Detection method, system and equipment for intelligent welding of pipe pile splicing and medium

Abstract

The invention relates to the technical field of image processing, and discloses a detection method, system and equipment for intelligent welding of pipe pile splicing and a medium. The method comprises the steps that noise reduction is conducted on a target weld seam image through a filtering algorithm, registration and splicing are conducted on the image through feature matching, and a weld seam unfolding image is obtained; segmenting the weld joint expanded image by using a deep learning network, and extracting a weld joint center line and surface features to obtain weld joint feature data; defect detection and recognition are conducted on the weld joint feature data, hierarchical calculation is conducted through a rule set, and weld joint quality data are obtained; parameter mapping is conducted according to the weld joint quality data, a welding parameter instruction is generated, welding process variables are adjusted, and target welding parameters are obtained; and the welding seam quality data and the target welding parameters are stored, and a quality report is generated through data analysis. The detection accuracy of intelligent welding for pipe pile splicing is improved.

Description

Technical Field

[0001] The present application relates to the field of image processing, and in particular to a detection method, system, equipment and medium for intelligent welding of pipe piles. Background Art

[0002] Pipe pile welding is a key process in large-scale engineering construction, and its quality directly affects the safety and durability of the project. Traditional pipe pile welding quality inspection mainly relies on manual visual inspection and sampling destructive testing. This method has problems such as strong subjectivity, low efficiency and high cost. With the development of technology, some automated detection methods have been introduced, such as ultrasonic detection and X-ray detection. These methods can improve the objectivity and efficiency of detection to a certain extent, but there are still problems such as incomplete detection coverage and lack of real-time performance.

[0003] Although the existing pipe pile welding quality inspection technology is constantly improving, it still faces some challenges. Most inspection methods are difficult to achieve comprehensive monitoring of the entire welding process, and can often only be inspected after welding is completed, and problems in the welding process cannot be discovered and handled in a timely manner. Secondly, the existing inspection data analysis methods are usually relatively simple, and it is difficult to deeply explore the complex relationships contained in the data, and it is impossible to provide strong data support for the continuous improvement of welding quality. In addition, the generation of quality reports often requires manual sorting and writing, which is not only time-consuming and labor-intensive, but also easily affected by subjective factors, making it difficult to ensure the consistency and objectivity of the report. Summary of the invention

[0004] The present application provides a detection method, system, equipment and medium for intelligent welding of pipe pile joints, which are used to solve the problems of insufficient data analysis and low efficiency in quality report generation in existing pipe pile joint welding quality detection technologies.

[0005] In a first aspect, the present application provides a detection method for intelligent welding of pipe pile joints, and the detection method for intelligent welding of pipe pile joints includes: acquiring images at pipe pile joints to obtain original weld images, calibrating the original weld images according to lighting parameters to obtain target weld images; denoising the target weld images through a filtering algorithm, and aligning and splicing the images through feature matching to obtain weld expansion images; segmenting the weld expansion images using a deep learning network, extracting weld center lines and surface features, and obtaining weld feature data; performing defect detection and identification on the weld feature data, and performing hierarchical calculations through a set of rules to obtain weld quality data; performing parameter mapping based on the weld quality data, generating welding parameter instructions, adjusting welding process variables, and obtaining target welding parameters; storing the weld quality data and target welding parameters, and generating a quality report through data analysis.

[0006] In the second aspect, the present application provides a detection system for intelligent welding of pipe pile joints, and the detection system for intelligent welding of pipe pile joints includes: an acquisition module, which is used to acquire images at the pipe pile joints to obtain an original weld image, and calibrate the original weld image according to lighting parameters to obtain a target weld image; a denoising module, which is used to denoise the target weld image through a filtering algorithm, and align and splice the images through feature matching to obtain a weld expansion image; a segmentation module, which is used to segment the weld expansion image using a deep learning network, extract the weld centerline and surface features, and obtain weld feature data; an identification module, which is used to perform defect detection and identification on the weld feature data, and perform hierarchical calculation through a set of rules to obtain weld quality data; a mapping module, which is used to perform parameter mapping based on the weld quality data, generate welding parameter instructions, adjust welding process variables, and obtain target welding parameters; a storage module, which is used to store the weld quality data and target welding parameters, and generate a quality report through data analysis.

[0007] The third aspect of the present application provides a computer device, wherein the memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor and the memory communicate through a bus, and when the machine-readable instructions are executed by the processor, the steps of the above-mentioned detection method for intelligent welding of pipe piles are performed.

[0008] A fourth aspect of the present application provides a computer-readable storage medium, in which instructions are stored. When the computer-readable storage medium is run on a computer, the computer executes the above-mentioned detection method for intelligent welding of pipe pile joints.

[0009] In the technical solution provided by the present application, by converting the weld quality data and the target welding parameters into a time series data stream and performing time marking, the comprehensive and continuous monitoring of the welding process is achieved, and the time series and integrity of the data are improved; multi-dimensional index construction and correlation calculation are performed on the data record blocks, thereby enhancing the data retrieval efficiency and correlation analysis capabilities; by extracting the time series characteristics and spatial characteristics of the data blocks, the dynamic changes and spatial distribution characteristics of the welding process are fully captured; the characteristic sequence data are classified according to the quality indicators and the statistical characteristic values ​​are calculated, thereby achieving a refined analysis of different quality indicators; through normalization processing and the construction of a quality evaluation matrix, the comparability between different quality indicators and the accuracy of the comprehensive evaluation are improved; the quality statistical data are divided into time windows and trend analysis is performed, thereby enhancing the grasp and prediction of the welding quality change trend; by extracting key quality Inflection points and generation of trend analysis data help to promptly detect abnormal changes in the welding process; comparing and analyzing quality prediction results with historical data improves the accuracy and reliability of quality assessment; by calculating quality deviation values ​​and extracting quality abnormality intervals, the ability to accurately identify welding quality problems is enhanced; causal analysis based on quality diagnosis data deepens the understanding of the causes of welding quality problems; structuring the cause analysis results and extracting key quality information improves the readability and practicality of the analysis results; by converting quality information into descriptive text and generating quality description data, an effective conversion from data to understandable information is achieved; finally, by formatting quality description data, generating chart indexes, and constructing a quality report structure, high-quality and professional welding quality reports are automatically generated, greatly improving the efficiency and consistency of report generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0011] Figure 1 A schematic diagram of an embodiment of a detection method for intelligent welding of pipe pile connections in an embodiment of the present application; Figure 2 A schematic diagram of an embodiment of a detection system for intelligent welding of pipe piles in an embodiment of the present application; Figure 3 It is a schematic diagram of the structure of a computer device in an embodiment of the present application. DETAILED DESCRIPTION

[0012] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 , an embodiment of the detection method for intelligent welding of pipe pile joints in the embodiment of the present application includes: Step S101, collecting images of the pipe pile joints to obtain an original weld image, and calibrating the original weld image according to illumination parameters to obtain a target weld image; Step S102, reducing noise on the target weld image by using a filtering algorithm, registering and splicing the image by feature matching, and obtaining a weld expansion image; Step S103, using a deep learning network to segment the weld seam unfolded image, extract the weld seam centerline and surface features, and obtain weld seam feature data; Step S104, performing defect detection and identification on the weld feature data, performing hierarchical calculations through a set of rules, and obtaining weld quality data; Step S105, performing parameter mapping according to the weld quality data, generating welding parameter instructions, adjusting the welding process variables, and obtaining target welding parameters; Step S106: storing the weld quality data and target welding parameters, and generating a quality report through data analysis.

[0013] It is understandable that the execution subject of the present application may be a detection system for intelligent welding of pipe piles, or a terminal or a server, which is not limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0014] Specifically, the image acquisition is performed on the pipe pile joint to obtain the original weld image, and the original weld image is calibrated according to the illumination parameters to obtain the target weld image. A high-resolution industrial camera array is used to perform all-round image acquisition around the pipe pile joint. During the operation of the welding robot, four industrial cameras are set up at a 90-degree angle and evenly distributed around the pipe pile. Each camera is equipped with a ring LED light source to provide stable lighting. The camera acquisition frequency is set to 30 frames per second, and the resolution is not less than 4096×3072 pixels to ensure that the subtle features of the weld can be captured. The acquired image data is transmitted to the image processing unit in real time via Gigabit Ethernet. To ensure image quality, the camera parameters, including exposure time, gain value and white balance, are monitored in real time during the acquisition process, and the output power of the LED light source is automatically adjusted according to the ambient light intensity. In addition, a calibration plate is placed within the field of view of each camera for subsequent image correction and size measurement. During the illumination calibration process, the image grayscale value distribution is extracted, the mean and variance of the light intensity in the image area are calculated, and the real-time illumination parameters are generated. It should be noted that illumination parameters refer to a set of key data that characterizes the illumination characteristics of an image, obtained by calculation during the image acquisition process. Specifically, illumination parameters include the mean and standard deviation of the light intensity in the image area. The mean light intensity reflects the overall brightness level of the image, which is obtained by calculating the average grayscale value of all pixels in the image; the standard deviation characterizes the discrete degree of brightness distribution in the image and reflects the contrast characteristics of the image. These two parameters together constitute the real-time illumination parameters.

[0015] In order to obtain these lighting parameters, this application uses four industrial cameras evenly distributed at a 90-degree angle around the pile for image acquisition. Each camera is equipped with a ring-shaped LED light source to provide stable lighting. In practical applications, the original image with a resolution of 4096×3072 is obtained, and then the grayscale value of the entire image area is counted and calculated at the pixel level to obtain real-time lighting parameters that reflect the lighting conditions of the current acquisition environment. By comparing and calculating the lighting parameters obtained in real time with the preset standard lighting values, an accurate compensation coefficient is obtained, thereby achieving accurate calibration of the image.

[0016] Extract the gray value distribution features from the image and calculate the illumination parameters. For the original image I(x,y), the calculation process is as follows: Average light intensity of the image area Calculation: ; in: is the gray value of the image at position (x, y); M×N is the total number of pixels in the image (4096×3072); This is the mean value of the real-time lighting parameters.

[0017] Image area light intensity variance Calculation:

[0018] in: It is the standard deviation in real-time illumination parameters, which is used to characterize the contrast characteristics of the image.

