Mechanical rubbing fault diagnosis method based on multi-wavelet kernel-support vector regression (SVR)

Abstract

The invention relates to a mechanical rubbing fault feature extraction method based on multi-wavelet kernel-support vector regression (SVR). The method comprises the following steps: 1) a vibration signal of a mechanical key component is measured and taken, and the signal is subjected to redundant lifting scheme wavelet (RLSW) decomposition; 2) each layer of detail signals and the last layer of approximation signals obtained by decomposition are subjected to adaptive singular value decomposition (ASVD) noise reduction; 3) a signal is reconstructed by redundant lifting wavelets, and a noise reduction signal is acquired; 4) with two interval three-time Hermite spline multi-wavelet scale functions as interpolation basis functions, a novel multi-wavelet kernel function is constructed; and 5) the vibration signals after noise reduction are divided into a training set and a prediction set, the training set is used to construct a prediction model, the prediction set is predicted, and a weak periodic impact component hidden in the original vibration signal is extracted by the prediction residual. Thus, the mechanical rubbing fault features can be extracted.

Description

technical field [0001] The invention belongs to the technical field of mechanical fault diagnosis, and relates to a method for extracting mechanical fault features, in particular to a mechanical fault diagnosis method based on a multi-wavelet kernel-support vector regression machine. Background technique [0002] By analyzing vibration signals, mechanical faults can be diagnosed without stopping the machine. The traditional method is to judge mechanical faults through the time domain and frequency domain indicators of mechanical vibration signals. Time-domain indicators include average value, effective value, peak value, peak-to-peak value, pulse index, margin index, skewness index and kurtosis index, etc. Frequency domain indicators include spectral center of gravity, mean square frequency, and frequency domain variance. These indicators can be used to roughly judge the mechanical state, but it is difficult to diagnose the specific fault type. Some rotating machinery fau...

AI Technical Summary

Problems solved by technology

However, what kind of waveform the mechanical fault characteristics correspond to cannot be known before diagnosis

[0006] 2) Traditional wavelet transform is based on Fourier transform. It is very difficult to construct wavelet basis functions that satisfy orthogonality, symmetry, short support and high vanishing moment at the same time.

And there are a large number of convolution operations, the amount of calculation is large, and it is difficult to realize online fault diagnosis

[0007] 3) The traditional wavelet transform adopts the Mallat algorithm, adopts down-sampling and zero-filling operations, and the length is halved every time the signal is decomposed, and there will be virtual frequency components, which are false signals that do not exist in the original signal, and are useful for mechanical fault feature extraction. bring difficulties

[0008] 4) The traditional wavelet boundary processing method is unreasonable. The mechanical vibration signal processed actually has a finite length. Generally, zero-filling operations and symmetric operations are used for boundary processing. Oscillations will appear at both ends of the decomposed signal, and false shocks will appear. Element

[0009] 5) Traditional wavelet threshold processing is not very reasonable

[0010] 6) The appropriate number of wavelet decomposition layers, there is no clear guideline

However, the signals generated by early mechanical friction faults often have small amplitude and high frequency, and the traditional threshold processing method will filter out this part of the useful signal

Images