Neuro-fuzzy control modelling for gas metal arc welding process

Weld quality features are difficult or impossible to directly measure and control during welding, therefore indirect methods are necessary. Penetration is the most important geometric feature since in most applications it is the most significant factor affecting joint strength. Observation of penetration is only possible from the back face of the full penetration weld. In all other cases, since direct measurement of depth of penetration is not possible, real time control of penetration in the Gas Metal Arc Welding (GMAW) process by sensing conditions at the top surface of the joint is necessary. This continues to be a major area of interest for automation of the process. The objective of this research has been to develop an on-line intelligent process control model for GMAW, which can monitor and control the welding process. The model uses measurement of the temperature at a point on the surface of the workpiece to predict the depth of penetration being achieved, and to provide feedback for corrective adjustment of welding variables. Neural Network and Fuzzy Logic technologies have been used to achieve a reliable Neuro-Fuzzy control model for GMAW of a typical closed butt joint having 60° Vee edge preparation. The neural network model predicts the surface temperature expected for a set of fixed and adjustable welding variables when a prescribed level of penetration is achieved. This predicted temperature is compared with the actual surface temperature occurring during welding, as measured by an infrared sensor. If there is a difference between the measured temperature and the temperature predicted by the neural network, a fuzzy logic model will recommend changes to the adjustable welding variables necessary to achieve the desired weld penetration. Large scale experiments to obtain data for modelling and for model validation, and various other modelling studies are described. The results are used to establish the relationships between the output surface temperature measurement, welding variables and the corresponding achieved weld quality criteria. The effectiveness of the modelling methodology in dealing with fixed or variable root gap has also been tested. The result shows that the Neuro-fuzzy models are capable of providing control of penetration to an acceptable degree of accuracy, and a potential control response time, using modestly powerful computing hardware, of the order of one hundred milliseconds. This is more than adequate for real time control of GMAW. The application potential for control using these models is significant since, unlike many other top surface monitoring methods, it does not require sensing of the highly transient weld pool shape or surface.