Method and system for detecting weld defect of fan tower based on laser radar

Abstract

The invention relates to the technical field of wind power operation and maintenance, in particular to a fan tower weld defect detection method and system based on laser radar, and the system comprises a data acquisition module, an anomaly analysis module and a defect identification module. According to the method, a laser radar is used for obtaining geometric-optical double response signals of welding seam space-time alignment; a signal abnormal region is extracted, a reflection saturation distribution index is constructed, and a geometric-reflectivity characteristic vector is generated; identifying the physical attribute of the abnormal region by using the classification model, judging the optical mirror interference as a pseudo defect, and judging the material physical deficiency as a real defect; and performing curved surface fitting repair on the pseudo defects, and performing physical dimension quantification on the real defects. According to the method, the relevance between the light intensity and the depth is excavated, oxide film interference and real defects are effectively distinguished, the problem that the false alarm rate of in-service tower drum detection is high is solved, and non-contact efficient scanning and digital reconstruction are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind power operation and maintenance, in particular to a wind turbine tower weld defect detection method and system based on laser radar. BACKGROUND

[0002] As the weakest link in the tower structure, welds are prone to defects such as cracks, pores, and incomplete fusion, which directly affects the safe operation of the entire wind turbine. Therefore, it is particularly important to regularly, efficiently, and accurately detect the welds of in-service wind turbine towers. Currently, the detection of wind turbine tower welds mainly relies on manual inspection or traditional automated detection equipment. However, existing detection techniques still have many limitations in practical applications and cannot meet the urgent needs of wind power operation and maintenance for high efficiency, high precision, and automated detection.

[0003] Firstly, for the automated detection of tower welds, existing technologies mostly adopt the operation mode of "tower rotation, fixed detection equipment". For example, in the factory manufacturing link of the tower, a motion control system is used to drive the roller frame to rotate the tower, and a fixed laser camera or visual sensor is used to collect image data. This method is effective in a factory environment, but for "in-service" towers that have been installed in wind farms and are in a straight and fixed state, the above detection scheme relying on rotating mechanisms cannot be implemented because the towers cannot be rotated.

[0004] To solve the problem of detecting in-service towers, the industry has begun to explore the use of wall-climbing robots carrying sensors for work. However, existing wall-climbing detection robots mostly carry ultrasonic sensors. Ultrasonic detection is a contact detection method that requires the probe to be in close contact with the rough weld surface and requires the application of coupling agent, resulting in slow scanning speed, low detection efficiency, and difficulty in covering a large tower surface area. In contrast, non-contact detection technology based on laser radar (or line structured light) has the advantages of fast speed and high precision, but it faces the problem of special optical interference in practical applications.

[0005] Specifically, wind turbine towers widely use carbon dioxide gas shielded welding technology, which easily produces a thin film of oxide precipitates composed of silicon and manganese elements on the weld surface. The oxide film is smooth and has strong mirror reflection characteristics. Traditional laser radar sensors rely on diffuse reflection signals for imaging, and when the laser beam hits the oxide film area, the light is reflected by the mirror and deviates from the receiver's field of view, resulting in signal loss or "holes" in the three-dimensional point cloud data collected. Existing detection algorithms often have difficulty distinguishing between data loss caused by optical characteristics and real physical defects (such as burn-through and deep pits), often misjudging the oxide film as a serious defect, resulting in a high false positive rate and increasing the cost and difficulty of subsequent manual review.

[0006] In summary, how to overcome the false alarms caused by the specular reflection of the oxide film on the weld seam and distinguish the real defects in the detection of weld seam defects by wall-climbing robots is a technical problem that urgently needs to be solved in this field.

[0007] To address this, a method and system for detecting weld defects in wind turbine towers based on lidar are proposed. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for detecting weld defects in wind turbine towers based on lidar. This method utilizes lidar to acquire spatiotemporally aligned geometric-optical dual-response signals of the weld; extracts signal anomaly regions, constructs a reflection saturation distribution index, and generates a geometric-reflectivity feature vector; uses a classification model to identify the physical properties of the anomaly regions, classifying optical specular interference as pseudo-defects and material physical deficiencies as real defects; repairs pseudo-defects through surface fitting, and quantifies the physical dimensions of real defects. This invention effectively distinguishes between oxide film interference and real defects by exploring the correlation between light intensity and depth, solving the problem of high false alarm rates in in-service tower detection, and achieving non-contact, high-efficiency scanning and digital reconstruction.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A method for detecting weld defects in wind turbine towers based on lidar, comprising: The lidar sensor emits a probe beam toward the weld surface and receives the echo signal, simultaneously demodulating a spatiotemporally aligned geometric-optical dual-response signal; the geometric-optical dual-response signal includes three-dimensional contour data and light intensity reflectivity distribution data. Signal continuity analysis is performed on the three-dimensional contour data to extract signal abnormal regions; based on the light intensity reflectance distribution data, the light signal scattering behavior in the signal abnormal regions is analyzed, a reflectance saturation distribution index is constructed, and it is determined whether there are abnormal phenomena caused by oxide films in the signal abnormal regions; the spatial change rate of the three-dimensional contour data and the light intensity change rate of the light intensity reflectance distribution data are calculated, and the two are spatially correlated and coupled to generate a geometric-reflectance feature vector. Based on the reflection saturation distribution index and the geometric-reflectivity feature vector, a classification model is used to identify the physical properties of signal abnormal regions and generate attribute discrimination indicators for defect identification; among them, regions with the attribute of optical mirror interference are judged as pseudo-defects, and regions with the attribute of material physical deficiency are judged as real defects.

[0010] Preferably, the step of synchronously demodulating the spatiotemporally aligned geometric-optical dual-response signal includes: driving a lidar sensor to emit a monochromatic coherent beam in the blue light band, the beam being reflected and scattered upon contact with the weld surface; using a photoelectric sensing array to simultaneously sense the spatial distribution of the diffuse reflection spot on the weld surface and the integral intensity of the echo energy within a single exposure time window; based on the principle of optical triangulation, resolving the spatial distribution of the diffuse reflection spot into height coordinate information in the three-dimensional contour data; based on the photoelectric conversion effect, mapping the integral intensity of the echo energy into the surface reflectivity value in the light intensity reflectivity distribution data; and using a hardware clock synchronization triggering mechanism to lock the height coordinate information and surface reflectivity value collected at the same time, constructing a point-to-point mapped weld surface physical property dataset.

[0011] Preferably, the step of performing signal continuity analysis on the three-dimensional contour data and extracting signal anomalous regions includes: performing spatial topology analysis on the three-dimensional contour data to identify discontinuous regions whose spatial curvature exceeds a preset physical deformation threshold; performing signal dispersion aggregation on the discontinuous regions to aggregate adjacent signal missing points whose dispersion meets a preset condition into connected regions; and filtering out connected regions whose area exceeds a preset area threshold to define them as the signal anomalous regions. The physical manifestations of the signal anomalous regions include: optical blind zones where the sensor does not receive effective echo signals due to total internal reflection of the light beam, and geometric shadow zones where the light beam is blocked due to surface geometric depressions.

[0012] Preferably, the step of constructing a reflection saturation distribution index based on the light intensity reflectance distribution data includes: taking the geometric centroid of the signal anomaly region as the center, extracting a region of interest covering the effective weld surface of the region and its surrounding area from the light intensity reflectance distribution data; statistically analyzing the light intensity energy distribution histogram within the region of interest, dividing the light intensity data into a saturated highlight region, a noise cutoff region, and a diffuse reflection median region; calculating the spatial distribution optical uniformity index of effective pixels within the region of interest excluding the noise cutoff region; and combining the proportion of the saturated highlight region, the proportion of the noise cutoff region, and the optical uniformity index through a weighted combination to generate a reflection saturation distribution index; the reflection saturation distribution index reflects whether the target area exhibits the binary optical characteristics of high reflectivity and total reflection blind zone coexisting, unique to oxide films.

