Rotary-percussive drill bit condition prediction using traditional feature engineering and neural network-based feature extraction

Abstract

Condition monitoring of replaceable components in underground drill rigs using machine learning is a difficult task, as the operating conditions may vary considerably between each hole. To model this nuanced data with acceptable performance, feature extraction must be performed, either by field experts or using automated machine learning architectures. This work compares the use of traditional feature extraction techniques to neural network-based automatic feature extraction for a rotary-percussive underground hydraulic drill rig. A dataset was created for the purpose of predicting the condition of drill bits with tungsten carbide button inserts, and consists of both operational pressures, as well as signals from an accelerometer and a microphone. Two feature extraction approaches are compared using data collected under controlled operating conditions. The first approach uses traditional features including kurtosis, FFT features, and wavelet features. Feature selection and bit condition prediction are performed using a Random Forest model. The second approach uses neural networks to automatically extract features from raw data. Convolutional neural networks and long short-term memory networks are used in the automatic feature extraction approach. The traditional feature extraction approach is sufficient for binary classification of bit condition, while the automatic neural network-based feature extraction approach is superior when prediction is scaled up to a more complex multi-class problem.