A Neural Network Approach for Gas Leakage Detection in Fluid Power Accumulators of Wind Turbines
Author
Bartz Rafaeli Neto, Denis
Term
4. term
Education
Publication year
2023
Abstract
Formålet med dette projekt er at forbedre vedligeholdelsen af offshore-vindmøller og reducere nedetid ved at opdage gaslækager i hydrauliske akkumulatorer. En hydraulisk akkumulator er en trykbeholder, der lagrer energi i olie ved hjælp af en gas; hvis gassen lækker, falder systemets ydelse. Arbejdet starter med at udvikle og validere en nøjagtig model af akkumulatoren baseret på data, der dækker forskellige driftsforhold. Modellen valideres med eksperimentelle målinger fra Hydrauliklaboratoriet på AAU. Dernæst udvikles en Fully Convolutional Network (FCN), en type neuralt netværk der kan genkende mønstre i tidsserier, til at opdage gaslækager ved at klassificere indgangssignaler og bestemme den tilhørende gas fortryk (precharge pressure). FCN’en trænes på tidsseriedata og vurderes på sin ydeevne. Studiet undersøger, hvordan valg af indgangsvariable, størrelse af det glidende vindue, hyperparametre og brugen af forskellige sensorer påvirker resultaterne. De eksperimentelle resultater viser, at når olietryk samt olie- og omgivelsestemperatur indgår, opnås 95% nøjagtighed i klassificeringen af fortrykket. Når der desuden monteres termoelementer på akkumulatorens overflade, forbedres ydeevnen betydeligt og når 100% nøjagtighed. Disse resultater peger på, at målrettet brug af sensorer og neurale netværk kan støtte bedre vedligeholdelsesbeslutninger og reducere nedetid.
This thesis aims to improve maintenance and reduce downtime in offshore wind turbines by detecting gas leaks in hydraulic accumulators. A hydraulic accumulator is a pressure vessel that stores energy in oil using a gas; if the gas leaks, system performance declines. The work begins by developing and validating an accurate accumulator model using data that cover a range of operating conditions. The model is validated with experimental measurements from the Hydraulics Laboratory at AAU. Next, a Fully Convolutional Network (FCN)—a neural network that learns patterns in time-series signals—is developed to detect gas leakage by classifying input signals and estimating the corresponding gas pre-charge pressure. The FCN is trained on time-series data and its performance is evaluated. The study examines how input variables, sliding window size, hyperparameters, and different sensor setups affect the results. Experiments show that including oil pressure as well as oil and ambient temperature yields 95% accuracy when classifying the pre-charge pressure. Adding thermocouples to the accumulator’s surface further improves performance to 100% accuracy. These findings indicate that targeted sensor use combined with neural networks can support better maintenance decisions and help minimize downtime.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Keywords
Documents