A low-voltage wire defect detection method and system based on machine vision

Abstract

The application discloses a kind of low-voltage wire defect detection method and system based on machine vision, it is related to machine vision technical field, the method includes: obtaining wire full circumference diameter surface image and executing adaptive brightness correction to eliminate light spot;Periodic texture and non-periodic defect characteristic signal are realized by multi-scale structure tensor decomposition Physical peeling;Deep cascade discriminant model is constructed to carry out defect positioning and classification determination, and according to the result output early warning instruction;The system includes multi-view image acquisition module and is deployed in the back-end processing unit of automated assembly line.The application can effectively eliminate high reflection background interference, significantly reduce the computing overhead while ensuring the detection accuracy, meet the real-time monitoring and structural defect accurate discrimination needs in high flow rate wire production scene by using the above technical scheme.

Description

Technical Field

[0001] This invention relates to the field of machine vision, and more particularly to a method and system for detecting defects in low-voltage conductors based on machine vision. Background Technology

[0002] Machine vision inspection technology, as a core branch of industrial automation, plays an increasingly important role in the quality monitoring and intelligent inspection of power equipment. Non-contact inspection of surface defects in low-voltage conductors has become a crucial link in ensuring the stable operation of power supply systems. As the largest and most widely distributed conductive carrier in the power distribution network, the integrity of the insulation layer and the structural stability of low-voltage conductors during their production process directly determine the cost and safety of subsequent operation and maintenance. By using image acquisition devices to capture the surface morphology of conductors and implementing intelligent analysis, the aim is to replace traditional manual visual identification with algorithms, thereby achieving automatic qualitative and quantitative monitoring of conductor quality defects under complex operating conditions.

[0003] However, in current implementations, a serious problem of feature coupling and semantic interference exists between the complex periodic textures and minute physical defects on the surface of multi-stranded low-voltage conductors with high reflectivity. The multi-strand strands on the conductor surface create dense shadow areas of varying depths, and the metal or insulating materials easily produce irregular specular reflections under strong light. This causes the algorithm to easily misinterpret normal strand boundaries as structural cracks or lose true defect feature signals in reflection blind zones, resulting in missed detections and false alarms for minute defects. This strong background noise caused by the interplay of material reflection and structural texture forces existing solutions to introduce extremely deep or parameter-intensive nonlinear computational models to improve recognition sensitivity. This not only significantly increases the computational load of the backend processing system but also causes the algorithm's response time to exceed the real-time threshold for online detection. Ultimately, this leads to a technical imbalance between improving detection accuracy and ensuring real-time performance, making it difficult to meet the stringent requirements for instantaneous defect capture and precise interception in high-volume, high-flow-rate conductor production scenarios.

[0004] This invention achieves the separation of periodic stranded texture and non-periodic defect features of wires through multi-scale structural tensor decomposition logic. It weakens specular reflection spots under strong light by using an adaptive brightness correction algorithm, effectively solving the interference of shadow areas and reflection spots on feature extraction of highly reflective material surfaces. In response to potential safety hazards such as loose stranding and broken strands, the system constructs a dedicated feature extraction logic and morphological discrimination mechanism. To reduce the computational load of the system, a lightweight deep cascade model is introduced, which ensures high resolution while significantly reducing the system's computational cost. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for detecting defects in low-voltage conductors based on machine vision, in order to solve the problems of existing technologies, such as difficulty in distinguishing periodic textures from small physical defects, high computational load, and insufficient detection accuracy.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] A machine vision-based method for detecting defects in low-voltage power lines includes the following specific steps:

[0008] S1: The multi-view image acquisition module acquires the full circumference surface image of the low-voltage conductor under motion. The multi-view image acquisition module contains multiple linear array image acquisition units that are uniformly distributed along the radial direction of the conductor, which are used to synchronously capture the multi-dimensional visual signals of the conductor during high-speed transmission.

[0009] S2: Perform adaptive brightness correction and denoising preprocessing on the full circumference surface image. Eliminate specular reflection spots generated by the metal surface and insulation layer of the conductor by constructing a local brightness mapping function, and extract the region of interest image of the conductor body using an improved edge operator.

