Study On Rotary Mechanical Fault Identification Based On Deep Learning Reconstruction Algorithm

With the development of computer artificial intelligence as well as the progress of science and technology,rotating machinery equipment is also developing towards the intelligent direction.the traditional artificial recognition method has been unable to meet the needs of modern rotating machinery equipment.Therefore,in the era of big data,studying a method of processing large rotating machinery data can’t wait.Deep auto-encoder neural network is an unsupervised learning algorithm for deep learning,and because of the advantages of the ability to process sophisticated data and the ability to automatically extract features,it has been established in the field of image processing and voice recognition,but it has yet to be developed in the field of fault recognition.This dissertation from the vibration signals of rotating machinery fault identification research background,the traditional artificial fault identification method needs a lot of problems such as data processing technical experience,and this paper studies the fault identification of rotating machinery based on deep auto-encoder reconstruction algorithm.First of all,this dissertation analyzes the basic principle of deep auto-encoder,and studies the feature extraction ability of deep auto-encoder algorithm which mainly studies the feature extraction ability of single auto-encoder in different iteration times and different hidden layer nodes,and then stack the derivative structure of auto-encoder to form deep autoencoder.The deep auto-encoder algorithm can automatically extract features from the original time domain data collected by the sensor,using public simulation data set of rolling bearing fault feature extraction and state recognition of DAE will work ability is verified and analyzed,and are compared with the BP neural network,and further validation of the method improved the fault recognition effect.a single DAE in dealing with large data automatically and complex data may appear low generalization,studying multilayer deep learning algorithm in order to overcome the limits of a single DAE and enhance the generalization,neural network with different activation functions usually show different characteristics and complementary learning behavior,different activation functions are used as hidden functions to design a series of auto-encoders with different characteristics.Combining deep learning and ensemble learning is for the characteristics of the unsupervised learning.Finally,reconstructive deep learning algorithm was applied to the laboratory planetary gear box of unbalanced data sets,and learn from vibration signals were collected to the depth of the feature,using support vector machine(SVM)for classification was used to fault classification precision and stability.The results show that the reconstruction algorithm is more effective than the available pattern recognition method in the abilities of eliminating the dependence on the artificial feature extraction and overcoming the limitations of the single deep learning model.