Study On The Uncertainty Of Optical Network Failure Prediction Based On Bayesian Deep Learning

The 5th generation mobile networks(5G)era is still based on optical network transmission and oriented towards 5G intelligent connection and cloud tube edge convergence.The development trend of optical network is increasingly complex and important,which puts forward a great challenge to the stability of the network.Once a failure occurs,will produce huge economic loss and serious business interruption.Before the failure breaks out,failure prediction is helpful to find the hidden danger accurately,determine the cause of the problem and minimize the repair time to realize the users "zero" perception of failure.In recent years,artificial intelligence technology has become the most effective tool for intelligent failure prediction,and gradually replaced the traditional failure prediction method.However,the prediction based on this method is usually a deterministic point estimation,which cannot effectively quantify the model uncertainty,making the confidence degree of prediction results impossible to measure.Aiming at the above problems,this paper proposes an optical network failure prediction framework,which can complete automatic annotation of massive data,failure prediction based on temporal series algorithm,and prediction uncertainty measurement based on Bayesian deep learning,thus achieving reliable prediction of optical network failures.The main innovations of this paper are as follows:Firstly,a data annotation scheme based on unsupervised algorithm is proposed to solve the problem that large-scale data annotation is difficult to be achieved manually.Unsupervised methods can be used to separate healthy and faulty data by actively discovering similarities and correlations between data without training.This paper uses the historical performance data set of live network to verify the scheme,and achieve efficient and accurate data annotation.Secondly,due to the occurrence of optical network failures,the data set is seriously unbalanced,and the equipment monitoring data has a high degree of time series correlation.Therefore,a prediction scheme based on LSTM is proposed.Through modeling and training of historical data set of live network,failure perception and troubleshooting can be realized in advance.Experimental results show that the scheme can achieve 99%prediction accuracy in this data set.Thirdly,although there are a lot of applied researches on artificial intelligence,the "black box" feature of this method makes it lack of quantification of the confidence in results while making predictions,especially when it is used in scenarios with high security requirements.To solve this problem,this paper proposes a failure prediction algorithm based on Bayesian deep learning.This algorithm,on the one hand,is based on LSTM neural network structure and can achieve high precision failure prediction.On the other hand,it can give a measure of the uncertainty of the prediction results by using Bayesian approximate reasoning of Dropout.Experimental results show that the proposed scheme can achieve 99%prediction accuracy in the data set we used,and uncertainty quantification can also indicate error prediction samples and new failure modes effectively.