[0010] S3: Executes feature separation logic based on multi-scale structural tensor decomposition. By combining the region of interest image and the gray-scale 3D model of the image gray level, it calculates the covariance matrix of the local gradient of the image and performs eigenvalue decomposition to physically separate the periodic stranded texture signal of the low-voltage conductor surface from the non-periodic defect feature signal.

[0011] S4: Construct a deep cascaded discrimination model, utilize the feature pyramid structure to fuse multi-level semantic information, extract features and characterize morphological parameters of defect feature signals, and realize automatic location and classification of insulation layer damage, scratches and structural strand breakage defects.

[0012] S5: Generate a defect distribution difference map based on the discrimination results, and output an inspection report and real-time early warning control instructions in combination with the preset quality and safety thresholds.

[0013] Preferably, in step S2, the adaptive brightness correction process uses a dynamic threshold algorithm based on voxel grid sampling, sets the voxel grid size to a preset size, and removes high-frequency noise points in the image through a statistical filtering algorithm, wherein the number of neighboring points is set to a first preset threshold and the standard deviation multiple is set to a second preset threshold.

[0014] Preferably, in step S3, the scale factor of the multi-scale structural tensor is configured within a preset range, and the eigenvalue response threshold satisfies the preset proportional relationship between each eigenvalue. The precise capture of microcracks in the fine insulating layer is achieved by extracting the evolution gradient of the tensor trace.

[0015] Preferably, in step S3, the feature separation logic is achieved by comparing the magnitude relationship between feature values. The image of the region of interest is combined with the image's grayscale to form a grayscale 3D model. The covariance matrix of the local gradient of the grayscale 3D model is calculated. This matrix describes the distribution law of grayscale changes within the local region, where grayscale value is the mathematical term for brightness in the algorithm processing, and the two are numerically equivalent. Eigenvalue decomposition is performed on the covariance matrix to obtain feature values ​​λ1, λ2, and λ3. λ1 ​​reflects the intensity of grayscale changes perpendicular to the stranding texture direction. In normal texture areas, the value of λ1 is large because there are dramatic and regular alternations of light and dark from convex to concave and back to convex again perpendicular to the stranding texture direction. However, in defective areas, λ1 shows randomness in direction and value, without dramatic and regular changes. λ2 is perpendicular to the λ1 direction, that is, along the stranding direction of the conductor. The value of λ2 in normal areas is small because the conductor... The grayscale changes along the stranded texture are very smooth, but for defects, the grayscale changes are more drastic, and λ2 will increase significantly. The direction of λ3 is perpendicular to the plane of λ1 and λ2. When the conductor surface has a normal periodic stranded texture, the grayscale changes in this local area are highly concentrated in the plane spanned by λ1 and λ2, with almost no dispersion in the third direction. The data points present a clean line or flat planar structure, and the value of λ3 is extremely small or close to zero. However, when defects such as burrs, barbs, and broken strands appear on the conductor surface, the originally clean line structure is destroyed, and the data points also produce a non-negligible dispersion in the third direction. The structure of the local area must be described by introducing a third dimension, and λ3 will increase significantly. By comparing the numerical relationship between λ1, λ2, and λ3, the periodic stranded texture signal of the low-voltage conductor surface and the non-periodic defect feature signal are physically separated into blind sources.

[0016] Preferably, in step S4, the depth-cascaded discrimination model uses an octree structure of preset depth for multi-scale spatial modeling, and reconstructs the geometric shape of the defect through Laplacian smoothing and hole repair logic, and performs least squares plane fitting repair on geometrically abnormal areas with geometric dimensions greater than a preset threshold.

[0017] Preferably, the classification process uses a support vector machine classifier based on radial basis kernel function to perform nonlinear mapping on the extracted multidimensional feature vectors, including area, perimeter, roundness, depth and edge curvature, to identify the defect category. The classification accuracy is limited to meet a preset accuracy threshold.

[0018] Preferably, the method also involves establishing a low-voltage conductor quality database containing historical test records of a preset scale, using a long short-term memory neural network to perform trend analysis and reliability prediction on the structural stability of the conductor, and predicting that the lead time meets a preset time threshold.

[0019] A machine vision-based low-voltage conductor defect detection system includes the following modules:

[0020] Multi-view image acquisition module: used to acquire full-circumference surface images of low-voltage conductors in motion. The multi-view image acquisition module contains multiple linear array image acquisition units evenly distributed along the radial direction of the conductor. It uses a blue narrowband light source with a preset center wavelength as the active illumination source. The effective pixel count of a single image acquisition unit meets the preset pixel threshold, and the scanning accuracy is limited to a preset accuracy range.

