Weld Process Modeling, Monitoring, Teaching, and Control using ...
Weld Process Modeling, Monitoring, Teaching, and Control using Artificial Neural Networks: Improving Navy and Commercial Shipbuilding for the Future Author name(s): Jerald E. Jones, Ph.D. (Membership: V) EnergynTech. Inc. P.O. Box 151241 Denver, CO 80215-9241 720-232-3490 303-279-5286 FAX [email protected]
Valerie L. Rhoades EnergynTech.Inc. P.O. Box 151241 Denver, CO 80215-9241 720-232-3490 303-279-5286 FAX [email protected]
Mark D. Mann EnergynTech.Inc. P.O. Box 151241 Denver, CO 80215-9241 720-232-3490 303-279-5286 FAX
Category: Information Systems Technical Paper and Presentation ABSTRACT: The future of welding for shipbuilding is increasing automation. As quality and productivity requirements continue to increase, automated systems will be required to meet those demands. What processes cannot be automated will require more highly skilled welders. Today, the shipbuilding industry depends heavily on manual welders to provide real-time process control and quality assurance. That workforce that existed in other industries, which have already moved to automated welding, has largely been replaced with computerized monitoring systems and control systems, and computer modeling for process optimization. A major reason why shipbuilding has lagged behind other industries in the U.S. and Europe is the nature of shipbuilding here. Mass produced ships, produced in Asia, are highly automated, but Asian shipbuilders depend on careful robot programming for much of their process control and quality monitoring. This is also true of the automobile industry in the U.S. and abroad. The commonality is production of large quantities of identical components. Shipbuildingin the U.S. and, to a lesser extent in Europe, has a large component of few-of-a-kind production. In this environment robot programming and re-programming for so many different kinds of parts is prohibitively expensive. For example, when Chrysler did a plant changeover for a new Dodge Neon model a few years ago, they reported the cost as $2.4 billion, a significant percentage of which was the cost of re-programming the dozens of robots in the assembly plant.
A DARPA Program a decade ago concentrated on flexible manufacturing – basically, robots that program themselves, depend heavily on sensors to be “smart” about their actions and decisions, and perform quality control with little, if any, human involvement. A technology was investigated at that time that depended heavily on computer models. This trend of Computer Aided Design (CAD), Computer Aided Manufacturing (CAM), and Finite Element Analysis modeling (FEA) has made automated fabrication much more prevalent in shipbuilding. But, the welding process has not been automated very successfully. Metal melting and solidification is a physical process which is described by a class of differential equations which have no closed form solution – it is known in Classical Mathematics as the “Stephan Problem”. The heat of fusion is the principal difficulty in modeling this process – it is created as a metal solidifies and is absorbed as a metal melts. The phase change of solid (very ordered, low energy, crystal lattice) to liquid (very disordered state of the atoms) causes a significant energy (heat) release (changing from liquid to the low energy state of solid) and the opposite during melting. Since there is no closed form solution, FEA models have to use iterative solutions to solve the coupled heat flow and fluid flow problem and can take hours to converge to a solution. In order to do process planning and optimization prior of welding, and real time monitoring for control and quality assurance, it is critical that a model exist which can predict the size and shape of a weld in real-time – that is, it must be able to respond in milliseconds or faster. Standard Artificial Neural Networks, such as the Backpropogation Algorithm, can produce models that are more than fast enough. However, these models do not have sufficient accuracy to be used in real-time control or monitoring of welding. In the DARPA Flexible Automation program, a group of Ph.D. Scientists and Engineers,began the development of a new Artificial Neural Network technology, called P/NA3. Today, Neural Networks produced using the P/NA3 technology are sufficiently accurate and fast to be able to run in real-time during welding or, embedded into specialized search algorithms, to optimize the welding process for planning of robotic welds. This technology has been used in the automobile industry; however, since robotic welding is relatively new to shipbuilding, that is a new area of application. As flexible automation application increases, the need for fast and accurate models of the welding process will grow significantly. This paper will describe the methods used to develop P/NA3 Neural Networks, and