Research On Fault Diagnosis Of Planetary Gearboxes Based On Deep Learning

Planetary gearboxes are extensively used in essential mechanical products represented by construction machinery,with its high transmission ratio,strong bearing capacity and excellent transmission stability.As the core transmission components,once it breaks down,it is likely to bring huge production losses and high maintenance costs.However,in the field of intelligent diagnosis of planetary gearboxes,the research on the equipment under non-stationary condition is not thorough.In view of this,it is critical to do research on processing methods of non-stationary vibration signal of planetary gearbox and develop intelligent fault diagnosis technology,which is more suitable for fault feature extraction and classification of planetary gearbox.This will be beneficial to establish intelligent fault diagnosis system of planetary gearboxes.Besides,this can realize early fault diagnosis and warning,cut the conservation cost of planetary gearboxes as well as avoid safety accidents.It has a high application value for the intelligent monitoring and health of construction machinery maintenance.Considering the key issues above,this paper takes the SD 16 bulldozer gearbox produced by Shantui Engineering Machinery Company and the fan gearbox produced by CRRC Wind Power Company as the research objects in order to achieve the research goal of fault feature extraction and intelligent identification of planetary gearbox under different working conditions.In this paper,a signal pre-processing method based on wavelet packet threshold denoising and a time-frequency feature extraction method for non-stationary signals based on language spectrum analysis are proposed,fault diagnosis models of planetary gearbox and transfer learning grounded in deep convolutional neural network are established.Meanwhile,the experimental verification research is carried out.(1)Aiming at the extracted problems of filtering out the noise disturbance in the non-stationary vibration signals of planetary gearbox,this paper investigates and analyses the commonly used signal filtering and noise elimination methods at home and abroad.Wavelet transform and wavelet packet transform are used to improve the method of noise elimination for the non-stationary vibration signal of planetary gearbox.Also,the signal-to-noise ratio(SNR)and mean square error(MSE)are used as the judgment indexes to carry out the de-noising contrast test,and the results of the de-noising of wavelet packet threshold are determined to be better.Furthermore,a series of comparative tests are designed through the selection of the basis function,the use of the soft and hard thresholds and the determination of the number of decomposition layers in the process of wavelet packet threshold de-noising.Accordingly,the de-no ising method of the soft thresholds of wavelet packet is determined:Using Sym4 function as the basis function,the best denoising effect is achieved for the non-stationary vibration signal with noise when the number of decomposition layers is 5.Finally,an effective filtering and noise reduction method for the non-stationary vibration signal of planetary gearbox is established.(2)In view of the non-stationary velocity characteristics of the vibration signals of the planetary gearbox under different working conditions,this paper investigates the processing methods of non-stationary signals in various fields.Based on the good performance of speech spectrum analysis in the extraction of speech features,a method for processing the non-stationary vibration signals of planetary gearbox is proposed.In practical applications,the selection of parameters such as window function,frame length and frame shift is adjusted according to the parameters such as frequency,strength and time-varying rate of the original vibration signal of the planetary gearbox collected,aiming to obtain the best time-frequency characteristics of speech spectrum.At the same time,the short-time Fourier transform(STFT)technique is used to map the segmented spectral features in order to complete the spectral map.This method can effectively extract the primary fault features of the non-stationary vibration signals of the planetary gearbox under different working conditions in the form of time-frequency images,which provides technical support for the further recognition and classification of the faults of the planetary gearbox.(3)Deep neural network can deeply mine the useful feature information in the data and intelligently classify the recognition objects.In order to improve the deep feature extraction ability and model diagnosis accuracy of the fault diagnosis model of planetary gearbox,the design and training of the fault diagnosis model of planetary gearbox grounded on deep convolutional neural network(CNN)are completed in this paper in combination with the theory of deep learning.The model combines the batch normalization regularization method and Adam parameter optimization algorithm on the ground of VGG16 convolutional neural network.It not only improves the generalization ability of the model,but also speeds up its convergence speed.It also has high recognition accuracy and diagnosis efficiency in fault diagnosis cases of planetary gearboxes with different structures under different working conditions.Therefore,the model realizes the deep fault feature extraction and the efficient identification of the health state of the specific research object.(4)Deep neural network can adaptively extract the deep features of fault information,but it is not suitable for the training of samples that do not conform to independently identically distribution and small data sets.In view of the difficulty in obtaining fault samples of planetary gearbox in engineering applications and the cost of research and development for the reconstruction of fault diagnosis model of similar models,which leads to low diagnosis accuracy and low production efficiency.In this paper,combined with transfer learning knowledge and taking SD16 bulldozer gearbox produced by Shantui Engineering Machinery Company as the experimental research object,the design and training of transfer learning fault diagnosis model based on convolutional neural network are completed.The prior knowledge is used to complete the fault diagnosis task of data sample shortage.Thus,the training cost of similar models is effectively cut while fault diagnosis efficiency is enhanced.