Processing noisy line spectrograms as digital pictures

Abstract Contiguous short-term spectral estimates of non-stationary acoustic signals are often displayed as intensity-modulated frequency-time pictures (‘spectrograms’). Some of the problems and advantages in storing and processing certain types of spectrograms as digital pictures are explored. The particular type of spectrogram discussed is from a source producing a time-variant ‘line’ spectrum consisting of one dominant primary line and several much weaker harmonic or quasi-harmonic lines. This type of spectrogram is commonly produced by rotating machinery. It is supposed that this signal is measured amidst additive broad band noise. The resulting spectrogram consists of a family of related meandering lines, the weaker members of which tend to be lost in the noise. In the proposed picture analysis scheme the raw spectral estimates are stored as a digital spectrogram. This spectrogram is then processed by line enhancement methods to extract the primary line to form a ‘family-template’. Using this template the weaker lines can be detected and extracted. By re-assembling the extracted components, a cleaned-up low-entropy frequency-time picture of the signal is produced which shows more structure and which is much more amenable to automatic analysis than the original spectrogram.