[0013] Preferably, the step of calculating the spatial rate of change of the three-dimensional contour data and the rate of change of light intensity reflectance distribution data to generate a geometric-reflectance feature vector includes: calculating the geometric shape gradient field of the three-dimensional contour data along the scanning direction and the light intensity energy gradient field of the light intensity reflectance distribution data along the scanning direction; constructing a geometric-light intensity correlation matrix describing the spatial cooperative change relationship between the geometric shape gradient field and the light intensity energy gradient field; performing eigenvalue decomposition on the geometric-light intensity correlation matrix to extract the principal component eigenvalues ​​characterizing the degree of alignment between the two boundaries; and concatenating the geometric shape gradient field, the light intensity energy gradient field, and the principal component eigenvalues ​​to generate the geometric-reflectance feature vector; the feature vector describes the consistency between the physical boundary and the optical boundary of the target region in the multi-dimensional feature space.

[0014] Preferably, the step of identifying the physical properties of the signal anomaly region using a classification model and generating an attribute discrimination indicator includes: establishing a multimodal signal sample library containing oxide film optical interference samples and real physical defect samples; training a support vector machine classifier using the multimodal signal sample library to establish a mapping relationship from input features to physical properties; inputting the reflectance saturation distribution index and geometric-reflectivity feature vector of the region to be tested into the classifier; if the classifier's determination result matches the characteristics of strong specular reflection and low optomechanical correlation, generating an attribute discrimination indicator characterizing an optical virtual image; if the classifier's determination result matches the characteristics of diffuse reflection scattering and high optomechanical correlation, generating an attribute discrimination indicator characterizing material deficiency.

[0015] Preferably, the step of identifying defects based on the attribute discrimination indicator and performing surface fitting repair on the region marked as a pseudo-defect includes: locking the boundary of the region marked as an optical virtual image by the discrimination indicator; extracting the effective three-dimensional contour data of the diffuse reflection zone of normal weld metal outside the boundary as a reference topological control point; using a non-uniform rational B-spline surface reconstruction algorithm to deduce and fill the virtual surface morphology data within the optical virtual image region based on the reference topological control point; generating a reconstructed three-dimensional weld model after eliminating optical interference to restore the continuous topological structure of the weld surface.

[0016] Preferably, the step of quantifying the physical dimensions of areas identified as real defects and outputting weld defect detection results specifically includes: based on the reconstructed 3D weld model, extracting the geometric parameters of all areas marked as material missing, and constructing a multidimensional defect feature vector containing the maximum normal depth, opening span, volume equivalent, and spatial coordinates; calling a pre-set wind turbine tower weld quality grading database, which stores physical limit thresholds defined based on industry standards; mapping the multidimensional defect feature vector to the physical limit thresholds, executing defect severity grading logic, and determining the hazard level of each real defect; using the reconstructed 3D weld model as a base, performing texture mapping on the determined hazard level and spatial coordinates to generate a weld defect topology distribution map; and outputting the weld defect topology distribution map and corresponding structured physical attribute data as the weld defect detection results.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention delves into the synchronous geometric-optical dual-response signal of the lidar sensor itself. By constructing a specular reflection saturation distribution index and a geometric-reflectivity coupled characteristic vector, this invention can accurately distinguish between "optical signal loss due to oxide film specular reflection" and "physical material loss due to weld porosity or burn-through" from a mathematical and physical perspective, without adding additional optical hardware load. This innovation completely solves the industry pain point of traditional laser triangulation methods in detecting carbon dioxide shielded welds, where "point cloud voids" are frequently misjudged as serious defects due to strong reflectivity, greatly improving the confidence of automated detection.

[0018] 2. This invention combines lidar non-contact scanning technology with a magnetic adsorption wall-climbing transport platform, and with a specific sensor installation angle (non-vertical preset incident angle), enables the system to achieve high-speed continuous scanning at heights of tens of meters without tower rotation or surface pretreatment. This not only fills the technological gap in "high-precision three-dimensional inspection of in-service towers," but also significantly improves the operational efficiency and coverage of maintenance and inspection.

[0019] 3. This invention, through material-optical property discrimination indicators, can not only filter out false defects but also utilize surface fitting repair technology to virtually restore the morphology of visually disturbed areas caused by oxide films, generating a noise-reduced 3D model of the reconstructed weld. Based on the structured data generated from this model, including maximum normal depth, opening span, and volume equivalent, a precise digital twin foundation is provided for the full life-cycle health management (PHM) of wind turbine towers. This allows the inspection results to be used not only for compliance assessment but also for subsequent stress analysis and life prediction. Attached Figure Description

[0020] Figure 1A flowchart of a method for detecting weld defects in wind turbine towers based on lidar, provided in an embodiment of the present invention; Figure 2 A structural diagram of a laser radar-based wind turbine tower weld defect detection system provided in an embodiment of the present invention; Figure 3 A schematic diagram of the data processing and feature fusion architecture provided for embodiments of the present invention. Detailed Implementation

[0021] 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.

[0022] Please see Figures 1 to 3 This invention provides a method for detecting weld defects in wind turbine towers based on lidar, the technical solution of which is as follows: A method for detecting weld defects in wind turbine towers based on lidar, comprising: The lidar sensor emits a probe beam toward the weld surface and receives the echo signal, simultaneously demodulating a spatiotemporally aligned geometric-optical dual-response signal; the geometric-optical dual-response signal includes three-dimensional contour data and light intensity reflectivity distribution data. Signal continuity analysis is performed on the three-dimensional contour data to extract signal abnormal regions; based on the light intensity reflectance distribution data, the light signal scattering behavior in the signal abnormal regions is analyzed, a reflectance saturation distribution index is constructed, and it is determined whether there are abnormal phenomena caused by oxide films in the signal abnormal regions; the spatial change rate of the three-dimensional contour data and the light intensity change rate of the light intensity reflectance distribution data are calculated, and the two are spatially correlated and coupled to generate a geometric-reflectance feature vector. Based on the reflection saturation distribution index and the geometric-reflectivity feature vector, a classification model is used to identify the physical properties of signal abnormal regions and generate attribute discrimination indicators for defect identification; among them, regions with the attribute of optical mirror interference are judged as pseudo-defects, and regions with the attribute of material physical deficiency are judged as real defects.

[0023] Example 1:

[0024] This embodiment applies to the operation and maintenance of the in-service tower of a 2.5MW onshore wind farm. The tower is 80 meters high, made of Q345D steel, and connected using carbon dioxide gas shielded welding. Due to the characteristics of the welding process, a large number of irregularly distributed silicon-manganese oxide precipitates (oxide film) remain on the weld surface. This film has an extremely smooth surface, which easily causes specular reflection during traditional laser inspection, leading to signal loss and being falsely reported as a "burn-through" defect.

[0025] This method also includes the following tools: Magnetic Adsorption Wall Climbing Transport Platform: It adopts a permanent magnet adsorption chassis with four independent drives. The adsorption force of a single wheel is set to 800N, which can carry a 15kg effective load and stably adsorb and move on the vertical and inverted conical tower wall.

[0026] Optical Inspection Module: The core sensor is a customized high-frequency line structured light lidar, mounted on a two-axis gimbal at the front of the robot. The optical axis and the weld seam normal direction are set at a preset incident angle of 22.5°. This angle is a "golden avoidance angle" determined through extensive optical simulation, which can effectively avoid more than 85% of planar mirror reflections directly entering the pupil.

