Detection of spatter signature for streaming data in the laser metal deposition process

Abstract

In recent years, Laser Metal Deposition (LMD) has experienced significant advancements. For process monitoring purposes, in-situ sensors are often used, which tend to produce noisy data, and due to the short processing window, these data need to be automatically analyzed in real-time to ensure their reliability for further processing. A simple Moving Average (MA) is commonly used to reduce signal peaks, which could otherwise skew the statistical properties of the data. Stabilization of the LMD process can be ascribed to the occurrence of spatters, which exhibit concept drift characteristics and are closely related to signal peaks. In this respect, this study aims to differentiate between two types of anomalies in data streams: point anomalies and concept drift, to eliminate the peaks that could cloak the performance in the actual signals during the process. To solve this issue, a two-step approach is being proposed. A differencing method is first applied to identify any potential point outliers, which are then verified to check if these identified observations are indeed peaks resulting from the spatters generation with a density-distance approach. The method's reliability and robustness were tested with overhang structures (3-axis printing) and impeller blade structures (5-axis printing). Results show that the existing method, the Drift Streaming Peaks-Over-Threshold method, is inferior compared to the proposed method in terms of F1-score, despite a decrease in performance as the inclination angle increases. These experiments ascertain the pertinence of the proposed method in processing incoming sensor data of LMD.