Analysis and Modeling of GTAW Weld Pool Geometry - Intelligent Computing for Sustainable Energy and Environment - page 291

Table 4. Simulation errors of BPNN model based on and

Training MSE Training ABE Training RE Testing MSE Testing ABE Testing RE
0.1766 0.2931 0.1272 0.4303 0.5531 0.1809
0.0011 0.0208 0.1433 0.0216 0.1075 0.5385
In this paper, PCA method was used to conduct dimension reduction from 11 to 4,
and obtain the soft-sensing model. Fig. 7 is the model simulation results based on the
PCA-BPNN model.
Table 5 shows the simulation errors. It was concluded that PCA-BPNN model is
better in sensing and based on the top-side characteristic parameters. This is
because all eleven auxiliary variables are used here, and more information about
penetration status is included in the model. Hence, the PCA-BPNN model can be
recommended as the preferred method in this paper to obtain real-time estimation for
and in GTAW process.

6 0.5

Process value
PCA-BP output

Process value
PCA-BP output

0.45

5

0.4

0.35

4

0.3

3 0.25

0.2

2

0.15

0.1

1

0.05

0 0

0 5 10 15

(a) ܹ ௕ (b) ܪ ௕

20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50

Sample No Sample No

Fig. 7. Simulation results of PCA-BPNN model based on all 11 inputs

Table 5. Simulation errors of BPNN model based on all 11 inputs

Training MSE Training ABE Training RE Testing MSE Testing ABE Testing RE

0.0613 0.1448 0.0688 0.2602 0.4679 0.1561

0.0005 0.0144 0.0929 0.0036 0.0470 0.2759

5 Conclusion

Based on the newly developed three dimensional vision monitoring system for the
GTAW process, 304stainless steel pipe with 2.03 mm thickness was used to conduct
experiments. The proposed novel top-side characteristic parameters are found to be
closely related to the back-side weld pool parameters that directly quantify the weld
penetration. Developed neural network models are effective in estimating the weld
penetration specified by the back side bead width and height. With the proposed top-
side characteristic parameters and neural network models, weld pool observation and
penetration prediction can be realized in real time conveniently.