Data-driven control of laser powder bed fusion

Abstract

There is growing attention to additive manufacturing (AM) processes in industry and academia. Specifically, one metal AM process - laser powder bed fusion (LPBF) - is particularly attractive as it can directly produce functional metal parts with fine feature resolution. However, LPBF suffers from quality control issues that hinder its wider adoption. These issues are not easily addressed, as the process is challenging to monitor and computationally expensive to model. Significant research efforts have been devoted to the improvement of LPBF quality outcomes: real-time process monitoring, control, and computational modeling have all seen progress in recent years. This thesis contributes to efforts in empirical modeling and real-time control of LPBF, approaching them from a mechatronics perspective. It first asks: what should be measured, what should be controlled, and how those things are empirically related. Then, given those empirically-modeled relations, the laser power is controlled in both a feedback and a feedforward manner to regulate the melt pool emission. But first, to make these studies possible, systems for melt pool monitoring and control were designed and implemented on a laboratory testbed. The designed control system allows for laser power modulation based on either the melt pool images or a pre-defined input profile, at 2 kHz. With the described system in place, both feedback and feedforward control schemes were experimentally demonstrated. With respect to feedback control, a melt pool signal reference was tracked on a part scale. First, transfer function models were identified from the experimental data. Then, feedback controller design and tuning were performed, accounting for process-related plant model variation, as well as modelling uncertainties. The feedback controller corrected in-layer and inter-layer melt pool signal drifts present during open-loop operation. With respect to feedforward control, this work makes two contributions. First, iterative learning control was shown to correct for repeating layer-to-layer disturbances in LPBF, but only if the geometry is repeating exactly. Second, certain geometric features, such as corners and narrow sections, were shown to create a disturbance that is observable in coaxial melt pool images. Further, the geometry-dependent model of this behavior was constructed, validated experimentally, and used for model-based feedforward control. The designed model-based feedforward controller reduced geometry-related signal deviations by 50%.