Addressing data imbalance in predictive maintenance : a performance evaluation of generative adversarial networks (GANs) for synthetic data generation
Loading...
Date
2025-10-20
Authors
Καμπουρόπουλος, Γεώργιος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Predictive maintenance (PdM) systems must infer rare failure events from predominantly healthy-operation, a setting where naive accuracy obscures risk and data scarcity undermines learning. This-thesis develops and tests a theoretical and methodological groundwork for evaluating synthetic-oversampling with generative adversarial networks (GANs) in high-imbalance PdM classification.
This study aims to construct a comparative framework for machine failure classification with the-use of 2 models, Random Forest and Support Vector Machine, under five sampling techniques:-(i) baseline with stratified cross-validation, (ii) random undersampling, (iii) SMOTE, (iv) a single-WGAN-GP synthesizing failures, and (v) a multi-WGAN-GP strategy that specializes generation-by failure subtype. The pipeline formalizes leakage-safe preprocessing (feature typing/encoding,-scaling and distributional transforms appropriate for margin-based models, and outlier handling-that preserves minority boundaries), with all fitting confined within folds. The study finds that-GAN Hyperparameter tuning and correct model evaluation play an important role and suggests-bayesian optimization solutions like Optuna’s TPE sampler, as well as the use of PR-AUC as the-primary metric for imbalanced classification. Complement metrics like ROC-AUC recall at fixed-false-positive rates may also be used to reflect operational trade-offs in PdM. The study examines-the types of GAN variations with the aim of with the aim of avoiding the usual training problems,-mode collapse and vanishing gradients, and articulates when subtype-specific generators can better-capture heterogeneous failure modes. Finally, the use of t-SNE visual diagnostics is examined to-assess manifold coverage and boundary fidelity of synthetic samples. The findings of this study-suggest that type-sensitive GAN oversampling can be a great solution for synthetic oversampling-in PdM.
Description
Keywords
Predictive maintenance, Imbalanced classification, Synthetic oversampling, WGAN- GP, Imbalanced performance evaluation