Single Sensor Source Separation for Acoustical Machine Diagnostics

Introduction

Acoustical recordings in the context of machine diagnosis usually consist of a multitude of signal components. In order to analyse the individual components of the ma- chine regarding their condition or even to detect faults it is advantageous to handle the signals separately accord- ing to their origin of generation. To become indepen- dent from recording situations, we propose single sen- sor source separation algorithms as preprocessing step for automatic machine diagnostic algorithms. Because of the underlying purely additive model, one promising method for source separation is the non-negative spectro- gram analysis, e.g. the non-negative matrix factorization (NMF). The basic NMF can be extended e.g. by sepa- rating steady state signals, transient signals, and even harmonic components with time-varying pitch and par- tial amplitudes. With such extensions, the non-negative spectrogram analysis can be easily adopted to a wide range of separation tasks. In this contribution an al- gorithm is proposed to separate a measured signal of an electrical machine, containing the superposition of a steady state motor noise signal and the transient signal of a defect bearing.


Paper: SpGnDiOhVo11.pdf

Slides: Slides.pdf


Experimental Results:

Separation of steady state and transient noise.

speed mix defect bearing motor noise
500 x(n) str sst
800 x(n) str sst
1000 x(n) str sst

Separation of noise and harmonic (increasing pitch).

mix defect bearing motor noise
x(n) str sst

Matlab Code:

NonNegativeMachineDiagnostics.m

NonNegativeSeparation.m


© by Martin Spiertz - 24.03.2011 - spiertz@ient.rwth-aachen.de