A method and system for intelligent real-time detection of PCCP welding quality

Abstract

The application provides a PCCP welding quality intelligent real-time detection method and system, relates to the field of pipe welding quality online detection and intelligent evaluation technology by using machine learning, and collects light energy data and multi-light source images of a spiral weld pool, dynamically adjusts exposure parameters by using energy differences in a visible light near-infrared wave band, suppresses strong light interference, and generates a weld surface image; a stress concentration area is located by scanning a weld thermal deformation area, combining speckle pattern changes, sound frequency changes and thin-walled steel cylinder elastic characteristics; the surface image and deformation data are input into a space-time convolutional neural network, a weld space structure diagram is constructed by fusing light energy changes, image details and spatial features, and a sensor position is adaptively corrected; the recess depth reconstructed by three-dimensional point clouds is compared, the recess degree and stress correlation are analyzed, a probability thermal diagram is generated to output a pressure failure risk level, and the probability early warning of the pressure failure risk can be realized.

Description

Technical Field

[0001] This application relates to the field of online detection and intelligent evaluation of pipeline welding quality using machine learning, and in particular to a method and system for intelligent real-time detection of PCCP welding quality. Background Technology

[0002] During the manufacturing process of spiral welds for thin-walled PCCP steel cylinders, the strong welding arc light severely interferes with image acquisition of the molten pool area, causing the penetration depth characteristics to be obscured. Therefore, there is an urgent need for a technology that can simultaneously detect penetration depth and thermal stress cracks in real time, requiring strong anti-interference capabilities, collaborative analysis of multiple physical quantities, and millisecond-level real-time response.

[0003] Currently, representative solutions employ a combination of pre- and post-injection phase image comparison and a multi-branch AI model. By acquiring weld images of the same location at the steel cylinder weld before and after injection, and inputting them into a comprehensive model that includes a leakage identification network, a weld identification network, and an image residual comparison network, the leakage location is marked and the risk level is assessed, claiming near-zero detection accuracy.

[0004] However, while this method is effective in static leakage detection, it has fundamental limitations. The reliance on offline operation with water injection and pressurization makes it impossible to synchronously report weld penetration defects during welding, leading to delayed quality intervention. Furthermore, the strong arc light causes overexposure of weld images, resulting in loss of molten pool details, and the residual comparison is affected by light and shadow noise, increasing the false negative rate. Summary of the Invention

[0005] The purpose of this application is to provide an intelligent real-time detection method and system for PCCP welding quality, so as to solve the problem of weak anti-interference ability in the prior art.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides an intelligent real-time detection method for PCCP welding quality, comprising:

[0007] Simultaneously acquire light energy data and multi-source images of the spiral weld seam molten pool area;

[0008] Based on the energy difference between the visible and near-infrared bands in the molten pool region, the exposure parameters of the multi-source image are dynamically adjusted to suppress interference from overly bright areas caused by strong light during welding, thereby generating a weld surface image with enhanced penetration characteristics.

[0009] Based on the scanning of the weld's thermal deformation region by the emitted coherent beam, the changes in speckle pattern and sound frequency generated during the weld's thermal deformation are captured. Combined with the elastic characteristics of the thin-walled steel cylinder, the deformation distribution of the weld is calculated, and the stress concentration area is located based on the deformation distribution.

[0010] The surface image of the weld and the deformation data of the stress concentration area are input into a spatiotemporal convolutional neural network. The variation characteristics of the light energy data, the brightness and darkness details of the multi-source image and the spatial characteristics of the deformation distribution are fused to construct a spatial structure map of the weld. The position deviation of the sensor is adaptively corrected based on the degree of bending of the steel cylinder.

[0011] By comparing the shape data of the spatial structure diagram with the weld depression depth distribution reconstructed from the three-dimensional point cloud, depression areas with insufficient penetration are identified. The correlation between the degree of depression in the depression area and the stress response at the abnormal deformation site is analyzed, and a probability heat map is generated. Based on the probability heat map, the defect location and pressure failure risk level of the spiral weld are output.

[0012] Optionally, the weld surface image and deformation data of the stress concentration area are input into a spatiotemporal convolutional neural network. The network integrates the variation characteristics of the light energy data, the brightness details of the multi-source image, and the spatial features of the deformation distribution to construct a spatial structure map of the weld. Based on the degree of bending of the rolled steel cylinder, the sensor's positional deviation is adaptively corrected, including:

[0013] The weld surface image and the deformation data of the stress concentration area are aligned according to spatial coordinates to form spatiotemporally correlated input data.

[0014] In a spatiotemporal convolutional neural network, a multi-branch feature extraction path is set up. The first branch extracts the trajectory of the change of the light energy data over time, the second branch extracts the light and dark contrast features of the weld surface image, and the third branch extracts the spatial topology of the deformation distribution.

[0015] By integrating the change trajectory, the light and dark contrast features, and the spatial topology through the cross-branch feature interaction module, an initial spatial structure diagram of the weld is generated.

[0016] Obtain the radius of curvature of the rolled steel cylinder, calculate the offset of the sensor installation position relative to the weld centerline, and generate a position compensation vector based on the offset.

[0017] Based on the position compensation vector, the spatial coordinates of each pixel in the initial spatial structure map are corrected, and the corrected spatial structure map is output.

[0018] Optionally, by using a cross-branch feature interaction module, the change trajectory, the light and dark contrast features, and the spatial topology are fused to generate an initial spatial structure map of the weld, including:

[0019] In the cross-branch feature interaction module, the peak positions of energy fluctuations at continuous time points of the change trajectory are extracted to generate energy fluctuation path lines. Based on the light-dark contrast features, gray-level transition boundaries are identified to generate light-dark boundary contours. The node connection relationships are analyzed according to the spatial topology to construct a deformable association network.

[0020] The correlation strength between the energy fluctuation path and the light-dark boundary contour is calculated based on the cross-attention mechanism, and the spatial correlation between the nodes in the deformable correlation network and the energy fluctuation path is calculated at the same time.

[0021] Based on the correlation strength and the spatial correlation, the boundary weights of the light-dark boundary contour and the node weights of the deformable correlation network are dynamically adjusted.

[0022] The weighted light-dark boundary contour and the deformation association network are superimposed on the energy fluctuation path line to generate a fused feature map. Multi-scale mesh reconstruction is performed on the fused feature map, wherein the fine mesh level retains high-frequency details and the coarse mesh level integrates low-frequency contours.

[0023] The reconstructed fused feature map is compensated for the feature offset caused by the change in grid scale, and the compensated three-dimensional grid data is output as the initial spatial structure map of the weld.

[0024] Optionally, by comparing the shape data of the spatial structure diagram with the weld depression depth distribution reconstructed from the 3D point cloud, depression areas with insufficient penetration are identified. The correlation between the depression degree of the depression areas and the stress response at the abnormal deformation sites is analyzed, and a probability heat map is generated. Based on the probability heat map, the defect location and pressure failure risk level of the spiral weld are output, including:

[0025] The spatial structure diagram is fitted with a surface to generate a three-dimensional shape surface, and three-dimensional point cloud data of the weld surface is obtained by laser scanning.

[0026] Calculate the height difference between the three-dimensional shape surface and the three-dimensional point cloud data at each spatial location, and mark the region where the height difference exceeds a threshold as a concave region;

[0027] Extract the depression depth of each point within the depression region and correlate it with the deformation amount at the corresponding position in the deformation distribution map to generate a correlation function between depth and deformation amount;

[0028] Based on the correlation function, the stress response value of the depression region under pressure is predicted, and the stress response value is mapped to a probabilistic thermodynamic value.

[0029] A probability heat map is generated based on the spatial distribution of the probability heat map values. The heat map values ​​are divided into low-risk, medium-risk, and high-risk levels, and the defect location and pressure failure risk level of the spiral weld are output.

[0030] Optionally, the depression depth of each point within the depression region is extracted and correlated with the deformation amount at the corresponding position in the deformation distribution map to generate a correlation function between depth and deformation amount, including:

[0031] Extract the depression depth of each point within the depression region, establish the spatial coordinate mapping relationship of the depression region, and match the data points of the depression depth with the deformation data points of the same spatial coordinates in the deformation distribution map.

[0032] Extract the geometric feature parameters of the recessed region, and construct a deformation function for the recessed region based on the material properties of the thin-walled structure. The deformation function includes a depth square term and a deformation gradient term.

[0033] Based on the geometric feature parameters, set the initial weight range of the depth square term and the initial weight range of the deformation gradient change term;

[0034] The matching data points of the indentation depth and deformation amount are input into the joint equation system. Combined with the initial weight range, the optimal weight coefficients of the depth square term and the deformation amount gradient change term are solved. Based on the optimal weight coefficients, the correlation function between depth and deformation amount is generated.

[0035] Optionally, based on the energy difference between the molten pool region in the visible and near-infrared bands, the exposure parameters of the multi-source image are dynamically adjusted to suppress interference from overly bright areas caused by strong light during welding, generating a weld surface image with enhanced weld penetration characteristics, including:

[0036] The first energy value in the visible light band and the second energy value in the near-infrared band of the molten pool region are acquired simultaneously by a dual-channel sensor, and the ratio of the first energy value to the second energy value is calculated.

[0037] Based on the preset range in which the ratio falls, a corresponding exposure compensation curve is selected, wherein the preset range is related to the energy difference range of the melt pool region.

[0038] Based on the exposure compensation curve, the exposure time and gain parameters of each independent light source in the multi-light source image are adjusted to reduce the exposure parameters of overly bright areas and increase the exposure parameters of dark areas.

