Blind Sources Separation (BSS) is a promising technique for signal processing and data analysis that allows the recovery of unknown signals (called sources) from observed signals mixed by an unknown propagation medium. The adjective 'blind' indicates that no assumption was done on the signals and the mixture. This lack of knowledge is compensated by assuming the independence of the source and the linearity of the propagation medium. We apply this technique to a test bench where various machines operate simultaneously in order to diagnose each element. So, by applying the superposition principle, the signals measured by every sensor positioned on each machine, are disrupted by the influence of other signals from the surrounding machines of the factory. Our goal was to remove the influence of the other machines, without having to stop them, which would be damaging for production. BSS methods provide an interesting alternative, since they permit in theory to solve our problem, i.e. to restore on every sensor the signature of its own machine. The sensor signals were assumed to be a convolutive mixture of independent signals arising from different physical phenomena. The method used is based on the N'guyen Jutten algorithm initially developed to separate speech signals. The algorithms used were adaptive and worked on the measured temporal signals from accelerometers or microphones. The signals were produced on a test bed carrying two low-power d.c. motors tired to the same structure, whose speed of rotation could be varied. The signature received by each sensor therefore contained the contributions of the two motors. The results indicate that this approach can be successfully applied to these signals for vibration analysis; acoustical analysis is more complex and will be discussed in detail. (C) 2000 Academic Press.
