A weld 3D image recognition method and system

Abstract

The application relates to the technical field of weld detection, in particular to a weld 3D image recognition method and system. The method comprises the following steps: scanning a weld to obtain a three-dimensional point cloud, and extracting a geometric consistency feature set of the weld from the three-dimensional point cloud; establishing a standard geometric consistency feature set of an ideal weld model; comparing the geometric consistency feature set with the standard geometric consistency feature set to generate a deviation heat map and a deformation score of the weld; and adjusting a preset weld defect judgment logic according to the deviation heat map and the deformation score. The method aims to solve the problems of false positives, missed detection, reduced detection accuracy and reliability of the existing weld three-dimensional image recognition method.

Description

Technical Field

[0001] This application relates to the field of weld inspection technology, and more specifically, to a method and system for 3D image recognition of welds. Background Technology

[0002] In modern industrial manufacturing, automated welding relies on 3D cameras and line lasers to scan weld seams and generate point clouds for geometric measurement and quality assessment. However, in high-speed, long-cycle continuous production environments, wear inevitably occurs in transmission components such as robot arms or linear modules, causing small, nonlinear deviations in the actual scanning trajectory as the path changes. These deviations, combined with pre-calibrated coordinate transformations, result in position-dependent cumulative errors when mapping the point cloud from the camera coordinate system to the workpiece coordinate system. This leads to slight bending, unevenness, and other distortions in the reconstructed 3D model relative to the actual weld seam. To accommodate long weld seams, segmented scanning and stitching are often used on-site; however, each segment of data itself carries the aforementioned errors, making it difficult for alignment and fusion algorithms to achieve high-precision seamless connections. Geometric discontinuities or local distortions easily occur in the stitched areas.

[0003] Existing defect judgment logic relies on an initial high-precision model and fixed thresholds, lacking the ability to self-evaluate and adapt to the geometric accuracy of the input model. Furthermore, frequent product model and process changes in factories mean that rule sets often rely on rapid adjustments based on experience, frequently ignoring the ongoing impact of equipment wear on data accuracy. This results in two typical biases: first, geometrical abrupt changes caused by splicing distortion are misjudged as out-of-tolerance defects, leading to false alarms, rework, and even scrap; second, real but small-scale defects (such as minor edge chipping or surface porosity) are masked by model distortion, making the features less significant and resulting in missed detections. In summary, equipment wear, cumulative errors from coordinate transformation, and segment splicing mismatch collectively weaken the reliability of the 3D recognition link, significantly reducing the detection accuracy and reliability of traditional methods on high-speed production lines, making it difficult to meet stringent quality control requirements.

[0004] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0005] This application discloses a 3D image recognition method and system for welds, aiming to solve the problems of existing 3D image recognition methods for welds in high-speed, continuous production environments, such as scanning trajectory deviation caused by wear of mechanical parts, systematic cumulative errors, and resulting geometric discontinuities or local distortions in the 3D weld model, which in turn lead to false alarms and missed detections in defect judgment logic, reducing the accuracy and reliability of detection.

[0006] The technical solution of this application is as follows: In a first aspect, this application discloses a 3D image recognition method for welds, comprising: The weld is scanned to obtain a three-dimensional point cloud, and the geometric consistency feature set of the weld is extracted from the three-dimensional point cloud. Establish a standard geometric consistency feature set for an ideal weld model; By comparing the geometric consistency feature set with the standard geometric consistency feature set, a deviation heatmap and deformation score of the weld are generated. The preset weld defect judgment logic is adjusted based on the deviation heat map and deformation score.

[0007] Furthermore, the weld 3D image recognition method also includes: Scan and acquire raw optical response data of the weld area; Microscopic optical features of the weld surface were extracted from the raw optical response data; The microscopic optical features of the weld surface are compared with a pre-set optical fingerprint database of deformation sources to identify the source of deformation. Adjust the weld defect judgment logic based on the source of deformation.

[0008] Furthermore, the microscopic optical features of the weld surface are compared with a pre-defined optical fingerprint database of deformation sources to identify the sources of deformation, including: Multi-dimensional feature enhancement processing is performed on the microscopic optical features of the weld surface to obtain enhanced features; Obtain the similarity between enhanced features and fingerprints from different deformation sources in the deformation source optical fingerprint database; Determine whether the similarity scores are all higher than a preset threshold and whether the differences between similar scores are less than a preset distinction threshold; When the similarity is higher than the preset judgment threshold and the difference between the similarities is less than the preset distinction threshold, auxiliary information of the weld area is collected. This auxiliary information includes equipment status and calibration / trajectory deviation, material batch and welding process parameters, thermal process data and MES / ERP records. The weld geometry and material combination are obtained. Based on the preset feature sensitivity weight table corresponding to the weld geometry and material combination, the enhancement features are weighted to obtain the weighted enhancement features. Based on the comparison results between the weighted enhanced features and the optical fingerprint database of deformation sources, and combined with auxiliary information, the source of deformation is obtained.

[0009] Furthermore, based on the comparison results between the weighted enhanced features and the optical fingerprint database of deformation sources, and combined with auxiliary information, the sources of deformation are obtained, including: Assess the reliability of the comparison results and the reliability of the auxiliary information; Adjust the weights of the comparison results and the auxiliary information based on the comparison confidence and the auxiliary confidence. Based on the adjusted weights of the comparison results and the weights of the auxiliary information, the comparison results and auxiliary information are integrated to obtain the source of deformation.

[0010] Furthermore, the 3D image recognition method for weld seams also includes: The three-dimensional point cloud is divided into multiple local regions along the centerline of the weld. Based on the original optical response data and point cloud data, the laser line width, the standard deviation of the laser intensity distribution, the local point cloud density, and the local surface fitting residual are combined according to preset weights to obtain the data reliability score of each local region. When extracting the geometric consistency feature set, the point cloud samples used to extract the geometric consistency feature set are weighted according to the data confidence score to obtain the confidence-weighted geometric consistency feature set. When generating the deviation heatmap, the deviation of each local area is weighted and mapped according to the data reliability score. When calculating the deformation score, the deviations of each local area are weighted and summarized based on the data reliability score.

[0011] Furthermore, the reliability of the comparison results and the reliability of the auxiliary information are assessed, including: Real-time monitoring of the working status of production equipment and the operating status of auxiliary information sources; The reliability of the comparison results is dynamically calculated based on the real-time working status and the real-time quality indicators of the raw optical response data. The reliability of auxiliary information is dynamically calculated based on the operating status of the auxiliary information source, the frequency of real-time data updates, the data integrity, and the degree of first match between the auxiliary information source and the current weld production batch or product model.

[0012] Furthermore, the 3D image recognition method for weld seams also includes: Acquire information on changes in the production line and calculate the damping coefficient and response time constant used to adjust the comparative reliability and auxiliary reliability. This information includes at least equipment status fluctuation indicators, quality fluctuation indicators of raw optical response data, and switching events related to the current weld production batch or product model. When the change information is lower than the preset first threshold, the fluctuations of the comparison reliability and the auxiliary reliability are smoothed according to the damping coefficient and the response time constant. When the change information exceeds the preset first threshold, the calculation frequency of the comparison reliability and auxiliary reliability is increased based on the damping coefficient and response time constant, and then limited within the preset upper limit of frequency. Based on the adjusted comparative confidence and auxiliary confidence, a strategy for maintaining the calculation frequency is determined.

[0013] Furthermore, the damping coefficients and response time constants used to adjust the contrast confidence level and auxiliary confidence level are calculated, including: Multi-scale time-frequency analysis is performed on the change information to identify the nonlinear characteristics, abrupt change points, and multi-peak distribution of the change information; Based on nonlinear characteristics, abrupt change points, and multi-peak distribution, adjust the weights and structure of the preset adjustment parameter change model; Based on the adjusted adjustment parameter variation model, the damping coefficient and response time constant are calculated; Based on the preset parameter change trend prediction rules, the damping coefficient and response time constant are corrected.

[0014] Furthermore, multi-scale time-frequency analysis is performed on the change information, including: Identify the production parameter characteristics of the current weld batch or product model; Obtain the switching information of weld production batch or product model in the change information, load the historical production parameter characteristics corresponding to the current weld production batch or product model, and preset the parameters for multi-scale time-frequency analysis based on the historical production parameter characteristics; Based on the characteristics of production parameters, the basis function type and decomposition level of the multi-scale time-frequency analysis are dynamically selected, and the frequency resolution of each scale of the multi-scale time-frequency analysis is adjusted. The system determines the second degree of matching between the production characteristic parameters and the preset parameter characteristics. When the second degree of matching is lower than the preset second threshold, it triggers parameter readjustment through multi-scale time-frequency analysis.