[0039] The adjusted multi-source image is fused at the pixel level to generate a single-frame fused image, wherein the texture details of the visible light band and the melting depth features of the near-infrared band are preserved during the fusion process.

[0040] The high-frequency components of the single-frame fused image are separated by a multi-scale decomposition method. The high-frequency components are then recombined after feature enhancement to generate a weld surface image with enhanced weld depth features.

[0041] Optionally, based on the emitted coherent beam scanning of the weld's thermal deformation region, the changes in speckle pattern and sound frequency generated during weld thermal deformation are captured. Combined with the elastic characteristics of the thin-walled steel cylinder, the deformation distribution of the weld is calculated, and stress concentration areas are located based on the deformation distribution, including:

[0042] A coherent beam of light is projected onto the weld surface to form an initial speckle pattern, and dynamic speckle image sequences and sound frequency signals during thermal deformation are acquired simultaneously.

[0043] The dynamic speckle image sequence is compared with the initial speckle pattern to extract the displacement change of each pixel from the initial state to the dynamic state, and the offset of the sound frequency signal relative to the reference frequency is calculated.

[0044] By combining the elastic parameters of the thin-walled steel cylinder, a joint function of the displacement change and offset is constructed. Based on the joint function, the deformation of each local area on the weld surface is calculated, and a deformation distribution map is generated based on the deformation.

[0045] Based on the difference in deformation gradient between adjacent regions in the deformation distribution map, the locations of gradient abrupt changes are identified and marked as stress concentration areas.

[0046] Secondly, this application provides an intelligent real-time detection system for PCCP welding quality, comprising:

[0047] The acquisition module is used to simultaneously acquire light energy data and multi-source images of the weld pool area of ​​the spiral weld.

[0048] The generation module is used to dynamically adjust the exposure parameters of the multi-source image based on the energy difference between the molten pool region in the visible light band and the near-infrared band, so as to suppress the interference of overly bright areas caused by strong light during welding and generate a weld surface image with enhanced penetration characteristics.

[0049] The positioning module is used to scan the thermal deformation area of ​​the weld seam based on the emitted coherent beam, capture the changes in speckle pattern and sound frequency generated when the weld seam is deformed by heat, and calculate the deformation distribution of the weld seam based on the elastic characteristics of the thin-walled steel cylinder, and locate the stress concentration area according to the deformation distribution.

[0050] The fusion module is used to input the weld surface image and the deformation data of the stress concentration area into a spatiotemporal convolutional neural network, fuse the variation characteristics of the light energy data, the brightness and darkness details of the multi-source image and the spatial characteristics of the deformation distribution, construct a spatial structure map of the weld, and adaptively correct the position deviation of the sensor based on the degree of bending of the steel cylinder.

[0051] The output module is used to compare the shape data of the spatial structure diagram with the weld depression depth distribution reconstructed from the three-dimensional point cloud, identify depression areas with insufficient penetration, analyze the correlation between the depression degree of the depression area and the stress response at the abnormal deformation site, generate a probability heat map, and output the defect location and pressure failure risk level of the spiral weld based on the probability heat map.

[0052] Thirdly, this application provides an electronic device, comprising:

[0053] Memory, used to store computer programs;

[0054] A processor is configured to execute the computer program to implement the steps of the intelligent real-time detection method for PCCP welding quality as described in the first aspect above.

[0055] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the intelligent real-time detection method for PCCP welding quality as described in the first aspect above.

[0056] This application provides an intelligent real-time detection method for PCCP welding quality. By simultaneously acquiring light energy data and multi-source images of the weld pool region of the spiral weld, and through synchronous capture of visible and near-infrared energy, it ensures the spatiotemporal alignment of multimodal data during the welding process, providing fundamental data support for subsequent dynamic anti-interference processing. Based on the energy difference in the weld pool region, it dynamically adjusts exposure parameters, adaptively suppressing overexposure of strong arc light by utilizing the energy difference between visible and near-infrared bands. This eliminates interference from welding spatter and fumes on image details, significantly improving the visualization clarity of weld depth features. Based on coherent beam scanning to capture speckle patterns and sound frequency changes, and through acoustic-optical co-sensing combined with a thin-walled steel cylinder elastic model, it can locate areas of thermal stress concentration in real time, allowing for early detection. It predicts the risk of microcrack initiation, compensating for the lack of physical quantities in pure image detection; a spatiotemporal convolutional neural network integrates multi-source data to construct a spatial structure map, fusing light energy change trajectories, image brightness and darkness details, and deformation space topology, enabling the construction of a joint representation of the three-dimensional geometry and mechanical state of the weld, solving the deformation misjudgment problem caused by traditional two-dimensional image analysis; at the same time, based on the adaptive correction of sensor pose deviation by the steel cylinder curvature, it can eliminate the positioning error of the detection equipment on the surface of the curled steel cylinder; by comparing the correlation between the three-dimensional point cloud indentation depth and deformation stress to generate a probability heat map, and by modeling the correlation between the indentation distribution of the weld depth and the stress response, it can quantify the pressure failure probability of the insufficient weld depth area, output the defect location and risk level, and achieve a leap from "defect identification" to "risk prediction". Attached Figure Description

[0057] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 A flowchart illustrating a method for intelligent real-time detection of PCCP welding quality provided in an embodiment of this application;

[0059] Figure 2 A schematic diagram of a specific process for a PCCP welding quality intelligent real-time detection method provided in this application embodiment;

[0060] Figure 3 A scene diagram illustrating an intelligent real-time detection method for PCCP welding quality provided in an embodiment of this application;

[0061] Figure 4 This is a schematic diagram of a PCCP welding quality intelligent real-time detection system provided in an embodiment of this application. Detailed Implementation

[0062] Research has found that existing PCCP welding quality inspection technologies rely on offline water injection and pressurization for static leakage detection, which cannot simultaneously provide feedback on weld penetration defects during the welding process, leading to delayed quality intervention. Furthermore, strong arc light causes overexposure of weld images, resulting in loss of molten pool details, and residual comparison is affected by light and shadow noise, increasing the false negative rate. These fundamental limitations severely hinder real-time, high-precision defect identification. Therefore, there is an urgent need for an intelligent real-time detection method that can simultaneously detect weld penetration defects during the welding process, reduce overexposure interference, and lower the false negative rate.

[0063] To address the aforementioned issues, this invention proposes an intelligent real-time detection method for PCCP welding quality, the core of which lies in the dynamic fusion of multi-source data and adaptive intelligent analysis. Specifically, firstly, light energy data and multi-source images of the weld pool area of ​​the spiral weld are simultaneously acquired. Exposure parameters are dynamically adjusted based on the energy difference between visible and near-infrared bands to suppress excessive brightness interference caused by strong welding light. Next, a coherent beam is emitted to scan the weld thermal deformation area, capturing speckle patterns and changes in sound frequency. Combining this with the elastic characteristics of thin-walled steel cylinders, the deformation distribution is calculated and stress concentration areas are located. Subsequently, the weld surface image and deformation data are input into a spatiotemporal convolutional neural network, fusing the characteristics of light energy changes, image brightness details, and spatial features of deformation distribution to construct a weld spatial structure map and adaptively correct sensor position deviations. Finally, the shape data of the spatial structure map is compared with the weld depression depth distribution reconstructed from the 3D point cloud to identify areas with insufficient weld penetration, analyze the correlation between the degree of depression and the stress response at abnormal deformation locations, and generate a probabilistic heat map to output defect location and pressure failure risk level. This method enables synchronous real-time feedback of weld depth defects during the welding process, effectively overcoming the problems of overexposure and loss of details caused by strong light interference. Furthermore, it reduces noise interference from residual comparison through multi-source data fusion, significantly improving detection accuracy and reducing the false negative rate.

[0064] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0065] The core of this application is to provide an intelligent real-time detection method for PCCP welding quality, and a flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes:

[0066] S101. Synchronously acquire light energy data and multi-source images of the spiral weld seam molten pool area;

[0067] In the above steps, light energy data refers to the quantitative information of light intensity detected in the molten pool area of ​​the spiral weld, including the energy level values ​​of visible light or infrared radiation, which are used to reflect the distribution and changes of light during the welding process; multi-source images are image sequences captured by different light sources such as visible light and infrared light under simultaneous or alternating illumination, which are used to capture the morphological characteristics and dynamic changes of the molten pool. Together, they provide the original data source for subsequent analysis.

[0068] In this embodiment, firstly, light energy data and multi-source images of the molten pool area of ​​the spiral weld are simultaneously acquired. In an actual spiral welding test, two independent light sources are arranged: a 630nm red visible light source and an 850nm infrared light source, with power settings of 100W and 80W respectively. Simultaneously, a photodiode is installed in the molten pool area as a light energy sensor, and a high-speed CMOS camera is installed as an image capture device. Then, a central synchronization controller sends a timestamp signal, and its internal synchronization trigger module simultaneously starts the sensor and camera with millisecond-level precision. This includes real-time measurement of the molten pool area by the sensor. The light intensity data of the pool area is sampled at a frequency of 1000Hz, and the output unit is mW / cm². For example, when there is strong light interference from the welding arc, the sensor may measure the light intensity value at the center of the molten pool to be 250mW / cm². The camera captures a sequence of images at 120fps to capture the outline of the molten pool under strong light. During the acquisition process, all data and images are marked with the same timestamp and stored in the data recording system. Finally, a synchronized light energy dataset and corresponding multi-image sequences are obtained. For example, a set of data contains 1000 light intensity sampling points and corresponding 100 frames of color images, ensuring that the two are consistent and aligned in the time dimension, supporting subsequent weld depth and crack detection.