[0021] Central parallel processing module: Executes all core algorithms and completes the entire process from image to defect determination result.

[0022] Automated actuators: including industrial robot arms, used to perform handling operations based on real-time early warning and control commands output by a central parallel processing module.

[0023] Preferably, the system is deployed on an automated power equipment testing production line, where an industrial robot arm performs the gripping and guiding of the wire. The robot's repeatability and positioning accuracy meet the preset accuracy requirements, the positioning error of the vision guidance system is controlled within the preset error range, and the testing time for a single wire segment is limited to below a preset time threshold.

[0024] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0025] This invention achieves the physical separation of the periodic stranded texture features and non-periodic minute defect features unique to wires through multi-scale structural tensor decomposition logic, effectively solving the problem of interference from high reflectivity materials and shadow areas of varying depths on feature extraction, and significantly reducing the missed detection rate of minute defects.

[0026] This invention employs a preprocessing algorithm based on adaptive brightness correction and a lightweight depth cascade model, which significantly reduces the computational overhead of the back-end processing system while ensuring high-resolution image processing performance, thus meeting the real-time monitoring needs of low-voltage conductors in high-volume, high-flow-rate production scenarios.

[0027] This invention addresses the unique multi-strand strand structure of low-voltage conductors by constructing a specialized feature extraction logic and morphological discrimination mechanism. This mechanism can not only accurately locate physical damage on the surface of the insulation layer, but also effectively identify structural safety hazards such as loose strands and broken strands, thereby improving the operational safety of the power supply system. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the overall technical solution architecture of a machine vision-based low-voltage conductor defect detection method proposed in this invention.

[0029] Figure 2This is a schematic diagram of the core principle framework of the feature separation logic based on multi-scale structural tensor decomposition in this invention;

[0030] Figure 3 This is a flowchart illustrating the logical process of defect extraction and localization classification based on a deep cascaded discriminant model in this invention.

[0031] Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow based on quality database and trend analysis in this invention. Detailed Implementation

[0032] 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.

[0033] Example 1

[0034] This embodiment provides a machine vision-based low-voltage conductor defect detection system and method. The system is deployed on a high-precision automated power equipment inspection production line and aims to solve the technical problems of low defect detection accuracy, poor real-time performance, and easy obscuring of minor damage by interference signals caused by the high reflectivity of the surface and the complex stranding structure of low-voltage conductors during high-speed continuous production.

[0035] In terms of system architecture, this embodiment constructs a comprehensive testing platform that integrates precision mechanical transmission, multi-dimensional optical imaging, parallel computing processing, and automated closed-loop control. In physical space, the system mainly consists of a workpiece conveying and guiding module, a multi-view image acquisition module, an active lighting source module, a central parallel processing module, and an automated actuator.

[0036] First, the workpiece conveying and guiding module uses a set of guide wheels with precision tension control function to guide the low-voltage wire to be tested through the detection center of the multi-view image acquisition module at a constant speed. In order to ensure the positional stability during the detection process, the repeatability of the guide wheel set is calibrated to ±0.01 mm.

[0037] Secondly, the multi-view image acquisition module, as the core visual sensor front end of this system, strictly follows the principle of radial symmetry in its physical layout. Specifically, the module contains four sets of linear image acquisition units, which are evenly distributed at a 90-degree angle along the radial circumference of the conductor. This ensures that the entire circumference of the low-voltage conductor in motion can be covered. Each image acquisition unit uses a high-resolution CMOS linear image sensor with an effective pixel count of no less than 5 million. The photosensitive size of each pixel has been miniaturized, so that the scanning accuracy of the system can be stabilized at an extremely high level of ±1 micrometer at the rated object distance. This high-precision sensor selection directly supports the subsequent imaging and capture capability of micro-cracks on the surface of the conductor.

[0038] The accompanying active lighting module uses a blue narrowband LED light source with a center wavelength of 450 nanometers. Compared with traditional white light sources, the blue narrowband light source has stronger scattering characteristics on the surface of metal and composite material insulation layers, which can effectively enhance the contrast between the extremely fine texture of the conductor surface and the background. The front end of the light source module is equipped with a corresponding narrowband filter, which only allows light signals within a range of ±10 nanometers of the center wavelength to pass through, thereby filtering out ambient light interference in the production line environment and improving the signal-to-noise ratio.

