An industrial weld non-destructive intelligent inspection system
By constructing a three-dimensional geometric model and digital twin carrier using multimodal sensing data, and combining reinforcement learning and Bayesian modeling, the accuracy problem of ultrasonic testing systems under complex working conditions was solved, achieving high-precision non-destructive testing and stable coupling agent control.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-12
- Publication Date
- 2026-04-07
Smart Images
Figure CN121298906B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of non-destructive testing, in particular to an industrial welding seam non-destructive intelligent testing system. BACKGROUND
[0002] In modern industrial manufacturing, welding as a key connection process is widely used in various fields; the quality of the weld directly relates to the safety, reliability and service life of the structure, so efficient and accurate non-destructive testing of the weld is of great significance; the ultrasonic testing technology has the advantages of non-destructive, strong penetration ability, high sensitivity and being suitable for various materials and complex structures, etc., however, in the actual application process, the effect of ultrasonic testing is significantly affected by many factors, when the workpiece surface has curvature change, ultrasonic wave is easy to refract, scatter and reflect path deviation in the propagation process, affecting the accuracy of defect positioning and quantitative analysis; different materials have different acoustic parameters, including sound velocity, density and attenuation coefficient, etc.; in addition, coupling agent as a key medium for effective sound energy transmission between probe and workpiece, if the coupling is poor or there is air gap locally, it will cause the reflection of sound wave energy to increase, the transmission to weaken, and even cause signal loss in severe cases, which seriously affects the stability and repeatability of the detection results.
[0003] Most of the existing ultrasonic testing systems are based on idealized assumptions such as flat workpieces, homogeneous materials and perfect coupling, and are often difficult to adapt to complex actual working conditions, resulting in decreased detection accuracy and increased misjudgment rate, especially in pipe girth welds, curved structures and other application scenarios; although some systems introduce artificial experience correction or simple gain compensation mechanism, they lack systematic modeling and dynamic correction ability for the coupling of the above factors, and it is difficult to realize accurate compensation of detection errors under complex working conditions. SUMMARY
[0004] (I) Technical problems solved
[0005] In view of the deficiencies of the prior art, the present application provides an industrial welding seam non-destructive intelligent testing system, which solves the problems raised in the background art.
[0006] (II) Technical solutions
[0007] To achieve the above purpose, the present application is realized by the following technical solutions:
[0008] An industrial welding seam non-destructive intelligent testing system, comprising:
[0009] A perception registration module acquires sensor data and obtains multi-modal perception data through a dynamic space-time registration algorithm; wherein the sensor data includes contour perception, spatial perception and morphological perception;
[0010] A weld seam model construction module converts the multi-modal perception data into a three-dimensional geometric model while integrating the obtained weld seam material and environmental parameters to construct a digital twin carrier;
[0011] A decision control module generates a coupling agent control strategy and a probe motion trajectory based on the constructed digital twin carrier, adopts topological path planning and constraint-enhanced reinforcement learning, and acquires weld seam detection data based on the probe motion trajectory;
[0012] The enhanced reinforcement learning is based on a virtual decision environment constructed by the digital twin, combines coupling agent control constraints, and verifies the reinforcement learning that meets the precision and usage optimization through virtual-real migration;
[0013] An AI algorithm optimization module acquires weld seam material, environmental parameters, and multi-modal perception data, establishes a mapping relationship of multi-factor application parameters, and quantifies the uncertainty of weld seam detection data through Bayesian modeling.
[0014] Further, the acquisition process of the sensor data is as follows:
[0015] Collecting weld seam area images, extracting sub-pixel level contour features of the weld seam through a gray center of gravity method;
[0016] Collecting the distance between the probe and the weld seam surface, the coupling agent thickness, and the spatial position of the probe, and simultaneously collecting coupling pressure distribution data to acquire spatial position and contact state information of the weld seam and the probe;
[0017] By emitting and receiving ultrasonic signals through a high-frequency ultrasonic probe, the peak amplitude, rising edge time, and center frequency characteristics of the defect echo are extracted, and the working environmental parameters are collected in combination with the environmental sensors.
[0018] Further, the acquisition process of the multi-modal perception data is as follows:
[0019] The acquisition of multi-source sensor data adopts a dynamic space-time registration algorithm for time synchronization and space registration to obtain multi-modal perception data.
[0020] Further, the process of converting the multi-modal perception data into a three-dimensional geometric model is as follows:
[0021] The multi-source data is preprocessed and unified into a three-dimensional point cloud data format; based on the preprocessed sub-pixel edge data, sampling points are obtained along the length direction of the weld at a step size of 0.1 mm, the Y coordinate of each sampling point is determined by the fitting curve of the sub-pixel edge, and the Z coordinate is initially set as the surface height of the base material, a two-dimensional contour point cloud of the weld surface is generated, the surface contour point cloud is stretched along the Z axis direction combined with the coupling agent thickness, the preliminarily generated three-dimensional point cloud is denoised, down-sampled and smoothed, a three-dimensional triangular mesh model of the weld area is constructed by using a Delaunay triangulation algorithm on the preprocessed three-dimensional point cloud, and a smooth surface is fitted by extracting the control vertex and weight factor of the defect three-dimensional point cloud.
[0022] Further, the process of constructing the digital twin carrier is:
[0023] Real-time environmental working condition data is collected by deploying multiple types of environmental sensors, a quantitative correlation model of environmental working condition data and detection is established, and the real-time data of the environmental sensor is bound to the detection period of the digital twin carrier through a time stamp.
[0024] Further, the process of the coupling agent control strategy is:
[0025] The multi-dimensional data in the digital twin carrier is converted into a virtual decision environment that can be called by topology path planning and reinforcement learning, and the influence factor and constraint boundary of coupling agent control are determined; the weld topology features are obtained based on the digital twin carrier, the coupling agent control area is divided, the detection path covering the entire weld is generated, and the intelligent agent generates real-time coupling agent control parameters in each control area of the topology path.
[0026] Further, the process of the probe motion trajectory is:
[0027] The weld three-dimensional topology data is extracted from the digital twin carrier, and is preprocessed in combination with the topology information in the digital twin and the hardware constraints, a main motion path covering the entire weld is generated, after the global main path is determined, local trajectory optimization is performed in combination with the real-time data in the digital twin, and the planned probe motion trajectory is simulated and verified in the entire process in the digital twin carrier.
