Intellige telligent nt We Weld lding ing and d We Weld lder - PowerPoint PPT Presentation

Mo Modelling elling of of Hum uman an We Welder er for or Intellige telligent nt We Weld lding ing and d We Weld lder er Traini aining* ng* YuMing Zhang University of Kentucky Lee Kvidahl Ingalls Shipbuilding NSRP Welding Panel Meeting Bethesda, Maryland May 4-5, 2016 For Unlimited Distribution. Research funded by the NSF under grant “NRI -Small: Virtualized Welding: A New Paradigm for Intelligent Welding Robots in Unstructured Environment,” IIS-1208420, Sept. 2012-August 31, 2016 and grant “Machine -Human Cooperative Control of Welding Process” CMMI-0927707, October 2009-Spet. 2013

Contents 1. Sensing of 3D Arc Weld Pool Surface: motivation, method, real time 2. Characterization: from numerous points to three characteristic parameters 3. Control of 3D Weld Pool Surface: control theory method 4. Human Welder Response: modeling and analysis, control using welder model 5. Welder Motion Response: human-robot system, speed adjustment, 3D adjustment 6. Future Directions

1. . Sen ensing sing of of 3D Arc We Weld d Poo ool Sur urface ace (Hu Human man Res esponse ponse Inpu put) t)

Weld pool: where complex phenomena originate; but only the surface is visible; the major feedback information available to human welders Measurement of weld pool surface temperature distribution: needs the emissivity to determine the temperature the infrared radiation but the emissivity is slope dependent Weld Penetration:

(1) Surface Specular: use laser reflection; (2) Arc Radiation: use laser reflection and intercept at a distance  low power continuous laser for continuous measurement, no need for a special camera Laser: 20 mW, 685 nm Y.M. Zhang, H.S. Song, and G. Saeed. Observation of a dynamic specular weld pool surface. Measurement Science & Technology, 17(6), 2006.

 Reflection law, surface constraint, error evaluation Reflected dots from image processing and reconstructed surface 180 Image processed dots Dots reflected by reconstructed surface 160 140 120 100 Y/mm 80 60 40 20 0 -40 -30 -20 -10 0 10 20 30 40 X/mm Hongsheng Song. Machine Vision Recognition of Three-Dimensional Specular Surface for Gas Tungsten Arc Weld Pool. ECE Department, University of Kentucky, 2007. XiaoJi Ma. Measurement of Dynamic Weld Pool Surface in Gas Metal Arc Welding Process. Department of Electrical and Computer Engineering, University of Kentucky, Feb. 2012.

Analy alytical tical Solu luti tion on Key for Real Time Measurement and Control W.J. Zhang, X.W. Wang, Y.M. Zhang, 2013. “Analytical Real -time Measurement of Three- dimensional Weld Pool Surface,” Measurement Science and Technology, 24(11), article Number 115011 (18pp), doi:10.1088/0957-0233/24/11/115011

2. . Ch Char arac acterization terization of of 3D We Weld ld Poo ool l Su Surface ace

 Characteristic parameters should be used rather than a large set of 3D coordinates. Should keep the fundamental information in the weld pool surface about the weld joint penetration. W.J. Zhang, Y.K. Liu, X. W. Wang, Y.M. Zhang. Characterization of three-dimensional weld pool surface in gas tungsten arc welding. Welding Journal, vol. 91, 2012.

Left: Measured 3D weld pool surface parameters from 36 experiments; Right: Least squares model fitting with 3-parameter model using the width, length, and convexity.     w 1.7906 W 0.5657 L 10.8057 C 0.9868 b

3. . Con ontr trol ol of of 3D We Weld d Poo ool Sur urface ace Modeling: how the characteristic parameters respond to the change in current and travel speed – extract the model from experimental data Control: Model predictive control algorithm Yukang Liu, YuMing Zhang. Control of 3D Weld Pool Surface. Control Engineering Practice, 21(11), 2013.

Welding Experiments Speed Disturbance 6 W b (mm) 4 2 0 0 20 40 60 80 100 120 Distance (mm) 7 8 Width (mm) Length (mm) 7 6 10*Convexity (mm) 6 5 Weld Pool Parameters Input Parameters 5 4 Current/10 (A) 4 Voltage/3 (V) 3 Speed (mm/s) 3 2 2 1 1 0 10 20 30 40 50 60 70 80 90 100 110 120 0 Time (s) 10 20 30 40 50 60 70 80 90 100 110 120 Time (s)

4. . Mo Modeling eling and d Analy nalysis sis of of Hum uman an We Welder er Res esponse ponse to to 3D We Weld Poo ool Sur urface face (mechanized welding, human adjusts the current)

 Skilled human welder holds the current regulator while observing the geometry of weld pool;  Adjusts the welding current to control the process to full penetration. Welding Parameters Argon flow Current/A Welding speed/mm/s Arc length/mm rate/L/min Experiment Parameters 57~81 1~2 3.5-4.5 11.8 Monitoring Parameters Manual control system of GTAW process Project Laser to weld pool Imaging plane to weld pool distance/mm angle/ ° distance/mm 35.5 24.7 101 Camera Parameters Shutter speed Frame rate/ fps Camera to imaging plane distance/mm /ms 4 30 57.8 Y.K. Liu, Y.M. Zhang, L. Kvidahl. Skilled Human Welder Intelligence Modeling and Control. Welding Journal, 93, 2014.

