Self-Learning Fuzzy Neural Networks and Computer Vision for Control of Pulsed GTAW
Neural network modeling and computer vision techniques are used to control the dynamics of the pulsed gas tungsten arc welding process
BY S. B. CHEN, L. WU, Q. L. WANG AND Y. C. LIU
ABSTRACT. The objective of this re- weld bead formation, and control of the of Mamdanl, et al. (Refs. 14-17), the search is to apply intelligent control size and changes in the weld pool, is still need for research to develop a fuzzy methodology to improve weld quality. a perplexing problem, whether for the logic controller that can learn from ex- Based on fuzzy logic and artificial neural control engineer or the welding technol- periences has been realized. The learn- network theory, a self-learning fuzzy and ogist (Refs. 1-7). In recent years, fuzzy ing task may include the identification of neural network control scheme has been logic controllers, which do not require the main control parameters or the de- developed for real-time control of pulsed analytical or accurate models of the con- velopment and tuning of the fuzzy mem- gas tungsten arc welding (GTAW). Using trolled process, have demonstrated a berships used in the control rules. an industrial TV camera as the sensor, the number of successful applications. Per- On the other hand, there has been an weld face width of the weld pool, i.e., the haps the most important application field increasing interest in studying artificial feedback signal in the closed loop sys- where fuzzy logic plays a significant role neural networks (ANN). The ANN is a tem, is obtained by computer image pro- is the control of complex industrial representation that attempts to mimic the cessing techniques. The computer vision processes. In general, these processes functionality of the brain. The present re- providing process status information in may be conventionally adjusted by a search on ANN shows that the method- real-time is an integral part of a self-learn- human operator due to their complexity. ology has been used on a more modest ing fuzzy neural control system. Such a Several industrial applications of fuzzy scale to develop nonlinear models and system enables adaptive altering of weld- logic control have been reported (Refs. controls. For example, recent studies ing parameters to compensate for chang- 8-13). These applications have mainly (Refs. 18-20) demonstrated the utility ing environments. The experiments on concentrated on simulating the perfor- and flexibility of the concept within the the control of the pulsed GTAW process mance of a skilled human operator in domain of process engineering. It has show that the scheme presented in this terms of linguistic rules. However, the been proven that any continuous func- paper can be used to control compli- process of learning and tuning linguistic tion can be approximated arbitrarily well cated variables such as encountered in rules to achieve the desired performance on a compact set by a feedforward artifi- welding processes. remains a difficult task. Starting with the cial neural network (Ref. 20). Generally, self-organizing control (SOC) techniques the ANN model and control system have Introduction the following characteristics: a strong ro- bustness, fault tolerance, universality, Owing to the uncertainties of phe- parallel distributed processing, and nomena such as metallurgy, heat transfer, learning and adaptive abilities. More- chemical reaction, arc physics and mag- KEY WORDS over, it is noted that Ref. 21 presented netization, the arc welding process is in- neural networks and fuzzy theory from a herently variable, nonlinear, time-vary- Self-Learning unified engineering perspective. The ing and strong coupling in its Fuzzy Logic combination of fuzzy logic and artificial input/output relationships. As a result, it Neural Networks neural networks oviously has great po- Intelligent Methodology is very difficult to obtain a practical and tential for various fields in industrial en- Real-Time Control controllable model of the arc welding gineering. Computer Vision process by classical modeling ap- The experiences of a skilled welder Topside Scanning proaches. Until now, control of weld could be summarized as a set of fuzzy Image Processing quality, such as weld joint penetration, logic control rules that are described in Weld Bead Width terms of IF (caused or conditions) THEN Pulsed GTAW S. B. CHEN, L. WU, Q. L. WANG and Y. C. (actions) rules. So we can develop a LIU are with Welding Division, Harbin Insti- fuzzy logic controller to imitate a skilled tute of Technology, Harbin, China. welder's operation of the welding
WELDING RESEARCH SUPPLEMENT I 201-S ...... Fu~ ~o.t~ollcr ~ !fC ...... "'-. ""'.... i er I~K U~ U
• 8(0 y(t)
Fig. 1 -- Principle diagram for the self-learning fuzzy neural control of Fig. 2 -- Fuzzy controller architecture. the arc welding process.