[0039] At the data transmission level, the image acquisition unit is connected to the central parallel processing module via the GigE Vision gigabit Ethernet protocol or the high-speed CameraLink interface. The central parallel processing module consists of a high-performance industrial server, which is equipped with a multi-core central processing unit and a graphics processing unit accelerator card with powerful floating-point computing capabilities. The module has a dedicated image processing pipeline and a deep learning inference engine deployed inside to execute subsequent complex feature separation and discrimination algorithms.

[0040] In terms of the system's workflow, this embodiment follows a logical closed loop from physical signal acquisition to digital signal analysis, and then to intelligent decision output. The specific method steps are as follows:

[0041] Step S1: Multi-view image acquisition process. When the low-voltage conductor enters the detection area under the traction of the workpiece conveying and guiding module, the synchronous controller sends a synchronous trigger command to the four linear array image acquisition units according to the preset encoder pulse signal. Under the continuous illumination of the 450-nanometer blue narrow-band light source, each acquisition unit synchronously captures the multi-dimensional visual signals of the conductor during high-speed transmission. Due to the use of linear array scanning, the system can project the surface of the continuously moving conductor into a large-format, high-resolution full-circumference unfolded image. Each line of scanning data corresponds to a tiny fault in the radial direction of the conductor. Through the combination of multiple sets of sensors, the complete digital reconstruction of the full-circumference surface features of the conductor is realized.

[0042] Step S2: Adaptive Brightness Correction and Denoising Preprocessing. The acquired raw image signal is first sent to the preprocessing unit of the central parallel processing module. For the irregular specular reflection spots generated by the metal surface and insulation layer of the conductor, i.e., the "flare" phenomenon, the preprocessing unit executes a dynamic thresholding algorithm based on voxel grid sampling. In this algorithm, the processing unit maps the two-dimensional image data to a virtual three-dimensional perception space and sets the voxel grid size to 0.05 mm x 0.05 mm x 0.05 mm. By constructing a local brightness mapping function, the processor calculates the average light intensity distribution within each grid and dynamically adjusts the compensation coefficient according to the brightness gradient of the surrounding pixels, thereby eliminating the occlusion of the underlying features by the spot. Subsequently, a statistical filtering algorithm is used to remove high-frequency noise points in the image caused by sensor thermal noise or environmental dust. The number of neighboring points is set to 50, and the standard deviation multiple is set to 2.0. On this basis, an improved edge operator is used to accurately extract the region of interest image where the conductor body is located, remove invalid background data, and reduce the subsequent computational burden.

[0043] Step S3: Feature separation process based on multi-scale structural tensor decomposition. This is the core innovation of this embodiment at the algorithm level. The processor performs structural tensor decomposition logic on the region of interest image, combines the region of interest image with the image's grayscale to generate a grayscale 3D model, and obtains the covariance matrix by calculating the local gradient of each pixel in the grayscale 3D model. This matrix describes the distribution law of grayscale changes in the local area. The processor performs eigenvalue decomposition to obtain the eigenvalues ​​λ1, λ2, and λ3 of the matrix. In this embodiment, by configuring the multi-scale scaling factor in the range of 0.03 to 0.1 mm, the sensitivity adjustment of textures of different sizes is realized. The algorithm distinguishes normal textures from abnormal defects by identifying the specific proportional relationship between eigenvalues, that is, λ1 is greater than λ2 and λ2 is greater than λ3 multiplied by 1.5. The normal periodic stranded texture on the surface of the conductor shows a high degree of directional consistency in the tensor space, while defects such as cracks and broken strands show directional randomness or abrupt changes in gradient. Through this physical-level blind source stripping technology, the system can strip the stranded texture signal in a strong background and retain only the defect feature signal that reflects the real physical damage.

