Answers are grounded in this paper and cite their sources.
The industrial usage of Directed Energy Deposition continues to lag behind earlier projections. One major reason for this is the complexity of process control, which is primarily driven by the high energy input. This often results in process instability, requiring multiple iteration loops before a component can be successfully manufactured. This reduces the overall economic efficiency. To address this challenge, various approaches for process monitoring have been developed with the aim of detecting defects at an early stage and enabling countermeasures to avoid additional costs. Increasingly, artificial intelligence is being applied in this context. However, defect classification methods typically require large datasets that include defective samples for training. Such datasets are often generated by deliberately inducing faults, leading to significant material consumption, increased costs, and extended production times. To overcome these limitations, this paper presents an autoencoder architecture trained exclusively on defect-free data that is capable of distinguishing between defect-free and defective samples using a reconstruction error threshold. The dataset consists of images captured by a CMOS industrial camera during a robot-based Directed Energy Deposition process. Defective samples were used solely for evaluating the autoencoder.
No takes yet. Share an insight, caveat, or question.
Cousin et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: