Machine Learning for Shovel Tooth Failure Detection

This was originally published in September 2019

Abstract

Current computer vision-based methods for identifying broken teeth on mining shovels suffer from a prohibitively high false positive rate (FPR) of 25%. We describe a 2-stage methodology for the detection of broken teeth that reduces the FPR to 5%. First, we used a Haar wavelet feature cascade based on the Viola-Jones object detection framework to detect the row of shovel teeth from the input image. The second stage is a classification step that takes the detections from stage 1 as input and produces a binary score indicating whether the equipment is intact or damaged. We evaluated two methods for stage 2: 1) Dynamic Time Warping with k-Nearest Neighbors (DTW—k-NN) and 2) Convolutional Neural Network (CNN). The accuracies of the two methods on an out-of-sample image set were 96.3 and 95.5%, respectively.