discuss the application of P/NA3 Neural Networks for welding optimization for planning and control. In addition, a P/NA3 Neural Network real-time weld quality monitoring system, that can process over 200,000 data points per second, will be presented. While this technology is being used for the mass production environment of the Automobile Industry, its introduction to shipbuilding, and application to few-of-a-kind manufacturing is completely new. In addition to being used for robotic welding, this same technology can be applied, in virtual reality training environments, to train more highly skilled welders. A welder can learn the exact relationship between their actions with the welding torch and the resulting weld shape and size – in a virtual environment. This adds a completely new dimension to welder training. While this
is possible by cutting a cross-section of a real weld – that is costly and time consuming, just to get one single cross-section. As shown in figure 5, in a virtual training environment, the P/NA3 Neural Network welding process model can allow the student to instantly “see” their weld in cross-section all along the weld. The paper will discuss the development, using Design of Experiments methods, of a P/NA3 Neural Network welding process model. This will include an advanced technology software that employs Artificial Intelligence technology to generate a mathematical representation of the weld shape. The Neural Network is developed using the input parameters of the welding process (e.g., voltage, wire feed rate, travel speed, torch angles, etc.) and the resulting weld shapes. The software that uses the neural network models for production pre-planning and process optimization will be examined. Real-time weld process control and real-time weld process quality assurance monitoring software will be discussed and examples provided, showing how this technology can be used in shipbuilding. This entails software that takes inputs of sensor data on multiple channels, uses the P/NA3 Neural Network welding process model to predict the weld quality. Then, in real-time the system can use the P/NA3 Neural Network process model to search for the parameter changes that will improve the weld quality – and dynamically adjusts the process “on-the-fly.” This real-time quality monitoring is done by the human brain in manual welding – the P/NA3 Neural Network welding process modeling technology can do that same task for robotic welding.
Figure 1.NAMeS Weld Measurement System, used to capture the intricate shape of a weld in a special mathematical structure that can be used to develop an Artificial Neural Network Model.
3
Figure 2.P/NA Neural Network Welding process model. The cross-section shape of the weld changes instantaneously as any of the process parameter slidebars is moved.
3
Figure 3.Engineering planning and optimization software with embedded P/NA Neural Network Welding process model. The response surface allows the engineer to quickly identify the process parameter values for a flexible automation welding robot which will produce the exact weld crosssection shape that is necessary to meet the mechanical property requirements. The search software can calculate the shape of over 10 million sets of weld parameter values in less than 1 second and can provide the engineer with the precise control parameters to produce a weld that is the exact shape and size required. Note: after applying the search algorithm system, the weld shape meets all of the requirements (see all green values in the “output window” of figure 2. The system calculates over 100 weld shape engineering outputs for each of the 10 million sets of weld parameters values, then selects the shape which is closest to the shape requested by the design engineer. Then the software can download these parameters into the robot controller 3 automatically. The P/NA Weld process model, allows an engineer to create anoptimized arc welding robot control program, just as a manufacturing engineer uses CAD data to automatically produce a control program for a machining system, such as an automated lathe or mill or laser.
3
Figure 4.The weld process monitoring / control system with embedded P/NA Neural Network model. The system can produce a 3D image of the weld. In the screen capture on the left, the plates being welded can be turned transparent, so the inside of the weld can be examined, similar to a radiograph of the weld. The system can monitor up to four robotic welding systems simultaneously, collecting nearly 1 million data points per second and converting that data to the precise weld shape, in real time, and compare that weld shape to the Mil. Spec. or Code. The 3D image is color coded: good (green), marginal (yellow), and reject (red) areas of the weld produced by the robot. This data can be instantly transmitted to a remote computer in an engineers’ office.
