Work

Mitigating the influence of anomaly detection errors on reinforcement learning for in-situ process optimization of directed energy deposition

Year: 2026

Type: article

Abstract: Laser-based directed energy deposition (DED-LB) enables additive manufacturing of complex metal parts through layer-by-layer material deposition. However, the process presents challenges in preventing defects due to thermal and melt pool dynamics, which can be addressed with in-situ monitoring and process optimization. Reinforcement learning (RL) enables in-situ process optimization policies that can simultaneously adjust multiple process parameters and dynamically respond to process variations. Anomaly-driven reinforcement learning (AD-RL) integrates anomaly detection (AD) models into RL to learn strategies that minimize deviations from normal process conditions, thereby avoiding anomalous process states that can lead to defects. Since the learned strategies depend on the detection accuracy of the AD model, it is necessary to investigate the influence of potential errors of AD models on the learned RL strategies. The sources of such errors can be systematic, if the AD model lacks sufficient detection capability resulting in repeated misclassifications, or stochastic, e.g., from measurement noise or other random perturbations. In this work, the error influence of AD models (hypershpere classifiers) on RL strategies (trained using proximal policy optimization) for DED-LB is analyzed. The systematic errors are introduced by varying the detection capability of the AD model through structured degradations, while stochastic errors are examined through randomly induced anomaly score perturbations during RL training. Based on the analysis, this work contributes to improving in-situ process optimization in DED-LB by suggesting how to mitigate the influence of potential AD errors.

Source: Procedia CIRP

DOI: 10.1016/j.procir.2026.03.174