Neural network control for wig welding processes.

Bucur, Gabriela ; Popescu, Catalin ; Popescu, Cristina 等

1. INTRODUCTION

The final purpose of a welding operation is to achieve a welding seam that satisfies many imposed conditions. These conditions results from technological general system analysis, system witch is composed by power source--welding source arc welding system, welding seam, manipulation system and control system.

The paper present the way to process the welding-arc voltage with neural network. That result can be applied in automated robotic welding based on the corection of the trajectory welding torch through the welding seam.

2. CONTROL POSITION SYSTEM

The values of welding parameters regime (welding current, arc voltage and speed penetration) depend on relative position between the components of welding joint. Some parameters are controlled by welding equipment and the others by the geometrical ratio between welding head and piece. For example, on WIG welding process, the arc voltage depends on electrode-piece distance (Stefanescu & Cupcea 2002).

For welding head position control we use an acquisition system for arc voltage values from welding process. So, the arc voltage is a measure of welding head position. This signal will be the input for a neural network structure especially made by the authors for modifies the position of welding head. This neural network will be the controller for position closed loop, shown in figure 1, where [U.sup.arc.sub.process] is process arc voltage, [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] is reference calculate arc voltage, a is output of neural network, q is controlled position of welding torch.

[FIGURE 1 OMITTED]

[FIGURE 2 OMITTED]

3. NEURAL NETWORK STRUCTURE

The neural network structure is presented in figure 2, where x is the reference signal, p--electrical arc voltage, [U.sup.arc.sub.process] , e is the adaptive estimation of signal x, [x.sup.*] is the neural network output, FLA1, 2 are the linear adaptive filters. This "neural controller", is composed by two interconnected adaptive linear filters. FLA1 work on interferences compensation principle and that filter extract the useful signal from welding process signal. This useful signal contains the information about the deviation of the trajectory welding torch through the welding seam. FLA 2 work on inverse modeling principle and that filter eliminate the error between reference signal and FLA1 output signal (Haykin, 1999).

4. EXPERIMENTAL STALL

The process signal (figure 3) is obtained from a real WIG welding process realized with an experimental stall, especially builder for that (Bucur et al., 2002). Characteristics:

* Electrical engine: MCC 12V.

* Welding source: LUD 450W ESAB ARISTO.

* Welding machine: AGA DW 22: 220V, 0.4 A, 50W.

* Welding process: WIG without addition material; wolfram electrode 2.4 mm; protection gas: argon; protection gas pressure: 10 bar; welding current: 65 A; welding voltage:

10 ... 30V; advance speed: 70 mm/min; oscillation frequency:

0.5 Hz; oscillation amplitude: [+ or -]1.5 mm.

* Process-computer interface: AX 5411.

[FIGURE 3 OMITTED]

[FIGURE 4 OMITTED]

5. ANALYTICAL COMPUTATION OF THE REFERENCE SIGNAL

The reference signal (figure 4) is analytically calculated for a welding seam with the next dimensions: L--seam width--1 mm; d--seam thickness = pieces high--2.3 mm; B--distance from welding torch to the seam--5.3 mm, [delta]--deviation of welding torch to the medium plan of the seam--0.

We know that the correspondence between the arc welding parameters: welding current IS, arc voltage [U.sub.arc] and length of electrical arc [1.sub.arc,] for WIG welding processes is done by:

[U.sub.arc] = [a.sub.1] [I.sub.S] + [a.sub.2] + [a.sub.3] / [I.sub.S] + [a.sub.4] [l.sub.arc.] (1)

For welding current [I.sub.S] = 65A (Bucur et al., 2002):

[U.sub.arc] = 8.412 + 0.3 x [l.sub.arc] (2)

A function that approximates the configuration of [U.sub.arc] is:

[U.sub.*] [ar.sup.cref] = -0.13s in([pi]t)--0.07cos(2[pi]t) + 3.62 [V]. (3)

6. NEURAL NETWORK SIMULATION

Neural network simulation was realized with the MATLAB program (Bucur & Bucur 2007). The program is:

time=0:0.2:27.8;
x = -0.13sin([pi]t)--0.07cos(2[pi]t) + 3.62;
plot(time,x); (figure 4)
p=[3.33984375
2.69775390625
2.55859375
2.49755859375
2.802734375
2.63427734375

3.4521484375
2.5341796875...];
plot(time,p); (figure 3)
t=x+p';
[w,b]=initlin(p',t);
[a,e]=adaptwh(w,b,p',t,0.00001);
[w,b]=initlin(e,x);
[a,e]=adaptwh(w,b,e,x,0.01);
plot(time,a,time,x) ; (figure 5)
plot(time,e) ; (figure 6)

[FIGURE 5 OMITTED]

[FIGURE 6 OMITTED]

In this program x is the reference signal, p is the process signal, w and b is adjusted weights, signal a is the output of neural controller and e is the error signal. The time of acquisition for arc voltage values is 28 seconds (Kharab & Guenther 2002).

We observe that time of transition is under 1 second and the deviation of equilibrium position for welding torch is near to zero value (figure 6).

7. CONCLUSION

This paper show the neural controller with linear adaptive filters can move the welding torch in a good position near to reference position. Also, during the WIG welding process, if the oscillation frequency is modified over the reference frequency applied to the neural network, we can demonstrate that the designed network can make the correction of the oscillation frequency of welding torch (Bucur et al., 2002).

In the future, the solution of neural network control of trajectory welding torch to the welding seam is very interesting, especially for MIG-MAG welding process, because will appear the burned drop transfer phenomena thru the welding arc.

8. REFERENCES

Bucur, G. & Bucur, L. (2007). Automatic WIG Welding Control using Feed Forward Neural Network. Annals of the University of Petrosani, Electrical Engineering, Vol.9, pp.260-265, ISSN 1454-8518

Bucur, G., Dumitrescu, St. & Micloni, V. (2002). Neural Network Control in Robotic Welding Processes. Journal of Symposium "35 de ani de activitate a Universitatii Petrol-Gaze la Ploiesti", Vol.LIV, No.2/2002, pp. 38-43, ISSN 1221/9371

Haykin, S. (1999). Neural Networks, Prentice Hall, ISBN 0-13273350-1, U.S.A.

Kharab, A. & Guenther, R. (2002). An Introduction to Numerical Methods in MATLAB, Chapman & Hall/CRC, ISBN 1-58488-281-6, U.S.A.

Stefanescu, C. & Cupcea, N. (2002). Control and Measurement Intelligent Systems, Blue Publishing House, ISBN 973-650058-6, Romania