tem for detecting the weld face verting the crisp variables e, ec and erto width of pulsed GTAW in real- fuzzy variables E, EC, ER by the corre- time. Fuzzy and neural net- sponding quantizing factors Ke, Kcand Kr, work algorithms were used in respectively; the Fuzzy Calculator is used .... designing a controller with to complete fuzzy inference calculating; self-learning or adaptive regu- and the "Defuzzier" means converting O.8 lation of welding parameters, the fuzzy variable U to the crisp variable such as welding current and u by a quantizing factor K u. In our travel speed, to obtain the de- scheme, the Fuzzy Calculator in Fig. 2 is sired weld bead formation. described in terms of an analytical for- 0.4 The organization of this mula instead of common fuzzy rule ta- paper is as follows: the analyt- bles, that is 0.2 ical formulation of fuzzy logic control is introduced; a self- U(t) = -[a(t)b(t)E(t) + (1-a(t))b(t)EC(t) learning fuzzy neural control 0 , , , I , , , I , i , i I i , . • ' .... J + (1-b(t))ER(t)] (2.1) 0 5 10 15 20 25 scheme for the arc welding ~rne (~) process is presented and the where U, E, EC and ER denote the fuzzy corresponding algorithms of variables corresponding to their crisp the fuzzy controller and neural Fig. 3 -- Simulation of three-term and two-term con- variables; control action, u(t); control network model are developed; trollers. error, e(t) = D-y(t); change in error, ec(t) and experimental systems and -- e(t)-e(t-l); acceleration error, er(t) = the results of real-time control ec(t)-ec(t- l ); and a(t) and b(t) ~ [0,1] de- process. This paper will develop an ana- of the weld face width in pulsed GTAW note the tuning parameters between E, lytical formula for fuzzy control instead are shown. of using common fuzzy linguistic rules, EC and ER. The fuzzy variables and their corresponding crisp variables differ in and it will be shown how to combine the Self-Learning Fuzzy Neural ANN for modeling of the dynamic weld- the transformation factors relating to their Control Scheme for the ing process and fuzzy control to produce universes of discourse. Contrary to the a self-learning neural scheme for real- Arc Welding Process common approach, all universes of dis- time control of the pulsed gas tungsten course are considered continuously in A schematic of a self-learning fuzzy arc welding (GTAW) process. this study, i.e., the E, ECand ER are con- neural control system for the arc welding Another critical difficulty in the dy- tinuous value variables. process is shown in Fig. 1. Where the namic control of the arc welding process The analytical formula (2.1) can be fuzzy controller FC maps the regulated is how to detect reliable information explained as follows (Refs. 22-24): We about changes in the weld pool, such as error e(t) into control action u(t), the arc first take a two-term fuzzy controller as welding process is denoted by WP, the weld bead width and joint penetration, an example. Generally, fuzzy rules relate output signal y(t) of the process is de- especially viewed from the top of the three universes of discourse, such as weld pool and in real-time. Although var- tected by measurement parts MS, PMN error E, change in error EC and control denotes a neural network model of WP ious efforts have been made to sense the action U. We can describe these uni- and MS, D denotes the desired value. The weld pool size in real-time from the top verses as a fuzzy set that contains lin- outputs of WP and MS both are denoted with ultrasonic detection, infrared sens- guistic variables {PB, PM, PS, O, NS, NM, by y(t) in the sense of neglecting the