摘要：AbstractWe study the effect 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 different 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 affects the ability to detect and predict failures. To reduce task dependence of the raw signal, we preprocess the same by computing the absolute difference 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 difference. We quantify effectiveness of detecting failures in terms of three measures: coefficient 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 one-to-many training sources are valid for detecting failures, signal preprocessing has a significant influence. Our results show that coefficient of variation of the Q-residual from a PCA model built using absolute difference 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 difference as signal type can be used to raise warning as early as nineteen days before failure.
