Deep Spatio-Temporal Anomaly Detection in Laser Powder Bed Fusion

Parametric and regression-based anomaly detection methods often fall short when faced with high-dimensional and spatio-temporally correlated data. A multitude of unrealistic assumptions built in such methods renders the decisions unreliable and prone to large errors, especially in uncertain settings. Model agnostic deep learning methods are emerging at an ever-growing pace as a result of recent advancements in high-performance computing technologies. Not only do they offer a great deal of flexibility when facing unstructured data, but also their generalization power makes them an appealing alternative to their physics-based or regression-based modeling counterparts. It has also been shown that relying on a single melt pool image to detect Additive Manufacturing (AM) process anomalies is usually ineffective and can result in significant inflation of the false alarm rate. In this study, we propose a configuration of convolutional long short term memory auto-encoders to learn a deep spatio-temporal representation from sequences of melt pool images collected from experimental AM builds. The extracted bottleneck tensors are unfolded and fed by an agglomerative clustering algorithm to annotate the anomalies. A dual control charting scheme composed of a Hotelling’s $T^{2}$ and a residual’s variance $S^{2}$ statistics is constructed and validated on training and validation normal samples, respectively. Evaluating the performance of the method against unseen test data demonstrates its high accuracy and low false alarm rate in various AM process running scenarios. Note to Practitioners—A prior technical knowledge of the specific AM technology is often necessary to model the physics-based relationships and detect the anomalies in the process. On the other hand, high-dimensional and spatio-temporally correlated co- axial images are widely captured during the laser powder bed fusion process. However, data-driven regression-based methods suffer from inadequacy and lack of generalizations in the feature extraction development phase. This paper proposes a framework to translate the collected data into simple and easy to understand statistical scalars in real-time. Practitioners will be equipped with generalized feature engineering and ultimately control chart methodologies that automate the anomaly detection process without requiring to understand the underlying physical relationships in the process.