Impact of noise model on the performance of algorithms for fault diagnosis in rolling bearings

Condition monitoring of rolling bearings is attracting much interest since most of the production slowdowns depends on the damaging of these components. Several algorithms for fault detection have appeared in the technical literature in the last decade. In most cases, performance is assessed over both synthetic and experimental data. Unfortunately, the computer simulations adopt signal models that are trivial and are not able to predict the actual performance on the field. In this work we propose a framework suitable to fairly, quantitatively and objectively compare different algorithms for fault detection in rolling bearings. The vibration signal is obtained through computer simulations. The signal entailed by the damage is generated through the model at "impact-delay-line" already available in the technical literature. The machine noise is generated as a wideband component with the possible superposition of narrowband components. The wideband component has been modeled as additive white Gaussian noise, additive white noise drawn from an alpha-stable distribution and additive noise stemming from an autoregressive process. Narrowband components are modeled through trains of Gaussian pulses. The performance of three well known algorithms for fault detection are compared in terms of capability in identifying the theoretical cyclic frequencies related to a damage. In these scenarios the behavior of fault detectors are definitely far from that predicted by classical wideband noise models like, for instance, additive white Gaussian noise.