dif- ing, pool image processing, and radi- NB}, where P, O and N are meant for pos- ference in their transformation coeffi- ographic sensing (Refs. 1-7), results for itive, zero and negative, and B, M and S cients. E(t) is the error between output of controlling weld quality have been met for big, medium and small. If we define PMN and MS. The structure and algo- with limited success. To achieve perfect these linguistic variables as follows: rithms of each component are explained control of weld quality, we should not below. only improve reliable measurement PB= 3, PM= 2, PS= 1,0=0, means, but also introduce advanced con- NS = -1, NM = -2, NB = -3 trol methodologies in real-time. Both The Fuzzy Controller should be given equal attention and de- then a simple fuzzy controller can be de- veloped simultaneously. In this study, a The fuzzy controller FC in Fig. 1 was computer vision system, i.e., a system constructed as shown in Fig. 2. In Fig. 2, scribed as shown in Table 1, and we can combining video camera and image pro- the "Differencor" is use to calculate Ae = summarize the fuzzy rules in the Table 1 cessing, was used in the closed loop sys- ecand g2e= er; the "Fuzzier" means con- as an analysis formula
202-s I MAY 1997 u <(E + EC)/2>
where
X,vl o U=
G(s) = 20 exp(-O.5s)/l(2s + 1) ifying the tuning parameters a(t) and b(t) where m and n denote orders of the arc (4s + 1)(2.2s + I)1 according to the neural network model of welding dynamic process, which could an uncertain system. be roughly estimated by experiments on The simulation results are shown in Fig. the system. The network structure of 3 where the symbol • indicates the unit- The Neural Network Model PMN PMN is shown in Fig. 4 step response of the controlled object The network functions of the PMN are G(s) regulated by three-term fuzzy con- The model PMN for the arc welding described as follows: troller (Equation 2.1 ) with the tuning fac- process WP and the measurement part tors a = 0.625 and b = 0.5625; and the MS, which contains computer vision sys- fli=l/fl] F (N1 ~-I/ symbol * indicates a two-term controller tems, can be realized by back-propaga- * exPL-/i__~lWlijXi* qlJ ]]l ' (Equation 2.2) with the tuning factor cc = tion networks with four layers and nodes 0.625. The other parameters in both con- N1-N2-N3-N4. The mapping relation- j=l ..... N 2 trollers were identical, that is Ke = 30, Kc ship of the model is described as (2.4) = 10, K r = 10 and Ku = 50. Obviously, the simulation results show that the three- YM (t+l) = fM (u(t), u(t- 1),...,u(t- m); term fuzzy controller is more advanced yM(t),...,yM (t-n)) (2.3) for high-order systems than the common two-term one. More details were illus- where defining k=l ..... N3 trated in Ref. 23. x T = Ix 1..... XNt]T = [u(t),...,u(t - m); In Fig. 1, the fuzzy controller (Equa- (2.5) tion 2.1 ) is adaptively regulated by mod- yM(t) ..... yM(t - n)]T
Table 1 -- Simple Fuzzy Control Table 2--Fuzzy Control Table 3--Fuzzy Control
E -3 -2 -1 0 +t +2 +3 E -3 -2 -1 0 +I +2 +3 E -3 -2 -1 0 +1 +2 +3 EC EC EC -3 -3 -3 -2 -2 -1 -1 0 -3 -3 -2 -1 -1 0 -1 +2 -3 -3 -3 -2 -2 -2 -2 -1 -2 -3 -2 -2 -1 -1 0 +1 -2 -3 -2 -1 0 0 +1 +2 -2 -2 -2 -2 -1 -1 -1 -1 -1 -2 -2 -1 -1 0 +I +1 -1 -3 -2 -1 0 +1 +1 +2 -1 -2 -1 -1 -1 0 0 0 0 -2 -1 -1 0 +1 +1 +2 0 -2 -2 -1 0 +1 +2 +2 0 -1 -1 0 0 0 +1 +1 +1 -1 -I 0 +1 +1 +2 +2 +1 -2 -1 -1 0 +1 +2 +3 +1 0 0 0 +1 +1 +1 +2 +2 -1 0 +1 +1 +2 +2 +3 +2 -2 -1 0 0 +1 +2 +3 +2 +1 +1 +1 +1 +2 +2 +2 +3 0 +1 +1 +2 +2 +3 +3 +3 -2 -1 0 +1 +1 +2 +3 +3 +1 +2 +2 +2 +2 +2 +3
WELDING RESEARCH SUPPLEMENT I 203-s Fig. 6 -- Experimental system. Fig. 7 -- Vision sensor system.
