Machine Learning for Time Series Anomaly Detection of Haul Truck Subsystems and Driver Behaviour

Thumbnail Image
Nicholson, Shelby
Mining , Haul truck , Machine learning
This thesis investigates machine learning techniques for anomaly detection in the operation of ultra-class mining haul trucks. Data was collected from three different models of haul truck, and two data sets were created. The first data set, focused on one of the three truck models, included the entirety of the 175 parameters accessible via the truck’s proprietary (OEM) monitoring system, sampled at 0.2 Hz, over some 11 hours of operation, and was gathered while several predefined truck maneuvers were performed. These maneuvers included aggressive (abrupt turns, large accelerations) and drowsy (slow veering, then sharp correction) driving behaviours. Subsystem anomaly detection on this first data set successfully applied Decision Tree and Random Forest Classifiers for detecting the anomaly labelled by the “High Brake Hydraulic Temperature” parameter but yielded poor performance detecting the anomaly labelled by the “Reduced Retard Level Switch” parameter. The second data set was from an Inertial Measurement Unit (IMU) (3-axis accelerations and 3-axis rotations), recording truck dynamics for all three truck models during numerous maneuvers. Driver behaviour anomaly detection on the second set successfully applied 1-Nearest Neighbour Classifiers and Dynamic Time Warping in recognizing left and right aggressive and drowsy turns. Attempts at unsupervised classification with Agglomerative Clustering on this second data set yielded mixed results. Overall, encouraging performance from this investigation indicates this is a fruitful area for further research on the detection of anomalies in mobile mining equipment.
External DOI