AI-based pre-alarm framework for early detection of abnormal gas-phase dynamics
Year: 2026
Type: article
Abstract: Autonomous underground production removes the people who traditionally noticed early signs of a hazard such as a machine fire, so on-board gas sensing has to assume part of that monitoring role. This paper asks whether the carbon monoxide (CO) channel on a production loader shows a precursor before the machine raises its own gas-sensor fault fag. Pre-fault detection is posed as a causal supervised classification task with a 48 h horizon, a 21-dimensional feature vector of rolling statistics and percentile-exceedance counts, and a random forest (RF) classifier, evaluated on 2,477 cleaned samples spanning 309 days of one loader’s register log, in which 18 fault events occur. Fitted and scored on the same record, the classifier separates the classes perfectly, with 100.00% accuracy and ROC AUC = 1.000, and all 16 observable event windows contain at least one threshold crossing; the median warning time among those windows is 27.67 h. A control given only the calendar date reproduces that result exactly, showing that the perfect in-sample score certifies temporal localisation rather than gas-based predictive skill. All 18 events fall inside a single 29.1-day period at the end of the record. The preselected chronological split at 31 October therefore contains no positive training examples and cannot support two-class supervised discrimination. Under leave-cluster-out cross-cluster validation with a 48 h embargo, ROC AUC falls to 0.5055 (block-label permutation = 0.46), accuracy is 84.70% compared with 85.26% for a constant-negative predictor, and 10 of 16 observable event windows contain at least one threshold crossing at η= 0.65.
Source: IFAC-PapersOnLine
Authors: Madeleine Martinsen, Mehrdad Mesgarpour
Institution: Mälardalen University
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7224079271