[0015] Secondly, this application also discloses a 3D image recognition system for welds, comprising: The feature extraction module is used to scan the weld to obtain a three-dimensional point cloud and extract the geometric consistency feature set of the weld from the three-dimensional point cloud. The benchmark establishment module is used to establish a standard geometric consistency feature set for an ideal weld model; The deformation assessment module is used to compare the geometric consistency feature set with the standard geometric consistency feature set to generate a deviation heat map and deformation score for the weld. The judgment logic adjustment module is used to adjust the preset weld defect judgment logic based on the deviation heat map and deformation score.

[0016] Beneficial effects This application discloses a 3D image recognition method for welds. The method involves scanning the weld to obtain a 3D point cloud and extracting a set of geometric consistency features from it. A standard set of geometric consistency features for an ideal weld model is then established. The geometric consistency feature set is compared with the standard set to generate a deviation heatmap and deformation score for the weld. Finally, the preset weld defect judgment logic is adjusted based on the deviation heatmap and deformation score. This method effectively solves the problem in existing technologies where scanning trajectory deviations and systematic cumulative errors caused by mechanical component wear lead to geometric discontinuities or local distortions in the 3D weld model, resulting in false alarms and missed detections in the defect judgment logic. By quantifying the degree of weld deformation and dynamically adjusting the defect judgment logic, this application significantly improves the accuracy and reliability of weld detection, avoiding unnecessary rework or product scrapping, and overcoming the shortcomings of existing technologies where the accuracy and reliability of detection results are greatly reduced. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a 3D image recognition method for weld seams provided in this application.

[0018] Figure 2 A flowchart of a weld seam 3D image recognition system provided in this application.

[0019] In the diagram: 1. Feature extraction module; 2. Benchmark establishment module; 3. Deformation evaluation module; 4. Judgment logic adjustment module. Detailed Implementation

[0020] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0021] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0022] Reference Figure 1 This application proposes a 3D image recognition method for welds, comprising: S1000: Scans the weld to obtain a 3D point cloud, and extracts the geometric consistency feature set of the weld from the 3D point cloud; S2000: Establishes a standard geometric consistency feature set for an ideal weld model; S3000: Compares the geometric consistency feature set with the standard geometric consistency feature set to generate a deviation heat map and deformation score for the weld. S4000: Adjusts the preset weld defect judgment logic based on the deviation heat map and deformation score.

[0023] Specifically, a "3D point cloud" refers to a dataset consisting of a large number of points with three-dimensional coordinate information, obtained through 3D scanning equipment (such as a line laser scanner). These points represent the geometry of the weld surface.

[0024] The "geometric consistency feature set" refers to a set of features extracted from a 3D point cloud that can describe the geometry and structural characteristics of a weld, such as weld height, width, bevel angle, and cross-sectional profile. Ideally, these features should possess a certain degree of stability or predictability.

[0025] An "ideal weld model" refers to a digital model that represents the perfect weld geometry, established based on design specifications or standard process parameters.

[0026] The "standard geometric consistency feature set" refers to the set of geometric features extracted from the ideal weld model and used as a reference.

[0027] A "deviation heatmap" is a visualization tool that visually represents the degree and spatial distribution of deviation between the actual geometry of a weld and an ideal model through color depth or chromatographic changes.

[0028] "Deformation score" is a quantitative indicator used to comprehensively evaluate the degree of geometric deformation of the weld as a whole or in a local area.

[0029] "Weld defect determination logic" refers to a series of rules, algorithms or models used to determine whether a weld has defects. It is usually based on preset geometric size thresholds and morphological features.

[0030] The implementation environment of this application typically includes an automated welding production line equipped with 3D scanning equipment (such as a line laser scanner mounted on a robot), a data processing unit (such as an industrial computer or server), and a control system for storing and executing defect determination logic.

[0031] This application first involves scanning the weld to obtain a three-dimensional point cloud, and then extracting a set of geometric consistency features from the point cloud. In one embodiment, a line laser scanner can be used to scan the weld point-by-point or line-by-line, acquiring three-dimensional point cloud data of the weld surface using the principle of laser triangulation. For example, a line laser scanner can be mounted on a robotic arm, which moves along the weld path while a line laser is projected onto the weld surface and the reflected light is captured by a camera, thereby generating a three-dimensional point cloud of the weld in real time. From this point cloud data, various geometric consistency features can be extracted. For example, by performing cross-sectional analysis on the point cloud data, parameters such as weld height, width, and bevel angle for each cross-section can be calculated, and these parameters can be combined to form a set of geometric consistency features. In another embodiment, a structured light scanner can be used to scan the weld area, acquiring a large area of ​​three-dimensional point cloud at once. Subsequently, point cloud processing algorithms, such as principal component analysis (PCA) or least squares fitting, are used to identify and extract key geometric elements such as the weld centerline and edge contours from the point cloud. Based on these geometric elements, the volume, surface area, curvature, and other characteristics of the weld can be further calculated to form a set of geometric consistency features.

[0032] Secondly, a standard geometric consistency feature set for the ideal weld model is established. In one implementation, an ideal 3D weld model can be constructed using CAD software based on the weld design drawings and process specifications. Subsequently, a set of standard geometric consistency features, such as ideal weld height, width, and bevel angle, is extracted from this ideal 3D model using the same method as extracting the geometric consistency features of the actual weld. These standard features will serve as the benchmark for subsequent comparisons. In another implementation, a high-precision scan can be performed on an actual weld that has undergone rigorous quality control and is deemed "perfect," obtaining its 3D point cloud. This high-precision point cloud is used as the ideal weld model, and its geometric consistency feature set is extracted from it as the standard geometric consistency feature set. This method can better reflect the ideal state in actual production.

[0033] Next, the geometric consistency feature set and the standard geometric consistency feature set are compared to generate a deviation heatmap and deformation score for the weld. In one implementation, the extracted actual weld geometric consistency features can be compared one by one with the standard geometric consistency features. For example, the difference between the actual weld height and the ideal weld height, and the difference between the actual weld width and the ideal weld width, can be calculated. These differences can be mapped onto the 3D model of the weld and a deviation heatmap can be generated using color coding (e.g., green indicates small deviation, red indicates large deviation) to visually display the spatial distribution of the deviation. Simultaneously, a weighted average or statistical analysis of the absolute values ​​of these differences can be performed to obtain a comprehensive deformation score, quantifying the overall deformation degree of the weld. In another implementation, a more complex geometric matching algorithm, such as the Iterative Closest Point (ICP) algorithm or a feature point-based registration algorithm, can be used to align the actual weld point cloud with the ideal weld model. After alignment, the distance from each point in the actual point cloud to the surface of the ideal model is calculated; these distances can be used to generate the deviation heatmap. Deformation scores can be obtained by performing statistical analysis on these distances (such as root mean square error, maximum deviation, etc.).

[0034] Finally, the preset weld defect judgment logic is adjusted based on the deviation heatmap and deformation score. In one embodiment, the corresponding thresholds in the weld defect judgment logic can be dynamically adjusted based on the local high deviation areas shown in the deviation heatmap and the overall deformation degree reflected by the deformation score. For example, if the deviation heatmap shows a slight but systematic bulge in a certain area of ​​the weld and the deformation score is high, the upper limit threshold of the weld height in that area can be appropriately relaxed to avoid false alarms. Conversely, if the deviation heatmap shows an abnormal depression in a certain area and the deformation score is low, the lower limit threshold of the depression depth in that area can be tightened to improve the defect detection rate. In another embodiment, an adaptive model based on machine learning can be established. This model takes the deviation heatmap and deformation score as input and outputs the adjusted defect judgment logic parameters. Through training with historical data, the model can learn how to optimally adjust the defect judgment logic under different deviations and deformation degrees, thereby achieving more intelligent defect identification. For example, when the deformation score indicates that the overall weld has slight shrinkage, the model can automatically adjust the judgment thresholds for weld width and height to adapt to this general deformation.

