Intelligent Detection,diagnosis And Prediction Methods For Wireless Internet Of Things

With cellular networks and wireless local area networks as the main access mode,Wireless Internet of Things(IoT)can expand the number of terminal devices and provide more flexible networking modes compared with IoT based on the fixed access mode.However,the huge scale and dynamic environment of wireless IoT greatly increase the difficulty and complexity of operation and maintenance(O&M).It is challenging to ensure the normal operation of wireless IoT in a comprehensive,accurate,and efficient manner when using the traditional manual measurement and analysis-based IoT O&M technique.Currently,academics and industry have studied anomaly detection,diagnosis and prediction based on artificial intelligence(AI)methods for wireless IoT,but the performance is still limited by the following reasons:(1)the density distribution of data in wireless IoT is extremely uneven,so that existing AI methods using simple density analysis for anomaly detection cannot obtain local anomaly information,and adjusting the anomaly detection model takes a lot of time,leading to the poor validity of the detection model;(2)the sensing cycle of data in wireless IoT is short,so that the sparsity and distribution differences of data make it difficult for existing AI methods to establish the anomaly diagnosis model,and the lack of labeled data prevents fast anomaly diagnosis,leading to the difficulty of establishing the diagnosis model;(3)data in wireless IoT are bursty and memorable,which results that it is difficult for existing AI methods to predict short-term bursty anomalous traffic accurately,and the prediction duration of exsiting AI methods is limited,leading to the poor adaptability of the prediction model.To address above challenges,according to data characteristics of wireless IoT,this dissertation focuses on anomaly detection,diagnosis and prediction of for wireless IoT by exploiting machine learning and deep learning theories.The main contributions of this dissertation are summarized as follows:(1)Anomaly detection method for wireless IoT based on adaptive twolayer density clusteringIn order to solve the anomaly detection problem for wireless IoT,this dissertation proposes an anomaly detection method based on adaptive two-layer density clustering.Firstly,in order to solve the problem that the density distribution of Key Performance Indicator(KPI)data is extremely uneven,which leads to the inability to find local anomaly information,a two-layer DensityBased Spatial Clustering of Applications with Noise(DBSCAN)method is used to cluster the KPI data twice.Then,the global information and local information of KPI data are mined by the two-layer DBSCAN method respectively,and the anomalies of the network are determined by comparing the density between data points.Secondly,in order to solve the problem that manual parameters adjustment in the DBSCAN method reduces the effectiveness of the anomaly detection model,the Gaussian kernel density estimation method is used to automatically determine the parameters required for DBSCAN.Experiment results show that the proposed method can effectively improve the anomaly detection accuracy,reduce the false alarm rate of anomaly detection,and reduce the cost of manual parameter adjustment.(2)Anomaly diagnosis method for wireless IoT based on deep transfer learning neural networksIn order to solve the anomaly diagnosis problem for wireless IoT,this dissertation proposes an anomaly diagnosis method based on deep transfer neural networks.Firstly,in order to solve the problem of sparse anomalous KPI data and large difference of feature distribution,which makes it difficult to establish anomaly diagnosis model by utilizing existing AI methods,the convolutional neural networks(CNNs)to extract high dimensional features of KPI data,and a maximum mean discrepancy(MMD)module is constructed based on high dimensional features to achieve the goal of domain feature adaptation by minimizing the distance of feature distribution of KPI data.Thus,sufficient KPI data with differentl feature distribution can be used to assist sparse abnormal KPI data to build an anomaly diagnosis model.Secondly,in order to solve the problem that unlabeled KPI data cannot be used for quickly anomaly diagnosis,virtualized labels are produced for unlabeled data by using the generator of domain adversarial neural networks,and the goal of domain feature transferring is achieved by making the discriminator unable to distinguish whether the extracted features belong to real or virtual labels,so that labeled KPI data can be used to classify unlabeled KPI data.Experiment results show that the proposed method can not only solve the problem that existing AI methods are difficult to build anomaly diagnosis models,but also can effectively improve the accuracy of anomaly diagnosis.(3)Anomaly prediction method for wireless IoT based on integrated learning neural networksIn order to solve the anomaly prediction problem for wireless IoT,this dissertation proposes an anomaly prediction method based on integrated learning neural networks.Firstly,in order to solve the problem that it is difficult to accurately predict short-term abnormal traffic using single-dimensional time series data,multi-dimensional historical time series data are utilized to perform similar time clustering and construct a training data set with features closer to those of the prediction date to solve the problem of low prediction accuracy due to the lack of information of abnormal traffic characteristics in short-term historical time series.Secondly,in order to solve the problem of limited prediction steps using multidimensional time series data,a single-dimensional time series prediction method based on the combination of convolutional neural networks and long short-term memory networks(LSTM)is designed.The CNNs can extract the spatial features of time series,and the LSTM can extract the continuity and memory features of time series,which can increase the prediction step size and achieve the goal of long-term prediction.Finally,the integrated learning method is used to select the optimal model among multiple prediction models to improve the stability and accuracy of the prediction method.Experiment results show that the proposed algorithm can obtain more accurate prediction results in both short-term and long-term prediction compared with existing AI methods.Based on the data characteristics in wireless IoT,this dissertation studies challenges of anomaly detection,diagnosis and prediction for wireless IoT.The proposed methods can effectively overcome the shortcomings of existing AI methods applied in wireless IoT,improve the performance of existing AI methods,and provide more intelligent and proactive O&M ideas for wireless IoT.