Making the black-box brighter: Interpreting machine learning algorithm for forecasting drilling accidents

We present an approach for interpreting a black-box alarming system forecasting accidents and anomalies during the drilling of oil gas wells. The interpretation methodology aims to explain local behavior accident predictive model engineers. explanatory uses Shapley additive explanations analysis features, obtained through Bag-of-features representation telemetry logs used phase. Validation shows that has 15% precision at 70% recall, overcomes metric values random baseline multi-head attention neural network. These results justify developed is better aligned with engineers, than state-of-the-art method. joint performance models allows engineers understand logic behind decisions particular moment, pay highlighted regions, correspondingly, increase trust level in alarms.