[0028] Further, the process of establishing the mapping relationship of the multi-factor coating parameters is:
[0029] The historical data and target working condition scenarios are called from the digital twin carrier, a virtual training environment is constructed, the intelligent agent is dynamically explored by simulating a plurality of complex working conditions in the digital twin, the trained strategy is imported into the actual detection system, and the results of the digital twin simulation and the actual detection are compared. The trained reinforcement learning model can take the weld material, environmental parameters and multi-modal perception data as input and output the optimal coupling agent coating parameters to form a dynamic mapping relationship.
[0030] Further, the process of quantifying the uncertainty of the weld joint detection data through Bayesian modeling is:
[0031] Based on historical detection data and domain knowledge, the prior probability distribution of each uncertainty factor is constructed, a hierarchical Bayesian model is constructed, the uncertainty of the weld joint detection data is decomposed into four levels of sensor layer, coupling agent layer, material layer and environment layer, each level is related to each other through probability distribution, multi-modal perception data is used as observation data, the prior distribution of each level is combined to construct the likelihood function; Markov chain Monte Carlo method is used to solve the posterior distribution, three types of uncertainty quantification indicators are extracted from the posterior distribution samples, in defect identification, the uncertainty quantification result is integrated into the digital twin carrier by adjusting the identification logic, and the detection parameters are dynamically adjusted based on the uncertainty quantification result.
[0032] (Three) beneficial effects
[0033] The present application provides an industrial weld joint nondestructive intelligent flaw detection system, which has the following beneficial effects:
[0034] (1) The present application constructs an ultrasonic, laser and visual multi-modal perception system through multi-dimensional, synchronously collects weld ultrasonic echo characteristics, three-dimensional contour data and surface image information, realizes information complementation in space and channel dimensions through a feature fusion module, effectively distinguishes similar defects such as pores and slag, quantifies various sources of uncertainty through Bayesian modeling, constructs a prior distribution through historical data, solves the posterior probability in combination with real-time observation data, accurately outputs the confidence interval of defect parameters, simulates the detection process through digital twin simulation, corrects trajectory deviation and coupling fluctuation in advance, reduces system error from the source, and solves the problem of distorted detection data under complex working conditions.
[0035] (2) The present application plans the probe trajectory through segmented Bezier curve, automatically enlarges the curvature of the curved section, starts reciprocating scanning in the defect section, adapts to various complex weld joint shapes, introduces a reinforcement learning algorithm in the coupling agent application link, dynamically optimizes the supply amount and regional allocation based on the weld joint material, such as the hardness of the fusion zone, environmental parameters and the uniformity of the perception data coupling agent, maintains the optimal coupling state without manual intervention, and at the same time, after the detection is completed, the data such as trajectory deviation, defect characteristics and correction records are returned to the twin carrier to update the weld joint material database and algorithm model. BRIEF DESCRIPTION OF DRAWINGS
[0036] Figure 1 The present application is a system schematic diagram. DETAILED DESCRIPTION
[0037] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.
[0038] Please refer to Figure 1 The embodiment provides an industrial weld nondestructive intelligent flaw detection system, and the detection system comprises:
[0039] A perception registration module is configured to acquire sensor data and obtain multi-modal perception data through a dynamic space-time registration algorithm, wherein the sensor data comprises contour perception, spatial perception and morphological perception.
[0040] The contour perception data is acquired through an industrial camera, and the 5000-pixel sub-pixel industrial camera is adopted to collect the image of the weld area, and the sub-pixel contour features of the weld, such as width, excess height and edge angle, are extracted through an improved gray center method and other algorithm processing, so that the high-precision perception of the weld contour is realized.
[0041] The spatial perception data is acquired by using a laser displacement sensor, such as a triangulation laser sensor, to collect the data of the coupling agent thickness of the probe and the weld surface, the spatial position of the probe and other data in real time, and to acquire the spatial position and contact state information of the weld and the probe in combination with the coupling pressure distribution data collected by the multi-unit pressure sensor array.
[0042] The morphological perception data is acquired by emitting and receiving ultrasonic signals through a high-frequency ultrasonic probe, extracting the peak amplitude, rising edge time and center frequency of the defect echo, and realizing the perception of the internal defect morphology of the weld; meanwhile, the working condition environment parameters collected by the environmental sensors such as temperature sensors and vibration sensors are used as supplementary information for morphological perception.
[0043] The multi-modal perception data is acquired as follows:
[0044] Through time synchronization and space registration, the multi-source sensor data is kept consistent in time sequence and spatial position.
[0045] The time synchronization is performed through the time sequence accurate alignment of the multi-sensor, the FPGA is used as the core hardware for time synchronization, the unique timestamp with a precision of 1 μs is allocated to all the devices participating in data collection, the trigger signal of the ultrasonic probe is used as the reference clock, the image collection time of the industrial camera, the sampling time of the laser and pressure sensor, and the monitoring time of the environmental sensor are all time sequence aligned with the reference clock, and finally the time deviation of the multi-source sensor data is controlled within 10 μs, so that the synchronization of the multi-modal data in the time dimension is realized.
[0046] Through the definition of multi-sensor coordinate system, calibration experiment and conversion matrix solving, and the spatial unified conversion of multi-modal data, spatial registration is performed;
[0047] The multi-sensor coordinate system is defined for image, laser and pressure, and ultrasound respectively; the calibration experiment is performed through the standard calibration board and the weld test block; the image data of the industrial camera, the distance data of the laser sensor, and the contact data of the pressure sensor are synchronously collected through the standard calibration board, and the conversion matrix of the image coordinate system and the laser and pressure coordinate system is solved through the algorithm; the defect echo data of the ultrasonic probe and the coupling agent thickness data of the laser sensor are synchronously collected through the weld test block, and the conversion matrix of the ultrasonic coordinate system and the laser and pressure coordinate system is solved; the profile perception data of the industrial camera is converted from the image coordinate system to the laser and pressure coordinate system by using the pre-solved conversion matrix; the form perception data of the ultrasonic probe is converted from the ultrasonic coordinate system to the laser and pressure coordinate system; all data are unified in the global coordinate system with the weld surface as the reference, and the multi-modal data are aligned in the spatial position.