[0069] In practical applications, during the inspection of a welding platform, to overcome the impact of strong arc light interference on weld inspection, operators adopted a multispectral synchronous acquisition scheme. Specifically, three sets of general-purpose light sources were symmetrically deployed at the weld location: Group A, a 760nm infrared light source with a power of 120W; Group B, an 850nm near-infrared light source with a power of 100W; and Group C, a 450nm blue light source with a power of 80W. Then, a high-speed camera continuously photographed the molten pool area at a rate of 200 frames per second, while photoelectric sensors recorded the radiation energy data in the visible light band (380 to 780 nm) and the infrared band (780 to 2500 nm) in real time. To synchronize data acquisition, the three sets of light sources were lit sequentially within 0.05 seconds, with the lighting interval of each light source set to 0.005 seconds. The acquisition period was determined by the formula "number of light sources × lighting interval + camera exposure time," which, when substituted, resulted in 3 sets × 0.005 seconds + 0.001 seconds = 0.016 seconds. When processing the acquired data, for example, when detecting the blue light image of group C, the original grayscale value of 186 at the edge of the molten pool is extracted. After subtracting the measured arc light background interference value of 152, the actual feature signal is calculated as 186 - 152 = 34 grayscale units. Simultaneously, combined with the thermal radiation data of group B at the 850 nm band, the irradiance recorded by the sensor at 2800 watts per square meter is converted into a temperature value. Using the formula "Temperature = Measured Irradiance × Calibration Coefficient + Ambient Temperature", the parameters 2800 × 0.35 + 25 = 1005 degrees Celsius are substituted. This series of operations, through precisely controlled spectral separation technology, allows the tiny thermal stress cracks in the heat-affected zone of the weld to be clearly displayed in a strong interference environment, effectively improving defect identification.

[0070] In the overall scheme of step S101 above, by simultaneously acquiring light energy data and multi-source images of the spiral weld molten pool area under strong arc light interference, the dynamic physical state and characteristic morphological information of the molten pool can be completely captured in the same cross-section at the same time. This effectively overcomes the defects of blurred molten pool outline and missing details caused by the strong arc light when acquiring image data in the traditional single method. It ensures that the molten depth characteristic signal and thermal stress crack distribution information are accurately aligned in the spatiotemporal dimension, thereby establishing a reliable data foundation for high-precision synchronous analysis of the internal molten depth evolution process of the weld and the initiation and propagation trend of external thermal stress cracks.

[0071] S102. Based on the energy difference between the visible light band and the near-infrared band in the molten pool area, the exposure parameters of the multi-source image are dynamically adjusted to suppress the interference of overly bright areas caused by strong light during welding, and to generate a weld surface image with enhanced penetration characteristics.

[0072] Optionally, step S102 may specifically include the following steps:

[0073] S1021. Simultaneously acquire the first energy value of the molten pool region in the visible light band and the second energy value in the near-infrared band using a dual-channel sensor, and calculate the ratio of the first energy value to the second energy value.

[0074] S1022. Select the corresponding exposure compensation curve according to the preset range in which the ratio falls, wherein the preset range is related to the energy difference range of the molten pool region.

[0075] S1023. Based on the exposure compensation curve, adjust the exposure time and gain parameters of each independent light source in the multi-light source image to reduce the exposure parameters of overly bright areas and increase the exposure parameters of dark areas.

[0076] S1024. Perform pixel-level fusion on the adjusted multi-source image to generate a single-frame fused image, wherein the texture details of the visible light band and the melting depth features of the near-infrared band are preserved during the fusion process.

[0077] S1025. The high-frequency components of the single-frame fused image are separated by a multi-scale decomposition method, and the high-frequency components are recombined after feature enhancement to generate a weld surface image with enhanced penetration depth features.

[0078] In the above steps, a dual-channel sensor refers to a detector capable of simultaneously measuring the energy values ​​of the visible light and near-infrared bands, and its output data is a quantized value of light intensity, in mW / cm²; the exposure compensation curve refers to the adjustment rules of exposure parameters based on the ratio of visible light to near-infrared light energy, dividing different ratio ranges through experimental data and matching corresponding curves to dynamically suppress strong light interference; pixel-level fusion refers to the process of superimposing the adjusted visible light and near-infrared images into a single frame according to weights, preserving the texture details of visible light and the melting depth contour of near-infrared; multi-scale decomposition refers to the technique of separating the high-frequency and low-frequency components of an image through mathematical methods and enhancing and recombining the high-frequency components.

[0079] In this embodiment, firstly, in step S1021, a dual-channel sensor is used to simultaneously acquire the first energy value of the molten pool region in the visible light band and the second energy value in the near-infrared band, and the ratio between the two is calculated. The sensor outputs light intensity data of the two bands in real time, with a sampling frequency of 1000Hz, and the ratio is calculated by the built-in processor. ,in The energy value of visible light. This represents the energy value of near-infrared light. For example, when the visible light energy is 180 mW / cm² and the near-infrared energy is 90 mW / cm², the ratio is calculated. R =2.0, this ratio is used to quantify the spectral energy difference in the melt pool region, providing a basis for subsequent exposure adjustments.

[0080] Secondly, in step S1022, the corresponding exposure compensation curve is selected based on the preset range of the ratio R. Three preset ranges are defined: R < 1.2 is the strong arc light interference range, matched with curve A; 1.2 ≤ R ≤ 2.5 is the medium interference range, matched with curve B; and R > 2.5 is the weak interference range, matched with curve C. The range thresholds are set based on welding experimental data, such as a significant reduction in near-infrared energy of the molten pool under strong arc light. For example, when R = 2.0, it falls into the medium interference range, and curve B is automatically selected. This curve defines a reduction in visible light channel exposure time and an increase in near-infrared gain.

[0081] Next, in step S1023, based on the selected exposure compensation curve, the exposure parameters of the visible light channel are first adjusted: the exposure time is shortened from the original 10 milliseconds to 2 milliseconds, and the gain is reduced from 15 dB to 8 dB. By shortening the photosensitivity time and reducing the signal amplification amplitude, the overly bright areas caused by the welding arc are directly suppressed. Then, the gain parameter for the near-infrared channel is increased: from 20 dB to 25 dB, enhancing the signal response in the dark areas of the infrared spectrum. This operation significantly reduces the pixel grayscale values ​​in overexposed areas, for example, the center point of the molten pool decreases from 250 to 180, while the details at the edges of the dark molten pool are enhanced, such as the boundary point grayscale value increasing from 30 to 70. For example, in a strong arc light interference experiment, the overexposed area of ​​the visible light image is reduced after adjustment, while the contrast of the melt depth contour in the near-infrared channel is improved, providing balanced dual-channel image data for subsequent fusion steps.

[0082] Then, in step S1024, the adjusted visible light and near-infrared images are weighted and superimposed to fuse, and the visible light image is assigned... Weighting to preserve weld surface texture details, near-infrared image allocation Weights are used to highlight the depth features of the molten pool. The fused grayscale value of each pixel is calculated using the formula: For example, if a pixel has a grayscale value of 150 representing the texture of a metal surface in a visible light image, and 80 reflecting the depth contour of a molten pool in a near-infrared image, the fusion result is: This process preserves the microstructural features of the visible light channel, such as the 0.1 mm weld seam texture, while enhancing the three-dimensional information of the molten pool in the near-infrared channel, ultimately generating a single-frame fused image.

[0083] Finally, in step S1025, the single-frame fused image is decomposed into three scales using the Gaussian pyramid algorithm: the original resolution layer, the half-scale layer, and the quarter-scale layer. High-frequency components, including molten pool edges and micro-crack features, are extracted from each layer. The pixel values ​​of these components are multiplied by a 1.5-fold enhancement factor; for example, the grayscale value of edge points is increased from 120 to 180. The enhanced high-frequency components are then reconstructed with the original low-frequency background. After reconstruction, the visual contrast of the molten pool boundary cracks increases; for example, the grayscale difference between the crack area and the background increases from 40 to 60. The signal-to-noise ratio of image noise in areas with strong light interference is improved to over 20 dB; for example, the background noise intensity decreases. The final output is a weld surface image with significantly enhanced molten depth features.

[0084] In practical applications, during laser welding, Optical Coherence Tomography (OCT) technology combines the principle of low-coherence interference with laser ranging to achieve real-time, high-precision monitoring of the weld pool depth. Its core process is as follows: The system first emits a low-power near-infrared detection laser beam, typically with a wavelength of 820-860 nm. This beam is split into two paths by a beam splitter—a measurement beam and a processing laser are coaxially focused on the bottom of the keyhole in the weld pool, while a reference beam acts on the weld plane as a reference height. When the measurement beam is reflected from the bottom of the keyhole, a phase shift occurs between it and the reference beam due to differences in their transmission paths. The system captures the interference signal using a spectrometer at a high frequency of 250 kHz and uses Fast Fourier Transform (FFT) to resolve the phase difference, ultimately calculating the keyhole depth with micron-level accuracy and an axial resolution of 20 μm. For example, in the welding test of 304 stainless steel, the error between the real-time measured weld depth value by the OCT system and the metallographic result is less than 3.9%. For instance, a metallographic value of 0.7055 mm corresponds to an OCT value of 0.6779 mm, verifying its reliability. This non-contact measurement method can dynamically adapt to the kHz-level oscillation of the molten pool. At the same time, through anti-interference design, such as enhancing the detection optical power and reducing optical aberrations, it overcomes the interference of strong welding arc light and metal vapor. Real-time data is fed back to the control system to automatically adjust the laser power or welding speed. For example, if a shallower penetration is detected, the power is increased by 10%, forming a closed-loop optimization of process parameters. This effectively suppresses defects such as insufficient fusion and porosity, and improves the stability of welding quality.

