Welding quality online detection method based on machine vision

Abstract

The invention discloses a welding quality on-line detection method based on machine vision, and particularly relates to the field of welding quality detection.The welding quality on-line detection method comprises the steps that multi-dimensional basic data collection is conducted, the two-dimensional form, heat distribution, micro texture, three-dimensional topological data and welding process parameters of a welding seam are obtained, and a full-dimensional original data set with time-space synchronization is constructed; performing feature decoupling on the data, respectively constructing geometric morphology, thermal, texture and three-dimensional topological feature functions, and converting unstructured data into structured feature values; integrating multi-dimensional features through a nonlinear deep fusion model, introducing thermal probability mapping, texture attention weighting and a three-dimensional residual compensation mechanism, and outputting a weld comprehensive quality value; and finally, positioning dominant defect dimensions and identifying defect types in combination with a dynamic threshold model to finish quality grading judgment. Multi-dimensional accurate detection and dynamic grading of the welding quality are achieved, different scene requirements are met, and the detection efficiency and the judgment accuracy are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of welding quality detection, more particularly, the present application relates to a welding quality online detection method based on machine vision. BACKGROUND

[0002] In the field of industrial manufacturing, welding as a key connection process, its quality directly affects the safety and service life of equipment structure, and welding quality online detection becomes the core link to ensure production efficiency and product reliability; with the development of intelligent manufacturing technology, machine vision has gradually become the mainstream technology direction of welding quality online detection due to its non-contact, high precision and real-time advantages.

[0003] Current mainstream solutions mostly take single-dimensional data acquisition as the core, such as acquiring two-dimensional morphology images of welds through visible light cameras, or collecting local temperature data with infrared devices, and a few solutions try to integrate the two types of data but do not form full-dimensional coverage; the feature extraction link mainly focuses on the numerical statistics of weld surface geometric parameters such as width and excess height, and multi-dimensional feature fusion generally adopts a simple linear weighting method to achieve simple superposition, and defect judgment relies on pre-set fixed threshold standards to complete quality grading, which meets the requirements of some scenes;

[0004] However, in actual use, it still has some disadvantages, such as:

[0005] 1. The existing technology mostly adopts single-dimensional data acquisition mode, only acquiring two-dimensional images or local thermal data of the weld surface, which cannot completely cover the full-dimensional information of geometric morphology, thermal characteristics, microscopic texture and three-dimensional topology, and is easy to miss internal hidden defects such as micro-cracks and stress concentration due to the one-sidedness of data acquisition, resulting in the lack of comprehensiveness and integrity of the detection results;

[0006] 2. The feature extraction link does not construct special quantitative functions for different dimensional defect attributes, only characterizes through simple parameter statistics, and it is difficult to accurately distinguish the essential differences of different types of defects such as geometric distortion, uneven heat distribution and texture abnormalities, and type misjudgment is easy to occur in the defect recognition process, and the overall recognition accuracy is insufficient;

[0007] 3. Multi-dimensional feature fusion generally adopts a simple linear weighting method, which does not fully consider the non-linear correlation between different dimensional features and the contribution difference of different defect types, cannot effectively couple the synergistic effect of multi-dimensional defects, and leads to a large deviation between the comprehensive quality evaluation results and the actual weld quality state, and the evaluation reliability is insufficient;

[0008] 4. Defect judgment relies on fixed and unchanged threshold standards, and does not dynamically adjust according to the actual application requirements such as weld type and workpiece service scene, which is insufficient for controlling critical defects in high safety requirement scenes, and over-intervenes in slight defects in ordinary scenes, which has poor judgment adaptability and easily affects production efficiency. SUMMARY

[0009] In order to overcome the above-mentioned defects of the prior art, the present application provides a welding quality online detection method based on machine vision, which solves the problems raised in the above background art through the following scheme.

