The number of basis vectors or modes to use in the model must be determined for each model. If too
few modes are used, the model will poorly represent the transformed vibration signal from the good
gear. If too many modes are used, more computation will be required producing more coefficients
than needed and the model may over fit the data. Several criteria are available. The number of
modes can be chosen to account for a fixed amount of rms or variance in the training set [40]. This
method requires selecting a cutoff criterion. The number of modes can be chosen by examining the
basis vectors and choosing only those that look like a signal [39]. This method is labor intensive and
requires selecting a cutoff criterion. Another way to select the number of modes is by statistical
hypothesis testing of the multiplicity of a noise eigenvalue in the singular values to distinguish
between noise and signal [41,42,43]. The authors chose to use the last method that is commonly
used in array signal processing for methods that require knowledge of the number of signals in the
data. This method makes the assumption that data are the sum of an unknown number of stationary
signals and ergodic Gaussian random noise.