Welding seam detection system, method and device based on ray back scattering

Abstract

The invention provides a welding seam detection system, method and device based on ray back scattering, and the system comprises a machine vision collection positioning module which is used for obtaining the original image data of the surface of a welding seam, recognizing the center line of the welding seam through an image processing algorithm, and outputting the space positioning information of a welding seam region; the ray back scattering detection module is in communication connection with the machine vision acquisition and positioning module, and is used for emitting rays to the welding seam area based on the spatial positioning information, acquiring back scattering signals and converting the back scattering signals into a two-dimensional grayscale image representing the internal density distribution of the welding seam; the multi-modal data fusion processing unit is used for carrying out space registration and feature extraction on the original image data and the two-dimensional grayscale image, constructing a unified feature vector and outputting a weld defect identification result through a preset defect identification model; and the moving mechanism is used for bearing the machine vision collecting and positioning module and the ray back scattering detection module.

Description

Technical Field

[0001] This application relates to the field of weld inspection technology, and in particular to a weld inspection system, method and inspection device based on X-ray backscattering. Background Technology

[0002] With the rapid development of large-scale equipment manufacturing and heavy structural engineering, welding technology has been widely applied in shipbuilding, marine engineering equipment, tunnel boring machines, large steel structures, and energy equipment. The welding quality of thick plates and multi-layer welds directly affects the overall strength, fatigue life, and service safety of critical structures. However, during the welding process, factors such as unstable heat input, changes in welding posture, and differences in material microstructure can easily lead to various defects such as cracks, porosity, slag inclusions, and lack of fusion inside and on the surface of the weld. If these defects are not detected in time, they often gradually expand during subsequent service, potentially causing structural failure or even major safety accidents. Therefore, how to continuously, accurately, and in real-time inspect the weld quality during welding operations has become a critical technical problem that urgently needs to be solved in the welding manufacturing field.

[0003] In related technologies, weld quality inspection mainly relies on post-weld non-destructive testing methods such as ultrasonic testing (UT) and radiographic testing (RT). These methods are typically offline and suffer from drawbacks such as detection lag, inability to provide real-time feedback on welding status, and high rework costs. Furthermore, traditional radiographic testing requires the placement of X-ray sources and receiving devices on both sides of the weld, which is often difficult to implement in large components or complex structures due to the inaccessibility of the back side. Ultrasonic testing, on the other hand, is highly dependent on coupling conditions and operational experience, and its signal-to-noise ratio drops significantly in multi-layered and ultra-thick weld scenarios, resulting in insufficient stability of the test results. In addition, existing online inspection schemes based on single visual or single physical field information often only acquire surface morphology or local features of the weld, failing to simultaneously reflect internal defects and appearance quality, leading to biased test results that cannot meet the needs for online, comprehensive, and intelligent assessment of weld quality in complex welding scenarios. Summary of the Invention

[0004] In view of this, it is necessary to provide a weld inspection system, method and inspection device based on X-ray backscattering that can overcome at least one of the above defects.

[0005] In a first aspect, embodiments of this application provide a weld inspection system based on X-ray backscattering, applied to weld inspection during the welding process, the system comprising: The machine vision acquisition and positioning module is used to acquire raw image data of the weld surface, identify the weld centerline through image processing algorithms, and output the spatial positioning information of the weld area. The X-ray backscattering detection module is communicatively connected to the machine vision acquisition and positioning module. It is used to emit X-rays to the weld area based on the spatial positioning information and acquire backscattering signals, and convert the backscattering signals into a two-dimensional grayscale image characterizing the density distribution inside the weld. The multimodal data fusion processing unit is used to perform spatial registration and feature extraction on the original image data and the two-dimensional grayscale image, construct a unified feature vector, and output the identification result of weld defects through a preset defect identification model. A mobile mechanism is used to carry the machine vision acquisition and positioning module and the ray backscattering detection module. The mobile mechanism is equipped with a synchronization trigger controller and a position sensor. The synchronization trigger controller is used to send a synchronization pulse signal to the vision module and the ray backscattering detection module according to the position signal fed back by the position sensor, so that the original image data and the two-dimensional grayscale image have corresponding timestamps. The real-time feedback unit is communicatively connected to the multimodal data fusion processing unit. It is used to generate a weld quality database containing defect type, location, and size parameters based on the identification results, and to send control commands to the external welding execution mechanism based on the evaluation value generated according to the severity of the defects.

[0006] In one embodiment, the machine vision acquisition and positioning module is equipped with a camera, which is used to acquire multi-exposure continuous images of the weld surface and the molten pool area during the welding process. The image processing algorithm uses a pyramid hierarchical search strategy to fuse multi-exposure images to eliminate welding arc interference and identify the weld centerline. The machine vision acquisition and positioning module transmits the identified spatial positioning information to the X-ray backscattering detection module in real time to constrain and guide the X-ray projection axis of the X-ray backscattering detection module to always be aligned with the center of the weld.

[0007] In one embodiment, the X-ray backscattering detection module includes a X-ray generating unit and a backscattering signal detector array; The rays generated by the ray generating unit are shaped into pencil beams by a collimator and enter the weld, and the pencil beams reciprocate along the cross-sectional direction of the weld. The backscatter signal detector array is symmetrically arranged on both sides of the ray generating unit to receive photon signals scattered back from inside the weld and convert them into analog electrical signals. The two-dimensional grayscale image is generated through signal harmonic processing, wherein the grayscale value of the two-dimensional grayscale image is linearly correlated with the density of detection points inside the workpiece.

[0008] In one embodiment, the multimodal data fusion processing unit includes a spatial registration module, a multimodal feature extraction module, and a fusion inference module; The spatial registration module is used to map the two-dimensional grayscale image and the original image data to a unified pixel coordinate system based on the position information of the moving mechanism and the sensor installation geometry. The multimodal feature extraction module is also used to extract density abrupt change features and surface geometric texture features inside the weld. The fusion inference module uses a dual-branch convolutional neural network to cascade and fuse the density abrupt change features inside the weld with the surface geometric texture features to construct a fusion feature vector, and then compares it with a preset defect feature library to output the recognition result.

[0009] In one embodiment, the moving mechanism is an automated guided vehicle (AGV), which is equipped with a high-precision odometer, an inertial measurement unit, and a track positioning sensor to achieve autonomous navigation and attitude compensation of the detected path. The chassis of the automated guided vehicle is equipped with a vibration damping platform, and the machine vision acquisition and positioning module and the X-ray backscatter detection module are wrapped with a protective cover with an integrated active circulation cooling device. The real-time feedback unit is connected to the fusion reasoning module and is used to generate a digital quality profile of the weld seam based on the identification results output by the multimodal data fusion processing unit, and to send compensation instructions for welding current, voltage or speed to the external welding execution mechanism based on the evaluation value generated according to the severity of the defects.

[0010] In one embodiment, the multimodal data fusion processing unit further includes a dynamic gain compensation module: The dynamic gain compensation module is used to adjust the signal receiving gain and integration time of the X-ray backscattering detection module according to the workpiece surface roughness and weld groove geometric parameters obtained by the machine vision acquisition and positioning module, so as to compensate for the fluctuation of the signal-to-noise ratio of the two-dimensional grayscale image caused by the change of workpiece thickness or the difference of surface scattering rate.