[0010] In order to achieve the above-mentioned purpose, the present application provides the following technical scheme: a welding quality online detection method based on machine vision, comprising:

[0011] S1: Multi-dimensional basic data acquisition: through the deployment of multi-modal acquisition equipment, full-dimensional raw data acquisition is performed on the weld area, and time protocol hardware synchronization system is used to realize the timing and spatial synchronization of each device, thereby constructing a weld full-dimensional raw data set containing geometry, heat, texture, three-dimensional topology and process parameters;

[0012] S2: Multi-dimensional feature quantization extraction and function construction: the full-dimensional raw data collected in S1 is subjected to feature decoupling and quantization calculation, and for the four dimensions of weld geometry, thermal characteristics, microscopic texture and three-dimensional topology, a geometry feature function, a thermal feature function, a texture feature function and a three-dimensional topology feature function are constructed respectively, so as to convert the unstructured raw data into structured feature values that can represent the defect degree;

[0013] S3: Deep fusion function modeling and calculation: based on the constructed geometry feature function, thermal feature function, texture feature function and three-dimensional topology feature function, a nonlinear deep fusion model is constructed, and a weld comprehensive quality value that can comprehensively reflect the overall quality level of the weld is outputted;

[0014] S4: Multi-dimensional defect tracing and dynamic grading determination: based on the weld comprehensive quality value outputted in S3 and the dimensional feature functions in S2, the abnormal coefficient of each dimensional feature function is calculated to locate the dominant defect dimension and identify the specific defect type, and a dynamic threshold model is constructed to complete the determination of the weld quality grade.

[0015] Technical effects and advantages of the present application:

[0016] 1. The multi-modal acquisition means is used to acquire full-dimensional data of the weld two-dimensional morphology, heat distribution, microscopic texture, three-dimensional point cloud and process parameters, and the time sequence and spatial synchronization mechanism is used to realize accurate matching of each dimensional data, so as to completely cover the quality characteristics of each dimension of the weld, effectively avoid defect omission caused by one-sided data, and ensure the comprehensiveness of the detection;

[0017] 2. Constructing exclusive quantification feature functions for the four core dimensions of geometry, heat, texture, and three-dimensional topology, extracting features through exclusive methods such as contour analysis, heat field statistics, gray level co-occurrence matrix calculation, and point cloud topology analysis, which can accurately distinguish the essential differences of different types of defects and significantly improve the accuracy of defect recognition;

[0018] 3. Building a nonlinear deep fusion model, introducing thermal probability mapping, texture attention weighting, and three-dimensional residual compensation mechanism, fully exploiting the nonlinear correlation and defect contribution difference of each dimension feature, realizing deep coupling and accurate fusion of multi-dimensional features, making the comprehensive quality evaluation result more consistent with the actual quality state of the weld, and improving the evaluation reliability;

[0019] 4. Establishing a dynamic threshold determination system, dynamically adjusting the determination threshold according to the weld type, workpiece service scene, and defect type, and clearly defining the determination boundary of different grade defects, which can not only ensure the strict control of fatal defects, but also avoid excessive intervention of slight defects, effectively balancing the detection accuracy and production efficiency, and adapting to diversified industrial application scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a schematic diagram of the overall structure of the present application.

[0021] Figure 2 is a schematic diagram of the S1-S3 process of the present application.

[0022] Figure 3 is a schematic diagram of the S4 process of the present application.

[0023] Figure 4 is a schematic diagram of the relationship between the geometric feature function value and the comprehensive quality value of the present application. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0025] Reference Figures 1-4 The welding quality online detection method based on machine vision shown in the present application comprises:

[0026] S1: Multi-dimensional basic data acquisition: through the deployed multi-modal acquisition equipment, the full-dimensional original data of the weld area is acquired, and the time sequence and spatial synchronization of each device are realized through the time protocol hardware synchronization system, and the full-dimensional original data set of the weld containing geometry, heat, texture, three-dimensional topology and process parameters is constructed;

[0027] This step is the data source basis of the whole detection process, through the deployment of multi-modal visual acquisition equipment and the split sub-steps to complete the directional acquisition, to obtain the geometric, thermal, texture and three-dimensional topology original data of the welding area, all the collected data are bound to the same welding seam detection area, that is, to limit its spatial coordinate range: 、 , to provide standardized input for subsequent feature extraction and fusion analysis, the timing and spatial consistency of the collection is guaranteed by an independent synchronization collection mechanism, the specific steps are as follows:

[0028] S101: Geometric shape basic data collection: based on the deployed 200 million pixel polarized visible light camera, matched with the ring polarized light source, the welding seam area is continuously imaged and collected for 5 frames, the collection area covers 50mm on both sides of the welding seam center, including the welding seam contour, the undercut area, the excess height area and the surrounding base material, to ensure complete capture of the welding seam contour and the surrounding base material area; the welding seam two-dimensional shape image sequence is collected, including the pixel gray scale distribution and the welding seam edge pixel coordinates of each frame, denoted as , wherein 、 is the pixel coordinate, is the frame number, the pixel value range is 0-255 (8-bit gray depth), and the image format is RAW to retain the original photosensitive data;