[0019] The difference between the real-time illumination parameters and the preset standard illumination values ​​is calculated to generate the compensation coefficient. The pixel values ​​of the original weld image are linearly transformed according to the compensation coefficient to obtain the brightness correction image. Then the brightness correction image is separated into RGB channels, the mean ratio of each channel pixel is calculated, the color calibration parameter matrix is ​​generated, and the pixel values ​​are channel mapped according to the color calibration parameter matrix to obtain the color calibration data. The color calibration data is reconstructed into an image matrix according to the pixel position, the contrast and sharpness of the image matrix are numerically adjusted, the enhancement coefficient is generated, and the image is pixel-level remapped according to the enhancement coefficient to obtain the target weld image. The target weld image is denoised by the filtering algorithm, and the image is registered and spliced ​​by feature matching to obtain the weld expansion image. In the denoising process, the target weld image is divided into overlapping sub-regions, the pixel values ​​of the sub-regions are counted, the regional noise intensity distribution is calculated, and the noise feature data is generated. The filter window size is set according to the noise feature data, the target weld image is scanned by a sliding window, the weight coefficient of the pixel value in the window is calculated, and the denoising processing data is obtained. Commonly used filtering algorithms include Gaussian filtering and median filtering. For feature matching and image stitching, the denoising data is divided into image blocks, the gradient feature points of the image blocks are extracted, the feature point description vector is established, and the feature matching data is generated. Then, the spatial distance of the feature points in the feature matching data is calculated, the unmatched point pairs are eliminated, and the position correspondence between the image blocks is established to obtain the image registration data. The image registration data is reconstructed according to the spatial position relationship, the pixel difference of the overlapping area of ​​adjacent image blocks is calculated, the image fusion weight is generated, and the stitching data is obtained. The stitching data is transformed and resampled, and the image edge is smoothed to obtain the weld expansion image. The weld expansion image is segmented using a deep learning network, the weld centerline and surface features are extracted, and the weld feature data is obtained. The improved U-Net network is used to perform semantic segmentation on the weld area. The network contains 5 layers of encoders and 5 layers of decoders, and each layer uses residual connections to improve the feature extraction capability. The weld area obtained by segmentation is subjected to morphological refinement processing to extract the weld centerline. A sampling point is taken every 1 mm along the weld centerline, and the cross-sectional contour curve is extracted in the direction perpendicular to the centerline. The geometric feature parameters of each contour curve are calculated, including weld width, excess height, undercut depth, etc. At the same time, the texture features of the weld surface are extracted using the grayscale co-occurrence matrix, including statistics such as energy, contrast, and entropy, which are used to evaluate the quality of weld formation. These features together constitute the weld feature data. Defect detection and identification are performed on the weld feature data, and the weld quality data is obtained by hierarchical calculation through a set of rules. During the defect detection process, a one-dimensional convolutional neural network is used to detect anomalies in the cross-sectional profile of the weld. The network contains 3 convolutional layers and 2 fully connected layers to identify defects such as excessive geometric dimensions and undercuts. Then, the improved YOLOv5 model is used to detect targets on the weld surface and identify surface defects such as pores, cracks, and slag inclusions.The model backbone network uses CSPDarknet53, and introduces the attention mechanism to improve detection accuracy. For each defect detected, its characteristic parameters such as location, size, type, etc. are extracted, and the severity score of the defect is calculated. Through a comprehensive evaluation system based on fuzzy rules, the weld quality is graded and evaluated according to the number, distribution and severity of various defects to obtain weld quality data.

[0020] Parameter mapping is performed based on the weld quality data to generate welding parameter instructions, adjust the welding process variables, and obtain the target welding parameters. A mapping relationship database between welding parameters and defect characteristics is constructed, and a time series prediction model is established using a long short-term memory network (LSTM) to predict the weld quality trend in real time. When defects or abnormal quality indicators are detected, welding parameter optimization suggestions are automatically generated according to the defect type and the prediction model, including the adjustment amount of parameters such as current, voltage, and welding speed. Then, the parameter optimization suggestions are converted into a specific control instruction sequence, and the instruction sequence is time-series encoded to generate welding parameter instructions. The adjustment amount of the process variables is calculated according to the welding parameter instructions, and the adjustment amount is dynamically compensated to obtain the target welding parameters. The weld quality data and the target welding parameters are stored, and a quality report is generated through data analysis. During the storage process, the data is converted into a time series data stream, the time series data stream is time-stamped, and the data is divided into blocks according to the timestamp to generate data record blocks. Multidimensional indexes are constructed for the data record blocks, the correlation between the data blocks is calculated, and the time series features and spatial features of the data blocks are extracted to obtain feature sequence data. Then, statistical analysis is performed on the characteristic sequence data to generate quality statistical data. Trend analysis and quality prediction are performed based on the quality statistical data to obtain quality prediction results. The quality prediction results are compared and analyzed with historical data to generate quality diagnosis data, and causal correlation analysis is performed to obtain cause analysis results. The cause analysis results are structured, key quality information is extracted, and the quality information is converted into descriptive text to generate quality description data, and a quality report is output based on the quality description data.

[0021] In a specific embodiment, the process of executing step S101 may specifically include the following steps: (1) An industrial camera is used to collect circumferential distribution data at the pipe pile joints, and the collected data is converted into a digital signal stream to obtain multiple groups of original weld images; (2) Timestamp multiple groups of original weld images, integrate and sort the data streams according to the time series, and obtain serialized original weld images through data transmission; (3) Extract the image grayscale value distribution from the serialized original weld image, calculate the mean and variance of the light intensity in the image area, and generate real-time lighting parameters; (4) Calculate the difference between the real-time illumination parameters and the preset standard illumination values ​​to generate a compensation coefficient. Perform a linear transformation on the pixel values ​​of the original weld image according to the compensation coefficient to obtain a brightness-corrected image. (5) Separate the RGB channels of the brightness-corrected image, calculate the mean ratio of the pixels in each channel, generate a color calibration parameter matrix, and perform channel mapping on the pixel values ​​according to the color calibration parameter matrix to obtain color calibration data; (6) The color calibration data is used to reconstruct the image matrix according to the pixel position, the contrast and sharpness of the image matrix are numerically adjusted to generate an enhancement coefficient, and the image is remapped at the pixel level according to the enhancement coefficient to obtain the target weld image.

[0022] Specifically, the industrial camera is used to collect the circumferential distribution of the pipe pile joints, and the collected data is converted into a digital signal stream to obtain multiple groups of original weld images. A high-resolution industrial camera array is used to collect all-round images around the pipe pile joints. During the operation of the welding robot, four industrial cameras are set up at a 90-degree angle and evenly distributed around the pipe pile. Each camera is equipped with a ring LED light source to provide stable lighting. The camera acquisition frequency is set to 30 frames per second, and the resolution is not less than 4096×3072 pixels to ensure that the subtle features of the weld can be captured. The collected image data is transmitted to the image processing unit in real time via Gigabit Ethernet. To ensure image quality, the camera parameters, including exposure time, gain value and white balance, are monitored in real time during the acquisition process, and the output power of the LED light source is automatically adjusted according to the ambient light intensity. Multiple groups of original weld images are timestamped, the data streams are integrated and sorted according to the time series, and the serialized original weld images are obtained through data transmission. Timestamp marking is a technology that adds time information to data, which gives each frame of the image an accurate time mark. In the integration and sorting process, images from different cameras are sorted according to timestamps to form an image sequence. The image grayscale value distribution is extracted from the serialized original weld image, the mean and variance of the image area light intensity are calculated, and the real-time lighting parameters are generated. The grayscale value distribution reflects the brightness of the image and is an important feature in image processing. The mean light intensity represents the overall brightness level of the image, while the variance reflects the discrete degree of the image brightness.

[0023] The difference between the real-time illumination parameters and the preset standard illumination values ​​is calculated to generate a compensation coefficient. The pixel values ​​of the original weld image are linearly transformed according to the compensation coefficient to obtain a brightness correction image. This process can be expressed by the following formula:

[0024] in, represents the corrected pixel value, is the preset standard illumination value. and are the mean and standard deviation of the real-time lighting parameters, is the original pixel value.

[0025] The brightness correction image is separated into RGB channels, the mean ratio of each channel pixel is calculated, and the color calibration parameter matrix is ​​generated. The pixel values ​​are channel mapped according to the color calibration parameter matrix to obtain color calibration data. RGB channel separation is the process of decomposing a color image into three monochrome channels of red, green, and blue. By calculating the mean ratio of each channel pixel, the overall color tendency of the image can be obtained. The generation process of the color calibration parameter matrix can be expressed by the following formula:

[0026] in, , , They are the calibration coefficients for the red, green, and blue channels, respectively, and their calculation is based on the deviation of each channel from the ideal white balance state. By mapping the original pixel values ​​through this matrix, the color balance of the image can be corrected. The color calibration data is reconstructed into an image matrix according to the pixel position, and the contrast and sharpness of the image matrix are numerically adjusted to generate an enhancement coefficient. The image is remapped at the pixel level according to the enhancement coefficient to obtain the target weld image. Contrast adjustment can be achieved by the following formula:

[0027] in, is the adjusted pixel value, is the input pixel value, is the contrast adjustment coefficient. Sharpness enhancement can be achieved through high-pass filtering, the core idea of ​​which is to enhance the high-frequency components of the image. Finally, through this series of processing, a fully corrected and enhanced target weld image is obtained.

[0028] In a specific embodiment, the process of executing step S102 may specifically include the following steps: (1) Divide the target weld image into overlapping sub-regions, perform pixel value statistics on the sub-regions, calculate the regional noise intensity distribution, and generate noise feature data; (2) Set the filter window size according to the noise characteristic data, perform sliding window scanning on the target weld image, calculate the weight coefficient of the pixel value in the window, and obtain the noise reduction processing data; (3) Divide the denoising data into image blocks, extract the gradient feature points of the image blocks, establish the feature point description vector, and generate feature matching data; (4) Calculate the spatial distance of the feature points in the feature matching data, eliminate the unmatched point pairs, establish the position correspondence between the image blocks, and obtain the image registration data; (5) Reconstruct the image registration data according to the spatial position relationship, calculate the pixel difference of the overlapping area of ​​adjacent image blocks, generate image fusion weights, and obtain spliced ​​data; (6) Perform coordinate transformation and resampling on the stitched data, perform smooth transition processing on the image edge, and obtain the weld unfolded image.

[0029] Specifically, the overlapping sub-region division method is adopted, and the pixel value statistics include calculating the statistical features such as the average grayscale value, standard deviation, maximum value and minimum value of each sub-region. These statistical features can reflect the brightness distribution within the sub-region. Through these statistical features, the regional noise intensity distribution is calculated to generate noise feature data. The noise intensity distribution reflects the spatial distribution characteristics of noise in the image. Setting the filter window size according to the noise feature data is an adaptive process. The selection of the filter window size directly affects the noise reduction effect. If the window is too large, the image details may be lost, and if the window is too small, the noise may not be effectively removed. Therefore, dynamically adjusting the filter window size according to the noise feature data can achieve a balance between retaining image details and removing noise.