[0085] In the overall scheme of step S102 above, the exposure parameters of the multi-source image are dynamically adjusted based on the energy difference between the visible and near-infrared bands of the molten pool region. This effectively suppresses interference from overly bright areas caused by strong welding arc light. The energy ratio of the molten pool region in the visible and near-infrared bands is simultaneously acquired using a dual-channel sensor. Based on a preset range, a corresponding exposure compensation curve is selected to precisely reduce the exposure time and gain parameters of overly bright areas while simultaneously increasing the exposure intensity of dark areas. Subsequently, the adjusted multi-source image is fused at the pixel level. While preserving the texture details in the visible band, the melt depth characteristics in the near-infrared band are enhanced. Then, the high-frequency components of the single-frame fused image are separated using a multi-scale decomposition method and feature enhancement and recombination are performed. Finally, a weld surface image with significantly enhanced melt depth characteristics is generated. This technology not only overcomes the inherent defects of blurred molten pool contours and loss of detail under strong arc light conditions, but also achieves high-fidelity synchronous presentation of molten pool surface texture and deep penetration morphology through dynamic control of spectral energy differences. This provides a high-contrast, low-noise image foundation for subsequent melt depth quantification and accurate analysis of thermal stress cracks.

[0086] S103. Based on the emitted coherent beam scanning of the weld thermal deformation area, capture the speckle pattern change and sound frequency change generated when the weld is deformed by heat, combine the elastic characteristics of the thin-walled steel cylinder, calculate the deformation distribution of the weld, and locate the stress concentration area according to the deformation distribution.

[0087] Optionally, step S103 may specifically include the following steps:

[0088] S1031. Project a coherent beam onto the weld surface to form an initial speckle pattern, and simultaneously acquire dynamic speckle image sequences and sound frequency signals during thermal deformation.

[0089] S1032. Compare the dynamic speckle image sequence with the initial speckle pattern, extract the displacement change of each pixel from the initial state to the dynamic state, and calculate the offset of the sound frequency signal relative to the reference frequency.

[0090] S1033. Combining the elastic parameters of the thin-walled steel cylinder, construct a joint function of the displacement change and the offset. Based on the joint function, calculate the deformation of each local area on the weld surface, and generate a deformation distribution map based on the deformation.

[0091] S1034. Based on the difference in deformation gradient between adjacent regions in the deformation distribution map, identify the gradient abrupt change location and mark the gradient abrupt change location as a stress concentration region.

[0092] In the above steps, the coherent beam refers to a laser beam with exactly the same wavelength and phase. After being projected onto the weld surface, it forms randomly distributed bright and dark spots due to light wave interference. The speckle pattern is an image composed of the above spots, and its variation is caused by the difference in the optical path due to the thermal deformation of the material, which is used to quantify displacement. The change in sound frequency refers to the shift in sound wave frequency generated by the vibration of the material due to welding thermal stress, which is used to correlate dynamic stress. The elastic characteristics of the thin-walled steel cylinder involve the physical properties of the material's resistance to deformation, including the elastic modulus and Poisson's ratio. These parameters determine the mathematical relationship between stress and deformation. The deformation distribution refers to a two-dimensional thermogram generated by calculating the displacement of each local area on the weld surface, with the degree of deformation indicated by color gradient. The stress concentration area is the location in the deformation distribution map where the deformation gradient between adjacent areas changes abruptly.

[0093] In this embodiment, firstly, in step S1031, a laser beam with a fixed wavelength and phase is emitted onto the weld surface using coherent beam projection technology, forming a stable initial speckle pattern based on the principle of laser interference. Simultaneously, a high-speed industrial camera continuously acquires a sequence of dynamic speckle images during the thermal deformation process, and a piezoelectric acoustic sensor captures the sound frequency signals generated by material vibration in real time. This step generates a synchronized optical image sequence and acoustic signal dataset, serving as the input for subsequent analysis. For example, in a thin-walled steel cylinder welding experiment, a laser beam with a wavelength of 632.8 nm covers a region of ±5 mm around the weld centerline, a high-speed camera records the speckle displacement trajectory at a rate of 120 frames per second, and an acoustic sensor detects that the sound wave frequency has shifted from a reference value of 5000 Hz to 5800 Hz, forming raw data containing optical displacement and acoustic stress characteristics.

[0094] Secondly, in step S1032, the previously output dynamic speckle image sequence and sound signal dataset are compared frame-by-frame with the initial pattern using a digital image correlation analysis algorithm and sub-pixel-level image matching technology, calculating the displacement change of each pixel. Simultaneously, a fast Fourier transform algorithm is used to convert the sound signal from the time domain to the frequency domain, and the offset of the actual frequency relative to the reference frequency is extracted. This step converts the optical and acoustic data into quantized displacement values ​​Δd and frequency offsets Δf, providing accurate input parameters for deformation calculation. For example, in the image analysis, the displacement change of a certain detection point in the weld thermal deformation area is measured to be 0.05 mm, and the sound signal spectrum analysis shows a sound frequency offset of 800 Hz. The output displacement and offset values ​​are used for subsequent function construction.

[0095] Next, in step S1033, the output displacement change and sound frequency offset parameters are combined with the material properties, elastic modulus, and Poisson's ratio of the thin-walled steel cylinder to construct a linear joint function to calculate the deformation of each local area on the weld surface. Its functional form is: ,in The material strain coefficient The acoustic stress conversion factor is... The change in displacement The first value is the audio frequency offset. Then, a bicubic interpolation algorithm is used to integrate the deformation of each region into a continuously distributed two-dimensional heatmap, with color gradients indicating the degree of deformation. This step integrates physical parameters with input data and outputs a visualized deformation distribution map. For example, for a 5 mm × 5 mm region, the input displacement change is 0.08 mm and the audio frequency offset is 600 Hz. The calculated deformation Δε = 0.072 is obtained from the megapascals per hertz (MPa). This area is shown in red on the thermogram, indicating a high-strain region with a deformation ≥ 0.1 mm.

[0096] Finally, in step S1034, based on the output deformation distribution heatmap, the Sobel edge detection algorithm is applied to scan the deformation gradient of adjacent regions and calculate the rate of change of deformation per unit distance. When the deformation gradient value exceeds a preset threshold, it is determined to be a gradient abrupt change location and marked as a stress concentration area. This step extracts gradient information from the heatmap and identifies high-stress risk points, outputting the marking results to complete the detection. For example, if the deformation of a certain segment of the weld fusion line suddenly increases from 0.06 mm to 0.15 mm, the gradient per unit distance reaches 0.09 mm, exceeding the threshold of 0.05 mm, and this location is identified as a potential crack initiation point.

[0097] In a practical application scenario involving the welding quality inspection of a thin-walled steel cylinder, a coherent optical scanning and acoustic signal fusion analysis technique was employed to monitor weld thermal deformation and locate stress concentration areas in real time. The specific operational procedure is as follows: First, a coherent laser beam with a wavelength of 632.8 nm is projected onto the weld surface, and an initial speckle pattern is acquired as a reference under stable room temperature conditions. Simultaneously, a high-speed industrial camera at a frame rate of 2000 fps is activated to record a sequence of dynamic speckle images during the welding thermal cycle. At the same time, a piezoelectric acoustic sensor acquires sound signals with a frequency range of 20 kHz to 100 kHz. Subsequently, the dynamic speckle image sequence is compared pixel-level with the initial speckle pattern. The displacement change Δd (in micrometers) of each pixel from the initial state to the dynamic state is calculated using a digital image correlation (DIC) algorithm. Simultaneously, the offset Δf (in Hertz) of the sound signal relative to the reference frequency of 400 kHz is extracted. Based on the material elastic parameters of the thin-walled steel cylinder, with elastic modulus E=210GPa, Poisson's ratio μ=0.3, and density ρ=7850kg / m³, a joint function of displacement change and sound frequency shift is constructed: deformation. ,in, The sound frequency to strain conversion coefficient was determined to be 0.15 through material calibration experiments. Based on this function, the deformation of each 1mm×1mm local area on the weld surface is calculated, and a full-field deformation distribution map is generated. Finally, by analyzing the gradient differences in deformation between adjacent areas in the deformation distribution map, locations where the gradient abruptly exceeds the threshold of 0.5mm⁻¹ are identified. For example, at a coordinate point x=15mm, the gradient value abruptly changes from 0.2 to 1.8, and this is marked as a stress concentration area, thereby achieving precise location of potential weld cracking risks.

[0098] In the overall scheme of step S103 above, an initial speckle pattern is formed by projecting a coherent beam onto the weld surface. Simultaneously, dynamic speckle image sequences and sound frequency signals during the thermal deformation process are acquired. The speckle image sequence is compared with the initial pattern at the pixel level to quantify the deformation of each local area. At the same time, the offset of the sound frequency signal relative to the reference frequency is analyzed. Based on parameters such as the elastic modulus and Poisson's ratio of the thin-walled steel cylinder, a joint function is constructed by fusing the displacement change and the sound frequency offset to accurately calculate the deformation of each area on the weld surface and generate a full-field deformation distribution map. By identifying the gradient abrupt change locations of adjacent areas in the deformation distribution map, stress concentration areas are located. This overcomes the sensor failure problem caused by high temperatures in traditional contact measurements and avoids the insufficient signal-to-noise ratio defect of single optical measurements under strong arc light interference. It provides full-field deformation data with micron-level accuracy and early stress concentration warning for the integrity assessment of welded structures.