<3> Learning algorithms of the model then weights modified as [ C (N3 PMN. Off-line and on-line learning algo- A W3ko(t) = h3 v 3 (t)f2k(t) + g3 A W3ko(t-1 ) rithms can be used to modify parameters W3ko(t+l ) = W3ko(t) + A W3ko(t) (3.4) of the PMN network. We can obtain the initial weights of the PMN by batch sam- q3o(t+l ) = qBo(t) + h~ v3(t) (3.5) (2.6) ple data [u(t),y(t)J. The off-line learning where Wli I and x i denote connection results of the PMN model can be used as weights and output functions for PMN a reference model for the uncertain con- A W2j k (t) = h2 V2k (t)flj (t) + g2 A W2jk(t-1 ) first-layer network, W2jk; and fli for the trolled objects. Using the on-line learn- W2jk(t+| ) = W21k(t) + A W2jk(t) (3.6) second-layer, and W3k r and f2k for the ing, we can modify network weights of q2k(t+ l) = q2k(t) + h 2 V2k (t) (3.7) third-layer, respectively. the PMN in real-time by the index (3.2) and the principle of error gradient de- A Wlij(t) = h I vq (t)x i + gl A Wti j (t-l) scent, that is Self-Learning Algorithms of Fuzzy Wzi i It+ 1) = W~i i (t) + A Wti i (t) (3.8) Neural Controller A W(t) ~ ~ /avv(t) qlj (t+ l ) = qlj (t)+ hlv l/t) (3.9) Self-learning algorithms for the fuzzy W(t+ 1) = w(t) + A w(t) (3.3) controller FC and neural network model where hi,sic (0,1) (i = 1,2, 3) are learn- PMN are developed as follows: The learning algorithms for the PMN are briefly described below. ing factors and momentum factors, re- spectively. The Equations 3.3-3.9 are Defining <1 > The index of control error one-step learning algorithms for PMN networks in a controlling period. N 2 v3(t) = (y(t) -yM(t))(1 -yM(t)) yA4(t) Je = ~,[D-y(t+l)] /2 <4> Adaptive-modifying of tuning para- t=l (3.1) meters a(t), b(t) of the FC. V 2k(t) = f2k(t)(1 - f2k(t)) Wjko(t) V 3 (t), <2> The index of model error Assuming the PMN network parame- k=l ..... N 3 ters are known variables that have been N 2 Je= ~,e (t+1)/2= obtained by off-line last one-step learn- t=1 N3 ing, we have the following algorithms to Vlj = fli(t)(1- flj(t))k~ 7 learning factors ha,h b E (0,1) ~Je~a(t) = lD-(y M It+l) + e)J 1~ YM (t+l)/aa(t)J = [D-y(t+l)] ~_controlveriod ~ 1~ YM (t+ l)/aa(t)], (ae/Sa is neglected) aY M (t+l)/aa(t) = l~ fM~u(t)] l~u(t)/~a(t)] au(t)/aa(t) = -b(t) lE(t) -EC(t)I Fig. 8 -- Welding current pulse series. 204-s I MAY 1997 that is formed to a digital image stored in the frame buffer of the image interface unit as a(t+ 1) = a(t) + h a (D -y(t+ 1))b(t)lE(t) a part of the computer memory by means -EC(t)l fOfm/OU (t)l (3.10) of the segment mapping, which could be written/read randomly. The welding cur- where rent was regulated by the interface unit of the welding power source, and regu- arm / au(t) = arm / ax, = lation of welding travel speed was real- ized by a single chip computer system. Image Processingof the Face of the Weld Pool - f~(,- f~)~/N~(W~,,f~, k=l ~. As is well known, the arc of the weld- ing process makes illumination and irra- Fig. ~) hna~e ot weld pool z~)ne. diation of surfaces of the weld specimen (3.11) and the weld pool complicated. Usually, Similarly the arc light is considered a strong inter- ference while sensing surface changes on b(t+ 1) = b(t) + h b (D-y(t+ 1)) the weld pool. After investigating many la(t)E(t) + (1-a(t))EC(t)-ER(t)] experiments, the intensity of the arc light fOfM/~u(t)l (3.12) was found to change from strong to weak when the welding current transformed Equations 3.10-3.12 are one-step self- from the pulse peak value to the base learning algorithms for the tuning para- value in pulsed GTAW. We chose the meters a(t), b(t)of the FC in a controlling pulse ripple and time sequence of the period. Essentially, it is to regulate the welding current shown in Fig. 8 for the fuzzy control rules as a human operator purpose of regulating and image pro- in real-time. cessing of the welding process. At the Fig. 1() lhe e\trac ted bead width. very moment when the pulse peak value Experimental Systems of the welding current was decreasing to for Pulsed GTAW To investigate the feasibility of the above fuzzy neural control scheme for 23O practical, uncertain objects, an experi- ment on control of weld bead width with the pulsed GTAW process was con- ducted. In this study, an attempt is made 180 to combine the fuzzy logic and neural > network techniques for the process dy- ¢ namics. 