Data analytics and machine learning methods applied to underground coalmine roof convergence data

This paper documents a process of data analytics and machine learning applied to coalmine roof convergence data. It follows a publication documenting the formation of the database. For each set of analyses Python scripts were utilised to extract data from an SQL database, with subsequent analyses also executed in Python, utilising open-source tools such as scikit-learn. Preliminary analyses with classic machine learning algorithms included multiple logistic regression, K-nearest neighbours, decision tree, random forest, support vector machine, kernel support vector machine, artificial neural network (ANN), naïve Bayes and multiple linear regression. Results indicated that relationships were complex and non-linear, with the artificial neural network most suitable. An ANN model was then refined through a standard process of hyperparameter optimisation and data augmentation to arrive at a final model. Due to the black box nature of the ANN further insight into how the parameters interacted was sought and found with Shapley additive explanations (SHAP values). SHAP values utilise Game Theory, where each input parameter or feature is a ‘player’ while the data set is the ‘team’. The SHAP value is the impact of each player on the target value, essentially establishing the contribution of each input parameter. The results from all analyses are delivered in the paper, together with insights, applications, strengths, and limitations of the methodology.