A standards-based systems engineering framework for AI-enabled in-situ monitoring in laser powder bed fusion

Deploying AI-based monitoring systems in laser powder bed fusion (PBF-LB/M) requires upstream decisions concerning process knowledge, data quality, and operational constraints. Many studies reviewed here report these decisions as task-specific choices, which makes it difficult to relate model performance to deployment requirements. This paper presents a standards-integrated systems engineering framework for AI monitoring in PBF-LB/M that combines ISO/IEC 25059, ISO/IEC 5259–4, ASTM E3353, and ISO/IEC 5338. The framework formalizes monitoring-target definition, label-taxonomy derivation, deployment-oriented data partitioning, and latency specification before model development and links these decisions to explicit verification evidence. We evaluate one instantiation through paired compliant and non-compliant conditions using the same data and model architectures. The comparisons show that upstream design choices affect deployment-relevant capability even when model capacity is held constant. Binary labeling yields a higher aggregate classification score but cannot distinguish the opposite corrective actions associated with keyhole-prone and lack-of-fusion-prone conditions. A mixed scan-pattern split obscures shifted-condition performance, while monitoring-scale analysis changes which architectures satisfy the latency constraint. Of the six quality-characteristic requirements specified for the case study, five satisfy the measured acceptance criteria. Robustness remains unmet because classification performance degrades under the tested compound shift, although a feature-space detector identifies the tested shifted condition with low computational overhead. Evidence for Intervenability is limited to the model-independent computational decision path because controller communication and physical actuation are not measured. The annotation protocol is evaluated through sensitivity analysis, process-scale consistency checks, and an independent expert assessment, but direct volumetric defect ground truth is unavailable. These results show how the framework converts upstream design decisions into traceable verification targets and exposes deployment limitations that aggregate model metrics can obscure. The framework is intended as a reusable design structure, while empirical validation in this study is limited to the present PBF-LB/M case.