[0030] The target weld image is scanned with a sliding window, and the weight coefficient of the pixel value in the window is calculated. This process can be expressed by the following formula:

[0031] in, It's pixels The weight coefficient of , ) is the window center coordinate, It's pixels The gray value of and are the standard deviations in the spatial domain and the grayscale domain, Represents the filter window. (u, v) are the coordinates of any pixel within the filter window.

[0032] Image segmentation is the process of dividing an image into several small blocks, each of which contains a certain number of pixels. In this embodiment, an overlapping segmentation method is adopted, that is, there is a certain overlapping area between adjacent blocks, so that the boundary information will not be missed in the subsequent feature matching. Extracting the gradient feature points of the image block is a key step in feature matching. Gradient feature points are points in the image where the grayscale changes significantly, usually located at the edge or corner of the object. Commonly used gradient feature point detection algorithms include SIFT (Scale-Invariant Feature Transform) and SURF (Speeded Up Robust Features). It can detect the key points in the image and generate feature vectors describing these key points. The purpose of establishing a feature point description vector is to quantitatively describe the feature points. The feature point description vector usually contains the gradient information of the area around the feature point, which can be used to uniquely identify a feature point. Through this process, feature matching data is generated. The feature matching data contains the position information and description vector of each feature point.

[0033] The purpose of calculating the spatial distance of feature points in feature matching data is to find pairs of feature points that match each other. Spatial distance calculation usually uses Euclidean distance or Manhattan distance. By setting a distance threshold, possible matching feature point pairs can be preliminarily screened out. Unmatched point pairs are eliminated to improve the accuracy of matching. Common methods include the RANSAC (Random Sample Consensus) algorithm, which can effectively remove false matches through random sampling and consistency checks. Through this process, the position correspondence between image blocks is established to obtain image registration data. Based on the spatial relationship between image blocks, multiple image blocks are spliced ​​into a complete image. During the splicing process, the overlapping areas between image blocks need to be processed. The pixel difference in the overlapping area of ​​adjacent image blocks is calculated to generate the image fusion weight. This process can be expressed by the following formula:

[0034] in, is the pixel value after fusion, is the i-th image block at position The weight of is the i-th image block at position The pixel value of is the number of overlapping image blocks. Coordinate transformation and resampling of the stitched data is to correct the geometric deformation of the image. Coordinate transformation usually includes operations such as translation, rotation and scaling, which can align different image blocks to the same coordinate system. Resampling is the process of recalculating pixel values ​​in a new coordinate system. Common methods include nearest neighbor interpolation, bilinear interpolation and bicubic interpolation. Smoothing the image edges is to eliminate stitching marks and make the image look more natural. Common methods include feathering and gradient blending. Through this process, the weld unfolded image is obtained. The weld unfolded image is the result of unfolding the cylindrical pipe pile joint into a plane image, which provides an intuitive view for subsequent weld quality inspection.

[0035] For example: preprocess the acquired target weld image. Assume that the diameter of a pipe pile is 1 meter and the circumference of the weld is about 3.14 meters. Use 4 high-resolution cameras to capture the image in a circle, and each camera covers an arc segment of about 90 degrees. The resolution of each original image is 4096×3072 pixels, corresponding to an actual size of about 80cm×60cm. In the image division stage, each image is divided into 16×12 grids, each grid size is 256×256 pixels, and there is a 50% overlap between adjacent grids. In this way, each image is divided into 192 sub-regions. The pixel value of each sub-region is counted, and the average gray value, standard deviation and other features are calculated. Assume that in a certain sub-region, the average gray value is 128 and the standard deviation is 20. These data constitute the noise characteristics of the region. Based on the noise characteristics, the filter window size is dynamically set. For example, in areas with strong noise, an 11×11 filter window may be used, while in areas with weak noise, a 5×5 window may be used. The entire image is scanned with a sliding window, and the new value of each pixel is obtained by the weighted average of all pixels in the window. The calculation of the weight coefficient takes into account the spatial position and gray value difference of the pixel, which ensures that the edge information of the image is retained while effectively removing noise. In the feature extraction stage, the SIFT algorithm is used to extract feature points for each sub-region. The advantage of the SIFT algorithm is that it is invariant to the rotation, scaling and brightness changes of the image, which is particularly important for the analysis of weld images. Each feature point is represented by its position coordinates and a 128-dimensional description vector. For example, in a 256×256 pixel sub-region, 50-100 feature points may be extracted. The nearest neighbor matching algorithm is used in the feature matching stage to match the feature points between different sub-regions. The distance ratio threshold is set to 0.7, that is, the match is considered successful only when the ratio of the nearest neighbor distance to the next nearest neighbor distance is less than 0.7. The setting of this threshold can effectively reduce mismatching. After the preliminary matching, the RANSAC algorithm is used to further eliminate mismatching points. The RANSAC algorithm can effectively remove matching pairs that do not conform to the overall transformation model through random sampling and consistency check. In the image stitching stage, the transformation matrix between image blocks is calculated based on the matching results. For two adjacent image blocks, operations such as translation, rotation, and scaling may be required to achieve precise alignment. In the overlapping area, pixel fusion is performed by weighted averaging. The setting of weights takes into account the distance from the pixel to the boundary. The farther the pixel is from the boundary, the greater its weight. This method can achieve a smooth transition and avoid obvious stitching marks. Global geometric correction and color balance are performed on the stitching results. Geometric correction aims to eliminate image deformation caused by lens distortion, and polynomial models are usually used for correction. Color balance is to eliminate possible color differences between different cameras, and histogram matching or color transfer techniques are usually used. The weld expansion image is obtained.This image clearly shows the details of the entire weld, including the width change of the weld, surface quality, possible defects, etc. For example, in the expanded image, you can clearly see the starting and ending points of the weld, the change trend of the weld width, and possible defects such as pores and cracks.

[0036] In a specific embodiment, the process of executing step S103 may specifically include the following steps: (1) Convert the weld seam unfolded image into a pixel matrix, divide the pixel matrix into multi-scale blocks, calculate the pixel mean and variance of each scale block, and generate image segmentation area data; (2) Calculate the grayscale gradient of the image segmentation area data, extract the first-order difference values ​​in the horizontal and vertical directions, and obtain edge feature data through bidirectional gradient weighted superposition; (3) Locate the extreme points of edge feature data, calculate the spatial continuity of local maximum points, select valid edge points according to the continuity criterion, generate edge point sequences, perform curve fitting on the edge point sequences, calculate the parameters of the main curve equation, and obtain the weld centerline data; (4) Construct a vertical sampling vector for the weld centerline data, extract the grayscale profile curve in the vertical direction, calculate the peak position and width parameters of the profile curve, generate a transverse characteristic curve, and parameterize the transverse characteristic curve to obtain contour feature data; (5) Divide the area where the contour feature data is located into overlapping sub-blocks, calculate the grayscale co-occurrence matrix of the pixels in the sub-blocks, extract the angular second-order moment, correlation, entropy value, and inverse difference moment, generate local texture features, and spatially combine the local texture features to obtain the texture feature matrix; (6) Perform principal component analysis on the texture feature matrix to extract the main feature components, map the main feature components with the contour feature data in feature space, generate normalized weight coefficients through nonlinear weighted combination, fuse the features according to the normalized weight coefficients, and obtain the weld feature data.

[0037] Specifically, the weld expansion image is converted into a pixel matrix, which is a form that converts a digital image into mathematical operations. The pixel matrix is ​​a two-dimensional array in which each element represents the pixel value of the corresponding position in the image. For grayscale images, each pixel is usually represented by an integer between 0 and 255, where 0 represents black and 255 represents white. The principle of multi-scale analysis is that different features in the image are more obvious at different scales. For example, the overall outline of the weld is easier to identify at a larger scale, while small defects are more obvious at a smaller scale. In actual operation, a pyramid structure can be used for multi-scale block division, that is, starting from the smallest block, gradually merging adjacent blocks to form larger blocks until a predetermined maximum scale is reached. The pixel mean and variance of each scale block are calculated to describe the statistical characteristics of each block. The pixel mean reflects the overall brightness level of the block, while the variance reflects the discreteness of the pixel values ​​within the block. The grayscale gradient is the rate and direction of change of the grayscale value in the image, which usually has a larger value at the edge of the image. Extracting the first-order difference values ​​in the horizontal and vertical directions is a common method for calculating the grayscale gradient. The first-order difference in the horizontal direction can be obtained by calculating the grayscale value difference of adjacent pixels in the horizontal direction, and the same is true for the vertical direction. This process can be implemented using convolution operations, and commonly used operators include Sobel operators, Prewitt operators, etc. Edge feature data is obtained by bidirectional gradient weighted superposition, which combines the gradient information in the horizontal and vertical directions. The purpose of weighted superposition is to obtain more accurate edge strength and direction information. For example, the gradient amplitude can be used as a weight to weight the gradient vectors in the horizontal and vertical directions to obtain the final gradient vector for each pixel.

