Gas leakage detection in hydraulic accumulator - A neural network approach -
Author
Racca, Alessandro
Term
4. semester
Education
Publication year
2019
Submitted on
2019-05-31
Abstract
This thesis investigates model-based detection of gas leakage in a gas-filled hydraulic accumulator within a wind turbine pitch system using a Nonlinear AutoRegressive neural network with exogenous input (NARX). To obtain supervised training data, a physics-based accumulator model is used to simulate the process; the network is trained in open loop with load flow rate as input and supply pressure as output to represent healthy behavior. As part of model-based fault detection and diagnosis, the healthy NARX model is then driven by faulty input-output data to generate residuals. Results show that the open-loop network tracks the faulty process too closely, predicting the actual output instead of the healthy output, and is therefore ineffective for detection; a closed-loop NARX configuration is concluded to be preferable for fault detection. The study outlines a neural-network-based, simulation-driven approach aimed at reducing downtime and enhancing safety in hydraulic pitch systems.
Dette speciale undersøger modelbaseret detektion af gaslækage i en gasfyldt hydraulisk akkumulator i en vindmølles pitchsystem ved hjælp af et Nonlinear AutoRegressive neural network with exogenous input (NARX). For at skaffe træningsdata til den superviserede læring anvendes en akkumulatormodel til at simulere processen; netværket trænes i åben sløjfe med lastflow som input og forsyningstryk som output for at beskrive et sundt system. Som led i modelbaseret fejldetektion og diagnose fodres den sunde NARX-model efterfølgende med fejlbehæftede input-output-data for at generere residualer. Resultaterne viser, at åben-sløjfe-netværket er så præcist, at det følger den fejlbehæftede proces og forudsiger det faktiske output i stedet for det sunde output, hvilket gør det uegnet til detektion; derfor vurderes en lukket-sløjfe NARX-konfiguration som bedre egnet til fejldetektion. Arbejdet peger på en neuralt netværksbaseret, simuleringsdrevet tilgang, der kan reducere nedetid og forbedre sikkerheden i hydrauliske pitchsystemer.
[This apstract has been generated with the help of AI directly from the project full text]