Research On Feature Extraction For Multicomponents In Mechanical Measuring

Vibration signals generated by a running mechanical system contain lots of fault information. By analysis into them, state changing of the parts in the mechanical equipment can be recognized, and the fault can be found out. Although time-frequency analysis methods have gained wide application in mechanical fault feature extraction and after decades of development, lots of researches into the feature extraction have been done for unstationary signals, however, demands from engineering applications still stay unsatisfied. With this background, signal analysis methods including wavelet transform, Hilbert-Huang transform(HHT), Chirplet transform, iterated Hilbert transform(IHT) are researched, signal processing technologies including reassignment and multi-taper spectrum estimation are studied, and denoising, demodulation, decomposition etc. feature extraction technologies are invistigated.Firstly, the theories for wavelet transform which is well known for its multiresolution characteristics is reviewed, including continuous wavelet transform and its allowance condition, discrete dyadic wavelet and its reconstruction condition, wavelet frame, multiresolution analysis and Mallat algorithm. In order to conquer the frequency aliasing when the signal is decomposed by wavelet transform, harmonic wavelet transform is employed to fulfill signal feature extraction. Since the coefficients of harmonic wavelet transform has the unique amplitude-frequency holding characteristic, an improved way to construct the time-frequency spectrum with its coefficients is proposed. And deduction indicates that cross section of the improved time-frequency spectrum is right the envelope of the component in the corresponding frequency band. Simulation and application show the efficiency of the demodulation based on harmonic wavelet.Since wavelet transform lacks local self-adaption, leakage will occur, which makes the signal to be smooth in global, therefore, Hilbert-Huang transform(HHT) is studied. Aiming at the problem of EMD's lacking strict orthogonality, a method based on Gram-Schmidt method to make the EMD process orthogonal is introduced. In fact, it didn't lead to orthogonal EMD. Hence the feasibility of obtaining IMFs by band pass filtering is investigated, then the orthogonal empirical mode decomposition(OEMD) based on band pass filtering is proposed, and its completeness and orthogonality are proved. Also, a fast implement algorithm, IMF binary searching, is proposed for OEMD. Applications into harmonic detection and singularity detection show that the mode aliasing and false mode are eliminated by OEMD, while applications into fault diagnosis for rotor experiment bench and gearbox indicate that for either stationary signal or unstationary signal, OEMD achieves good analysis results.Chirplet transform is an extension for wavelet transform. Firstly, the principle of chirplet transform and its self-adaptive decomposition are presented. The superiority of chirplet self-adaptive decomposition is demonstrated by simulation. To solve the problem of feature extraction from a great background noise, chirplet best path pursuit which combines chirplet transform and best path pursuit is introduced, and also, its drawback is presented. Therefore, the combination of EMD with chirplet best path pursuit is proposed to extract complex multi-components from great background noise. Its efficiency and feasibility are revealed by simulation. And it's applied into the analysis of the vibration signal from an oil transfer pump, which gives correct diagnosis result.Demodulation is a traditional and effective way to extract mechanical fault features. Therefore, demodulation for multi-components is researched. The most frequently used demodulation methods are reviewed here, so are the limitations for each one. Then the model for multi-components employed in IHT, iterated procedure to get each component's amplitude envelope and instantaneous frequency(IF), and detailed method to design the filters for IHT are elaborated. Analysis shows that there are certain limitations to directly use the phase information got in the IHT procedure to calculate the IF of each component. Instead, a smoothed IF estimation method(SIFE) is proposed, and its efficiency is later indicated by simulation. IHT and SIFE are used for early fault diagnosis for rolling bearing and good results are achieved, which provides mechanical fault diagnosis with a new method.Reassignment is able to improve the time-frequency concentration of binary time-frequency representation, while multi-tapering is of good estimation variance. The principle and property of reassignment are investigated, and the reassigned forms of different time-frequency and time-scale representations are presented. Application into mechanical fault feature extraction also shows reassignment is efficient in feature extraction for multi-components. Then multi-tapering proposed by Thomson is presented, and so is its extension into unstationary signal analysis. Combining the two preceding methods, merits of each one are still there, and with the combination, characteristic information of signals can be extracted from very noisy background. It is applied into the diagnosis of a rotor un-centering fault, outer fault of a rolling bearing and a gearbox fault. A comparison with reassigned time-frequency spectrum is done, which says the combination is successful in using into mechanical fault feature extraction.Finally, a summarization of the thesis and expectations for the future research ends the whole thesis.