【摘要】 激光粉末床熔融(LPBF)是制造高性能、轻量化、多功能复杂金属构件的重要增材制造方法。LPBF过程的在线监测与诊断控制是实现产品质量评价、提升质量可靠性与一致性的重要保障。多材料LPBF面临不同材料物理特性差异大、可打印性多变等挑战，对缺陷的在线监控提出了更为迫切的要求。本文综述了LPBF过程在线监测、诊断与控制的最新研究进展，通过分析未熔合、匙孔孔隙、裂纹等典型内部缺陷的特征和形成机理，总结了缺陷的在线监测需求；围绕光、热、声、电磁等传感信号，回顾了近年来LPBF在线监测技术的发展，并分析了当前LPBF在线监测、缺陷诊断与控制存在的局限性；重点探讨了人工智能(AI)在提升LPBF在线监测单一信号与多源信号的诊断能力及缺陷反馈控制方面的重要作用，并对AI驱动的LPBF在线监测与控制的未来发展进行了展望。更多还原

【Abstract】 Significance Laser powder bed fusion(LPBF) is a key additive manufacturing process for producing high-performance, lightweight, and multifunctional complex metal components.However, the transient multi-physics nature of LPBF poses significant challenges to process stability, often leading to defects such as lack of fusion, porosity, and cracking, which can severely compromise the structural integrity and service reliability of fabricated parts.For advanced applications like multi-material LPBF, the fabrication process becomes even more intricate due to variations in thermo-physical properties among different materials, further intensifying the demand for robust in situ monitoring.In situ monitoring and diagnostic evaluation are therefore essential for assessing product quality and enabling stable, repeatable LPBF production.Artificial intelligence(AI), with its powerful capability to extract critical features from high-dimensional, noise-contaminated, and multi-source heterogeneous datasets, plays a pivotal role in enhancing the accuracy, intelligence, and real-time capability of defect diagnosis.This review first summarizes the requirements for in situ monitoring based on an analysis of defect formation mechanisms.Subsequently, it comprehensively reviews recent advances and limitations in LPBF in situ sensing technologies.The current state of AI applications in defect diagnosis is thoroughly discussed, and perspectives on the future development of AI-driven LPBF monitoring and control are provided.Progress Research on defect formation mechanisms in LPBF indicates that coupled multi-physics interactions, melt-pool behavior under cyclic heating-cooling and extremely non-equilibrium solidification, and various instabilities are the primary causes of defects.For multi-material LPBF, the process involves disparate material properties, and a single set of process parameters is often insufficient, making lack of fusion and cracking prevalent at material interfaces and leading to interfacial failure.The highly variable printability in multi-material LPBF results in more complex process signatures, imposing stricter requirements on the sensitivity, accuracy, and reliability of in situ monitoring as well as the precision of process control.Accordingly, LPBF in situ defect monitoring primarily targets unstable factors during fabrication, abnormal variations in physical fields, and changes in molten pool geometry and dimensions, enabling the acquisition of key signatures associated with defect initiation and evolution.These feature data form the basis for subsequent defect diagnosis and process control.Current LPBF in situ monitoring technologies encompass multiple sensing modalities, including optical, thermal, acoustic, and electromagnetic signals.Optical monitoring based on industrial cameras provides layer-wise information on powder spreading quality, surface morphology, and geometric features, while high-speed cameras capture transient melt-pool dynamics, spatter, and plume evolution on micro-to millisecond timescales.In thermal monitoring, photodiodes acquire full-field time-series signals of molten pool radiation intensity via point-wise measurements; thermal tomography and infrared thermography enable full-field temperature mapping during and after deposition.To mitigate measurement errors caused by dynamically changing surface emissivity, dualwavelength ratio thermometry, which synchronously acquires radiation at two different wavelengths, can effectively improve temperature measurement accuracy.Furthermore, emerging techniques such as multispectral thermometry, thermionic emission monitoring, and embedded-sensor thermometry are under active development.For acoustic monitoring, laser ultrasonics enables near-surface and internal defect detection, whereas acoustic emission monitoring captures acoustic signatures generated by process instabilities.Coil-based or magnetoresistive sensor-based eddy current testing is a major approach in electromagnetic monitoring for in situ defect detection; meanwhile, near-field microwave imaging and related techniques are gradually being introduced into LPBF monitoring.In the realm of defect diagnosis and control, AI has become an indispensable tool for improving diagnostic accuracy and enabling intelligent control.For single-signal defect diagnosis, machine learning and deep learning models(e.g., SVM, random forests, CNN/Transformer variants, and RNN/LSTM hybrids) have been widely applied to analyze different sensing signals for LPBF defect diagnosis.These approaches perform feature extraction, classification of defect-inducing factors, and defect identification with high reported accuracy.To overcome the limitations of single-signal monitoring and to reduce false alarms and missed detections, multi-signal fusion methods are being urgently developed.Multi-signal data fusion for defect diagnosis includes AI-driven integration of heterogeneous signals(e.g., optical, thermal, acoustic) for defect identification, as well as the fusion of multi-layer time-series signals to analyze inter-layer defect evolution features.Such fusion strategies can significantly improve diagnostic reliability and enable cross-validation.Furthermore, physics-informed neural network-based approaches are emerging as a promising route to enhance model accuracy and interpretability under limited labeled data by embedding physical constraints, thereby reducing data dependence while substantially improving generalization capability.Finally, existing studies demonstrate that AI-enabled adaptive parameter adjustment can effectively improve build quality, paving the way towards “dark factory” intelligent manufacturing for LPBF.Conclusions and Prospects Significant progress has been achieved in LPBF in situ monitoring, diagnosis, and control.Currently, an LPBF in situ monitoring framework spanning optical, thermal, acoustic, and electromagnetic signals has been established, enabling the acquisition of process signatures across different spatiotemporal scales.While the capabilities of existing monitoring techniques continue to advance, emerging sensing approaches are also being actively explored.The monitoring paradigm is gradually evolving from pure process monitoring towards integrated process and outcome monitoring.AI has been extensively applied to defect diagnosis, markedly improving diagnostic efficiency and accuracy.Multi-signal data fusion effectively addresses the limited observability of single-signal monitoring and enhances both the accuracy and generalization of diagnostic methods.Physics-constrained diagnostic approaches strengthen model generalizability and interpretability, representing a promising future direction.In defect control, monitoring-informed adaptive parameter regulation and closed-loop control can effectively improve the reliability and consistency of build quality, laying the groundwork for intelligent manufacturing.However, owing to the spatiotemporal multi-scale nature of LPBF, existing monitoring approaches still struggle to simultaneously satisfy the requirements of high spatiotemporal resolution, large monitoring coverage, and long-duration observation.Current techniques are primarily designed to capture process signatures rather than directly measure defects.Moreover, diagnostic capabilities remain limited, and effective defect control is still insufficient, highlighting the critical need to further develop AI-driven LPBF in situ monitoring and control methods.Future research should prioritize: 1) enhancing the capability to sense multi-source physical signals and advancing emerging sensing technologies; 2) fully leveraging multi-physics signal data and simulation data through multi-signal fusion to enable real-time defect prediction and diagnosis; 3) advancing digital twin and related technologies to achieve accurate quality assessment and real-time defect remediation, thereby establishing a new paradigm for intelligent quality control of complex, integrated LPBF structures.更多还原