A comparative study of different algorithms using contrived failure data to detect robot anomalies

Abstract

Unexpected downtime remains a costly, preventable burden in manufacturing. To mitigate and eliminate unexpected downtime, manufacturers have incorporated machine learning to diagnose equipment faults and determine the equipment remaining useful life. However, such models suffer from a lack of failure data and context knowledge surrounding data gathered from production (i.e. rich labeled sets). The purpose of this paper is to conduct an algorithm comparison study over a previously collected contrived anomaly data set from an industrial robot. The goal of the study is to measure the effectiveness of algorithms to eliminate bias and variance from the classification results. The tested system is a 6-DOF collaborative robot from Universal Robots, on which an anomalous condition is artificially induced on a robot to simulate robot overload. The different algorithms are assessed based on their accuracy in determining the overloading case on the robot. From the analysis of data-driven machine learning and deep learning models, a deep learning regression was determined as the best model from the assessment of the data both qualitatively (low overfitting and bias) and quantitatively (98% overall accuracy). As part of the future research direction, different failure cases should be created based on wear applied to specific robot joint components and an adaptability assessment carried out for the model to other robot paths.