[0027] Signal processing unit: Equipped with NVIDIA Jetson AGX Orin edge computing module, responsible for real-time demodulation of synchronous geometric-optical dual response signals.

[0028] As one embodiment of the present invention, refer to Figure 1 A flowchart of a method for detecting weld defects in wind turbine towers based on lidar, referring to... Figure 2 A structural diagram of a wind turbine tower weld defect detection system based on lidar, referring to... Figure 3 A schematic diagram of the data processing and feature fusion architecture.

[0029] Further, the step of simultaneously demodulating the spatiotemporally aligned geometric-optical dual-response signal includes: driving a lidar sensor to emit a monochromatic coherent beam in the blue light band, the beam being reflected and scattered upon contact with the weld surface; using a photoelectric sensing array, simultaneously sensing the spatial distribution of the diffuse reflection spot on the weld surface and the integral intensity of the echo energy within a single exposure time window; based on the principle of optical triangulation, resolving the spatial distribution of the diffuse reflection spot into height coordinate information in the three-dimensional contour data; based on the photoelectric conversion effect, mapping the integral intensity of the echo energy into the surface reflectivity value in the light intensity reflectivity distribution data; and using a hardware clock synchronization triggering mechanism to lock the height coordinate information and surface reflectivity value collected at the same time, constructing a point-to-point mapped dataset of the physical properties of the weld surface.

[0030] Specifically, the built-in semiconductor laser of the lidar emits a blue monochromatic coherent beam with a wavelength of 450 nanometers. Compared to traditional red light, according to the optical diffraction limit theory, the focused spot diameter of short-wavelength blue light on the metal surface is reduced to 69% of that of red light (approximately 25 micrometers), significantly improving the lateral spatial resolution. Simultaneously, the refractive index (n≈1.8-2.2) of the silicon-manganese oxide produced by CO2 welding differs significantly from that of the weld metal in the 450nm band. Combined with a 22.5° tilted incidence design, this results in a typical "two-mode" optical response when the beam contacts a smooth oxide film. The laser power is dynamically set between 20 and 50 milliwatts, utilizing the high sensitivity of blue light to the surface finish of microscopic surfaces to achieve a typical "two-mode" optical response when the beam contacts a smooth oxide precipitate film: strong directional specular reflection (signal saturation) occurs when the beam is incident perpendicularly or nearly perpendicularly; while total internal reflection occurs when the beam is incident at an angle due to the smooth surface (signal loss, below the noise floor). This extreme brightness-dark binarization response contrasts sharply with the uniform diffuse reflection signal generated by the beam inside rough real defects (such as pores and cracks), providing a high-confidence optical discrimination benchmark for subsequent algorithms.

[0031] It employs a custom high dynamic range (HDR) CMOS image sensor with a dynamic range exceeding 120dB; the single exposure time window is set to 50 microseconds. Within this extremely short time window, the sensor utilizes its 4096 lateral pixel units to simultaneously capture the diffuse reflection spot shape and energy distribution of the laser line on the weld surface. The embedded processing unit performs two signal processing operations in parallel: on the one hand, it uses a centroid extraction algorithm to calculate the sub-pixel position of the energy center of the light spot on the sensor pixel, and calculates the height coordinate information (Z-axis) with a vertical resolution of 0.01 mm based on a triangulation model; on the other hand, based on the photoelectric conversion effect, it calculates the sum of the gray integrals of all pixels in the area covered by the light spot, and linearly maps it to a surface reflectivity value (intensity I) of 0 to 1024 levels; for example, the reflectivity value of the oxide film area often reaches saturation (>1000) instantaneously due to specular reflection, while the value drops sharply (<50) at the micropores due to light trapping.

[0032] By using a high-precision magnetic encoder on a wall-climbing robot to send a hardware trigger signal every 0.1 millimeters of movement, the height coordinates and surface reflectivity values ​​acquired within the same exposure frame are forcibly locked, ensuring that the timestamp deviation between the two is less than 1 microsecond. Ultimately, a dataset of weld surface physical properties is constructed, which scans 2000 contours per second, with each contour containing 4096 point pairs of data. This provides accurate spatiotemporal alignment data support for subsequent algorithms to distinguish between "optical virtual images" and "physical defects."

[0033] This invention employs a "monochromatic coherent beam (blue light)" and "single-exposure synchronous demodulation." Compared to red lasers, short-wavelength blue light significantly reduces subsurface scattering noise on metal surfaces, improving the positioning accuracy of the light spot centroid. More importantly, the hardware-level clock synchronization triggering mechanism ensures strict alignment of light intensity and depth data on a microsecond-level timescale, eliminating spatial misalignment caused by the jitter of the wall-climbing robot's movement. This provides a high-confidence physical data foundation for subsequent "optical-mechanical coupling analysis."

[0034] Further, the step of performing signal continuity analysis on the three-dimensional contour data and extracting signal abnormal regions includes: performing spatial topology analysis on the three-dimensional contour data to identify discontinuous regions whose spatial curvature exceeds a preset physical deformation threshold; performing signal dispersion aggregation on the discontinuous regions to aggregate adjacent signal missing points whose dispersion meets preset conditions into connected regions; filtering out connected regions whose area exceeds a preset area threshold and defining them as the signal abnormal regions; the physical manifestations of the signal abnormal regions include: optical blind zones where the sensor does not receive effective echo signals due to total internal reflection of the light beam, and geometric shadow areas where the light beam is blocked due to surface geometric depressions.

[0035] Specifically, spatial topology analysis is performed on the high-density 3D point cloud (point spacing set to 0.05 mm) acquired by lidar. Based on welding process standards, the preset physical deformation threshold is a change in the angle between the normal vectors of adjacent sampling points exceeding 45 degrees or a height step exceeding 0.5 mm. When scanning to the weld edge undercut or oxide film reflective area, abrupt curvature changes or invalid depth values ​​(NaN) are detected, not only identifying isolated anomalies but also delineating spatially discontinuous topological fracture zones.

[0036] Subsequently, signal discreteness aggregation was performed on these discontinuous regions. An aggregation radius of 1.0 mm was set, and discrete signal loss points within this Euclidean distance range were treated as fragments of the same event and merged. For example, for a set of discontinuous lost signals caused by the micro-roughness of surface oxide scale, morphological closing operations were used to fuse them into a complete connected region, avoiding misclassification of the same large signal loss area as multiple tiny noise points.

[0037] Next, feature filtering was performed, setting a significant area threshold of 2.0 square millimeters to automatically filter out tiny spots caused by airborne dust or circuit thermal noise. Ultimately, two typical signal anomaly regions were successfully identified in a single scan: a continuous strip-shaped area of ​​15 square millimeters, which was accurately defined as an "optical blind zone" due to the specular total internal reflection of the laser caused by the smooth surface of the oxide film; and a deep pit with a diameter of about 1.2 millimeters and a depth of 1.5 millimeters, which was accurately defined as a "geometric shadow area" because the laser incident light path was blocked by the steep pit walls, preventing the sensor from receiving the echo.

[0038] This invention implements an efficient pre-screening mechanism through "spatial topology analysis" and "signal discreteness aggregation," avoiding complex tensor calculations on all pixels of the entire image and instead focusing only on anomalous regions with significant physical characteristics (such as total loss or high gradient). This strategy significantly reduces the computational load on edge computing units, making real-time detection possible on computationally limited embedded wall-climbing platforms and improving overall response speed.