[0035] This application's 3D image recognition method for welds utilizes an adaptive defect judgment logic adjustment mechanism. Traditional methods rely on preset fixed geometric dimension criteria, failing to perceive and compensate for measurement system errors and minute deformations of the weld itself. When the robot arm's motion trajectory deviates, causing distortion of the 3D point cloud data, traditional logic incorrectly identifies these measurement errors as weld defects or masks the true defects. In contrast, this application first acquires a 3D point cloud by scanning the weld and extracts a set of geometric consistency features from it. These features objectively reflect the actual geometric shape of the weld. Subsequently, a standard set of geometric consistency features for an ideal weld model is established, providing a reliable benchmark for subsequent deviation assessment. Crucially, this application generates a deviation heatmap and deformation score for the weld by comparing the actual geometric consistency feature set with the standard geometric consistency feature set. The deviation heatmap visually displays the degree of deviation of each region of the weld from the ideal state, while the deformation score quantifies the overall deformation level.

[0036] Most importantly, this application can dynamically adjust the preset weld defect judgment logic based on these deviation heatmaps and deformation scores. This means that when the system detects geometric deviations caused by measurement errors or general deformation, it will not simply classify them as defects, but will intelligently adjust the defect judgment threshold or rules to avoid false alarms. At the same time, for real defects that may be masked by model deformation, the detection rate can be improved by finely adjusting the judgment logic, effectively reducing missed detections.

[0037] For example, in existing technologies, if the movement deviation of the robotic arm causes the reconstructed weld model to be generally slightly wider, the defect judgment logic with a fixed threshold may classify all welds with widths slightly exceeding the upper limit as defects, leading to a large number of false alarms. This application, however, identifies this general widening trend based on deviation heatmaps and deformation scores, and adjusts the upper limit for weld width judgment accordingly, thereby avoiding false alarms while still accurately identifying locally abnormally wide defects. Therefore, the weld 3D image recognition method of this application can significantly improve the accuracy and reliability of weld inspection, effectively cope with complex and changing production environments, reduce unnecessary rework and product scrap, and provide a more advanced and intelligent solution for quality control in the industrial manufacturing field.

[0038] This application further proposes that the method also includes: S5000: Scans and acquires raw optical response data of the weld area; S6000: Extracting microscopic optical features of weld surface from raw optical response data; S7000: Compares the microscopic optical features of the weld surface with a pre-set optical fingerprint database of deformation sources to identify the source of deformation; S8000: Adjust the weld defect judgment logic based on the source of deformation.

[0039] The scanning and acquisition of raw optical response data of the weld area refers to the use of optical sensors such as high-resolution cameras, spectrometers, or multispectral imaging devices to perform non-contact scanning of the weld surface to obtain optical information such as reflection, absorption, and scattering under different wavelengths, polarization states, or illumination conditions. The raw optical response data can include information such as the color, brightness, texture, gloss, and spectral characteristics of the weld surface, which reflects the microstructure and material composition of the weld surface.

[0040] Extracting microscopic optical features from raw optical response data refers to analyzing the acquired raw optical response data using techniques such as image processing, signal analysis, or machine learning to identify and quantify microscopic structural features on the weld surface that are difficult to detect with the naked eye. Microscopic optical features can include, but are not limited to, optically distinguishable properties such as surface roughness, grain boundaries, oxide layer thickness, microcracks, porosity, inclusions, microstructural changes in the heat-affected zone, and surface contaminants. These features are the specific optical manifestations of potential deformation sources that lead to weld deformation or defects.

[0041] Subsequently, the microscopic optical features of the weld surface are compared with a pre-established optical fingerprint database of deformation sources to identify the sources of deformation. The optical fingerprint database of deformation sources is a pre-built database that stores typical microscopic optical feature patterns or "fingerprints" of weld surfaces corresponding to known deformation sources (e.g., specific types of material defects, abnormal welding process parameters, thermal stress, oxidation, contamination, porosity, inclusions, etc.). By matching or calculating the similarity between the microscopic optical features extracted from the current weld and the patterns in the fingerprint database, potential sources of deformation causing the current weld deformation or defects can be identified. For example, a specific spectral response pattern may indicate the presence of an oxide layer on the weld surface, while another texture feature may be related to abnormal grain growth within the material.

[0042] Finally, the weld defect judgment logic is adjusted based on the source of deformation. Once the specific source of deformation is identified, the preset weld defect judgment logic can be modified or optimized accordingly. For example, if the deformation is identified as mainly caused by thermal stress, the judgment threshold for certain geometric deviations can be appropriately relaxed, or the sensitivity to cracks in specific directions can be increased; if the deformation is identified as caused by material contamination, it may be necessary to more rigorously inspect surface defects and adjust the corresponding defect level. This adjustment makes defect judgment more intelligent and accurate, avoiding a "one-size-fits-all" judgment method.

[0043] This solution introduces microscopic optical feature analysis of weld surfaces to directly pinpoint the causes of deformation. The system collects raw optical response data, extracts features such as roughness, grain size, and oxide layer / contaminants, and compares this data with a deformation source optical fingerprint database to identify specific sources such as porosity, inclusions, lack of fusion, and hot cracks. Based on this, the defect judgment threshold and feature weights are adjusted according to the source, avoiding a one-size-fits-all approach based solely on geometric deviations, and significantly improving the targeting, accuracy, and reliability of defect identification.

[0044] In some preferred embodiments, the following specific example illustrates the situation: Suppose that during a welding process, a slight depression is found on the weld surface through 3D point cloud scanning, and its geometric deviation score is near the critical value. If the judgment is based solely on the above geometric consistency characteristics, it may be difficult to determine whether the depression is caused by normal cooling shrinkage or by more serious internal defects (such as porosity or slag inclusions).

[0045] At this point, the weld area is scanned using a high-resolution camera to acquire its raw optical response data. Subsequently, the microscopic optical characteristics of the weld surface are extracted from this data, such as surface reflectivity, texture uniformity, and the presence of abnormal spectral absorption peaks. For example, if the analysis reveals an abnormally low surface reflectivity and the presence of specific spectral absorption peaks in the recessed area, these characteristics would highly match the optical fingerprints of "porosity defects" in a pre-defined deformation source optical fingerprint database.

[0046] Based on this, the system identifies the deformation source of the indentation as a "porosity defect." According to this identification result, the preset weld defect judgment logic will be adjusted. For example, the judgment threshold for geometric deviations caused by "porosity defects" may be tightened, or the weight of geometric features related to porosity may be increased when calculating the deformation score. In this way, even slight geometric indentations, once combined with the deformation source of "porosity defects" identified by microscopic optical features, can be more accurately judged as defects by the system, thus avoiding missed judgments that may occur due to relying solely on geometric features, and significantly improving the accuracy and reliability of defect identification.

[0047] Furthermore, in another embodiment of this application, step S7000 includes: S7100: Multi-dimensional feature enhancement processing is performed on the microscopic optical features of the weld surface to obtain enhanced features; S7200: Acquire the similarity between enhanced features and fingerprints from different deformation sources in the deformation source optical fingerprint database; S7300: Determine whether the similarity scores are all higher than the preset judgment threshold and whether the differences between similarity scores are less than the preset distinction threshold; S7400: When the similarity is higher than the preset judgment threshold and the difference between the similarities is less than the preset distinction threshold, auxiliary information of the weld area is collected. The auxiliary information includes equipment status and calibration / trajectory deviation, material batch and welding process parameters, thermal process data and MES / ERP records. S7500: Obtain the weld geometry and material combination, and weight the enhancement features according to the preset feature sensitivity weight table of the corresponding weld geometry and material combination to obtain weighted enhancement features; S7600: Based on the comparison results between the weighted enhanced features and the optical fingerprint database of deformation sources, combined with auxiliary information, the deformation source is obtained.

[0048] The process involves multi-dimensional feature enhancement of the microscopic optical characteristics of the weld surface. This aims to extract and strengthen microstructural, textural, and spectral features closely related to weld deformation by applying techniques such as image processing, signal analysis, or machine learning, thereby obtaining enhanced features. These enhanced features can more clearly characterize the actual state of the weld, providing richer and more discriminative information for subsequent deformation source identification.

[0049] Obtaining the similarity between enhanced features and fingerprints from different deformation sources in the deformation source optical fingerprint database can be achieved by calculating the distance between feature vectors (e.g., Euclidean distance, Mahalanobis distance) or by employing pattern recognition algorithms (e.g., support vector machines, neural networks). The deformation source optical fingerprint database pre-stores standard optical fingerprints corresponding to various known deformation sources (such as thermal stress, mechanical stress, material defects, etc.) as a benchmark for identification.

