An experimental study of detection in nonstationary noise

The performance of any system for detecting a noise-like signal in additive background noise is critically dependent on the statistics of the signal and the noise. In analyzing the performance of a given detector it is necessary to make assumptions regarding these statistics. The situations in which both the noise and the signal are Gaussian and stationary (or very nearly so) are quite numerous, and the assumption that such is the case has led to useful results. There are, however, situations where this assumption is not justified: fading signals and fluctuations in noise intensity are commonly encountered in practice.1¿7 In such cases, for short observation intervals the signal and the noise may often be regarded as Gaussian and approximately stationary and detectors can then be designed and analyzed on this assumption. However, if the observation interval is lengthened sufficiently in an attempt to increase the sensitivity of the detector, the inputs to the detector will eventually be nonstationary and the performance of the detector will differ from that expected for stationary inputs.