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

Abstract

This repository contains the supplementary materials and experimental data supporting the research article: "Sensor-Augmented Retrieval of System Dynamics Model Segments for Resource-Constrained Environments". The dataset is divided into two primary components: the source model environment (search space) and the raw experimental benchmarks evaluating retrieval performance under various sensor anomaly conditions. 1. Model Corpus, Queries, and Scenarios This section contains the definitions, source code, and configurations used to build the search environment and simulate IoT telemetry. System Dynamics Models: Contains the source code and serialized JSON structures of 18 industrial engineering models covering domains such as HVAC systems, industrial boilers, chemical reactors, and wind turbines. These models integrate operational variables for Temperature, Pressure, Humidity, FlowRate, and Vibration directly into their underlying equations. Segment Decomposition: The 18 models were programmatically decomposed into 51 typed structural segments, primarily focusing on Stock-Flow subsystems and discrete feedback loops. Domain-Specific Queries: A benchmark set of 32 search queries engineered with dense technical language (e.g., "boiler steam pressure generation safety valve fuel combustion"). Each query is manually annotated to establish a deterministic ground truth, targeting exactly two to three highly specific semantic segments. Sensor Anomaly Scenarios: 10 parameterized evaluation scenarios that simulate realistic operational environments. These range from standard, stable operation (0 anomalies) to catastrophic systemic disruptions (up to 5 simultaneous anomalies). 2. Experimental Benchmarks File: edge_retrieval_full_benchmark-run1.csv … edge_retrieval_full_benchmark-run5.csv Each file is the raw execution log of one complete end-to-end pass of the evaluation matrix (3 models × 3 strategies × 10 scenarios × 32 queries = 2,880 evaluations), reported as 90 model × strategy × scenario rows. The matrix was executed five times, for 14,400 evaluations in total. It provides a comparative analysis of three ONNX configurations: a full-precision baseline (Nomic-768D), the FP32 export of the compact model (MiniLM-384D) and its INT8-quantized variant (MiniLM-384D-Int8). Including both MiniLM precisions separates the effect of quantization from that of architecture and dimensionality. It benchmarks three distinct operational context fusion strategies: LinearFusion, ContextBoost, and LateReranking. Key Finding - Resource Efficiency: INT8 quantization reduces the model file from 91 MB to 23 MB, resident memory from 100.4 ± 1.7 MB to 42.7 ± 1.0 MB (2.4×) and one-off corpus indexing from 101.0 ± 23.4 ms to 49.9 ± 13.5 ms (2.0×, the ratio holding within 1.96–2.19 across the five runs). Against Nomic-768D (548 MB on disk, 7,728 ± 54 MB resident, 23.03 ± 0.45 s indexing) the reductions are 181× in resident memory and 462× in indexing time. Key Finding - Retrieval Accuracy: Quantization carries no measurable cost in retrieval quality. Averaged per strategy, the two MiniLM precisions differ by at most 0.014 in P@3 and 0.039 in MRR, and the sign of the difference is not consistent across strategies. The advantage over the larger baseline is therefore attributable to architecture rather than precision (semantic-only MRR: 0.885 for INT8, 0.846 for FP32, 0.774 for Nomic-768D). Key Finding - Anomaly Resilience: Score-level fusion of an unnormalized context term degrades sharply as anomalies accumulate, with LinearFusion P@3 falling to 0.083 (Nomic-768D) and 0.219 (MiniLM-384D-Int8) in the AllSensors_High scenario. LateReranking holds P@3 at the semantic-only level in every scenario (0.594 and 0.604 respectively); since the candidate pool K = 3 equals the evaluation cutoff, this stability is invariant by construction, and MRR falls below the semantic-only reference. Sensor context as formulated here preserves ranking quality rather than improving it. P@3 and MRR are deterministic and bit-identical in all five runs, so the repetitions quantify variability in the timing and memory measurements alone. Note on Data Formats Please note that the configuration files (system_dynamics_models_corpus.txt, benchmark_queries_and_scenarios.txt, and benchmark_sensors.txt) are provided as raw text snippets rather than standalone, executable files. They contain the original C# data structure initializations (e.g., lists, arrays, and dictionaries) exactly as they were used in the evaluation framework's source code.