Intelligent Interpretation of Machine Condition Data

Abstract

This dissertation argues that classification is an effective tool in the prediction of machine condition. A system based on continuous learning can be developed to automate the laborious process of interpretation of symptoms derived from collected data into various normal and fault modes. In order to defend these arguments, the study seeks to explain, how prediction works and how the results can be evaluated. The study explores the philosophy of condition monitoring in assuring the safe and uninterrupted operation of machines. Condition monitoring provides essential information for the maintenance and operability of process plants. Vibration monitoring is considered as one of the most important techniques to offer adequate and reliable information to maintain rotating machines in a condition, where they can perform their required functions without failure for a specified time period, when used under specified conditions. In addition of detection and collection of data that indicate the state of a machine, condition monitoring includes the examination of symptoms and syndromes to determine the nature of faults or failures. High confidence level is required in both diagnostics and prognostics, because misinterpretation of condition related data may lead into severe economic consequences. The diagnostics of machine condition is laborious and challenging. It requires a lot of analyst’s effort and time to detect an anomaly and even more to identify a fault mode. This study presents methods to predict the current condition of machines using training data collected from various normal and fault modes on the same machine or substantially similar machines. Learning algorithms offer possibilities to increase the confidence level of prediction. The study presents results on practical experiments to demonstrate the principles of continuous learning processes. The experiments rely on data collected from wind turbine gearboxes, which are extremely difficult to be diagnosed, because of the large amount of data and symptoms. The study proofs that significant improvements to current confidence level of prediction can be achieved by the use of learning systems