[0039] Further, the step of constructing a reflection saturation distribution index based on the light intensity reflectance distribution data includes: taking the geometric centroid of the signal anomaly region as the center, extracting a region of interest covering the effective weld surface of the region and its surrounding area from the light intensity reflectance distribution data; statistically analyzing the light intensity energy distribution histogram within the region of interest, dividing the light intensity data into a saturated highlight region, a noise cutoff region, and a diffuse reflection median region; calculating the spatial distribution optical uniformity index of effective pixels within the region of interest excluding the noise cutoff region; and combining the proportion of the saturated highlight region, the proportion of the noise cutoff region, and the optical uniformity index through a weighted combination to generate the reflection saturation distribution index; the reflection saturation distribution index reflects whether the target area exhibits the unique binary optical characteristics of high reflectivity and total reflection blind zone coexisting with the oxide film.

[0040] Before extracting the light intensity signal of the abnormal signal region and its neighborhood, the process includes performing adaptive light intensity normalization: obtaining the current working height of the wall-climbing adsorption robot and the pipe diameter parameters of the wind turbine tower, and fitting the tower surface reflection model of the current scanning area in real time; calculating the local incident angle of each sampling point relative to the probe beam based on the tower surface reflection model; performing vignetting correction and normalization on the original light intensity reflectivity distribution data using a reverse compensation algorithm to eliminate the light intensity background difference caused by the geometric curvature of the tower; and calculating the reflection saturation distribution index based on the normalized light intensity data.

[0041] This invention successfully decouples the light intensity attenuation caused by geometry from the light intensity variation caused by material properties by establishing a reflection model of the tower's curved surface. This eliminates the halo effect (bright center, dark edges) commonly seen in lidar scanning of curved surfaces, ensuring that oxide films located at weld edges or the small diameter of the tower are not missed due to natural light intensity attenuation. This correction significantly improves the consistency and reliability of the reflection saturation distribution index across the entire field of view, guaranteeing uniform detection accuracy for complex curved weld surfaces.

[0042] Specifically, when the lidar's 3D scanning module detects a 12 square millimeter data loss hole (i.e., an abnormal signal area) in the heat-affected zone of the weld, it immediately retrieves the synchronously acquired light intensity reflectivity distribution data. Centered on the geometric centroid of this abnormal area, a 20-pixel × 20-pixel rectangular neighborhood covering the abnormal area and its surrounding effective weld surface is extracted as the region of interest (ROI). The original light intensity data for this region has a quantization level of 1024.

[0043] Histogram analysis was performed on 400 sampling points within the ROI, dividing the light intensity signal into three intervals: a saturated specular region (intensity > 1000), a noise cutoff region (intensity < 20, corresponding to a signal blind zone), and a diffuse reflection intermediate value region. (Oxide film case) Statistical data shows that this region exhibits an extreme "bimodal distribution" characteristic: approximately 35% of the signal points fall into the saturated specular region (spectral reflection), while another 55% fall into the noise cutoff region (total internal reflection blind zone), with the intermediate diffuse reflection region accounting for a very low percentage. This phenomenon of "coexistence of high brightness and extreme darkness" is consistent with the optical characteristics of oxide films. (Real pore case) In contrast, if this region represents a real pore, the light beam undergoes diffuse reflection within the rough pore walls. Statistical analysis shows that 90% of the signal intensity is concentrated in the diffuse reflection intermediate value region between 200 and 600, with the combined percentage of the saturated specular region and the noise cutoff region being only 10%.

[0044] Texture features within the neighborhood are analyzed using the gray-level co-occurrence matrix principle. Only effective pixels within the region of interest (ROI), excluding the noise cutoff area, are selected during calculation. For the oxide film, the surrounding effective area exhibits a continuous, strong reflective transition with smooth texture, resulting in an optical uniformity index as high as 0.92. However, for the pores, the uneven inner walls cause fragmented light spots, leading to extremely chaotic texture in the effective signal area, resulting in an optical uniformity index of only 0.35.

[0045] A weighted combination was applied, assigning a weight of 0.3 to the saturated highlight area, a weight of 0.3 to the noise cutoff area, and a weight of 0.4 to optical uniformity. (Oxide film case) The generated reflection saturation distribution index quantization value was 0.89, exceeding the preset physical characteristic threshold of 0.75, thus determining that this area conforms to the physical characteristic of "directional reflection from oxide film." (Porosity case) The generated index value was 0.20, and it was determined that it did not conform to the directional reflection characteristic, confirming a genuine physical material deficiency.

[0046] The "spectral reflection saturation distribution index" proposed in this invention transforms the qualitative optical characteristics (bright or dark) of oxide precipitate films into a quantitative mathematical index. By statistically analyzing the proportion of "high reflectivity" and "weak reflectivity," the index can cover all optical performances of the oxide film under different incident angles (vertical overexposure or tilted complete darkness). This technique allows, like an experienced quality inspector, to identify false defects by observing reflective properties.

[0047] Further, the step of calculating the spatial rate of change of the three-dimensional contour data and the rate of change of light intensity reflectance distribution data to generate a geometric-reflectance feature vector includes: calculating the geometric shape gradient field of the three-dimensional contour data along the scanning direction and the light intensity energy gradient field of the light intensity reflectance distribution data along the scanning direction; constructing a geometric-light intensity correlation matrix describing the spatial cooperative change relationship between the geometric shape gradient field and the light intensity energy gradient field; performing eigenvalue decomposition on the geometric-light intensity correlation matrix to extract the principal component eigenvalues ​​characterizing the degree of alignment between the two boundaries; and concatenating the geometric shape gradient field, the light intensity energy gradient field, and the principal component eigenvalues ​​to generate the geometric-reflectance feature vector; the feature vector describes the consistency between the physical boundary and the optical boundary of the target region in the multi-dimensional feature space.

[0048] The constructed geometry-light intensity correlation matrix is ​​a 2x2 covariance matrix, with its elements defined as follows: the two elements on the main diagonal are the variances of the geometric gradient field and the light intensity energy gradient field, respectively; the two elements on the side diagonal are the covariances of the geometric gradient field and the light intensity gradient field. The covariance calculation is limited to a 20x20 pixel neighborhood window of the target region. Singular value decomposition (SVD) is performed on this matrix, and its maximum singular value is extracted as the principal component eigenvalue representing the degree of boundary alignment between the two. The physical meaning of this eigenvalue is that when a real defect exists, geometric abrupt changes and light intensity abrupt changes are highly synchronized, and the eigenvalue approaches 1; while for oxide film pseudo-defects, the two are uncorrelated, and the eigenvalue approaches 0. The final generated geometry-reflectivity feature vector is a three-dimensional vector containing the mean of the geometric gradient, the mean of the light intensity gradient, and this principal component eigenvalue. Singular value decomposition is performed on this matrix, and its maximum singular value is extracted as the principal component eigenvalue representing the degree of boundary alignment. The larger the maximum singular value, the stronger the linear correlation between the changes in geometric shape and the changes in light intensity and energy in spatial distribution.

[0049] Specifically, two typical "suspected defect" areas within the weld heat-affected zone were first identified: area A, a real transverse microcrack (0.5 mm wide, 1.2 mm deep), and area B, a long strip of residual oxide scale. When calculating the geometric gradient field, differentiation was performed along the scan line direction perpendicular to the weld orientation. For area A (microcrack), the 3D contour data showed a steep depth step at the crack edge, with a calculated geometric gradient peak as high as 6.5 mm / mm, indicating severe physical deformation. For area B (oxide scale), although there was signal loss within it, the effective depth data around the boundary of the lost area showed a smooth transition, with a calculated geometric gradient value of only 0.2 mm / mm, indicating a smooth physical surface.

