Research On Fault Diagnosis Of Quayside Crane Reducer Based On Support Vector Machine

The reducer is an important part of the port crane,its running state has an important impact on the safety and reliability of the port crane.In order to prevent the sudden failure of the gearbox,the method of regular maintenance is usually used in engineering.However,due to the contingency of gearbox failure and the scattered life distribution,in this case,the way of scheduled maintenance is easy to cause "excessive maintenance" and "insufficient maintenance".Therefore,it is necessary to introduce on-the-spot maintenance instead of scheduled maintenance.However,condition based maintenance must be supported by perfect condition monitoring technology and fault diagnosis algorithm.This paper focuses on the research of fault diagnosis algorithm.Based on the current signal processing method and machine learning algorithm,starting with the vibration acceleration signal of gearbox,the research of gearbox fault diagnosis is carried out with support vector machine as the core algorithm.It mainly solves the following problems:(1)Firstly,to solve the problem of low signal-to-noise ratio of the collected vibration signal,the time-frequency signal is transformed into the time-frequency signal by S-transform,and the time-frequency signal is reconstructed by the threshold value of singular value in the time-frequency domain,then the reconstructed time-frequency signal is transformed back to the time-domain signal by the inverse transform of S-transform.Through the above steps,the original time domain signal can be denoised.From the experimental data,this method can not only effectively remove the added Gaussian noise,but also suppress the original background noise.(2)Secondly,based on the study of the failure mechanism of two main parts of gearbox gear and bearing,this paper puts forward the idea of multi type feature fusion.Four time-domain features are extracted from two aspects of long-term stability and sensitivity to signal change;seven frequency bands are divided by gear meshing relationship and bearing characteristic frequency,and the ratio of energy amplitude of each frequency band to that of the whole frequency band is extracted as the frequency-domain feature;the first six IMF components with the most concentrated energy are extracted by analyzing the energy distribution of IMF components obtained by EMD decomposition The energy moment is taken as the characteristic quantity.The 17 dimensional vector composed of the above three kinds of features is used as the feature of the original signal,which solves the problem that the recognition rate of the traditional single type feature training model is low on the one to many classification problem.(3)Since it is very difficult to train and recognize the support vector machine model directly with 17 dimension eigenvectors,principal component analysis is introduced to reduce the dimension of the input vector.In this paper,PCA dimension reduction is carried out after extracting features of simulated fault signals collected by fault test-bed,and the relative data accuracy is maintained after 17 dimension vector is reduced to different dimensions.It is found that when 17 dimension is reduced to 4 dimension,93% accuracy can still be maintained.Therefore,the final output dimension of PCA dimension reduction is determined as 4 dimension in this paper.(4)Finally,50 sets of features are extracted from 6 kinds of data signals(including 3 kinds of bearing fault,2 kinds of gear fault and 1 kind of normal)collected by the fault test-bed,and 300 sets of feature samples are obtained.40 groups were used for training model and 10 groups for testing model.The combination of different kernel functions and penalty factors is tested from two aspects of running time and classification accuracy respectively.Finally,Gaussian radial basis function and penalty factor C = 50 are selected as the optimal combination.Based on the common faults of the gearbox test-bed prefabrication in Shanghai Zhenhua heavy industry group,this paper deals with the signal noise reduction,feature extraction,multi feature fusion and dimensionality reduction,and finally determines the best combination of support vector machine kernel function and parameters.After the training,the SVM model is tested,and the classification accuracy is 93.3%,which achieves the expected effect.