We present a model-independent, data-driven approach to quantify dynamical changes in nonlinear, possibly chaotic, processes with application to machine failure forewarning. From time-windowed data sets, we use time-delay phase-space reconstruction to obtain a discrete form of the invariant distribution function on the attractor. Condition change in the system's dynamic is quantified by dissimilarity measures of the difference between the test case and baseline distribution functions. We analyze time-serial mechanical (vibration) power data from several large motor-driven systems with accelerated failures and seeded faults. The phase-space dissimilarity measures show a higher consistency and discriminating power than traditional statistical and nonlinear measures, which warrants their use for timely forewarning of equipment failure.
(c) 2004 American Institute of Physics