[0044] Step S4: Deep Cascaded Discriminant and Morphological Reconstruction Process. The extracted defect feature signals are fed into a deep cascaded discriminant model. This model utilizes a feature pyramid structure to fuse multi-level semantic information, taking into account both macroscopic structural strand breakage defects and microscopic insulation layer microcracks. The model employs a depth of 9 octrees for multi-scale spatial modeling, enabling extremely detailed representation of the defect's geometric morphology. During the identification process, the system pre-repairs the defect area using Laplacian smoothing and hole repair logic to eliminate fragmented areas caused by light and shadow interference. For geometrically abnormal regions with a diameter greater than 0.2 mm, the processor performs least-squares plane fitting to further pinpoint the edge contour of the defect. Subsequently, the system extracts a 5-dimensional feature vector including area, perimeter, roundness, depth, and edge curvature, and inputs it into a support vector machine classifier based on radial basis kernel function. This classifier divides the complex feature space into multiple hyperplanes through nonlinear mapping, thereby achieving automatic localization and classification of different defect categories such as insulation layer damage, scratches, broken strands, and extrusion. Its measured classification accuracy is stable at over 99%.

[0045] Step S5: Quality assessment and early warning output process. The central processing module generates a defect distribution difference map in real time based on the classification results and performs multi-level assessments in conjunction with preset quality and safety thresholds. If the detected defect indicators exceed the safety red line, the system immediately outputs real-time early warning control commands to the automated actuator. In this embodiment, the automated actuator includes an industrial robot arm with a repeatability accuracy of ±0.01 mm. Based on the positioning information from the vision guidance system, the positioning error of the robot arm is strictly controlled within 50 micrometers. It performs the grasping, rejection, or marking operations of the wires. At the same time, all detection data is synchronized in real time to the low-voltage wire quality database. This database stores more than 100,000 historical records and uses an internally deployed long short-term memory neural network to predict the long-term trend of the structural stability of the wires. This prediction mechanism can predict the potential aging trend of production equipment or the quality fluctuation of material batches 500 hours in advance, thereby realizing the leap from "post-event detection" to "pre-event prevention".

[0046] The entire inspection process exhibits extremely high timeliness. Thanks to the parallel architecture of the hardware and the linear optimization of the algorithm, the time required for the complete inspection of a single wire segment, including image acquisition, algorithm processing and instruction issuance, is limited to less than 15 seconds, which fully meets the high-speed production requirements of industrial-grade production lines.

[0047] Example 2

[0048] Based on Example 1, this example further details an optimized implementation scheme for specific ultra-fine diameter low-voltage conductors in extreme environments. Due to their smaller diameter and tighter stranding texture on the surface, ultra-fine diameter conductors place more stringent requirements on the spatial resolution and noise suppression capabilities of the vision system.

[0049] At the hardware configuration level, this embodiment has made secondary enhancements to the multi-view image acquisition module. First, the scanning frequency of the four linear array image acquisition units has been increased to a higher frequency to match the instantaneous displacement compensation of the conductor under vibration. In addition to using a blue narrowband light source with a center wavelength of 450 nanometers, the light source module also introduces a set of coaxial light illumination components. The coaxial light components directly and perpendicularly illuminate the surface of the conductor through a semi-transparent and semi-reflective mirror. This forms complementary illumination with the radially distributed active light source. This combined illumination scheme can significantly reduce the shadow dark areas caused by the drastic curvature change of the surface of the fine conductor, so that the microscopic scratches on the surface of the insulation layer have higher brightness and darkness contrast when imaging.

[0050] To address strong vibration interference, the system architecture of this embodiment introduces a real-time feedback displacement sensor module. This module uses a laser displacement meter to monitor the radial runout data of the conductor during the transmission process in real time, and compensates the offset to the image stitching unit of the central parallel processing module in real time. In the algorithm processing stage, the adaptive brightness correction process in step S2 introduces a spatiotemporal joint filtering algorithm in the spatial domain. The processor not only considers the pixel distribution of the current frame image, but also refers to the voxel grid energy state of the three frames before and after, thereby effectively filtering out dynamic artifacts caused by mechanical vibration.

[0051] In step S3, the feature separation logic fine-tunes the scale factor of the multi-scale structure tensor to a narrower range of 0.01 to 0.05 mm, taking into account the extremely short stranding period of ultra-fine diameter wires. By further optimizing the eigenvalue response threshold, the ratio of λ2 to λ3 is increased to 2.0 times, thereby enabling the system to generate a strong response to microcracks with a width of only 20 micrometers. This parameterized dynamic adjustment ensures the system's adaptability when switching between different product specifications.