[0011] In one embodiment, the system further includes an adaptive scan deflection controller: The adaptive scanning deflection controller is communicatively connected to the machine vision acquisition and positioning module. It is used to calculate the scanning path of the pencil beam in real time based on the curvature and offset of the identified weld centerline in the spatial coordinate system, and drive the collimator to perform asymmetric angle deflection to ensure that the scanning center of the pencil beam always coincides with the actual physical centerline of the weld.

[0012] In one embodiment, the system further includes an online dynamic calibration unit; The online dynamic calibration unit is used to acquire the reference backscatter signal collected by the moving mechanism in the non-weld seam area, and to correct the detection sensitivity of the X-ray backscatter detection module in real time by combining the data fed back by the current ambient temperature sensor, so as to eliminate image grayscale drift caused by workpiece background thickness fluctuation and temperature drift.

[0013] Secondly, embodiments of this application provide a weld inspection method based on X-ray backscattering, applied to the weld inspection system based on X-ray backscattering as described in the first aspect, the method comprising: The original image data of the weld surface is acquired, and the weld centerline is identified through image processing algorithms and the spatial positioning information of the weld area is output. Based on the spatial positioning information, rays are emitted into the weld area and backscattered signals are collected. The backscattered signals are then converted into a two-dimensional grayscale image characterizing the density distribution inside the weld. Spatial registration and feature extraction are performed on the original image data and the two-dimensional grayscale image to construct a unified feature vector, and the identification result of weld defects is output through a preset defect identification model. A synchronization pulse signal is sent according to the position signal so that the original image data and the two-dimensional grayscale image have corresponding timestamps; Based on the identification results, a weld quality database containing defect type, location, and size parameters is generated, and control commands are sent to external welding execution agencies based on the evaluation value generated according to the severity of the defects.

[0014] Thirdly, embodiments of this application provide a detection device, comprising: Processor; and A memory having computer-readable instructions stored thereon for controlling the processor to execute the X-ray backscattering-based weld inspection method as described in the second aspect.

[0015] This application provides a weld inspection system, method, and device based on backscattering radiation, which can simultaneously acquire, time / space register, multimodal fusion recognition, and real-time feedback of internal density information and surface morphology information of welds during welding operations. By using a machine vision module to precisely locate the weld centerline and guide the projection of backscattered radiation, combined with pencil collimation and a backscatter detector array for highly sensitive reception of scattered photons, this system can obtain a two-dimensional grayscale image reflecting internal defects from a single side. Furthermore, this internal information is registered with a high dynamic range multi-exposure visual image in a unified pixel coordinate system and fused with features at the convolutional neural network level to achieve joint identification, location, and quantitative assessment of internal and surface defects such as cracks, porosity, inclusions, and lack of fusion. Attached Figure Description

[0016] Figure 1This is a schematic diagram of a weld inspection system based on X-ray backscattering provided in one embodiment of this application.

[0017] Figure 2 This is a schematic diagram of a backscattering module provided in one embodiment of this application.

[0018] Figure 3 This is a schematic flowchart of a weld inspection method based on X-ray backscattering provided in an embodiment of this application.

[0019] Figure 4 This is a schematic diagram of a detection device provided in one embodiment of this application.

[0020] Explanation of main component symbols Weld inspection system based on X-ray backscattering 10 Machine vision acquisition and positioning module 11 X-ray backscattering detection module 12 Multimodal data fusion processing unit 13 Mobile agency 14 Real-time feedback unit 15 Detection device 20 Processor 21 Memory 22 Method steps S100-S500 Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0022] It should be noted that, in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0023] It should be noted that in the embodiments of this application, the terms "first," "second," etc., are used only for descriptive purposes and should not be construed as indicating or implying relative importance, nor as indicating or implying order. Features specified as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0024] Based on the embodiments described 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.

[0025] With the rapid development of large-scale equipment manufacturing and heavy structural engineering, welding technology has been widely applied in shipbuilding, marine engineering equipment, tunnel boring machines, large steel structures, and energy equipment. The welding quality of thick plates and multi-layer welds directly affects the overall strength, fatigue life, and service safety of critical structures. However, during the welding process, factors such as unstable heat input, changes in welding posture, and differences in material microstructure can easily lead to various defects such as cracks, porosity, slag inclusions, and lack of fusion inside and on the surface of the weld. If these defects are not detected in time, they often gradually expand during subsequent service, potentially causing structural failure or even major safety accidents. Therefore, how to continuously, accurately, and in real-time inspect the weld quality during welding operations has become a critical technical problem that urgently needs to be solved in the welding manufacturing field.

[0026] In related technologies, weld quality inspection mainly relies on post-weld non-destructive testing methods such as ultrasonic testing (UT) and radiographic testing (RT). These methods are typically offline and suffer from drawbacks such as detection lag, inability to provide real-time feedback on welding status, and high rework costs. Furthermore, traditional radiographic testing requires the placement of X-ray sources and receiving devices on both sides of the weld, which is often difficult to implement in large components or complex structures due to the inaccessibility of the back side. Ultrasonic testing, on the other hand, is highly dependent on coupling conditions and operational experience, and its signal-to-noise ratio drops significantly in multi-layered and ultra-thick weld scenarios, resulting in insufficient stability of the test results. In addition, existing online inspection schemes based on single visual or single physical field information often only acquire surface morphology or local features of the weld, failing to simultaneously reflect internal defects and appearance quality, leading to biased test results that cannot meet the needs for online, comprehensive, and intelligent assessment of weld quality in complex welding scenarios.

[0027] In view of this, the weld inspection system, method, and device based on X-ray backscattering provided in this application can realize the synchronous acquisition, temporal / spatial registration, multimodal fusion recognition, and real-time feedback of weld internal density information and surface morphology information during the welding operation. By using a machine vision module to precisely locate the weld centerline and guide the projection of backscattered rays, combined with pencil collimation and a backscatter detector array for highly sensitive reception of scattered photons, this system can obtain a two-dimensional grayscale image reflecting internal defects from a single side. Furthermore, this internal information is registered with a high dynamic range multi-exposure visual image in a unified pixel coordinate system and fused with features at the convolutional neural network level to achieve joint identification, location, and quantitative assessment of internal and surface defects such as cracks, porosity, inclusions, and lack of fusion.

[0028] Figure 1 This is a schematic diagram of a weld inspection system based on X-ray backscattering provided in an embodiment of this application. Figure 1 The weld inspection system 10 based on X-ray backscattering shown includes at least the following components: a machine vision acquisition and positioning module 11, a X-ray backscattering detection module 12, a multimodal data fusion processing unit 13, a moving mechanism 14, and a real-time feedback unit 15.

[0029] In this embodiment of the application, the machine vision acquisition and positioning module 11 is used to acquire the original image data of the weld surface, and to identify the weld centerline and output the spatial positioning information of the weld area through the image processing algorithm.

