Doktora Tezleri

Permanent URI for this collectionhttps://hdl.handle.net/20.500.14365/8833

Browse

Browsing Doktora Tezleri by Subject "Alan Uyarlaması"

Filter results by typing the first few letters
Now showing 1 - 1 of 1
  • Thumbnail Image
    Doctoral Thesis
    Dönen elektrik makinelerinde arıza teşhisi ve denetimsiz alan uyarlaması için padé yaklaşımlı sinir ağları
    (2026) Kılıçkaya, Sertaç; Eren, Levent
    Fault diagnosis is a core component of condition monitoring in industrial machinery, where data-driven methods are increasingly used to address complex dynamics and varying operating conditions. Although Convolutional Neural Networks (CNNs) perform well under controlled settings, their effectiveness often depends on deep architectures and degrades under varying operating conditions. To address these challenges, this dissertation proposes Padé Approximant Neural Networks (PadéNets), inspired by Padé rational approximants, for supervised fault diagnosis and unsupervised domain-adaptive learning. Using constant-speed vibration and acoustic datasets from the University of Ottawa, one-dimensional PadéNets operating directly on raw waveforms consistently outperformed CNN and Self-Organized Operational Neural Network (Self-ONN) baselines across accelerometer and microphone channels. Classification accuracies of 99.96 %, 98.26 %, and 97.61 % were achieved for accelerometers located from closest to farthest from the drive end, while 98.33 % accuracy was obtained for the microphone channel, using compact architectures with approximately 145 K trainable parameters. To address domain shifts caused by load and speed variations, PadéNet feature extractors were also integrated into Deep CORAL, Domain-Adversarial Neural Network (DANN), and Conditional Domain-Adversarial Network (CDAN) frameworks. In cross-load transfer experiments on the CWRU dataset using raw vibration signals and in constant-to-variable speed transfer experiments on the University of Ottawa dataset using log-Mel spectrograms, CDAN with PadéNet encoders achieved target-domain accuracies of up to 99.28 % and 97.06 %, respectively. Overall, the results demonstrate that PadéNet-based architectures provide an effective balance between accuracy, robustness, and computational efficiency.