A Novel Approach for Fault Diagnosis of Hydraulic Pitch Systems in Wind Turbines
Authors
Riisager, Kris Kristensen ; Bender, Niels Christian
Term
4. term
Education
Publication year
2016
Submitted on
2016-06-01
Pages
215
Abstract
Specialet undersøger fejldetektion og -diagnose (FDD) i et hydraulisk pitch-system ved hjælp af kunstige neurale netværk (ANN). Et hydraulisk pitch-system er et system, der bruger tryksat væske til at styre bevægelser. Der udvikles en matematisk model af systemet, og modellens parametre bestemmes ud fra eksperimenter. Modellen bruges til at udvikle og afprøve flere FDD-metoder og til at vurdere, hvilke forsøg der er nødvendige. Der bygges en testrig med systemets hovedkomponenter, som gør det muligt at efterligne lækagefejl og generelt evaluere de udviklede FDD-metoder. Hovedkonklusionen er, at en modelbaseret metode med diagnose via et focused time-delay ANN (et neuralt netværk, der udnytter tidsforskudte målinger) giver den bedste teoretiske ydeevne. Metoden afprøves på testriggen for at blive valideret. De vigtigste resultater viser, at estimater af lækager bliver mindre præcise på den fysiske opstilling sammenlignet med simulation, mens en fastklemt servoventil og en fejl i en tryktransducer (tryksensor) diagnosticeres korrekt.
This thesis explores fault detection and diagnosis (FDD) in a hydraulic pitch system using artificial neural networks (ANN). A hydraulic pitch system uses pressurized fluid to control motion. A mathematical model of the system is developed, and its parameters are identified from experiments. The model is used to design and test several FDD methods and to determine which experiments are necessary. A test rig with the main components of the pitch system is built to emulate leakage faults and to evaluate the developed FDD methods. The main conclusion is that a model-based approach with diagnosis via a focused time-delay ANN (a neural network that uses time-shifted inputs) shows the best theoretical performance. This approach is validated on the test rig. The key results show that leakage estimation becomes less accurate on the physical setup compared with simulation; however, a stuck servo valve and a pressure transducer failure are correctly diagnosed.
[This abstract was generated with the help of AI]
Keywords
Documents