[0030] Specifically, the machine vision acquisition and positioning module 11 includes a high dynamic range industrial camera (CCD or CMOS), a lens, a directional light source, and an image acquisition card. During the welding process, the industrial camera continuously acquires images of the weld surface and the molten pool area in multi-exposure or high frame rate mode. The image acquisition is triggered by hardware and the synchronization interface is constrained by the synchronization trigger controller. The image processing algorithm includes a pyramid hierarchical search strategy and Hu invariant moment template matching for coarse-fine positioning of the weld area. Subsequently, fine-grained positioning is performed on the region of interest (ROI) to output spatial positioning information such as the weld centerline coordinates, weld contour, and weld groove geometric parameters. This module further performs distortion correction, exposure fusion, illumination compensation, and image enhancement on the original image to provide high-quality visual input for subsequent registration and feature extraction.

[0031] Understandably, the machine vision acquisition and positioning module 11, by providing accurate spatial positioning information and high-quality surface images, can constrain the ray projection area of ​​the backscattering device, reduce ineffective radiation and imaging errors, and significantly reduce the backend computation and improve the overall detection speed through ROI pre-clipping and pyramid positioning, thereby meeting the dual requirements of real-time performance and robustness in online "welding and testing" scenarios.

[0032] In this embodiment, the X-ray backscatter detection module 12 is communicatively connected to the machine vision acquisition and positioning module 11, and is used to emit X-rays to the weld area based on spatial positioning information and acquire backscatter signals, and convert the backscatter signals into a two-dimensional grayscale image characterizing the density distribution inside the weld.

[0033] Specifically, the X-ray backscatter detection module 12 includes a X-ray generating unit (such as an X-ray tube and a high-voltage power supply unit), a collimator, a backscatter signal detector array, a front-end signal conditioning and analog-to-digital conversion unit, a filtering and image reconstruction processing unit, and necessary cooling and shielding structures. The X-ray generating unit is shaped into a pencil beam by the collimator and performs reciprocating or trajectory scanning in the weld cross-section direction according to the center line and ROI provided by the vision module. The backscatter detector array converts the received rebound photons into photoelectric signals, which are amplified by photomultiplier or equivalent detector and then sampled by the analog-to-digital converter. The front-end signal processing includes pulse shaping, energy window segmentation, and noise suppression. Subsequently, a two-dimensional grayscale image reflecting the local electron density / matter density difference is generated through stitching and reconstruction algorithms. This module also supports dynamic gain, integral time adjustment, and online calibration based on non-weld areas to maintain image consistency.

[0034] Understandably, by tightly coupling X-ray backscatter detection with the spatial positioning of machine vision, this module can acquire internal density distribution information under conditions of approaching the weld surface on one side, overcoming the limitations of traditional radiographic RT in the case of inaccessible back side. At the same time, the design of collimated pencil beam and detector array improves spatial resolution and acquisition signal-to-noise ratio, so that density abrupt features such as cracks, pores, and slag inclusions can be effectively displayed in two-dimensional grayscale images, providing reliable internal features for subsequent multimodal recognition.

[0035] In this embodiment, the multimodal data fusion processing unit 13 is used to perform spatial registration and feature extraction on the original image data and the two-dimensional grayscale image, construct a unified feature vector, and output the identification result of weld defects through a preset defect identification model.

[0036] Specifically, the multimodal data fusion processing unit 13 includes a time synchronization and timestamp alignment module, a spatial registration module, a multimodal feature extraction module, a dynamic gain compensation module, a fusion inference module, and a result post-processing module. The time synchronization module assigns a unified timestamp to the visual frame and the backscatter frame based on the trigger signal of the moving mechanism and the time encoder, and performs delay compensation. The spatial registration module maps the two types of images to a unified pixel / physical coordinate system based on the pose of the moving mechanism, the sensor mounting geometry, and the calibration matrix. The feature extraction module uses a convolutional neural network (CNN) or its variants to extract the density change features of the backscatter image and the geometric, texture, and edge features of the visual image, respectively. The fusion inference module uses a dual-branch or multi-channel deep fusion network to cascade, attention-weighted, or feature-aligned fusion of the extracted features, and outputs the defect category, pixel or physical coordinate position, size estimate, and confidence level. The result post-processing module performs grade determination, pose mapping, visualization annotation, and encapsulation into a structured record for the recognition results.

[0037] Understandably, this multimodal data fusion processing unit, through spatiotemporal alignment and deep learning-driven cross-modal feature fusion, can compensate for the blind spots and noise sensitivity of single sensor detection, thereby improving the accuracy and robustness of defect detection. At the same time, the modular design facilitates online parameter adjustment (such as dynamic gain and integration time) and model upgrades, ensuring that the system can maintain stable recognition performance under different materials, thicknesses, and complex working conditions.

[0038] In this embodiment, the moving mechanism 14 is used to carry the machine vision acquisition and positioning module 11 and the ray backscatter detection module 12. The moving mechanism 14 is equipped with a synchronization trigger controller and a position sensor. The synchronization trigger controller is used to send a synchronization pulse signal to the vision module and the ray backscatter detection module according to the position signal fed back by the position sensor, so that the original image data and the two-dimensional grayscale image have corresponding timestamps.

[0039] Specifically, the mobile mechanism 14 is preferably an automated guided vehicle (AGV) or a mobile platform with track positioning function, equipped with a high-precision odometer, encoder, inertial measurement unit (IMU) and track / landmark positioning sensor for real-time acquisition of position and attitude information; the chassis of the mobile mechanism 14 is equipped with a vibration-damping mounting platform to reduce the impact of mechanical vibration on the acquisition accuracy, and the outer shell is an electromagnetic shield and protective cooling cover to resist welding arc light, high temperature and spatter; the synchronous trigger controller, in conjunction with the position sensor, sends hardware trigger pulses to the vision module and backscatter module at preset detection points or at fixed step distances, and synchronously writes the trigger information and global timestamp into the data acquisition subsystem; the mobile mechanism 14 also supports asymmetric collimator deflection drive to cooperate with the adaptive scanning deflection controller to adjust the pencil beam trajectory.

[0040] Understandably, by integrating precise pose measurement and synchronous triggering mechanisms into the mobile mechanism 14, the system can achieve repeated positioning and time alignment along the weld seam, ensuring a one-to-one correspondence between visual and backscattered data at the same weld seam location, thus providing a reliable physical basis for spatial registration and defect location; at the same time, the vehicle-based mobile solution supports long-distance continuous detection and improves the flexibility and automation of on-site deployment.

[0041] In this embodiment, the real-time feedback unit 15 is communicatively connected to the multimodal data fusion processing unit 13, and is used to generate a weld quality database containing defect type, location and size parameters based on the identification results, and send control commands to the external welding execution mechanism based on the evaluation value generated according to the severity of the defects.

[0042] Specifically, the real-time feedback unit 15 includes a defect level determination module, a weld quality database module, a visualization and alarm module, and a communication interface module. The defect level determination module converts the identification results into defect severity values ​​and processing suggestions based on a regularized weighted evaluation function. The weld quality database module encapsulates defect image slices, two-dimensional grayscale images, timestamps, pose information, and identification metadata in a unified data structure and establishes an index for retrieval and traceability. The visualization module generates image annotations, three-dimensional position mappings, and quality reports for weld defects. The communication interface module sends control commands, alarms, or adjustment suggestions to the welding actuator or welding robot (e.g., adjusting welding current, voltage, feed speed, or triggering welding stop) via industrial fieldbus or Ethernet (such as EtherNet / IP, Modbus, OPC UA, etc.) and supports uploading historical data to the enterprise server according to plan for long-term archiving and quality statistical analysis.

