On-line 3D Inspection System For Line-heating Technique Based On Deep Learning

The shipbuilding industry has always been an important factor to measure a country’s development and technology,as well as an important indicator of the economy.Since the reform and opening-up,China has established the strategic goal of becoming the world’s top shipbuilding country.What’s more,China’s shipbuilding industry is developing by leaps and bounds in the late years.According to the statistics of the Ministry of Industry and Information Technology(as of January 15,2021),China’s shipbuilding completed,new orders and handheld orders respectively accounted for 43.1%,48.8% and 44.7% of the global total,which means that China’s shipbuilding capacity is the first around the world.The processing and manufacturing of hull panels plays an important role in the whole shipbuilding,and the processing man-hours account for 25%～35% of the man-hours of the whole shipbuilding.Today the processing of the hull shell mainly uses the Line-heating Technique,which formed the steel plate plastic by heating and tracking cooling.In the precision and quality testing after molding,the sample box and living collaterals is still used for testing,but this detection method is inefficient: accuracy cannot be guaranteed;templates are vulnerable to wear and damage because they are easily affected by the environment.However,it would have been expensive to redesign and make samples again.Therefore,it is of great significance and economic value to develop an automatic Line-heating Technique testing system.In this paper,aiming at the inspection problem of the traditional Line-heating Technique in the shipyard and combining the 3D visual technology and deep learning approach,the point cloud registration is proposed to complete the detection task of water-fire bending plate forming firstly and then a new point cloud registration model based on deep learning is designed;finally,an online 3D inspection system of Line-heating Technique based on deep learning is developed.First of all,this paper introduces the method of non-contact collection of 3D point cloud models and the appropriate point cloud outlier removal and down-sampling algorithm are selected for the actual test model.Then,the point cloud registration network model based on PointConv and the Siamese network structure is introduced in detail;moreover,the validity of the model is verified by experimental comparison.And the experimental results show that this model is not only superior to traditional point cloud registration algorithms such as ICP,but also superior to the learning-based Pointnet LK algorithm.Eventually,On-line 3D inspection system for line-heating technique based on deep learning is designed and developed which is convenient for practical use in processing.In this paper,the problem of line-heating bending plate forming detection is transformed into the point cloud registration problem,and then an online inspection system for line-heating technique is developed by deep learning and 3D vision algorithms.As a computer-aided tool,this system can reduce the workload in the actual operation,effectively improve the detection accuracy and help the processing personnel to formulate the next processing plan.