[0052] In the deep cascaded discrimination model in step S4, this embodiment adopts a deeper feature pyramid and introduces an attention mechanism. The attention mechanism can guide the model to focus on the stranded junction where the conductor surface is under the greatest stress, because these areas are often high-incidence areas where stress concentration leads to insulation layer damage. By increasing the weight of key areas, the system can improve the ability to capture minor defects by about 15% while ensuring a low false alarm rate.

[0053] Regarding the interaction of the actuators, this embodiment achieves zero-latency synchronization between the processor and the industrial robot arm through the industrial Ethernet protocol. The end effector of the robot arm is replaced with a flexible gripping mechanism to prevent secondary mechanical damage when handling ultra-fine diameter wires. The vision guidance system instructs the robot to perform precise positioning with an accuracy of ±0.01 mm by calculating the three-dimensional spatial coordinates of the defect in real time.

[0054] Furthermore, the quality database in this embodiment possesses the dual attributes of edge computing and cloud synchronization. The on-site generated inspection reports are uploaded to the cloud server through the industrial gateway. A larger-scale expert system is used to perform multi-dimensional modeling of the structural stability of the conductors. When processing massive amounts of records, the long short-term memory neural network improves the convergence speed of training by introducing a residual connection structure, which further stabilizes the lead time for predicting the reliability of the production line at more than 500 hours, providing accurate data support for the company's production scheduling plan.

[0055] Example 3

[0056] This embodiment focuses on describing a distributed low-voltage conductor defect detection scheme, which is suitable for detection scenarios with multiple production lines running in parallel in large-scale factories. In this implementation, the system adopts a three-level architecture of "end-edge-cloud" to achieve optimized allocation of computing resources.

[0057] At the edge of the three-tier architecture, namely the front-end perception layer, each pipeline is independently configured with the multi-view image acquisition module and active illumination source module described in Embodiment 1. Each acquisition unit has a built-in dedicated digital signal processing chip for performing basic image compression and preliminary noise reduction. This distributed preprocessing design greatly reduces the transmission pressure on the backbone network. The raw video stream acquired by the acquisition unit is converted into 16-bit binary two's complement form in real time, and preliminary adaptive brightness mapping is completed inside the digital signal processing chip to initially eliminate the effect of specular reflection.

[0058] At the edge of the three-tier architecture, namely the edge computing layer, every four pipelines share one edge computing node. This node is equipped with an AI inference module with a high-performance graphics processor. The edge computing node executes the core algorithm logic, namely the multi-scale structural tensor decomposition in step S3 and the deep concatenation discrimination in step S4. Since the data from multiple pipelines converge here, a multi-threaded parallel processing pipeline is built inside the node. Virtualization technology is used to allocate an independent processing core to each pipeline. When performing feature extraction, the system uses the lightweight Mobilenet-v3 as the backbone network of the deep concatenation model, which significantly reduces the inference latency of a single frame image while ensuring the feature extraction depth.

[0059] On the cloud side of the three-tier architecture, namely the management and decision-making layer, the system has established a unified quality database across production lines. This database not only collects the geographical location and category of defects, but also records the environmental humidity, temperature, and raw material batch information during production. By performing nonlinear correlation analysis on these multidimensional data through a long short-term memory neural network, the system can identify the influence of environmental factors on the quality of conductor insulation. For example, the system may find that when the environmental humidity exceeds 80%, the probability of extrusion defects in conductors of a certain material increases by 20%. Based on this in-depth analysis, the reliability early warning with a 500-hour lead time has higher decision-making value.

[0060] To address the synchronization issues in a distributed architecture, this embodiment introduces time-sensitive networking technology. When executing the real-time early warning control in step S5, all control commands are timestamped at the nanosecond level. After receiving the command, the industrial robot arm can accurately compensate for the minute delays caused by network transmission, ensuring that defect location can still be achieved with an error of less than 50 micrometers on a high-speed production line.

[0061] In terms of specific algorithm evolution, this embodiment adds a self-learning step in step S4. When the support vector machine classifier encounters a suspected defect that is difficult to determine, the system will automatically transmit the image sample to the cloud-based human expert review interface through an encrypted tunnel. The result of the human review will be used as a new training sample to periodically update the model parameters of the edge computing nodes, thereby enabling the entire system to have continuously evolving detection accuracy as the running time increases.

