Data-Driven Vibration-Based Condition Monitoring: Fundamentals, Applications, and Challenges
Abstract
Vibration-Based Condition Monitoring (VBCM) is commonly utilized in Prognostics and Health Management (PHM) due to its non-destructive nature and inherent advantages over alternative forms of condition monitoring. Furthermore, the rapid evolution of sensor fabrication and the rise of the Internet of Things (IoT) have facilitated large-scale VBCM systems across diverse domains, including industry, transportation, healthcare, agriculture, and wildlife monitoring. The recent advancements in computing technologies have significantly expanded the potential for VBCM by leveraging the synergy between signal processing and Machine Learning (ML). Accordingly, data-driven VBCM has emerged as a paradigm shift, improving the performance and reliability of VBCM systems. To this end, addressing various attributes of data-driven VBCM becomes increasingly important since it represents the core of current and future VBCM systems. The work presented in this thesis addresses the main aspects of VBCM, including signal processing fundamentals, feature extraction, availability of labeled data, computational complexity, and power efficiency. The methods employed in this thesis span the fields of Digital Signal Processing (DSP) and ML techniques (supervised, Deep Learning (DL)), including signal preprocessing, signal denoising, signal frequency-domain analysis, signal time-frequency domain analysis, feature extraction, signal companding (compression-expansion), and 1-dimensional (1D) convolutional reconstruction autoencoders. These methods address extraction of effective condition-related features, limited availability of labeled data, noise removal, complexity considerations in VBCM systems, and power efficiency of power-constrained sensor nodes in remote VBCM. By addressing the aforementioned problems, the end-to-end performance of VBCM systems can be improved in terms of the size of training data, the reliability of the monitoring process, system delay, memory requirements, and power consumption. To ensure the explainability of the extracted features, the developed methods for the extraction of condition-related features are based on signal processing since feature engineering using signal processing creates explainable features that link meaningfully to signal conditions or classes compared to DL-based features. The thesis also contributes to the VBCM literature by providing a comprehensive tutorial on signal processing fundamentals, an overview of a typical signal-based ML pipeline, and an application-independent review of feature extraction techniques. The work presented in this thesis presents efficient solutions to the main challenges that face the practical deployment of real-world VBCM systems