Evaluation of Time-Domain Acoustic Signature in TIG Welding of 5083 Aluminum Alloy: A Methodological Comparison of Feature Reduction Approaches

• A machine learning model is developed for identifying weld conditions in aluminum TIG welding. • Desired weld condition is obtained experimentally by varying the weld parameters. • Feature selection methods were applied to sound signature to optimize input features for models. • Among the models tested, decision tree, using Chi-Square method achieved highest accuracy of 86.7%. • The results suggest that machine learning models can be used for real-time monitoring systems. Despite of Tungsten inert gas (TIG) welding being a traditional technique for joining aluminium and aluminium alloys, the formation of porosity, burn-through, and cracks as a result of inadequate heat control is a serious threat to the structural integrity of the joints. Non-destructive post-welding inspection is expensive, less productive and undergoes tedious procedure. In this scenario, condition monitoring of the weld using machine learning algorithms is a remedy to detect the defects instantly. In the present study, a machine learning model was developed to identify weld conditions such as good weld, porosity, and burn-through in TIG welding of aluminium alloy. Process parameters, including welding current, weld speed, wire speed, gas flow rate, and root gap, are varied to get the desired welding condition. The sound signatures of weld conditions are captured using a microphone with a sample rate of 10 kHz. The feature selection and feature reduction are performed on the sound signature data using ANOVA, MRMR, Chi-Square, reliefF, and Kruskal-Wallis methods. Neural network, ensemble, CART, and SVM techniques are used in the development of the classification models. The medium and fine tree models showed a maximum classification accuracy of 85.6% using the ANOVA method whereas Medium tree model with MRMR showed maximum classification accuracy of 85%. For the Chi-Square and ReliefF methods, the maximum classification accuracy obtained is 86.7% and 85.6% for the fine and medium trees. For bagged tree model, a maximum accuracy of 86.1% was obtained with Kruskal-Wallis method. Among all, decision tree with Chi-Square method is found to be the best, with a classification accuracy of 86.7%. The findings of the study are promising for developing online monitoring systems to predict weld quality with high accuracy.