[0038] The purpose of locating the extreme points of edge feature data is to find the exact position of the edge. Extreme points refer to points where the gradient amplitude reaches the local maximum in the gradient direction. These points usually correspond to the edges in the image. The spatial continuity of local maximum points is calculated to distinguish between real edges and isolated extreme points caused by noise. Spatial continuity can be judged by checking whether the positions and directions of adjacent extreme points meet certain continuity conditions. According to the continuity criterion, valid edge points are screened and edge point sequences are generated, which further improves the accuracy of edge detection. The continuity criterion can include the continuity of the gradient direction, the similarity of the gradient amplitude, etc. Through these criteria, some isolated points that are unlikely to belong to the real edge can be removed, and those points that are more likely to constitute a continuous edge can be retained. The edge point sequence is curve fitted, the master curve equation parameters are calculated, and the weld centerline data is obtained. Curve fitting is to fit discrete edge points into a continuous curve, and this curve represents the centerline of the weld. The master curve equation parameters can be expressed by the following formula:

[0039] in, is the ordinate of the weld centerline, is the horizontal axis, ( ， ， ， ) are the polynomial coefficients, is the amplitude of the sine term, is the angular frequency, is the phase. This equation combines polynomials and sine functions, and can better describe the various shapes of the weld centerline, including straight segments, curved segments, and wavy shapes. Constructing a vertical sampling vector for the weld centerline data and extracting the grayscale profile curve in the vertical direction is to analyze the width and morphological characteristics of the weld. The vertical sampling vector refers to the sampling line extending vertically from the weld centerline to both sides. By extracting the grayscale values ​​along these sampling lines, the grayscale profile curve is obtained. The peak position and width parameters of the profile curve are calculated to generate the transverse characteristic curve. This process can be described by the following formula:

[0040] in, The distance from the weld centerline is The gray value at and are the amplitudes of the two Gaussian distributions, and is the center position of the two Gaussian distributions, and is the standard deviation of the two Gaussian distributions, is the background gray value. This model assumes that the profile of the weld can be described by the superposition of two Gaussian distributions, which can better fit the actual weld profile. The transverse characteristic curve is parameterized to obtain the contour feature data. Parametric description refers to the use of a set of parameters to represent the shape characteristics of the transverse characteristic curve, such as the peak position, peak width, peak height and other parameters of the curve. These parameters constitute the contour feature data, which intuitively reflect the characteristics of the weld width, height and shape. The area where the contour feature data is located is divided into overlapping sub-blocks for local texture analysis. The use of overlapping sub-blocks can ensure that the texture information of the boundary area is not missed. Calculating the gray level co-occurrence matrix of the pixels in the sub-block is a commonly used texture analysis method. The gray level co-occurrence matrix describes the distribution relationship of pixels of different gray levels in the image in spatial position. Extract the angular second-order moment, correlation, entropy value, and inverse difference moment to generate local texture features. These statistics are all calculated based on the gray level co-occurrence matrix, and they describe the texture characteristics of the image from different angles. The second-order angular moment reflects the uniformity of the image, the correlation reflects the linear dependence of the local grayscale, the entropy value reflects the complexity of the image, and the inverse difference moment reflects the local change of the image.

[0041] The local texture features are spatially combined to obtain a texture feature matrix. Spatial combination refers to combining the texture features of different sub-blocks according to their spatial position relationship in the original image to form a complete texture feature description. The texture feature matrix contains the texture information of the entire weld area and can reflect the subtle structural characteristics of the weld surface. The texture feature matrix is ​​subjected to principal component analysis to extract the main feature components. Principal component analysis (PCA) is a dimensionality reduction technique that can convert high-dimensional feature data into principal components of lower dimensions while retaining the main information of the data. Through PCA, the most representative feature components can be extracted from the texture feature matrix, which contain most of the variation information of the original data. The main feature components are mapped to the contour feature data in feature space, aiming to combine texture features and shape features to form a weld feature description. Feature space mapping can be understood as projecting different types of features into the same feature space so that they can be processed and analyzed uniformly. Through nonlinear weighted combination, normalized weight coefficients are generated. This process takes into account the importance of different features in weld quality assessment. Nonlinear weighting can better handle the complex relationship between features, while normalization ensures the scale consistency of different features. The features are fused according to the normalized weight coefficient to obtain the weld feature data. This fused feature data comprehensively reflects the characteristics of the weld shape, width, surface texture and other aspects, providing comprehensive data support for weld quality assessment.

[0042] In a specific embodiment, the process of executing step S104 may specifically include the following steps: (1) Convert the weld feature data into a numerical matrix, decompose the numerical matrix into multiple dimensions, combine the feature vectors according to their distribution characteristics in different dimensions, and generate feature classification data; (2) Group the feature classification data according to geometric features, calculate the statistical distribution parameters of each group of features, extract the spatial position relationship and morphological features, and generate defect candidate area data; (3) Locate the boundary of the defect candidate area data, extract the gray value distribution characteristics in the area, calculate the gray gradient change rate and directional parameters, generate the defect contour descriptor, perform shape analysis on the defect contour descriptor, and obtain the defect type data; (4) Perform spatial cluster analysis on defect type data, calculate defect density distribution and connectivity, extract topological structure characteristics of defect groups, generate defect distribution maps, calculate regional defect index based on defect distribution maps, and obtain defect assessment data; (5) Compare the defect assessment data with the rule set, calculate the weight coefficient of each assessment indicator, perform nonlinear fusion on the multi-dimensional assessment results, generate comprehensive scoring data, establish a quality grade mapping relationship based on the comprehensive scoring data, and obtain the grading result data; (6) Perform time series correlation analysis on the grading result data, extract the quality change trend characteristics, associate and map the quality trend characteristics with the defect assessment data, generate quality status descriptors, and construct weld quality data based on the quality status descriptors.

[0043] Specifically, converting weld feature data into a numerical matrix is ​​to convert various weld features extracted in the early stage into a form that can be used for mathematical operations. A numerical matrix is ​​a multidimensional array in which each row represents a sample point and each column represents a feature. For example, for a weld with a length of 1000 mm, if a sample point is taken every 1 mm and 10 features are extracted from each point, a 1000×10 numerical matrix will be obtained. Multidimensional decomposition of the numerical matrix is ​​to reduce the dimension of the data and extract the most representative features. Commonly used multidimensional decomposition methods include principal component analysis (PCA) and singular value decomposition (SVD). Taking PCA as an example, it projects the original data into a new feature space by calculating the eigenvalues ​​and eigenvectors of the covariance matrix. In this new feature space, the main changes in the data are retained, while the minor changes are ignored, thereby achieving data dimensionality reduction. Feature classification data is generated by combining the distribution characteristics of the eigenvectors in different dimensions. It is to recombine the reduced-dimensional data for subsequent classification and analysis. The combination method can be based on the projection value of the feature vector on each principal component or on the similarity between the feature vectors. Through this combination, similar features can be clustered together to form different feature categories. The feature classification data is grouped according to geometric features in order to identify areas where defects may exist. Geometric features include the width, height, and contour shape of the weld. By grouping these geometric features, areas with similar geometric characteristics can be classified into one category. The statistical distribution parameters of each group of features are calculated, including mean, variance, skewness, kurtosis, etc. These statistical parameters can reflect the overall distribution and discreteness of each group of features. The spatial position relationship and morphological features are extracted, taking into account the spatial distribution and morphological information of the weld features. The spatial position relationship can be described by calculating the distance and angle between feature points. Morphological features can be extracted through a series of morphological operations, such as expansion, corrosion, opening operation, closing operation, etc. These features can help identify abnormal areas in the weld, such as protrusions, depressions, etc. Through these steps, defect candidate area data is generated. The boundary of the defect candidate area data is located in order to accurately determine the boundary of the area where defects may exist. Boundary location can be achieved through edge detection algorithms, such as the Canny edge detector or the Sobel operator. The grayscale value distribution features within the region are extracted, including the grayscale histogram, mean, standard deviation, etc. of the pixels in the region. These features can reflect the distribution and change trend of the grayscale values ​​in the region.

[0044] The grayscale gradient change rate and directional parameters can be calculated using the following formula:

[0045]

[0046]

[0047] in, is the grayscale gradient amplitude, is the grayscale gradient direction, is the gray gradient change rate, is the image grayscale function, is the distance along the gradient direction. These parameters can reflect the rate and direction of change of grayscale values ​​in the region, which helps to identify the boundary and shape characteristics of defects. Generate defect contour descriptors, perform shape analysis on defect contour descriptors, and obtain defect type data. Defect contour descriptors can include parameters such as the perimeter, area, circularity, and rectangularity of the contour. Through the combination of these parameters, the shape of the defect can be quantitatively described, and then the type of defect, such as pores, cracks, slag inclusions, etc., can be inferred. Spatial clustering analysis of defect type data is performed to study the spatial distribution characteristics of defects. Commonly used spatial clustering algorithms include K-means, DBSCAN, etc. Through clustering analysis, defects that are similar in space can be grouped together to form a defect group. Calculate defect density distribution and connectivity, which can reflect the distribution and aggregation degree of defects in the weld. Extract the topological structure characteristics of the defect group and generate a defect distribution map. Topological structure characteristics include the shape, size, direction, etc. of the defect group. The defect distribution map is a visual representation method that intuitively shows the distribution of defects in the weld. The regional defect index is calculated according to the defect distribution map to obtain defect assessment data. The regional defect index can comprehensively consider the number, size, distribution and other factors of the defects to give a quantitative assessment result. The defect assessment data is compared with the rule set, which is a series of evaluation rules based on welding standards and expert experience. Through the comparison, the severity and impact range of each defect type can be determined. The weight coefficient of each evaluation index is calculated, considering the degree of influence of different types of defects on welding quality. The setting of the weight coefficient can be based on expert experience or automatically learned through machine learning methods. The multi-dimensional evaluation results are nonlinearly fused to generate comprehensive scoring data. Nonlinear fusion can better handle the complex relationship between different evaluation indicators. Commonly used methods include support vector machine (SVM), random forest, etc. A quality grade mapping relationship is established based on the comprehensive scoring data to obtain the grading result data. The quality grade mapping relationship maps continuous score values ​​to discrete quality grades, such as excellent, good, qualified, unqualified, etc. A time series correlation analysis is performed on the grading result data to extract the quality change trend characteristics. Time series correlation analysis takes into account the changes in welding quality over time and can be achieved through time series analysis methods, such as autoregressive models, moving average models, etc. The quality change trend characteristics can reflect the dynamic changes in quality during the welding process and help to detect potential problems in a timely manner. The quality trend characteristics are associated with the defect assessment data to generate a quality status descriptor. The quality status descriptor is a quality representation that not only includes the current quality status, but also the quality change trend. The weld quality data is constructed based on the quality status descriptor, which is the final quality assessment result and can be used to guide the control and optimization of the welding process.

[0048] For example: a pipe pile weld with a length of 5 meters is inspected. Assuming that a sample point is taken every 10 mm and 20 features are extracted from each point, the initial weld feature data is a 500×20 matrix. These 20 features include weld width, height, surface roughness, grayscale mean, grayscale variance, etc. This matrix is ​​input into the PCA algorithm. Assuming that the first 5 principal components explain 95% of the variance, these 5 principal components are retained and the original 500×20 matrix is ​​reduced to a 500×5 matrix. In this way, each sample point is represented by 5 main features. According to the distribution of these 5 features, the sample points can be divided into several categories, such as normal area, width abnormal area, height abnormal area, etc.

