Kevin Omede
Dynamic Model-Context Alignment for Smart Manufacturing: A Digital Twin-Driven Framework for Continuous Adaptation of Open-Source Models with Conversational XAI.
Rel. Alessandro Simeone. Politecnico di Torino, Master of science program in Industrial Production And Technological Innovation Engineering, 2026
Abstract
Artificial intelligence models deployed statically on manufacturing edge devices rapidly become suboptimal as operational context evolves: sensor drift, process changes, hardware upgrades, and seasonal patterns continuously alter the effective model-context alignment. Existing approaches select models based solely on task type and ignore the structured constraints exposed by the company’s Digital Twin (DT), leaving no mechanism for autonomous re-alignment once deployed. We propose a closed-loop framework—Dynamic Model-Context Alignment (DMCA)—that continuously re-evaluates model selection as the DT updates its operational profile. The cycle comprises five stages: (1) DT profiling via Asset Administration Shell (AAS) parsing, (2) multi-criteria model selection from HuggingFace, (3) edge deployment with ONNX optimization, (4) concept drift detection via ADWIN, and (5) automatic re-alignment with human-in-the-loop confirmation through a conversational XAI copilot.
We present an empirical benchmark of 8 open-source time-series foundation models (Chronos, Moirai, TimesFM, PatchTST, MOMENT, Lag-Llama) on two industrial datasets (NASACMAPSSFD001 and ETT-h1), executed on a GPU NVIDIA T4 via serverless cloud infrastructure at a total cost of $0.30
Relators
Academic year
Publication type
Number of Pages
Additional Information
Course of studies
Classe di laurea
Ente in cotutela
Aziende collaboratrici
URI
![]() |
Modify record (reserved for operators) |
