Research Of Laser Vision Seam Detection And Tracking System Based On Depth Hierarchical Feature

In recent years,with the coordinated development of various technologies such as robotics and information technology,manual welding methods with many shortcomings are not suitable for modern manufacturing,and are replaced by welding processing systems based on welding robots,but limited by the intelligence level of the current robots,most robots still remain in the fixed mode of "manual teaching-memory reappearance" when carrying out welding operations.In the unstructured welding environment,due to factors such as thermal deformation,vibration displacement,and clamping error,there is an inevitable deviation between the actual trajectory of the weld and the teaching trajectory,but the robot in this working mode cannot make corresponding corrective actions according to the change of the trajectory,thus affecting the final welding quality and accuracy.For this reason,this paper studies the weld seam tracking technology required for robots to realize environmental information perception and self-cognition functions,adopts the “Sensing-Decision-Execution” model,based on the laser vision sensor which uses laser as an information carrier and visual perception as a means of design,by developing acquisition module of laser vision information,information processing module of welding seam image,control module of weld seam tracking and other functional units,a laser vision seam detection and tracking system based on depth hierarchical feature was constructed.The research of this topic has won the funding from National Science and Technology Major Project “development and industrial engineering of five thousand robots with completely independent intellectual property rights for machine tool automatic production”(No.2015ZX04005006)and Guangdong Province Science and Technology Major Project "research on integrated technology for CNC machine tools and robots "(No.2014B090920001).The contents of the research mainly include:(1)Based on the analysis of the overall framework,working mechanism and design requirements of the automatic seam tracking system,a laser vision sensor for real-time acquisition of weld seam position information was developed.According to the spatial distribution relationship between the laser vision sensor and the robot in the vision measurement system,combining the perspective projection imaging model,a measurement mathematical model with the function of reconstructing 3D position information from 2D image information was deduced and established in detail.Based on Halcon software,the calibration algorithm of the industrial camera mounted on the sensor is studied,so as to realize the mapping of the two-dimensional visual coordinate to the three-dimensional space position in the camera coordinate system.The calibration algorithm of the laser plane was studied.A large number of homogenous laser spots were acquired using planar targets by position and orientation sampling,the plane parameters were obtained by plane fitting using orthogonal regression method.For the Eye-in-Hand system with the sensor as the eye and the robot as the hand,a hand-eye calibration algorithm by measuring a fixed point from different poses is proposed to solve the hand-eye relationship matrix between the camera and the robot,and finally achieve the conversion of two-dimensional vision coordinates in the pixel coordinate system to three-dimensional space coordinate in the robot base coordinate system.(2)In view of the fact that image processing algorithms in most weld seam tracking systems are extremely sensitive to noise such as spatters and arcs,it is difficult to quickly and rapidly detect weld seams and accurately position them from time-series images that contain strong noise pollution,leading to a problem of reduced tracking accuracy,the welding seam detection and tracking algorithm was studied from two aspects: weld feature extraction,weld detection and positioning.Based on the study of several key basic theories related to deep convolutional neural networks,combined with the strong feature expression capability and self-learning function of deep convolutional neural networks,a weld seam detection and tracking algorithm based on deep hierarchical features was proposed.The proposed image processing algorithm extracts the target's high-level deep abstract features through the deep convolutional neural network structure,which overcomes the limitations of traditional methods which consider from pixel-level feature analysis and single geometric feature recognition or statistical decision,and reduces the ambiguity of the features;Based on the continuity of the motion of the feature points of the adjacent frames and the correlation of the laser stripe structure information,a real-time seam searching and positioning strategy based on multi-correlation filter cooperative detection mechanism is proposed to solve the technical problem of intelligent detection and positioning of seam under large noise.(3)The automatic tracking and control system of welding seam shows obvious coupling,strong nonlinearity and timely degeneration.In order to avoid the periodic chattering problem of welding torch following the calculation trajectory as far as possible,the control system should have some dynamic compliance and self-adaptability.Based on the traditional PID control method with wide adaptability and high control precision,drawed on the experience of the immune feedback response process of organisms,and designed the fuzzy controller which can learn the unpredicted information about environment in the control process,a fuzzy immune adaptive intelligent control algorithm suitable for complex welding environment was proposed.The designed intelligent tracking controller has robustness and global self-search ability,which can optimize the control parameters in real time,give the characteristics of domain adaptive scaling and fuzzy rules self-adjusting to the system,and achieve smooth trajectory tracking.(4)A welding seam tracking experiment platform was built,and a series of weld seam automatic tracking experiments were performed using the proposed algorithm in the actual welding environment.The experimental results showed that under the strong arc and splash interference,the sensor measurement frequency reached 20 Hz,the average absolute tracking error was less than 0.25 mm,the maximum tracking error didn' t exceed 1 mm,and the welding torch end ran smoothly during welding.This proves that the system can realize real-time seam tracking,accurately and reliably detect the Welding seam position from different degrees of noise pollution images and achieve smooth welding.Compared with the best accuracy range 0.3mm ~ 0.5mm that can be achieved by most current automatic weld tracking systems,it has greater performance advantages and can meet high quality welding requirements.