Machine independent approach towards condition monitoring

Current age has been primarily revolutionized by the increased use of rotary machines in our everyday tasks. We are technologically civilized society; everything we see, do or even imagine cannot be realized without the machines. At present, and in the future, the comfort and existence of technologically advance life heavily relies on reliable performance of countless machines around us. To avoid the catastrophic consequences associated with machine failures, research in Machine Health Monitoring (MHM) has attracted many researchers and engineers all over the world. In machines, domain knowledge is required to model parts of the system to find characteristic features for faults detection. Thus, it readily demands a detailed investigation of known and potential faults to obtain their mathematical models which is financially inefficient, cumbersome and nearly impossible for a large number of different existing and future machines, operating in different environments. This thesis identifies stated research problems and proposes a novel self-learning and adaptive modular solution, which attunes to diverse machine applications and is a vital contribution towards Machine Independent Condition Monitoring (MICM) framework. Using Gaussians models over the frequency contents of captured vibrations from the machines, the proposed research continuously monitors the machines for steady state operating conditions; and anomalous situations are detected as soon as anomaly score crosses a certain threshold. To address different application environments and fault scenarios, different size and weightage based bin formation of frequencies representing vibration generated by the machines has been proposed. Fuzzy logic inspired fault memberships to track the growth of the fault and complex anomaly plot to investigate the anomalous contents (either they belong to amplitude anomaly or amplitude anomaly) have been introduced to select transform parameters for fault diagnostic at the later stages. After anomaly detection the next stage is fault diagnosis. Translation invariant features obtained from spectral transform of captured vibrations are used with Artificial Neural Network (ANN) for the autonomous fault diagnosis. To consider the faults signatures with different degrees of non-stationarity, originating from different fault sources Wavelet Transform (WT) and vibration imaging techniques have been introduced to enhance spectral features suiting different transients’ operating conditions. Finally, MICM framework has been devised for autonomous fault detection and classification. Proposed framework comprises of essential and optional modules. Optional modules can be enabled autonomously and selected to obtain improved and reliable fault classification. Using essential modules, translation invariant spectral features are obtained and presented to ANN for autonomous selection of appropriate features based upon the trained Weights and Feature Weight Profile (FWP) of input layer of the network. Optional modules are used as pre and post processors and are autonomously selected by trying different combinations with essential modules during FWP selection. The selection of the parameters of the optional and essential modules depends upon the nature of fault source and is vital to reliable fault classification under worst noise conditions. The presented work in the thesis, with classification accuracies surpassing the existing techniques, signifies the contributions of the conducted research and opens up a new area of research.