Unsupervised Online Anomaly Detection of Metal Additive Manufacturing Processes via a Statistical Time-Frequency Domain Approach

Fault detection techniques in metal additive manufacturing (AM) have explored a variety of monitoring methods to flag anomalies as they occur during the sintering process. Although many in-situ techniques are able to adeptly detect these abnormalities, several utilize machine learning black box methods that do not easily transfer to varying print geometries. An approach that is adaptable to a multitude of geometries holds an advantage in determining anomalies for more complex cross-sections and raster patterns. To address this lack of a geometry agnosticism, we propose a method that detects faults using the frequency content of the melt pool image response through an unsupervised approach. Scan line length and scan speed extracted from known geometry can be translated to associated frequencies via a spectrogram. We examine three specific geometries to determine detection performance on each by comparing the frequency content to the nominal response. A deviation from the expected performance will signify that an anomaly has occurred. We verify this approach is feasible for fault detection and is accurate in detecting anomalies that are hard to observe in the image time series. A feasible geometry agnostic method and the current interpretability will be discussed in this paper. The results reached in this paper strongly indicate that the approach is promising, has potential for improvement, and that a geometrically independent method is sensible. With further work, a generic algorithm applicable on any geometry will be achievable.