[0062] Through the detailed description of the above embodiments, it can be seen that the present invention, through the precise layout of the hardware architecture and the deep integration of the algorithm process, not only achieves high-resolution capture of the entire circumference surface of low-voltage conductors, but also completely solves the signal interference problem under complex textures through innovative structural tensor decomposition technology. From microscopic pixel-level brightness correction to macroscopic production trend prediction, the present invention constructs a multi-level, all-round quality assurance system, which significantly improves the intelligence and automation level of low-voltage conductor detection.

[0063] Example 4

[0064] This embodiment further discloses the implementation details of the present invention in a specific highly flexible production scenario, particularly the adaptive adjustment mechanism for frequent switching of conductor materials.

[0065] In this highly flexible scenario, the system's central parallel processing module has a built-in material feature library. Whenever the production line switches wire types, the system automatically calls the preset parameter group corresponding to that type. For example, for dark insulation layers, the power of the active lighting source module will automatically increase by 30% to compensate for the material's absorption of blue light; while for transparent insulation layers with high gloss, the adaptive brightness correction algorithm in step S2 will increase the dynamic threshold and expand the voxel grid sampling radius to more effectively suppress secondary reflection noise.

[0066] When performing feature separation based on multi-scale structural tensor decomposition, the system monitors the baseline gray-scale entropy value of the image in real time. If the stranding texture on the surface of the conductor changes due to the adjustment of the manufacturing process, the processor will automatically extract the evolution gradient of the tensor trace and dynamically adjust the proportional relationship between the feature response thresholds λ1, λ2, and λ3. This closed-loop adjustment mechanism ensures that no matter how complex the surface texture of the conductor is, the stripped defect feature signal always maintains a high signal-to-noise ratio.

[0067] For the determination of structural strand breakage defects, this embodiment introduces a three-dimensional reconstruction auxiliary module in step S4. This module uses two-dimensional images from four perspectives to reconstruct the three-dimensional point cloud of the conductor surface through epipolar geometric constraints. When a strand breakage occurs, the local outer diameter of the conductor will undergo significant geometric deformation. The depth-cascaded discrimination model performs spatial voxelization processing on the point cloud data through an octree structure and uses Laplacian smoothing to remove pseudo-feature points caused by fine burrs. Subsequently, through the least squares plane fitting repair algorithm, the system can calculate the deviation of the normal vector between the broken strand and the normal surface. If the deviation exceeds a preset angle and the diameter of the affected area is greater than 0.2 mm, the system will immediately determine it as a serious structural defect.

[0068] At the level of automation execution, in order to meet the grasping requirements of wires of different diameters, the end effector of the industrial robot arm integrates a force sensor. When executing the warning control command in step S5, the robot arm not only relies on the 50-micron-level positioning error of the vision guidance system for spatial tracking, but also ensures that the clamping force can firmly grasp the wire without damaging the insulation layer through torque closed-loop control. This deep electromechanical coupling design reflects the comprehensive performance advantages of this system as a high-end power testing equipment.

[0069] Finally, the quality database in this embodiment has powerful data visualization capabilities. The system can generate a "defect heat map" in real time, showing the distribution density of defects in the length and circumference of the conductor. Combined with trend analysis of long short-term memory neural networks, the system can identify whether a specific guide wheel on the production line has periodic scratch defects due to wear, thereby guiding maintenance personnel to perform precise equipment maintenance. This big data-driven quality management increases the conductor qualification rate from 95% in traditional solutions to over 99.9%, greatly reducing the safety hazards of power grid operation.

[0070] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for detecting defects in low-voltage conductors based on machine vision, characterized in that, Includes the following steps: S1, acquires the full circumference surface image of the low-voltage conductor under motion state through the multi-view image acquisition module; S2 performs adaptive brightness correction and noise reduction preprocessing on the full circumference surface image. It eliminates specular reflection spots generated by the metal surface and insulation layer of the low-voltage conductor by constructing a local brightness mapping function, and extracts the region of interest image of the main body of the low-voltage conductor using an improved edge operator. S3 executes feature separation logic based on multi-scale structural tensor decomposition, combines the image of the region of interest with the grayscale of the image to form a grayscale three-dimensional model, calculates the local gradient covariance matrix in the grayscale three-dimensional model and performs eigenvalue decomposition, and performs physical-level blind source stripping of the periodic stranded texture signal and non-periodic defect feature signal on the surface of the low-voltage conductor. S4. Construct a deep cascaded discrimination model, utilize the feature pyramid structure to fuse multi-level semantic information, extract features and characterize morphological parameters of defect feature signals, and realize automatic location and classification of defects in low-voltage conductors. S5 generates a defect distribution difference map based on the discrimination results, and outputs a detection report and real-time early warning control instructions in combination with the preset quality and safety thresholds.

