Research On Gearbox Fault Diagnosis And Machine Learning Algorithm

The gear is treated as core component of the power and drivetrain system, it is significant to monitor and diagnostic it’s operational condition. In reality, the actual gearbox fault type is unknown, we may face the risk of a variety of failures that may occur at any time, it is of great shortage to monitor a single fault. Meanwhile, in a complex gearbox system, multiple pairs of gears, meshing simultaneously.will create a strong background noise, this noise will cause great inconvenience for gear feature extraction in early weak fault.This paper research feature extraction of fault gear and machine learning based on the common failure types of failure mechanism from angles of stationary and non-stationary with vibration signals.the results as follows:(1) Hilbert transform is used to extract modulation signal on the basis of the bandpass filter, the actual helicopter vibration data proved the obvious effect.(2) The method that improving the sampling frequency can greatly enhance average number of useful signal and does not get large attenuation. In engineering, this article puts forward that angle domain average have good effect to elimate noise in the case of rotation speed fluctuation, the anle averaging signal has a high signal-to-noise ratio.(3) Residual signal is useful o extract weak fault signal feature. Amplitude domain analysis with residual signal can get a better result than TSA signal when analysing actual gear pair. Besides, residual signal has a higher sensitivity for fault gear feature.(4)This paper used wavelet to reduce noise and simulated the amplitude modulation and frequency modualation characteristics of gear failure in the background of gaussian noise and analyzed the effect of different threshold criterion and the mother wavelet function and the layer number of decomposition for wavelet threshold de-noising. The result proved that choosing sym8mother wavelet decomposition layers for4to6layersis best for noise reduction.At the same time, the method of continuous wavelet transform was used in actual vibration data, various scales energy feature extraction method for fault feature extraction and Euler distance method was carried out between the same types and different types of gear Euclidean distance between the feature vector. In view of the root crack fault, using time domain synchronous average method to carry on the continuous wavelet transform, energy feature extraction has a good diagnosis effect, compared with the wavelet denoising method, average time synchronous average method has a better diagnosis effect. At the same time, in view of the tooth root no characterization of the crack fault or the meshing frequency fault type, with the continuous wavelet transform, energy of the residual signal feature extraction method has better diagnosis effect, the method of the sensor measurement points for different between the same type gear, still has good resolving power.Finally, the residual signal by wavelet packet analysis method for fault diagnosis, the experimental results show that under the premise of less calculation time, with the residual signal wavelet packet decomposition and power feature extraction, using Euclidean distance comparison method with continuous wavelet analysis method have the same effect of diagnosis.(5) In view of the fact that practical engineering fault samples are difficult to collect This article puts forward that Support Vector Machine (SVM) learning algorithm is very suitable for type statistics which does not need long time to statistic fault type gears’ thresholds. In the meshing frequency representation of fault type, with the residual signal,the method of wavelet packet decomposition and power feature extraction based on SVM one-class learning has a high accuracy of judging gears’type. for those types that meshing frequency has characterization, with the TSA signal,the method of wavelet packet decomposition based on SVM one-class learning has the high accuracy of judgement.