3
Figure 5. The Virtual welder training system with embedded P/NA Neural Network Welding process model. The student can move the torch to any location along the weld and see instantly the cross section of the weld at that location. Since the system stores all of the welders movements and torch angles, the student can “see” instantly the relationship between their welding technique and the shape of weld that they are producing, at any point along the weld.
Valerie L. Rhoades EnergynTech.Inc. P.O. Box 151241 Denver, CO 80215-9241 720-232-3490 303-279-5286 FAX [email protected]
Mark D. Mann EnergynTech.Inc. P.O. Box 151241 Denver, CO 80215-9241 720-232-3490 303-279-5286 FAX
Category: Information Systems Technical Paper and Presentation ABSTRACT: The future of welding for shipbuilding is increasing automation. As quality and productivity requirements continue to increase, automated systems will be required to meet those demands. What processes cannot be automated will require more highly skilled welders. Today, the shipbuilding industry depends heavily on manual welders to provide real-time process control and quality assurance. That workforce that existed in other industries, which have already moved to automated welding, has largely been replaced with computerized monitoring systems and control systems, and computer modeling for process optimization. A major reason why shipbuilding has lagged behind other industries in the U.S. and Europe is the nature of shipbuilding here. Mass produced ships, produced in Asia, are highly automated, but Asian shipbuilders depend on careful robot programming for much of their process control and quality monitoring. This is also true of the automobile industry in the U.S. and abroad. The commonality is production of large quantities of identical components. Shipbuildingin the U.S. and, to a lesser extent in Europe, has a large component of few-of-a-kind production. In this environment robot programming and re-programming for so many different kinds of parts is prohibitively expensive. For example, when Chrysler did a plant changeover for a new Dodge Neon model a few years ago, they reported the cost as $2.4 billion, a significant percentage of which was the cost of re-programming the dozens of robots in the assembly plant.
A DARPA Program a decade ago concentrated on flexible manufacturing – basically, robots that program themselves, depend heavily on sensors to be “smart” about their actions and decisions, and perform quality control with little, if any, human involvement. A technology was investigated at that time that depended heavily on computer models. This trend of Computer Aided Design (CAD), Computer Aided Manufacturing (CAM), and Finite Element Analysis modeling (FEA) has made automated fabrication much more prevalent in shipbuilding. But, the welding process has not been automated very successfully. Metal melting and solidification is a physical process which is described by a class of differential equations which have no closed form solution – it is known in Classical Mathematics as the “Stephan Problem”. The heat of fusion is the principal difficulty in modeling this process – it is created as a metal solidifies and is absorbed as a metal melts. The phase change of solid (very ordered, low energy, crystal lattice) to liquid (very disordered state of the atoms) causes a significant energy (heat) release (changing from liquid to the low energy state of solid) and the opposite during melting. Since there is no closed form solution, FEA models have to use iterative solutions to solve the coupled heat flow and fluid flow problem and can take hours to converge to a solution. In order to do process planning and optimization prior of welding, and real time monitoring for control and quality assurance, it is critical that a model exist which can predict the size and shape of a weld in real-time – that is, it must be able to respond in milliseconds or faster. Standard Artificial Neural Networks, such as the Backpropogation Algorithm, can produce models that are more than fast enough. However, these models do not have sufficient accuracy to be used in real-time control or monitoring of welding. In the DARPA Flexible Automation program, a group of Ph.D. Scientists and Engineers,began the development of a new Artificial Neural Network technology, called P/NA3. Today, Neural Networks produced using the P/NA3 technology are sufficiently accurate and fast to be able to run in real-time during welding or, embedded into specialized search algorithms, to optimize the welding process for planning of robotic welds. This technology has been used in the automobile industry; however, since robotic welding is relatively new to shipbuilding, that is a new area of application. As flexible automation application increases, the need for fast and accurate models of the welding process will grow significantly. This paper will describe the methods used to develop P/NA3 Neural Networks, and discuss the application of P/NA3 Neural Networks for welding optimization for planning and control. In addition, a P/NA3 Neural Network real-time weld quality monitoring system, that can process over 200,000 data points per second, will be presented. While this technology is being used for the mass production environment of the Automobile Industry, its introduction to shipbuilding, and application to few-of-a-kind manufacturing is completely new. In addition to being used for robotic welding, this same technology can be applied, in virtual reality training environments, to train more highly skilled welders. A welder can learn the exact relationship between their actions with the welding torch and the resulting weld shape and size – in a virtual environment. This adds a completely new dimension to welder training. While this
