Model-Based Features

Mathematical models of time-domain signals represent the underlying dynamics of the mechanisms that generated the signals. Autoregressive models are particularly useful for applications such as rotating machinery, which produce sharp peaks in their power spectra. Features extracted from these models can reflect behavioral differences that arise from healthy or faulty conditions. [1]

Diagnostic Feature Designer generates autoregressive models from the imported signal and extracts features in the following areas.

• Autoregressive Model Features — These features characterize the raw differences in the autoregressive (AR) model parameters that the app estimates. Specify Model Order and Number of Peak Modes. You can then create features from the estimated model coefficients and from the natural frequencies and damping factors for the number of peak modes that you specify.

• Model Fit Features — These features characterize the goodness of the model fit. In the presence of anomalies, mechanical changes can introduce additional dynamics that are not accurately captured by a low order model. Poor model fit implies possible faulty behavior. The features include mean squared error, mean absolute error, and Akaike’s information criterion (AIC), which is a measure of model quality.

• Model Residual Features — These features provide statistics related to the prediction error, that is, the differences between the predicted model output and the measured output. Residuals represent the portion of the signal that is not explained by the model. The features include the residual mean, residual variance, residual RMS, and residual kurtosis.

For more information about autocorrelation modeling, see levinson.

References

[1] Wang, Wenyi, and Albert K. Wong. “Autoregressive Model-Based Gear Fault Diagnosis.” Journal of Vibration and Acoustics, vol. 124, no. 2, Apr. 2002, pp. 172–79. DOI.org (Crossref), https://doi.org/10.1115/1.1456905.