Pattern recognition in acoustic signature inspection of railroad wheels

There is a need for an automatic testing system for railroad wheels in service. Earlier work established the feasibility of using acoustic signatures for this purpose, to the extent that such a system has now been installed and is under test. An important conclusion from previous work was that the best detection method was a comparison of sounds from the two members of a wheel set. Pattern recognition techniques were used to detect the difference in the sounds of these two wheels. Preliminary results indicate that these techniques are adequate to discriminate good wheels from flawed ones. Various comparison schemes can be used depending on the features selected. Among features considered are spectral cross correlation, and the number of resonances common to the signals of both wheels. Linear decision functions and a least-squares procedure are used in the classification. The results obtained are encouraging. [Work supported by the Association of American Railroads.]