MECHANICAL FAULT ANALYSIS AND DETECTION USING NON STATIONARY DECOMPOSITION FOR VIBRATION SIGNALS

Abstract

Machine health monitoring plays an increasingly crucial role in automated industries, particularly in the context of meeting Industry 4.0 standards. One significant aspect is the detection and diagnosis of faults in rotating machines by implementing continuous machine health monitoring systems. These systems can proactively detect and classify issues related to rotating elements in real-time, allowing for timely maintenance and repairs. Bearing faults and shaft imbalances are common problems that accounts for 50% of motor failures. This can significantly impact machine performance and lead to premature failures. Through continuous monitoring and analysis of vibration patterns, temperature fluctuations, stator current, acoustic noise or any other relevant parameters, an early sign of bearing faults and shaft imbalances can be identified. This will enable timely corrective actions to prevent catastrophic assembly line failures. Integrating machine health monitoring with advanced analytics and predictive maintenance algorithms, can help achieve higher levels of efficiency, productivity, and cost savings by minimizing unplanned downtime and extending the lifespan of critical machinery. There has been significant contribution in this field but a major challenge remains in terms of fault detection and severity identification under varying load and rotational speed. The changing speed impacts frequency content and pattern changing the fault characteristic frequency which hinders consistent fault detection. To overcome this challenge robust algorithm incorporating speed information or extracting features which are independent of speed is essentially an area of research. This thesis presents a v comprehensive investigation of the rolling bearing faults and shaft unbalance faults, including their characteristics and fault signatures in the vibrational signals. The presented work proposes two methods based on non-stationary signal decomposition to tackle variational speed problem. The first work introduces an intelligent framework for fault detection using a single sensor. It utilizes Gramian multi-resolution dynamic mode decomposition to process vibration signals. Initially, the vibration signals are transformed using a gram matrix, which converts the one dimensional data into a snapshot matrix that evolves with time, preserving the temporal variation. This transformed data is then subjected to spatial temporal decomposition through multi-resolution dynamic mode decomposition (MrDMD). It decomposes the system dynamics into hierarchically evolving fast and slow modes, enabling the identification of transient fault characteristics. To handle noise from sensors and the environment, a robust least square dynamic mode decomposition algorithm is applied at each level of MrDMD. The resulting mode matrix is further processed by colour coding, effectively converting it into an image format for analysis and classification. The second work fuses vibration signals from sensors placed at three different locations in the frequency domain. This fusion process ensures that maximum spectral information is retained, enabling a more comprehensive analysis. The fused signal is then subjected to decomposition using an energy-preserving maximum overlap discrete wavelet transform, resulting in a multi-scale matrix. Further, to evaluate the severity of the shaft unbalance the decomposed scale matrix is encoded into a contour plot, using the mean absolute deviation of individual scales as iso-reference lines. Finally, the images generated from both the methods are used for classification using different convolutional neural networks. The proposed methodology is evaluated on publicly vi available datasets, from University of Ottawa for bearing fault identification and Fraunhofer Institute for Integrated Circuits for shaft unbalance and severity detection. The results show an overall classification accuracy of 96.83% for bearing fault characteristic and accuracy of 97.05% for unbalance severity detection. The effectiveness of the method is evaluated by comparing the accuracy of fault detection and analysing the performance metrics such as sensitivity and specificity. The finding and result demonstrates the potential of the proposed methodology in improving the reliability and maintenance practices of rotating machinery systems, ultimately leading to enhanced operational efficiency and reduced downtime. The performance surpasses the results achieved by previous studies in terms of adaptability of real-time operation and accuracy.