Monitoring of valve condition in reciprocating pumps by analysis of vibration signatures
Abstract
The purpose of the research was to investigate the connection between machinery condition and vibration signatures for the check valves in a set of reciprocating pumps. A set of vibration time histories were recorded from pumps used to load the accumulator of a high pressure water system, powering aluminum extrusion presses. Vibration signal features, such as rms values and peak amplitudes, and signal models, such as autoregressive and Prony series models, are used to extract information from the signals. Statistical pattern recognition methods and artificial neural nets are employed to determine which of the resulting set of signal parameters are most indicative of pump condition. The best combination of classifier and signal parameters are then chosen. This research shows, for the pumps under investigation: (1) condition information is readily interpreted within the frequency domain. Even the recording from one individual impact can be properly classified with less than 7% likelihood of error. (2) k-Nearest neighbor classifiers are superior in this application to any other classifier tested. (3) Prony models may be used with little loss of classification ability when compared with raw data, and have superior performance when compared to AR model coefficients; this indicates that the frequency information, resulting from the AR decomposition, is more sensitive to pump condition than the AR coefficients themselves. (4) As few as 5 frequency coefficients are needed for effective classification.
Degree
Ph.D.
Advisors
Davies, Purdue University.
Subject Area
Mechanical engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.