Scholarlink

<p>Large-scale sequential datasets in transport, safety, and environmental monitoring are commonly analysed using forecasting, classification, anomaly detection, and machine-learning workflows. These approaches are designed to predict future states, identify deviations from learned background distributions, or model changing system conditions over time. A prior question, however, often remains under-specified: whether pooled analysis is already structurally misleading because observations generated under heterogeneous regimes have been merged into a composite reference profile that does not correspond to any single operational or physical state. In this paper, anomaly detection is treated as a constrained detection/classification problem whose validity depends on the coherence of the reference surface used to define unusualness. This study introduces and evaluates an operator-level, feasibility-first framework based on admissible-state reasoning. The framework examines whether pooled summaries obscure regime-conditioned feasible structure and whether pooled-versus-regime distortion can be summarised in a portable and auditable way using the Pooling Artefact Index (PAI). It is applied across three public demonstrators: U.S. airline delay causes (2003–2022; 318,017 month-airport-carrier records), urban traffic crashes (2013–2025; 209,306 events), and global near-real-time weather observations (2024–2026; 119,847 records). Regimes are defined using transparent observable structure, and empirical admissible envelopes are estimated within each regime before comparison with pooled reference envelopes. The framework is implemented as a pre-modelling diagnostic: it declares the relevant clocks and regimes, estimates pooled and regime-conditioned envelopes, quantifies their divergence using PAI, and then determines whether downstream prediction, monitoring, or anomaly assessment should proceed on pooled or regime-specific reference surfaces. Across domains, regime conditioning yields tighter and more interpretable feasible envelopes, whereas pooled summaries generate composite envelopes that do not correspond cleanly to any single operational or physical state. Median PAI was 0.24 for airlines, 0.06 for traffic, and 5.56 for weather under the continuous-envelope variant. The airline and traffic demonstrators provide the primary compositional cases, whereas the weather demonstrator is interpreted more cautiously as a secondary continuous-envelope illustration. These findings support admissible-state reasoning as a portable pre-analysis framework for heterogeneous sequential systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Admissible-state reasoning across public operational and physical systems: regime-conditioned feasibility and the pooling artefact index

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月28日

查看全文
  • Adegoke Adefolalu

摘要

Large-scale sequential datasets in transport, safety, and environmental monitoring are commonly analysed using forecasting, classification, anomaly detection, and machine-learning workflows. These approaches are designed to predict future states, identify deviations from learned background distributions, or model changing system conditions over time. A prior question, however, often remains under-specified: whether pooled analysis is already structurally misleading because observations generated under heterogeneous regimes have been merged into a composite reference profile that does not correspond to any single operational or physical state. In this paper, anomaly detection is treated as a constrained detection/classification problem whose validity depends on the coherence of the reference surface used to define unusualness. This study introduces and evaluates an operator-level, feasibility-first framework based on admissible-state reasoning. The framework examines whether pooled summaries obscure regime-conditioned feasible structure and whether pooled-versus-regime distortion can be summarised in a portable and auditable way using the Pooling Artefact Index (PAI). It is applied across three public demonstrators: U.S. airline delay causes (2003–2022; 318,017 month-airport-carrier records), urban traffic crashes (2013–2025; 209,306 events), and global near-real-time weather observations (2024–2026; 119,847 records). Regimes are defined using transparent observable structure, and empirical admissible envelopes are estimated within each regime before comparison with pooled reference envelopes. The framework is implemented as a pre-modelling diagnostic: it declares the relevant clocks and regimes, estimates pooled and regime-conditioned envelopes, quantifies their divergence using PAI, and then determines whether downstream prediction, monitoring, or anomaly assessment should proceed on pooled or regime-specific reference surfaces. Across domains, regime conditioning yields tighter and more interpretable feasible envelopes, whereas pooled summaries generate composite envelopes that do not correspond cleanly to any single operational or physical state. Median PAI was 0.24 for airlines, 0.06 for traffic, and 5.56 for weather under the continuous-envelope variant. The airline and traffic demonstrators provide the primary compositional cases, whereas the weather demonstrator is interpreted more cautiously as a secondary continuous-envelope illustration. These findings support admissible-state reasoning as a portable pre-analysis framework for heterogeneous sequential systems.