Outlier Detection Methods Based On Top-k(σ) Algorithm And Neural Network In Sensor Networks

Wireless sensor network can real-time monitoring, sensing and collecting all environmental information in the network distribution area. When does the wireless sensor network impact by external events(such as forest fires, air pollution, etc.), or the sensor node itself fault(such as low battery, electromagnetic interference, etc.), it must timely and accurately detect the change of data stream from sensor nodes. It is great significance both for the external emergency warning prevention and sensor network health monitoring itself. In this paper, we studied and discussed outlier detection in wireless sensor networks(WSNs). The main research results are given as follows:1. After improving the top-k algorithm, a top-k(σ) outlier detection algorithm for WSNs was proposed in this paper. Different from top-k algorithm, the proposed algorithm uses the data distribution collected by the sensor nodes to construct appropriate data grid, and puts the data sets into the grid after normalization, then sets a distance threshold σ to reconstruct the PC list(populated-cells list). This algorithm sorts the numbers of data points in each cell and those of its neighborhood respectively, as well as computes the Euclidean distance R_D between two data subsets, and compares the value of R_D with σ so as to verify the degree of deviation of the subset from the normal data sets. Thus the top-k(σ) algorithm can improve the precision of the outliers detection.2. Two outlier detection methods for WSNs based on BP neural network and linear neural network in this paper. Latest historical data with fixed length of data window is used to train a neural network model, and then these methods can predict the sensor data of the next time. A confidence interval with probability p is calculated with the help of the model residual. The new measurement will be identified as normal one if it falls inside the prediction interval. Otherwise, it will be classified as an abnormal record.3. The design of the relevant indicators to assess performance of the algorithm. For given several datasets, the simulation results under MATLAB platform show that, the threshold σ has great effect on the performance of outlier detect algorithm. When σ?[2.5, 3], the top-k(σ) algorithm have higher detection accuracy and lower false positive rate. If σ = 3, for the given five data sets, the average accuracy of outlier detection of top-k(σ) algorithm is 93.70%, which is 4.94% higher than that of top-k algorithm, and the average false positive rate of top-k(σ) algorithm is 4.48% lower than that of top-k algorithm.4. In order to compare and demonstrate the performance of the proposed methods, we finished the simulation experiments in MATLAB environment. The experiment results showed that the detection rate of outlier detection based on linear neural network reached 97.9%, and the false positive rate is less than 0.76%. While the detection rate of outlier detection based on BP neural network reached 96.7%, the false positive rate is less than 0.84%.