[0049] For each category, calculate the statistical distribution parameters. For example, for the width abnormal area, calculate its average width, width standard deviation, maximum width, minimum width, etc. At the same time, extract the spatial location information and morphological features of these areas, such as the length, position, shape, etc. of the area. This information constitutes the defect candidate area data. Perform a detailed analysis on each defect candidate area. Perform boundary positioning, and then calculate the grayscale gradient and directional parameters in the area. For example, for an area suspected of cracks, it may be found that the grayscale gradient is particularly large in a certain direction and the rate of change is also high. Based on these features, generate defect contour descriptors, such as the length, width, direction, etc. of the crack.

[0050] Perform spatial cluster analysis on defects in the entire weld. For example, it may be found that crack defects are particularly concentrated in a certain section of the weld. Calculate the defect density and connectivity of this area to generate a defect distribution map. Based on the map, calculate the regional defect index, such as if the defect density in this area is three times the average level of the entire weld. Compare these defect assessment data with preset rules. For example, the rules may stipulate that the length of a single crack shall not exceed 5mm, and the total length of cracks in every 100mm weld shall not exceed 10mm. Based on these rules, calculate the severity of each defect type. Then, perform a weighted summation of different types of defects to obtain a comprehensive score.

[0051] Assume that the comprehensive score range is 0-100, where 0-60 is unqualified, 61-75 is qualified, 76-90 is good, and 91-100 is excellent. Based on this mapping relationship, the quality grade of the weld can be obtained. Perform a time series analysis on the quality grade of the entire welding process. For example, it may be found that the quality grade has a downward trend in the middle and late stages of welding. Combine this trend information with specific defect data to generate a quality status descriptor, such as "Quality declines in the late stage of the welding process, mainly manifested in an increase in crack-type defects. It is recommended to check whether the welding speed and heat input are appropriate."

[0052] In a specific embodiment, the process of executing step S105 may specifically include the following steps: (1) Analyze the weld quality data into a feature matrix, perform time series decomposition on the feature matrix, calculate the changing rules of quality characteristics on different time scales, and generate quality trend data; (2) Perform multi-dimensional correlation analysis on quality trend data, extract the mapping relationship between quality indicators and process parameters, calculate parameter sensitivity distribution, and obtain parameter impact data; (3) Group the parameter influence data according to the process variable type, calculate the coupling coefficient of each group of parameters, build the correlation network between the parameters, generate the parameter constraints, perform boundary calculation based on the parameter constraints, and obtain the parameter adjustment range; (4) Discretize the parameter adjustment range, calculate the quality prediction value corresponding to each parameter point, generate a parameter response surface, search for the optimal point based on the parameter response surface, and obtain a parameter optimization solution; (5) Convert the parameter optimization scheme into a control instruction sequence, perform time encoding on the instruction sequence, generate welding parameter instructions, and calculate the adjustment amount of the process variable according to the welding parameter instructions; (6) Dynamically compensate the adjustment amount of the process variable, calculate the compensated parameter value, quantize and map the compensated parameter value, generate a control data stream, and output the target welding parameter according to the control data stream.

[0053] Specifically, the feature matrix is ​​a multidimensional array in which each row represents the quality state at a time point and each column represents a quality feature. For example, for a welding process lasting 5 minutes, if data is recorded once per second and 10 quality features are recorded each time, a 300×10 feature matrix will be obtained. The purpose of time series decomposition of the feature matrix is ​​to analyze the changing laws of quality features at different time scales. Time series decomposition methods include Fourier transform, wavelet transform, etc. Taking wavelet transform as an example, it can decompose the time series into components of different frequencies, thereby revealing the changing patterns of quality features at different time scales. For example, it may be found that some quality features show periodic fluctuations on a short time scale, while there is a trend of slow increase or decrease on a long time scale. Calculate the changing laws of quality features on different time scales to generate quality trend data. Further analysis and summary of the time series decomposition results. The changing laws can include trends, periodicity, mutation points, etc. For example, it may be found that the weld width shows a trend of slow increase throughout the welding process, while the surface roughness fluctuates briefly after each welding parameter adjustment. These changing laws constitute the quality trend data. Multi-dimensional correlation analysis of quality trend data aims to reveal the relationship between different quality indicators and the relationship between quality indicators and process parameters. Correlation analysis methods include Pearson correlation coefficient, Spearman rank correlation coefficient, etc. Through correlation analysis, it can be found which quality indicators have strong correlations and which process parameters have significant effects on quality indicators. The mapping relationship between quality indicators and process parameters is extracted to establish a quantitative relationship between quality indicators and process parameters. The mapping relationship can be established by regression analysis, neural network and other methods. For example, it may be found that the weld width is positively correlated with the welding current and negatively correlated with the welding speed. Calculate the parameter sensitivity distribution and obtain parameter influence data. Parameter sensitivity reflects the degree of influence of process parameter changes on quality indicators. Sensitivity analysis methods include local sensitivity analysis, global sensitivity analysis, etc. Through sensitivity analysis, it can be determined which process parameters have the greatest impact on welding quality, thereby providing guidance for parameter optimization.

[0054] Process variable types can include current parameters, voltage parameters, speed parameters, gas flow parameters, etc. By grouping, the mutual influence between parameters of the same type and the interaction between parameters of different types can be analyzed. The coupling coefficient of each group of parameters is calculated to build an association network between parameters. The coupling coefficient reflects the degree of mutual influence between parameters. The association network can intuitively display the relationship between parameters in a graphical way, where nodes represent parameters and edges represent the coupling relationship between parameters. Through this network, the complex interactions between parameters can be clearly seen.

[0055] Generate parameter constraints, which include physical limitations of parameters, process requirements, equipment capabilities, etc. For example, the welding current cannot exceed the maximum output capacity of the equipment, and the welding speed must be as high as possible while ensuring penetration. Perform boundary calculations based on the parameter constraints to obtain the parameter adjustment range. This process can be expressed by the following formula:

[0056] in, It is The adjustment range of each parameter is It is The value of the parameter, and are the lower and upper bounds of the parameter, respectively. It is Constraints, is the total number of parameters, is the total number of constraints. This formula describes the adjustable range of each parameter when all constraints are met. The parameter adjustment range is discretized to convert the continuous parameter space into a finite set of discrete points. The discretization method can be uniform division or non-uniform division based on the importance of the parameter. After discretization, it is possible to search in a limited combination of parameters, which greatly reduces the computational complexity of optimization. The quality prediction value corresponding to each parameter point is calculated to generate a parameter response surface. The quality prediction value can be calculated by the mapping relationship between the quality index and the process parameter established previously. The parameter response surface is a multidimensional space, in which the independent variables are the various process parameters and the dependent variable is the predicted quality index. This surface intuitively shows the relationship between the process parameters and the welding quality. The optimal point search is performed according to the parameter response surface to obtain the parameter optimization scheme. The optimal point search methods include gradient descent method, genetic algorithm, particle swarm optimization algorithm, etc. Through these algorithms, the parameter combination that makes the quality index reach the optimal value can be found on the parameter response surface. This parameter combination is the parameter optimization scheme. The parameter optimization scheme is converted into a control instruction sequence, and the optimization result is converted into an executable control command. The control instruction sequence includes the target values ​​of each process parameter and the process control instructions to achieve these target values. The instruction sequence is time-coded to generate welding parameter instructions. Time coding takes into account the time sequence and dynamic characteristics of parameter adjustment to ensure the smoothness and continuity of parameter adjustment. The adjustment amount of the process variable is calculated according to the welding parameter instruction. The adjustment amount refers to the difference between the current parameter value and the target parameter value. When calculating the adjustment amount, the dynamic characteristics of the parameters need to be considered. For example, some parameters may need to be adjusted slowly to avoid drastic fluctuations in the welding process.

[0057] Dynamic compensation is performed on the adjustment amount of the process variable, taking into account real-time feedback and environmental changes during the welding process. Dynamic compensation can fine-tune the original adjustment amount based on the real-time monitoring of the welding state. For example, if it is detected that the temperature rise of the weld is too fast, the increase in current may need to be appropriately reduced. The compensated parameter values ​​are calculated, quantized and mapped to generate a control data stream. Quantization mapping is the conversion of continuous parameter values ​​into discrete control signals that can be executed by the equipment. The control data stream is a series of timed control signals that are directly used to drive the various actuators of the welding equipment. The target welding parameters are output according to the control data stream, which is the last step of the entire parameter optimization and control process. The target welding parameters are the final execution parameters after optimization, encoding, compensation and quantization. They will be directly applied to the welding equipment to control the welding process.

[0058] For example, a welding process lasting 30 minutes is analyzed. Assuming that data is recorded every 5 seconds and 15 quality features are recorded each time, the initial weld quality data is a 360×15 feature matrix. These 15 features include weld width, height, surface roughness, penetration, porosity, etc. This matrix is ​​input into the wavelet transform algorithm to perform time series decomposition on each quality feature. For example, the analysis of weld width may find that: there are periodic fluctuations on the 1-minute scale, which may be related to the pulse frequency of the welding current; there is a slow increase trend on the 5-minute scale, which may be related to heat accumulation; there is a sudden change on the 15-minute scale, which may be related to the replacement of welding wire. These findings constitute quality trend data. Multi-dimensional correlation analysis is performed on quality trend data. By calculating the Pearson correlation coefficient, it may be found that the correlation coefficient between weld width and welding current is 0.85, and the correlation coefficient with welding speed is -0.72, which shows that welding current and welding speed are the two main factors affecting weld width. Meanwhile, it can be found that the correlation coefficient between surface roughness and gas flow rate is -0.63, which suggests that adjusting the gas flow rate may improve the surface quality.

[0059] Based on these correlations, a mapping relationship between quality indicators and process parameters is established. For example, a multivariate linear regression model is established for weld width: Weld width = 5.2 + 0.03 × welding current - 0.5 × welding speed + 0.1 × arc voltage Through sensitivity analysis, calculate the degree of influence of each parameter on the quality index. For example, it may be found that the welding current has the highest sensitivity, followed by the welding speed, and then the arc voltage. These sensitivities constitute the parameter influence data. Divide the parameters into current parameter group, speed parameter group and voltage parameter group, and calculate the coupling coefficient of the parameters in the group. For example, it may be found that there is a strong coupling relationship between the pulse current and the base current, which need to be adjusted at the same time. Construct a parameter association network, and it may be found that there is a strong interaction between the current parameter group and the speed parameter group, while the interaction with the voltage parameter group is weak. Generate parameter constraints based on equipment capabilities and process requirements. For example: 100A≤Welding current≤300A;0.5m / min≤Welding speed≤2m / min; 18V≤arc voltage≤30V; welding heat input = welding current×arc voltage / welding speed≤3kJ / mm. According to these constraints, the parameter adjustment range is calculated. Discretize the parameter adjustment range. For example, discretize the welding current with a step size of 10A, the welding speed with a step size of 0.1m / min, and the arc voltage with a step size of 1V. For each discrete point, use the previously established mapping relationship to predict the quality index and generate a parameter response surface. Use a genetic algorithm to search for the optimal point on the parameter response surface. Assume that the optimal parameter combination obtained after the algorithm converges is: welding current 220A, welding speed 1.2m / min, arc voltage 26V. This is the parameter optimization solution. Convert this solution into a control instruction sequence. Considering that the welding current cannot change suddenly, design a slow increase process: starting from the current value of 200A, increase by 2A per second until the target value of 220A is reached. Similarly, the welding speed starts from 1.0m / min and increases by 0.02m / min per second, and the arc voltage starts from 24V and increases by 0.2V per second.