2. The machine vision-based low-voltage conductor defect detection method according to claim 1, characterized in that, The multi-view image acquisition module includes four sets of linear image acquisition units, ensuring full circumference surface coverage of low-voltage conductors in motion. It also uses a blue narrowband light source with a preset center wavelength as the active illumination source. The effective pixel count of a single image acquisition unit meets a preset pixel threshold, and the scanning accuracy is limited to a preset accuracy range.

3. The machine vision-based low-voltage conductor defect detection method according to claim 1, characterized in that, In step S2, the adaptive brightness correction process first combines the acquired two-dimensional image with brightness mapping to the three-dimensional perception space, sets the voxel grid size to a preset size, constructs a local brightness mapping function through the voxel grid, calculates the average brightness within each grid, and dynamically adjusts the compensation parameters of the brightness mapping function according to the average brightness around the grid. The grid brightness is adjusted according to the compensation parameters to eliminate the shading of the light spot on the underlying features of the low-voltage conductor.

4. The machine vision-based low-voltage conductor defect detection method according to claim 1, characterized in that, In step S3, the feature separation logic based on multi-scale structural tensor decomposition is as follows: For each pixel in the extracted region of interest image, a gray-scale three-dimensional model is formed by combining the corresponding gray value of the pixel. The local gradient covariance matrix is ​​calculated to describe the distribution law of gray-scale changes in the image. Feature decomposition is performed on the obtained covariance matrix to obtain eigenvalues ​​λ1, λ2, and λ3. By comparing the numerical relationship between each eigenvalue, the periodic stranded texture signal on the surface of the low-voltage conductor and the non-periodic defect feature signal are physically separated.

5. The machine vision-based low-voltage conductor defect detection method according to claim 1, characterized in that, In step S4, the depth-cascaded discrimination model uses an octree structure of preset depth to perform multi-scale spatial modeling, performs multi-scale spatial voxel modeling on the surface of the conductor acquired from multi-view images, removes pseudo-feature points caused by burrs in the point cloud through Laplacian smoothing, and then reconstructs the three-dimensional geometric shape of the defect through hole repair logic to obtain the complete three-dimensional contour of the defect area.

6. The machine vision-based low-voltage conductor defect detection method according to claim 5, characterized in that, For geometrically abnormal regions that have been reconstructed in three dimensions and whose geometric dimensions are greater than a preset threshold, least squares plane fitting repair is performed to obtain the fitting plane of the abnormal region. The normal vector deviation between the point cloud of the abnormal region and the fitting plane is calculated. If the normal vector deviation exceeds a preset angle threshold, the abnormal region is determined to be a structural defect.

7. The machine vision-based low-voltage conductor defect detection method according to claim 1, characterized in that, The classification process employs a support vector machine classifier based on radial basis function kernel function to perform nonlinear mapping on the extracted multidimensional feature vectors, including area, perimeter, roundness, depth, and edge curvature, to identify the defect type.

8. The machine vision-based low-voltage conductor defect detection method according to claim 1, characterized in that, It also involves establishing a low-voltage conductor quality database containing historical test records of a preset scale, and using a long short-term memory neural network to perform trend analysis and reliability prediction on the structural stability of low-voltage conductors.

9. A machine vision-based low-voltage conductor defect detection system, applied to the machine vision-based low-voltage conductor defect detection method according to any one of claims 1-8, characterized in that, include: A multi-view image acquisition module is used to acquire full-circumference surface images of low-voltage conductors in motion. The central parallel processing module completes the entire process from image to defect determination result; Automated actuators, including industrial robot arms, are used to perform handling operations on low-voltage wires based on real-time early warning control commands output by a central parallel processing module.

Images