain sufficient spatial context support when detecting millimeter-level microcracks, reducing spurious feature interference caused by environmental noise.

[0020] In one embodiment of the present invention, the heterogeneous defect fusion diagnostic unit employs a contour envelope-based identification algorithm when determining foreign object suspension defects. The system first extracts the outer boundary of the suspected foreign object region and calculates its fractal dimension and roundness index. If the fractal dimension exceeds a preset threshold of 1.5 and the roundness significantly deviates from a standard circle, it is marked as an uncontrolled target such as a tree branch, kite string, or bird's nest, and infrared thermal imaging information is used to determine whether the foreign object has caused a local hot spot.

[0021] Furthermore, the edge execution and cloud synchronization unit also integrates an adaptive bitrate adjuster. Under fluctuating network bandwidth conditions, the adjuster dynamically adjusts the image compression ratio based on the current link quality. For defect-free, normal conductor areas, a high compression ratio is used to save bandwidth. For video segments with suspected defects, it automatically switches to lossless compression mode and enables multipath transmission to ensure the highest-resolution original visual information is available for cloud-based expert review.

[0022] As one embodiment of the present invention, the overhead insulated conductor surface defect visual inspection system based on industrial vision intelligence also includes a power management and status monitoring module. This module monitors the remaining power, internal temperature, and load rate of the core computing module of the inspection platform in real time. When the ambient temperature exceeds 65 degrees Celsius or the power is below 20%, the system automatically reduces the sampling frequency and switches to a low-power sleep mode, retaining only the basic environment-triggered wake-up function.

[0023] Furthermore, the structured light source module of the multi-dimensional optical information acquisition unit consists of 120 high-brightness light-emitting diodes, divided into four independent control quadrants. Based on feedback from the image sensor, the system dynamically adjusts the light source intensity in different quadrants, thereby forming a uniform diffuse reflection light field on the surface of the conductor, effectively suppressing metallic reflections or specular reflections from high-gloss paint surfaces.

[0024] In one embodiment of the present invention, the adaptive environment perception and enhancement unit also utilizes a combination algorithm of median filtering and bilateral filtering during image preprocessing. Median filtering is used to remove impulse noise from the image, while bilateral filtering, while smoothing the image background, can effectively preserve the gradient information of the crack edge of the insulated wire, avoiding the loss of details in the subsequent feature extraction process.

[0025] Furthermore, the attention weighting mechanism in the multi-scale feature collaborative extraction unit includes two dimensions: spatial attention and channel attention. The spatial attention module extracts the mask region where the main conductor is located by calculating the correlation between pixels. The channel attention module learns the nonlinear dependencies between feature channels through global average pooling and fully connected layers, giving higher response gain to channels carrying key defect information, thereby improving the system's detection sensitivity for fine scratches by more than 2 times.

[0026] As one embodiment of the present invention, the heterogeneous defect fusion diagnostic unit also introduces spatiotemporal consistency constraint logic. When processing continuous video streams, the system performs statistical voting on the detection results of the same conductor segment in multiple adjacent frames. Only when a certain defect feature is stably identified for more than 5 consecutive frames will the system confirm it as a real defect and trigger an alarm, effectively filtering out instantaneous false alarms caused by birds flying by or fleeting light and shadow.

[0027] Furthermore, the edge execution and cloud synchronization unit records the geographic information system coordinates of the inspection trajectory in real time during task execution. The system marks the detected defect locations on the power vector map, forming an intuitive defect distribution heat map. Based on this heat map, the cloud platform, combined with local meteorological disaster records and line operating load, uses association rule mining algorithms to analyze the root causes of defects and provides targeted replacement suggestions to the maintenance department.

[0028] As one embodiment of the present invention, the deep neural network model used in this system incorporates data augmentation techniques during the training phase. By randomly rotating, scaling, translating, and adding simulated rain and fog occlusion effects to the original images, the number of synthesized samples generated reaches over 500,000, significantly enhancing the model's generalization ability under extreme weather conditions.

[0029] Furthermore, the multi-dimensional optical information acquisition unit integrates an inertial measurement unit to sense the attitude angle changes of the inspection platform in real time. When a large roll or pitch motion is detected, the system automatically adjusts the camera's electronic shutter speed and, in conjunction with a digital image stabilization algorithm, controls the pixel offset of the conductor in the image to within 2 pixels, ensuring imaging stability under high-speed inspection conditions. Compared with the prior art, the beneficial effects achieved by the present invention are: 1. This solution integrates a multi-dimensional optical information acquisition unit and an adaptive environmental perception enhancement unit to construct a comprehensive dynamic compensation mechanism. Utilizing a pulse-synchronized structured light source and a motion blur restoration algorithm, it effectively overcomes the image quality degradation caused by complex lighting interference in the field and the swaying of the inspection platform. Compared to traditional single-view detection methods, this solution achieves high-resolution capture of the entire circumferential surface of the conductor, solving the robustness problem of target extraction in complex backgrounds from a physical perception perspective, and greatly reducing the interference of environmental noise on subsequent detection logic.

