|Abstract||Many industrial facilities, amongst others, are very sensitive to any sudden hazards that can be expensive and resource-costly; therefore, they should be monitored over the day. In this work, new Machine Learning (ML)-based solutions are explored, with an emphasis on compact neural networks methods, for monitoring sensitive equipment using different data measurement types collected from condition monitoring sensors. Consequently, the proposed thesis work tackles a challenging topic that solves problems of high interest to the academic and industrial engineering community. ML algorithms have been thoroughly employed to build numerous structural damage detection systems and Structural Health Maintenance (SHM). ML-based techniques for facility condition monitoring are investigated and presented in a comparative analysis.
Using convolutional layers, network parameters are greatly decreased in Convolutional Neural Networks (CNNs) by local connectivity and exchanging weights, which comprise a group of kernels with a limited receptive area, unlike a typical neural network with a complete connection through each node.
Training of conventional, deep 2-Dimensional (2D) CNNs requires an intensive process in order to obtain an appropriate generalization capacity. This typically involves large-scale datasets, which in turn raises the computational issues considerably. One-Dimensional (1D) CNNs, on the other hand, have recently been proposed in many 1D signal processing applications, including gear condition monitoring. They have recently been designed to overcome these disadvantages by working directly and more effectively on 1D signals.
Nevertheless, CNNs involve homogenous configurations that entirely rely on the linear neuron model. It is evident that, in many complex problems, the required learning performance can only be achieved by deep CNNs with immense complexity. Self-Organized Operational Neural Networks (Self-ONNs), however, are recently introduced to overcome the drawback of the convolutional neurons targeting a greatly complex and nonlinear solution space. With minimum level of complexity of the network and least data to train, Self-ONNs can model multi-modal and sophisticated functions and boost diversity by involving heterogeneity with a flexible set of operators that can be optimized.
In this study, the design and implementation of compact 1D CNNs, as well as Self-ONNs targeting gear cracking fault detection and diagnosis are explored for three different sensor types: acoustic, current, and vibration. It is shown that the presented contribution can detect cracking fault occurrences as well as diagnose the fault severity condition. The proposed approach can effectively deal with limited training data acquired in a physical motor setup designed for this study. The performance of the system is continuously evaluated twice; over 1D CNNs and Self-ONNs, in terms of predetermined metrics for validation on real current, vibration, and acoustic signals collected at a lab in Qatar University. The results indicate to the outstanding performance of Self-ONNs in contrast to 1D CNNs in challenging problems among the three signal types for gear fault detection and level percentage diagnosis. Moreover, further analysis indicates the fastest detection of the first-time occurrence of a faulty frame that can be achieved. The proposed thesis topic presents a social, health, economic, and environmental impact, in addition to scientific and academic dissemination. It is also aligned with different priority themes of Qatar's national priority research themes