Computing Intelligence Technique and Multiresolution Data Processing for Condition Monitoring

Condition monitoring (CM) of rotary machines has gained increasing importance and extensive research in recent years. Due to the rapid growth of data volume, automated data processing is necessary in order to deal with massive data efficiently to produce timely and accurate diagnostic results. Artificial intelligence (AI) and adaptive data processing approaches can be promising solutions to the challenge of large data volume. Unfortunately, the majority of AI-based techniques in CM have been developed for only the post-processing (classification) stage, whereas the critical tasks including feature extraction and selection are still manually processed, which often require considerable time and efforts but also yield a performance depending on prior knowledge and diagnostic expertise.

To achieve an automatic data processing, the research of this PhD project provides an integrated framework with two main approaches. Firstly, it focuses on extending AI techniques in all phases, including feature extraction by applying Componential Coding Neural Network (CCNN) which has been found to have unique properties of being trained through unsupervised learning, capable of dealing with raw datasets, translation invariance and high computational efficiency. These advantages of CCNN make it particularly suitable for automated analyzing of the vibration data arisen from typical machine components such as the rolling element bearings which exhibit periodic phenomena with high non-stationary and strong noise contamination. Then, once an anomaly is detected, a further analysis technique to identify the fault is proposed using a multiresolution data analysis approach based on Double-Density Discrete Wavelet Transform (DD-DWT) which was grounded on over-sampled filter banks with smooth tight frames. This makes it nearly shift-invariant which is important for extracting non-stationary periodical peaks. Also, in order to denoise and enhance the diagnostic features, a novel level-dependant adaptive thresholding method based on harmonic to signal ratio (HSR) is developed and implemented on the selected wavelet coefficients. This method has been developed to be a semi-automated (adaptive) approach to facilitate the process of fault diagnosis. The developed framework has been evaluated using both simulated and measured datasets from typical healthy and defective tapered roller bearings which are critical parts of all rotating machines. The results have demonstrated that the CCNN is a robust technique for early fault detection, and also showed that adaptive DD-DWT is a robust technique for diagnosing the faults induced to test bearings. The developed framework has achieved multi-objectives of high detection sensitivity, reliable diagnosis and minimized computing complexity.