Upcoming interferometric detectors of gravitational waves will generate data at large rates over a long time. The data will consist of not only the strain measurement but also the output of system and environmental monitors. Tests for detecting non-stationarity that require a statistical model for the ambient noise may not be the right choice for analyzing this huge data set since, in most cases, the models would have to be estimated iteratively from the data itself. A solution is proposed in the form of a {\em robust} time-frequency test for detecting non-stationarity whose reliability, in terms of its false alarm rate, is almost independent of the nature of the ambient noise. As a measure of its performance, the detection efficiency of this test is compared against that of an ideal test that requires prior information about both the statistical distribution of the noise and also the frequency band of the burst. When supplemented with an approximate knowledge of the burst duration, this test can detect, at the same false alarm rate and detection probability, bursts that are about 3 times larger in amplitude than those that the ideal test can detect. Apart from being robust, this test has properties which make it suitable as an online monitor.