| A new analytical approach is developed for detecting incipient faults of rotating machinery whose periodical characteristics generate time series data representable as cyclostationary processes. The new approach is a higher-order statistical (HOS) method as nonstationary time series estimation, in addition to stationary and nonlinear estimation, provide the basis for enhanced feature information of the random fault mechanisms under study. An algorithm selects and combines different transformed estimates of the raw time series, second-order cumulant spectrum (nonstationary), power spectrum (stationary), and bispectrum (nonlinear), for investigation of incipient fault discrimination and classification power of multivariate classifiers using different extracted feature information sets. The HOS approach (cumulant spectrum, bispectrum, and power spectrum), is tested and evaluated against a traditional power spectrum approach with simulated and actual experimental data. Robustness of the HOS approach is first investigated in simulated time series signals with amplitude and phase modulation indices and differing levels of additive Gaussian noise as parameters. Simulations show that use of HOS features improves incipient fault detection capability of a linear classifier and is less sensitive to Gaussian noise within the signal environment. Actual vibration signals from a rotating drill wear monitoring study are also analyzed. The drills are used in the manufacturing of electronic circuit cards from epoxy-glass composite. Combining HOS features with power spectrum features improved the overall classification performance of parametric and non-parametric classifiers. Additionally, the HOS approach is less sensitive to changes in drilling process parameters such as circuit card construction and chip load. The pattern recognition analyses performed in this research provide strong statistical evidence that HOS estimation and feature extraction is beneficial for discrimination and classification of incipient failures of rotating tools, a difficult mechanical system monitoring problem. |
