Multi-Sensor Monitoring of Directed Energy Deposition for Anomaly Detection and Part Quality Control

Additive Manufacturing (AM) technologies are currently transitioning from prototyping and research tools towards widespread manufacturing methodologies for high-value industries. Metal-based AM has begun implementation in specialised cases to produce complex, lightweight, and functionalised parts for high-value applications. Currently, the development, production, inspection, and certification of these parts is long and intensive, involving many man-hours of specialist work. This thesis investigates how to optimise this process to lower barriers for further adoption of AM through multi-sensor process monitoring, machine learning and process modelling. Laser-beam Directed Energy Deposition (DED-LB) is a form of metal AM used to produce free-form components, surface claddings, and facilitate component repair, making it a valuable tool for industry. The layer-wise fabrication, complex thermal, material, and environmental interactions can allow for the formation of process-induced defects, or undesirable geometric features. These anomalies are not easily detected during manufacture and may remain undetectable without specialised inspection of completed components. Process monitoring has recently been employed to understand the formation of anomalies in AM fabricated components. The current approach is to implement a single sensor or to monitor one aspect of the fabrication, which fails to capture the full range of interactions involved in the creation of complex components. This thesis implements a new multi-axis monitoring system designed to provide continuous and wholistic measurements of melt pool shape, size, and position throughout deposition. To demonstrate this system’s capabilities, intersecting thin-walled Ti-6Al-4V samples are produced at angles of 90⁰, 60⁰, and 30⁰ to mimic complex geometries. A processing algorithm is developed to interpret data from multiple sensors, drawing information regarding various anomalies from the synthesis of results. This multi-axis monitoring allows sample height distortion to be tracked, identifying anomalies such as bulges and depressions, with the latter correlating to severe porosity formation in intersection angles of 30°. This enabled early identification of severe defects within the first 10% of the build. Internal porosity in 30° intersections is determined to result from depletion of laser energy and powder flow and is eliminated through modification of the toolpath to remove fillets. This thesis highlights the importance of multi-axis monitoring and demonstrates the benefits of data fusion in observing process-induced defects. While anomalies can be identified from the monitoring data, to improve the quality of a component, the conditions causing that anomaly must be mitigated. Hence, this thesis employs an advanced multi-physics Finite Element Model (FEM) to inform a new processing paradigm. When producing a component, industry practice is to determine appropriate processing parameters through trial-and-error parameter search studies, testing many combinations of laser power, speed, spot size, etc. in simple depositions. After inspection, parameters are selected and applied at constant values throughout deposition, failing to account for the highly dynamic thermal conditions experienced during fabrication. To address this shortfall, this thesis employs a thermally calibrated FEM to determine adaptive processing parameters. In this work, high laser power (2 kW) and speed (1500 mm/min) parameters are applied in conjunction with no interlayer cooling as initial conditions to produce depositions with poor quality and substantial heat accumulation. The simulation uses these parameters as a starting point, and in conjunction with an optimisation route, predicts dynamic laser power and cooling intervals that lead to successful builds. The combination of pre-determined power reduction and interval cooling enable high-quality fabrications and consistent Widmanstätten microstructure throughout the deposit. It is noted that the simulation model is unable to account for factors such as fluid flow and oxygen ingress under localised shielding. Therefore, samples produced with localised shielding exhibited microhardness values as large as 677 ± 44 HV0.5, accompanied by brittle failure during deposition. When repeated under controlled atmosphere, microhardness values were consistent with literature, ranging from 315 ± 8 to 337 ± 9 HV0.5. This simulation-informed approach presents a novel method of process design, minimising lengthy process selection methods and mitigating defect formation prior to fabrication. As noted, experienced conditions may differ from simulated conditions during DED-LB, and process monitoring of these conditions generates large data files. Interpretation of the monitoring data is required to be reliable, fast, and involve minimal human intervention, else the scalability of AM fabrication will always be limited by the availability of human experts. Machine Learning (ML) algorithms demonstrate an exciting opportunity for the rapid interpretation of large and complex datasets, such as those generated by process monitoring, and have enabled breakthroughs in many fields in recent years. Here, the context-driven Long Short-Term Memory (LSTM) algorithm is demonstrated for its applicability to the prediction of multiple feature classes simultaneously for parts fabricated by DED-LB, an approach not identified in literature. The LSTM considers the context of datapoints in a time sequence, which is relevant to the evolution of the melt pool in DED-LB, where previous states influence the current state. To determine model architectures with highest performance, 120 network variations are trained per input data scenario (single or multi-source data). Few consistent trends are identified, reinforcing the necessity for hyperparameter optimisation routines when developing an ML tool. The performance of the algorithms is examined when trained on data from any single process monitoring data source, or from a combination of sources. For single-source algorithms, LSTMs trained on data from the coaxial camera performed best, despite lower spatial resolution and increased noise, likely due to the increased sampling rate. The use of multi-source data provides greater depth of information, allowing relationships for sparse data to be better learned than for any single data source, with up to 100-fold improvements in the true-positive rates for ‘Depression’ and ‘Porosity’ classes. However, the presented architectures are shown to not yet be appropriate for real-time prediction of features. The combination of research investigations in this thesis creates a new framework of advanced, and automated DED-LB production. This next generation framework will improve the understanding of defect and feature formation in DED-LB, enhance and accelerate the determination of process regimes for consistent part production, and refine the interpretation of monitoring data for rapid assessment of component quality.<p></p>