is possible by cutting a cross-section of a real weld – that is costly and time consuming, just to get one single cross-section. As shown in figure 5, in a virtual training environment, the P/NA3 Neural Network welding process model can allow the student to instantly “see” their weld in cross-section all along the weld. The paper will discuss the development, using Design of Experiments methods, of a P/NA3 Neural Network welding process model. This will include an advanced technology software that employs Artificial Intelligence technology to generate a mathematical representation of the weld shape. The Neural Network is developed using the input parameters of the welding process (e.g., voltage, wire feed rate, travel speed, torch angles, etc.) and the resulting weld shapes. The software that uses the neural network models for production pre-planning and process optimization will be examined. Real-time weld process control and real-time weld process quality assurance monitoring software will be discussed and examples provided, showing how this technology can be used in shipbuilding. This entails software that takes inputs of sensor data on multiple channels, uses the P/NA3 Neural Network welding process model to predict the weld quality. Then, in real-time the system can use the P/NA3 Neural Network process model to search for the parameter changes that will improve the weld quality – and dynamically adjusts the process “on-the-fly.” This real-time quality monitoring is done by the human brain in manual welding – the P/NA3 Neural Network welding process modeling technology can do that same task for robotic welding.
Figure 1.NAMeS Weld Measurement System, used to capture the intricate shape of a weld in a special mathematical structure that can be used to develop an Artificial Neural Network Model.
3
Figure 2.P/NA Neural Network Welding process model. The cross-section shape of the weld changes instantaneously as any of the process parameter slidebars is moved.
3
Figure 3.Engineering planning and optimization software with embedded P/NA Neural Network Welding process model. The response surface allows the engineer to quickly identify the process parameter values for a flexible automation welding robot which will produce the exact weld crosssection shape that is necessary to meet the mechanical property requirements. The search software can calculate the shape of over 10 million sets of weld parameter values in less than 1 second and can provide the engineer with the precise control parameters to produce a weld that is the exact shape and size required. Note: after applying the search algorithm system, the weld shape meets all of the requirements (see all green values in the “output window” of figure 2. The system calculates over 100 weld shape engineering outputs for each of the 10 million sets of weld parameters values, then selects the shape which is closest to the shape requested by the design engineer. Then the software can download these parameters into the robot controller 3 automatically. The P/NA Weld process model, allows an engineer to create anoptimized arc welding robot control program, just as a manufacturing engineer uses CAD data to automatically produce a control program for a machining system, such as an automated lathe or mill or laser.
3
Figure 4.The weld process monitoring / control system with embedded P/NA Neural Network model. The system can produce a 3D image of the weld. In the screen capture on the left, the plates being welded can be turned transparent, so the inside of the weld can be examined, similar to a radiograph of the weld. The system can monitor up to four robotic welding systems simultaneously, collecting nearly 1 million data points per second and converting that data to the precise weld shape, in real time, and compare that weld shape to the Mil. Spec. or Code. The 3D image is color coded: good (green), marginal (yellow), and reject (red) areas of the weld produced by the robot. This data can be instantly transmitted to a remote computer in an engineers’ office.
3
Figure 5. The Virtual welder training system with embedded P/NA Neural Network Welding process model. The student can move the torch to any location along the weld and see instantly the cross section of the weld at that location. Since the system stores all of the welders movements and torch angles, the student can “see” instantly the relationship between their welding technique and the shape of weld that they are producing, at any point along the weld.
