Conditional time-frequency moments are discussed and applied to the Westlan
d helicopter test data. The conditional moments we use have a simple physic
al interpretation, namely, they are the mean, median and mode frequencies a
t a given time, and the spread about the mean frequency at a given time (th
e 'instantaneous spectral bandwidth'). We show that they characterise the f
aults well and can differentiate between different fault classes. Using spe
ctrograms for their calculation enables real-time computation. Accordingly,
these conditional moments are a sparse feature set that can potentially se
rve as an on-line diagnostic tool for detecting and monitoring developing f
aults. (C) 2000 Academic Press.