Non-local Mean Algorithm With Applications To Impact Feature Enhancement For Rolling Bearing

Rolling bearings are among the most important parts of rotating machinery.They are usually subjected to high running speed and heavy load,and thus being the most vulnerable parts.As such,bearings have become one main subject within fault diagnosis community.Although a lot of efforts have been devoted to develop various techniques for bearing fault detection,results in field applications are not yet satisfied.Up to date,many researchers are still seeking and developing new approaches to bearing fault diagnostic.Fault feature enhancement is the base of fault diagnostic and determines diagnosis accuracy.For feature enhancement,there exist a wide range of approaches like minimum entropy deconvolution(MED),spectral kurtosis(SK)and so on.Such methods have reported promising results for the cases presented in some literature.Their deficiency,however,is obvious in the case of real field applications.In this dissertation,a novel approach of non-local mean algorithm(NLM)is introduced from image processing community,with aim to provide a bearing fault feature enhancement characterized by anti-noise and excellent diagnosis capability.NLM is originally proposed for two-dimensional(2D)image processing,which utilizes weighted average of pixels and self-similarity of structures to eliminate noise.NLM is reformulated here to deal with one-dimensional(1D)signal like vibration signal from bearings.A feature enhancement model is proposed for detect bearing local damages which induce cyclic impacts.Results on experimental data demonstrate that NLM is able to relieve the noise and highlight the abnormal impact characteristics representing bearing fault information.The processing results of NLM are significantly dependent on such parameters as ?,M and P,of which inappropriate value will lead to an unsatisfied result.Particle swarm algorithm is employed to assist the determination of such three parameters in an adaptive way instead of manual operation,resulting in the development of an adaptive NLM.Adaptive NLM exploits the reciprocal of the kurtosis of the resultant filtered signals as objective function.A more promising result is expected by using the adaptive NLM than traditional NLM in the sense of feature enhancement of cyclic impacts embedded in bearing vibrations.Experimental results validate the proposed method.In the case of strong background noise,NLM is inadequate to extract local fault induced cyclic impacts from bearing vibrations due to the average processing used in NLM.Aimed at this issue,a new approach call weight envelop is proposed.Weight envelop make direct use of the weight curve of NLM as the envelope of processed signal instead of obtaining envelop from the NLM filtered signal.Omitting the averaging process mitigates the smearing phenomenon in NLM and a better result in the case of low SNR(signal-to-noise rate)is expected.The frequency spectrum of the resulted weight envelop(termed as weight envelop spectrum)is then given to determine the health condition of bearings.The effectiveness of the method of weight envelop is examined by using both synthetic and experimental data.Results reveal that weight envelops are less computationally expensive in addition to its advantages for signals of low SNR in comparison with NLM approach.Weight envelop spectrum is able to extract fault feature in the case of low SNR.However,its efficacy should be reinforced with respect to weak features generated by incipient bearing faults.The combination of weight envelop spectrum with other prevalent methods such as MED,SK and wavelet packet decomposition is explored in order to harness their respective advantages.More different the fault induced shocks from the noise,more efficient the weight envelop spectrum.As such,an efficient preprocessing will benefit the proposed weight envelop spectrum.On the other hand,weight envelop spectrum can be treated as the post-processing of those methods.A great deal of case studies demonstrates the combinations and indicates the advantages in weak feature enhancement for incipient bearing faults.Finally,an intelligent fault diagnosis system prototype is constructed on the base of Lab View platform.The prototype consists mainly of three parts including signal processing and filtering,time domain analysis,and frequency domain analysis.Users can obtain vibration data information by the comparison of characteristic parameters as well as time domain and frequency domain processing,which renders bearing condition monitoring and fault diagnosis to be implemented in a comparatively easy and intuitive fashion.For practical applications,this prototype is to be enhanced,but it provides a platform for the field applications of the family of the NLM approaches investigated in the present dissertation.