[0043] Understandably, the real-time feedback unit 15 transforms the detection process from passive recording to proactive intervention and quality closure: when a defect exceeding the threshold is detected, the system not only generates a traceable quality file for post-analysis, but also immediately sends a correction instruction to the welding end or issues a stop welding / manual intervention alarm, thereby reducing rework, reducing production losses and improving product reliability; at the same time, the structured data storage and indexing mechanism facilitates subsequent quality traceability, statistical analysis and continuous model training and optimization.

[0044] In this embodiment, the machine vision acquisition and positioning module 11 is equipped with a camera, which is used to acquire multiple continuous images of the weld surface and the molten pool area during the welding process. The image processing algorithm uses a pyramid hierarchical search strategy to fuse the multiple exposure images to eliminate welding arc interference and identify the weld centerline. The machine vision acquisition and positioning module transmits the identified spatial positioning information to the X-ray backscattering detection module in real time to constrain and guide the X-ray projection axis of the X-ray backscattering detection module to always align with the weld center.

[0045] Specifically, the machine vision acquisition and positioning module 11 includes a high dynamic range industrial camera (CCD or CMOS), a programmable zoom lens, a directional lighting device, and an image acquisition card. Under synchronous triggering, the camera continuously acquires image frames in short, medium, and long exposure sequences or more. The acquired frames can be used in the camera or on the host computer for exposure synthesis and arc suppression through pixel-level fusion (based on a pyramid hierarchical search strategy). Subsequently, a coarse-fine two-level positioning strategy using Hu invariant moment or invariant moment template matching is adopted to quickly locate the weld ROI on the synthesized image and extract spatial parameters such as centerline, weld width, and bevel angle. The module sends this spatial positioning information to the X-ray backscattering detection module via a real-time communication protocol (such as industrial Ethernet or serial / CAN bus) to guide the projection axis and synchronize with the trigger.

[0046] Understandably, this module addresses the high dynamic range issue in welding scenarios through multi-exposure fusion, providing high-quality surface images and accurate centerline coordinates. This reduces back-end computation (through ROI cropping) and significantly improves the positioning accuracy of backscatter projection, ensuring that single-sided backscatter imaging aligns with the weld center and meets the real-time requirements of online "welding and testing simultaneously".

[0047] In this embodiment, the X-ray backscatter detection module 12 includes a X-ray generating unit and a backscatter signal detector array. The X-rays generated by the X-ray generating unit are collimated into a pencil beam and injected into the weld, with the pencil beam reciprocating along the weld cross-section. The backscatter signal detector array is symmetrically arranged on both sides of the X-ray generating unit to receive photon signals scattered from inside the weld and convert them into analog electrical signals. A two-dimensional grayscale image is generated through signal modulation processing, wherein the grayscale value of the two-dimensional grayscale image is linearly correlated with the density of detection points inside the workpiece.

[0048] Specifically, the X-ray backscattering detection module 12 consists of a high-voltage driven X-ray generating unit (e.g., an X-ray tube), a collimator (or a multi-leaf aperture), a symmetrically arranged multi-channel detector array (e.g., a photomultiplier tube array, a silicon photodiode array, or a combination of scintillation and photodetectors), a front-end pulse shaping and amplification circuit, an analog-to-digital converter, and a back-end reconstruction and correction unit. The collimator shapes the divergent beam into a narrow-band pencil beam and scans it along the weld cross-section direction according to the controller trajectory. After receiving the rebound photons, the detector array amplifies, performs energy window segmentation and noise filtering. The image is sampled by the analog-to-digital converter and reconstructed into a two-dimensional grayscale image through splicing / geometric correction and normalization. After calibration and scaling, the grayscale value of the image is approximately linearly related to the local electron density or material density of the workpiece.

[0049] Understandably, by using pen-shaped collimation and symmetrical multi-point detection, high spatial resolution and high signal-to-noise ratio internal density images can be obtained when approaching the inspected part from one side. The combination of signal front-end and reconstruction algorithm can enhance the visibility of small density abrupt changes (such as micropores and cracks), provide a reliable source of internal features for multimodal algorithms, and adapt to different materials and thickness conditions by dynamically adjusting the gain / integration time.

[0050] In this embodiment, the multimodal data fusion processing unit 13 includes a spatial registration module, a multimodal feature extraction module, and a fusion inference module. The spatial registration module maps the two-dimensional grayscale image and the original image data to a unified pixel coordinate system based on the position information of the moving mechanism and the geometric relationship of the sensor installation. The multimodal feature extraction module further extracts the density abrupt change features inside the weld and the surface geometric texture features. The fusion inference module performs cascaded fusion of the density abrupt change features inside the weld and the surface geometric texture features based on a dual-branch convolutional neural network to construct a fused feature vector, and compares it with a preset defect feature library to output the recognition result.

[0051] Specifically, the spatial registration module calculates the pixel-to-physical coordinate transformation matrix based on the pre-calibrated intrinsic and extrinsic parameter matrices, the pose of the moving mechanism (provided by the odometer / IMU / track sensor), and data collected by the calibration board. It also performs distortion correction, scale transformation, and pixel-level mapping on the backscattered image and the visual image. The multimodal feature extraction module contains two parallel convolutional network branches: the backscattered branch focuses on extracting gray-level gradients, local contrast anomalies, and density boundaries, while the visual branch focuses on extracting edge, texture, melt pool morphology, and brightness features. The fusion inference module concatenates the features from the two branches into a unified fusion feature vector through a feature alignment layer, an attention weighting mechanism, and a cascaded fusion layer. This vector is then input into the classification / regression head to output the defect category, center coordinates, size estimate, and confidence score. Simultaneously, the recognition results are compared with a pre-set defect feature library for similarity assessment to assist in the judgment.

[0052] Understandably, this unit compensates for the blind spots of a single sensor by using spatiotemporal alignment and deep feature fusion, thereby improving the detection rate of low-contrast, small, or complex geometric defects. The modular and configurable network structure facilitates rapid adaptation to new materials or welding processes in the field through incremental training or transfer learning.

[0053] In this embodiment, the mobile mechanism 14 is an automated guided vehicle (AGV). The AGV is equipped with a high-precision odometer, an inertial measurement unit, and a track positioning sensor for autonomous navigation and attitude compensation along the detection path. A vibration-damping platform is mounted on the chassis of the AGV, and a protective cover with an integrated active cooling system surrounds the machine vision acquisition and positioning module and the X-ray backscatter detection module. A real-time feedback unit is connected to the fusion inference module, used to generate a digital quality profile of the weld seam based on the recognition results output by the multimodal data fusion processing unit, and to send compensation commands for welding current, voltage, or speed to the external welding execution mechanism based on the evaluation value generated according to the severity of the defects.

