Research On Wireless Network Failure Prediction Based On Deep Learning With Logs

With the development of social economy and advancement of science and technology,wireless networks provide people with more and more rapid and high-quality services.Wireless networks play an increasingly important role in production and life,and people's dependence on electronic products further promotes their dependence on the wireless network.What follows is a rapid increase of the scale of wireless networks and the difficulty of maintenance.How to effectively use existing resources to predict possible future failures and respond in advance has long been an important research field in wireless network management.However,there are still deficiencies in the current wireless network failure prediction research.Most of the current work uses traditional data mining methods,relying heavily on researchers' understanding of the entire wireless network system,which will encounter much difficulty if applied to large-scale wireless networks.At the same time,in the current wireless network failure prediction field,researchers use more structured data such as alarm data and key performance indicators,and instead of semi-structured data such as logs.Aiming at these two problems,this dissertation designs a log numerical transformation method and two wireless network fault prediction algorithms.Feasibility and effectiveness of the algorithm are verified in experiments.Specifically,the main research contents and results can be summarized as follows:First,learning from natural language processing,we convert logs to matrixes.The wireless network log data is directly treated as a simple text.Through data cleaning,the log data is converted into an "article" composed of words.Using two sliding windows,log samples can be extracted from the log.The sequence in each log sample is a word sequence.Using the word embedding method at the same time,each word sequence can be replaced with a matrix so that the entire log text can be converted into numerical representation that can be calculated.This paper designs two kinds of log data extraction methods for two failure prediction models.Second,after the log sequences in log samples are embedded and converted into matrices,the convolutional neural network is taken advantage of to detect the feature through multiple convolutional layers and pooling layers.The layers automatically extract the features of the log information to analyze the relationship between the current log sequence and future system status,so as to predict possible future wireless network failures.In this process,researchers only need little domain knowledge.Third,with the convolutional neural network based prediction model in the second part,we extract the future log sequence while extracting the failure label from the second sliding window.Using the sequence-to-sequence model to model the relationship between the current logs and future logs,a log prediction model is obtained so that future log sequences can be predicted based on the current log Afterwards,taking use of future log sequences and corresponding failure labels,a failure diagnosis model is obtained based on convolutional neural network to model the relationship between future logs and future system states,so as to achieve the purpose of predicting future failures.