130 Components of Control Systems S The block diagram for the control sys- tems of the pulsed GTAW process is 80 shown in Fig. 5. The systems are com- 100 150 200 250 300 350 4E posed of an IBM-PC/AT 386 computer for ColumnCoixel) self-learning control and image process- ing algorithms, an industrial TV camera Fig. I I -- A line gray distribution of image. used as a vision sensor to pick up the top image of the weld pool, an image inter- face unit, an image monitor and an alter/direct pulse arc welding power i .o 1 source. The photographs of the experi- ment system and sensor equipment are shown in Figs. 6 and 7, respectively. The TV camera was mounted on the back of the welding torch, at an angle of 30 deg w with the horizon. The working platform I I with the workpiece holder could be reg- ulated in a horizontal plane (500 x 200 250 mm), i.e., which is with two-degrees of freedom. The image of the weld pool was picked up by the TV camera and trans- Fig. 12 -- Specimen size (mm). WELDING RESEARCH SUPPLEMENT I 205-s 80 g lOO A 70 _o. eo 12o o 100 5o | 40 80 30 . , . I .... I .... L .... i .... i .... I , , , - I .... i ~ . , , i .... i .... , .... , .... i .... ~ .... b .... , 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 81 Time 0leo) Time(sec) Fig. 13 -- Test for the process of welding current/bead width. A -- Regulating current; B -- weld face width. i(t) = Ivg = 60 A as in Fig. 8, the illumi- other parameters. For simplicity, we es- in the control scheme for pulsed GTAW are shown in Fig.14. nated image of the welding pool was tablished a SISO model of the welding The simulation of the process shows clearest. At this moment, the weld pool pool dynamics for the control of pulsed that the control system is able to adap- began to cool down and solidify incom- GTAW. The input and output of the tively attain the optimized control and pletely due to the heat inertia of the model were the welding current and the desired welding specification by self- molten process. An image of the welding bead width of the top of the welding learning algorithms, i.e., welding cur- pool zone is shown in Fig. 9. The image pool, respectively. An experiment for rentpeak value is approximately 180 A, consists of four parts from top to bottom: testing the dynamics of the welding cur- weld bead width y(t) to 6 mm, and tun- welding torch, tungsten electrode, weld rent/bead width process was completed ing factors aft) to 0.63 and b(t) to 0.67. pool and welded joint. A static digital under the following conditions: bead-on- image with various noises entered into plate welds on low-carbon steel plate of 2-mm thickness, 100 x 250 mm; tungsten Experimental Results on the Control the image interface. Using slide-median of Pulsed GTAW filtering algorithm of one-dimension, a electrode diameter of 3 mm; shielding line pixel gray distribution was obtained gas of argon at 8 mL/min flow rate; con- Based on the system scheme in Fig. 5, in a frame filtered image shown in Fig. stant travel speed of 9.5 cm/min; and arc experiments to control weld bead width 11. Searching the distributed zones of the voltage of 12 to 30 V DC. The parame- with pulsed GTAW have been performed pixel gray in the weld pool image, line by ters of welding current pulsed series in on the system shown in Fig. 6. The bead- line, along the tungsten axis, the maxi- Fig. 8 were designed as follows: on-plate experiments were conducted mum width of the weld pool was ob- Tvg = 200 ms, Iv~ = 30 A, Tp = under the following conditions: The tained, which is indicated by a real-line 350 ms, T I = 40 ms, Tb = specimens