[0048] After time synchronization and spatial registration, the profile perception and spatial perception, i.e. coupling agent thickness distribution, contact pressure distribution, probe spatial posture, and multi-source data of form perception are fused into unified multi-modal perception data.
[0049] The multi-modal perception data is converted into a three-dimensional geometric model by the weld model construction module, and the weld material and environmental parameters are integrated to construct a digital twin carrier;
[0050] Three-dimensional geometric model construction:
[0051] The profile perception, spatial perception and form perception are preprocessed and unified into three-dimensional point cloud data format; for the profile perception data, the two-dimensional coordinates of the sub-pixel edge are converted into three-dimensional spatial coordinates with the weld base material surface as the reference, and each point is assigned with profile attribute labels such as width value and excess height value; for the spatial perception data, the thickness value measured by the laser is mapped to the Z-axis of the three-dimensional coordinates, and the pressure distribution data is associated to the corresponding three-dimensional point according to the sensor unit position, and the translation and rotation parameters of the probe are taken as global transformation factors to ensure the position consistency of the spatial data;
[0052] For the form perception data, the three-dimensional depth of the defect is calculated according to the propagation time and sound speed of the ultrasonic echo, and the horizontal position of the profile perception is combined to assign the defect point with defect attribute labels such as peak amplitude and defect type; the environmental parameters are taken as global attributes and marked in the detection period of the entire model;
[0053] Based on the pre-processed sub-pixel edge data, sampling along the weld length direction X-axis with a step of 0.1 mm, the Y-coordinate of each sampling point is determined by the fitting curve of the sub-pixel edge, such as polynomial fitting, and the Z-coordinate is initially set as the surface height of the base material, generating a two-dimensional contour point cloud of the weld surface;
[0054] The preliminarily generated three-dimensional point cloud is subjected to denoising, downsampling and smoothing processing, and a statistical filtering algorithm is adopted to eliminate outliers caused by sensor interference, such as laser reflection abnormal points and pressure sudden change points; through voxel grid filtering, the point cloud density is reduced to 5 points / mm 3 , balancing the model accuracy and calculation efficiency; a Gaussian filter is adopted, with a 3x3x3 window, to eliminate local fluctuations of the point cloud and ensure the smoothness of the surface reconstruction;
[0055] The pre-processed point cloud is subjected to Delaunay triangulation algorithm to construct a three-dimensional triangular mesh model of the weld area; this algorithm can ensure the spatial consistency of the mesh, so that the edge length deviation of the triangular facet is ≤0.05 mm, accurately restoring the macroscopic geometric shape of the weld, such as the excess height and the groove angle; for defect areas and other parts that need to be finely expressed, NURBS, non-uniform rational B-spline surface algorithm, is adopted. By extracting the control vertices and weight factors of the defect point cloud, a smooth surface is fitted, so that the three-dimensional morphology of the defect, such as the extension direction of the crack and the spherical profile of the porosity, is accurately presented in the model;
[0056] The morphologically perceived ultrasonic defect data, peak amplitude, rising edge time and center frequency, are associated with the three-dimensional model; according to the propagation time t of the ultrasonic echo and the coupling agent sound speed v, the defect depth is calculated, and the horizontal position of the contour perception is marked in the three-dimensional model to mark the spatial coordinates (X, Y, Z) of the defect;
[0057] A digital twin carrier is constructed;
[0058] The material parameters of different areas of the weld, including grain size, microstructure, hardness distribution and longitudinal wave speed, are obtained in advance, and the material parameters are associated with the corresponding areas of the three-dimensional geometric model, so as to give the digital properties to each area, and to establish a quantitative association model between the environmental parameters and the detection system. The three-dimensional geometric model, the material parameters of the weld and the environmental parameters are integrated in a unified digital twin platform; the macro shape of the weld and the spatial position of the defect are presented in the three-dimensional model through the geometric dimension; the material properties in different areas are superimposed on the geometric model in the form of color gradient or attribute label through the material dimension; and the parameters such as temperature and vibration are visualized in the form of real-time numerical value or dynamic curve in the model interface through the environmental dimension, and drive the physical simulation in the model, such as ultrasonic beam propagation simulation under temperature change;
[0059] A decision control module generates a coupling agent control strategy and a probe motion trajectory based on the constructed digital twin carrier, adopts topology path planning and constraint enhanced reinforcement learning, and acquires weld detection data based on the probe motion trajectory;
[0060] The enhanced reinforcement learning is based on a virtual decision environment constructed by the digital twin, combined with coupling agent control constraints, and verified by virtual-real migration to meet the accuracy and consumption optimization of reinforcement learning.
[0061] Coupling agent control strategy:
[0062] By building a virtual decision basis, multidimensional data is converted into a virtual decision environment that can be called by topology path planning and reinforcement learning, and the influence factors and constraint boundaries of coupling agent control are determined.
[0063] The core of topology path planning is to divide the coupling agent control area based on the topological characteristics of the weld, generate a detection path that covers the entire weld and takes into account efficiency and accuracy, and provide regional optimization targets for reinforcement learning. Topology node extraction and path partitioning; When extracting the key nodes of the weld topology, the inflection points, defect centers, material partition boundaries, and width mutation points of the weld are used as the core topology nodes. For example, a straight-line weld with 3 defects contains 5 core nodes, including the start point, 3 defect centers, and the end point.
[0064] When dividing the control area, the weld segment between the core nodes is divided into independent coupling agent control areas, such as node 1-node 2 as area A and node 2-node 3 as area B. Each area is marked with area characteristic parameters, such as area A with a weld width of 8mm, a reinforcement of 1.2mm, no defects, and an environmental interference coefficient of 1.0; area B with a weld width of 10mm, a reinforcement of 0.8mm, a gas hole defect, and an environmental interference coefficient of 1.1. Optimal detection path generation and coupling agent demand mapping.
[0065] The improved A* algorithm is used for path optimization algorithm selection, with the minimum total path length, coverage of all control areas, and highest priority of defect area detection as the objective function, and the heuristic function is set, such as the heuristic weight of the defect area being twice that of the non-defect area. Through path and coupling agent demand mapping, the generated optimal path is mapped to the coupling agent demand level of each control area.
[0066] For example, when planning a path with precise demand level around a defect, the probe moving speed needs to be reduced, such as 1mm / s, and more coupling agent adjustment time needs to be reserved; the path speed of the base material flat area can be increased to 3mm / s, and the coupling agent is supplied with the minimum consumption.