[0054] Specifically, the automated guided vehicle integrates a differential odometer, encoder, and IMU to achieve high-frequency pose updates, and is supplemented by track / visual positioning sensors for closed-loop positioning correction; the vibration damping platform on the chassis adopts a spring-damped or air-floating vibration damping structure to reduce the impact of operating vibration on image acquisition; the protective cover integrates an air-cooled or liquid-cooled circulation system, an electromagnetic shielding layer, and an anti-splash structure to ensure stable operation of the sensors in high-temperature and strong arc light environments; the real-time feedback unit writes the defect records (including timestamps, poses, image slices, and confidence levels) output by fusion inference into the local quality database, and generates a defect severity score based on a weighted evaluation function. When the score exceeds a preset threshold, a compensation command is issued to the welding actuator through an industrial protocol (such as dynamically adjusting welding current, voltage, wire feed speed, or triggering welding stop / manual intervention).

[0055] Understandably, mobile inspection platforms, while ensuring long-term continuous inspection capabilities, maintain data acquisition accuracy and equipment reliability through high-precision pose measurement and vibration reduction / protection design; real-time feedback closed-loop control helps to correct process parameters in a timely manner, reduce rework, and improve production efficiency and product qualification rate.

[0056] In this embodiment, the multimodal data fusion processing unit 13 further includes a dynamic gain compensation module. The dynamic gain compensation module is used to adjust the signal receiving gain and integration time of the X-ray backscattering detection module according to the workpiece surface roughness and weld groove geometric parameters obtained by the machine vision acquisition and positioning module, so as to compensate for the fluctuation of the signal-to-noise ratio of the two-dimensional grayscale image caused by changes in workpiece thickness or differences in surface scattering rate.

[0057] Specifically, the dynamic gain compensation module receives the surface roughness estimate, bevel angle, and local reflectivity information output by the vision module in real time, and calculates the target gain and integration time parameters by combining the backscattering historical signal statistics (such as background average and variance); it then sends these parameters to the backscattering front end via communication to adjust the detector preamplification factor, ADC sampling parameters, and integration window length; at the same time, in the image post-processing stage, it performs local gain mapping and histogram equalization or gamma correction on different blocks to unify contrast and grayscale distribution.

[0058] Understandably, by using prior surface / geometric information to drive gain and integration time adjustment, the detectability and contrast of backscattered images can be maintained at points of abrupt changes in material thickness or significant differences in surface scattering, reducing the probability of false detection / false detection and improving detection consistency across materials / thicknesses.

[0059] In this embodiment, the system further includes an adaptive scanning deflection controller. The adaptive scanning deflection controller is communicatively connected to the machine vision acquisition and positioning module, and is used to calculate the scanning path of the pencil beam in real time based on the curvature and offset of the identified weld centerline in the spatial coordinate system, and drive the collimator to perform asymmetric angle deflection to ensure that the scanning center of the pencil beam always coincides with the actual physical centerline of the weld.

[0060] Specifically, the adaptive scanning deflection controller receives information on the curvature, lateral offset, and local oscillation of the weld centerline, calculates the required collimator angle or deflector drive based on kinematic and geometric compensation algorithms, and outputs control signals in real time to drive the servo or piezoelectric deflection mechanism to change the projection angle or scanning trajectory. The controller also combines the instantaneous speed and attitude information of the moving mechanism to perform time correction to avoid motion blur.

[0061] Understandably, this controller can dynamically align the projected beam with the true center when the weld is bent, misaligned, or offset during installation, thus avoiding imaging area shifts or defect location errors caused by geometric deviations, thereby improving detection integrity and positioning accuracy.

[0062] In this embodiment, the system further includes an online dynamic calibration unit. The online dynamic calibration unit acquires the reference backscattered signal collected by the moving mechanism in the non-weld seam area, and, in conjunction with data from the current ambient temperature sensor, corrects the detection sensitivity of the X-ray backscattering detection module in real time to eliminate image grayscale drift caused by fluctuations in workpiece background thickness and temperature drift.

[0063] Specifically, the adaptive scanning deflection controller receives information on the curvature, lateral offset, and local oscillation of the weld centerline, calculates the required collimator angle or deflector drive based on kinematic and geometric compensation algorithms, and outputs control signals in real time to drive the servo or piezoelectric deflection mechanism to change the projection angle or scanning trajectory. The controller also combines the instantaneous speed and attitude information of the moving mechanism to perform time correction to avoid motion blur.

[0064] Understandably, this controller can dynamically align the projected beam with the true center when the weld is bent, misaligned, or offset during installation, thus avoiding imaging area shifts or defect location errors caused by geometric deviations, thereby improving detection integrity and positioning accuracy.

[0065] The workflow of the X-ray backscattering-based weld inspection system 10 is described below with an exemplary embodiment. The workflow includes steps such as data acquisition, preprocessing, spatiotemporal alignment, multimodal feature extraction and fusion reasoning, defect quantification and grade determination, and real-time feedback.

[0066] Specifically, the entire process includes the following steps: Data acquisition (visual + backscatter) Specifically, the machine vision acquisition and positioning module periodically... (Assuming the data collection time interval is) The k-th frame visual image is acquired based on the global reference time t0, and the acquisition time is denoted as t0. Simultaneously, the backscatter sensor acquires backscatter data for the k-th frame at the corresponding trigger time. The original timestamp of each sensor i in the k-th frame is denoted as... Considering device response latency ,but

[0067] Understandably, this is achieved by recording data at the acquisition end. And save The system can obtain a comparable original time baseline, providing a basis for subsequent time synchronization.

[0068] Visual preprocessing (multi-exposure fusion and noise reduction) Specifically, the visual channel acquires images of the weld surface and molten pool using short / medium / long multi-exposure sequences. Then, a pyramid hierarchical search strategy is used to perform pixel-level fusion of the multi-exposure images to expand the dynamic range and suppress arc artifacts. Distortion correction and geometric correction are also applied to the backscattered / visual original frames, respectively. A bilateral filter can be used to suppress backscattered image noise (or high-frequency noise in the visual image), and its specific implementation formula is as follows:

[0069] in: For the original image, This is the filtered image; Represents pixel coordinates, For A neighborhood window centered on the center; The spatial domain Gaussian kernel (controlling spatial distance weights) has the following parameters: ; The Gaussian kernel for the range (grayscale) (controlling grayscale similarity) has the following parameters: ; This is the normalization factor.

[0070] Understandably, bilateral filtering preserves edge information while smoothing homogeneous regions, making it suitable for enhancing the edge features of small density abrupt changes (such as cracks and micropores) in backscattered images, thus facilitating subsequent feature extraction.

[0071] Backscatter signal reconstruction (grayscale image generation and calibration) Specifically, the analog signal from the backscatter detector array is amplified at the front end, divided into energy windows, and converted from analog to digital before being stitched together to reconstruct a two-dimensional grayscale image. The grayscale values ​​are then mapped to relative electron density or density exponent (linear scaling) using a pre-calibrated curve to obtain an image representing the internal density distribution. The calibration relationship can be expressed as a linear approximation:

[0072] in The image grayscale value, This refers to the local material density or equivalent electron density. The calibration coefficient is obtained from experiments on standard samples.

[0073] It is understandable that calibration can unify the grayscale values ​​under different acquisition conditions to a physical meaning, thereby enabling comparison and quantification between different times / different workpieces.