[0029] S102: Thermal feature basic data collection: based on the deployed uncooled infrared thermal imaging camera, the sampling frequency is matched with the visible light camera, and the welding seam area is continuously imaged and collected for 5 frames of thermal distribution at a sampling frequency of 30 frames / second, covering the welding pool and the heat affected zone, the welding seam multi-frame thermal distribution image, the real-time temperature value of each pixel point and the inter-frame temperature change data are collected, and the welding seam thermal distribution image and the temperature field data matrix obtained are denoted as , wherein 、 is the infrared pixel coordinate, the matrix value is the real-time temperature value of the corresponding pixel point, the unit is ℃, the data output format is 16-bit TIFF, the time stamp, welding current and voltage parameters of each frame collected are recorded synchronously, which are used for subsequent correlation analysis of thermal features and welding process parameters;

[0030] S103: Texture feature basic data collection: based on the deployed high-resolution line scan camera, matched with the linear LED light source, the scanning direction is perpendicular to the extension direction of the welding seam; the welding seam surface is scanned and collected along the direction perpendicular to the extension direction of the welding seam, the step length is 0.01mm, covering all the micro areas on the welding seam surface, the welding seam surface full-area gray texture matrix, i.e. the welding seam micro texture matrix, is collected, denoted as , including pixel gray gradient and neighborhood gray difference, wherein Line pixel column number, corresponding to 100mm scan width, Scan line number, N is the weld length, matrix value is the texture gray gradient (range 0-4095), data is transmitted in real time to the edge computing module in the form of binary stream, ensuring the integrity of the capture of micro texture features.

[0031] S104: Three-dimensional topological basic data acquisition: based on the deployed laser triangulation point cloud scanner, scanning frequency 10kHz, point cloud density 100 points / mm², measurement range 0-50mm, Z-axis accuracy ±0.005mm, laser line width ≤0.1mm, full-area three-dimensional surface scanning acquisition is carried out on the weld area, covering the weld surface three-dimensional topography, concave or convex area and normal vector distribution, three-dimensional coordinate point cloud data (including three-dimensional coordinates of each point, spatial relationship of neighborhood points) of the weld are collected, surface normal vector matrix (including normal vector angle of each point) is collected, wherein the three-dimensional point cloud data is denoted as Weld plane physical coordinates, unit mm; Height value, unit mm, covering all spatial positions of the weld, surface normal vector matrix is denoted as Physical coordinates, Angle between normal vector and X-axis, Angle between normal vector and Y-axis, point cloud data format is PLY, including color information matched with

[0032] S105: Synchronous acquisition deployment: based on the deployed PTP precise time protocol hardware synchronization system, the acquisition devices of S101 to S104 are deployed in time sequence and space synchronization, and the synchronized data set supporting subsequent analysis is collected:

[0033] Time sequence synchronization: through the output of trigger signal by industrial-grade synchronization controller, the trigger delay of each device acquisition is controlled to be ≤0.2ms, and the multi-dimensional synchronized original data of the same detection time is collected, including Frame or scan line corresponding relationship;

[0034] Space synchronization: chessboard calibration board is used to complete the external parameter calibration of each device, and the pixel and physical coordinate mapping parameter set is collected, including mapping matrix of and , mapping coefficient of and physical coordinates, ensuring that the space registration error is ≤0.05mm;

[0035] ​​​​​​Welding process parameter acquisition: Core process parameters of the welding process are synchronously acquired via the industrial bus of the welding equipment, including: welding current. Welding voltage Welding speed (Ratio of weld length to welding time), shielding gas flow rate ; Form a set of welding process parameters ;

[0036] Data binding: Assign a unique identifier (ID) to each set of synchronously collected data, and associate it with the weld segment number, collection timestamp, and welding process parameter set to obtain a complete and traceable original dataset. ;

[0037] S2: Multi-dimensional feature quantization extraction and construction: The feature decoupling and quantization calculation are performed on the full-dimensional raw data collected by S1. For the four dimensions of weld geometry, thermal properties, micro texture and three-dimensional topology, geometric feature function, thermal feature function, texture feature function and three-dimensional topology feature function are constructed respectively, and the unstructured raw data is transformed into structured feature values ​​that can characterize the degree of defects.

