Monitoramento do processo de manufatura aditiva por deposição à arco, utilizando algoritmo de aprendizagem e fusão de sensores

Abstract

This work presents the development of a process-monitoring system for Wire Arc Additive Manufacturing (WAAM) to support the inspection stage of fabricated parts by predicting and indicating potentially defective regions. The solution integrates a data acquisition and instrumentation system, a machine-learning algorithm, sensor fusion, anomaly clustering, and cross-validation with radiographic imaging. The experimental plan comprised the deposition of walls in carbon steel and stainless steel, using different deposition strategies, controlled induction of instabilities, and the acquisition of time series for the process variables: welding voltage and current, shielding-gas flow rate, wire feed speed, travel (deposition) speed, arc luminous intensity, and deposition acoustic signals. Anomaly detection was designed through interactive analysis pipelines, a learning algorithm based on an LSTM Autoencoder, and two indicators of anomalous events: an Amplitude Threshold (large-magnitude events in the reconstruction error) and a Volatility Threshold (persistent fluctuation events). Subsequently, a sensor-fusion strategy and clustering (K-Means) of hybrid features (statistical and wavelet-based) were applied, together with dimensionality reduction (PCA), to obtain groups with signatures/profiles consistent with regions exhibiting an increased probability of defect occurrence. System validation was performed by overlaying the predictions onto the radiographic images of the fabricated parts, thereby enabling comparison between the regions indicated by the prediction system and the defective areas actually observed.