Machine Condition Diagnostics Based on Enhanced System Identification Technologies

In modern industries, condition monitoring is considered as the most promising technology to ensure the productivity, reliability and safety of machines. This thesis focuses on developing effective and efficient approaches to monitor the changepoint of rotating machines at an early stage based on the resonant modulation and demodulation techniques. Having had a literature review relating to the resonant modulation based machine condition monitoring, it has been identified that potential technical deficiencies in achieving accurate condition monitoring may lie in the lack of a good understanding of resonant modulation and effective fault detection and diagnosis, which often leads to difficulties in optimising post-processing signal processing approaches to extract accurate fault features.

Based on the hypothesis of linear systems, the resonant modulation in outputs of the systems vary under different input excitations (periodic, approximately periodic impulsive, and quasi-stationary). Different inputs can lead to different modulation characteristics. Performing an effective demodulation of these responses is the most beneficial part for utilising the resonant modulation methods in vibration-based condition monitoring. The characteristics of the vibration signals are expected to be effectively analysed by different signal processing approaches for early fault detection and diagnosis.

From the perspective of the system identification, the stochastic subspace identification (SSI) approach is employed to select the optimal central frequencies rather than selecting the most impulsive frequency band, which is targeted by the method similar to Kurtogram. Through simulation and experimental studies, it shows that the SSI is more robust to the strong background noise (SNR≤-30dB) in selecting the optimal frequency bands than the conventional Kurtogram method. Based on the frequency bands selected by the SSI, a novel method, named ensemble average of autocorrelation signals (EAAS), was developed to demodulate the periodic impacts induced resonant responses. The simulation study shows that EAAS can identify the incipient bearing faults from the noisy vibration signals at a SNR less than -35dB and the performance of the EAAS is also verified by the experiments on ball bearings of induction motors. For the aperiodic impacts induced resonant modulation, the output responses are not periodic and are terminologically named as cyclostationary signals. Two novel approaches, ensemble average of autocorrelated envelopes (EAAE) and phase linearisation based modulation signal bispectrum (PL-MSB), were developed to characterise the cyclostationary responses due to the nonstationary impacts. The simulation studies demonstrated that the proposed EAAE and PL-MSB can handle the extremely poor signals (SNR≤-30dB) to extract the strong fault signatures for the purpose of accurate diagnosis of bearing faults. The experimental studies on tapered roller bearings show that the proposed two methods outperform the conventional envelope for early fault detection and diagnosis under an extremely low SNR and large random slippages.