[0038] It should be further explained that the full amount of raw data collected by S1 specifically includes: Based on image analysis algorithms, feature decoupling, quantization calculation, and function construction are performed on the full set of raw data collected by S1. This transforms the unstructured raw data into structured feature values ​​that characterize the degree of weld defects. Each feature function corresponds to a single-dimensional defect attribute, providing independent and accurate variable support for subsequent deep fusion modeling. The specific steps are as follows:

[0039] S201: Constructing geometric morphological characteristic functions:

[0040] Data collected based on S101 The two-dimensional morphological image sequence of the weld is constructed by contour extraction, curvature entropy calculation, morphological analysis and multi-parameter coupling to construct a composite feature function that reflects the geometric morphological defects of the weld. The core geometric indicators such as contour fluctuation, undercut ratio and uniformity of reinforcement height are quantified. All benchmark values ​​are standard thresholds calibrated based on 10,000 combination weld samples.

[0041] It should be further explained that multi-dimensional feature quantization extraction includes:

[0042] Contour extraction and curvature calculation: for Inter-frame registration was performed on each frame of the image to eliminate robotic arm jitter error. An improved Canny operator was used, with a low threshold of 25 and a high threshold of 85, to extract the weld contour of each frame, denoted as: , For the contour point sequence index, is the total number of contour points in single frame; each corresponding image coordinate , i.e. the horizontal coordinate of the kth point on the contour is , and the vertical coordinate is ;

[0043] Polynomial fitting is performed on the contour point sequence to obtain the continuous function expression of the contour, i.e. the contour curve: , m is the horizontal coordinate, is the corresponding vertical coordinate, representing the geometric shape of the tth frame contour;

[0044] The curvature sequence of the contour curve is calculated , is the first order derivative of m, representing the contour tangent slope; is the second order derivative, representing the contour bending trend, and the mean value of the curvature of each frame contour is calculated , is the horizontal coordinate of the kth contour point; Curvature entropy calculation: the contour curvature entropy is calculated based on the mean value of 5 frames of curvature:

[0045] The greater the entropy value, the more significant the weld contour curvature fluctuation, and the more serious the geometric distortion;

[0046] Undercut and excess height parameter calculation: the undercut area in is extracted by morphological closing operation (structure element size 5x5), and the ratio of undercut area to total weld area is calculated , is the total number of pixels in the undercut area, is the total number of pixels in the weld area.

[0047] Based on the contour curve , the excess height value of each frame of weld (excess height is the difference between the vertical coordinates of the highest point of the contour and the base surface of the base material) is calculated, and the standard deviation of 5 frames of excess height is calculated , representing the uniformity of excess height distribution;

[0048] Geometric shape feature function construction: combined with the curvature entropy standard value of qualified weld , the undercut area ratio standard value , the excess height standard deviation and the curvature mean value standard value , the geometric shape feature function is constructed: , wherein, is the geometric shape qualified reference, indicates the existence of geometric defects, and the greater the value, the more serious the defect degree.

[0049] S202: construct the thermal feature function: ​

[0050] Based on the weld heat distribution matrix collected in S102 And the welding process parameters in S105 , through Gaussian filtering denoising, thermal parameter statistics, process parameter correlation analysis to construct the characteristic function, the specific steps are as follows:

[0051] Thermal field data preprocessing: Gaussian filtering (kernel size 7x7, σ=2.0) is performed on each frame of temperature matrix to eliminate thermal imaging noise and obtain the denoised temperature matrix

[0052] Thermal parameter calculation: calculate the variance of the filtered temperature field: , is the total number of pixels of the infrared image, , is the summation index; is the average temperature of the t-th frame, which represents the uniformity of temperature distribution;

[0053] Based on the temperature difference between adjacent frames, the average cooling rate of the weld area is calculated , is the total duration of 5 frames; is the average temperature of the first frame, is the average temperature of the fifth frame, combined with the welding current in correct the cooling rate, the correction coefficient , is the standard welding current, and the corrected cooling rate is obtained ;

[0054] Thermal characteristic function construction: combined with the standard value of the temperature variance of the qualified weld , the corrected cooling rate standard value , the thermal characteristic function is constructed: , is the thermal characteristic qualified reference, indicates the presence of thermal defects, and the larger the value, the more significant the deviation of the thermal distribution or cooling characteristics from the qualified standard;

[0055] S203: Construct the texture characteristic function:

[0056] Based on the weld microtexture matrix collected in S103 , the characteristic function is constructed by gray level co-occurrence matrix calculation and texture parameter extraction; the specific steps are as follows:

[0057] Gray level co-occurrence matrix calculation: for ​​​Perform block processing (block size 50×50 pixels), calculate the gray-level co-occurrence matrix of each block (sampling distance 2 pixels, angle coverage 0° / 45° / 90° / 135°), and eliminate local texture fluctuation interference;

[0058] Texture Parameter Extraction: Extracting three core texture parameters from the gray-level co-occurrence matrix: setting , (For grayscale indexes)

[0059] Contrast (CON): Characterizes the degree of difference in grayscale among texture pixels. G represents the total number of gray levels. The probability value of the gray-level co-occurrence matrix is ​​calculated as follows: First, [the following is a list of values]... The 12-bit grayscale values ​​from 0 to 4095 are normalized to 16 grayscale levels to simplify calculations. Then, following the spatial rules of a sampling distance of 2 pixels and angle coverage of 0° / 45° / 90° / 135°, the effective pixels in the texture matrix are traversed to count the grayscale level of the current pixel. Gray levels of neighboring pixels at corresponding spatial locations The frequency of each grayscale combination is calculated, and finally, the frequency of each grayscale combination is divided by the total frequency of all combinations to obtain the probability of that grayscale combination occurring. .

[0060] Entropy (ENT): Characterizes the degree of disorder in texture information. ;

[0061] Correlation (COR): Characterizes the gray-level correlation of texture pixels. , The grayscale mean;

[0062] Texture feature function construction: combining the standard values ​​of texture parameters of qualified welds. , , Construct texture feature functions: , As a benchmark for qualified texture features, This indicates the presence of microscopic texture defects.

[0063] S204: Constructing three-dimensional topological characteristic functions:

[0064] Based on S104 data acquisition 3D point cloud data and The normal vector matrix is ​​used to construct eigenfunctions through voxelization, topological entropy calculation, and normal vector deviation analysis; the specific steps are as follows:

[0065] Point cloud data voxelization: for Voxelization (voxel size 0.02 mm³) is performed to convert the 3D point cloud into a voxel mesh. , , , Using voxel center coordinates facilitates topological feature calculation; based on S101 Mapping the two-dimensional contour of the weld to three-dimensional space, defining the voxel range of the weld region, and counting the number of all voxels within that range. The total volume of the weld element That is, the total volume of the weld voxels is the sum of the volumes of all voxels within the weld region; subsequently, within the weld voxel region, the volume is determined by the height of the base metal reference plane. Using this as a reference, voxel height was selected. and For voxels with an absolute difference greater than 0.1 mm (the acceptable weld height deviation threshold), count the number of such defect voxels. The total volume of the defective voxels That is, the total volume of defective voxels is the sum of the volumes of all voxels with abnormal heights;

[0066] 3D topology parameter calculation:

[0067] Topological entropy : ,in, The total number of voxels. , For summation index; topological entropy is used to characterize the degree of irregularity of a three-dimensional surface;

[0068] Defect volume ratio Extracting voxel mesh The depressions or protrusions of the voxel To determine the acceptable weld height deviation threshold, calculate the ratio of the volume of the depression or protrusion to the total volume of the weld. ;

[0069] Normal vector deviation angle :based on Calculate the angle between the normal vector at each point and the reference normal vector (perpendicular to the plane of the plate). Find the average deviation angle of all points. , The total number of points in the point cloud represents the degree of surface stress concentration.

[0070] Root mean square height : Characterizes surface smoothness;

[0071] Construction of 3D topological characteristic functions: combining the standard value of the topological entropy of qualified welds. Standard value of defect volume ratio Standard value of normal vector deviation angle and the high root mean square standard value , the three-dimensional topological characteristic function is constructed: , is the three-dimensional topological qualified reference, indicates that there is a three-dimensional topological defect, and the larger the value is, the more serious the defect is;

[0072] S3: Deep fusion function modeling and calculation: based on the constructed geometric morphological characteristic function, thermal characteristic function, texture characteristic function and three-dimensional topological characteristic function, a nonlinear deep fusion model is constructed, and a weld comprehensive quality value which can comprehensively reflect the overall quality level of the weld is outputted;