[0050] Simultaneously, the light intensity energy gradient field is calculated. For region A (microcrack), due to multiple reflections and attenuation after the laser beam penetrates deep into the crack, the echo intensity drops sharply at the crack edge, forming a distinct optical shadow. The calculated light intensity energy gradient value is 180 levels / pixel, exhibiting a strong light-dark boundary. For region B (oxide scale), due to the alternating occurrence of overexposure caused by specular reflection and black spots caused by total internal reflection, its light intensity energy gradient value is exceptionally large, reaching 450 levels / pixel. Subsequently, a geometry-light intensity correlation matrix is ​​constructed to quantify the synergy of the spatial distribution of the two gradients. For the real crack, the location of its depth abrupt change highly coincides with the edge of the light intensity shadow at the sub-pixel level, and the boundary alignment between the two is extremely high.

[0051] Eigenvalue decomposition was performed on the matrix to extract principal component eigenvalues ​​to characterize this alignment. The results show that the principal component eigenvalue of region A is as high as 0.92 (normalized value), indicating a very high consistency between the geometric and optical boundaries; while the principal component eigenvalue of region B is only 0.15, indicating that the extremely high light intensity variation is not supported by a corresponding depth variation, and the two are spatially decoupled. Finally, the geometric gradient value (6.5), light intensity gradient value (180), and principal component eigenvalue (0.92) are concatenated to generate a high-dimensional feature vector containing multidimensional physical information. This vector, as a digital fingerprint, clearly distinguishes physical cracks from optical artifacts in the multidimensional feature space, providing input data with decisive discriminative power for subsequent classifiers.

[0052] This invention constructs a "geometry-reflectivity coupled feature vector," and through the "geometry-light intensity correlation matrix" and eigenvalue decomposition, it mathematically reveals the intrinsic difference between real defects (where depth and light intensity change abruptly) and pseudo-defects (where only light intensity changes abruptly). This feature extraction method based on tensor physics is more robust than simple image texture analysis, maintaining extremely high discrimination accuracy even when oil or rust interferes with the weld surface.

[0053] Further, the step of identifying the physical properties of the signal anomaly region using a classification model and generating attribute discrimination indicators includes: establishing a multimodal signal sample library containing oxide film optical interference samples and real physical defect samples; training a support vector machine classifier using the multimodal signal sample library to establish a mapping relationship from input features to physical properties; inputting the reflectance saturation distribution index and geometric-reflectivity feature vector of the region to be tested into the classifier; if the classifier's determination result matches the characteristics of strong specular reflection and low optomechanical correlation, generating an attribute discrimination indicator characterizing an optical virtual image; if the classifier's determination result matches the characteristics of diffuse reflection scattering and high optomechanical correlation, generating an attribute discrimination indicator characterizing material deficiency.

[0054] The step of analyzing the light signal scattering behavior in the signal anomaly area also includes performing frequency domain texture filtering: transforming the three-dimensional contour data and light intensity reflectivity distribution data from the spatial domain to the frequency domain, extracting the main texture frequency of the fish scale pattern formed by the weld surface process; calculating the spectral response characteristics of the signal anomaly area to be determined; if the spectral response shows high-frequency random noise or low-frequency DC component, it is determined to be an oxide precipitate covering layer that has disrupted the texture periodicity; if the spectral response shows a phase abrupt change or spectral leakage at the main texture frequency, it is determined to be a real physical defect that breaks the continuity of the fish scale pattern; and fusing the spectral response characteristics as an auxiliary discrimination criterion into the geometry-reflectivity feature vector.

[0055] This invention overcomes the limitations of relying solely on spatial domain geometric continuity by utilizing the inherent process characteristics of welds (periodic fish-scale pattern) for frequency domain diagnosis. Even in cases of partial loss of depth data (blind spots), the presence of defects can still be inferred by analyzing the integrity of the texture period. This method exhibits extremely high sensitivity in identifying micro-cracks or incomplete fusion defects that, while not causing large-area voids, interrupt the weld texture direction. Simultaneously, it can accurately identify thin oxide films that merely cover the texture but do not damage its structure.

[0056] Specifically, a multimodal signal sample library containing 5000 sets of typical weld features was first constructed. This sample library covers real tower data collected from multiple operating wind farms, including 3000 sets of manually verified silicon-manganese oxide optical interference samples (labeled as negative samples) and 2000 sets of real physical defect samples covering porosity, cracks, and lack of fusion (labeled as positive samples). A support vector machine (SVM) classifier was trained using radial basis function (RBF) kernel function to establish a high-dimensional feature space mapping. After iterative optimization, its classification accuracy on the validation set reached 99.2%.

[0057] During training, a grid search method was used to optimize the hyperparameters of the support vector machine. The search range for the radial basis function kernel parameter was 0.001 to 1, and the search range for the penalty coefficient was 1 to 1000. The final optimal parameter combination was a penalty coefficient of 100 and a kernel parameter of 0.01. To address the imbalance problem (more oxide film samples than true defect samples), a class weight factor was introduced into the loss function, assigning a higher penalty weight to defect samples. The specific performance requirements for the model on the test set were: a recall rate of ≥99.5% for detecting true defects and a false positive rate of less than 0.2%.

[0058] Furthermore, before inputting the data into the classifier, a spatiotemporal parallax consistency check is performed: a spatiotemporal data cube containing 3 to 7 frames is constructed using multi-view sequences generated by the wall-climbing robot's micro-movements. A sparse optical flow algorithm is used to track the motion of feature points at the edges of abnormal regions, and the relative displacement vector stability index of these feature points is calculated. If the stability index is less than 0.5, meaning the feature point drifts drastically with slight changes in viewpoint, it is determined to be a specular reflection virtual image; if the stability index is greater than 0.8, meaning the feature point's spatial position is stable, it is determined to be a physical defect. This index is used as a confidence parameter and jointly input with the aforementioned feature vectors into the classifier.

[0059] In actual operation of the wall-climbing robot, when a reflective spot located at the crest of a weld seam was detected, the algorithm calculated a reflection saturation distribution index of 0.88 in real time (indicating strong specular reflection), while the principal component eigenvalue in the geometric-reflectivity feature vector was only 0.12 (indicating that depth changes are highly uncorrelated with light intensity changes). The classifier combined these features and input them into the mapping function, determining that it falls within the decision boundary of "strong specular reflection and low optical-mechanical correlation." It immediately generated an attribute discrimination indicator (status code: 0) representing an "optical virtual image," instructing the backend system to automatically filter out this data gap and avoid a false alarm.

[0060] Next, when a microcrack at the weld toe, barely perceptible to the naked eye, was detected, the input features showed a reflectance saturation distribution index of 0.25 (consistent with diffuse metallic reflection characteristics), while the principal component eigenvalue of the geometry-reflectivity feature vector was as high as 0.94 (indicating a high alignment between the geometric depth abrupt change and the edge of the optical shadow). The classifier determined that it fell into the region of "diffuse scattering and high photomechanical correlation," generating an attribute discrimination indicator (status code: 1) representing "material deficiency," and then locked the coordinates and initiated the subsequent physical size quantization process.

