Adaptive Preprocessing of Nonstationary Signals
Abstract
As the number and bandwidth of sensors increase, an acute demand for preprocessing sensor data obtained for machine-based decision making arises. Especially in a data fusion context, the data from numerous sensors must first be preprocessed to prevent saturation of the decision making mechanism-albeit man or machine. presented is a general preprocessing approach which provides a compact representation (feature vector) of sensor data. The approach, supported by a signal decomposition theorem, adaptively models in recursive fashion, the detrended sensor data as an autoregressive (AR) process of sufficiently high order. Provisions are included to accommodate nonstationary data by incorporating an information-theoretic transition detector to identify the segments of near-stationary data segments which are scale invariant, translation invariant, normalized, and represent sufficient statistics. Furthermore, the merit of the preprocessor is quantitatively determined in a continuous manner from the resulting innovations (modeling error process). Specific application results utilizing nonstationary radar data demonstrate the ability to simultaneously reduce data and maintain information content, without requiring a priori statistics and/or expert rules.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 09, 1989
- Accession Number
- ADA209646
Entities
People
- M. D. Eggers
- T. S. Khoun
Organizations
- Massachusetts Institute of Technology