Unsupervised WSN Data Anomaly Detection Method Based On LSTM

Wireless Sensor Networks(WSN)is the main way to collect data from the perceptual world.Several sensors work together to sense and collect information in the target area from the physical world.However,there are many challenges for anomaly detection of data acquired from wireless sensor networks.Firstly,the anomaly detection algorithm needs to have low algorithm complexity and memory space.Secondly,the data collected by wireless sensor network nodes are large in scale and high in dimension,which greatly reduces the efficiency of data anomaly detection.In addition,the samples collected by nodes are generally unlabeled and the data need to be labeled manually.Therefore,the method of supervised learning cannot be used well due to the high cost of labeling.Aiming at the anomaly detection of wireless sensor Network data under the unsupervised framework,this paper proposes an unsupervised detection method based on Long Short-Term Memory Network(LSTM):LSTM-SVDD(Support Vector Data Description,SVDD).The benchmark dataset Traffic Dataset,SWAT and WADI were used for experimental simulation and verification.The experimental results show that the proposed LSTM-SVDD algorithm has significantly improved the detection accuracy compared with the traditional anomaly detection method,and has practical application value.The main work of this paper is as follows:(1)Aiming at the problem of lack of labels and high dimensions of data,an unsupervised framework anomaly detection algorithm based on LSTM is proposed.LSTM is used to analyze and capture the dependence between sequences and extract the corresponding key information,which can assist the realization of the next single classification task,so as to effectively improve the accuracy of anomaly detection of wireless sensor network data.(2)Aiming at the online anomaly detection problem,a simplified anomaly detection objective function is defined to combine the training model parameters,so that the model can accelerate the training speed on the basis of extracting the significant features from the time series data,so as to realize fast and efficient anomaly detection and avoid the memory occupation caused by data storage.In addition,based on the generality of LSTM architecture,another variant GRU is introduced and compared with the algorithm based on LSTM framework.According to the results,the overall performance of the LSTM-SVDD algorithm in this paper has certain advantages.