Training data selection criteria for detecting failures in industrial robots

We study the e↵ect of source and type of training data on detecting failures in industrial robots using Principal Component Analysis (PCA). Specifically, using field data across multiple robots performing di↵erent tasks, we compare two scenarios: first, where training data obtained from a single robot is used to evaluate multiple robots (one-to-many), and second, where each robot is evaluated on the basis of its own training data (one-to-one). We further investigate if the data preprocessing prior to running PCA a↵ects the ability to detect and predict failures. To reduce task dependence of the raw signal, we preprocess the same by computing the absolute di↵erence between successive measurements and compare the results with a PCA model that is built using raw signal alone and another that is built from a combined signal having both raw measurements and their absolute di↵erence. We quantify e↵ectiveness of detecting failures in terms of three measures: coe cient of variation of the Q-residual obtained by projecting the test data on the PCA model, number of samples above a data-driven confidence threshold, and lead time, measured as the number of days prior to failure when the residual error rises above a given threshold. Specifically, we show that while both one-to-one and oneto-many training sources are valid for detecting failures, signal preprocessing has a significant influence. Our results show that coe cient of variation of the Q-residual from a PCA model built using absolute di↵erence between measurements serves as a robust descriptor for predicting and detecting failure in robots in the one-to-many training scenario. With the same signal, when using number of samples above threshold, we find that one-to-one training source is able to detect failure in robots. Finally, with lead time, we find that one-to-one training scenario with absolute di↵erence as signal type can be used to raise warning as early as nineteen days before failure.