[0067] In the simulation of path execution in the digital twin carrier, if the uniformity of the coupling agent η < 0.85 in a certain area path, such as the path of the weld bead height mutation is too steep, the topology node adjustment link is returned, a transition node is added, such as adding a node at the height mutation, the control area is refined, and the path is regenerated until the path of all areas meets the coupling agent control basic condition η ≥ 0.85.
[0068] In each control area of the topology path division, the agent is trained to meet the constraint conditions and optimization control target, that is, the coupling agent control system generates real-time coupling agent control parameters, including supply amount, supply rate and adjustment time;
[0069] S301: Historical data pre-training, fast initialization strategy, extract 1000 groups of state and optimal action data from the digital twin historical database, such as the optimal supply amount of 15 microliters per second in a certain area at 25 degrees Celsius without vibration, train the agent in a supervised learning manner, and let the agent quickly master the basic control logic, such as the supply amount of a high bead height area is 20% more than that of a flat area, to avoid a large number of invalid actions in the initial exploration stage.
[0070] S302: Dynamic working condition intensive training, adapt to complex scenes, simulate 20 typical complex working conditions in digital twin, including low temperature-10 degrees Celsius, high vibration 5 hertz, high temperature 55 degrees Celsius, and high humidity, high bead height weld with crack defects, etc. Let the agent dynamically explore the control action in each control area:
[0071] Sa1: The agent outputs a set of control actions, such as supply amount 18 microliters per second, rate of change and 1 microliter per second, according to the current state of the area, such as temperature 10 degrees Celsius and uniformity 0.82 2 ;
[0072] Sa2: The digital twin simulates the effect of the action in real time, and feeds back the new state, such as uniformity improved to 0.92 and pressure stabilized at 20 Newton;
[0073] Sa3: Calculate the score of this action according to the reward function, and update the strategy parameters of the agent using the time difference algorithm, such as adjusting the correlation weight of temperature and supply amount;
[0074] Sa4: Repeat the above process 10000 times until the agent can adjust the coupling agent state to uniformity ≥ 0.85 and supply amount ≤ 25 microliters per second within 100 milliseconds in more than 95% of the working conditions, and the training reaches a convergent state.
[0075] S304: Virtual-real migration verification to ensure actual effectiveness; import the trained strategy into the actual detection system, compare the digital twin simulation results with the actual detection results: if the actual detection of the coupling agent uniformity in a certain area is 5% lower than the digital twin simulation value, such as the simulation uniformity is 0.9 and the actual uniformity is only 0.855, the actual detection state data is returned to the digital twin, the environmental disturbance coefficient of the virtual environment is corrected, such as from 1.0 to 1.05, and the intelligent agent is retrained; repeatedly adjust until the deviation of virtual and actual results is not more than 3%, ensure that the strategy can be effectively stabilized in the actual scene.
[0076] The generation of regional coupling agent control strategy, after training, the intelligent agent can output regional and dynamic coupling agent control strategy according to the real-time state of the topological path. In the regional basic strategy, each control region corresponds to a set of basic control parameters, such as the basic supply amount of the flat base material region is 12 microliters per second, the supply rate change rate is ±1 microliter per second 2 , the adjustment time is 50 milliseconds; the basic supply amount of the defect-containing region is 18 microliters per second, the rate change rate is ±0.5 microliters per second 2 , the adjustment time is 30 milliseconds; the basic supply amount of the low-temperature residual high region is 22 microliters per second, the rate change rate is ±0.8 microliters per second 2 , the adjustment time is 20 milliseconds; through the dynamic adjustment mechanism, the digital twin real-time monitors the working condition changes, such as the temperature from 25 degrees Celsius to 5 degrees Celsius, the intelligent agent will automatically correct the strategy, and the supply amount of the low-temperature residual high region will be increased from 22 microliters per second to 24 microliters per second, to ensure that the uniformity is not less than 0.85. Emergency strategy: if the digital twin detects a sudden abnormality, such as the coupling agent uniformity in a certain area drops to 0.7, the intelligent agent will immediately output emergency actions, temporarily increase the supply amount to 30 microliters per second, which is the maximum hardware capacity and lasts for 50 milliseconds, and reduce the probe moving speed from 2 mm / s to 1 mm / s until the coupling agent state returns to normal.
[0077] Probe motion trajectory:
[0078] Weld 3D topology data extraction and trajectory planning foundation construction, the planning of the probe motion trajectory needs to take the actual geometric characteristics of the weld as the core basis. First, extract the weld 3D topology data from the digital twin carrier and preprocess it to lay a precise spatial reference for trajectory planning. Retrieve the complete 3D geometric parameters of the weld from the digital twin carrier, including the total length of the weld, the width variation along the length direction, such as gradually changing from 8mm to 12mm, the reinforcement distribution, such as 1mm reinforcement in the straight section and 1.5mm reinforcement at the inflection point, the 3D coordinates of defects, such as a 5mm deep and 2mm diameter pore 200mm away from the weld start point, and the weld orientation type, such as straight line segment, 90° bend segment, arc segment, and the flatness data of the base material surface, such as local protrusion ≤0.2mm. Divide the extracted 3D data into multiple probe motion units according to geometric feature consistency, i.e. the width, reinforcement, and orientation variation amplitude in each unit ≤5%.
[0079] For example: Unit 1: from the weld start point to 150mm, straight orientation, width 8±0.2mm, reinforcement 1±0.1mm, no defects, defined as a straight and defect-free unit.
[0080] Unit 2: from 150mm to 250mm, straight orientation, width 8.5±0.3mm, reinforcement 1.2±0.1mm, containing 1 pore, i.e. a defect unit.
[0081] Unit 3: from 250mm to 350mm, 90° bend segment, width 9±0.3mm, reinforcement 1.3±0.1mm, no defects, i.e. a curved unit.
[0082] The purpose of dividing the motion units is to allow subsequent trajectory planning to adjust parameters as needed, avoiding trajectory deviation caused by sudden changes in weld characteristics.