[0030] 2. This solution innovatively adopts a multi-scale feature collaborative extraction and heterogeneous defect fusion diagnostic architecture. Through the hierarchical perception capability and attention weighting mechanism of deep neural networks, it achieves accurate capture of multi-scale targets, ranging from micrometer-level cracks to decimeter-level foreign object suspension. The system no longer relies on single morphological features but comprehensively utilizes multi-dimensional information such as texture, color, and geometric topology, combined with a large-scale standard defect sample library for intelligent diagnosis. This combination of deep learning and prior knowledge significantly improves the system's accuracy in identifying hidden and subtle defects, reducing the false negative and false alarm rates to extremely low levels, providing reliable data support for preventative maintenance of power distribution networks.

[0031] 3. This solution achieves efficient allocation of computing resources and a closed-loop business process through edge execution and cloud synchronization units. Utilizing model quantization and compression technology, highly complex deep learning models are successfully deployed on embedded edge terminals, ensuring real-time local processing of massive amounts of high-definition images and resolving the issue of poor real-time performance caused by data transmission latency. Simultaneously, through the collaboration of real-time edge alerts and long-term state trend analysis in the cloud, a complete link from real-time perception to intelligent decision-making is constructed. The design of the closed-loop self-learning module ensures that the system can continuously self-optimize as the environment changes, exhibiting strong engineering applicability and technological leadership, and possessing significant economic and social value for promoting the digital transformation of intelligent power system inspection. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the overall technical solution architecture of the visual inspection system for surface defects of overhead insulated conductors based on industrial visual intelligence proposed in this invention. Figure 2 This is a schematic diagram of the core principle framework of the multi-scale feature collaborative extraction unit in this invention; Figure 3 This is a schematic diagram of the logic flow of the adaptive environment perception and enhancement unit in this invention; Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between the edge execution and cloud synchronization units in this invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Example 1 Please refer to the attached document. Figure 1 This embodiment discloses a visual inspection system for surface defects of overhead insulated conductors based on industrial visual intelligence. Built upon a modern power Internet of Things (IoT) inspection framework, this system aims to address the low recognition rate of minute surface damage on insulated conductors in complex field environments through high-dimensional optical perception and deep artificial intelligence algorithms. The system's overall operational architecture is logically rigorous, with efficient data flow between units via a high-speed industrial bus and wireless communication links.

[0035] Please refer to the attached document. Figure 1The multi-dimensional optical information acquisition unit, serving as the physical front-end of the entire system, is responsible for the synchronous capture of raw image data and spatial geometric information. During dynamic inspection, the high-definition camera array integrated within this unit employs three sets of high-speed industrial cameras positioned at 120-degree angles to each other. This surround layout ensures that the circumferential surface of the overhead insulated wire can be projected completely and without blind spots onto the photosensitive area of ​​the image sensor. The photosensitive chip of the high-speed industrial camera uses global shutter technology, which effectively suppresses motion tearing. Its maximum sampling frame rate reaches 120 frames per second, ensuring clear edge contours are still obtained even when the inspection platform moves at high speed. Deeply coupled to the high-definition camera array is a structured light source module, which consists of 120 high-brightness light-emitting diodes (LEDs) physically divided into four independent control quadrants. The structured light source module uses pulse width modulation (PWM) technology, and its modulation frequency is precisely synchronized with the electronic shutter timing of the high-speed industrial camera at the microsecond level through a hardware trigger signal. The moment the camera turns on the exposure, the structured light source module is driven by a pulsed current that is three times higher than that of continuous illumination, generating extremely high instantaneous illuminance. This creates a high-contrast reflection field on the surface of the conductor, allowing for stable surface texture images even under strong sunlight at noon or in complete darkness at night.

[0036] In the precision adjustment mechanism of the multi-dimensional optical information acquisition unit, the autofocus function is driven by real-time distance data provided by a time-of-flight sensor. This time-of-flight sensor operates in the 940nm infrared band, emitting pulsed light and measuring its round-trip flight time to calculate the absolute displacement deviation between the outer circumferential surface of the conductor and the focal plane of the lens in real time. This displacement data is fed back in real time to a micro-stepping motor driving the lens assembly, with its focusing response time strictly controlled within 50 milliseconds. This closed-loop control mechanism ensures that the imaging focal plane is always locked on the surface of the insulating layer, effectively avoiding defocus blur caused by conductor sag or vibration of the inspection platform. Furthermore, the unit also senses changes in the attitude angle of the inspection platform through an integrated inertial measurement unit. When a roll or pitch motion exceeds a preset threshold, the system automatically activates a digital image stabilization algorithm to control the pixel displacement deviation of the conductor in the image to within 2 pixels, ensuring consistency in subsequent algorithm processing.

