Partial Least Square Methodology for Error Detection and Prevention (PLS/EDP), Without Knowledge of Intermediate Stages of Incipient Error
Incipient errors in agents are challenging to detect until they have reached a high magnitude, due lack of sensor training data, or models of agents, which define the intermediate stages of the incipient error. In this study, we build a methodology named Partial Least Squares for Error Detection and Prevention (PLS/EDP) to estimate the incipient error’s magnitude, without using training sensor data or models related to its intermediate states. We then extend the PLS/EDP to observe and detect the start of incipient errors, before a large magnitude of the errors are reached. The PLS/EDP methodology is based on an adaptation of Partial Least Squares (PLS) method. Distance metrics, with a fuzzy interface are used to observe trends in agents, with respect to cluster centers formed by different states of the agent. The PLS/EDP methodology is tested with two case studies. The first case study pertains to incipient wear error in gears. The second case study pertains to incipient changes in temperature and pressure in injection molding. In both the case studies, the models are built without the intermediate states of incipient error, and PLS/EDP is made to predict these intermediate states, and detect the change in state of agent, at its early stages. Further, the methodology is tested with varying amounts of noise in training data, to observe its effect on its classification and tracking abilities. PLS/EDP could predict incipient error states with a level of significance of α = 0.05. Simulations on moving incipient errors, show that the moving average algorithm can track the fuzzy memberships to clusters and detect the onset of errors, much before the full magnitude of incipient errors are reached. The distance between cluster centers Cs¬, only changes marginally with large decreases in noise, which shows that the classification and tracking abilities of PLS/EDP are only slightly affected by it.
Nof, Purdue University.
Industrial engineering|Artificial intelligence
Off-Campus Purdue Users:
To access this dissertation, please log in to our