Research On Intelligent Fault Diagnosis Of Gearbox Based On ART2 Neural Network

At present,China's industrial development is rapid,and the number of gearboxes in mechanical equipment is growing at a high speed,but it also poses a huge challenge to the fault diagnosis of the gearboxes.In order to adapt to the development brought by big data and improve the fault diagnosis efficiency of the gearboxes,it is necessary to continuously improve the intelligence level of fault diagnosis.At present,there are many intelligent fault diagnosis methods applied to gearbox pattern recognition,which are mainly divided into pattern recognition based on supervised learning and pattern recognition based on unsupervised learning.However,pattern recognition based on supervised learning is mainly based on training samples because it is difficult to obtain sample data with universality.Pattern recognition of supervised learning of data is difficult to use in practical applications.Under this premise,this paper mainly studied the pattern recognition method for unsupervised learning of gearbox fault diagnosis.Firstly,an Adaptive Reonance Theory 2(abbreviation:ART2)neural network based on unsupervised learning was introduced.And the ART2 neural network was used to express the running state of the gearbox,and the ART2 neural network realized the pattern recognition of the gearbox running state without the training sample.In the extraction of eigenvalues,the relative wavelet packet energy was used as the extraction method of eigenvalues.Then the ART2 neural network was used to analyze the extracted eigenvalues.The minimum of intra-class distance criterion function was used to select the alert value to improve classification accuracy of the neural network.The proposed method was analyzed and verified by the bearing life vibration data measured by the test bench.The analysis results showed that the method could reflect the state transition process of the bearing from normal to failure,and verify the feasibility of the method.Secondly,because the ART2 neural network only activates neurons with the largest output value when activating neurons,it ignores the effects of other neurons with larger output values,and these neurons with larger output values may also meet the warning value,resulting in a single calculation result.Especially in the case where the pattern characteristics of the input samples are similar,the classification accuracy is difficult to guarantee.Based on the above analysis,this paper introdoced a classification method of ART2 neural network combined with K-means algorithm,and introduced K-means algorithm into ART2 neural network to repair this defect of ART2 neural network.The feasibility of the proposed method was verified by simulation data.The results showed that this method successfully solved the problem that ART2 neural network only selected the largest neuron of output value and improved the classification accuracy.Finally,because the vibration signal collected from the gearboxes in actual work is noise-containing.Therefore,this paper introduced the wavelet threshold denoising method to preprocess the signal under the premise of combining the ART2 neural network of K-means algorithm.The proposed method was validated by multiple seted of experimental data.The experimental results showed that the proposed method successfully solved the problem of noise interference and further improved the classification accuracy.