[0099] This application provides a schematic flowchart of a PCCP welding quality intelligent real-time detection method, as shown in the embodiments below. Figure 2 As shown, it includes the following:

[0100] S104. Input the weld surface image and the deformation data of the stress concentration area into a spatiotemporal convolutional neural network, fuse the variation characteristics of the light energy data, the brightness and darkness details of the multi-source image and the spatial characteristics of the deformation distribution, construct a spatial structure map of the weld, and adaptively correct the position deviation of the sensor based on the degree of bending of the steel cylinder.

[0101] Optionally, step S104 may specifically include the following steps:

[0102] S1041. Align the weld surface image with the deformation data of the stress concentration area according to spatial coordinates to form spatiotemporal correlated input data;

[0103] S1042. In the spatiotemporal convolutional neural network, a multi-branch feature extraction path is set up. The first branch extracts the trajectory of the change of the light energy data over time, the second branch extracts the light and dark contrast features of the weld surface image, and the third branch extracts the spatial topology of the deformation distribution.

[0104] S1043. By using the cross-branch feature interaction module, the change trajectory, the light and dark contrast features, and the spatial topology are integrated to generate an initial spatial structure diagram of the weld.

[0105] Specifically, step S1043 may include the following processes: In the cross-branch feature interaction module, extract the peak positions of energy fluctuations at continuous time points of the change trajectory, generate energy fluctuation path lines, identify grayscale transition boundaries based on the light-dark contrast features, generate light-dark boundary contours, analyze node connection relationships according to the spatial topology, and construct a deformable association network; calculate the association strength between the energy fluctuation path lines and the light-dark boundary contours based on the cross-attention mechanism, and simultaneously calculate the spatial correlation between nodes in the deformable association network and the energy fluctuation path lines; dynamically adjust the boundary weights of the light-dark boundary contours and the node weights of the deformable association network according to the association strength and the spatial correlation; superimpose the weighted light-dark boundary contours and the deformable association network onto the energy fluctuation path lines to generate a fused feature map; perform multi-scale mesh reconstruction on the fused feature map, where fine mesh levels retain high-frequency details and coarse mesh levels integrate low-frequency contours; compensate the feature offset caused by mesh scale changes in the reconstructed fused feature map, and output the compensated three-dimensional mesh data as the initial spatial structure map of the weld.

[0106] S1044. Obtain the radius of curvature of the rolled steel cylinder, calculate the offset of the sensor installation position relative to the weld centerline, and generate a position compensation vector based on the offset.

[0107] S1045. Based on the position compensation vector, correct the spatial coordinates of each pixel in the initial spatial structure map, and output the corrected spatial structure map.

[0108] In the above steps, the weld surface image refers to two-dimensional image data acquired through a vision sensor, containing the light intensity distribution and texture details of the weld surface. Deformation data of stress concentration areas refers to the output information on three-dimensional deformation and gradient changes, identifying locations with high stress risk. Spatiotemporal convolutional neural networks are deep learning architectures that integrate spatial and temporal features, extracting spatial topology and temporal patterns from the data through convolution operations. The variation characteristics of light energy data reflect the fluctuation pattern of material thermal radiation intensity over time. The brightness and darkness details of multi-source images refer to the grayscale contrast features formed on the weld surface under illumination from different angles. The spatial characteristics of deformation distribution describe the distribution pattern and structural correlation of weld surface deformation in two-dimensional or three-dimensional space. The spatial structure diagram of the weld is a three-dimensional geometric model generated by integrating multi-source features, representing the three-dimensional morphology of the weld. The degree of curvature of the steel cylinder refers to the amount of surface deformation determined by the workpiece's radius of curvature. The sensor's positional deviation represents the spatial offset between the sensor's installation position and the actual weld centerline.

[0109] In this embodiment, firstly, step S1041 aligns the weld surface image with the deformation data of the stress concentration area using a unified spatial coordinate system. Specifically, camera calibration parameters are used to establish the conversion relationship between image pixel coordinates and world coordinates. Simultaneously, the coordinates of the stress concentration points marked in the deformation distribution map are mapped to the same coordinate system, forming a spatiotemporally correlated input dataset. For example, in the inspection of thin-walled steel cylinders, when the camera focal length is set to 18.838 mm and the object distance is set to 115.113 mm, the image pixel coordinates are converted into three-dimensional spatial coordinates using a pinhole imaging model, and precisely matched with the stress points at coordinates of 0.08 mm and 0.15 mm in the deformation distribution map, achieving sub-millimeter level alignment accuracy.

[0110] Secondly, in step S1042, three independent feature extraction paths are deployed in the spatiotemporal convolutional neural network. The first branch uses a temporal convolutional network to process light energy data, capturing the energy fluctuation trajectory of the audio signal through a sliding window and extracting the peak offset at consecutive time points. The second branch uses a lightweight Ghost Net convolutional network to extract the light and dark contrast features of the weld surface image, calculates the gray-level gradient through depthwise separable convolution, and identifies boundaries where the gray-level transition exceeds 50%. The third branch uses a graph convolutional network to analyze the spatial topology of the deformation distribution map, constructing an association network with stress concentration points as nodes and deformation gradients as edge weights. For example, if the audio signal frequency is 5800 Hz at a certain moment, the first branch extracts its offset trajectory relative to the reference value of 5000 Hz; the second branch identifies a weld edge where the gray-level value jumps from 100 to 160; and the third branch analyzes the node connections where the deformation gradient in a certain region reaches 0.09 mm, outputting three types of features for subsequent fusion.

[0111] Next, cross-modal fusion is performed in step S1043. First, the peak positions of consecutive time points on the energy fluctuation path are extracted to generate the path. Based on the contrast features, gray-level transition boundaries with gradients exceeding 0.05 per pixel are identified to generate light-dark boundary contours. The node connection strength is calculated according to the deformation association network; edges with gradient abrupt changes > 0.09 / mm are marked as strongly associated. Subsequently, the association strength between the energy path and the light-dark contour is calculated through a cross-attention mechanism. Specifically, the dot product of the two types of feature vectors is calculated and normalized to obtain the association weight matrix. For example, the weight of the overlapping area between the path and the contour is increased to 0.85. At the same time, the spatial correlation between the deformation nodes and the energy path is calculated. The Euclidean distance formula is used to calculate the distance from the node to the path. When the distance is less than 1 mm, the correlation coefficient is taken as 0.7. Next, the weighted shading contour weights (0.85) and deformed network node weights (0.7) are superimposed on the energy path line to generate a fused feature map. Features are then integrated through multi-scale mesh reconstruction: the fine mesh layer has a resolution of 0.1 mm to preserve high-frequency details, while the coarse mesh layer has a resolution of 1 mm to integrate low-frequency contours. Bilinear interpolation is used to adjust the mesh scale during the reconstruction process. Finally, feature offset caused by mesh transformation is compensated using the formula: current mesh scale difference × 0.02 mm. For example, 0.02 mm compensation is added when converting from a fine mesh to a coarse mesh. The resulting 3D mesh data serves as the initial spatial structure map.

[0112] Then, the sensor position deviation is calculated in step S1044. First, the radius of curvature of the steel cylinder is obtained, and the offset is calculated based on the angle between the sensor axis and the weld centerline. The calculation formula is as follows: ,in Radius of curvature (unit: millimeters). The angle value is in radians. The angle value needs to be converted to radians using the following formula: radian angle A three-dimensional position compensation vector is generated based on the offset. For example, when the steel cylinder radius is 2000 mm and the included angle is 2.095 degrees, the calculation process is as follows: radian radian, Millimeters, generating compensation vector Millimeters.

[0113] Finally, the initial spatial structure map is geometrically corrected in step S1045. The compensation vector is applied to the spatial coordinates of each pixel, and the correction formula is as follows: ,in, , These are the corrected coordinates. , , The coordinates before correction. , , This is the correction amount. To verify the deviation between the corrected stress concentration point and the actual location, the deviation value is calculated using the Euclidean distance formula: ,in, These are the coordinates of the measured point. These are the theoretical coordinates. For example, the initial coordinates of a point (100.0, 50.0, 0.0) are corrected to (173.2, 50.0, 0.0), while the actual weld position is (173.0, 50.0, 0.0). The deviation is calculated as follows: The maximum positional deviation was reduced from 0.34 mm to 0.15 mm, and the corrected spatial structure diagram was output.

[0114] In practical applications, in an automated welding quality inspection system for thin-walled steel cylinders, the weld surface image with a resolution of 2048×2048 pixels is first spatially aligned with the stress concentration deformation data stored in a 1mm×1mm grid, with the image corresponding to a 200mm×200mm actual area. The deformation data is resampled to the same resolution using bilinear interpolation, ensuring each pixel contains grayscale value, deformation amount, and light energy fluctuation value (W / m²). The system initiates three feature extraction branches: the first branch uses a 3-dimensional convolutional kernel to scan the energy data, detecting a continuous three-second sequence [152,280,186] to determine the peak time t=2s, generating an energy fluctuation path line; the second branch uses an edge detection operator to calculate pixel gradients, synthesizing the horizontal gradient 34 and vertical gradient 19 of a certain region to mark the light and dark boundaries; the third branch uses grids with deformation >500με as nodes, establishing connection edges when the distance between adjacent nodes is ≤3mm and the deformation gradient difference is ≥200με / mm, constructing a deformation association network. The spatial correlation between the energy path and the light / dark boundary is calculated during the fusion stage: when the energy peak point is 0.5mm from the light / dark boundary, the correlation weight w = 1 / (1+0.5²) = 0.8; simultaneously, the node importance is adjusted by multiplying the weight by 1.5 when the distance of the deformed node from the energy path is ≤2mm and by multiplying it by 0.7 when d >2mm. The weighted light / dark boundary and the deformed network are superimposed on the energy path to generate a fused feature map, and then multi-scale reconstruction is performed: the fine mesh layer (0.1mm / cell) retains high-frequency details such as contours with gradients ≥39, the coarse mesh layer (1mm / cell) integrates node connections, and scale errors are eliminated through offset compensation to output the initial 3D structure map. Based on the radius of curvature R = 1.2m measured by the laser rangefinder, combined with the arc length L = 0.6m at the center of the weld, the central angle θ = L / R = 0.5rad is obtained. When the sensor is offset from the center of the weld... When the value is 3mm, generate the position compensation vector: The coordinates of each pixel in the initial image are corrected: for example, the original coordinates (100-1.44, 200+2.63) = (98.56, 202.63) mm. Through multi-source data fusion and curvature adaptive calibration, micron-level 3D modeling of weld seams is achieved.

