Sensor-Augmented Retrieval of System Dynamics Model Segments for Resource-Constrained Environments

Introduction: Efficient extraction of System Dynamics models in industrial IoT environments requires not only structural similarity but also the integration of real-time operational context. This paper presents a sensor-enhanced extraction framework that merges the context of real-time IoT anomalies—such as temperature, pressure, and vibration—with segment-level semantic representations (embeddings). The main focus is on providing semantic search for resource-constrained environments by using optimized INT8-quantized embedding models.
Methods: We conducted extensive benchmarks using a new corpus of 18 industrial engineering models decomposed into 51 typed segments. To assess the feasibility for implementation under resource constraints, we compared the performance and resource consumption of a full-precision model (Nomic-768D, 11.7 GB memory) against an INT8-quantized model (MiniLM-384D, 23 MB memory) on a local workstation.
Results: Experimental results show that the quantized model not only reduces memory consumption by 510 times, but also achieves +14% higher Mean Reciprocal Rank (MRR = 0.885 vs. 0.774) with a significant reduction in latency from 22,244 ms to 17 ms. Furthermore, the study confirms that under high sensor bias (loads), such as thermal anomalies, the LateReranking fusion strategy effectively recovers the accuracy (P@3) from 0.219 to 0.594.
Conclusions: These findings suggest that the combination of lightweight quantized models and context-aware reranking is a viable path to implementing reliable semantic search capabilities in industrial IoT ecosystems where hardware resources are limited.