A STUDY OF MACHINE LEARNING FRAMEWORK FOR ENABLING EARLY DEFECT DETECTION IN WIRE ARC ADDITIVE MANUFACTURING PROCESSES

This paper presents the study on the performance of a variety of proposed time-domain acoustic features-based frameworks for the detection of geometrically defective print segments during the Wire Arc Additive Manufacturing (WAAM) process. Specifically, we investigate into a variety of acoustic features, namely the Root Mean Square of Pressure (RMSP), Energy, Mean Amplitude, Kurtosis, Zero Crossing Rate (ZCR), Skewness, Crest Factor and Peak-to-peak, and print process parameters, namely Torch Speed (TS) and Wire Feed Rate (WFR) combined with Machine Learning (ML) frameworks for detecting geometrically defective print segments. Experiments carried out on Inconel 718 show that among the studied frameworks, using acoustic features and process parameters with Random Forest (RF) performs best in terms of F1 score at 89%, while using acoustic features and process parameters with Support Vector Machine (SVM) performs best in picking out defective segments based on the Confusion Matrix. These findings serve as our first step in developing an intelligent sensing system for the early identification of defective beads in the WAAM printing process, so that appropriate intervention can be implemented to save printing resources and material costs. In addition, the proposed approach has the advantage of detecting defects within a more localized region for more targeted intervention.