Διδακτορικές Διατριβές

Μόνιμο URI για αυτήν τη συλλογήhttps://repository.ihu.gr/handle/11544/47648

Περιηγούμαι

Πλοήγηση Διδακτορικές Διατριβές ανά Θέμα "Neural Networks"

A B C D E F G H I J K L M N O P Q R S T U V W X Y Z
Α Β Γ Δ Ε Ζ Η Θ Ι Κ Λ Μ Ν Ξ Ο Π Ρ Σ Τ Υ Φ Χ Ψ Ω
Φιλτράρετε τα αποτελέσματα πληκτρολογώντας τα πρώτα γράμματα
Τώρα δείχνει 1 - 1 από 1
  • Μικρογραφία εικόνας
    Τεκμήριο
    Machine Learning based methods for monitoring and control in modern Industry 4.0 systems
    (ΔΙΠΑΕ, 2026-06-10) Gravanis, Georgios; Diamantaras, Konstantinos; Papadopoulou, Simira; Salampasis, Michail; Σχολή Μηχανικών, Τμήμα Μηχανικών Πληροφορικής και Ηλεκτρονικών Συστημάτων
    This thesis presents the results of a systematic research conducted about the integration of Machine Learning (ML) methods in modern Industry 4.0 systems. The work is motivated by the increasing availability of industrial operational data and by the limitations of conventional approaches when dealing with nonlinear dynamics, scarce fault labels, Machine learning black-box behavior, and in general by the benefits of fast online decision support in industrial environments. More specifically, within this thesis is explored how ML can advance five complementary directions of modern Industry 4.0 systems i.e. Supervised Fault Detection and Diagnosis (FDD) in nonlinear processes, explainability of FDD model decisions, unsupervised anomaly detection in manufacturing systems, ML-augmented advanced control, and surrogate property modeling enabling digital twin development. For supervised fault detection and diagnosis, dynamic neural-network models based on Long Short Term Memory (LSTM) and Time Delay Neural Network (TDNN) architectures are combined with dimensionality-reduction techniques and evaluated on the Tennessee Eastman Process (TEP) benchmark problem. The best configurations achieve approximately 98.5% average classification accuracy, demonstrating strong capability for real-time fault recognition in nonlinear industrial processes. To improve classifier transparency, ML explainability methods i.e. Integrated Gradients and SHAP are applied to the FDD framework, showing that, for most disturbances, the explanations generated are consistent providing physically meaningful insights from a process-engineering perspective. The thesis further proposes an unsupervised anomaly-detection framework for manufacturing machinery based on k-nearest-neighbor outlier detection and correlation with a Key Quality Indicator. Applied to an industrial yogurt packaging machine, the method provides guidance for identifying hidden malfunctions and ranking the most probable root-cause machinery subsystems. Except for supervised and unsupervised fault and anomaly detection, it is explored whether ML can be part of advanced control methods. To this extend, a ML model is embedded in a nonlinear model predictive control algorithm for the energy management in a multi-node microgrid. In this case TDNN based state-of-charge predictors replace mechanistic models and reduce cumulative energy transfer between nodes by about 6.5% on average while remaining feasible for real-time implementation. Finally, it is investigated whether ML can be utilized to create surrogate models capturing thermophysical properties for Digital Twin development. In this thesis Autoencoder and Support Vector Regressor models are developed for the prediction of CO2 and CH4 diffusivities in liquid water, using experimental and molecular dynamics simulated supported datasets. The results show that the proposed models capture the nonlinear dependence of diffusivity on pressure and temperature, providing accurate predictions across broad operating conditions. Overall, the current thesis demonstrates that ML can enhance modern Industry 4.0 systems in a trustworthy, scalable, and application-oriented manner.