A method for providing predictive maintenance of a device, comprises the
steps of modeling as a time series of a discretely sampled signal
representative of occurrences of a defined event in the operation of the
device, the time series being modeled as two-state first order Markov
processes with associated transition probabilities, wherein one state
applies when the number of the occurrences exceeds a certain threshold,
and the other state applies when the number of the occurrences falls below
the certain threshold; computing the four transition probabilities the
last N states S.sub.n, where N is a predetermined number, conducting a
supervised training session utilizing a set of J devices, which have
failed due to known causes and considering the two independent
probabilities and, the training session comprising computing the
two-dimensional feature vectors for the initial M windows of N scans,
computing the two-dimensional feature vectors for the final N number of
scans, plotting a scatter-diagram of all 2D feature vectors, and deriving
a pattern classifier by estimating the optimal linear discriminant which
separates the two foregoing sets of vectors; and applying the classifier
to monitor the persistence of occurrences of the defined event in the
operation of the device.
