Self-Organized Predictor of Methane Concentration Warnings in Coal Mines

Abstract

Coal mining operation continuously balances the trade-off between the mining productivity and the risk of hazards like methane explosion. Dangerous methane concentration is normally a result of increased cutter loader workload and leads to a costly operation shutdown until the increased concentrations abate.

We propose a simple yet very robust methane warning prediction model that can forecast imminent high methane concentrations at least 3 minutes in advance, thereby giving enough notice to slow the mining operation, prevent methane warning and avoid costly shutdowns.

Our model is in fact an instance of the generic prediction framework able to rapidly compose a predictor of any future events upon the aligned time series big data. The model uses fast greedy backward-forward search applied subsequently upon the design choices of the machine learning model from the data granularity, feature selection, filtering and transformation up to the selection of the predictor, its configuration and complexity.

We have applied such framework to the methane concentration warning prediction in real coal mines as a part of the IJCRS’2015 data mining competition and scored $3^{rd}$ place with the performance just under 85 %. Our top model emerged as a result of the rapid filtering through the large amount of sensors time series and eventually used only the latest 1 minute of aggregated data from just few sensors and the logistic regression predictor. Many other model setups harnessing multiple linear regression, decision trees, naive Bayes or support vector machine predictors on slightly altered feature sets returned nearly equally good performance.