[0050] In practical applications, determining whether the similarity scores are all above a preset threshold and whether the differences between similarity scores are less than a preset discrimination threshold is for effective screening of preliminary identification results. The preset threshold ensures the reliability of the identification results; that is, only when the similarity reaches a certain level is a corresponding deformation source considered to exist. The preset discrimination threshold assesses the degree of difference in similarity between fingerprints from different deformation sources. When the difference is small, it indicates that accurate differentiation is difficult based solely on optical features, and more information needs to be introduced.

[0051] When the similarity scores are all above a preset judgment threshold, and the differences between similarity scores are all below a preset distinction threshold, it indicates that a single microscopic optical feature of the weld surface is insufficient to accurately distinguish the specific source of deformation. In this case, auxiliary information for the weld area needs to be collected. Auxiliary information includes equipment status and calibration / trajectory deviations, material batches and welding process parameters, thermal process data, and MES / ERP records. This information provides supplementary judgment criteria from multiple dimensions such as production process, equipment operation, and material properties, helping to make more accurate decisions when optical features are blurred.

[0052] Furthermore, the weld geometry and material composition are acquired, and the enhanced features are weighted according to a preset feature sensitivity weighting table for the corresponding weld geometry and material composition to obtain weighted enhanced features. The feature sensitivity weighting table is preset based on the response degree of various microscopic optical features to specific deformation sources under different weld structures and material properties. Through weighting, the recognition process can be made more targeted, improving the recognition accuracy of specific deformation sources.

[0053] Therefore, based on the comparison results between the weighted enhanced features and the optical fingerprint database of deformation sources, combined with auxiliary information, the source of deformation is finally obtained. This means that, based on the comparison of optical features, various auxiliary data in the production process are comprehensively considered to achieve a more accurate and comprehensive identification of the source of deformation.

[0054] This application's solution enhances the microscopic optical features of the weld surface through multi-dimensional feature enhancement, effectively improving feature recognition and enabling more effective capture of subtle deformation-related changes. Subsequently, by calculating the similarity between the enhanced features and fingerprints in the deformation source optical fingerprint database, the possible deformation type of the weld can be preliminarily determined. When optical features alone are insufficient to clearly distinguish the deformation source (i.e., when the similarity is high and the differences are small), auxiliary information is introduced for comprehensive judgment, thereby compensating for the limitations of a single optical feature. Furthermore, by considering the weld's geometry and material composition, and using a feature sensitivity weighting table to weight the enhanced features, the recognition process can be optimized for different working conditions, further improving the accuracy and robustness of the recognition. Therefore, this application can more comprehensively and accurately identify the deformation source of the weld.

[0055] In another embodiment of this application, step S7600 is further proposed to include: S7610: Evaluate the reliability of the comparison results and the reliability of the auxiliary information; S7620: Adjust the weights of the comparison results and the auxiliary information based on the comparison confidence and the auxiliary confidence; S7630: Based on the adjusted weights of the comparison results and the weights of the auxiliary information, the comparison results and auxiliary information are fused to obtain the source of deformation.

[0056] Specifically, assessing the reliability of the comparison results refers to the quantitative evaluation of the reliability of the comparison results between the weighted enhanced features and the deformation source optical fingerprint database. This reliability can be calculated based on various factors, such as the performance indicators of the comparison algorithm, the discriminative power of the comparison features, the stability of the comparison results, and the accuracy of historical data verification. Its purpose is to quantify the reliability of the comparison results. Simultaneously, assessing the reliability of auxiliary information refers to the quantitative evaluation of the reliability, completeness, timeliness, and relevance of the collected auxiliary information (including equipment status and calibration / trajectory deviations, material batches and welding process parameters, thermal process data, and MES / ERP records). For example, the data update frequency, data completeness, and degree of first match with the current weld production batch or product model of the auxiliary information source can all be used as criteria for assessing the reliability of the auxiliary information. Its purpose is to ensure the effectiveness and usability of the auxiliary information.

[0057] The adjustment of the weights of the comparison results and auxiliary information based on the comparison reliability and auxiliary reliability can be understood as dynamically allocating their influence in the final deformation source identification based on the assessment of the reliability of the two information sources. For example, when the comparison reliability is high and the auxiliary reliability is low, the weight of the comparison results will be increased, while the weight of the auxiliary information will be decreased; and vice versa. This dynamic adjustment mechanism ensures that, in the case of uneven information quality, the system can prioritize the more reliable information sources.

[0058] In practical applications, the deformation source is obtained by fusing the adjusted comparison results and auxiliary information based on their respective weights. Specifically, this involves using a weighted fusion algorithm to integrate the weighted comparison results and auxiliary information. For example, weighted averaging, weighted voting, or machine learning-based weighted fusion models can be employed. The aim is to comprehensively utilize the advantages of both information sources, considering their respective reliability, to arrive at the most accurate and reliable deformation source identification result.

[0059] This application's solution effectively addresses the potential blind spots in information fusion inherent in traditional methods by introducing a reliability assessment mechanism for both comparison results and auxiliary information. By quantitatively evaluating the reliability of comparison and auxiliary information, the system can intelligently identify inputs with higher quality and greater reliability, and dynamically adjust their contribution weight in the final deformation source identification accordingly. It is precisely this reliability-based weight adjustment that enables the system to adaptively optimize when faced with data of varying quality or relevance, preventing the inadequacy of a single information source from affecting the overall accuracy of the identification.

[0060] In some preferred embodiments, the following specific example illustrates the situation: In a weld inspection, the comparison results between the weighted enhanced features and the optical fingerprint database of deformation sources showed that the weld deformation had a high similarity to both "welding parameter fluctuations" and "equipment vibration." However, the comparison algorithm's confidence level (comparison confidence level) for "welding parameter fluctuations" was only 0.8, while its confidence level for "equipment vibration" was only 0.6. Meanwhile, the collected auxiliary information (such as equipment status records) showed recent anomalies in the vibration data of the welding equipment, but its data integrity (auxiliary confidence level) was only 0.7, while the welding process parameter records were complete and stable, with an auxiliary confidence level of 0.9.

[0061] At this point, the system adjusts the weights based on this reliability information. For example, for the deformation source "welding parameter fluctuation," the weight of the comparison result might be set to 0.8, while the weight of the auxiliary information might be set to 0.9. For the deformation source "equipment vibration," the weight of the comparison result might be set to 0.6, while the weight of the auxiliary information might be set to 0.7. Subsequently, the system merges the comparison result and the auxiliary information using a weighted average based on these adjusted weights. For example, if the indication intensity of the comparison result for "welding parameter fluctuation" is X, and the indication intensity of the auxiliary information for "welding parameter fluctuation" is Y, then the final merged indication intensity of "welding parameter fluctuation" is (0.8*X+0.9*Y) / (0.8+0.9). In this way, the system can comprehensively consider the reliability of different information sources and ultimately arrive at a more convincing deformation source. For example, it may ultimately identify "welding parameter fluctuation" as the primary deformation source, while "equipment vibration" is a secondary or uncertain factor.

[0062] In another embodiment of this application, a weld seam 3D image recognition method further includes: S9000: Divide the three-dimensional point cloud into multiple local regions along the weld centerline. Based on the original optical response data and point cloud data, the laser line width, standard deviation of laser intensity distribution, local point cloud density, and local surface fitting residual are combined according to preset weights to obtain the data reliability score of each local region. S10000: When extracting the geometric consistency feature set, the point cloud samples used to extract the geometric consistency feature set are weighted according to the data confidence score to obtain the confidence-weighted geometric consistency feature set. S11000: When generating a deviation heatmap, the deviation of each local area is weighted and mapped according to the data reliability score; S12000: When calculating the deformation score, the deviations of each local area are weighted and summarized based on the data reliability score.

