Research On Abnormal Data Detection Method Of Agricultural Wireless Sensor Networks

Agricultural wireless sensor network has become one of the most important data sources of agricultural big data.However,due to the influence of manufacturing process,network transmission and human interference,single or multiple sensor nodes in wireless sensor networks will inevitably produce anomalies in the process of data acquisition.In sensor node,screening the outliers in the collected agricultural data can ensure the efficient development of agricultural data analysis.At the same time,in order to obtain the real-time data about farms,it is necessary to introduce the real-time anomaly detection method to process the real-time data stream from sensors.On the other hand,in large-scale wireless sensor networks,how to detect simultaneous abnormal multiple sensor nodes and reduce the consumption of computing resources is an urgent problem to be solved in the application of agricultural wireless sensor networks.In order to solve the above problems,this paper proposed off-line and real-time anomaly detection methods of sensor data and real-time anomaly detection methods of multi-sensor data,taking agricultural data such as temperature and humidity of air and soil as objects.The purpose of this paper is to provide reference for improving the quality of perceived data from wireless sensor networks in agriculture.The main works include:(1)Off-line anomaly detection of sensor data based on convolution neural network.In view of the abnormal problem in the sampling data of sensor nodes,the idea of image recognition was introduced into the field of anomaly detection of agricultural data,and an off-line anomaly detection method of sensor data based on convolution neural network was proposed.The discrete sensor sampling data were drawn as continuous data of images in segments.The sensor data images data set was constructed by the data augmentation method of image rotation,the convolution neural network was constructed to classify the images with outliers,and off-line anomaly detection of sensor data was realized.The experimental results show that this method achieved average accuracy of 0.9457 in anomaly detection,and with the increase of the segment length of sensor data,the accuracy of anomaly detection tends to decrease.(2)Real-time anomaly detection method of sensor data based on improved convolution neural network.In order to the lack of real-time performance and high consumption of detection time of off-line anomaly detection,a real-time anomaly detection method for sensor data based on improved convolution neural network was proposed.The standardized sensor data in agricultural were transformed as polar coordinates,the sliding window mechanism of adaptive data feature was introduced to divide the polar coordinates data into a subset and the data was reconstructed as matrix format,and the real-time data stream was processed by iterative updating of the sliding window.Broad learning model was used to optimize traditional convolution neural network,the improved convolution neural network model was used to realize anomaly detection.Comparative experiments with support vector machine,random forest and convolution neural network show that the proposed method was improved performance of real-time detection,and achieved accuracy of 0.9854 in anomaly detection.The proposed method has more advantages when handling the data with high fluctuation.The anomaly detection time of proposed method lower than traditional convolution neural network,it could meet the needs of sensor node in real-time data anomaly detection.(3)Real-time anomaly detection method for multi-sensor data based on broad learning.Aiming at the scenario of simultaneous anomalies of multiple sensor nodes in agricultural wireless sensor networks,a real-time anomaly detection method of multi-sensor data based on broad learning was proposed.The observation vector was constructed based on the discrete multi-sensor real-time sampling data,the observation vector was constructed as the adjacent difference vector by the adjacent element difference method,and the broad learning model was used for real-time anomaly detection.The experimental results show that the detection accuracy of the proposed method decreases slightly with the increase of the number of abnormal sensors.The proposed method achieved average accuracy of0.9922 and low time of real-time anomaly detection in scenarios where 10% to 50% of the sensors were abnormal at the same time.