[0115] In the overall scheme of step S104 above, multispectral dynamic exposure control is used to precisely suppress interference from overly bright areas of the molten pool, effectively suppressing overly bright areas of the molten pool and enhancing the molten depth characteristics, thereby achieving synchronous high-fidelity characterization of molten depth quantification and surface texture; through joint analysis of speckle displacement and acoustic frequency shift, micron-level thermal deformation field reconstruction and precise positioning of stress concentration areas are completed; by using a spatiotemporal convolutional neural network to fuse multi-source data and compensate for positional deviations, a three-dimensional dynamic structure map of the weld with sub-millimeter precision is constructed, ultimately achieving non-destructive, synchronous, and full-field detection of the molten depth and thermal stress cracks of the spiral weld.

[0116] S105. By comparing the shape data of the spatial structure diagram with the weld depression depth distribution reconstructed from the three-dimensional point cloud, depression areas with insufficient penetration are identified. The correlation between the depression degree of the depression area and the stress response at the abnormal deformation point is analyzed, and a probability heat map is generated. Based on the probability heat map, the defect location and pressure failure risk level of the spiral weld are output.

[0117] Optionally, step S105 may specifically include the following steps:

[0118] S1051. Perform surface fitting on the spatial structure diagram to generate a three-dimensional shape surface, and simultaneously obtain three-dimensional point cloud data of the weld surface through laser scanning.

[0119] S1052. Calculate the height difference between the three-dimensional shape surface and the three-dimensional point cloud data at each spatial location, and mark the area where the height difference exceeds the threshold as a concave area.

[0120] S1053. Extract the depression depth of each point in the depression area and perform correlation fitting with the deformation amount at the corresponding position in the deformation distribution map to generate a correlation function between depth and deformation amount.

[0121] Step S1053 may specifically include the following processes: extracting the depression depth of each point within the depression region, establishing a spatial coordinate mapping relationship for the depression region, and matching the data points of the depression depth with the deformation data points of the same spatial coordinates in the deformation distribution map; extracting the geometric feature parameters of the depression region, constructing a deformation function for the depression region based on the material properties of the thin-walled structure, the deformation function including a depth square term and a deformation gradient change term; setting the initial weight range of the depth square term and the initial weight range of the deformation gradient change term according to the geometric feature parameters; inputting the matched depression depth data points and deformation data points into a joint equation system, and solving for the optimal weight coefficients of the depth square term and the deformation gradient change term in combination with the initial weight range; and generating a correlation function between depth and deformation based on the optimal weight coefficients.

[0122] S1054. Based on the correlation function, predict the stress response value of the depression region under pressure, and map the stress response value to a probabilistic thermodynamic value.

[0123] S1055. Generate a probability heat map based on the spatial distribution of the probability heat map values, divide the probability heat map into low-risk, medium-risk, and high-risk levels according to the heat value range of the probability heat map, and output the defect location and pressure failure risk level of the spiral weld.

[0124] In the above steps, the spatial structure diagram refers to a digital model representing the geometric morphology of the weld surface generated through 3D modeling. 3D point cloud data is a dense set of points obtained through laser scanning, accurately recording the 3D coordinates of each spatial location on the weld surface. The height difference is the difference in vertical distance between the 3D shaped surface and the point cloud data at the same coordinate point. A recessed area is a local surface deformation region where the height difference exceeds a set threshold. The recess depth refers to the vertical distance between the lowest point of the recessed area and the surrounding normal surface. Deformation is a quantified value of the degree of deformation produced by the material under stress. The correlation function is a fitting equation describing the mathematical relationship between recess depth and deformation. The stress response value predicts the magnitude of internal stress in the recessed area under pressure. The probabilistic thermal value is a risk probability value mapped from 0 to 1 after normalizing the stress response value. The probabilistic thermal map visualizes the spatial distribution of the probabilistic thermal values, using color gradients to represent the level of risk. The risk level is divided into low, medium, and high levels of pressure failure risk based on the thermal value range.

[0125] In this embodiment, firstly, NURBS surface fitting is performed on the spatial structure diagram in step S1051. Specifically, the node vectors are calculated using the Hartley-Judd chord length parameterization algorithm, and a non-uniform node sequence is generated based on the geometric distance between adjacent grid control points to ensure the smoothness of the parameter space mapping. The B-spline basis functions are calculated based on the de Boer-Cox recursive formula, and the NURBS surface equation is constructed by combining the control point weights. ,in, For the coordinates of the control points, As a weighting factor, and They are respectively Direction order and The system employs basis functions; control points are solved using an inverse method, and the shape points of the meshed structure diagram are input into a system of linear equations. The position and weight of the control points are iteratively optimized using a chasing method to generate a continuous and smooth 3D shape surface. Simultaneously, a laser scanner acquires 3D point cloud data of the weld surface with a point spacing of 0.05 mm, ensuring that it shares the same coordinate system as the surface model. For example, the spatial structure diagram of a spiral weld contains 1024×1024 grid points. A smooth surface is generated through NURBS fitting, and the laser scanner simultaneously acquires 500,000 spatial point cloud data points. Both coordinate systems have the weld center as the origin (X0, Y0, Z0).

[0126] Secondly, the height difference of the spatial location is calculated in step S1052, and abnormal areas are marked. Specifically, this is implemented by: for each point cloud coordinate... At the corresponding parameter position of the surface model The interpolation value is the surface height. Calculate the height difference The dynamic threshold is set to 0.3 mm, and the dynamic threshold is based on the material's yield strength. If the critical value of deformation is determined, Millimeters are marked as depressions, and adjacent depressions are integrated to form independent depression regions. For example, a certain point Corresponding surface height millimeters Millimeter-wide exceedances are marked, while of The millimeter limit was not exceeded; the final output consists of 3 independent recessed areas, with a maximum area of ​​8.5 square millimeters.

[0127] Next, a mathematical relationship between the depression depth and the deformation is established through step S1053. First, the spatial coordinates and depth values ​​of each point within the depression area are extracted. Mapping to the deformation distribution map to obtain the deformation amount at the same location. Deformation functions are constructed based on the characteristics of thin-walled structures. ,in, For the deformation gradient, Set as weighting coefficient; Given an initial range of 200 sets of matching data, construct a system of equations. The optimal coefficients are solved using the least squares method, and the objective function is: Solving for , Fitting error For example, region 1 millimeters per millimeter, calculated The value is in millimeters, consistent with the measured deformation of 0.58 millimeters.

[0128] Then, the bearing stress is predicted in step S1054 and converted into a probabilistic thermodynamic value. The indentation depth is then... Input the previously associated function to obtain the deformation amount. Combined with the elastic modulus of the material Gipascal, calculate stress response value ; through normalization formula ,in, It is the calculated stress response value, in megapascals (MPa), representing the magnitude of stress expected to be generated in the depression area under pressure. This is the safety stress threshold, measured in megapascals (MPa). When the stress is below this value, there is no risk; it is set to 50 MPa here. This is the material's yield strength, measured in megapascals (MPa). The material will yield when the stress reaches or exceeds this value; here, it's set to 300 MPa. P is the normalized probabilistic thermodynamic value, ranging from 0 to 1. A P value closer to 0 indicates lower risk, while a value closer to 1 indicates higher risk. This is mapped to a probabilistic thermodynamic value, where... Megapascals is the safety threshold. Megapascals (MPa) is the yield limit. For example, in region 1... At millimeters, Megapascals, then This indicates low risk.

[0129] Finally, a spatial probability heatmap is generated and risk levels are assigned in step S1055. This is achieved by mapping the P-value of each depression point to a color spectrum: blue (P < 0.3), yellow (0.3 ≤ P < 0.7), and red (P ≥ 0.7). Based on the heatmap value range, low-risk (P < 0.3, no treatment required), medium-risk (0.3 ≤ P < 0.7, re-inspection within 6 months), and high-risk (P ≥ 0.7, immediate discontinuation and repair required) are assigned. For example, if the P-values ​​of three depression areas in a weld are 0.275, 0.63, and 0.82, the corresponding output coordinates are: low-risk (101.0, 51.3), medium-risk (102.5, 49.8), and high-risk (103.7, 50.1), along with a corresponding risk level report.

