Signatures of criticality in mining accidents and recurrent neural network forecasting model

Abstract

We report signatures of criticality in mining accident data obtained from the Mine Accident, Injury and Illness Report form (MSHA Form 7000-1). This work builds on the hypothesis that workplace accident statistics follow self-organized criticality (Mauro et al., 2018). “1/f noise,” a distinct feature of critical systems, is extracted from this database and is used to forecast accident trends using a long short-term memory (LSTM) recurrent neural network (RNN). The algorithm used for extracting this noise is applicable to data available in any standard worker’s compensation database. We also report a Pareto distribution in the number of accidents in relation to employee mine experience, implying a strong correlation between experience and susceptibility to accidents.