[0037] Please refer to the attached document. Figure 3The adaptive environment perception and enhancement unit receives raw high-depth image data transmitted by the multi-dimensional optical information acquisition unit. Due to the extreme complexity of the field inspection environment, images are often affected by uneven lighting, cluttered backgrounds, and dynamic blurring. This unit first uses a Gaussian mixture model to dynamically model the image background. This model establishes five Gaussian distribution components for each pixel. Through long-term background learning, it accurately extracts vegetation, buildings, blue sky, or clouds from the background. Using background subtraction, the system can completely cut out the main area of ​​the conductor from the complex natural background, thereby significantly reducing the computational cost of subsequent neural networks. To address the problem of local overexposure or loss of detail in dark areas, this unit calculates the local histogram statistical characteristics of the image and applies adaptive histogram equalization technology. This technology dynamically adjusts the contrast mapping curve based on the grayscale distribution characteristics within the local window, ensuring optimal visibility of the minute textures on the insulation layer surface under various lighting conditions.

[0038] To address motion blur caused by conductor oscillation, the adaptive environment perception and enhancement unit utilizes instantaneous acceleration parameters fed back from the accelerometer, combined with the degradation mechanism of the optical system, to construct a precise Wiener filter restoration function. This function performs deconvolution restoration on the degraded image in the frequency domain, and its calculation logic is as follows:

[0039] in, This represents the spectrum of the restored image. This represents the spectrum of the original image containing blur interference. It is the degradation transfer function calculated based on the motion vector of the inspection platform, and This is the preset signal-to-noise ratio adjustment factor. Through the calculation of this formula, the system can compensate for pixel diffusion caused by wire oscillation, significantly restore the edge sharpness of the image, and make millimeter-level fine cracks clearly visible in the image again.

[0040] The adaptive environment perception and enhancement unit also includes a color deviation repair subunit. In imaging insulated wires at different times or with different materials, changes in the color temperature of the light source can cause color shifts in the image, interfering with the accuracy of defect identification. The color deviation repair subunit calculates the gain compensation coefficients for the red, green, and blue channels by detecting a standard white balance reference area in the image. The system first extracts reference pixels in the wire's shadow area or a known neutral gray area, and uses a total internal reflection white balance algorithm to correct the color matrix, thereby eliminating the negative impact of ambient color temperature on the determination of insulation aging and fading. Subsequently, the system uses a combination of median filtering and bilateral filtering algorithms for noise suppression. Median filtering effectively removes thermal noise and impulse noise generated by the photosensitive chip when operating at high temperatures, while bilateral filtering, while smoothing the background, effectively preserves the gradient information of the insulated wire crack edges based on a dual weight of spatial proximity and pixel value similarity, preventing defect features from being over-smoothed.

[0041] Please refer to the attached document. Figure 2 The multi-scale feature collaborative extraction unit is the core deep learning hub of the system. This unit constructs a feature evolution framework based on deep neural networks, employing a hierarchical, progressive convolutional operation structure to address defects with varying spatial spans. In the low-level feature extraction stage, the system utilizes small 3x3 convolutional kernels for high-density scanning, focusing on capturing minute cracks, minor abrasions, and tiny carbonization points caused by electrical sparks on the surface of the insulated wire. These low-level features possess extremely high spatial resolution and are crucial for identifying early-stage defects. In the mid-level feature layer, the system introduces dilated convolution technology, increasing the receptive field without increasing the number of parameters to perceive the distribution patterns of large-area insulation peeling, long-distance cracking, and material aging. In the high-level feature layer, the system extracts semantic information about the overall abnormal orientation of the wire, significant distortion of its geometric contour, and the presence of external foreign objects through multi-layer pooling and stride convolution.