[0074] Time synchronization and timestamp correction (time axis alignment) Specifically, taking the backscatter channel as a reference (or any channel as a reference), calculate the difference between the visual and backscatter timestamps of the k-th frame:

[0075] This difference is then used to correct the timestamps for the visual (or backscattered, depending on the reference strategy) ends, for example, to correct the visual end as follows:

[0076] Understandably, through the aforementioned time deviation compensation, the system can unify the data from different sensors to the same moment, reduce the registration error caused by trigger delay or transmission delay, and ensure the correspondence between the visual frame and the backscatter frame in physical location.

[0077] Spatial registration (pixel / physical coordinate mapping) Specifically, based on the pose (odometer / IMU / track positioning) provided by the mobile mechanism and the sensor mounting geometry, the transformation matrix from the visual pixels and backscattered pixels to the physical coordinate system of the weld is calculated. , Then, pixel-level or feature-level alignment is performed to obtain multimodal peer samples at the same physical point.

[0078] Understandably, accurate spatial registration is a prerequisite for achieving pixel-level feature fusion and precise defect localization, and registration errors directly affect the accuracy of size estimation and location positioning.

[0079] Multimodal feature extraction and fusion inference (deep learning) Specifically, let the features extracted by backscattering be The features extracted visually are ,in To unify the spatial dimensions (after registration). This represents the number of channels in each branch. The features from the two branches are concatenated along the channel direction to obtain the fused feature:

[0080] Then Input is transformed into linear features and activated to obtain output features or the final classification / regression head:

[0081] in For linear transformation of weights and biases, The activation function is (e.g., softmax for multi-class classification or sigmoid / softplus for regression confidence).

[0082] Understandably, dual-branch convolutional networks can retain and enhance internal density abrupt changes (backscattering advantage) and surface geometry / texture (visual advantage) information respectively. Through the fusion of splicing and attention / alignment mechanisms, they can ultimately improve the recognition performance of complex or low-contrast defects.

[0083] Defect quantification, severity assessment and real-time feedback Specifically, the fusion reasoning module outputs a series of candidate defects. Each defect includes: type Physical center coordinates Size (e.g., length) ,Width or equivalent diameter and confidence level These parameters are standardized and weighted to generate a comprehensive defect evaluation value. A common weighted model is expressed as:

[0084] in: For size normalization metrics (e.g., converting pixel area or longest axis length to millimeters and normalizing to [0,1]); Output the confidence score for the model (value range [0,1]); This is a category risk mapping function (mapping defect categories to severity scores, such as cracks being mapped to 1.0, lack of fusion to 0.8, inclusions to 0.6, etc.). Weight The weights are non-negative and usually satisfy the following conditions: It can be optimized according to engineering requirements to balance the impact of size, confidence level and defect type.

[0085] Understandable To comprehensively evaluate the indicators, different levels of response can be triggered when different thresholds are exceeded (such as minor alarm, production parameter adjustment, or immediate welding stoppage and manual judgment), and all defects and their parameters are written into the weld quality database to achieve traceability.

[0086] Supplementary Explanation (Formula Parameters and Calibration) Specifically, to change the size of the pixel scale Convert to physical size Calibration scale factor can be used :

[0087] in The confidence level is obtained from the camera's internal geometry through a calibration process. Can be determined simultaneously by classification confidence probability Combined with regression uncertainty estimates (e.g., based on Monte Carlo dropout or variance regression). Categorical risk function. It can be set by quality engineers based on the failure impact or obtained through statistical learning of historical failure conditions.

[0088] The weld inspection system based on X-ray backscattering proposed in this application achieves simultaneous online acquisition and pixel-level registration of internal density and surface morphology information of welds by tightly coupling single-sided backscattering imaging with high dynamic range machine vision. Utilizing technologies such as collimated pencil beam and multi-channel detector array, vision-guided projection constraints, adaptive scan deflection, and dynamic gain compensation, the system can obtain high signal-to-noise ratio internal grayscale images and retain minute density abrupt changes in weld seams, even in scenarios where the back side is inaccessible or in thick plates and multi-layer welds. Combined with a dual-branch deep fusion network for cross-modal feature-level fusion, it effectively improves the detection rate, classification accuracy, and positioning precision of minute or low-contrast defects such as cracks, porosity, inclusions, and lack of fusion.

[0089] Furthermore, the system engineering design includes AGV movement detection, synchronous triggering and high-precision pose compensation, online dynamic calibration and real-time feedback closed loop, enabling the detection process to be continuous, real-time, and robust. Detection results can instantly generate structured quality files and automatically issue welding parameter adjustment or welding stop commands based on defect severity, thereby significantly reducing rework and downtime, improving production pass rates, and achieving quality traceability and continuous optimization. Overall, this invention balances unilateral online feasibility, deep multimodal sensing capabilities, and deployment in production environments, demonstrating good engineering applicability and promotional value.

[0090] Please refer to the following: Figure 2 , Figure 2 This is a schematic diagram of a backscattering module provided in one embodiment of this application.

[0091] The X-ray backscattering module provided in this application constructs a tightly coupled precision detection closed loop in its hardware architecture. The system primarily uses an Automated Guided Vehicle (AGV) as its mobile carrier, with a machine vision acquisition and positioning module and a X-ray backscattering detection module integrated on its chassis via a vibration-damping mounting platform. In terms of physical arrangement, industrial cameras are fixed around the periphery of the X-ray generating unit; this parallel arrangement ensures that the vision sensor can lock onto the weld area before or simultaneously with the X-ray probe. To cope with the harsh physical environment of the welding site, the entire detection module is encased in a protective cover with electromagnetic shielding and incorporates an active circulating cooling system to ensure that the X-ray tube maintains a constant temperature and stable operating state even under high temperature and strong electromagnetic interference conditions.

[0092] In the workflow, the system achieves precise alignment of heterogeneous signals through a hardware synchronous trigger controller. As the AGV travels along the weld seam trajectory, the position sensor provides real-time displacement data, which the trigger controller then sends trigger pulses to the camera, enabling high dynamic range imaging of the molten pool and weld seam surface. Simultaneously, the pencil beam generated by the X-ray generator performs a lateral reciprocating scan driven by the collimator, penetrating deep into the workpiece to collect density scattering signals. This collaborative mechanism of "vision-guided, X-ray-following" allows the spatial positioning information output by the vision module to constrain the scanning range of the X-ray in real time, ensuring that the pencil beam is always accurately projected onto the central area of ​​the weld seam, greatly improving the effectiveness of unilateral non-contact detection.

[0093] The data processing center is handled by a high-performance PC processor, responsible for receiving and processing two modalities of data from the front end. The processor first uses image processing algorithms to denoise and extract features from the visual images, while simultaneously converting the backscattered analog signal into a two-dimensional grayscale image. Through pre-defined spatial registration logic, the system maps the grayscale image representing internal defects and the optical image representing surface morphology to a unified pixel coordinate system based on the physical installation geometry of the sensors. Subsequently, these registered multimodal features are input into a dual-branch convolutional neural network model, where the AI ​​model automatically determines whether there are defects such as porosity and slag inclusions inside the weld, and whether there are cracks or undercuts on the surface, providing the identification results in real time.