[0073] Based on the image analysis fusion algorithm, the geometric morphological, thermal, texture and three-dimensional topological characteristic functions constructed in S2 are integrated, a nonlinear deep fusion model is constructed, the limitation of traditional weighted fusion is broken through, and through the coupling of geometric constraint, thermal probability mapping, texture attention weighting and three-dimensional residual compensation, the multi-dimensional defect feature is quantified into a single comprehensive quality value Q, realizing the overall quantitative representation of the welding quality. The specific mathematical function is as follows:

[0074]

[0075] , wherein, is a fusion dimension index parameter, only used as a summation index; k=1 corresponds to , k=2 corresponds to , k=3 corresponds to , and k=4 corresponds to ; is the Gaussian kernel width, an engineering calibration value, used for smoothing constraint of thermal feature probability mapping;

[0076] is the thermal probability mapping term, which maps the thermal feature value to a probability weight based on Gaussian distribution: when , the value of this term is 1, without weight attenuation; when deviates from 1, the value of this term is Gaussian attenuation, which weakens the excessive influence of thermal defects on the comprehensive value, and adapts to the physical law that welding quality thermal features are auxiliary, and geometry and topology are core;

[0077] is the texture attention weighting term, which introduces a hyperbolic tangent function to construct an attention mechanism: when , that is, the texture feature is qualified, , without additional weighting; when , that is, the texture defect, , with nonlinearly increasing, strengthening the contribution of microscopic texture defects to the comprehensive quality, and adapting to the detection needs of hidden defects such as pores and micro-cracks;

[0078] The three-dimensional residual compensation term integrates the coupling relationship between three-dimensional topological features and geometric and thermal features: The residual value of the three-dimensional topological defect represents the degree to which the three-dimensional feature deviates from the acceptable standard; multiplied by Achieve coupled compensation for geometric and three-dimensional defects; the more severe the geometric defect, the higher the weight of the three-dimensional residual. As a thermal correction factor, the more uneven the heat distribution, the smaller the correction factor, which weakens the overcompensation of the three-dimensional residual and fits the physical nature of the coupling of heat, force and form in welding.

[0079] For the normalization constraint term, the square root of the sum of fourth powers is used as the normalization denominator to strengthen the constraint on multi-dimensional extreme defects: when any dimension When the value deviates significantly from 1, the denominator increases non-linearly. This prevents the overall value Q from becoming uncontrollable due to single-dimensional defects and ensures that the range of Q values ​​remains stable. And Q=1 is the standard for acceptable welding quality;

[0080] It should be further noted that in the experiment verifying the deep fusion function model, the geometric morphological feature function values ​​were used as control variables, as shown in the table below:

[0081]

[0082] S4: Multi-dimensional Defect Source Tracing and Dynamic Grading Judgment: Based on the comprehensive weld quality value output by S3 and the characteristic functions of each dimension of S2, the dominant defect dimension is located and the specific defect type is identified by calculating the abnormal coefficient of the characteristic functions of each dimension. A dynamic threshold model is then constructed to complete the determination of the weld quality level.

[0083] Based on the comprehensive quality value Q output by S3 and the feature functions of each dimension constructed by S2, a three-layer judgment system is built, which includes dimensional defect localization, dynamic threshold grading, and defect impact assessment. This system breaks through the limitations of traditional single threshold judgment and achieves accurate source tracing of defect types, dynamic adaptation of quality levels, and differentiated output of processing strategies. The specific steps are as follows:

[0084] S401: Dimensional Defect Location and Type Identification:

[0085] Based on the degree of anomaly of the S2 feature functions deviating from the acceptable benchmark, combined with a pre-trained defect pattern library (containing 5000 sets of mapping samples between feature anomalies and defect types), the dominant defect dimension is accurately located and the defect type is subdivided, and the corresponding defect level is associated:

[0086] Calculate the anomaly coefficients for each dimension: , , , ,in The greater the abnormality coefficient, the greater the contribution of the dimension to the quality deviation.

[0087] If , it is determined to be a geometry-dominant defect, and the parameter matching mode library of S201 is combined:

[0088] If and , it is determined to be a flash and residual height uneven defect, which is a slight defect.

[0089] If , it is a weld contour distortion defect, which is a serious defect.

[0090] If , it is determined to be a heat-dominant defect, and the parameter matching mode library of S202 is combined:

[0091] If , it is a heat distribution uneven defect, which is a slight defect.

[0092] If , it is a grain coarsening defect caused by abnormal cooling rate, which is a serious defect.