for the test were low-carbon crossing the weld pool in Fig.10. To ob- 650 ms, I b = 20 A, Tg = 60 ms steel plate of 2 mm thickness shown in tain a reliable weld bead width, the dif- The pulse peak value of the welding cur- Fig.12; a dumbbell specimen was used ference was used in the algorithm of one- rent Ip is the regulating variable. for imitating sudden changes in heat transfer or heat sink conditions during order five-points to calculate the sudden The test result for the process is shown change boundaries of the pixel gray dis- the welding process; the tungsten elec- in Fig. 13. Based on the above experi- trode diameter was 3 mm with an acute tributions in the weld pool image and ob- mental data and using the batch testing tained the maximum weld bead width in angle of 45 deg and the tip diameter not data of input- output pairs and an off-line larger than 0.5 mm; the shielding gas of real-time. The whole image processing learning algorithm, a neural network period was approximate 0.7 s. argon was at 8 mL/min flow rate; con- model of the process as shown in Fig. 4 stant travel speed of 9.5 cm/min; and an was obtained. Its nodes N1, N2, N3 and arc voltage of 12 to 30 V DC. The weld- Experiment Results N4 are 5, 10, 10, and 1, respectively, ing current pulse series shown in Fig. 8 which realizes the following mapping: was chosen for the experiment. The pa- Modeling and Simulation of the Welding Process rameters of the pulsed current were de- YM (t+l) = signed as: By analyzing the technology of the fM (u(t),u(t-l),u(t-2), YM (t), YM (t-l)) (4.1) Tvg = 200 ms, Ivg = 30 A, Tp = pulsed GTAW process and the test data 350 ms, T 1 = 40 ms, T b = from standard conditions, it is known that YM (t+l) adds a pseudo-random se- 650 ms, I b = 20 A, Tg = 60 ms and the pulse peak value of the welding the main factors influencing weld pool quence in the simulation model of the current 1. is the regulating variable to changes are welding current and welding pulsed GTAW process WP in Fig. 1. control t~e weld bead width, D = 5.5 travel speed, with other parameters such Using the self-learning algorithms as pre- mm. The quantizing factors of the FC as plate thickness and root opening being viously mentioned, and the quantizing were chosen as Ke = 4, Kc = Kr = 2 and normal. Usually, a robotic welding sys- factors of the FC chosen as Ke = 4, K c = tem only regulates the welding current Ku = 200. The control period was deter- Kr= 2 and Ku =200, the simulation results under a fixed welding travel speed and mined as one second. The experiment re- 206-s I MAY 1997 80 A 140 7O 130 t® gE 120 O .~ 5o ~ 110 40 100 30 I , , , , i , , , , i , , , , i .... i 90 20 40 60 80 101 20 40 80 80 100 Time (sec) Trne (sec) 0.04 C 0.1 D 0.035 0.05 0.03 0.025 0 E g g 0.02 -0.05 0.015 -0.1 0.01 W 0.005 -0.15 o -0.2 . , . i . , , , i , , , , i , , , , i .... ; -0.035 .... ' .... ' .... ' .... f .... J 0 20 40 60 80 10 0 20 40 60 80 103 Time (sec) Time (see) 0.7 0.7 0.65 0.6 .E c S 0.55 0.5 .... i .... , , . . , .... , • • , , I 0.5 , , , , i , , , , i , , , , i , , , , i , , , , i 20 40 60 80 103 0 20 40 60 80 100 nine (see) Time (sec) Fig. 14 -- Simulation of the control scheme for pulsed GTAW. A -- Welding current; B -- weld face width; C -- control error; D -- error change; E -- tuning factor a(t); F -- tuning factor b(t). suits on the control system are shown in process by the self-learning fuzzy neural process, the tuning parameters a(t) and Figs.15 and 16. control scheme presented in this paper, b(t) of the fuzzy controller FC are adap- The goal was to control both the top Fig. 16A shows that the results of con- tively modified as seen in Fig. 16C and and bottom of the weld bead width, even trolling weld bead width are satisfactory. D. Figures 15B, 16E and 16 F are pho- under severe interference such as differ- The control error is maintained within tographs of the weld bead. ences in heat transfer and heat sink con- 0.3 mm and the regulation of welding