In-situ defect detection in laser-directed energy deposition with machine learning and multi-sensor fusion

Abstract

Early defect identification in laser-directed energy deposition (L-DED) additive manufacturing (AM) is pivotal for preventing build failures. Traditional single-modal monitoring approaches lack the capability to fully comprehend process dynamics, leading to a gap in multisensor monitoring strategies. This research proposes a novel in-situ monitoring method using a multi-sensor fusion-based digital twin (MFDT) for localized quality prediction, coupled with machine learning (ML) models for data fusion. It investigates acoustic signals from laser-material interactions as defect indicators, crafting a ML-based pipeline for rapid defect detection via feature extraction, fusion, and classification. This approach not only explores acoustic features across multiple domains, as well as coaxial melt pool images for ML model training, but it also introduces a novel MFDT framework that combines data from coaxial melt pool vision cameras and microphones, synchronized with robotic movements, to predict localized quality attributes. The key novelty in this research is the exploration of intra-modality and cross-modality multisensor feature correlations, revealing key vision and acoustic signatures associated with varying process dynamics. A comprehensive understanding of how multi-sensor signature varies with process dynamics improves the effectiveness of the proposed multi-sensor fusion model. The proposed model outperforms conventional methods with a 96.4 % accuracy, thereby setting a solid foundation for future self-adaptive quality improvement strategies in AM.