Research On Unmanned Aerial Vehicle Anomaly Detection Based On Improved TCN

With the development of information technology,drones have gained attention from various countries due to their cost-effectiveness and convenience.They have gradually transitioned from military drones to civilian drones,and have made progress in multiple fields.However,with the widespread application and development of drones,the frequent occurrence of drone faults has also attracted the attention of domestic and foreign researchers.Drone data is collected through sensors in the form of time series,and the drone flight dataset has the characteristics of being large and complex.For the processing of drone flight datasets,relying solely on expert experience is not enough to accurately diagnose drone anomaly types and make effective maintenance processing.In recent years,deep learning research based on neural networks has made significant progress in the field of anomaly detection technology and has gradually become a key technology for anomaly detection.Therefore,for time series with a large and complex amount of data such as drones,whether the feature values of the time series can be effectively extracted has become the key to drone anomaly detection.Compared with recurrent neural networks,temporal convolutional neural networks(TCN)can effectively avoid problems such as gradient explosion,gradient vanishing,and lack of memory retention when dealing with large amounts of time axis information.This article proposes a research on unmanned aerial vehicle anomaly detection based on temporal convolutional neural networks,which have the characteristics of memorizing historical information and adapting to sequence models.Firstly,this article proposes three types of common anomalies in unmanned aerial vehicles:zero drift,random point jamming,and constant gain ratio,and completes anomaly injection on the simulated normal dataset.Design two types of experiments for time series anomaly detection,namely classification and regression,based on the common problems and abnormal datasets of time series anomaly detection.This article conducts experimental research on the classification problem of drone data anomaly detection based on temporal convolutional neural networks.Based on the spatiotemporal characteristics of temporal convolutional neural networks,a channel attention mechanism was designed to extract the output results of each hidden layer of the neural network,and global pooling operations were performed on the obtained output results.By using attention mechanism to self allocate weights to the output results of each hidden layer,the network can focus on more important information in the anomaly detection dataset,thereby improving the accuracy of drone anomaly detection classification by 2.8%.After completing the classification experiment of drone anomaly detection,this article designs anomaly detection regression experiments for three types of anomaly datasets,enabling drones to perform real-time detection of flight data collected by sensors during task execution,and complete anomaly judgment and regression.In response to the problem of information loss caused by hollow convolutional features,the feature extraction module of temporal convolutional neural networks has been improved by introducing dense connections,improving the efficiency of information exchange between channels,enhancing feature reuse,and thereby improving the R2 parameter accuracy of drone anomaly detection regression problems by 2.2%.