Development And Validation Of Soft Computing Based Models For Pulsed Gas Metal Arc Welding Process.Arc Welding Is A Very Complex Thermo-Metallurgical Process. : And It Is Very Difficult To Develop A Physical Model Of Such A Process. Apart From Physical Model, One Can Also Use Soft-Computing Techniques, Which Apply Numerical Interpolation Or Approximation Of Discrete Data. The Main Weld Quality Control Tests Are The Destructive Tests, Which Are Carried Out On Welds Obtained From The Production Line Of The Product; And As A Result It Increases Production Cost And Makes This Approach Unattractive. Therefore, On-Line Monitoring Or Prediction Of Welding Quality Is Very Crucial To Maintain Proper Weld Quality And Detect Welding Defects Quickly. In This Research, A Series Of Experiments, By Applying Response Surface Method, Were Carried Out. The Welding Process Parameters Considered Were Pulse Voltage, Background Voltage, Pulse Frequency, Pulse Duty Factor, Wire Feed Rate And Table Feed Rate, And The Weld Quality Parameters Considered Were Tensile Strength, Bead Geometry, Distortion, And Deposition Efficiency. During The Experiments, Acquired Arc Signals (Current And Voltage) Are Analyzed For Extracting Weld Quality Sensitive Features Using Time-Domain And Wavelet Packet Analyses. Process Parameters And/Or Various Combinations Of Signal Features Were Used To Formulate Different Modeling Strategies, Which Were Then Used For Indirect Prediction Of Weld Quality Using Different Soft-Diagnostic Tools Such As Back Propagation Neural Network (Bpnn) And Six Different Radial Basis Function Networks (Rbfns). Comparative Evaluation Of All Artificial Neural Network (Ann) Models For Different Weld Quality Parameters For Various Strategies Has Been Reported. The Results Show That The Ann Model Trained With Root Mean Square Values Of Wavelet Packet Coefficients Of Current Signal Along With The Process Parameters, Gives Superior Prediction Of Weld Quality As Compared To Simple Use Of Process Parameters With/Without The Statistical Properties Of The Untransformed Arc Signal. Bpnn Model Produces Superior Prediction As Compared To Rbfn Models For Ultimate Tensile Strength, Bead Geometry And Deposition Efficiency. In Rbfn Models, Sigmoid Activation Function For The Output Layer With Full Training Algorithm Was Found To Be The Best, And Produced Superior Prediction Of Distortion. Furthermore, A Neuro-Ga Model Is Developed For Determination Of Optimal Input Parameter Setting For Any Desired Combination Of Quality Parameters.