[0063] Specifically, the 3D point cloud is divided into multiple local regions along the weld centerline to allow for refined evaluation of different parts of the weld. The data reliability score for each local region is obtained by comprehensively considering various data quality indicators. These indicators include laser linewidth, the standard deviation of laser intensity distribution, local point cloud density, and local surface fitting residuals. Laser linewidth reflects the accuracy and clarity of the laser scan; the standard deviation of laser intensity distribution indicates the consistency of surface reflectivity or noise level; local point cloud density reflects the completeness and density of data acquisition; and local surface fitting residuals measure the consistency between the point cloud data and the local geometric model, i.e., the smoothness and noise level of the point cloud. These indicators can be weighted and synthesized according to the actual application scenario and data characteristics using preset weights to obtain a quantitative data reliability score. For example, the weights can be obtained based on empirical values, expert knowledge, or by training a machine learning model.

[0064] When extracting geometric consistency features, the point cloud samples used for extraction are weighted according to the data confidence score. This means that point cloud samples in high-confidence regions are given higher importance during feature extraction, while point cloud samples in low-confidence regions are given lower importance. This weighting process can effectively reduce the negative impact of low-quality data on the feature extraction results, making the extracted geometric consistency feature set more representative and accurate.

[0065] When generating a deviation heatmap, weighted mapping of deviations in each local area based on data reliability scores allows deviations in high-reliability areas to be more prominently displayed in the heatmap, while deviations in low-reliability areas are appropriately downplayed. This helps users more intuitively focus on the true deviations reflected in areas with reliable data quality.

[0066] When calculating the deformation score, the deviations of each local area are weighted and summarized based on the data reliability score. This ensures that the final deformation score more accurately reflects the overall deformation of the weld. By assigning higher weights to deviations in high-reliability areas, score distortion caused by local low-quality data can be effectively avoided, thereby improving the reliability of the deformation score.

[0067] This solution introduces a data reliability score to compensate for the basic solution's insufficient consideration of differences in raw data quality. Local regions are divided along the weld centerline and quantified with scores. High-scoring point clouds are given higher weights during feature extraction to suppress interference from low-quality data. Deviation heatmaps and deformation scores are generated using score-weighted mapping and aggregation, making the results closer to actual deformation and reducing misjudgments. The score can be combined with indicators such as laser linewidth, brightness stability, point cloud density, and local fitting residuals, integrating them into the feature extraction, visualization, and decision-making processes to improve recognition reliability.

[0068] As one specific implementation method: Suppose that during the scanning of a weld section, slight vibration of the scanning equipment in a local area results in low density of the 3D point cloud data and large surface fitting residuals in that area. According to the solution in this application, the 3D point cloud is first divided into multiple local regions along the weld centerline, and a data reliability score is calculated for each local region. In the vibrating region, due to the low point cloud density and large fitting residuals, its data reliability score will be calculated as a low value. When extracting the geometric consistency feature set of the weld, point cloud samples from the low-reliability region will be assigned lower weights, while point cloud samples from the high-reliability region will be assigned higher weights, thus ensuring that the extracted feature set is not severely contaminated by low-quality data. Furthermore, when generating the deviation heatmap, the deviation in the low-reliability region will be weighted and mapped, and its display intensity in the heatmap will be correspondingly reduced, avoiding overemphasis on local deviations caused by data quality issues. When calculating the deformation score, the deviation in the low-reliability region will also be assigned a lower weight in the summarization, so that the final deformation score can more accurately reflect the overall deformation of the weld, rather than being dominated by local low-quality data. In this way, even in the presence of local data quality issues, the method of this application can still provide reliable weld deformation assessment results.

[0069] Furthermore, in another embodiment of this application, S7610 may specifically include the following steps.

[0070] S7611: Real-time monitoring of the working status of production equipment and the operating status of auxiliary information sources; S7612: Dynamically calculate the comparison reliability of the comparison results based on the real-time working status and the real-time quality indicators of the raw optical response data; S7613: The auxiliary reliability of auxiliary information is dynamically calculated based on the operating status of the auxiliary information source, the real-time data update frequency, the data integrity, and the degree of first match between the auxiliary information source and the current weld production batch or product model.

[0071] The real-time monitoring of the operating status of production equipment and auxiliary information sources aims to obtain immediate information about the current production environment and data sources. The operating status of production equipment can include uptime, load, fault records, and maintenance history, reflecting the stability of the equipment during weld scanning and data acquisition. The operating status of auxiliary information sources focuses on whether the systems or sensors providing auxiliary information are functioning properly, such as whether the data interface of the MES / ERP system is working smoothly and whether sensors are online.

[0072] Real-time quality metrics for raw optical response data can include signal-to-noise ratio, sharpness, completeness, and consistency. For example, when the production equipment is operating stably and the quality metrics of the raw optical response data are high, the deformation source results obtained through optical feature comparison can be considered to have high reliability; conversely, if the equipment malfunctions or the data quality deteriorates, the reliability of the comparison should decrease accordingly. This dynamic calculation method allows the reliability assessment to reflect the actual situation in a timely manner.

[0073] Furthermore, the operational status of the auxiliary information source ensures the availability of the auxiliary information. The frequency of real-time data updates reflects the timeliness of the auxiliary information; more frequent updates generally mean newer and more reliable information. Data integrity measures whether the auxiliary information is missing key fields or contains outliers. The degree of matching between the auxiliary information and the current weld production batch or product model refers to the correlation between the auxiliary information and the currently identified weld in terms of production batch, product model, etc. The higher the matching degree, the greater the reference value of the auxiliary information for the current weld, and the higher its reliability. By comprehensively considering these factors, the reliability of the auxiliary information can be quantitatively assessed.

[0074] This solution overcomes the inaccuracies of fixed weights under complex operating conditions by dynamically evaluating the "reliability of comparison results" and "reliability of auxiliary information" in real time. The system monitors equipment status and optical data quality, lowering the reliability of comparison results upon the occurrence of jitter / noise; it also evaluates the operation, update frequency, completeness, and matching degree with the current batch of auxiliary information sources, reducing their weight if abnormalities are found. This makes subsequent weight allocation and information fusion more accurate and robust, improving the reliability of deformation source and defect determination.

[0075] In another embodiment of this application, the weld 3D image recognition method further includes: S7614: Obtain information on changes in the production line and calculate the damping coefficient and response time constant for adjusting the comparative reliability and auxiliary reliability. The information on changes includes at least equipment status fluctuation indicators, quality fluctuation indicators of raw optical response data, and switching events related to the current weld production batch or product model. S7615: When the change information is lower than the preset first threshold, the fluctuations of the comparison confidence and the auxiliary confidence are smoothed according to the damping coefficient and the response time constant. S7616: When the change information is higher than the preset first threshold, the calculation frequency of the comparison confidence and auxiliary confidence is increased according to the damping coefficient and response time constant, and then limited within the preset frequency upper limit. S7617: Determine the maintenance strategy for the calculation frequency based on the adjusted comparative confidence and auxiliary confidence.

[0076] Specifically, production line change information can be understood as a comprehensive indicator reflecting the stability of the production process. This includes at least equipment status fluctuation indicators, raw optical response data quality fluctuation indicators, and switching events related to the current weld production batch or product model. Equipment status fluctuation indicators refer to the real-time amplitude or frequency of changes in key parameters (such as speed, temperature, and voltage) of production equipment during operation, such as welding robots and scanning equipment. Raw optical response data quality fluctuation indicators refer to the real-time changes in clarity, completeness, and noise levels of the raw optical response data acquired through scanning. Switching events related to the current weld production batch or product model refer to the situation where the weld being processed on the production line switches from one batch or model to another. This switching is usually accompanied by significant changes in process parameters, material properties, etc.

[0077] The damping coefficient and response time constant are key parameters used for dynamically adjusting reliability calculations and handling fluctuations. The damping coefficient controls the system's response speed and overshoot; for example, a larger damping coefficient makes the system response smoother and reduces oscillations. The response time constant characterizes the time required for the system to reach a new steady state; for example, a smaller response time constant means the system can adapt to changes more quickly. These parameters can be calculated based on historical data, machine learning models, or expert experience.

[0078] When the change in information is below a preset first threshold, it indicates that the production line is in a relatively stable state. At this point, based on the calculated damping coefficient and response time constant, the fluctuations in the comparative and auxiliary reliability are smoothed. Smoothing methods such as moving average, exponential smoothing, or Kalman filtering can be used to eliminate short-term noise and random fluctuations, making the reliability assessment results more stable and reliable.

