Overshooting defects in shipbuilding subassemblies are essential to ensure the final product’s overall integrity and safety. In this work, we focus on the automatic detection of overshooting defects in simple and T-shaped sub-assemblies by employing reconstruction-based unsupervised learning on 3D point clouds. To this purpose, we implemented and compared four state-of-the-art architectures, including a Variational Autoencoder (VAE), FoldingNet, a Dynamic Graph CNN (DGCNN) autoencoder, and a PointNet++ autoencoder. These architectures were trained exclusively on defect-free samples, anticipating the possibility of overshooting defects occurring in different locations and with varying geometric patterns that are difficult to characterize explicitly in advance. Those defects are then identified by applying an Isolation Forest to the reconstruction error features, enabling fully unsupervised anomaly detection and allowing us to study how the detection performance changes with the contamination parameter. The results show that reconstruction-based anomaly detection on point clouds is a viable strategy for identifying defects in an industrial environment and the importance of choosing architectures that balance detection performance, stability across different geometries, and computational cost.

La irrupción de los modelos de lenguaje de gran tamaño (LLM) ha abierto nuevas vías para la interacción con sistemas industriales complejos, en particular con plataformas SCADA. Este trabajo presenta una propuesta de arquitectura para la integración de agentes basados en LLM con sistemas SCADA mediante el protocolo abierto MCP (Model Context Protocol), comparándolo con la alternativa de las llamadas skills, instrucciones en lenguaje natural que orquestan herramientas preexistentes. Se identifican las limitaciones estructurales del uso de herramientas genéricas de bajo nivel, especialmente la alucinación de identificadores en cadenas largas y repetitivas, y se justifica el diseño de herramientas específicas de alto nivel. Se describe una arquitectura modular replicable y se reportan los resultados obtenidos en un caso de estudio industrial, donde la propuesta permite reducir significativamente el número de invocaciones requeridas y aumentar la fiabilidad de las respuestas frente al enfoque basado en skills.

Language-model agents are increasingly proposed to operate and query industrial systems, yet their adoption meets a persistent obstacle: scaling improves reliability, but not to the level an industrial environment demands, where a wrong decision can cause losses. This work argues a position, structure before scale, whereby, without denying the role of model capability, the greatest lever for reliability on the plant floor is the structure imposed on the model, which moreover does not require the infrastructure cost of a larger model. The position is built from recent literature, which converges on three enablers (knowledge graphs, agents and the MCP protocol) and on a well-documented reliability ceiling, and is illustrated with one deployment: operating a SCADA system through an agent and MCP, where encapsulating brittle coordination in code cuts model invocations from 20--50 to a single one.