[0042] To ensure sufficient contextual support for detecting millimeter-level microcracks, the multi-scale feature collaborative extraction unit introduces a feature pyramid architecture. This architecture upsamples high-level feature maps carrying rich semantic information through top-down paths and lateral connections, and then fuses them pixel-by-pixel with low-level feature maps containing high-resolution detail information. This multi-level feature coupling mechanism allows the system to see both global foreign object hanging contours and localized minute scratches. Simultaneously, this unit integrates an attention weighting mechanism, including a spatial attention module and a channel attention module. The spatial attention module generates a saliency probability map by calculating the spatial correlation between pixels, locking the algorithm's attention resources onto the mask region where the main conductor is located, automatically ignoring interference from surrounding trees or poles. The channel attention module compresses the information of each feature channel through a global average pooling layer, and then learns the non-linear dependencies between channels through two fully connected layers, assigning higher weight gains to channels carrying key defect information. Through this dual attention mechanism, the system's sensitivity to detecting minute scratches is improved by more than 2 times, effectively filtering out false features generated by complex backgrounds.

[0043] Please refer to the attached document. Figure 1 The heterogeneous defect fusion diagnostic unit performs deep fusion and decision-making on the high-dimensional feature vectors output by the multi-scale feature collaborative extraction unit. This unit internally stores a standard defect sample library of over 100,000 samples, covering carbonization trajectories caused by arc discharge, aging and pulverization caused by long-term ultraviolet radiation, circumferential and longitudinal cracks caused by mechanical stress, and external foreign objects such as snow, bird nests, and kite strings. The diagnostic logic is based on a multi-branch classification network, constructing a multi-dimensional defect judgment space by nonlinearly mapping the extracted texture statistics, color space deviation, and geometric topological parameters.

[0044] In the specific diagnostic process, the system determines the specific attributes of the defect by calculating the Mahalanobis distance between the feature vector to be tested and the centroids of each category in the sample database. The calculation logic is as follows:

[0045] in, Represents Mahalanobis distance, The currently detected feature vector, Let be the feature mean vector of a certain type of standard defect, and SS be the covariance matrix of the features of this type of defect. Compared with ordinary Euclidean distance, Mahalanobis distance considers the correlation between the feature dimensions and can more accurately describe the similarity between the target and the known defect category. For situations where multiple defects may coexist under complex working conditions, such as wires having both aging cracks and foreign objects hanging on them, the heterogeneous defect fusion diagnostic unit uses evidence theory to handle conflict resolution. The system treats the preliminary judgments output by different branch networks as independent sources of evidence, calculates the confidence of each judgment result through evidence synthesis rules, and thus outputs a logically consistent and self-contradictory comprehensive diagnostic report.

[0046] For foreign object suspension defects, the heterogeneous defect fusion diagnostic unit employs a contour envelope-based recognition algorithm. The system first extracts the outer boundary of the suspected foreign object region and calculates its fractal dimension and roundness index. If the fractal dimension exceeds a preset threshold of 1.5 and the roundness significantly deviates from a standard circle, it is marked as an uncontrolled target. At this point, the system also combines infrared thermal imaging information for multimodal verification to determine whether the foreign object has caused a local increase in contact resistance and triggered an overheating point. Furthermore, this unit introduces spatiotemporal consistency constraint logic. When processing continuous inspection video streams, the system statistically votes on the detection results of the same conductor segment in adjacent frames. Only when a defect feature is stably identified for more than 5 consecutive frames and its spatial location is logically consistent will the system confirm it as a genuine defect and trigger an alarm. This multi-frame confirmation mechanism effectively filters out instantaneous false alarms caused by birds flying rapidly by, flickering light and shadow, or sudden noise from sensors.

[0047] Please refer to the attached document. Figure 4 The edge execution and cloud synchronization unit achieves optimal allocation of computing resources. To meet real-time requirements, this unit is deployed on the embedded computing platform of the inspection terminal, employing hardware acceleration technology to deeply optimize the trained deep neural network model. Through tensor processing, the model is quantized and pruned, compressing the model parameter size from several hundred megabytes to less than 15% of its original size while ensuring an accuracy loss of less than 1%. This enables the system to process no less than 60 frames of 1080-pixel high-definition images per second locally, achieving real-time inference of the detection logic.

[0048] At the data flow level, the edge execution and cloud synchronization unit integrates an adaptive bitrate regulator. When bandwidth fluctuations occur in 5G mobile communication networks or dedicated power wireless frequency bands, the regulator dynamically adjusts the image compression ratio based on real-time quality feedback from the current link. When a normal conductor area is detected, the system uses lossy compression with a high compression ratio, retaining only basic inspection records to save expensive wireless traffic. Once a suspected defect is detected, the system immediately switches to lossless compression mode, initiates multipath transmission, and pushes a data packet containing precise geographic coordinates, timestamps, defect levels, and a high-resolution thumbnail of the defect area to the cloud. After receiving the data, the cloud server not only has it reviewed by experts but also activates a defect evolution prediction model based on a long short-term memory network. This model combines historical inspection data, local meteorological disaster records, and line load curves to analyze the deterioration rate of the defect under different humidity, temperature, and wind speed conditions, assesses its failure probability within the next month, and provides the maintenance department with scientific maintenance priority recommendations.