[0060] In the process of executing these instructions, the welding status is monitored in real time. If it is detected that the temperature of the weld rises too fast, the current increase is dynamically compensated, reducing the increase per second from 2A to 1.5A. Finally, the compensated parameter values ​​are converted into specific control signals, such as the duty cycle of the PWM wave, the pulse frequency of the stepper motor, etc., to form a control data stream.

[0061] In a specific embodiment, the process of executing step S106 may specifically include the following steps: (1) Convert the weld quality data and target welding parameters into a time series data stream, time-stamp the time series data stream, divide the data into blocks according to the timestamp, and generate data record blocks; (2) Construct a multidimensional index for the data record blocks, calculate the correlation between the data blocks, extract the temporal and spatial features of the data blocks, and obtain feature sequence data; (3) Classify the characteristic sequence data according to the quality indicators, calculate the statistical characteristic values ​​of each indicator, normalize the statistical characteristic values, generate the indicator distribution characteristics, construct the quality assessment matrix based on the indicator distribution characteristics, and obtain the quality statistical data; (4) Divide the quality statistical data into time windows, calculate the quality change trend within the window, extract the key quality inflection points, generate trend analysis data, perform quality prediction based on the trend analysis data, and obtain the quality prediction results; (5) Compare and analyze the quality prediction results with historical data, calculate the quality deviation value, extract the quality abnormality interval, generate quality diagnosis data, perform causal correlation analysis based on the quality diagnosis data, and obtain the cause analysis results; (6) Structuralize the cause analysis results, extract key quality information, convert the quality information into descriptive text, generate quality description data, and output a quality report based on the quality description data.

[0062] Specifically, the weld quality data and target welding parameters are converted into a time series data stream, and the discrete data points are organized in time order to form a continuous data stream. The time series data stream is a special data structure that contains not only the value of the data, but also the time information of the data generation. In the process of pipe pile welding, the weld quality data may include indicators such as weld width, depth, surface roughness, etc., while the target welding parameters include current, voltage, welding speed, etc. The time series data stream is time-stamped to give each data point an accurate timestamp. The accuracy of the timestamp depends on the frequency of data acquisition, which can usually be accurate to milliseconds. The purpose of time marking is to provide a basis for subsequent time series analysis and data synchronization. Data is divided into blocks according to the timestamp to generate data record blocks. The purpose of data block division is to divide the continuous data stream into several relatively independent units to facilitate subsequent parallel processing and local analysis. The size of the data record block can be set according to actual needs, for example, the data per minute can be used as a block, or each completed welding cycle can be used as a block.

[0063] Multidimensional indexing is performed on data record blocks to improve data retrieval efficiency. Multidimensional indexing is an efficient way of organizing data, which allows data to be quickly located and accessed from multiple dimensions. In the scenario of pipe pile welding, the dimensions of the index can include time, welding position, quality indicator type, etc. Commonly used multidimensional index structures include R-tree, KD-tree, etc. After constructing a multidimensional index, complex queries such as "find all records with abnormal weld width within a certain time period" can be quickly performed. The correlation between data blocks is calculated to find the connection between different data blocks. The calculation methods of correlation include correlation coefficient, mutual information, etc. By calculating the correlation, long-term trends and periodic changes in the welding process can be found. For example, it may be found that several data blocks after each replacement of welding wire have a high correlation, which may reflect the impact of welding wire replacement on welding quality. The temporal and spatial features of the data blocks are extracted to obtain feature sequence data. The temporal features reflect the law of data change over time, including trends, periodicity, mutation points, etc. The spatial features reflect the distribution of data at different positions of the weld. Feature extraction methods include Fourier transform, wavelet analysis, principal component analysis, etc. Through feature extraction, the original high-dimensional data can be compressed into a more compact and meaningful form.

[0064] The characteristic sequence data are classified according to quality indicators in order to analyze different types of quality indicators separately. Quality indicators may include geometric characteristics (such as weld width and depth), mechanical properties (such as tensile strength and hardness), surface quality (such as roughness and porosity), etc. Analyzing each type of indicator separately can more accurately evaluate different aspects of welding quality. Calculate the statistical characteristic values ​​of various indicators, including mean, variance, skewness, kurtosis, etc. These statistical characteristic values ​​fully describe the distribution of the data. For example, the mean reflects the overall level of the indicator, the variance reflects the degree of dispersion of the data, the skewness reflects the symmetry of the data distribution, and the kurtosis reflects the degree of sharpness of the data distribution.

[0065] The statistical characteristic values ​​are normalized to generate indicator distribution characteristics. The purpose of normalization is to convert indicators of different dimensions to the same scale for easy comparison and comprehensive analysis. Common normalization methods include minimum-maximum normalization, Z-score standardization, etc. The normalized indicator distribution characteristics can intuitively reflect the relative level and fluctuation of each quality indicator. A quality assessment matrix is ​​constructed based on the indicator distribution characteristics to obtain quality statistical data. The quality assessment matrix is ​​a two-dimensional table in which rows represent different quality indicators and columns represent different statistical characteristics. Through this matrix, all aspects of welding quality can be fully displayed. Quality statistical data is a further processing and refinement of the quality assessment matrix, which may include comprehensive quality index, qualified rate of various indicators, etc.

[0066] The quality statistical data is divided into time windows in order to analyze the quality change trend over time. The size of the time window can be set according to actual needs, for example, it can be in hours, days or weeks. In each time window, the quality change trend within the window is calculated. Trend calculation methods include linear regression, exponential smoothing, etc. In this way, the rising, falling or stable stages of quality can be identified. Extract key quality inflection points and generate trend analysis data. The key quality inflection point refers to the moment when the quality trend changes significantly, such as from rising to falling, from stable to fluctuating, etc. The identification of inflection points can be achieved through curve fitting and derivative analysis. Trend analysis data includes the time of each inflection point, the quality level, and the amplitude and speed of trend change.

[0067] Quality prediction is performed based on trend analysis data to obtain quality prediction results. Quality prediction methods include time series analysis, machine learning algorithms, etc. The prediction results include not only the quality level in the future, but also the confidence interval of the prediction. These prediction results can be used to warn of possible quality problems in advance and guide the formulation of production plans. The quality prediction results are compared and analyzed with historical data to calculate the quality deviation value. The quality deviation value reflects the gap between the actual quality and the expected quality. Through comparative analysis, the accuracy of the prediction can be evaluated, and abnormal conditions in the welding process can also be discovered.

[0068] Extract the quality anomaly interval and generate quality diagnostic data. The quality anomaly interval refers to the time period when the quality index exceeds the normal range. The judgment of anomalies can be based on statistical methods (such as the 3σ principle) or rules formulated by experts. Quality diagnostic data includes information such as the type, duration, and severity of the anomaly. Perform causal association analysis based on the quality diagnostic data to obtain the cause analysis results. Causal association analysis aims to find the root cause of quality anomalies. Analysis methods include correlation analysis, decision tree analysis, Bayesian network, etc. The cause analysis results include not only possible causes, but also the confidence level and related evidence of each cause.

[0069] The cause analysis results are structured and key quality information is extracted. The purpose of structured organization is to transform complex analysis results into a clear and organized form. Key quality information includes the most significant quality problems, the most likely causes, the most effective improvement measures, etc. Convert quality information into descriptive text and generate quality description data. Convert the digitized analysis results into human-readable text descriptions. The descriptive text should be concise and highlight the key points, while taking into account professionalism and readability. Quality description data may include overall quality evaluation, description of major problems, cause analysis, improvement suggestions, etc. Output a quality report based on the quality description data. The quality report is the final output of the entire analysis process. It should comprehensively and systematically summarize the welding quality status and provide a basis for management decisions and process improvements. The content of the report should include multiple parts such as data analysis results, chart presentation, text description and suggestions.

[0070] For example, suppose that a welding process lasting 8 hours is analyzed. During the welding process, data is recorded once a second, including 10 quality indicators (such as weld width, depth, surface roughness, etc.) and 5 welding parameters (such as current, voltage, welding speed, etc.). In this way, a time series data stream containing 28,800 rows and 15 columns is obtained. This data stream is time-stamped, and each data point has a timestamp accurate to milliseconds. Then, the data is divided into blocks of 5 minutes to obtain 96 data record blocks. A multidimensional index is constructed for these 96 data blocks. The index dimensions include time, welding position (assuming that the welding process is carried out along the circumference of the pile, it can be expressed by angle) and quality indicator type.

[0071] By calculating the correlation between data blocks, it may be found that the correlation between adjacent blocks is high, while the correlation between blocks that are far apart is low. This may reflect the existence of short-term quality fluctuations in the welding process. Extract the temporal and spatial characteristics of each data block. For example, it may be found that the weld width shows periodic changes in time and gradient changes in space (gradually widening from one end of the pipe pile to the other end). Classify the characteristic sequence data according to quality indicators, and calculate the statistical characteristic values ​​of each type of indicator. For example, for the weld width, the average value is calculated to be 10mm, the standard deviation is 0.5mm, the skewness is 0.1, and the kurtosis is 3. These statistical values ​​are normalized to obtain the indicator distribution characteristics. Construct a quality assessment matrix, where each row of the matrix represents a quality indicator and each column represents a statistical feature.

[0072] The quality statistics data are divided into time windows, with 1 hour as a window. The quality change trend is calculated in each window. It may be found that the weld width has a significant upward trend between the 3rd and 4th hours. This time point is marked as a critical quality inflection point. Based on these trend data, time series analysis methods (such as ARIMA model) are used to predict the quality and predict the change trend of the weld width in the next 2 hours. The predicted results are compared with the actual data to calculate the quality deviation value. Assume that the predicted average weld width is 0.3mm higher than the actual value, and this deviation is recorded. Extract the quality abnormal interval, for example, it is found that the weld depth is lower than the standard value within 15 minutes of the 6th hour. This abnormal interval is marked and quality diagnostic data is generated.

