An electronic engine-health-monitoring system is based on (1) computation of frequency-phase-normalized (FPN) spectra from the digitized outputs of vibration sensors and (2) the use of these spectra as feature vectors, which are presented to an artificial neural network for recognition of features associated with incipient failures. The normalization in question is with respect to the instantaneous frequency and phase of rotation of the engine shaft [customarily denoted the "sync" (short for "synchronous") frequency and phase, respectively]. FPN spectra are useful for extracting that dynamical information that is most useful for detection and classification of failures.
Most machinery failures are preceded by growing tolerances, imbalances, and bearing-element wear, which give rise to subtle modifications of vibration waveforms. The vibration waveforms, and thus the subtle waveform modifications, are corrupted by such benign phenomena as environmental noise, fluid/structural interactions, nonlinear coupling, and feedthrough of vibrations from nearby machinery, all of which compound the diagnostic task. An important element of a vibration-based engine-monitoring system is the ability to extract true defect "signatures" from vibration-sensor outputs that also contain signatures of benign phenomena.
Other, similar systems have utilized frequency normalization; in particular, scaling of the frequency variables of conventional power spectral densities (PSDs) to sync frequencies, to provide more robust representations and simplify analyses of sync-related spectral components. However, the frequency-normalization processes in most such systems discard information on the relative phases of spectral components at various frequencies. These phase relationships are well hidden (they cannot be identified from conventional PSDs). They arise from nonlinear interactions among mechanical components; often, such nonlinear interactions are caused by mechanical defects. In the present system, the relative-phase information is preserved in the frequency-normalization process and is utilized to extract additional information about defects.
The raw vibration-sensor output is sampled and processed according to the phase-synchronized enhancement method (PSEM) (see figure). In this method, the quasi-periodic vibration signal is first sampled at constant time intervals. The sync component of the signal is assumed be characterized by small fluctuations in frequency and phase about a constant, pure sync tone; the instantaneous sync frequency and phase fluctuations are obtained by applying a demodulation technique to quadrature sinusoid representations of the sync signal. The estimated phase fluctuations are converted to time fluctuations (realignment time) to obtain slightly nonuniform time intervals for resampling the vibration signal at uniform phase intervals. The uniform-phase samples are equivalent to uniform-time samples of constant-frequency sync and sync-related components. Once the sync frequency component becomes discrete in the resampled signal, all other sync-related components automatically become discrete.
The phase relationships of interest are those between the sync signal and its harmonics. These relationships can be quantified by use of a hyperspectrum that comprises a hierarchy of joint moments between a reference spectral component (in this case, the sync component) and each of the harmonics. A hyperspectrum can be computed by use of fast Fourier transforms of signals that have been preprocessed by the PSEM.
The performance of the system has been studied in an application to vibration measurements from a bearing test rig, in an effort to identify spectral signatures that could be correlated with wear marks on ball bearings. The results of the study thus far indicate that the inclusion of phase information in FPN spectra yields signatures that enable discrimination of subtle changes that are not represented by spectral energy density or amplitude alone. As a result, a neural network can provide a more reliable and robust representation and classification of patterns for autonomous engine-health monitoring.
This work was done by Jong Jen-Yi of AI Signal Research, Inc., for Marshall Space Flight Center. MFS-26529