[0083] Global trajectory planning based on topology constraints, determine the overall motion framework, global trajectory planning aims to fully cover the weld, no missed detection, and high efficiency and low consumption. Combine the topology information in the digital twin with hardware constraints to generate a main motion path that covers the entire weld. The core constraints of the topology path are set by clearly defining the hard constraint conditions of the probe motion to ensure that the trajectory meets the engineering reality. The vertical distance between the probe and the weld surface needs to be stable at 0.5±0.05mm to ensure the thickness of the coupling agent layer and avoid damaging the weld due to being too close or causing coupling failure due to being too far away. The probe needs to be aligned with the center of the weld in the horizontal direction with an offset ≤0.3mm to avoid the sound beam deviating from the detection area.
[0084] The maximum moving speed of the probe during motion constraint is ≤5mm / s to prevent the coupling agent from splashing or not being supplied in time, and the minimum moving speed is ≥0.5mm / s to ensure sufficient ultrasonic signal sampling; the maximum curvature radius at the turning point is ≥10mm to avoid the probe from detaching from the weld due to inertia.
[0085] The trajectory needs to cover the entire area of the weld, including the fusion zone, i.e. the weld center extending 2mm to both sides; the defect unit needs to cover at least the center of the defect ±3mm to ensure that the defect is completely scanned.
[0086] The global backbone path generation adopts a topology node connection and path smoothing method to generate a global trajectory: by extracting topology nodes on the center line of each motion unit, 1 node is taken every 50mm for straight units, 1 node is taken every 10mm for defect units, and 1 node is taken every 5mm for curved units. The node density increases with the complexity of the weld, for example, unit 1, 150mm straight section extracts 3 nodes, unit 2, 100mm defect section extracts 10 nodes;
[0087] By connecting adjacent nodes with segmented Bezier curves, linear Bezier curves are used for straight units to ensure uniform motion, cubic Bezier curves are used for curved units to smoothly transition turns and avoid trajectory jumps, and quadratic Bezier curves are used for defect units to form a slight reciprocating scanning arc around the defect center, ensuring that the defect is detected multiple times;
[0088] Global path verification is performed, the backbone path is simulated in the digital twin carrier, and it is checked whether there are any constraints, such as whether the curvature radius of the curved unit is ≥10mm, whether the coverage of the defect unit meets the standard, if there are problems, adjust the node position, such as increasing the node spacing of the curved unit to enlarge the curvature, until the global path meets all the constraints.
[0089] After the global backbone path is determined, local trajectory optimization needs to be performed for complex areas such as defect units, weld inflection points and material mutation areas, combined with real-time data in the digital twin, to ensure detection accuracy; local trajectory optimization for defect units: for the defects marked in the digital twin, plan a focused local trajectory:
[0090] When adjusting the motion speed, the probe moving speed of the defect unit is reduced from the conventional 3mm / s to 1mm / s, ensuring that the ultrasonic probe has enough time to collect defect echo signals and avoiding missing defects due to excessive speed;
[0091] During the motion trajectory process, based on the three-dimensional coordinates of the defect center, a composite trajectory of cross and ring scanning is planned: first, scan the defect center ±3mm cross in the longitudinal direction along the weld length, then scan the defect center ±2mm cross in the transverse direction along the weld width, and finally, do ring scanning with the defect center as the center and 2mm as the radius, covering the possible extension area around the defect;
[0092] Probe posture fine-tuning: according to the defect depth, such as a 5mm deep pore, the optimal incidence angle of the ultrasonic beam is simulated through digital twinning, and the tilt angle of the probe is fine-tuned by ≤3° to ensure that the sound beam is perpendicular to the defect surface, thereby improving the echo signal strength; during local trajectory optimization of the bending / inflection unit, for areas where the weld orientation changes, the probe is prevented from detaching or decoupling; during speed gradient adjustment, the probe speed is gradually reduced from 3mm / s to 1.5mm / s at a distance of 5mm before approaching the inflection point, such as a 90° bend, and then gradually increased to 3mm / s after passing the inflection point, to prevent inertia from causing the probe to deviate;
[0093] If the digital twinning simulation finds that the curvature radius of the original curved trajectory is too small, such as 8mm < 10mm, then two additional topology nodes are added on both sides of the inflection point, a cubic Bezier curve is refitted, and the curvature radius is enlarged to 12mm, to ensure smooth turning of the probe;
[0094] During coupling pressure coordination, the coupling agent state data in the digital twinning is combined, such as a high excess height in the bending unit, and the fine-tuning parameters of the probe contact pressure are set simultaneously during trajectory planning, the pressure is increased from 20N to 22N when turning, and the fit of the probe and the weld is enhanced to prevent the coupling agent from accumulating or losing due to centrifugal force; local trajectory optimization of material mutation units: for the material boundary between the fusion zone and the base material, the trajectory is adjusted to adapt to the change in sound velocity; within a range of 2mm on both sides of the center of the fusion zone weld, the movement step of the probe is reduced from the regular 0.5mm to 0.2mm, increasing the number of sampling points to ensure that the ultrasonic signal differences caused by material changes are captured; if the digital twinning detects local protrusions in the fusion zone, such as 0.3mm, the trajectory in that area is offset by 0.1mm in the direction perpendicular to the protrusion, maintaining the distance between the probe and the weld surface at 0.5mm, to prevent sudden changes in coupling pressure.
[0095] Obtain weld detection data:
[0096] Before the probe starts moving according to the preset trajectory, the weld topological parameters of the current trajectory segment, such as width and excess height, are obtained through digital twinning, and the probe posture is calibrated simultaneously: based on the spatial coordinates of the trajectory, the vertical distance between the probe and the weld surface is adjusted to 0.5±0.05mm to match the coupling agent thickness requirement, and the horizontal direction is aligned with the trajectory centerline with a deviation of ≤0.3mm, to ensure that the sound beam is perpendicular to the weld along the trajectory direction; after starting, the probe movement mechanism moves at a constant speed according to the trajectory speed, such as 3mm / s for straight sections and 1mm / s for defect sections, and receives real-time feedback on the trajectory position, i.e. encoder accuracy ±0.01mm, to avoid data misalignment caused by deviation;
[0097] During the trajectory movement, each sensor collects ultrasonic data according to the trajectory step and the sampling frequency. The ultrasonic probe emits 10MHz high-frequency sound waves according to the trajectory step, synchronously receives the echo signal, and collects characteristics such as peak amplitude, rising edge time, and center frequency. One set of original ultrasonic data is generated at each step. Due to the reciprocating scanning of the trajectory, the sampling frequency is increased to 2 times to ensure that there is no omission of defect signals. Laser and pressure data: the laser displacement sensor collects the thickness of the coupling agent in real time at a frequency of 1kHz, corresponding to the coupling state of each position of the trajectory. The multi-unit pressure sensor synchronously records the contact pressure. The data of the two are bound according to the trajectory timestamp to reflect the coupling stability of each point of the trajectory. The industrial camera triggers shooting at key nodes of the trajectory, such as inflection points and defect centers. One frame of weld contour image is collected every 50mm of the trajectory to supplement the visual features of the trajectory coverage area and assist in verifying the spatial correlation of the ultrasonic data.
