Application of Vibration-Based Machine Learning Model for Fault Diagnosis in Rotating Machines

Abstract

Reliable operation of rotating machinery is essential for industrial process efficiency. While vibration-based condition monitoring (VCM) is the standard for fault detection, it typically requires specialist knowledge and manual interpretation that can vary between analysts. This thesis addresses this limitation by developing an automated, robust vibration-based machine learning (VML) framework capable of diagnosing both rotor and bearing faults with minimal intervention. The research first extends an artificial neural network (ANN) model using optimised time- and frequency-domain vibration parameters to classify rotor faults (misalignment, shaft cracks, rub) and anti-friction bearing defects. Experimental validation was conducted on a multi-rotor test rig across three distinct operating speeds: 450 RPM (below the first critical speed), 900 RPM (between critical speeds), and 1350 RPM (above the second critical speed). The model demonstrated 100% classification accuracy for all fault types under these varying dynamic conditions. To facilitate industrial application, a hierarchical three-step diagnostic methodology is developed. Using a unified ANN configuration and feature set, the system sequentially performs: (1) Fault Detection (classifying machine state as healthy or faulty), (2) Fault Diagnosis (identifying the specific defect type), and (3) Fault Location (pinpointing the defect's position). This structured approach streamlines maintenance decision-making and is validated with 100% accuracy in detection, diagnosis, and location across all tested speeds. Furthermore, the study integrates the poly-coherent composite spectrum (pCCS) technique to fuse multi-sensor vibration data. This approach preserves amplitude and phase information while reducing noise, allowing for a significant reduction of input features: specifically, from 24 parameters across four sensors to a compact set of eight standardised parameters. Finally, the methodology's generalisability is proven on a second, mechanically distinct test rig featuring a split roller bearing. The standardised parameters and identical ANN model successfully transferred to this new configuration, accurately identifying healthy, misalignment, unbalance, and inner-race bearing fault conditions with 100% accuracy at speeds of 12 Hz and 18 Hz, and 96.7% accuracy for unbalance detection at 6 Hz. This confirms the framework's robustness and its potential as a universal diagnostic tool for diverse rotating machinery.