[0094] Finally, the system achieves in-depth application of detection results and a quality closed loop through a real-time feedback unit. Identified defect information, image summaries, and spatial locations are automatically archived into a digital quality database, generating traceable weld quality files. More importantly, the feedback unit can calculate evaluation values ​​in real time based on the severity of defects and send compensation commands to the external welding robot, dynamically adjusting welding current, voltage, or travel speed. This closed-loop control logic, from perception to decision-making to execution, not only achieves "testing while welding" but also significantly reduces subsequent rework costs through real-time process correction, ensuring automated and intelligent management of the welding quality of large steel structural components.

[0095] Please refer to the following: Figure 3 , Figure 3 This is a schematic flowchart of a weld inspection method based on X-ray backscattering provided in an embodiment of this application. Figure 3 The weld inspection method based on X-ray backscattering, as shown, includes at least the following steps: S100: Acquire the original image data of the weld surface, identify the weld centerline using an image processing algorithm, and output the spatial positioning information of the weld area; S200: Based on the spatial positioning information, emit X-rays to the weld area and collect backscattered signals, converting the backscattered signals into a two-dimensional grayscale image characterizing the density distribution inside the weld; S300: Perform spatial registration and feature extraction on the original image data and the two-dimensional grayscale image, construct a unified feature vector, and output the identification result of weld defects through a preset defect identification model; S400: Send a synchronization pulse signal according to the position signal so that the original image data and the two-dimensional grayscale image have corresponding timestamps; S500: Generate a weld quality database containing defect type, location, and size parameters based on the identification result, and send control commands to an external welding execution mechanism based on the evaluation value generated according to the severity of the defects.

[0096] S100: Acquire raw image data of the weld surface, identify the weld centerline through image processing algorithms, and output the spatial positioning information of the weld area.

[0097] In this embodiment of the application, the weld detection method based on X-ray backscattering includes the following steps in step S100: acquiring the original image data of the weld surface, identifying the weld centerline through an image processing algorithm, and outputting the spatial positioning information of the weld area. For details, please refer to the attached document. Figure 1 , Figure 2 The relevant descriptions will not be repeated here.

[0098] S200: Based on spatial positioning information, it emits rays to the weld area and collects backscattered signals, converting the backscattered signals into a two-dimensional grayscale image that characterizes the density distribution inside the weld.

[0099] In this embodiment, the weld detection method based on X-ray backscattering includes the following steps in step S200: emitting X-rays towards the weld area based on spatial positioning information and acquiring backscattered signals; converting the backscattered signals into a two-dimensional grayscale image characterizing the internal density distribution of the weld. For details, please refer to the attached document. Figure 1 , Figure 2 The relevant descriptions will not be repeated here.

[0100] S300: Spatial registration and feature extraction are performed on the original image data and the two-dimensional grayscale image to construct a unified feature vector, and the identification results of weld defects are output through the preset defect identification model.

[0101] In this embodiment, the weld detection method based on X-ray backscattering includes the following steps in step S300: spatial registration and feature extraction of the original image data and the two-dimensional grayscale image, constructing a unified feature vector, and outputting the weld defect identification result through a preset defect identification model. For details, please refer to the attached document. Figure 1 , Figure 2 The relevant descriptions will not be repeated here.

[0102] S400: Sends a synchronization pulse signal based on the position signal so that the original image data and the two-dimensional grayscale image have corresponding timestamps.

[0103] In this embodiment of the application, the weld detection method based on X-ray backscattering includes, in step S400: sending a synchronization pulse signal according to the position signal so that the original image data and the two-dimensional grayscale image have corresponding timestamps. For details, please refer to the following documentation. Figure 1 , Figure 2 The relevant descriptions will not be repeated here.

[0104] S500: Generates a weld quality database containing defect type, location, and size parameters based on the identification results, and sends control commands to external welding execution agencies based on the evaluation value generated according to the severity of the defects.

[0105] In this embodiment, the weld inspection method based on X-ray backscattering includes the following steps in step S500: generating a weld quality database containing defect type, location, and size parameters based on the identification results, and sending control commands to an external welding execution mechanism based on the evaluation value generated according to the severity of the defects. For details, please refer to the attached document. Figure 1 , Figure 2 The relevant descriptions will not be repeated here.

[0106] Figure 4 This is a detection device 20 provided in one embodiment of this application. For example... Figure 4 As shown, the detection device 20 includes at least the following components: processor 21 and memory 22.

[0107] In this embodiment, the memory 22 is used to store executable instructions of the processor 21, which, when configured to execute instructions, implement... Figure 3 The weld inspection method based on X-ray backscattering is shown in the figure.

[0108] In one embodiment of this application, the program operating in the detection device 20 may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). The information processed by these devices is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (Flash ROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0109] It should be noted that a portion of the detection device 20 described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0110] It should be noted that the "computer" mentioned here refers to the computer built into the detection device 20, which uses hardware including an operating system and peripheral devices. Furthermore, "computer-readable recording media" refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage devices such as hard drives built into the computer.

[0111] Furthermore, a "computer-readable recording medium" can include: a medium that dynamically stores a program for a short period of time, such as a communication line used when transmitting a program via a network such as the Internet or a communication line such as a telephone line; or a medium that stores a program for a fixed period of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining with programs already recorded in the computer.

[0112] It is understood that the embodiments of this application provide a weld inspection system, method, and device based on X-ray backscattering. This system, method, and device can simultaneously acquire, time / space register, multimodal fusion recognition, and provide real-time feedback of internal density information and surface morphology information of the weld during welding operations. By precisely positioning the weld centerline using a machine vision module to guide the projection of backscattered rays, and combining pencil collimation with the high-sensitivity reception of scattered photons by a backscattering detector array, this system can obtain a two-dimensional grayscale image reflecting internal defects from a single side. Furthermore, this internal information is registered with a high dynamic range multi-exposure visual image in a unified pixel coordinate system and fused with features at the convolutional neural network level to achieve joint identification, location, and quantitative assessment of internal and surface defects such as cracks, porosity, inclusions, and lack of fusion.

[0113] This embodiment integrates the detection device onto a mobile AGV equipped with an odometer / IMU / track positioning system, and includes a synchronous trigger controller, position sensor, vibration damping platform, and protective cooling cover. This enables multi-angle, continuous scanning detection and high-precision attitude compensation for long welds. The dynamic gain compensation module, adaptive scan deflection controller, and online dynamic calibration unit can adjust the receiving gain, integration time, and scanning trajectory in real time to accommodate changes in workpiece thickness, surface roughness, or ambient temperature drift. This maintains a high signal-to-noise ratio in the backscattered image and alignment between the scanning center and the actual weld centerline, significantly improving detection robustness and the detectability of subtle defect features. The multimodal data fusion processing unit, based on a dual-branch / multi-channel deep network, performs cascaded fusion and outputs defect type, location, size, and confidence level. The results are then encapsulated in a searchable weld quality database using a unified data structure, supporting visual annotation, historical traceability, and incremental updates.