[0079] When the change in information exceeds a preset first threshold, it indicates that the production line may have undergone significant changes or is in an unstable state, such as equipment failure, material change, or product model switching. In this case, to promptly capture and respond to these changes, the calculation frequency of comparative reliability and auxiliary reliability is increased based on the damping coefficient and response time constant. This means the system will perform reliability assessments more frequently to ensure accurate real-time reliability information is provided even in rapidly changing environments. Simultaneously, to avoid over-calculation and resource waste, the calculation frequency will be limited within a preset upper limit.

[0080] Therefore, based on the adjusted comparison reliability and auxiliary reliability, a strategy for maintaining the computation frequency can be determined. This strategy may include maintaining a low computation frequency in a stable state to save resources, increasing the computation frequency during periods of drastic change to ensure real-time performance, and gradually reducing the computation frequency after the changes stabilize to achieve a dynamic balance.

[0081] This solution introduces a production line change perception and response mechanism: when equipment status, raw optical data quality, or batch / model switching changes, the system calculates the damping coefficient and response time constant accordingly, serving as a "regulator" for reliability assessment. For small changes, it smooths reliability fluctuations to filter out random noise; for larger changes, it increases the calculation frequency and sets an upper limit to quickly track changes while avoiding excessive resource consumption. This achieves robust adaptation to dynamic operating conditions and improves the accuracy and robustness of defect detection.

[0082] In some preferred embodiments, the following specific example illustrates the situation: Suppose that automated welding of the car body frame is underway on a welding production line in a car manufacturing plant. This production line periodically switches between different car body models, and the welding equipment (such as welding robots) may experience slight performance drift after prolonged operation. At the same time, the laser scanner may be affected by ambient light interference at certain times, leading to a decrease in the quality of the raw optical response data.

[0083] According to the scheme of this application, the system continuously acquires information on changes in the production line. For example, when the production line switches from "car model A" to "SUV model B", the system will identify the change information as "a switching event related to the current weld production batch or product model". Simultaneously, the system monitors the operating parameters of the welding robot. If slight fluctuations are detected in its vibration amplitude or positioning accuracy, they will be recorded as "equipment status fluctuation indicators". If the raw optical response data acquired by the laser scanner shows local blurring or a decrease in signal-to-noise ratio, it will be recorded as "raw optical response data quality fluctuation indicators".

[0084] Based on these changes, the system will calculate in real time the damping coefficient and response time constant used to adjust the comparative confidence level and the auxiliary confidence level.

[0085] Specifically: 1. In a steady state (changes below a preset first threshold): Assuming the production line stably produces "Car Model A" for a period of time, the equipment operates smoothly, and the data quality is good, the changes are below the preset first threshold. The system will smooth the real-time calculated comparison and auxiliary reliability based on the calculated damping coefficient and response time constant. For example, using an exponentially weighted moving average method makes the reliability curve smoother, filtering out unavoidable minor fluctuations in daily production, and ensuring that the weld defect judgment logic has higher accuracy and a lower false alarm rate in a steady state.

[0086] 2. When a switching event occurs (change information exceeds a preset first threshold): When the production line switches from "Sedan Model A" to "SUV Model B," this is a significant change event, causing the change information to exceed a preset first threshold. At this time, the system will immediately increase the calculation frequency of the comparison reliability and auxiliary reliability based on the calculated damping coefficient and response time constant. For example, it will increase the calculation frequency from once per second to five times per second to ensure that when welding of the new product model begins, the reliability of new weld features and auxiliary information can be quickly and accurately assessed, thereby adjusting the weld defect judgment logic in a timely manner and avoiding potential defect identification problems caused by model switching. Simultaneously, the calculation frequency will be limited to a preset upper limit, such as ten times per second, to avoid system overload.

[0087] 3. When equipment performance drifts or data quality deteriorates (changes exceed a preset first threshold): If the welding robot experiences continuous performance drift, or ambient light interference causes a continuous decline in the quality of the original optical response data, these changes will also exceed a preset first threshold. The system will also increase the reliability calculation frequency and determine a strategy to maintain the calculation frequency based on the adjusted comparative reliability and auxiliary reliability. For example, during periods of performance drift or data quality deterioration, a higher calculation frequency is maintained; once the problem is resolved or the production line stabilizes, the calculation frequency is gradually reduced to a normal level.

[0088] In another embodiment of this application, a sub-step S7614 is further proposed: calculating the damping coefficient and response time constant used to adjust the contrast confidence level and the auxiliary confidence level, including: A1: Perform multi-scale time-frequency analysis on the changing information to identify the nonlinear characteristics, abrupt change points, and multi-peak distribution of the changing information; A2: Adjust the weights and structure of the preset adjustment parameter change model based on nonlinear characteristics, abrupt change points, and multi-peak distribution; A3: Calculate the damping coefficient and response time constant based on the adjusted adjustment parameter change model; A4: Based on the preset parameter change trend prediction rules, the damping coefficient and response time constant are corrected.

[0089] Specifically, multi-scale time-frequency analysis of changing information refers to using techniques such as wavelet transform, empirical mode decomposition (EMD), or variational mode decomposition (VMD) to decompose the original changing information into different time and frequency scales. This decomposition effectively reveals the energy distribution and temporal evolution of the changing information across different frequency components, thereby identifying its inherent nonlinear characteristics, such as abrupt changes caused by sudden equipment failures or process parameter switching, and multi-peak distributions formed by the superposition of multiple factors. The aim is to comprehensively and deeply understand the complex dynamic characteristics of changing information.

[0090] The process of adjusting the weights and structure of a pre-defined model based on nonlinear characteristics, abrupt changes, and multi-peak distributions can be understood as using these characteristics to optimize a pre-defined mathematical model after identifying the complex features of the changing information. This model aims to describe how the damping coefficient and response time constant change with the changing information. For example, if significant nonlinear characteristics are identified, the model structure may be adjusted to include nonlinear terms; if abrupt changes exist, the model weights may be adjusted to give higher priority to recent data, or a piecewise linear / nonlinear model may be introduced. The goal is to enable the model to more accurately reflect the actual physical process and improve the accuracy of parameter calculations.

[0091] In practical applications, the damping coefficient and response time constant are calculated based on the adjusted parameter variation model. Specifically, an optimized model is used, taking real-time change information as input, and the model's calculation logic directly outputs the corresponding damping coefficient and response time constant. For example, the model can be built based on machine learning algorithms (such as neural networks or support vector machines) or adaptive control theory, and its parameters (weights and structure) have been calibrated according to the complex characteristics of the changing information. The aim is to provide a data-driven and adaptive parameter calculation mechanism.

[0092] Furthermore, adjusting the damping coefficient and response time constant based on preset parameter change trend prediction rules involves, after initially calculating the damping coefficient and response time constant, combining historical data and domain knowledge to apply a set of preset rules to predict the future change trends of these parameters. For example, if it is predicted that the fluctuation index of a certain equipment status will continue to deteriorate, the damping coefficient can be fine-tuned in advance to make it more robust in subsequent confidence level adjustments. The purpose is to further improve the predictability and stability of the parameters and avoid over-adjustment caused by short-term fluctuations.

[0093] This application's solution, by introducing multi-scale time-frequency analysis, can deeply uncover the hidden nonlinear characteristics, abrupt change points, and multi-peak distributions within the production line's changing information, thus overcoming the computational biases that traditional methods may suffer from due to their inability to fully capture these complex dynamics. This more refined understanding of the changing information allows for the adaptive adjustment of the weights and structure of the pre-defined adjustment parameter change model based on these identified complex characteristics. This model adjustment mechanism ensures that the computational model used maintains a high degree of consistency with the actual operating state of the production line in real time, avoiding errors caused by model rigidity. Based on this, the accuracy and adaptability of the damping coefficient and response time constant calculated from the adjusted model are significantly improved. Finally, by combining this with pre-defined parameter change trend prediction rules for correction, the predictability and stability of these parameters are further enhanced, enabling them not only to respond to current changes but also to predict future trends, thus providing a more solid and reliable foundation for the dynamic adjustment of comparative and auxiliary reliability.

[0094] In some preferred embodiments, the following specific example illustrates the situation: Suppose that during the weld production process, it is necessary to calculate the damping coefficient and response time constant used to adjust the contrast confidence level and auxiliary confidence level. First, the system acquires information about changes in the production line, such as data from equipment vibration sensors, noise level fluctuations from laser scanners, and material batch switching events recorded in the MES system. To analyze this information in depth, wavelet transform can be used to perform multi-scale time-frequency analysis on this data. Through wavelet decomposition, the original signal can be decomposed into detail components and approximate components of different frequency sub-bands, thereby clearly identifying abrupt changes in high-frequency noise (such as instantaneous equipment failures), nonlinear patterns in mid-frequency fluctuations (such as slow drift of process parameters), and multi-peak distributions in low-frequency trends (such as inherent characteristic differences caused by different production batches).