In this study, we have only investi- ditions. Mainly, in this experiment, the current (peak value) in Fig. 16B is identi- gated the controllability of welding cur- effects of controlling the bead face width cal with experiments with a human op- rent and weld face width, and only indi- of the weld pool were investigated, erator, i.e., the fuzzy controller is adap- rectly obtained a desired root surface which indirectly correlates the root sur- tively regulating welding current (peak width. If the requirement of the underside face width on thin specimens. With the value) in the narrower section of the formation is critical, the design scheme welding peak current constant at 180 A, specimen due to the adverse changes of shown in Fig. 17 is for the control of the Fig. 15 shows that the weld bead width scattered heat conditions, which is just root surface width. BWNM indicates the changes drastically from the desired similar to the behavior of a skilled back weld width network model of value. With the manipulation of the welder. With real-time learning of the pulsed GTAW combined with the infor- WELDING RESEARCH SUPPLEMENT I 207-s A 150 140 13o 120 l 110 1oo 0 10 ~n 30 40 50 60 70 8O Time (sec) Fig. 15 -- Constant normal welding. A -- Bead face width; B -- top of the specimen. 80 lqO B A 130 7O 120 60 ~ 110 100 a 50 40 80 30 7O 60 ,,,i .... i .... , .... i .... i .... , .... , .... i 20 0 10 20 30 40 50 60 70 80 0 Time (sec) Time (sec) o.t i¢: D 0.62 C 0.71 0.615 0.708 0.61 0.706 ~ 0.704 ,,~ 0.605 0.7 0.595 0.~ .... , .... i, ,.. , .... I .... , .... i, , , • .., , , , , ..... i , , , , .... b , , , ~ .... , . . • , .... J 0.~ 0.59 10 20 30 40 50 60 70 8( 10 20 30 40 50 60 70 80 Time (sec) Time (NO) Fig. 16 -- Welding results. A -- Bead face width; B -- welding current; C -- tuning factor a(t); D -- tuning factor b(t); E -- top of the specimen; F -- underside of the specimen. 208-s I MAY 1997 mation of weld pool half-length h(t). Ik If the welding travel speed is chosen D e(t) -~ u(t) ~ y(t) as the regulating variable for control of desired weld bead width, the model of """% the process will appear as a time-delay due to the mechanical inertia of the reg- ulating speed mechanism. Such a system with uncertainties and time delays is dis- cussed in Ref. 25. e (t) y(t) Discussion 4 The experimental results to control l b (t) welds with changing heat transfer cir- cumstances demonstrated that the self- learning fuzzy neural control scheme Fig. 17 -- Control scheme for the root surface width. shown in Fig. 1 is feasible for controlling weld pool dynamics for pulsed GTAW. Comparing Fig. 15 with Fig. 16, the re- sults show that the regulating actions of techniques will be improved corre- 13. Mandic, N. J., Schart, F. M., and Mam- the control system are similar to the in- spondingly. dani, E. H. 1985. Practical application of a re- alistic fuzzy rule-based controller to the dy- telligent behavior of a skilled welder. Acknowledgment namic control of a robot arm. lEE Proceedings, The results are mainly influenced by Pt. D. 132(4). the time needed for completing the con- 14. Procyk, T. J., and Mamdani, E.H. 1979. trol algorithm and image processing, This work was supported by the Na- A linguistic self-organizing process controller, which could be improved by parallel tional Natural Science Foundation of Automatica, 15:15-30. processing of the neural networks and by China, No. 59575057 and No. 15. Berenji, H. R., and Khedkar, P. 1992. 