[0114] In summary, the embodiments of this application have the following beneficial effects: On the one hand, they achieve single-sided online inspection capability in complex, large components or scenarios where the back side is inaccessible, making up for the shortcomings of traditional RT / UT offline or double-sided restricted inspection; on the other hand, they significantly improve the accuracy and positioning precision of defect detection through vision-backscatter multimodal fusion and AI inference, and ensure the real-time performance, stability and adaptability of the system in industrial welding sites through engineering designs such as synchronous triggering, AGV pose compensation, dynamic gain and online calibration; in addition, the system links the inspection results to the welding actuator in real time and builds a digital quality archive, which helps to adjust welding parameters in a timely manner, reduce rework costs and achieve quality traceability and closed-loop production control, and has good engineering application value and promotion prospects.

[0115] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A weld inspection system based on X-ray backscattering, applied to weld inspection during the welding process, characterized in that, The system includes: The machine vision acquisition and positioning module is used to acquire raw image data of the weld surface, identify the weld centerline through image processing algorithms, and output the spatial positioning information of the weld area. The X-ray backscattering detection module is communicatively connected to the machine vision acquisition and positioning module. It is used to emit X-rays to the weld area based on the spatial positioning information and acquire backscattering signals, and convert the backscattering signals into a two-dimensional grayscale image characterizing the density distribution inside the weld. The multimodal data fusion processing unit is used to perform spatial registration and feature extraction on the original image data and the two-dimensional grayscale image, construct a unified feature vector, and output the identification result of weld defects through a preset defect identification model. A mobile mechanism is used to carry the machine vision acquisition and positioning module and the ray backscattering detection module. The mobile mechanism is equipped with a synchronization trigger controller and a position sensor. The synchronization trigger controller is used to send a synchronization pulse signal to the vision module and the ray backscattering detection module according to the position signal fed back by the position sensor, so that the original image data and the two-dimensional grayscale image have corresponding timestamps. The real-time feedback unit is communicatively connected to the multimodal data fusion processing unit. It is used to generate a weld quality database containing defect type, location, and size parameters based on the identification results, and to send control commands to the external welding execution mechanism based on the evaluation value generated according to the severity of the defects.

2. The weld inspection system based on X-ray backscattering according to claim 1, characterized in that, The machine vision acquisition and positioning module is equipped with a camera, which is used to acquire multi-exposure continuous images of the weld surface and the molten pool area during the welding process. The image processing algorithm uses a pyramid hierarchical search strategy to fuse multi-exposure images to eliminate welding arc interference and identify the weld centerline. The machine vision acquisition and positioning module transmits the identified spatial positioning information to the X-ray backscattering detection module in real time to constrain and guide the X-ray projection axis of the X-ray backscattering detection module to always be aligned with the center of the weld.

3. The weld inspection system based on X-ray backscattering according to claim 2, characterized in that, The backscattering detection module includes a radiation generating unit and a backscattering signal detector array; The rays generated by the ray generating unit are shaped into pencil beams by a collimator and enter the weld, and the pencil beams reciprocate along the cross-sectional direction of the weld. The backscatter signal detector array is symmetrically arranged on both sides of the ray generating unit to receive photon signals scattered back from inside the weld and convert them into analog electrical signals. The two-dimensional grayscale image is generated through signal harmonic processing, wherein the grayscale value of the two-dimensional grayscale image is linearly correlated with the density of detection points inside the workpiece.

4. The weld inspection system based on X-ray backscattering according to claim 3, characterized in that, The multimodal data fusion processing unit includes a spatial registration module, a multimodal feature extraction module, and a fusion inference module; The spatial registration module is used to map the two-dimensional grayscale image and the original image data to a unified pixel coordinate system based on the position information of the moving mechanism and the sensor installation geometry. The multimodal feature extraction module is also used to extract density abrupt change features and surface geometric texture features inside the weld. The fusion inference module uses a dual-branch convolutional neural network to cascade and fuse the density abrupt change features inside the weld with the surface geometric texture features to construct a fusion feature vector, and then compares it with a preset defect feature library to output the recognition result.

5. The weld inspection system based on X-ray backscattering according to claim 4, characterized in that, The moving mechanism is an automated guided vehicle (AGV), which is equipped with a high-precision odometer, an inertial measurement unit, and a track positioning sensor to achieve autonomous navigation and attitude compensation of the detected path. The chassis of the automated guided vehicle is equipped with a vibration damping platform, and the machine vision acquisition and positioning module and the X-ray backscatter detection module are wrapped with a protective cover with an integrated active circulation cooling device. The real-time feedback unit is connected to the fusion reasoning module and is used to generate a digital quality profile of the weld seam based on the identification results output by the multimodal data fusion processing unit, and to send compensation instructions for welding current, voltage or speed to the external welding execution mechanism based on the evaluation value generated according to the severity of the defects.

6. The weld inspection system based on X-ray backscattering according to claim 1, characterized in that, The multimodal data fusion processing unit also includes a dynamic gain compensation module: The dynamic gain compensation module is used to adjust the signal receiving gain and integration time of the X-ray backscattering detection module according to the workpiece surface roughness and weld groove geometric parameters obtained by the machine vision acquisition and positioning module, so as to compensate for the fluctuation of the signal-to-noise ratio of the two-dimensional grayscale image caused by the change of workpiece thickness or the difference of surface scattering rate.

7. The weld inspection system based on X-ray backscattering according to claim 3, characterized in that, The system also includes an adaptive scan deflection controller: The adaptive scanning deflection controller is communicatively connected to the machine vision acquisition and positioning module. It is used to calculate the scanning path of the pencil beam in real time based on the curvature and offset of the identified weld centerline in the spatial coordinate system, and drive the collimator to perform asymmetric angle deflection to ensure that the scanning center of the pencil beam always coincides with the actual physical centerline of the weld.

8. The weld inspection system based on X-ray backscattering according to claim 1, characterized in that, The system also includes an online dynamic calibration unit; The online dynamic calibration unit is used to acquire the reference backscatter signal collected by the moving mechanism in the non-weld seam area, and to correct the detection sensitivity of the X-ray backscatter detection module in real time by combining the data fed back by the current ambient temperature sensor, so as to eliminate image grayscale drift caused by workpiece background thickness fluctuation and temperature drift.

9. A weld inspection method based on X-ray backscattering, applied to the weld inspection system based on X-ray backscattering as described in any one of claims 1 to 8, characterized in that, The method includes: The original image data of the weld surface is acquired, and the weld centerline is identified through image processing algorithms and the spatial positioning information of the weld area is output. Based on the spatial positioning information, rays are emitted into the weld area and backscattered signals are collected. The backscattered signals are then converted into a two-dimensional grayscale image characterizing the density distribution inside the weld. Spatial registration and feature extraction are performed on the original image data and the two-dimensional grayscale image to construct a unified feature vector, and the identification result of weld defects is output through a preset defect identification model. A synchronization pulse signal is sent according to the position signal so that the original image data and the two-dimensional grayscale image have corresponding timestamps; Based on the identification results, a weld quality database containing defect type, location, and size parameters is generated, and control commands are sent to external welding execution agencies based on the evaluation value generated according to the severity of the defects.

10. A detection device, characterized in that, include: processor; as well as The memory stores computer-readable instructions for controlling the processor to execute the X-ray backscattering-based weld inspection method as described in claim 9.

Images