[0049] This system also includes a closed-loop self-learning module, which constructs a closed loop of knowledge evolution from the edge to the cloud. The system periodically collects false alarms and missed alarms reported by each inspection terminal, and uses these negative samples to incrementally train the model in the cloud using a semi-supervised learning algorithm. The updated model weight parameters are periodically pushed to each inspection terminal via wireless network, enabling continuous evolution of the system's detection performance. In addition, the power management and status monitoring module monitors the remaining battery power, internal core temperature, and embedded processor load rate of the inspection platform in real time. When the ambient temperature exceeds 65 degrees Celsius or the battery power falls below the 20% warning threshold, the system automatically activates a protection protocol, reducing power consumption by decreasing the image sampling frequency and shutting down non-core computing modules, retaining only basic environmentally triggered wake-up functions to ensure the device's survivability under extreme conditions.

[0050] The structured light source module in the multi-dimensional optical information acquisition unit also has dynamic light field adjustment capabilities. Based on the brightness histogram feedback acquired by the image sensor, the system adjusts the output intensity of the light source in four independent quadrants in real time. When severe metallic reflection or specular reflection caused by high-gloss paint appears on the surface of the conductor, the system automatically reduces the light source intensity at the corresponding angle and increases diffuse reflection compensation in other quadrants, thereby forming a uniform light field distribution on the imaging surface and effectively eliminating feature occlusion caused by reflection.

[0051] During task execution, the edge execution and cloud synchronization unit utilizes a fusion algorithm combining GPS and inertial navigation to record the geographic information system coordinates of the inspection trajectory in real time. The system dynamically marks the detected defect locations on a power vector map, creating an intuitive defect distribution heatmap. The cloud platform uses association rule mining algorithms to analyze the coupling relationship between defect distribution and geographical environment and climate region. For example, in coastal areas with high salt spray, the corrosion and aging frequency of insulation layers is significantly higher than in inland areas. Based on this, the cloud system generates targeted, differentiated maintenance strategies, greatly improving the intelligence level of power grid asset management.

[0052] The deep neural network model used in this system incorporates extremely high-intensity data augmentation techniques during the initial training phase. The research team randomly rotated, scaled, and transformed the contrast of tens of thousands of original guideline images, and used generative adversarial networks to synthesize simulated occlusion effects under rain, snow, dense fog, and sandstorms, resulting in a final training sample size exceeding 500,000 images. This large training set ensures the model has extremely strong generalization capabilities and maintains high robustness even under extremely harsh outdoor weather conditions.

[0053] The multi-scale feature collaborative extraction unit employs a gating mechanism-based fusion strategy to handle the interaction between low-level and high-level features during feature fusion. The gating mechanism dynamically determines the fusion ratio based on the noise level of the current image. In images with good lighting and high signal-to-noise ratio, the system assigns higher fusion weights to low-level detail features to achieve maximum accuracy in locating minute defects. Conversely, in low-light or heavily blurred images, the system automatically switches to a fusion mode primarily based on high-level semantic features, utilizing global topological logic to infer potential defect locations, thus ensuring detection continuity under various extreme conditions.

[0054] Example 2 Building upon Example 1, Example 2 further refines the hardware redundancy and algorithm anti-interference design of the industrial vision-based intelligent overhead insulated conductor surface defect visual inspection system under complex electromagnetic environments. Considering that the inspection platform is often surrounded by extremely high-intensity electromagnetic fields during high-voltage power grid inspections, the internal circuitry of the multi-dimensional optical information acquisition unit undergoes comprehensive electromagnetic shielding. All high-speed signal transmission lines use double-shielded twisted-pair cables and are equipped with opto-isolators for physical isolation between electrical signals and the computing system, preventing electrical noise interference from electromagnetic pulses on high-definition image sequences.

[0055] Please refer to the attached document. Figure 1In Example 2, the multi-dimensional optical information acquisition unit further incorporates a multispectral imaging module. In addition to a high-definition camera in the visible light band, this unit integrates a set of thermal imaging sensors operating in the long-wave infrared band. The thermal imaging sensors can detect localized temperature rises in the insulating layer due to internal discharge or leakage current. In the heterogeneous defect fusion diagnostic unit, the texture topology features under visible light and the hotspot distribution features under infrared light are spatiotemporally aligned and fused. The system further confirms the severity of the defect by calculating the overlap probability between the defect area identified in the visible light image and the high-temperature anomaly area in the thermal imaging image. For example, if the visible light image shows a slight crack in the insulating layer, while the corresponding infrared image shows a significant temperature abrupt change at that location, the system will automatically raise the alarm level for that defect to the highest level.