[0098] By associating the timestamp and spatial coordinates of the trajectory, the multimodal original data is accurately associated. The encoder time of the trajectory movement is used as the reference to uniformly calibrate the sampling timestamps of ultrasonic, laser, and image data, with a deviation of ≤1μs. This avoids data misplacement caused by sensor response delay. Each set of collected data is bound to the three-dimensional coordinates (X / Y / Z) of the trajectory, such as the ultrasonic data at X=200mm, Y=5mm of the trajectory. This corresponds to the weld depth information marked at this coordinate point. The laser data marks the coupling agent thickness at this coordinate to form a one-to-one mapping relationship between the trajectory coordinates and the multimodal data, which can trace back the complete detection data at a certain position.
[0099] The associated original data is preprocessed in real time. Ultrasonic data is filtered by a band-pass filter to eliminate noise and extract the peak, time, and frequency characteristics of the defect echo. Laser data calculates the uniformity of the coupling agent at each point of the trajectory and compares the thickness deviation of adjacent 3 points. Pressure data calculates the stability within 100ms. Image data extracts the weld width and excess height of the trajectory corresponding area through sub-pixel algorithm and corrects the slight deviation of the trajectory coordinates, such as updating the data association coordinates when the actual weld center deviates from the trajectory by 0.1mm.
[0100] When the trajectory enters the defect unit and the digital twin pre-marked area, the data enhancement mechanism is triggered: the probe movement speed is reduced to 1mm / s, the ultrasonic sampling step is reduced to 0.1mm / step, and the sampling density is increased. The sampling frequency of the laser sensor is increased to 2kHz to monitor the thickness change of the coupling agent in real time, ensuring the coupling stability in the defect area. After the collection is completed, the ultrasonic data of the trajectory segment is compared with the pre-stored defect characteristics of the digital twin. If the signal difference is >10%, the trajectory is retraced by 5mm for re-collection until the data consistency meets the standard.
[0101] After the trajectory ends, all coordinate and data pairs are integrated to form a complete weld detection dataset: the data coverage of each segment of the trajectory is checked, such as flat segments ≥ 99%, defect segments 100%, and missing trajectory points, such as missing samples caused by vibration, are controlled to be backtracked and resampled along the original trajectory; the integrated data is stored in order according to the trajectory and is associated with the weld model in the digital twin to provide a two-dimensional detection basis of position and characteristics for subsequent defect analysis and model iteration.
[0102] The AI algorithm optimization module obtains weld material, environmental parameters, and multi-modal perception data, uses a reinforcement learning algorithm to establish a mapping relationship of multi-factor coating parameters, and quantifies the uncertainty of weld detection data through Bayesian modeling;
[0103] Establish a mapping relationship of multi-factor coating parameters:
[0104] Historical data and target working conditions are retrieved from the digital twin carrier, such as high temperature, high vibration, low temperature, and large excess weld, a virtual training environment is constructed, a combination of historical material, environmental, and perception data is input, a supervised learning is used to initialize the agent strategy, the agent quickly masters the basic logic, such as the supply amount in the fusion zone is 20% higher than that in the base material zone; 1000 typical working conditions are simulated, the agent explores actions in the virtual environment, such as adjusting the supply amount, rate, and coating area, and iterates the strategy parameters according to the reward function, and a preliminary mapping sketch of material, environment, perception, and coating parameters is formed; several complex working conditions are simulated in the digital twin, such as -10℃ low temperature + 5Hz vibration, 55℃ high temperature + crack defect, high humidity, and large excess weld, the agent outputs coating actions according to the current material, environment, and perception state, such as supply amount 18μL / s, rate change rate +1μL / s 2 , and defect area coating is selected;
[0105] The digital twin provides real-time feedback on action effects, such as coupling agent uniformity from 0.8 to 0.92, ultrasonic peak amplitude enhancement, and reward score calculation;
[0106] The time difference algorithm is used to update the agent strategy parameters, and after 10000 iterations, the agent can output coating actions that meet the uniformity ≥ 0.85 + usage ≤ 25μL / s within 100ms in more than 95% of the working conditions; the trained strategy is imported into the actual detection system, and the results of digital twin simulation and actual detection are compared:
[0107] If the actual coupling agent uniformity is 5% lower than the simulation value, such as simulation 0.9, actual 0.855, the actual data is fed back to the digital twin, and the associated model of material, environment, and perception in the virtual environment is corrected;
[0108] Iterate until the virtual-real result deviation is ≤3%, ensuring that the strategy is stable in the actual scene; the trained reinforcement learning model can take the weld material, environmental parameters, and multi-modal perception data as input and output the optimal coupling agent application parameters, supply amount, supply rate, and application area, forming a dynamic mapping relationship:
[0109] When the input fusion zone material has a hardness of 250 HV, a sound speed of 5900 m / s, an environmental temperature of -5°C, and a coupling agent uniformity of 0.8, the model outputs a supply amount of 22 μL / s, a rate change rate of +0.8 μL / s, and an application area selection of the fusion zone. 2
[0110] When the input base material zone material has a hardness of 180 HV, a sound speed of 5950 m / s, an environmental temperature of 25°C, and a coupling agent uniformity of 0.9, the model outputs a supply amount of 12 μL / s, a rate change rate of ±1 μL / s, and an application area selection of the base material zone. 2
[0111] Through this mapping relationship, dynamic adaptability is achieved. If the weld material ages, such as the hardness increasing to 260 HV or the environmental parameters changing, such as the temperature rising to 30°C, the model will automatically adjust the output application parameters, always ensuring optimal coupling agent application effect, providing a stable coupling basis for the accuracy of ultrasonic testing.