[0093] If , it is determined to be a texture-dominant defect, and the parameter matching mode library of S203 is combined:

[0094] If and , it is a pore or micro-crack defect, the pore defect is a slight defect, and the micro-crack defect is a fatal defect.

[0095] If , it is a slag inclusion or spatter defect, which is a serious defect.

[0096] If , it is determined to be a three-dimensional topology-dominant defect, and the parameter matching mode library of S204 is combined:

[0097] If , it is a concave or convex defect, when the concave or convex depth is less than 0.2mm, it is a slight defect, and when the concave or convex depth is greater than or equal to 0.2mm, it is a serious defect.

[0098] If , it is a weld surface stress concentration defect, which is a serious defect.

[0099] If there are two or more dimensions , it is determined to be a coupled defect, and it is simultaneously determined to be a fatal defect.

[0100] S402: Dynamic threshold grading decision: Combine the weld type, workpiece service scenario and historical detection data to build a dynamic grading threshold model to ensure that the decision result adapts to different industrial online detection needs, which are as follows:

[0101] Define the basic threshold matrix , which corresponds to the initial critical value of slight defect and serious defect, respectively;

[0102] Introduce scene correction coefficient : The quality requirement of pressure parts is higher, take ; Take for ordinary structural parts, calculate the dynamic threshold:

[0103] Slight defect dynamic threshold: ;

[0104] Serious defect dynamic threshold: ;

[0105] Introduce defect type correction coefficient : Take for fatal defect, take for serious defect, take for slight defect, calculate the corrected comprehensive quality value ;

[0106] Final grading decision rule:

[0107] Qualified level: and no single dimension , indicating that the defects in each dimension are within the allowable range, and there is no slight and above defect, which can continue online detection;

[0108] Slight defect level: or single dimension , indicating that there is a slight defect, and continue detection after real-time labeling of defect information;

[0109] Serious defect level: or single dimension , indicating that there is a serious defect, which immediately triggers online detection warning, labels defect details and prompts subsequent repair process;

[0110] Fatal defect level: identified as fatal defect type by S401, regardless of value, it is directly judged as fatal defect, which immediately triggers emergency warning and suspends online detection, avoiding unqualified products flowing into the next process;

[0111] Data anomaly level: or any , prompting that the acquisition equipment has calibration deviation, which needs to suspend online detection and output equipment maintenance prompt.

[0112] Secondly: the embodiment of the present application discloses only the structure involved in the embodiment of the present application, other structures can refer to the general design, and in the case of no conflict, the same embodiment and different embodiments of the present application can be combined with each other;

[0113] Finally: the above only describes the preferred embodiments of the present application and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A machine vision-based online welding quality inspection method, characterized in that, include: S1: Multi-dimensional basic data acquisition: Through the deployment of multi-modal acquisition equipment, the raw data of the weld area is acquired in all dimensions, and the timing and spatial synchronization of each device is realized through the time protocol hardware synchronization system to build a raw dataset of the weld in all dimensions, including geometry, thermal, texture, three-dimensional topology and process parameters. S2: Multi-dimensional feature quantization extraction and construction: The feature decoupling and quantization calculation are performed on the full-dimensional raw data collected by S1. For the four dimensions of weld geometry, thermal properties, micro texture and three-dimensional topology, geometric feature function, thermal feature function, texture feature function and three-dimensional topology feature function are constructed respectively, and the unstructured raw data is transformed into structured feature values ​​that can characterize the degree of defects. S3: Deep Fusion Function Modeling and Calculation: Based on the constructed geometric morphology feature function, thermal feature function, texture feature function and three-dimensional topology feature function, a nonlinear deep fusion model is constructed to output a comprehensive weld quality value that can comprehensively reflect the overall quality level of the weld. S4: Multi-dimensional Defect Source Tracing and Dynamic Grading Judgment: Based on the comprehensive weld quality value output by S3 and the characteristic functions of each dimension of S2, the dominant defect dimension is located and the specific defect type is identified by calculating the abnormal coefficient of the characteristic functions of each dimension. A dynamic threshold model is then constructed to complete the determination of the weld quality level.

2. The online welding quality inspection method based on machine vision according to claim 1, characterized in that: The full-dimensional original dataset includes: It includes a sequence of two-dimensional weld morphology images, weld thermal distribution images and temperature field data matrix, weld microtexture matrix, weld three-dimensional coordinate point cloud data and surface normal vector matrix, and also incorporates process parameters during the welding process; the data in each dimension are matched through a temporal and spatial synchronization mechanism.

