Nonlinearity detection for condition monitoring utilizing higher-order spectral analysis diagnostics
In this dissertation, we investigate the theory and application of higher-order spectral analysis techniques to condition monitoring in shipboard electrical power systems. Monitoring and early detection of faults in rotating machines, such as induction motors, are essential for both preventive maintenance and to avoid potentially severe damage. As machines degrade, they often tend to become more nonlinear. This increased nonlinearity results in the introduction of new frequencies which satisfy particular frequency selection rules; the exact selection rule depends on the order of the nonlinearity. In addition, the phases of the newly generated frequencies satisfy a similar phase selection rule. This results in a phase coherence, or phase coupling, between the “original” interacting frequencies and the “new” frequencies. This phase coupling is a true signature of nonlinearity. Since the classical auto-power spectrum contains no phase information, the phase coupling signature associated with nonlinear interactions is not available. However, various higher-order spectra (HOS) are capable of detecting such nonlinear-induced phase coupling. The efficacy of the various proposed HOS-based methodologies is investigated using real-world vibration time-series data from a faulted induction motor driving a dc generator. The fault is controlled by varying a resistor placed in one phase of the three-phase line to the induction motor. First, we propose a novel method using a bispectral change detection (BCD) for condition monitoring. Even though the bicoherence is dominant and powerful in the detection of phase coupling of nonlinearly interacting frequencies, it has some difficulties in its application to machine condition monitoring. Basically, the bicoherence may not be able to distinguish between intrinsic nonlinearities associated with healthy machines and fault-induced nonlinearities. Therefore, the ability to discriminate the fault-only nonlinearities from the intrinsic nonlinearities is very important. The proposed BCD method can suppress the intrinsic nonlinearities of a healthy machine by nulling them out and thereby identify the fault-only nonlinearities. In addition, most machines contain rotating components, and the vibration fields they generate are periodic. These periodic impulse train signals may produce artificially high bicoherence values and can lead to ambiguous indications of faults in machine condition monitoring. The proposed BCD method can remove the artificially high bicoherence values caused by periodic impulse-train signals. With these advantages, the proposed BCD method is a new and sensitive indicator for condition monitoring. Second, we propose a novel method to estimate, from a measured single time-series data record, complex coupling coefficients in order to quantify the “strength” of nonlinear frequency interactions associated with rotating machine degradation. The estimation of the coupling coefficients is based on key concepts from higher-order spectral analysis and least mean-square-error analysis. The estimated coupling coefficients embody the physics of the nonlinear interactions associated with machine degradation and provide a quantitative measure of the “strength” of the nonlinear interactions. In addition, as an auto-quantity method utilizing a single time-series data record, the proposed method adds supplemental fault signature information to conventional tools. Such knowledge has the potential to advance the state-of-the-art of machine condition monitoring. Third, we propose a bispectral power transfer analysis methodology to quantify power transfer between nonlinearly interacting frequency modes associated with machine degradation. Our proposed method enables us to identify the relative amounts of power transferred by various nonlinear interactions, and thereby identify the predominant interactions. Such knowledge provides important new signature, or feature, information for machine condition monitoring diagnostics.