[0112] Quantify the uncertainty of weld detection data:
[0113] The four types of uncertainty sources of weld detection data include sensor errors, coupling agent state fluctuations, weld material non-uniformity, and environmental interference. Sensor errors: the peak amplitude, rising time, and other characteristics of the ultrasonic probe have measurement noise, such as signal fluctuations caused by electronic component thermal noise. The coupling agent thickness measurement of the laser displacement sensor has a random error of ±0.005 mm. The sub-pixel profile extraction of the industrial camera has a deviation of ≤0.02 mm.
[0114] Coupling agent uniformity changes, such as local thickness deviation leading to ultrasonic beam propagation path distortion, causing uncertainty in defect positioning data. The sound speed of the coupling agent changes with temperature, such as the sound speed change rate of propylene glycol type coupling agent being 0.002 / °C, affecting the accuracy of defect depth calculation.
[0115] The sound speed and density of the fusion zone, heat-affected zone, and base material differ, such as the sound speed of the steel fusion zone being 1% to 3% lower than that of the base material, leading to random fluctuations in ultrasonic echo propagation time and reflection intensity. Microscopic organizational defects within the weld, such as micro pores and grain boundaries, introduce signal-induced uncertainty.
[0116] Temperature fluctuation in-10-55℃ affects the sensor performance, such as the piezoelectric ceramic sensitivity of the ultrasonic probe; vibration in 0-5Hz leads to the fluctuation of the contact pressure between the probe and the weld, making the coupling state unstable, and further affecting the consistency of the detection data.
[0117] Based on historical detection data and domain knowledge, the prior probability distribution of each uncertainty factor is constructed. For the error of ultrasonic sensor, the ultrasonic echo data of 1000 groups of defect-free welds are statistically analyzed, and the prior distribution of peak amplitude is fitted as a normal distribution , wherein is the average amplitude, is the standard deviation of amplitude, and the prior distribution of rising edge time is a lognormal distribution; for the fluctuation of coupling agent state, the coupling agent thickness data under different uniformity are collected, and the prior distribution of coupling agent thickness is fitted as a Beta distribution, reflecting the probability characteristics of uniformity; combined with the experimental data of temperature and sound velocity, the prior distribution of sound velocity is established as a linear regression distribution, which is a probability model of linear change of sound velocity with temperature; for the inhomogeneity of weld material, through metallographic analysis and sound velocity measurement, the sound velocity distribution of fusion zone, heat affected zone and base material is statistically analyzed and fitted as a normal distribution , , ; the prior distribution of microstructure defects is fitted as a Poisson distribution through defect rate statistics, such as the number of pores per square meter of weld; for environmental interference, the measurement error of temperature sensor is statistically analyzed as a normal distribution N(0, 0.05 2 ), and the amplitude error of vibration sensor is a Rayleigh distribution, and the correlation model between environmental parameters and detection data error is established, such as the probability distribution of ultrasonic sound velocity change of 0.002 times per 1℃ change of temperature.
[0118] A hierarchical Bayesian model is constructed to decompose the uncertainty of weld detection data into four levels: sensor layer, coupling agent layer, material layer and environment layer, and each level is related to each other through probability relationship:
[0119] The sensor layer describes the measurement error distribution of ultrasonic, laser and camera sensors, such as the probability distribution of ultrasonic peak amplitude and laser coupling agent thickness; the coupling agent layer describes the influence of coupling agent state on sensor data, such as the associated probability model of coupling agent uniformity and ultrasonic amplitude, and the associated probability model of coupling agent sound velocity and temperature; the material layer describes the influence of weld material on ultrasonic propagation, such as the associated probability model of defect depth and ultrasonic rising edge time, coupling agent sound velocity, and the probability distribution of sound velocity in different material regions; the environment layer describes the influence of temperature and vibration on sensors and coupling agents, such as the probability distribution of temperature and vibration amplitude, and the associated probability model of vibration and coupling pressure fluctuation.
[0120] The multi-modal perception data including ultrasonic features, laser thickness, camera profile, environmental parameters, etc. are taken as observation data, and a likelihood function is constructed in combination with the prior distribution of each level. The posterior distribution of the model parameters is calculated by combining the observation data with the prior distribution through Bayes' theorem, i.e. the probability distribution of the parameters after considering the observation data;
[0121] The posterior distribution solving and uncertainty quantification are solved by using the Markov Chain Monte Carlo method: the initial value of the model parameter is initialized, the average value, standard deviation and the like of the historical data are used, for each parameter, a candidate value is randomly generated in the neighborhood of its prior distribution, the acceptance probability of the candidate value is calculated, if the probability is greater than a random number, the candidate value is accepted, otherwise the original parameter value is retained, the above process is repeated a large number of times, the preheating samples are discarded, and the remaining samples are used to estimate the statistical characteristics of the posterior distribution;
[0122] Uncertainty quantification index output:
[0123] Three types of uncertainty quantification indexes are extracted from the posterior distribution samples, the mean and the standard deviation: the posterior mean of each detection data such as defect depth and ultrasonic amplitude is calculated, representing the most likely true value and the standard deviation, representing the uncertainty size, the 95% confidence interval is calculated, reflecting that the data has a 95% probability of falling within the interval, and the posterior probability density curve is drawn to intuitively show the uncertainty distribution pattern of the detection data.
[0124] In defect recognition, the recognition logic is adjusted in combination with the uncertainty quantification result, for the area with small uncertainty, the conventional threshold is used for defect judgment, for the area with large uncertainty, the recognition threshold is increased and the detection times are increased to reduce the misjudgment rate; in defect positioning, the posterior mean is taken as the final position and the confidence interval is taken as the positioning error range.
[0125] Uncertainty fusion of digital twin carrier:
[0126] The uncertainty quantification result is integrated into the digital twin carrier to realize two-dimensional modeling of certainty and uncertainty, and the confidence interval of the position of each defect is marked in the three-dimensional geometric model of digital twin, the detection effect under different uncertainty scenarios is simulated, and the basis for detection strategy optimization is provided, such as increasing the coupling agent supply amount in the area with large uncertainty.
[0127] Based on the uncertainty quantification result, the parameters of the detection system are dynamically adjusted: if the uncertainty of the ultrasonic amplitude in a certain area is large, the transmission voltage of the ultrasonic probe is automatically increased to improve the signal-to-noise ratio and reduce the uncertainty. If the uncertainty of the coupling agent thickness is large, adjust the coupling agent supply strategy, such as increasing the supply amount, reducing the probe moving speed, and improving the uniformity of the coupling agent.