[0061] Before using the classification model to identify the physical properties of signal anomaly regions, a spatiotemporal parallax consistency verification is performed: Utilizing the mechanical micro-motion characteristics of the wall-climbing adsorption robot during its operation, continuous multi-frame three-dimensional contour data and light intensity reflectivity distribution data for the same weld area are acquired to construct a spatiotemporal data cube; feature extraction is performed on the spatiotemporal data cube to lock the edge contour of the signal anomaly region to be determined as a tracking feature set; the sparse optical flow algorithm is used to track the motion trajectory of this edge contour in continuous multi-frame images, calculating its relative displacement vector relative to the weld background texture; if the relative displacement vector shows that the boundary of the region undergoes drastic non-monotonic drift or morphological flickering with slight changes in the detection angle, it is determined to conform to the instability characteristics of specular reflection virtual images; if the relative displacement vector approaches zero or maintains a linear translation consistent with the scanning motion, it is determined to be a physically existing real defect, and this verification result is used as input features into the classification model.

[0062] This invention transforms the unavoidable mechanical vibrations of wall-climbing adsorption robots operating at heights into a detection advantage. By introducing a time dimension to construct a spatiotemporal data cube, it utilizes the extremely strong angular sensitivity of the oxide film's specular reflection—that is, a slight change in the incident angle causes a drastic shift in the position of the bright spot, creating a stark contrast with the spatial stability of real defects. This spatiotemporal consistency verification mechanism significantly improves robustness against dynamic lighting interference, effectively eliminating false signals that appear as defects in a single frame but are unstable across multiple frames.

[0063] This invention utilizes a classification model trained on a "multimodal signal sample library" to automate and intelligently manage the discrimination process. Compared to traditional techniques that rely on manually setting fixed thresholds (e.g., grayscale > 85%), the machine learning-based classifier can adapt to changes in optical characteristics under different tower materials and welding processes. It solidifies expert discrimination experience into an algorithm model, solving the problems of strong subjectivity and poor consistency in manual detection, and ensuring the uniformity of detection standards.

[0064] Further, the steps of identifying defects based on the attribute discrimination indicator and performing surface fitting repair on the areas marked as pseudo-defects include: locking the boundary of the area marked as an optical virtual image by the discrimination indicator; extracting the effective three-dimensional contour data of the diffuse reflection zone of normal weld metal outside the boundary as a reference topological control point; using a non-uniform rational B-spline surface reconstruction algorithm to deduce and fill the virtual surface morphology data within the optical virtual image area based on the reference topological control point; generating a reconstructed three-dimensional weld model after eliminating optical interference to restore the continuous topological structure of the weld surface.

[0065] Specifically, after marking an irregular data-deficient area of ​​approximately 25 square millimeters in the weld reinforcement region as an "optical virtual image" (i.e., a pseudo-defect), the repair subroutine immediately starts. First, a boundary locking operation is performed, and a morphological edge detection algorithm is used to precisely depict the pixel-level contour of this data-deficient area, defining it as a "topological void to be repaired." Then, using this contour as a reference, an equidistant expansion of 3.0 millimeters is performed outwards into the normal weld region, extracting approximately 1200 valid three-dimensional contour data points within the annular buffer zone. These data points not only contain information on the diffuse reflection height of the weld metal but also implicitly represent the radial and axial curvature variation trends of the weld reinforcement cap (reinforcement height), and are established as "reference topological control points."

[0066] The selection of reference topological control points must meet the quality filtering conditions: eliminating noise points with excessively large gradient values ​​and light intensity saturation points. During the fitting process of the non-uniform rational B-spline surface, bicubic spline basis functions are used, and a tangent plane continuity constraint is introduced. Specifically, at the defect boundary, the angle between the normal vector of the tangent plane of the reconstructed surface and the normal vector of the surrounding normal weld surface is forced to be less than 0.5 degrees. The control vertices are iteratively solved using the least squares method until the root mean square error between the fitted surface and the reference control points converges to within 0.02 mm, thereby ensuring that the repaired virtual surface will not produce non-physical wrinkles or steps.

[0067] Next, the Non-Uniform Rational B-Spline (NURBS) surface reconstruction engine was invoked. The engine used the aforementioned control points to construct a weight matrix, and under the constraint of maintaining the continuity of the boundary tangent vector (G1 continuity), performed smooth interpolation extrapolation into the cavity. This generated a virtual surface patch that perfectly matched the curvature of the surrounding weld. This patch contained approximately 800 virtual height coordinate values, and the normal deviation between its fitted surface and the theoretical weld surface was strictly controlled within 0.02 mm. Finally, this virtual data was backfilled into the original depth map, replacing the originally invalid NaN values. After this processing, the "deep pit" that was originally presented on the digital model due to oxide film reflection was successfully restored to a full, continuous weld protrusion. The generated reconstructed 3D weld model not only eliminated visual noise but also provided a high-quality digital twin base without geometric singularities for subsequent tower stress concentration calculations based on finite element analysis (FEA), ensuring the accuracy of the structural health assessment.

[0068] This invention introduces NURBS surface reconstruction, achieving a leap from "defect detection" to "morphology restoration." For oxide film regions misidentified as defects, instead of ignoring them, they are actively repaired. The advantage of this is that it provides a topologically continuous and surface-smooth "digital twin model" for subsequent wind turbine tower stress concentration analysis (FEA), avoiding stress calculation singularities caused by data gaps and enhancing the engineering application value of the data.

[0069] Further, the steps of quantifying the physical dimensions of areas identified as actual defects and outputting weld defect detection results specifically include: based on the reconstructed 3D weld model, extracting the geometric parameters of all areas marked as material missing, and constructing a multidimensional defect feature vector containing the maximum normal depth, opening span, volume equivalent, and spatial coordinates; calling a pre-set wind turbine tower weld quality grading database, which stores physical limit thresholds defined based on industry standards; mapping the multidimensional defect feature vector to the physical limit thresholds, executing defect severity grading logic, and determining the hazard level of each actual defect; using the reconstructed 3D weld model as a base, performing texture mapping on the determined hazard level and spatial coordinates to generate a weld defect topology distribution map; and outputting the weld defect topology distribution map and corresponding structured physical attribute data as the weld defect detection results.

[0070] Specifically, in the reconstructed 3D model of the clean weld, a real defect region identified as "material missing" was first located at a vertical height of 45.2 meters in the tower. Based on sub-millimeter precision point cloud data, the algorithm automatically calculated the geometric deviation of this defect relative to the fitted surface of the base material, constructing a precise multi-dimensional defect feature vector. This vector includes: a maximum normal depth of 1.25 mm, an opening span of 3.6 mm, a material missing volume equivalent of 4.2 cubic millimeters, and spatial coordinates in a global coordinate system. .

[0071] Subsequently, the core processor calls a pre-built wind turbine tower weld quality grading database, which embeds the ISO 5817:2014 weld quality grading standard. According to the standard's definition for 20 mm thick steel plates, an undercut depth exceeding 0.5 mm is a warning threshold, and exceeding 1.0 mm is a scrap threshold. The specific defect hazard level classification logic is as follows: Level 1 severe defect, corresponding to scrap or immediate shutdown status, defined as the detection of a through crack, or a single defect with a maximum normal depth greater than 1.0 mm, or a single pore volume greater than 5 cubic millimeters; Level 2 medium defect, corresponding to the need for repair welding status, defined as a defect depth between 0.5 mm and 1.0 mm, or a crack length between 1.5 mm and 3.0 mm; Level 3 minor defect, corresponding to the periodic monitoring status, defined as a defect depth less than or equal to 0.5 mm and a volume less than 1 cubic millimeter; Level 4 is the qualified status, i.e., no defects detected. Based on the extracted multidimensional geometric parameters, the above rules are automatically mapped to generate structured physical attribute data containing defect ID and level color code.

[0072] The mapping comparison logic was executed, and it was found that the measured depth of 1.25 mm had exceeded the scrap limit. Therefore, the hazard level of the defect was determined to be "Level I serious defect".

