Research On Fault Diagnosis Technology Based On Deep Learning

With the growing development of China’s manufacturing industry,the industrial field has gradually become automated,large and systematic.In the process of unmanned factory,the composition of machinery and equipment has become more and more complicated,functions have become more and more perfect,and the safety of equipment has been receiving increasing attention.Fault diagnosis is the most commonly used method for feature extraction from vibration signals collected by mechanical equipment.However,today’s development of mechanical equipment tends to be high-precision,high-speed,and high-efficiency,accompanied by continuous development of data acquisition and storage technologies.The fault signal gradually exhibits the characteristics of "mechanical big-data".Traditional fault diagnosis methods are difficult to deal with massive fault data.The deep learning algorithm is a branch of artificial intelligence because of its multi-hidden layer network and adaptive feature extraction capability.The ability to mine the essential characteristics of the data and use all the characteristics of the original signal without discarding the original data information accurately characterizes the complicated mapping relationship between the observed data and the fault category compared to the traditional method.This paper deals with the failure based on deep learning.Diagnostic techniques are studied.Firstly,from the principle of Deep Belief Nets(DBN),the use of standard handwritten digit sets for DBN restrictions The feature extraction capabilities of the Restricted Boltzmann Machine(RBM)part and the network fine-tuned by BP The classification ability is studied,and the influence of the number of hidden layer nodes,the learning rate and the number of iterations on the feature extraction ability is analyzed through experiments,and the setting method of the main parameters is determined,which lays the foundation for the DBN-based fault diagnosis technology.Then,the application process of fault diagnosis based on DBN method was studied,the computational ability of data sets under different lengths of divided samples was studied,the problem of no basis for manual adjustment of parameters was improved through parameter optimization methods,and the world recognized failures.Diagnostic field standard bearing data sets-Case Western Reserve University bearing data sets are calculated and compared with other methods to obtain higher accuracy.Based on this,aiming at the problem of excessively long training time in DBN network,a distributed parallel computing platform is set up,and the computer cluster distributed computing environment of Matlab computing environment is set up,and the process of DBN parameter optimization is validated.The parallel computing environment can effectively improve the computational efficiency of DBN network training.Finally,based on the Java web development fault diagnosis system,a preliminary model of online fault diagnosis was established,and the real-time uploaded data can be obtained according to the diagnostic model to obtain the equipment status and visually presented,and a preliminary exploration was made for practical application in the future.