 In general, the human intelligent model can be written as:       I k ( )= ( f W k ( 3), L ( k 3), C ( k 3), I k ( 1)) f f f  Following linear model can be identified using standard least squares method:         I k ( )= 0.16 W ( k 3) 0.082 L ( k 3)+1.81 C ( k 3)+0.26 I k ( 1) f f f 4 Measured dCurrent Linear Estimated dCurrent 2 dCurrent 0 -2 -4 0 200 400 600 800 1000 1200 1400 1600 Sample Number Linear modeling result.

5 Measured dCurrent Measured dCurrent Measured dCurrent 4 3 Linear Linear Linear 4 ANFIS Estimated dCurrent ANFIS Estimated dCurrent ANFIS Estimated dCurrent 2 3 2 dCurrent dCurrent dCurrent 2 1 1 0 0 0 -1 -1 -2 -2 -2 1360 1380 1400 1420 1440 1460 500 520 540 560 580 600 50 100 150 Sample Number Sample Number Sample Number Model comparison between linear and ANFIS model. Model Comparison between Neuro-Fuzzy Model and linear model Average Model Maximum Model RMSE /A Error /A Error /A Linear Model 0.52 0.79 3.15 ANFIS Model 0.50 0.76 3.03 Y.K. Liu, W.J. Zhang, Y.M. Zhang. Dynamic neuro-fuzzy-based human intelligence modeling and control in GTAW. IEEE Transactions on Automation Science and Engineering, 12, 2015.

Model Comparison Linear Models         Novice welder I k ( )= 0.049 W k ( 3) 0.0049 ( L k 3)+1.73 ( C k 3)+0.72 I k ( 1)         Skilled welder I k ( )= 0.16 W ( k 3) 0.082 L ( k 3)+1.81 C ( k 3)+0.26 I k ( 1) f f f

Nonlinear Model Comparison Novice Welder Skilled Welder 0 3 -0.5 2 -1 2 2  In normal cases the skilled welder's  I (A) -1.5  I (A) 1 0 0 adjustments are minimal which can prevent -2 -2 -2 0 -2.5 large oscillation and overshoot novice welder 7 7 -1 -3 6 6 model suffers; -3.5 -2 5 5 6 6 5 -4 5 4 4 4 4 -3 L (mm) L (mm) 3 W (mm) 3 W (mm) 3 3 (a)  In other cases where the convexity is Novice Welder Skilled Welder 0.4 1 either considerably small or large, the 0.35 0.8 0.3 2 2 adjustment made by the skilled welder is 0.25  I (A) 0.6  I (A) 0 0 0.2 larger than that of the novice welder, which 0.15 -2 -2 0.4 0.1 can provide shorter settling time than novice 7 7 0.05 0.2 6 6 0 welder does. 5 0 5 -0.05 6 6 5 5 -0.1 4 4 4 4 L (mm) L (mm) -0.2 3 3 3 W (mm) 3 W (mm) (b)  The skilled welder model does provide Novice Welder Skilled Welder 3 better adjustment than the novice welder. 1 2.5 2 2 0.8 2  I (A)  I (A) 0 0 0.6 1.5 -2 -2 0.4 1 7 7 0.5 6 0.2 6 Nonlinear model surface of the neuro-fuzzy human welder model (left: 0 5 5 6 6 novice welder, right: skilled welder) for convexity = (a) 0.10 mm (b) 0 5 5 -0.5 4 4 0.18mm (c) 0.26mm. Previous response is zero for all cases. 4 4 L (mm) L (mm) (c) 3 3 3 W (mm) 3 W (mm)

Control Experiments: Varying Initial Current A B A B A B A B 8 8 Current/10 (A) Current/10 (A) Voltage/3 (V) 7 Voltage/3 (V) 7 Speed (mm/s) Speed (mm/s) 6 6 Input Parameters 5 Input Parameters 5 4 4 3 3 A B A B 2 2 1 1 0 0 0 10 20 30 40 50 60 70 80 90 100 110 10 20 30 40 50 60 70 80 Time (s) Time (s) (a) (b) Control experiment result with different initial current (a) 52A; (b) 54A.

5. We . Welder der Mo Moti tion on Res esponse ponse to to We Weld Poo ool Sur urfac face (weldi lding ng speed ed adjus ustment ment, , 3D opera ratio tion n adjustment) ustment) 1. Equipment for experimental data: human-robot system 2. Extract good response from not-perfect performance of human welder 3. Quality evaluation model 4. Supervised learning using good data Y.K . Liu, Y.M. Zhang, L. Kvidahl, 2014. “Skilled Human Welder Intelligence Modeling and Control: Part I- Modeling,” Welding Journal, 93: 46s -52s. Y.K . Liu, Y.M. Zhang, L. Kvidahl, 2014. “Skilled Human Welder Intelligence Modeling and Control: Part II- Analysis and Control Applications,” Welding Journal, 93(5): 162s -170s.