[0128] The above-described embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented by software, the above-described embodiments can be implemented in whole or in part in the form of a computer program product. A person of ordinary skill in the art can be aware that units and algorithm steps of the examples described in connection with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether the functions are performed by hardware or software depends on the specific application and design constraints of the technical solutions.
[0129] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, and can be located in one place or distributed on multiple network units. Part or all of the units can be selected to achieve the purpose of the embodiments.
[0130] The above describes only specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.
Claims
1. An intelligent non-destructive testing system for industrial welds, characterized in that: The system includes: The perception registration module acquires sensor data and obtains multimodal perception data through a dynamic spatiotemporal registration algorithm; the sensor data includes contour perception, spatial perception, and morphological perception. The weld model construction module transforms multimodal sensing data into a three-dimensional geometric model, while integrating the acquired weld material and environmental parameters to construct a digital twin carrier. The process of constructing the digital twin carrier is as follows: By deploying multiple types of environmental sensors to collect environmental condition data in real time, a quantitative correlation model between environmental condition data and detection is established, and the real-time data of environmental sensors is bound to the detection period of the digital twin carrier through timestamps. The decision control module, based on the constructed digital twin carrier, uses topology path planning and constraint-enhanced reinforcement learning to generate coupler control strategies and probe motion trajectories, and acquires weld detection data based on the probe motion trajectory. The enhanced reinforcement learning is based on a virtual decision-making environment constructed from digital twins, combined with coupling agent control constraints, and verified through virtual-real transfer to achieve reinforcement learning that satisfies the optimization of accuracy and usage. The AI algorithm optimization module acquires weld material, environmental parameters, and multimodal perception data, establishes a mapping relationship between multi-factor smearing parameters, and quantifies the uncertainty of weld detection data through Bayesian modeling. The process of establishing the mapping relationship of multi-factor application parameters is as follows: Historical data and target working conditions are retrieved from the digital twin carrier to construct a virtual training environment. Several complex working conditions are simulated in the digital twin for the intelligent agent to explore dynamically. The trained strategy is imported into the actual detection system. By comparing the results of the digital twin simulation with the actual detection, the trained reinforcement learning model can take weld material, environmental parameters, and multimodal perception data as inputs and output the optimal coupling agent application parameters to form a dynamic mapping relationship. The process of quantifying the uncertainty of weld seam detection data through Bayesian modeling is as follows: Based on historical detection data and domain knowledge, prior probability distributions of various uncertainty factors are constructed. A hierarchical Bayesian model is built to decompose the uncertainty of weld detection data into four levels: sensor layer, coupling agent layer, material layer, and environment layer. Each level is interconnected through probability distributions, and multimodal sensing data is used as observation data. Combining the prior distributions of each level, a likelihood function is constructed. The Markov chain Monte Carlo method is used to solve the posterior distribution. Three types of uncertainty quantification indicators are extracted from the posterior distribution samples. During defect identification, the identification logic is adjusted based on the uncertainty quantification results, and the uncertainty quantification results are integrated into the digital twin carrier. Based on the uncertainty quantification results, the detection parameters are dynamically adjusted.
2. The intelligent non-destructive testing system for industrial welds according to claim 1, characterized in that: The process of acquiring the sensor data is as follows: Images of the weld area were acquired, and subpixel-level contour features of the weld were extracted using the gray-scale centroid method. The distance between the probe and the weld surface, the thickness of the coupling agent, and the spatial position of the probe are collected. At the same time, the coupling pressure distribution data are collected to obtain information on the spatial position and contact state of the weld and the probe. By transmitting and receiving ultrasonic signals using a high-frequency ultrasonic probe, the peak amplitude, rise time, and center frequency characteristics of the defect echo are extracted, while environmental parameters are collected using environmental sensors.
3. The intelligent non-destructive testing system for industrial welds according to claim 1, characterized in that: The process of acquiring the multimodal sensing data is as follows: The acquisition of multi-source sensor data employs a dynamic spatiotemporal registration algorithm for time synchronization and spatial registration, resulting in multimodal sensing data.
4. The industrial weld non-destructive intelligent flaw detection system according to claim 2, characterized in that: The process of converting multimodal sensing data into a three-dimensional geometric model is as follows: Multi-source data is preprocessed and unified into a 3D point cloud data format. Based on the preprocessed subpixel edge data, sampling points are obtained along the weld length direction with a step size of 0.1mm. The Y coordinate of each sampling point is determined by the fitting curve of the subpixel edge, and the Z coordinate is initially set to the height of the base material surface to generate a 2D contour point cloud of the weld surface. Combined with the coupling agent thickness, the surface contour point cloud is stretched along the Z-axis. The initially generated 3D point cloud is denoised, downsampled, and smoothed. The Delaunay triangulation algorithm is used on the preprocessed 3D point cloud to construct a 3D triangular mesh model of the weld area. By extracting the control vertices and weight factors of the defect 3D point cloud, a smooth surface is fitted.
5. The intelligent non-destructive testing system for industrial welds according to claim 1, characterized in that: The process of generating the coupling agent control strategy is as follows: The multi-dimensional data in the digital twin carrier is transformed into a virtual decision-making environment for topology path planning and reinforcement learning, and the influencing factors and constraint boundaries of coupling agent control are determined. Based on the digital twin carrier, the weld topology features are obtained, the coupling agent control area is divided, and a detection path covering the entire weld is generated. In each control area of the topology path, the agent is trained to generate real-time coupling agent control parameters.
6. The industrial weld non-destructive intelligent flaw detection system according to claim 5, characterized in that: The process of generating the probe's motion trajectory is as follows: The three-dimensional topology data of the weld is extracted from the digital twin carrier and preprocessed. Combined with the topology information and hardware constraints in the digital twin, a main motion path covering the entire weld is generated. After the global main path is determined, local trajectory optimization is performed by combining real-time data in the digital twin. The planned probe motion trajectory is then simulated and verified in the digital twin carrier.
Citation Information
Patent Citations
Welding robot track planning method
CN118438446A
Ultrasonic scanning microscope detection system device positioning device
CN211978773U