3. The online welding quality inspection method based on machine vision according to claim 1, characterized in that: The geometric morphological feature functions include: Contour extraction is performed on the two-dimensional morphological image sequence of the weld, and the curvature and mean curvature of the contour curve are calculated. The contour curvature entropy is calculated based on the mean curvature of multiple frames. The undercut area is extracted and the undercut area ratio is calculated. The standard deviation of the remaining height is calculated in combination with the contour curve. Then, by combining the standard values ​​of the curvature entropy, undercut area ratio, standard deviation of remaining height and mean curvature of qualified welds, geometric morphological feature functions are constructed by performing nonlinear operations such as squaring, hyperbolic tangent transformation, exponential operation and product coupling on each parameter.

4. The online welding quality inspection method based on machine vision according to claim 1, characterized in that: The thermal characteristic function includes: The weld thermal distribution image and temperature field data matrix are preprocessed by filtering and noise reduction. The variance of the temperature field is calculated to characterize the uniformity of temperature distribution. The average cooling rate of the weld area is calculated based on multi-frame temperature data, and the cooling rate is corrected by combining welding process parameters. Then, by combining the standard value of temperature variance of qualified welds and the standard value of corrected cooling rate, a thermal characteristic function is constructed by weighting the parameters.

5. The online welding quality inspection method based on machine vision according to claim 1, characterized in that: The texture feature function includes: The micro-texture matrix of the weld is normalized to gray level, and a gray-level co-occurrence matrix is ​​constructed according to the set sampling distance and angle. The contrast, entropy and correlation coefficient of the gray-level co-occurrence matrix are calculated. Then, combined with the standard values ​​of contrast, entropy and correlation coefficient of qualified welds, the texture feature function is constructed by performing nonlinear mapping operation on each parameter.

6. The online welding quality inspection method based on machine vision according to claim 1, characterized in that: The three-dimensional topological feature functions include: The three-dimensional coordinate point cloud data of the weld is processed into voxels, and the ratio of the total volume of defect voxels to the total volume of weld voxels is calculated. The mean angle of each normal vector in the normal vector matrix of the weld surface is calculated to characterize the surface flatness. Then, by combining the standard value of the voxel volume ratio of qualified welds and the standard value of the mean angle of normal vectors, a three-dimensional topological feature function is constructed by coupling the parameters.

7. The online welding quality inspection method based on machine vision according to claim 1, characterized in that: The nonlinear deep fusion model includes: This paper integrates geometric morphological feature functions, thermal feature functions, texture feature functions, and 3D topological feature functions. A thermal probability mapping mechanism is introduced to adjust the weights of the thermal feature functions. A texture attention weighting mechanism is used to enhance the contribution of micro-texture defects. A 3D residual compensation mechanism is combined to achieve coupling compensation of 3D topological features with other dimensional features. At the same time, a normalization constraint term is introduced to constrain the extreme values ​​of multi-dimensional features. Finally, the comprehensive weld quality value is output through the coupling operation of the numerator and denominator, thus constructing a nonlinear deep fusion model.

8. The online welding quality inspection method based on machine vision according to claim 1, characterized in that: The process of locating the dominant defect dimension and identifying specific defect types includes: Calculate the anomaly coefficients of the feature functions of each dimension, including geometry, thermal, texture, and 3D topology, that deviate from the qualified benchmark. Determine the dominant defect dimension based on the proportion of the anomaly coefficients. For the dominant defect dimension, match the core parameters corresponding to the feature function of that dimension with a preset defect pattern library to identify the specific defect type. If the anomaly coefficients of two or more dimensions reach the set proportion, it is determined to be a coupled defect.

9. The online welding quality inspection method based on machine vision according to claim 1, characterized in that: The construction of the dynamic threshold model includes: A basic threshold for minor and severe defects is set, and a scenario correction coefficient is introduced to adjust the basic threshold to obtain a dynamic threshold. A defect type correction coefficient is introduced and the overall weld quality value is corrected according to the defect type. The corrected overall weld quality value is compared with the dynamic threshold. Based on the degree of deviation of the characteristic functions of each dimension from the qualified benchmark, the judgment rules for weld quality level are formulated, and the construction of the dynamic threshold model is completed.

Images