Research On Intelligent Fault Diagnosis Method Of Gear Based On MSVDD Model

Gearbox is widely used in mechanical transmission,which can play the role of acceleration and deceleration,clutch control and power distribution.The working state of the gearbox is related to the performance of the whole mechanical equipment.Because the working environment of the gearbox is bad,it often runs uninterrupted under the condition of high load and high speed,so it is of great significance to diagnose its fault.Nowadays,most of the fault diagnosis researches deal with the vibration signal from three aspects: First,the traditional time-frequency analysis method and the extraction of gear fault related features are used to analyze the running state of gear.The second is to manually extract and screen the features from the vibration signals according to the expert experience so as to carry out pattern recognition;The third is to automatically learn the features of the original signal to achieve the purpose of pattern recognition.In view of this,this paper firstly extracted gear fault features based on the Angle Domain Synchronous Averaging technique,and qualitatively analyzed its correlation with gear fault.Then,a Support Vector Data Description(SVDD)algorithm and a Convolutional Neural network are used Networks(CNN),the multi-classification SVDD(MSVDD)model and the optimized CNN model are respectively built for gear fault diagnosis.The main research work is as follows:(1)In view of the characteristics of strong noise and strong non-linearity of gear vibration signal,the preprocessing technology and feature optimization method suitable for gear vibration signal are studied.Gear vibration signals often have a lot of noise pollution.In this paper,synchronous average technology is selected to improve the signal signal-to-noise ratio.The evaluation index T-D(TSA-Diff),namely the ratio of the average power of the difference between the synchronous average signal and the differential signal to the average power of the differential signal,is used to evaluate the quality of the synchronous average effect.At the same time,the validity of synchronous average signal feature in gear fault diagnosis is verified by experimental data.Aiming at the problems of feature redundancy in gear multi-domain state feature concentration and weak fault characterization ability of gear,three feature optimization methods were studied and compared.The high-dimensional features were visualized by dimensionality reduction using TSNE tool.The experimental results show that the feature subsets selected by the feature optimization method of Maximal Relevance Minimal Redundancy(MRMR)have good characteristics of in-class compaction and inter-class separation.(2)SVDD algorithm is often used to solve single-valued classification problems.However,support vector can not effectively improve the ability to distinguish boundary samples in multi-valued classification problems such as fault diagnosis.A gear intelligent diagnosis method based on MSVDD model is studied.In view of the "one-vs-rest" idea in SVM,this paper generalizes SVDD to the field of multi-classification.In order to solve the problem of low diagnosis accuracy of the sample in the alias domain in MSVDD model,combining the effects of penalty factor C and Gaussian kernel parameter σ on the hypersphere model,Adaptive Chaotic Simplified Particle Swarm Optimization(ACSPSO)algorithm is used to optimize C andσ.And the fitness function is designed to ensure that the optimization results can make the MSVDD model include as many target samples as possible without the phenomenon of aliasing domain.Finally,the rationality of fitness function design and the validity of MSVDD model are verified by experimental data.(3)In view of the fact that the feature extraction of signal depends on expert experience in fault diagnosis research,a gear intelligent diagnosis method based on CNN optimization was studied.In this paper,a CNN model suitable for gear fault diagnosis is designed from the perspective of convolutional neural network.In view of the shortcomings of CNN model,three optimization methods,batch normalization,residual module and global mean pooling,are introduced to optimize the CNN model,and the effectiveness of the combination of these three optimization methods is verified through experimental data.That is to say,the optimized CNN model not only has fast convergence speed and high fault identification accuracy,but also can effectively reduce the total parameters of the network model.After several tests,the model shows good stability in performance.In addition,the condition factors can still be excluded in the data set of multiple working conditions to maintain a high fault identification accuracy.Finally,the performance of MSVDD model and optimized CNN model is analyzed from two aspects of time cost and noise resistance.The experimental results show that the time cost of training MSVDD model is much lower than that of training optimized CNN model.The noise resistance of the optimized CNN model is better than that of MSVDD model.Therefore,a suitable diagnostic model should be selected according to the specific application scenarios in engineering application.