[0095] Once these nonlinear characteristics, abrupt changes, and multi-peak distributions are identified, the system adjusts a pre-defined adaptive fuzzy logic model based on this information. For example, if a significant abrupt change is detected, the weights of the fuzzy rules are dynamically adjusted to make the model more sensitive to recent data changes; if a specific multi-peak distribution is identified, the shape of the membership function of the fuzzy set is optimized to better match these distributions. Based on this adjusted adaptive fuzzy logic model, real-time change information is input into the model to calculate the damping coefficient and response time constant at the current moment.

[0096] Finally, to further improve parameter stability, the system incorporates preset parameter change trend prediction rules for correction. For example, based on historical data analysis, if a device's condition fluctuation index shows an upward trend in three consecutive batches, even if the currently calculated damping coefficient is already high, the prediction rule will slightly adjust it upwards to proactively address potentially worsening device conditions, ensuring smoother and more reliable subsequent reliability adjustments. In this way, the obtained damping coefficient and response time constant not only accurately reflect the current state but also possess a certain predictive capability, making the adjustment of weld defect judgment logic more intelligent and efficient.

[0097] In another embodiment of this application, a sub-step of A1 is further proposed: performing multi-scale time-frequency analysis on the change information, including: A11: Identify the production parameter characteristics of the current weld production batch or product model; A12: Obtain the switching information of weld production batch or product model in the change information, load the historical production parameter characteristics corresponding to the current weld production batch or product model, and preset the parameters for multi-scale time-frequency analysis based on the historical production parameter characteristics; A13: Based on the characteristics of production parameters, dynamically select the basis function type and decomposition level of the multi-scale time-frequency analysis, and adjust the frequency resolution of each scale of the multi-scale time-frequency analysis. A14: Determine the second degree of matching between the production characteristic parameters and the preset parameter characteristics. When the second degree of matching is lower than the preset second threshold, trigger parameter readjustment of multi-scale time-frequency analysis.

[0098] Specifically, before conducting multi-scale time-frequency analysis of the variation information, it is first necessary to identify the production parameter characteristics of the current weld production batch or product model. These production parameter characteristics may include, but are not limited to, process parameters directly related to weld formation and quality, such as welding speed, current, voltage, wire feed speed, shielding gas type and flow rate, preheating temperature, and cooling rate. The combination of these parameters defines the production conditions for a specific batch or model and influences the specific manifestation of the variation information.

[0099] Furthermore, the system acquires switching information for weld production batches or product models from the change information. This switching information can originate from records in the Production Management System (MES) or Enterprise Resource Planning (ERP) system, indicating the weld batch or product model currently in production. Once a switch is detected, the system loads the historical production parameter characteristics corresponding to the current weld production batch or product model. These historical production parameter characteristics are pre-stored production data related to a specific batch or model, along with corresponding change information analysis experience. Based on these historical production parameter characteristics, parameters for multi-scale time-frequency analysis can be preset, such as the initial basis function type, number of decomposition levels, and frequency resolution range.

[0100] Based on the characteristics of production parameters, the basis function type and decomposition level of the multi-scale time-frequency analysis are dynamically selected, and the frequency resolution of each scale is adjusted. Basis function types can include wavelet basis functions (such as Haar wavelets, Daubechies wavelets, Morlet wavelets, etc.), S-transform basis functions, and Hilbert-Huang Transform (HHT) empirical mode decomposition (EMD). Different basis functions have different sensitivities to different types of signal characteristics. For example, wavelet basis functions may be more suitable for changes with sharp abrupt changes or local features; HHT may be more advantageous for non-stationary or nonlinear signals. The number of decomposition levels determines the level of analysis refinement; more levels result in a stronger ability to capture high-frequency details. Adjusting the frequency resolution ensures that key frequency components of the changing information can be effectively captured at different scales. The dynamic selection and adjustment of these parameters aim to enable multi-scale time-frequency analysis to better adapt to the specific changing information characteristics of the current production batch or product model.

[0101] In addition, the system will determine a second degree of matching between the production characteristic parameters and the preset parameter characteristics. This second degree of matching can be obtained by calculating the similarity or distance between the current production parameter characteristics and historical preset parameter characteristics, such as using Euclidean distance, cosine similarity, or a matching score based on a machine learning model. The preset parameter characteristics are typical production parameter models built based on a large amount of historical data and expert experience. When the second degree of matching is lower than a preset second threshold, it indicates a significant difference between the current production conditions and historical experience, triggering parameter readjustment for multi-scale time-frequency analysis. Parameter readjustment can be achieved by re-optimizing parameters such as basis function type, number of decomposition levels, and frequency resolution using optimization algorithms (such as genetic algorithms or particle swarm optimization algorithms) or rule-based expert systems to ensure the accuracy and robustness of the analysis.

[0102] This solution identifies the production parameter characteristics of the current batch / model and loads historical characteristics based on switching information as the initial settings for multi-scale time-frequency analysis. Subsequently, it dynamically selects the basis function and decomposition level based on real-time characteristics, and fine-tunes the frequency resolution at each scale to more accurately capture nonlinear features, abrupt changes, and multi-peak distributions. When the real-time characteristics do not match the preset characteristics well, parameter readjustment is triggered. This achieves adaptive analysis parameters, improves the reliability of analysis under complex operating conditions, and provides a robust basis for calculating damping coefficients and response time constants.

[0103] As a specific implementation method, the following example illustrates the process: Suppose that on a certain welding production line, quality inspection needs to be performed on the welds of different car body models. Different car body models may use different batches of steel, welding process parameters (such as welding current and speed), and different geometric structures. These factors will cause the production line variation information (such as equipment status fluctuation indicators and quality fluctuation indicators of raw optical response data) to exhibit different time-frequency characteristics.

[0104] When implementing the scheme of this application: First, when the production line switches from "model A" to "model B", the system will identify that the current weld production batch or product model is "model B" and identify its corresponding production parameter characteristics. For example, model B may use high-strength steel, and the welding current and speed are higher than those of model A.

[0105] Next, the system acquires this switching information and loads the historical production parameter characteristics of "Model B" from the database. Based on these historical characteristics, the system presets the initial parameters for multi-scale time-frequency analysis. For example, if historical data shows that the changes in "Model B" contain a lot of high-frequency noise and obvious transient impacts, the system may preset the use of the Daubechies wavelet basis function and set a higher decomposition level.

[0106] Then, based on the real-time production parameter characteristics of "Model B", the system dynamically selects the basis function type and decomposition level for multi-scale time-frequency analysis, and adjusts the frequency resolution at each scale. For example, if real-time monitoring reveals that high-frequency vibration components are more prominent, the system may further optimize the selection of wavelet basis functions, increase the number of high-frequency decomposition levels, and improve the frequency resolution in the high-frequency band to capture these vibration features more precisely.

[0107] Finally, the system continuously assesses the second degree of matching between the production characteristic parameters of "Model B" and the preset parameter characteristics. If, during the production process, due to some reason (e.g., a temporary change in material batches by the material supplier, resulting in slight deviations in material properties), the matching degree between the current production parameter characteristics and the preset parameter characteristics of "Model B" falls below the preset second threshold, the system will immediately trigger parameter readjustment for multi-scale time-frequency analysis. At this time, the system may re-evaluate and optimize the basis functions, the number of decomposition levels, and the frequency resolution to ensure that even when production conditions undergo small but significant changes, multi-scale time-frequency analysis can still accurately identify nonlinear characteristics, abrupt changes, and multi-peak distributions in the changing information, thereby providing the most accurate input for the calculation of the damping coefficient and response time constant.

[0108] Reference Figure 2 The specific embodiments of this application also disclose a 3D image recognition system for welds, including: Feature extraction module 1 is used to scan the weld to obtain a three-dimensional point cloud and extract the geometric consistency feature set of the weld from the three-dimensional point cloud. Benchmark establishment module 2 is used to establish a standard geometric consistency feature set for an ideal weld model; Deformation assessment module 3 is used to compare the geometric consistency feature set with the standard geometric consistency feature set to generate a deviation heat map and deformation score for the weld. The judgment logic adjustment module 4 is used to adjust the preset weld defect judgment logic based on the deviation heat map and deformation score.

