Research On Deep Generative Variational Bayes Network For Fault Diagnosis

Modern mechanical systems has undergone a qualitative leap in degree of complexity,precision and intelligence.It has been more and more closely coupled with electrical systems,hydraulic systems,control systems,information systems,etc.This putting forward more rigorous standard on the stability and reliability of mechanical systems.And the fault monitoring and diagnosis technology is facing greater challenges.In this paper,we aimed to solving the ubiquitous problem of fault monitoring and diagnosis method based on artificial intelligence which has weak theoretical support,strong dependence on feature extraction,high dependence on training data,low utilization rate of current deep artificial intelligence model and poor universality in practical engineering application.Therefore a fault monitoring and diagnosis model which integrating variational optimization,bayesian probability network,deep neural network,normalizing flow transformation and related algorithm of sampled data estimation is proposed by us.The main research work of this paper is as follows:(1)The main limitations of existing fault monitoring and diagnosis methods are analyzed and the main reasons are discussed by us.We introduced some successful generative models in the field of artificial intelligence and analyzed the most theoretical,stable and representative model of Variational Auto-Encoder in detail.We focus on four points of VAE including generative structure of implicit variables,variational posterior inference method,variational lower bound on the marginal likelihood and stochastic gradient variational bayes evaluation algorithm,which are the theoretical foundation of our fault monitoring and diagnosis method based on deep intelligent model.(2)We derivated the vibration signal state evaluation method based on single generative structure of implicit variables in VAE.And the principle of state evaluation based on the training set of observation data which is vibration spectrum and marginal probability distribution of vibration spectrum is proposed by us.We also proposed a denoising method based on VAE in terms of information entropy for the low signal-to-noise ratio of vibration signal and confirmed it’s effectiveness and superiority by an experiment with simulated and real vibration signal.Furthermore,we proposed a calculation method of marginal probability density in VAE with Annealing Importance Sampling algorithm for the lack of accurately enough calculation method of marginal probability density in VAE.Combined with the above research,we proposed our vibration signal state evaluation model which named AIS-VAE.Finally,we conducted an experiment of bearing state evaluation in it’s whole degradation life to prove the effectiveness of our model.(3)We derivated the modeling method of joint probability distribution of vibration spectrum and the failure when AIS-VAE faced with multi-fault vibration signal,and proposing an fault monitoring and diagnosis model with semi-supervised variational bayes network based VAE.Following this we proposed the VAE-BGN model which has the ability of fault diagnosis by extension the single generative structure of implicit variables in AIS-VAE to bayes generative network which has joint posterior inference.For the generally fault label missing in practical engineering,we integrated explicit classifier to posterior inference network and anew derivated the joint ELBO and semi-supervised AEVB algorithm(S-AEVB)of VAE-BGN which can suitable for semi-supervised learning.For the low accuracy of variational inference of multi-fault vibration signal with gaussian posterior distribution which lead to the problem of transition fuzzy aliasing of posterior distribution boundary in different vibration signal,we integrated normalizing flow transformation to our model so as to realize arbitrarily complex and flexible variational inference,then we confirmed that the accuracy of variational posterior inference has been greatly enhanced through normalizing flow transformation by conducting an experiment.For the lacking of optimal classification of the fault signal in theory by explicit classifier in VAE-BGN as the explicit classifier always act as a role of posterior optimization,we seek to use the gibbs sampling algorithm in VAE-BGN as the optimal implicit classifier.Finally,we contrastively analyzed and tested the performance of our VAE-BGN model through a gear box fault diagnosis experiment.