Local defect detection in bearings in the presence of impulsive noise and spectral overlapping of informative and non-informative impulses

Predictive maintenance has been pointed out as one of the key elements of industry 4.0 philosophy. Measurement of diagnostic signals and advanced analytics allow to detect damage, identify its type, localize it in the machine, and even make some prognosis of remaining life. The efficiency of diagnostic procedures depends on many factors. Recently, research focus has been put on the presence of impulsive disturbance in the raw observation that makes signal processing challenge. The impulsive noise interferes with the operation of commonly used diagnostic methods, impulses excite the resonance frequency of the machine and spread broadband, covering any information about local damage. To be more precise, a local fault is also impulsive (so broadband) but with smaller energy (thus not so broadband as disturbances). What is more, the center frequency of damage may depend on damage location and fault type. In the consequence, in case of impulsive disturbance and impulsive content related to local damage, one may notice spectral overlapping of these components, what additionally makes the analysis more complex. We considered four cases with different levels of overlap, see Fig. 1.

Figure 1. Spectrograms of signals with no overlapping (top left), partial overlapping (top right), full overlapping with different center frequency (bottom left), and full overlapping with similar center frequency (bottom right).

How spectral overlapping may affect diagnostic procedures? We demonstrate the comparative analysis related to the results for three selected techniques, namely the Alpha selector, Conditional Variance-based selector (CVB), and Spearman selector. Each method generates a selector that is used for data filtering to extract only the information from the signal that is most related to the local damage. An ideal case would provide a band-pass filter with weight 1 for the informative frequency band (cyclic impulses) and 0 for noisy bands (e.g. non-cyclic impulses).

Figure 2. Exemplary results of filtration by the three mentioned selectors (for the signal with no overlap – top left case in Fig. 1).

To make the comparative study for tested selectors we used the Envelope Spectrum Based Indicator (ENVSI). It enables the numerical comparison of the envelope spectrum of filtered signals. To properly validate our results, we carried out the Monte Carlo simulations to highlight the effectiveness of the algorithms and indicate the differences between the corresponding results depending on a carrier frequency of periodic impulses and amplitudes of non-periodic disturbances. By analyzing the obtained results (Fig. 3) it is reasonable to consider the level of spectral overlapping of the informative and non-informative impulsive components. The presented results indicate that this factor has some influence on the results of all considered methods. However, one can observe significant differences in these methods, depending on the level of overlapping and also on the amplitudes of the non-cyclic disturbances.

Figure 3. ENVSI values obtained from the Monte Carlo simulations.

This activity has received funding from European Institute of Innovation and Technology (EIT), a body of the European Union, under the Horizon 2020, the EU Framework Programme for Research and Innovation. This work is supported by EIT RawMaterials GmbH under Framework Partnership Agreement No. 18253 (OPMO – Operational Monitoring of Mineral Crushing Machinery).

By Justyna Hebda-Sobkowicz, Jakub Nowicki, Radosław Zimroz and Agnieszka Wyłomańska, Faculty of Geoengineering, Mining and Geology and Faculty of Pure and Applied Mathematics, Wrocław University of Science and Technology