[0073] A causal analysis of this anomaly may reveal that the welding current dropped significantly during these 15 minutes. Further analysis determined that the current drop was caused by power supply fluctuations. This analysis result was structured and key information was extracted: the weld depth abnormality was related to power supply fluctuations. This information was converted into descriptive text: "In the 15 minutes of the 6th hour, the weld depth was lower than the standard value. The cause analysis showed that this may be related to the power supply fluctuations observed during the same period. It is recommended to check and stabilize the power supply and add an uninterruptible power supply device if necessary." Based on all the analysis results and descriptive text, a complete quality report is generated. The report includes overall quality assessment, trend of changes in key quality indicators, abnormal conditions and cause analysis, prediction results and suggestions, etc.

[0074] In a specific embodiment, the process of executing the step of structuring the cause analysis results may specifically include the following steps: (1) Perform hierarchical analysis on the cause analysis results, classify the data items, group them according to quality indicators, process parameters, and defect characteristics, and generate a structured data table; (2) Perform correlation calculation on the structured data table, extract the dependency relationship between data items, establish a hierarchical structure diagram of data items, and obtain the data organization structure; (3) Divide the numerical data in the data organization structure into intervals, calculate the numerical distribution characteristics, mark abnormal data points, generate data statistical results, and extract key quality information based on the data statistical results; (4) Perform semantic mapping on key quality information, convert numerical features into descriptive sentences, generate semantic templates, construct text structures based on the semantic templates, and obtain quality description texts; (5) Divide the quality description text into chapters according to quality assessment, abnormality analysis, and improvement suggestions, integrate the contents of each chapter, generate a report framework, fill in the specific content according to the report framework, and obtain quality description data; (6) Format the quality description data, generate chart indexes, associate and map text content with charts, build a quality report structure, and output a quality report.

[0075] Specifically, the cause analysis results are analyzed in layers, and the complex analysis results are broken down into a hierarchical structure that is easier to understand and process. The process of layered analysis includes identifying the main problems, determining the root causes of the problems, and analyzing the relationship between the causes. In the scenario of pipe pile welding, the main problems include unstable weld quality and fluctuations in welding parameters, while the root causes involve equipment failure, operating errors, material problems, and other aspects. The data items are labeled by category in order to classify different types of data for subsequent analysis and processing. The methods of category labeling include rule-based labeling and automatic labeling based on machine learning. In the scenario of pipe pile welding, the data items are grouped according to quality indicators, process parameters, and defect characteristics. Quality indicators include weld width, depth, strength, etc.; process parameters include welding current, voltage, speed, etc.; defect characteristics include pores, cracks, slag inclusions, etc. Through this grouping, the relationship and influence between different types of data can be clearly seen. Generate a structured data table to organize the results of layered analysis and category labeling into an ordered table form. Each row of the structured data table represents a specific problem or phenomenon, and each column represents an attribute or feature. For example, the columns of the table include "problem description", "relevant quality indicators", "affected process parameters", "observed defect characteristics", "possible causes", "severity", etc. The correlation calculation of the structured data table aims to discover the intrinsic connection between the data items. The methods of correlation calculation include correlation coefficient analysis, principal component analysis, factor analysis, etc. Through these analyses, it is revealed which quality indicators have strong correlations, which process parameters have significant effects on specific defect characteristics, etc. For example, it is found that there is a strong positive correlation between welding current and weld width, while there is a negative correlation between welding speed and weld depth. The dependency relationship between data items is extracted, and a hierarchical structure diagram of data items is established to obtain the data organization structure. The extraction of dependency relationships is achieved through methods such as causal analysis and Bayesian networks. The hierarchical structure diagram is a tree-like or mesh-like graphical representation that shows the hierarchical relationship and influence path between data items. In this diagram, the top node is the final quality indicator, the middle layer is the various factors affecting the quality, and the bottom layer is the specific process parameters and operation details. Through this structure diagram, the mechanism and importance of each factor can be clearly seen. The numerical data in the data organization structure is divided into intervals to convert continuous numerical data into discrete intervals for easy statistics and analysis. Interval division methods include equal width division, equal frequency division, cluster division, etc. Selecting a suitable division method requires considering the distribution characteristics of the data and the actual application requirements. For example, for welding current, it is divided into three intervals: "low current", "normal current" and "high current". Calculate the numerical distribution characteristics, including statistics such as mean, variance, skewness, and kurtosis. These characteristics reflect the central tendency, degree of dispersion, symmetry, and sharpness of the data. Through these characteristics, the overall distribution of the data can be quickly understood.For example, the mean of the weld width reflects the overall level, the variance reflects the stability of the width, and the skewness reflects the tilt direction of the width distribution. Mark abnormal data points to identify those data that deviate significantly from the normal range. The identification methods of abnormal data points include statistical methods (such as the 3σ principle) and machine learning methods (such as the isolation forest algorithm). In the scenario of pipe pile welding, abnormal data points represent emergencies or potential problems in the welding process and require special attention. Generate data statistical results and extract key quality information based on the data statistical results. The data statistical results include descriptive statistics of various indicators, the distribution of abnormal points, and the correlation between various parameters. Key quality information is the most important and valuable information extracted from these statistical results, which is directly related to the evaluation and improvement of welding quality. For example, key quality information includes "the weld width fluctuates greatly, mainly concentrated in the middle and late stages of the welding process", "the welding speed in the high current range is significantly negatively correlated with the weld depth", etc.

[0076] Semantic mapping is performed on key quality information to convert numerical features into descriptive statements. Semantic mapping is a process of converting quantitative information into qualitative descriptions, which needs to consider the actual meaning and application scenarios of the data. For example, a numerical feature such as "the mean value of weld width is 10 mm and the standard deviation is 0.5 mm" is converted into a descriptive statement such as "the weld width is generally stable, but there are small fluctuations." Generate semantic templates and create a series of standardized sentence structures to describe various quality situations. Semantic templates include "[Quality indicator] overall [evaluation word], but [problem description]", "[Process parameter] and [quality indicator] are in a [relationship type] relationship, and [improvement measures] are recommended", etc. These templates are automatically filled in according to actual data to generate natural and fluent descriptions. Construct a text structure based on the semantic template to obtain a quality description text. The construction of the text structure needs to consider factors such as logical order, importance ranking, and causal relationship. The quality description text includes overall quality assessment, description of major problems, cause analysis, and improvement suggestions. In this structured way, the integrity and order of the quality report are ensured. The quality description text is divided into chapters according to quality assessment, abnormal analysis, and improvement suggestions. In order to make the report structure clearer and facilitate readers to quickly locate the required information. The quality assessment section summarizes the overall welding quality status; the abnormal analysis section describes the problems found and their possible causes in detail; the improvement suggestion section proposes targeted solutions. Integrate the contents of each chapter to generate a report framework. The report framework is the skeleton of the quality report, which determines the overall structure and information organization of the report. Fill in specific content according to the report framework to obtain quality description data. Fill the previously generated quality description text, data statistics, key quality information, etc. into the corresponding positions in the report framework. The filling process needs to pay attention to the coherence and logic of the content to ensure smooth connection between the various parts. Format the quality description data to improve the readability and aesthetics of the report. Formatting includes setting fonts, font sizes, paragraph spacing, page margins, etc., and also includes adding headers, footers, directories, chapter titles and other elements. Good typesetting makes the report more professional and easy to read. Generate a chart index to facilitate readers to quickly find charts in the report. Chart indexes usually include chart numbers, titles and page numbers. In the quality report of pipe pile welding, the charts include weld quality trend chart, parameter correlation heat map, defect distribution map, etc. The text content is associated with the chart to ensure the consistency and complementarity between the text description and the chart display. The association mapping is achieved by inserting chart references in the text or citing relevant text paragraphs in the chart description. This association helps readers better understand the data and analysis results. Construct the quality report structure and output the quality report.

[0077] The above describes the detection method for intelligent welding of pipe pile joints in the embodiment of the present application. The following describes the detection system for intelligent welding of pipe pile joints in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the detection system for intelligent welding of pipe piles includes: The acquisition module is used to acquire images of the pipe pile joints to obtain the original weld image, and calibrate the original weld image according to the illumination parameters to obtain the target weld image; The denoising module is used to denoise the target weld image through a filtering algorithm, register and splice the image through feature matching, and obtain the weld expansion image; A segmentation module is used to segment the weld seam unfolded image using a deep learning network, extract the weld seam centerline and surface features, and obtain weld seam feature data; The recognition module is used to detect and identify defects in weld feature data, and obtain weld quality data by performing hierarchical calculations through a set of rules; A mapping module is used to perform parameter mapping based on weld quality data, generate welding parameter instructions, adjust welding process variables, and obtain target welding parameters; The storage module is used to store weld quality data and target welding parameters and generate a quality report through data analysis.

[0078] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein instructions are stored in the computer-readable storage medium, and when the instructions are executed on a computer, the computer executes the steps of the detection method for intelligent welding of pipe piles.

[0079] Based on the same technical concept, the embodiment of the present application also provides an electronic device. Figure 3 As shown, it is a schematic diagram of the structure of the electronic device 300 provided in the embodiment of the present application, including a processor 301, a memory 302, and a bus 303. Among them, the memory 302 is used to store execution instructions, including a memory 3021 and an external memory 3022; the memory 3021 here is also called an internal memory, which is used to temporarily store the operation data in the processor 301, and the data exchanged with the external memory 3022 such as a hard disk. The processor 301 exchanges data with the external memory 3022 through the memory 3021. When the electronic device 300 is running, the processor 301 communicates with the memory 302 through the bus 303.

Claims

1. A detection method for intelligent welding of pipe piles, characterized in that: The detection method for intelligent welding of pipe pile joints comprises: Capturing images of pipe pile joints to obtain an original weld image, and calibrating the original weld image according to illumination parameters to obtain a target weld image; Denoising the target weld image by using a filtering algorithm, registering and splicing the image by feature matching, and obtaining a weld expansion image; Using a deep learning network to segment the weld unfolded image, extract the weld centerline and surface features, and obtain weld feature data; Perform defect detection and identification on the weld feature data, perform hierarchical calculations through a set of rules, and obtain weld quality data; Perform parameter mapping according to the weld quality data, generate welding parameter instructions, adjust welding process variables, and obtain target welding parameters; The weld quality data and target welding parameters are stored, and a quality report is generated through data analysis.

