Transient behavior of machinery is becoming increasingly recognized as a major factor in predicting life of components and exposing degradation or faults that are not observable from steady state measurements. The methods developed here are directed to trawling archival data records to reconstruct a sufficiently informative experiment in lieu of dedicated testing. The central novel observation is the modification to subspace system identification methods to permit their application to multiple concatenated, but non-contiguous data sets, which jointly can be sufficiently rich for identification while individual records are not. This is a break from traditional system identification, where one uses a single contiguous synchronous input-output record.