[0056] Please refer to the attached document. Figure 3 In the adaptive environment perception and enhancement unit, Example 2 employs a depth-prior-based dehazing algorithm to address common dense fog conditions in the field. This algorithm utilizes a convolutional neural network to predict the transmittance map of the image and combines it with an atmospheric scattering model to reconstruct clear conductor surface information. For rainy conditions, this unit integrates a rain removal subunit based on spatiotemporal features. By analyzing pixel motion vectors between consecutive frames, it identifies and removes scratch-like interference caused by raindrops rapidly passing through the lens, ensuring that the underlying features of the image are not obscured by meteorological noise.

[0057] Please refer to the attached document. Figure 2 In Example 2, the multi-scale feature collaborative extraction unit introduces deformable convolution technology. Traditional rectangular convolution kernels, when processing irregularly shaped defects such as spreading cracks or twisted foreign objects, have fixed sampling points, making it difficult to perfectly fit the target shape. Deformable convolution, by learning additional pixel offsets, allows the sampling position of the convolution kernel to adaptively deform according to the actual shape of the defect, thereby more accurately extracting the topological structure of irregular defects and further improving the system's ability to capture targets with complex shapes.

[0058] Please refer to the attached document. Figure 1 In Embodiment 2, the heterogeneous defect fusion diagnostic unit adds correlation tracking logic for dynamic targets. When the inspection platform traverses a continuous defect area at high speed, the system establishes a unique dynamic sequence number for each detected defect point. The Kalman filter algorithm is used to predict the pixel coordinates of the defect in the next frame, combined with the Hungarian algorithm for cross-frame matching. This correlation tracking ensures that for continuous scratches existing over long distances, the system only performs a logical report once and automatically calculates the total length of the scratches and the cumulative damaged area in the diagnostic report, providing a more accurate material consumption assessment for later maintenance.

[0059] Please refer to the attached document. Figure 4In Example 2, the edge execution and cloud synchronization unit employs a distributed collaborative computing mode. When multiple inspection robots simultaneously perform tasks in a certain area, each edge terminal shares computing resources through a dedicated power wireless network. If one terminal experiences excessive computational load due to processing complex video streams, it can split some feature extraction tasks and send them to surrounding terminals with lower loads for collaborative processing. Regarding cloud synchronization, Example 2 introduces a defect data storage system based on blockchain technology. Each detected defect record, diagnostic report, and corresponding original image hash value are written into a distributed ledger. This ensures the authenticity and immutability of the inspection data, providing legally valid original data support for subsequent liability determination and equipment life assessment.

[0060] Example 2 further expands the functionality of the power management and status monitoring module. This module integrates a piezoelectric energy harvesting device, capable of powering small sensors using induced electrical energy generated by the mechanical vibration or strong electromagnetic field during the inspection platform's movement. Simultaneously, the system establishes a digital twin-based operational status model of the inspection terminal, synchronously simulating the wear and tear of each hardware unit in the cloud. When the system detects a trend of declining image signal-to-noise ratio in the high-definition camera due to sensor aging, it automatically issues a hardware warning, prompting maintenance personnel to replace the corresponding sensing module during the next maintenance cycle.

[0061] The adaptive environment perception and enhancement unit also includes a dynamic contrast enhancement subunit specifically designed for handling low-contrast imaging during rainy days or twilight. This subunit independently controls contrast gain at different spatial frequency levels by constructing a Laplacian pyramid. The high-frequency levels focus on enhancing the perception of minute bumps and depressions on the conductor surface, while the low-frequency levels focus on maintaining overall image brightness stability. This processing approach makes the image visually more layered, greatly assisting the multi-scale feature collaborative extraction unit in recognizing shallow scratches.

[0062] The multi-scale feature collaborative extraction unit adds a global context aggregation module at the top of the feature pyramid. This module extracts statistical global information of the entire image through a large-span global pooling operation and feeds it back as a bias term to the feature extraction branches at each scale. This helps the system identify abnormal logic in the background. For example, when similar textures appear throughout the entire field of view, the system can determine that it is lens contamination rather than a large-area wire defect, thus eliminating systematic false alarms caused by environmental factors at the root.

[0063] The heterogeneous defect fusion diagnostic unit also incorporates a defect growth algorithm based on topological connectivity during processing. When the system locates a suspected crack point on the surface of the insulation layer, the algorithm uses that point as a seed and automatically extends it in its neighborhood according to the gradient descent direction. If the grown path conforms to the morphological characteristics of a physical crack, it is confirmed as a real defect. This automatic point-to-line growth mechanism significantly improves the system's detection accuracy for slender cracks hidden in complex background textures.