[0073] Finally, using the reconstructed 3D weld model as a digital base, texture mapping technology was employed to render the defective area as a highlighted red warning block, along with dimension labels, generating an intuitive weld defect topology distribution map. This map, along with a structured JSON file containing all physical attribute data, was synchronously output to the ground maintenance terminal, enabling engineers to accurately grasp the 3D shape and precise location of the defect without climbing the tower, providing indisputable data support for developing grinding and repair welding plans.

[0074] This invention generates a "weld defect topology distribution map" and structured data, achieving standardized output of detection results. It maps abstract algorithm results to physical parameters (such as volume and depth) that conform to industry standards, allowing maintenance personnel to directly use the detection conclusions for decision-making. This intuitive and structured output format establishes a data link from "intelligent detection" to "intelligent operation and maintenance."

[0075] This invention innovatively introduces light intensity reflectivity data as a verification dimension for depth data by constructing a "synchronous geometric-optical dual response signal," thus solving the problem of optical characteristic interference. Without changing the main structure of the lidar hardware, through in-depth mining at the signal level, it effectively distinguishes between "invisible" (oxide film specular reflection) and "absent" (material physical defects), thereby significantly reducing the false alarm rate of wind turbine tower detection from the industry average of 20%-30% to less than 1%, greatly reducing the cost of ineffective manual verification.

[0076] Example 2:

[0077] This embodiment provides a lidar-based system for detecting weld defects in wind turbine towers. The system mainly consists of a data acquisition module, an anomaly analysis module, and a defect identification module, and is implemented on a specific hardware platform. The hardware carrier of the system includes a magnetic adsorption wall-climbing platform, an optical detection module, and a signal processing unit. The magnetic adsorption wall-climbing platform uses a four-wheel independently driven permanent magnet adsorption chassis, with a single-wheel adsorption force set at 800N, capable of stably adsorbing and moving a 15kg effective load on the vertical and inverted conical tower wall surfaces. The core sensor of the optical detection module is a customized high-frequency line structured light lidar, mounted on a two-axis gimbal at the front of the robot. The optical axis and the weld normal direction are set at a preset incident angle of 22.5°, which effectively avoids more than 85% of direct sunlight entering the pupil due to planar specular reflection. The signal processing unit is equipped with an NVIDIA Jetson AGX Orin edge computing module, responsible for running the algorithm logic for anomaly analysis and defect identification.

[0078] As one embodiment of the present invention, refer to Figure 1 A flowchart of a method for detecting weld defects in wind turbine towers based on lidar, referring to...Figure 2 A structural diagram of a wind turbine tower weld defect detection system based on lidar, referring to... Figure 3 A schematic diagram of the data processing and feature fusion architecture.

[0079] The data acquisition module utilizes the aforementioned optical detection module to perform specific acquisition tasks. This module drives the built-in semiconductor laser of the lidar to emit a blue monochromatic coherent beam with a wavelength of 450 nanometers, and the laser power is dynamically set between 20 and 50 milliwatts. This module leverages the high sensitivity of short-wavelength blue light to the microscopic surface finish of metals, exciting a typical "dual-mode" optical response upon contact with a smooth oxide precipitate film: specular reflection saturation during perpendicular incidence and total internal reflection escape (signal cutoff) during oblique incidence. The module employs a high dynamic range (HDR) CMOS image sensor with a dynamic range exceeding 120 dB, simultaneously capturing the diffuse reflection spot morphology and energy distribution within a 50-microsecond single exposure time window using 4096 lateral pixel units. By using the hardware trigger signal emitted by the high-precision magnetic encoder of the wall-climbing robot every 0.1 mm of movement, this module forcibly locks the height coordinate information and surface reflectivity value collected within the same exposure frame, ensuring that the timestamp deviation between the two is less than 1 microsecond. This allows for the synchronous demodulation of spatiotemporally aligned geometric-optical dual response signals, constructing a dataset of physical properties of the weld surface that scans 2000 contours per second.

[0080] The anomaly analysis module processes the collected data in real time. First, spatial topology analysis is performed on a high-density 3D point cloud with a point spacing of 0.05 mm. When the angle between the normal vectors of adjacent sampling points exceeds 45 degrees or the height step exceeds 0.5 mm, discontinuous regions are identified. By setting an aggregation radius of 1.0 mm, discrete signals are aggregated into connected components, defining optical blind spots caused by total internal reflection escape or geometric shadow areas caused by occlusion. For signal anomalous regions, this module performs optical feature extraction based on the region of interest: the neighborhood is truncated with the centroid of the anomalous region as the center, the co-occurrence ratio of saturated highlight areas and noise cutoff areas is statistically analyzed, and the reflection saturation distribution index is generated by combining the texture optical uniformity of the effective pixel area. The geometric shape gradient field of the 3D contour data and the light intensity energy gradient field of the light intensity reflectivity distribution data are calculated respectively, a geometric-light intensity correlation matrix is ​​constructed and the principal component feature values ​​are extracted to generate a geometric-reflectivity feature vector. Spatiotemporal parallax consistency verification is performed. The sparse optical flow algorithm is used to track the motion trajectory of the edge contour of the anomalous region in consecutive frames, and the displacement vector is calculated to quantify its stability relative to the background texture, generating a stability confidence feature.

[0081] The defect identification module also runs within the signal processing unit, outputting the final detection results. This module incorporates a Support Vector Machine (SVM) classifier trained on 5000 sets of typical weld features (including 3000 oxide film samples and 2000 real defect samples). The module inputs the reflectance saturation distribution index, geometric-reflectivity feature vector, and spatiotemporal parallax stability confidence score into the classifier. When the determination result is "optical virtual image," the module initiates a Non-Uniform Rational B-Spline (NURBS) surface reconstruction algorithm, using reference topological control points from external normal weld data to generate a virtual surface patch with a fitting deviation controlled within 0.02 mm for filling and repair. When the determination result is "material missing," the module calculates the maximum normal depth, opening span, and volume equivalent of the defect, and compares it with a pre-set ISO5817:2014 standard database. If the defect depth exceeds the standard threshold (e.g., 1.0 mm), the module determines it as a serious defect and uses texture mapping technology to generate a highlighted weld defect topology distribution map and structured physical attribute data, which is then output to the ground maintenance terminal via wireless network.

[0082] This system achieves precise spatiotemporal synchronization of geometric and optical data through the data acquisition module, providing multi-dimensional physical evidence for defect judgment and overcoming the limitation that single-depth data cannot identify materials. The anomaly analysis module constructs a reflection saturation distribution index and a geometric-reflectivity feature vector to mathematically explore the differences between the optical fingerprint of oxide film specular reflection and the physical characteristics of real defects, effectively solving the problem of point cloud missing misjudgment caused by strong reflective interference. The defect identification module uses a classification model to achieve intelligent and high-precision discrimination, significantly reducing the false alarm rate. This enables the system to adapt to the complex optical environment of in-service towers, improving the automation level and detection confidence of wind power operation and maintenance.

[0083] 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 of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting weld defects in wind turbine towers based on lidar, characterized in that, include: The lidar sensor emits a probe beam toward the weld surface and receives the echo signal, simultaneously demodulating the spatiotemporally aligned geometric-optical dual response signal; The geometric-optical dual response signal includes three-dimensional contour data and light intensity reflectivity distribution data; Signal continuity analysis is performed on the three-dimensional contour data to extract signal anomaly regions; based on the light intensity reflectance distribution data, the light signal scattering behavior in the signal anomaly regions is analyzed to construct a reflection saturation distribution index; the spatial change rate of the three-dimensional contour data and the light intensity change rate of the light intensity reflectance distribution data are calculated, and the spatial change rate and the light intensity change rate are spatially correlated and coupled to generate a geometric-reflectance feature vector; Based on the reflection saturation distribution index and the geometric-reflectivity feature vector, a classification model is used to identify the physical properties of signal abnormal regions, generate attribute discrimination indicators, and perform defect identification; among them, regions with optical mirror interference as physical properties are judged as pseudo-defects, and regions with material physical defects as physical properties are judged as real defects.

