Features extraction and selection fori ndustrial systems condition monitoring

Abstract

This thesis presents the development of advanced feature extraction and selection techniques for condition monitoring and fault diagnosis in industrial systems, with a particular focus on rotating machinery and oil & gas infrastructures. The primary contribution lies in proposing a novel wrapper-based feature selection criterion that jointly optimizes overall classification accuracy, class-specific accuracy, and model stability. Two real-world experimental applications were considered to validate the proposed methodology. The first involved a benchmark database of vibration and current signals acquired from bearing fault scenarios in synchronous motors, where an Adaptive Time-Varying Morphological Filtering (ATVMF) technique was applied for advanced signal pre-processing and feature extraction. The second application involved a comprehensive 3W dataset comprising temporal measurements collected via industrial sensors from oil and gas wells, where data normalization between 0 and 1 was implemented to enhance data consistency prior to analysis. In both cases, feature selection was performed using hybrid metaheuristic optimization algorithms combined with machine learning classifiers such as Random Forest. The experimental results demonstrated that integrating ATVMF with the proposed selection criterion notably improves fault detection and classification performance in multi-class and variable-condition scenarios. Furthermore, the framework ensures robustness and computational efficiency, making it suitable for real-time industrial monitoring systems