[0064] Example 2's edge execution and cloud synchronization unit further refines the long-term state trend analysis function. The cloud system aggregates all inspection data from a specific area over the past three years to establish a spatiotemporal evolution model of insulated conductor aging. This model can identify areas with abnormally accelerated aging rates and automatically correlate them with surrounding industrial pollution sources or extreme weather frequencies. Through this deep mining of big data, maintenance departments can shift from reactive emergency repairs to predictive, precise maintenance, significantly reducing the overall cost of power grid operation.

[0065] In the structured light source control of the multi-dimensional optical information acquisition unit, Example 2 employs an adaptive dimming strategy based on image entropy. The system calculates the local entropy value of the acquired image in real time. If the entropy value of a certain area is too low, it indicates severe loss of image details. At this time, the system automatically fine-tunes the incident angle and intensity of the light source in that quadrant until the image entropy value is maximized, ensuring that each image uploaded to the algorithm unit carries the richest surface physical information. This refined light source adjustment capability enables the system to exhibit extremely high adaptability and robustness when dealing with insulated wires made of materials with extreme optical properties such as high reflectivity and strong absorption.

[0066] Through the detailed descriptions of Embodiments 1 and 2 above, the visual inspection system for surface defects of overhead insulated conductors based on industrial visual intelligence constructed by this invention has achieved extremely high levels in terms of perception dimension, processing depth, and business closed-loop capability. The system utilizes advanced hardware synchronization technology to solve imaging quality issues, a multi-scale neural network architecture to solve feature extraction issues, fusion diagnostic logic to solve judgment accuracy issues, and edge computing and cloud synchronization technologies to solve real-time and intelligent management issues. The specific algorithms, parameter values, hardware configurations, and logical flows mentioned in each embodiment are all for the optimal transformation of the inventive concept and should not be considered as limitations on the scope of protection of this invention. In actual engineering implementation, conventional equivalent substitutions and local optimizations of the above technical details, based on specific voltage levels, inspection environments, and computing power cost requirements, should all be included within the scope of the claims of this invention.

[0067] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A visual inspection system for surface defects of overhead insulated conductors based on industrial visual intelligence, characterized in that, include: The multi-dimensional optical information acquisition unit is used to acquire serialized image data of the entire circumference surface of the overhead insulated conductor during dynamic inspection by integrating three sets of high-speed industrial cameras at 120-degree angles to each other. It uses a structured light source module composed of 120 high-brightness light-emitting diodes to provide pulsed synchronous illumination and works with a time-of-flight sensor to acquire spatial depth information between the conductor and the camera. The adaptive environment perception and enhancement unit is used to dynamically model the image background using a Gaussian mixture model, extract the main body region of the conductor by background subtraction, perform adaptive histogram equalization based on the statistical characteristics of the local histogram of the image, and use acceleration parameters to construct a Wiener filter restoration function to perform deconvolution repair on the degraded image affected by the sway of the conductor. A multi-scale feature collaborative extraction unit is used to construct a feature evolution framework based on deep neural networks. It uses a hierarchical and progressive convolutional operation structure to extract the texture topology, color variation features and geometric contour deformation of the conductor surface under different receptive fields, and uses an attention weighting mechanism that includes spatial and channel dimensions to accurately locate significant defect areas. The heterogeneous defect fusion diagnostic unit is used to perform deep fusion of the feature vectors output by the multi-scale feature collaborative extraction unit. It combines the preset standard defect sample library to identify the type and severity of insulation layer damage, aging, cracks and external foreign object suspension. It determines the defect attributes by calculating the Mahalanobis distance between feature vectors and uses evidence theory to handle the conflict of multiple defects. The edge execution and cloud synchronization unit is used to quantize and prune the model using the tensor processing unit, perform local real-time inference on the inspection terminal, and synchronize key defect samples, diagnostic results and geographic information system coordinates to the remote monitoring cloud through the wireless transmission network. It also uses the long short-term memory network to perform long-term defect evolution prediction. The adaptive environment perception and enhancement unit also includes a color deviation correction subunit, used for: Reference pixels are extracted from the main area of ​​the conductor, and the gain compensation coefficients of the red, green and blue channels are calculated using the total internal reflection white balance algorithm to eliminate the interference of ambient color temperature on the determination of the insulation layer color. Median filtering is used to remove impulse noise from the image, and bilateral filtering is used to smooth the image background while preserving the gradient information of the crack edge of the insulated wire.