59635160. Learning and tuning fuzzy logic controllers enhancing computational speed. through reinforcements. IEEE Trans. Neural References Networks, 3(5): 724-740. Conclusions and Further Work 16. Jang, J. R. 1992. Self-learning fuzzy 1. Bentley, A. E., and Marburger, S. J. 1992. controllers based on temporal back propaga- Arc welding penetration control using quanti- tion. Ibid., 3(5): 714-723. The following conclusions resulted tative feedback theory. Welding Journal 17. Lee, C. C. 1991. A self-learning rule- from the above investigation on the ap- 71(11): 397-s to 405-s. based controller employing approximate rea- plication of fuzzy logic and artificial 2. Carlson, N. M., Johnson, J. A., and soning and neural net concepts. Int. J. Intelli- gent Systems, 5(3): 71-83. nueral networks to the control of the Kunerth, D. C. 1990. Control of GMAW: de- tection of discontinuities in the weld pool. 18. Nguyen, D. H., and Widrow, B. 1990. welding process: Ibid., 69(7):256-s to 263-s. Neural networks for self-learning control sys- 1 ) The self-learning fuzzy neural con- 3. Chen, W., and Chin, B. A. 1990. Moni- tems. IEEE Control Systems Magazine, 10: trol scheme presented in this paper can toring joint penetration using infrared sensing 18-23. be used for real-time control of the bead techniques. Ibid., 69(9):181-s to 185-s. 19. Hunt, K. J., et al. 1992. Neural net- width with the pulsed GTAW. 4. Nagarajan, S., Chert, W. H., and Chen, works for control systems -- a survey. Auto- 2) The self-learning fuzzy neural con- B. A. 1989. Infrared sensing for adaptive arc matica, 28(6) 1083-1112. trol in the experimental system has ex- welding. Ibid., 68(11 ): 462-s to 466-s. 20. Willis, M. J., et al. 1992. Artificial neural networks in process estimation and hibited satisfactory adaptive behavior to 5. Rokhlin, S. I., and Guu, A. C. 1990. Computerized radiographic sensing and con- control. Automatica, 28:1181-1187. control a variable process such as arc trol of an arc welding process. Ibid., 69(3):83- 21. Kosko, B. 1992. Neural Networks and welding. Artificial neural networks can s to 97-s. Fuzzy Systems. Prentice Hall, Englewood also be used to model the dynamics of 6. Elmer, J. W., Giedt, W. H., and Eagar, T. Cliffs, N.J. the welding process. W. 1990. The transition from shallow to deep 22. Long, S. Z., and Wang, P. Z. 1982. Self- 3) The arc can be utilized to illumi- penetration during electron beam welding. regulation problem of fuzzy control rules. Ibid., 69(5):167-s to 176-s. Fuzzy Mathematics, 3:105-111. nate the weld pool and the weld face- 23. Li, S. Y., and Hu, H. Z. 1987. Design during the tranformation of the pulse 7. Wang, Q. L., Yang, C. L., and Gang, Z. 1993. Separately excited resonance phenom- and study of a class three-dimensional fuzzy peak value of the welding current to the enon of the weld pool and its application. controller. Proc. of Int. Symp. on Fuzzy Sys- base value. Ibid., 72(9):455-s to 462-s. tems and Knowledge Engineering. 2: 4) Computer vision techniques can be 8. Batur, C., and Kasparian, V. 1991. Adap- 447-448. used to provide dynamic information on tive expert control. Int. J. Control, 64(4): 24. Abdelnour, G. M., et al. 1991. Design the weld pool in feedback control of the 367-381. of a fuzzy controller using input and output 9. Bernard, J. A. 1988. Use of rule-based mapping factors. IEEE Systems, lan, and Cy- arc welding process. bernetics, 21 : 952-960. Further investigations will be made to system for process control. IEEE Control Sys- tems Magazine, 8:3-13. 25. Chen, S. B., Wu, L., and Wang, Q. I. perfect the fuzzy neural network to con- 10. King, P. J., and Mamdani, E. H. 1977. 1997. Self-learning fuzzy neural control of un- trol weld quality such as fine formation The application of fuzzy control system to in- certain system with time-delays. IEEESystems, of the weld bead and full joint penetra- dustrial process. Automatica, 13: 235-242. Man, and Cybernetics, 27: tion with pulsed GTAW. A multivariable 11. Kickert, W. J. M., and Mamdani, E. H. feedback control system will be designed 1978. Analysis of a fuzzy logic controller. for the arc welding process. Action en- Fuzzy Sets and Systems, 1:29-44. 12. Kickert. W. J. M., and Van LautoLemke, tries of the control system will contain H. R. 1976. Application of a fuzzy controller welding current, are voltage and travel in warm water plant. Automatica, 12: speed. Image sensor and processing 301-308. WELDING RESEARCH SUPPLEMENT I 209-s