Multi-State Recognition Method for Coal Mine Ventilation Systems via Abductive Learning
TL;DR
An abductive learning-based framework for multi-state recognition is proposed, integrating data-driven modeling with domain knowledge-guided reasoning in a unified cross-disciplinary approach, significantly outperforming traditional signal-processing methods, classical machine learning, and state-of-the-art multivariate time-series anomaly detectors.
Abstract
Coal mine ventilation systems are critical to underground safety, whose operational status directly affects the health of personnel and the stability of production. Anomaly detection in such systems faces severe class imbalance, transient switching patterns that mimic faults, and insufficient utilization of domain knowledge. To bridge data-driven learning and domain reasoning, this paper proposes an abductive learning-based framework for multi-state recognition, integrating data-driven modeling with domain knowledge-guided reasoning in a unified cross-disciplinary approach. A symmetric autoencoder extracts normal patterns from multi-sensor data, with reconstruction errors statistically modeled via a multivariate gaussian distribution to quantify anomaly likelihood. An abductive label correction mechanism incorporates hierarchical domain rules to refine pseudo-labels, mitigating bias from scarce fault samples. Additionally, a vibration-based switching-state recognition module identifies transitional operations using sliding-window statistical features, reducing false alarms. The proposed multi-state recognition method was experimentally validated using real-world operational data collected from mine ventilators. In the three-class classification task, the F1 scores for the normal, abnormal, and switching states reached 0.9620, 0.9431, and 0.8562, respectively, significantly outperforming traditional signal-processing methods, classical machine learning, and state-of-the-art multivariate time-series anomaly detectors (e.g., Anomaly Transformer, TranAD). Comparative experiments with changepoint detection methods and generalization tests on public bearing fault datasets further confirm the robustness and generalization capability of the proposed method, indicating its practical applicability in real industrial scenarios. Nevertheless, the current framework still relies on inspection records and maintenance experience for logical label generation and switching-threshold determination, which may limit its direct application to equipment without sufficient domain knowledge.