2. The detection method for intelligent welding of pipe pile joints according to claim 1 is characterized in that: The method of collecting images of the pipe pile joints to obtain an original weld image, and calibrating the original weld image according to illumination parameters to obtain a target weld image includes: The industrial camera is used to collect circumferential distribution data at the pipe pile joints, and the collected data is converted into a digital signal stream to obtain multiple groups of original weld images; Time stamping the multiple groups of original weld images, integrating and sorting the data streams according to the time sequence, and obtaining serialized original weld images through data transmission; Extracting image grayscale value distribution from the serialized original weld image, calculating the mean and variance of the light intensity of the image area, and generating real-time illumination parameters; The real-time illumination parameter is calculated to be different from the preset standard illumination value to generate a compensation coefficient, and the pixel value of the original weld image is linearly transformed according to the compensation coefficient to obtain a brightness-corrected image; Separating the brightness-corrected image into RGB channels, calculating the mean ratio of pixels in each channel, generating a color calibration parameter matrix, and performing channel mapping on pixel values ​​according to the color calibration parameter matrix to obtain color calibration data; The color calibration data is used to reconstruct an image matrix according to pixel positions, the contrast and sharpness of the image matrix are numerically adjusted to generate an enhancement coefficient, and the image is remapped at the pixel level according to the enhancement coefficient to obtain a target weld image.

3. The detection method for intelligent welding of pipe pile joints according to claim 1 is characterized in that: The target weld image is denoised by a filtering algorithm, and the image is registered and spliced ​​by feature matching to obtain a weld expansion image, including: Dividing the target weld image into overlapping sub-regions, performing pixel value statistics on the sub-regions, calculating regional noise intensity distribution, and generating noise feature data; According to the noise characteristic data, the filter window size is set, the target weld image is scanned by a sliding window, and the weight coefficient of the pixel value in the window is calculated to obtain the noise reduction processing data; Divide the noise reduction processed data into image blocks, extract gradient feature points of the image blocks, establish feature point description vectors, and generate feature matching data; Calculating the spatial distance of the feature points in the feature matching data, eliminating unmatched point pairs, establishing a position correspondence between image blocks, and obtaining image registration data; Reconstructing the image registration data according to the spatial position relationship, calculating the pixel difference of the overlapping area of ​​adjacent image blocks, generating image fusion weights, and obtaining spliced ​​data; The stitching data is subjected to coordinate transformation and resampling, and the image edge is subjected to smooth transition processing to obtain a weld expansion image.

4. The detection method for intelligent welding of pipe pile joints according to claim 1 is characterized in that: The method of using a deep learning network to segment the weld seam unfolded image, extracting the weld seam centerline and surface features, and obtaining weld seam feature data includes: Converting the weld expansion image into a pixel matrix, dividing the pixel matrix into multi-scale blocks, calculating the pixel mean and variance of each scale block, and generating image segmentation area data; Calculating the grayscale gradient of the image segmentation area data, extracting the first-order difference values ​​in the horizontal and vertical directions, and obtaining edge feature data through bidirectional gradient weighted superposition; Positioning the extreme points of the edge feature data, calculating the spatial continuity of the local maximum points, screening the effective edge points according to the continuity criterion, generating an edge point sequence, performing curve fitting on the edge point sequence, calculating the parameters of the main curve equation, and obtaining the weld centerline data; Constructing a vertical sampling vector for the weld centerline data, extracting a grayscale profile curve in the vertical direction, calculating the peak position and width parameters of the profile curve, generating a transverse characteristic curve, and performing parameterized description on the transverse characteristic curve to obtain contour feature data; Divide the area where the contour feature data is located into overlapping sub-blocks, calculate the gray level co-occurrence matrix of the pixels in the sub-blocks, extract the angular second-order moment, correlation, entropy value, and inverse difference moment, generate local texture features, and spatially combine the local texture features to obtain a texture feature matrix; The principal component analysis is performed on the texture feature matrix to extract the main feature components, the main feature components and the contour feature data are mapped into feature space, a normalized weight coefficient is generated through nonlinear weighted combination, the features are fused according to the normalized weight coefficient, and the weld feature data is obtained.

5. The detection method for intelligent welding of pipe pile joints according to claim 1 is characterized in that: The defect detection and identification of the weld feature data is performed, and the weld quality data is obtained by performing hierarchical calculation through a set of rules, including: Converting the weld feature data into a numerical matrix, performing multi-dimensional decomposition on the numerical matrix, combining the feature vectors according to distribution characteristics in different dimensions, and generating feature classification data; The feature classification data are grouped according to geometric features, statistical distribution parameters of each group of features are calculated, spatial position relationships and morphological features are extracted, and defect candidate area data are generated; The defect candidate area data is positioned at the boundary, the gray value distribution characteristics in the area are extracted, the gray gradient change rate and the directional parameter are calculated, a defect contour descriptor is generated, and the defect contour descriptor is analyzed in shape to obtain defect type data; Performing spatial cluster analysis on the defect type data, calculating defect density distribution and connectivity, extracting topological structure characteristics of defect groups, generating defect distribution maps, calculating regional defect indexes based on the defect distribution maps, and obtaining defect assessment data; Compare the defect assessment data with the rule set, calculate the weight coefficient of each assessment indicator, perform nonlinear fusion on the multi-dimensional assessment results, generate comprehensive scoring data, establish a quality grade mapping relationship based on the comprehensive scoring data, and obtain grading result data; A time series correlation analysis is performed on the grading result data to extract quality change trend characteristics, the quality trend characteristics are correlated and mapped with the defect assessment data to generate a quality status descriptor, and weld quality data is constructed according to the quality status descriptor.

6. The detection method for intelligent welding of pipe pile joints according to claim 1, characterized in that: The method of performing parameter mapping according to the weld quality data, generating welding parameter instructions, adjusting welding process variables, and obtaining target welding parameters includes: Parsing the weld quality data into a feature matrix, performing time series decomposition on the feature matrix, calculating the changing rules of quality features on different time scales, and generating quality trend data; Performing multi-dimensional correlation analysis on the quality trend data, extracting the mapping relationship between quality indicators and process parameters, calculating parameter sensitivity distribution, and obtaining parameter impact data; The parameter influence data are grouped according to the process variable type, the coupling coefficient of each group of parameters is calculated, an association network between the parameters is constructed, parameter constraints are generated, and boundary calculations are performed according to the parameter constraints to obtain a parameter adjustment range; Discretizing the parameter adjustment range, calculating the quality prediction value corresponding to each parameter point, generating a parameter response surface, searching for the optimal point based on the parameter response surface, and obtaining a parameter optimization solution; Converting the parameter optimization scheme into a control instruction sequence, performing time-series coding on the instruction sequence, generating a welding parameter instruction, and calculating the adjustment amount of the process variable according to the welding parameter instruction; Dynamically compensate the adjustment amount of the process variable, calculate the compensated parameter value, quantize and map the compensated parameter value, generate a control data stream, and output the target welding parameter according to the control data stream.

7. The detection method for intelligent welding of pipe pile joints according to claim 1 is characterized in that: The storing of the weld quality data and target welding parameters and generating a quality report through data analysis include: Converting the weld quality data and the target welding parameters into a time series data stream, time-stamping the time series data stream, dividing the data into blocks according to the timestamps, and generating data record blocks; Constructing a multi-dimensional index for the data record blocks, calculating the correlation between the data blocks, extracting the temporal characteristics and spatial characteristics of the data blocks, and obtaining characteristic sequence data; Classifying the characteristic sequence data according to quality indicators, calculating statistical characteristic values ​​of various indicators, normalizing the statistical characteristic values, generating indicator distribution characteristics, constructing a quality assessment matrix according to the indicator distribution characteristics, and obtaining quality statistical data; Divide the quality statistical data into time windows, calculate the quality change trend within the window, extract key quality inflection points, generate trend analysis data, perform quality prediction based on the trend analysis data, and obtain quality prediction results; Compare and analyze the quality prediction results with historical data, calculate the quality deviation value, extract the quality abnormality interval, generate quality diagnosis data, perform causal association analysis based on the quality diagnosis data, and obtain cause analysis results; The cause analysis results are structured, key quality information is extracted, the quality information is converted into descriptive text, quality description data is generated, and a quality report is output based on the quality description data, including: hierarchical analysis of the cause analysis results, category labeling of data items, grouping according to quality indicators, process parameters, and defect characteristics, and generating a structured data table; performing correlation calculations on the structured data table, extracting dependencies between data items, establishing a hierarchical structure diagram of data items, and obtaining a data organization structure; dividing the numerical data in the data organization structure into intervals, calculating numerical distribution characteristics, and marking abnormal data points. Generate data statistical results, extract key quality information based on the data statistical results; perform semantic mapping on the key quality information, convert numerical features into descriptive statements, generate semantic templates, build a text structure based on the semantic templates, and obtain a quality description text; divide the quality description text into chapters according to quality assessment, abnormality analysis, and improvement suggestions, integrate the contents of each chapter, generate a report framework, fill in specific content according to the report framework, and obtain quality description data; format the quality description data, generate a chart index, associate and map the text content with the chart, build a quality report structure, and output a quality report.

8. A computer device, characterized in that: include: A processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor and the memory communicate via the bus, and when the machine-readable instructions are executed by the processor, the steps of the detection method for intelligent welding of pipe piles are performed as described in any one of claims 1 to 7.

9. A detection system for intelligent welding of pipe piles, used to implement the detection method for intelligent welding of pipe piles as described in any one of claims 1 to 7, characterized in that: The detection system for intelligent welding of pipe piles comprises: The acquisition module is used to acquire images of the pipe pile joints to obtain an original weld image, and calibrate the original weld image according to the illumination parameters to obtain a target weld image; A noise reduction module is used to reduce the noise of the target weld image by a filtering algorithm, and to register and splice the image by feature matching to obtain a weld expansion image; A segmentation module is used to segment the weld seam expansion image using a deep learning network, extract the weld seam centerline and surface features, and obtain weld seam feature data; An identification module is used to detect and identify defects in the weld feature data, and to perform hierarchical calculations through a set of rules to obtain weld quality data; A mapping module, used to perform parameter mapping according to the weld quality data, generate welding parameter instructions, adjust welding process variables, and obtain target welding parameters; The storage module is used to store the weld quality data and target welding parameters and generate a quality report through data analysis.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, a detection method for intelligent welding of pipe piles as described in any one of claims 1 to 7 is implemented.

Images