Fuzzy inference and fusion for health state diagnosis of hydraulic pumps and motors

An in-line health monitoring system has been developed for hydraulic pumps that utilize feature extraction, fuzzy inference systems, and knowledge fusion. These techniques were applied to hydraulic pump diagnostics in order to create a robust system for the determination and classification of pump degradation. A thorough examination of available measurements identified three critical parameters: pump pressure, case drain flow, and case drain temperature as strong indicators of impending pump failure. Using machinery vibration analysis techniques, dynamic analysis of the high frequency "noise" content of pump pressure and case drain signals was performed to provide reliable diagnostic features. A physics-based approach was also implemented to model performance characteristics of the pump. An Automated Health Assessment and Fusion stage was then implemented to correlate changes in these diagnostic features to pump health using one of five linguistic degradation categories, determined using a Mamdani-type fuzzy logic inference architecture. Fuzzy logic-based classification was performed for each analysis approach and fused using Bayesian knowledge fusion techniques. This fusion also incorporated a priori information, in the form of historical failure rate data and operational hours, to provide a real-time health output supplemented with a confidence interval. This paper reviews the technical approach implemented for a hydraulic pump in-line health monitoring system, specifically addressing the application and utility of fuzzy logic inference in hydraulic pump health assessment. Also briefly reviewed here are details regarding the advanced feature extraction process, Bayesian knowledge fusion, and the transition of the algorithms to a prototype embedded system. Finally, results are presented using pump degradation data from accelerated endurance tests provided by The Air Force Research Laboratory (AFRL) to validate the approach. These results demonstrate a robust inference system that accurately tracks damage based on the three measurement inputs.