[0130] In practical applications, during the automated quality inspection of spiral welds on thin-walled steel cylinders, the system first performs surface fitting on the spatial structure diagram to generate a three-dimensional shape surface. Simultaneously, a line laser scanner acquires three-dimensional point cloud data of the weld surface with a precision of 0.05 mm. Then, the system calculates the millimeter-level height difference between the three-dimensional shape surface and the point cloud data at each spatial location. When the height difference at a certain coordinate point exceeds a set threshold of 0.3 mm, for example, if the height difference at coordinate position (-15.2, 28.7) is detected to be -0.45 mm, the area is marked as a concave region. Next, the concavity depth data of each point within the concave region is extracted, and a mapping relationship between this depth and the deformation amount at the same spatial coordinates in the deformation distribution diagram is established. Based on the material properties of thin-walled structures, a joint equation system is constructed, including a depth square term characterizing the influence of depth and a gradient change term reflecting the deformation gradient, with initial weight ranges set to -10 to 10 and 0.1 to 2, respectively. Taking a certain detection point as an example: a concavity depth of -0.45 mm corresponds to a deformation of 680 micro-strain; such matching data points are input into the equation system. The optimal weights were determined using the least squares method, yielding a depth squared term coefficient of -2.3 and a gradient term coefficient of 1.1, thus forming the correlation function. Where d represents the depth of the depression, This represents the change in deformation per unit distance. Based on this function, the stress response under pressure conditions is predicted. For example, when a depression area is subjected to a pressure of 20 MPa, the calculated local stress concentration factor reaches 1.8. This value is converted into a probabilistic thermal value ranging from 0 to 1; for example, a stress value of 80 MPa is mapped to a thermal value of 0.92. A probabilistic thermal map is generated through gridding mapping, where thermal values ​​less than 0.3 are displayed in blue indicating low-risk areas, 0.3 to 0.7 in yellow indicating medium-risk areas, and greater than 0.7 in red indicating high-risk areas. For example, three high-risk coordinates were detected: (-15.2, 28.7) thermal value 0.92, (7.8, -12.3) thermal value 0.87, and (-3.4, 5.6) thermal value 0.78. Finally, a spiral weld defect location map and risk level report are output, achieving a visualized early warning of pressure failure risk.

[0131] In the overall scheme of step S105 above, sub-millimeter-level synchronous detection of weld penetration quantification and thermal stress cracks is achieved. The three-dimensional dynamic structure of the weld is accurately reconstructed through multispectral dynamic exposure control and acoustic-optical joint stress field analysis. Based on the depth comparison between the spatial structure map and the three-dimensional point cloud data, areas with insufficient weld penetration are identified. The dynamic coupling relationship between the indentation depth and the stress response is decoupled through the deformation-defect correlation function. A probabilistic heat map is constructed in combination with the characteristics of thin-walled structures, and the defect location and three levels of low / medium / high pressure failure risk of the spiral weld are directly output, realizing millisecond-level closed-loop prediction of welding defects and structural risks.

[0132] The following is a complete embodiment for steps S101 to S105:

[0133] like Figure 3 As shown, in a large-scale water pipeline welding inspection project, an intelligent inspection system was deployed to monitor the quality of spiral welds in real time. This system first simultaneously collects light energy data and multi-source images in the weld pool area: using three sets of universal light sources with wavelengths of 450nm blue light, 760nm infrared, and 850nm near-infrared, a photoelectric sensor records the radiation energy values ​​in the 380-2500nm spectral range at a frequency of 1000 times per second; simultaneously, a high-speed industrial camera captures images of the weld pool at a rate of 200 frames per second. When the sensor detects that the energy ratio of the visible light band (380-780nm) to the near-infrared band (780-2500nm) exceeds 2.0 (e.g., a measured value of 2.8), the system automatically reduces the exposure time of the blue light source from 0.001 seconds to 0.0003 seconds and increases the gain of the near-infrared source by 20dB, effectively suppressing interference from strong arc light and generating a weld surface image with 40% improved clarity at the weld pool boundary.

[0134] Subsequently, a 632.8 nm coherent laser beam was projected into the weld hot deformation zone, and dynamic speckle image sequences and acoustic frequency signals were simultaneously acquired during the welding process. By comparing the initial speckle pattern, the displacement change Δd = 1.8 μm at a certain coordinate point (x, y) = [15.2, 28.7] was calculated, and an acoustic frequency shift Δf = 22 kHz was detected. Based on the parameters of the thin-walled steel cylinder (elastic modulus 210 GPa, Poisson's ratio 0.3), the deformation calculation formula was applied: ε =1−0.321.8×10−6×210×109+0.15×22000=423 με When the deformation gradient difference between adjacent 1 mm² regions exceeds 200 με / mm, the region is marked as a stress concentration area. After inputting the weld image and stress concentration area deformation data into a spatiotemporal convolutional neural network, the system automatically extracts: the energy path formed by the peak fluctuation of light energy within 5 seconds; the light and dark boundaries generated by edge detection of the image; and the connection relationship of deformation network nodes, with a spacing ≤ 3 mm and a gradient difference ≥ 200 με / mm. The spatial correlation between the energy path and the light and dark boundaries is calculated through a cross-attention mechanism, and a weight w = 1 / (1 + 0.5²) = 0.8 is assigned when the distance between the two is 0.5 mm. After generating a three-dimensional structure map through multi-scale mesh reconstruction, the central angle θ = 0.5 radians is calculated based on the pipe curvature radius of 1.2 meters and the weld arc length of 0.6 meters. Coordinate compensation is performed for the sensor position deviation of 3 mm: new coordinates .

[0135] Finally, using line laser scanning with a precision of 0.05 mm, recessed areas with height differences > 0.3 mm were marked on the spatial structure map. Based on the correlation function: The stress concentration factor at this location is predicted to be 1.8 under a pressure of 20 MPa, corresponding to a thermal value P = 0.92. In the generated probabilistic heatmap, areas with thermal values ​​≥ 0.7 (marked in red) are identified as high-risk defect areas. A risk level report containing five high-risk coordinate points is output, providing accurate early warning for pipeline pressure safety.

[0136] The intelligent real-time detection method for PCCP welding quality provided in this application adjusts exposure parameters in real time through a multispectral dynamic exposure control mechanism, and synchronously preserves the surface texture and deep penetration characteristics of the weld pool by utilizing the energy difference between visible light and near-infrared light. It combines coherent beam speckle displacement and acoustic offset analysis techniques, and integrates the elastic parameters of the thin-walled steel cylinder to achieve millisecond-level localization of stress concentration areas. Based on a spatiotemporal convolutional neural network that integrates multi-source sensor data, it dynamically weights energy paths, light and dark boundaries, and deformation node features through a cross-attention mechanism, and combines the real-time compensation of mechanical deviations by the steel cylinder's curling curvature to construct a sub-millimeter-level precision three-dimensional dynamic structural map of the weld. Finally, it compares the structural map with point cloud data to identify areas with insufficient weld penetration. By decoupling the coupling relationship between depth and stress through a deformation-defect correlation function, it generates a probability heatmap to directly output defect location and low / medium / high three-order pressure failure risk levels, forming a millisecond-level early warning closed loop. This significantly improves the detection rate of welding defects while reducing the false alarm rate, achieving non-destructive, synchronous, and accurate full-field detection and risk prediction of spiral weld penetration and thermal stress cracks.

[0137] Figure 4 This is a schematic diagram of a specific embodiment of a PCCP welding quality intelligent real-time detection system provided in this application, referring to... Figure 4 The system may include:

[0138] Acquisition module 41 is used to simultaneously acquire light energy data and multi-source images of the spiral weld molten pool area;

[0139] The generation module 42 is used to dynamically adjust the exposure parameters of the multi-source image based on the energy difference between the molten pool region in the visible light band and the near-infrared band, so as to suppress the interference of overly bright areas caused by strong light during welding and generate a weld surface image with enhanced weld depth characteristics.

[0140] The positioning module 43 is used to scan the thermal deformation area of ​​the weld based on the emitted coherent beam, capture the changes in speckle pattern and sound frequency generated when the weld is deformed by heat, calculate the deformation distribution of the weld in combination with the elastic characteristics of the thin-walled steel cylinder, and locate the stress concentration area according to the deformation distribution.

[0141] The fusion module 44 is used to input the weld surface image and the deformation data of the stress concentration area into the spatiotemporal convolutional neural network, fuse the variation characteristics of the light energy data, the brightness and darkness details of the multi-source image and the spatial characteristics of the deformation distribution, construct the spatial structure map of the weld, and adaptively correct the position deviation of the sensor based on the degree of bending of the steel cylinder.

[0142] Output module 45 is used to compare the shape data of the spatial structure diagram with the weld depression depth distribution of the three-dimensional point cloud reconstruction, identify depression areas with insufficient penetration, analyze the correlation between the depression degree of the depression area and the stress response at the deformation abnormality, generate a probability heat map, and output the defect location and pressure failure risk level of the spiral weld based on the probability heat map.

[0143] The intelligent real-time detection system for PCCP welding quality in this application is used to implement the aforementioned intelligent real-time detection method for PCCP welding quality. Therefore, the specific implementation of the intelligent real-time detection system for PCCP welding quality can be found in the embodiment section of the intelligent real-time detection method for PCCP welding quality above. The specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0144] This application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of any of the above-described intelligent real-time detection methods for PCCP welding quality.

[0145] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described intelligent real-time detection methods for PCCP welding quality.

[0146] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0147] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the intelligent real-time detection method for PCCP welding quality described above.

