Feature Design, Selection, and Model Optimization for Fault Classification in a Hydraulic Rock Drill

Abstract

Early detection and accurate identification of faults in equipment and systems enables maintenance professionals to operate more efficiently, saving resources and time. This work presents two data-driven methodologies for classifying fault conditions in a hydraulic reciprocating rock drill. Labeled fault data was collected from a test cell where 11 different operating conditions, including 10 faulty conditions, were recorded. Three hydraulic pressure signals were measured at 50khz from different locations in the rock drill. The presented solutions both leverage a Wavelet Scattering Network to extract features from the pressure signals. The first method classifies the faults by voting. The output from two optimized Neural Networks trained using different methods for combining the scattering features, global pooling and median, were merged to produce the 99.04% accuracy which was submitted 2022 PHM Conference Data Challenge. The second method is presented as an improvement and leverages a minimum redundance maximum relevance technique to remove confounding features by ranking them against the operator in addition to the faults. Linear Discriminant Analysis (LDA) combined the remaining features to maximize the distance between classes, and a single optimized Neural Network was used to perform the classification.