2. The method for detecting weld defects in wind turbine towers based on lidar according to claim 1, characterized in that, The steps for synchronously demodulating spatiotemporally aligned geometric-optical dual-response signals include: driving a lidar sensor to emit a monochromatic coherent beam in the blue light band, which is reflected and scattered upon contact with the weld surface; using a photoelectric sensing array to simultaneously sense the spatial distribution of diffuse reflection spots on the weld surface and the integral intensity of echo energy within a single exposure time window; based on the principle of optical triangulation, resolving the spatial distribution of diffuse reflection spots into height coordinate information in three-dimensional contour data; based on the photoelectric conversion effect, mapping the integral intensity of echo energy into surface reflectivity values ​​in light intensity reflectivity distribution data; and using a hardware clock synchronization triggering mechanism to lock the height coordinate information and surface reflectivity values ​​collected at the same time, constructing a point-to-point mapped dataset of weld surface physical properties.

3. The method for detecting weld defects in wind turbine towers based on lidar according to claim 1, characterized in that, The steps of performing signal continuity analysis on the three-dimensional contour data and extracting signal anomalous regions include: performing spatial topology analysis on the three-dimensional contour data to identify discontinuous regions whose spatial curvature exceeds a preset physical deformation threshold; performing signal dispersion aggregation on the discontinuous regions to aggregate adjacent signal missing points whose dispersion meets a preset condition into connected regions; and filtering out connected regions whose area exceeds a preset area threshold to define them as the signal anomalous regions. The physical manifestations of the signal anomalous regions include: optical blind zones where the sensor does not receive effective echo signals due to total internal reflection of the light beam, and geometric shadow zones where the light beam is blocked due to surface geometric depressions.

4. The method for detecting weld defects in wind turbine towers based on lidar according to claim 1, characterized in that, The steps for constructing the reflectance saturation distribution index based on the light intensity reflectance distribution data include: Centered on the geometric centroid of the signal anomaly region, a region of interest (ROI) covering the signal anomaly region and the surrounding effective weld surface is extracted from the light intensity reflectivity distribution data. A histogram of light intensity energy distribution within the ROI is plotted, dividing the light intensity data into a saturated highlight region, a noise cutoff region, and a diffuse reflection median region. The spatial distribution optical uniformity index of effective pixels within the ROI, excluding the noise cutoff region, is calculated. Through a weighted combination, the proportion of the saturated highlight region, the proportion of the noise cutoff region, and the optical uniformity index are combined to generate a reflection saturation distribution index. This reflection saturation distribution index reflects whether the target region exhibits the unique binary optical characteristics of a oxide film, characterized by a coexistence of high reflectivity and total reflection blind zones.

5. The method for detecting weld defects in wind turbine towers based on lidar according to claim 1, characterized in that, The steps of calculating the spatial rate of change of three-dimensional contour data and the rate of change of light intensity reflectance distribution data to generate a geometric-reflectance feature vector include: calculating the geometric shape gradient field of the three-dimensional contour data along the scanning direction and the light intensity energy gradient field of the light intensity reflectance distribution data along the scanning direction; constructing a geometric-light intensity correlation matrix describing the spatial cooperative change relationship between the geometric shape gradient field and the light intensity energy gradient field; performing eigenvalue decomposition on the geometric-light intensity correlation matrix to extract the principal component eigenvalues ​​characterizing the degree of alignment between the two boundaries; and concatenating the geometric shape gradient field, the light intensity energy gradient field, and the principal component eigenvalues ​​to generate the geometric-reflectance feature vector; the feature vector describes the consistency between the physical boundary and the optical boundary of the target region in a multi-dimensional feature space.

6. The method for detecting weld defects in wind turbine towers based on lidar according to claim 1, characterized in that, The steps of identifying the physical properties of the signal anomaly region using a classification model and generating attribute discrimination indicators include: establishing a multimodal signal sample library containing samples of optical interference from oxide films and samples of real physical defects; training a support vector machine classifier using the multimodal signal sample library to establish a mapping relationship from input features to physical properties; inputting the reflectance saturation distribution index and geometric-reflectivity feature vector of the region to be tested into the classifier; if the classifier's determination result matches the characteristics of strong specular reflection and low optomechanical correlation, generating an attribute discrimination indicator characterizing an optical virtual image; if the classifier's determination result matches the characteristics of diffuse reflection and high optomechanical correlation, generating an attribute discrimination indicator characterizing material deficiencies.

7. The method for detecting weld defects in wind turbine towers based on lidar according to claim 1, characterized in that, The steps of identifying defects based on the attribute discrimination indicator and performing surface fitting repair on the areas marked as pseudo-defects include: locking the boundary of the area marked as an optical virtual image by the discrimination indicator; extracting the effective three-dimensional contour data of the diffuse reflection zone of normal weld metal outside the boundary as a reference topological control point; using a non-uniform rational B-spline surface reconstruction algorithm to deduce and fill the virtual surface morphology data within the optical virtual image area based on the reference topological control point; and generating a reconstructed three-dimensional weld model after eliminating optical interference to restore the continuous topological structure of the weld surface.

8. The method for detecting weld defects in wind turbine towers based on lidar according to claim 1, characterized in that, The steps for quantifying the physical dimensions of areas identified as real defects and outputting weld defect detection results specifically include: based on the reconstructed 3D weld model, extracting the geometric parameters of all areas marked as material missing, and constructing a multidimensional defect feature vector containing the maximum normal depth, opening span, volume equivalent, and spatial coordinates; calling a pre-set wind turbine tower weld quality grading database, which stores physical limit thresholds defined based on industry standards; mapping the multidimensional defect feature vector to the physical limit thresholds, executing defect severity grading logic, and determining the hazard level of each real defect; using the reconstructed 3D weld model as a base, performing texture mapping on the determined hazard level and spatial coordinates to generate a weld defect topology distribution map; and outputting the weld defect topology distribution map and corresponding structured physical attribute data as the weld defect detection result.

9. A system for detecting weld defects in wind turbine towers based on lidar, characterized in that, include: Data acquisition module: The lidar sensor emits a probe beam to the weld surface and receives the echo signal, and simultaneously demodulates the spatiotemporally aligned geometric-optical dual response signal; the geometric-optical dual response signal includes three-dimensional contour data and light intensity reflectivity distribution data; Anomaly Analysis Module: Performs signal continuity analysis on the three-dimensional contour data to extract signal anomaly regions; analyzes the light signal scattering behavior within the signal anomaly regions based on the light intensity reflectance distribution data, constructs a reflectance saturation distribution index, and determines whether there are anomalies caused by oxide films in the signal anomaly regions; calculates the spatial change rate of the three-dimensional contour data and the light intensity change rate of the light intensity reflectance distribution data, and spatially couples the two to generate a geometric-reflectance feature vector; Defect identification module: Based on the reflection saturation distribution index and the geometric-reflectivity feature vector, the module uses a classification model to identify the physical properties of signal abnormality areas and generates attribute discrimination indicators for defect identification; among them, areas with the attribute of optical mirror interference are judged as pseudo defects, and areas with the attribute of material physical deficiency are judged as real defects.

Images