[0148] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0149] The above provides a detailed description of the intelligent real-time detection method and system for PCCP welding quality provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for intelligent real-time detection of PCCP welding quality, characterized in that, include: Simultaneously acquire light energy data and multi-source images of the spiral weld seam molten pool area; Based on the energy difference between the visible and near-infrared bands in the molten pool region, the exposure parameters of the multi-source image are dynamically adjusted to suppress interference from overly bright areas caused by strong light during welding, thereby generating a weld surface image with enhanced penetration characteristics. Based on the scanning of the weld's thermal deformation region by the emitted coherent beam, the changes in speckle pattern and sound frequency generated during the weld's thermal deformation are captured. Combined with the elastic characteristics of the thin-walled steel cylinder, the deformation distribution of the weld is calculated, and the stress concentration area is located based on the deformation distribution. The surface image of the weld and the deformation data of the stress concentration area are input into a spatiotemporal convolutional neural network. The variation characteristics of the light energy data, the brightness and darkness details of the multi-source image and the spatial characteristics of the deformation distribution are fused to construct a spatial structure map of the weld. The position deviation of the sensor is adaptively corrected based on the degree of bending of the steel cylinder. By comparing the shape data of the spatial structure diagram with the weld depression depth distribution reconstructed from the three-dimensional point cloud, depression areas with insufficient penetration are identified. The correlation between the degree of depression in the depression area and the stress response at the abnormal deformation site is analyzed, and a probability heat map is generated. Based on the probability heat map, the defect location and pressure failure risk level of the spiral weld are output.

2. The method according to claim 1, characterized in that, The weld surface image and deformation data of the stress concentration region are input into a spatiotemporal convolutional neural network. The network fuses the variation characteristics of the light energy data, the brightness details of the multi-source image, and the spatial features of the deformation distribution to construct a spatial structure map of the weld. Based on the degree of bending of the rolled steel cylinder, the sensor's positional deviation is adaptively corrected, including: The weld surface image and the deformation data of the stress concentration area are aligned according to spatial coordinates to form spatiotemporally correlated input data. In a spatiotemporal convolutional neural network, a multi-branch feature extraction path is set up. The first branch extracts the trajectory of the change of the light energy data over time, the second branch extracts the light and dark contrast features of the weld surface image, and the third branch extracts the spatial topology of the deformation distribution. By integrating the change trajectory, the light and dark contrast features, and the spatial topology through the cross-branch feature interaction module, an initial spatial structure diagram of the weld is generated. Obtain the radius of curvature of the rolled steel cylinder, calculate the offset of the sensor installation position relative to the weld centerline, and generate a position compensation vector based on the offset. Based on the position compensation vector, the spatial coordinates of each pixel in the initial spatial structure map are corrected, and the corrected spatial structure map is output.

3. The method according to claim 2, characterized in that, By integrating the change trajectory, the light and dark contrast features, and the spatial topology through the cross-branch feature interaction module, an initial spatial structure diagram of the weld is generated, including: In the cross-branch feature interaction module, the peak positions of energy fluctuations at continuous time points of the change trajectory are extracted to generate energy fluctuation path lines. Based on the light and dark contrast features, gray-level transition boundaries are identified to generate light and dark boundary contours. The node connection relationships are analyzed according to the spatial topology to construct a deformation association network. The deformation association network uses stress concentration points as nodes and deformation gradients as edge weights. The correlation strength between the energy fluctuation path and the light-dark boundary contour is calculated based on the cross-attention mechanism, and the spatial correlation between the nodes in the deformable correlation network and the energy fluctuation path is calculated at the same time. Based on the correlation strength and the spatial correlation, the boundary weights of the light-dark boundary contour and the node weights of the deformable correlation network are dynamically adjusted. The weighted light-dark boundary contour and the deformation association network are superimposed on the energy fluctuation path line to generate a fused feature map. Multi-scale mesh reconstruction is performed on the fused feature map, wherein the fine mesh level retains high-frequency details and the coarse mesh level integrates low-frequency contours. The reconstructed fused feature map is compensated for the feature offset caused by the change in grid scale, and the compensated three-dimensional grid data is output as the initial spatial structure map of the weld.

4. The method according to claim 1, characterized in that, By comparing the shape data of the spatial structure diagram with the weld depression depth distribution reconstructed from the 3D point cloud, areas with insufficient penetration are identified. The correlation between the degree of depression in these areas and the stress response at abnormal deformation sites is analyzed, generating a probability heatmap. Based on the probability heatmap, the defect location and pressure failure risk level of the spiral weld are output, including: The spatial structure diagram is fitted with a surface to generate a three-dimensional shape surface, and three-dimensional point cloud data of the weld surface is obtained by laser scanning. Calculate the height difference between the three-dimensional shape surface and the three-dimensional point cloud data at each spatial location, and mark the region where the height difference exceeds a threshold as a concave region; Extract the depression depth of each point within the depression region and correlate it with the deformation amount at the corresponding position in the deformation distribution map to generate a correlation function between depth and deformation amount; Based on the correlation function, the stress response value of the depression region under pressure is predicted, and the stress response value is mapped to a probabilistic thermodynamic value. A probability heat map is generated based on the spatial distribution of the probability heat map values. The heat map values ​​are divided into low-risk, medium-risk, and high-risk levels, and the defect location and pressure failure risk level of the spiral weld are output.

5. The method according to claim 4, characterized in that, Extract the depression depth at each point within the depression region and correlate it with the deformation amount at the corresponding position in the deformation distribution map to generate a correlation function between depth and deformation amount, including: Extract the depression depth of each point within the depression region, establish the spatial coordinate mapping relationship of the depression region, and match the data points of the depression depth with the deformation data points of the same spatial coordinates in the deformation distribution map. Extract the geometric feature parameters of the recessed region, and construct a deformation function for the recessed region based on the material properties of the thin-walled structure. The deformation function includes a depth square term and a deformation gradient term. Based on the geometric feature parameters, set the initial weight range of the depth square term and the initial weight range of the deformation gradient change term; The matching data points of the indentation depth and deformation amount are input into the joint equation system. Combined with the initial weight range, the optimal weight coefficients of the depth square term and the deformation amount gradient change term are solved. Based on the optimal weight coefficients, the correlation function between depth and deformation amount is generated.

6. The method according to claim 1, characterized in that, Based on the energy difference between the molten pool region in the visible and near-infrared bands, the exposure parameters of the multi-source image are dynamically adjusted to suppress interference from overly bright areas caused by strong light during welding, generating a weld surface image with enhanced weld penetration characteristics, including: The first energy value in the visible light band and the second energy value in the near-infrared band of the molten pool region are acquired simultaneously by a dual-channel sensor, and the ratio of the first energy value to the second energy value is calculated. Based on the preset range in which the ratio falls, a corresponding exposure compensation curve is selected, wherein the preset range is related to the energy difference range of the melt pool region. Based on the exposure compensation curve, the exposure time and gain parameters of each independent light source in the multi-light source image are adjusted to reduce the exposure parameters of overly bright areas and increase the exposure parameters of dark areas. The adjusted multi-source image is fused at the pixel level to generate a single-frame fused image, wherein the texture details of the visible light band and the melting depth features of the near-infrared band are preserved during the fusion process. The high-frequency components of the single-frame fused image are separated by a multi-scale decomposition method. The high-frequency components are then recombined after feature enhancement to generate a weld surface image with enhanced weld depth features.

7. The method according to claim 1, characterized in that, Based on the scanning of the weld's thermal deformation region using the emitted coherent beam, the changes in speckle pattern and sound frequency generated during weld thermal deformation are captured. Combined with the elastic characteristics of the thin-walled steel cylinder, the deformation distribution of the weld is calculated, and stress concentration areas are located based on this deformation distribution, including: A coherent beam of light is projected onto the weld surface to form an initial speckle pattern, and dynamic speckle image sequences and sound frequency signals during thermal deformation are acquired simultaneously. The dynamic speckle image sequence is compared with the initial speckle pattern to extract the displacement change of each pixel from the initial state to the dynamic state, and the offset of the sound frequency signal relative to the reference frequency is calculated. By combining the elastic parameters of the thin-walled steel cylinder, a joint function of the displacement change and offset is constructed. Based on the joint function, the deformation of each local area on the weld surface is calculated, and a deformation distribution map is generated based on the deformation. Based on the difference in deformation gradient between adjacent regions in the deformation distribution map, the locations of gradient abrupt changes are identified and marked as stress concentration areas.

8. A PCCP welding quality intelligent real-time detection system, characterized in that, include: The acquisition module is used to simultaneously acquire light energy data and multi-source images of the weld pool area of ​​the spiral weld. The generation module is used to dynamically adjust the exposure parameters of the multi-source image based on the energy difference between the molten pool region in the visible light band and the near-infrared band, so as to suppress the interference of overly bright areas caused by strong light during welding and generate a weld surface image with enhanced penetration characteristics. The positioning module is used to scan the thermal deformation area of ​​the weld seam based on the emitted coherent beam, capture the changes in speckle pattern and sound frequency generated when the weld seam is deformed by heat, and calculate the deformation distribution of the weld seam based on the elastic characteristics of the thin-walled steel cylinder, and locate the stress concentration area according to the deformation distribution. The fusion module is used to input the weld surface image and the deformation data of the stress concentration area into a spatiotemporal convolutional neural network, fuse the variation characteristics of the light energy data, the brightness and darkness details of the multi-source image and the spatial characteristics of the deformation distribution, construct a spatial structure map of the weld, and adaptively correct the position deviation of the sensor based on the degree of bending of the steel cylinder. The output module is used to compare the shape data of the spatial structure diagram with the weld depression depth distribution reconstructed from the three-dimensional point cloud, identify depression areas with insufficient penetration, analyze the correlation between the depression degree of the depression area and the stress response at the abnormal deformation site, generate a probability heat map, and output the defect location and pressure failure risk level of the spiral weld based on the probability heat map.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the steps of the intelligent real-time detection method for PCCP welding quality as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the intelligent real-time detection method for PCCP welding quality as described in any one of claims 1 to 7.

Images