Unsupervised online anomaly detection in laser powder bed fusion via statistical time-series methods: An experimental assessment and feasibility study

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Metal additive manufacturing (AM) processes are susceptible to in-process faults that compromise part quality or interrupt printing. Early defect detection enables process intervention, reducing lead time and material waste. This work presents a statistical time-series framework for online anomaly detection in laser powder bed fusion (LPBF). Several online statistical time-series methods are postulated and assessed, leveraging statistical detection criteria and an unsupervised dataset to identify anomalies based on user-defined statistical thresholds. To enable online detection and minimize training data requirements, the melt pool area from coaxial camera images is extracted to create a compressed univariate time-series representation. Two stochastic model structures, autoregressive (AR) and recursive AR (RAR), learn expected nominal behavior. Melt pools are declared anomalous through statistical hypothesis testing, with detection sensitivity determined by the statistics of the nominal response and user-defined type I and II error probabilities. Three model-based detection methods are proposed and assessed against baseline time-series thresholding and autoencoder (AE) reconstruction error approaches. The detection tests were evaluated on natural and artificial anomalies across two raster geometries. The two best-performing methods were the RAR parameter-based confidence interval test and the sequential probability ratio test (SPRT), achieving false alarm rates of 0 . 11 % and 0 . 25 % , respectively, outperforming both baseline threshold and AE-based approaches. The RAR method leverages extended sample history for improved predictive accuracy, while the SPRT adapts its decision window length for rapid detection when anomalies are clear and extended sampling within an “uncertainty zone” when ambiguity remains. These statistical time-series structural health-monitoring approaches have yet to be used in LPBF metal AM processes. The results demonstrate that simple, interpretable features like melt pool size, combined with lightweight statistical algorithms, can deliver accurate, fast online anomaly detection from sparse unsupervised training datasets.