[0109] This 3D weld image recognition system aims to solve the problems of false alarms and missed detections caused by scanning system errors and fixed defect judgment logic in existing technologies. Through the collaborative work of its various internal modules, it achieves accurate extraction and comparison of weld geometric consistency features, generates intuitive deviation heatmaps and quantified deformation scores, and dynamically adjusts the weld defect judgment logic accordingly, thereby effectively improving the accuracy and reliability of weld inspection. The system integrates data acquisition, feature analysis, deviation assessment, and intelligent decision-making to form a closed-loop quality control mechanism, capable of adapting to complex and changing production environments and reducing unnecessary rework and product scrap.

[0110] The above are merely embodiments of this application and are not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for 3D image recognition of weld seams, characterized in that, include: The weld is scanned to obtain a three-dimensional point cloud, and the geometric consistency feature set of the weld is extracted from the three-dimensional point cloud. Establish a standard geometric consistency feature set for an ideal weld model; By comparing the geometric consistency feature set with the standard geometric consistency feature set, a deviation heatmap and deformation score of the weld are generated. The preset weld defect judgment logic is adjusted based on the deviation thermogram and the deformation score. The method further includes: Scan and acquire raw optical response data of the weld area; Microscopic optical features of the weld surface were extracted from the raw optical response data; The microscopic optical features of the weld surface are compared with a preset optical fingerprint database of deformation sources to identify the source of deformation. Adjust the weld defect determination logic according to the deformation source; The step of comparing the microscopic optical features of the weld surface with a preset optical fingerprint database of deformation sources to identify the source of deformation includes: The microscopic optical features of the weld surface are subjected to multi-dimensional feature enhancement processing to obtain enhanced features; Obtain the similarity between the enhanced features and fingerprints from different deformation sources in the deformation source optical fingerprint database; Determine whether all the similarities are higher than a preset judgment threshold and whether the differences between the similarities are less than a preset distinction threshold; When the similarity is higher than the preset judgment threshold and the difference between the similarities is less than the preset distinction threshold, auxiliary information of the weld area is collected. The auxiliary information includes equipment status and calibration / trajectory deviation, material batch and welding process parameters, thermal process data and MES / ERP records. The weld geometry and material combination are obtained, and the enhancement features are weighted according to a preset feature sensitivity weight table corresponding to the weld geometry and material combination to obtain weighted enhancement features. Based on the comparison results between the weighted enhanced features and the deformation source optical fingerprint database, and in conjunction with the auxiliary information, the deformation source is obtained.

2. The weld 3D image recognition method according to claim 1, characterized in that, The deformation source is obtained by comparing the weighted enhanced features with the deformation source optical fingerprint database and combining the auxiliary information, including: Assess the reliability of the comparison results and the reliability of the auxiliary information. Based on the comparison confidence level and the auxiliary confidence level, adjust the weights of the comparison results and the auxiliary information. Based on the adjusted weights of the comparison results and the weights of the auxiliary information, the comparison results and the auxiliary information are fused to obtain the source of deformation.

3. The weld 3D image recognition method according to claim 1, characterized in that, Also includes: The three-dimensional point cloud is divided into multiple local regions along the centerline of the weld. Based on the original optical response data and point cloud data, the laser line width, the standard deviation of the laser intensity distribution, the local point cloud density, and the local surface fitting residual are combined according to preset weights to obtain the data reliability score of each local region. When extracting the geometric consistency feature set, the point cloud samples used to extract the geometric consistency feature set are weighted according to the data trust score to obtain a trust-weighted geometric consistency feature set. When generating the deviation heatmap, the deviation of each local area is weighted and mapped according to the data reliability score; When calculating the deformation score, the deviations of each local area are weighted and summarized based on the data reliability score.

4. The weld 3D image recognition method according to claim 2, characterized in that, The evaluation of the comparison reliability of the comparison results and the evaluation of the auxiliary information's auxiliary reliability include: Real-time monitoring of the working status of production equipment and the operating status of auxiliary information sources; The comparison reliability of the comparison results is dynamically calculated based on the real-time working status and the real-time quality index of the original optical response data. The reliability of the auxiliary information is dynamically calculated based on the operating status of the auxiliary information source, the real-time data update frequency, the data integrity, and the degree of matching between the auxiliary information and the current weld production batch or product model.

5. The weld 3D image recognition method according to claim 4, characterized in that, The method further includes: Obtain production line change information, calculate the damping coefficient and response time constant for adjusting the comparison reliability and the auxiliary reliability, and the change information includes at least equipment status fluctuation index, quality fluctuation index of raw optical response data, and switching events related to the current weld production batch or product model. When the change information is lower than a preset first threshold, the fluctuations of the comparison confidence and the auxiliary confidence are smoothed according to the damping coefficient and the response time constant. When the change information is higher than the preset first threshold, the calculation frequency of the comparison confidence and the auxiliary confidence is increased according to the damping coefficient and the response time constant, and then limited within the preset frequency upper limit. Based on the adjusted comparison confidence level and the auxiliary confidence level, a maintenance strategy for the calculation frequency is determined.

6. The weld 3D image recognition method according to claim 5, characterized in that, The calculations used to adjust the damping coefficients and response time constants of the comparative confidence level and the auxiliary confidence level include: Multi-scale time-frequency analysis is performed on the change information to identify its nonlinear characteristics, abrupt change points, and multi-peak distribution. Based on the nonlinear characteristics, abrupt change points, and multi-peak distribution, adjust the weights and structure of the preset adjustment parameter change model; Based on the adjusted adjustment parameter change model, the damping coefficient and the response time constant are calculated; The damping coefficient and the response time constant are corrected according to the preset parameter change trend prediction rules.

7. The weld 3D image recognition method according to claim 6, characterized in that, The multi-scale time-frequency analysis of the change information includes: Identify the production parameter characteristics of the current weld batch or product model; Obtain the switching information of weld production batch or product model in the change information, load the historical production parameter characteristics corresponding to the current weld production batch or product model, and preset the parameters of the multi-scale time-frequency analysis according to the historical production parameter characteristics; Based on the characteristics of the production parameters, the basis function type and decomposition level of the multi-scale time-frequency analysis are dynamically selected, and the frequency resolution of each scale of the multi-scale time-frequency analysis is adjusted. The second degree of matching between the production characteristic parameters and the preset parameter characteristics is determined. When the second degree of matching is lower than the preset second threshold, the parameter readjustment of multi-scale time-frequency analysis is triggered.

8. A 3D image recognition system for weld seams, characterized in that, include: The feature extraction module is used to scan the weld to obtain a three-dimensional point cloud, and extract the geometric consistency feature set of the weld from the three-dimensional point cloud. The benchmark establishment module is used to establish a standard geometric consistency feature set for an ideal weld model; The deformation assessment module is used to compare the geometric consistency feature set with the standard geometric consistency feature set to generate a deviation heat map and deformation score for the weld. The judgment logic adjustment module is used to adjust the preset weld defect judgment logic according to the deviation heat map and the deformation score; It is also used to scan and acquire raw optical response data of the weld area; Microscopic optical features of the weld surface were extracted from the raw optical response data; The microscopic optical features of the weld surface are compared with a preset optical fingerprint database of deformation sources to identify the source of deformation. Adjust the weld defect determination logic according to the deformation source; It is also used to perform multi-dimensional feature enhancement processing on the microscopic optical features of the weld surface to obtain enhanced features; Obtain the similarity between the enhanced features and fingerprints from different deformation sources in the deformation source optical fingerprint database; Determine whether all the similarities are higher than a preset judgment threshold and whether the differences between the similarities are less than a preset distinction threshold; When the similarity is higher than the preset judgment threshold and the difference between the similarities is less than the preset distinction threshold, auxiliary information of the weld area is collected. The auxiliary information includes equipment status and calibration / trajectory deviation, material batch and welding process parameters, thermal process data and MES / ERP records. The weld geometry and material combination are obtained, and the enhancement features are weighted according to a preset feature sensitivity weight table corresponding to the weld geometry and material combination to obtain weighted enhancement features. Based on the comparison results between the weighted enhanced features and the deformation source optical fingerprint database, and in conjunction with the auxiliary information, the deformation source is obtained.

Images