Fault Diagnosis For The Rotating Machinery Based On The EMD And HHT

Nowadays, the rotating machinery has been widely used in almost all of the industry sections. Therefore it is extremely significant to detect the fault of the rotating machinery. Generally acknowledged, extracting fault feature is the critical step of fault diagnosis. To extract fault feature effectively, signal processing-based methods are widely used today. Due to the fact that most fault vibration signals of the rotating machinery present non-stationary properties, it is essential to choose appropriate signal processing methods that are suitable for non-stationary signals to extract fault feature.The time-frequency analysis methods are widely used in rotating machinery fault diagnosis because they can provide both time and frequency domain information of a signal simultaneously. In this paper, three of the most general basic/classical methods are compared:Fourier Transform, Hilbert Transform and Wavelet Transform. It is shown that when being applied to analyze typical wavelet bases, the three methods above are equable in mathematical containment, and, even, so far as the relevant parameters matched to each other, the results are equable in essence. The equivalence which is suitable for the phase, amplitude, and all other physical parameters derived from the complex expression of time-frequency analysis contradicts the assertion that methods based on Hilbert or wavelet are superior to the Short-Time Fourier Transform (STFT) conditionally. The present paper takes frequency amplitude as an example to demonstrate the practical application of SFT, Hilbert Transform and Wavelet analysis in vibration signal processing. In addition, it is proposed in theory and experiment that the key role to optimize the result is to find or choose the exact time-frequency resolution so as to control the core of the filter (time domain) or transform function (frequency domain) adroitly no matter what approaches are adopted.Traditional time-frequency methods (signal convoluted with the pre-specified basis function), however, are limited because the pre-specified basis function can not adapt to various data sources. Consequently, the data mismatched the basis function are assigned to harmonic waves automatically. Meanwhile, the process of convolution embraces integration procedure, which makes the results more bounded. Recently, N. E. Huang put forward a new time-frequency methodology dealing with non-stationary and non-linear signal-Hilbert-Huang Transform (HHT), consisting of Empirical Mode Decomposition (EMD) and Hilbert Transform. The new path shows more adaptability than FFT. Despite its effectiveness in practical circumstances, most of the underlying mathematical problems have still not been treated. Thus, the EMD method is open to questions and improvement such as the criterion of IMFs’sifting, end effects, model aliasing and etc. The present paper, to prevent over decomposition, advocates that the mean curve of envelope is permitted to fluctuate slightly on the whole length of data, as well as locally large fluctuation of sifting criterion. Together with the introduction of linear trend extraction, the problem with decomposition of IMFs is well resolved. On the other hand, the divergent phenomena occurs when decomposing non-stationary signals with the EMD method, and such divergence definitely results in corrupting the whole data series and makes the decomposition distorted. In this paper, the modification of rectifying the IMFs’ends is outlined and the end effects are well overcome. Another difficulty with the EMD method is model aliasing (one IMF contains several local oscillation with distinct frequencies and scales). The prior measure is to assign a subjective frequency or scale for detecting. Though it works in most circumstances, the adaptability of the algorithm is degraded. In this paper, temporary frequency is introduced so as to enlarge the frequency ratio among the respective signals, and so circumvented the model aliasing problem.Signals extracted from rolling bearings and gears are featured with modulation and noises, thus the demodulation on such signals is significant in the fault diagnosis for the rotating machinery. Typical envelope demodulation need a preset central frequency as its natural one and then conducts band-pass filtering to the original data. Dependently, frequency components consist of several time domain signals may sometimes not be filtered near the sideband, subsequently, frequency that is not fit for analysis occurs on the demodulation spectrum. In this paper, according to simplified mathematical model of rolling bearings and gears, several fault analysis are proposed, including waveform parameter analysis based on EMD, and resonant demodulation analysis based on HHT. Experiments from simulation and open data from internet prove that the adaptability of these methods enables the efficiency of abstracting parameters of time and frequency domain fault features from rolling bearings and gears, thus enhances the accuracy of fault diagnosis.