2. The visual inspection system for surface defects of overhead insulated conductors based on industrial visual intelligence according to claim 1, characterized in that, The multi-dimensional optical information acquisition unit also includes: Miniature stepper motors are used to drive the movement of lens groups in high-speed industrial cameras. The multi-dimensional optical information acquisition unit performs automatic focusing in the following ways: acquiring real-time distance data provided by the time-of-flight sensor; calculating the absolute displacement deviation between the outer circumferential surface of the conductor and the focal plane of the lens; and driving a micro stepper motor based on the absolute displacement deviation to control the focusing response time to within 50 milliseconds. The multi-dimensional optical information acquisition unit also uses an integrated inertial measurement device to sense the attitude angle change of the inspection platform. When the attitude angle change exceeds a preset threshold, a digital image stabilization algorithm is used to control the pixel offset of the wire in the image to within 2 pixels.

3. The visual inspection system for surface defects of overhead insulated conductors based on industrial visual intelligence according to claim 1, characterized in that, The specific process of Wiener filtering restoration performed by the adaptive environment perception and enhancement unit is as follows: Obtain the spectrum of the original image containing blur interference; Calculate the degradation transfer function based on the motion vector of the inspection platform; A gain compensation operator is constructed by utilizing the proportional relationship between the square of the magnitude of the degradation transfer function and a preset signal-to-noise ratio adjustment factor; The original image spectrum, the reciprocal of the degradation transfer function, and the gain compensation operator are multiplied together to obtain the restored image spectrum.

4. The visual inspection system for surface defects of overhead insulated conductors based on industrial visual intelligence according to claim 1, characterized in that, The multi-scale feature collaborative extraction unit also utilizes a feature pyramid architecture to perform feature fusion: Upsampling of high-level feature maps is performed using a top-down approach; The upsampled high-level feature map is fused pixel-by-pixel with the low-level feature map containing high-resolution detail information by lateral connection. At the top of the feature pyramid, the global context aggregation module is used to extract statistical information of the entire image, and the statistical information is fed back as a bias term to the feature extraction branches at each scale.

5. A visual inspection system for surface defects of overhead insulated conductors based on industrial visual intelligence according to claim 1, characterized in that, The execution process of the attention weighting mechanism is as follows: The spatial attention module is used to calculate the spatial correlation between pixels and generate a saliency probability map for locking the main body of the wire; The channel attention module compresses the information of each feature channel through a global average pooling layer, and learns the non-linear dependencies between channels through a fully connected layer, assigning weight gain to the channels carrying key defect information.

6. The visual inspection system for surface defects of overhead insulated conductors based on industrial visual intelligence according to claim 1, characterized in that, The specific process by which the heterogeneous defect fusion diagnostic unit calculates the Mahalanobis distance is as follows: Obtain the difference vector between the feature vector to be tested and the feature mean vector of the standard defect; Obtain the inverse of the covariance matrix of the standard defect features; Perform matrix multiplication operations on the transpose of the difference vector, the inverse of the covariance matrix, and the difference vector in sequence; The square root of the calculation result is performed to obtain the Mahalanobis distance value used to determine the defect attribute.

7. A visual inspection system for surface defects of overhead insulated conductors based on industrial visual intelligence according to claim 1, characterized in that, The heterogeneous defect fusion diagnostic unit is also used for: Extract the outer boundary of the suspected foreign object region and construct its contour envelope; Calculate the fractal dimension and roundness index of the contour envelope. When the fractal dimension exceeds 1.5 and the roundness deviates from the standard circle, the suspected foreign object area is marked as an uncontrolled target. A spatiotemporal consistency constraint logic is introduced to statistically vote on the detection results of the same section of wire in a continuous video stream. When the defect feature is stably identified for more than 5 consecutive frames, an alarm signal is triggered.

8. A visual inspection system for surface defects of overhead insulated conductors based on industrial visual intelligence according to claim 1, characterized in that, The edge execution and cloud synchronization unit also includes an adaptive bitrate adjuster for: The image compression ratio is dynamically adjusted based on the link quality feedback of the current wireless transmission network. Lossy compression is performed when a normal conductor region is detected. When a suspected defect is detected, switch to lossless compression mode and enable multipath transmission.

9. A visual inspection system for surface defects of overhead insulated conductors based on industrial visual intelligence according to claim 1, characterized in that, Also includes: The closed-loop self-learning module is used to collect false alarm cases and missed alarm cases reported by the inspection terminal. It incrementally optimizes the model in the cloud through a semi-supervised learning algorithm and pushes the updated weight parameters to the inspection terminal regularly. The power management and status monitoring module is used to monitor the remaining power, internal temperature and computing module load rate of the inspection platform in real time. When the internal temperature exceeds 65 degrees Celsius or the remaining power is less than 20%, the detection system reduces the sampling frequency and switches to a low-power sleep mode.

Images