Research On Spectrum Anomaly Detection Of Tiantong-1 S-band Downlink Based On Object Detection

The development of wireless communication services has made the spectrum resources more and more tense,but the existing spectrum utilization efficiency is not high,and the dynamic spectrum resource allocation strategy is considered to be an effective solution to the above-mentioned problems.However,compared with static allocation,the radio usage environment under dynamic resource allocation is more complicated,and the problem of mutual interference between signals is more prominent.Especially in the satellite communication system,because the satellite is located far away from the ground,the communication distance is long,the transmission loss is large,and due to the limited spectrum resources,the satellite frequency band is getting closer and closer to the ground mobile communication frequency band,and the mutual interference situation is more severe..In this regard,this article builds two spectrum anomaly detection algorithms applied in different scenarios.First of all,in a scenario where only a small range of the spectrum is needed for anomaly detection.Aiming at the problem of using supervised learning for anomaly recognition tasks that require a large amount of labeled data and limited computing resources on the satellite,this paper uses a lightweight fully connected neural network to achieve anomaly warning and classification,and we can get rich through the spectrum monitoring system.Model training is performed on experimental samples of,and a generative adversarial architecture is introduced during training to achieve semi-supervised learning.The unsupervised learning autoencoder is used to extract the spectral features,and the classifier structure is trained with label data accounting for 1/4 of the total sample size.A good spectrum anomaly warning and classification is achieved with a small number of label data sets,and a correct 97.3% detection rate is achieved when the false alarm probability is16%.For anomaly detection tasks in a larger frequency band,it is difficult for the existing anomaly detection methods to uniformly identify the location of the frequency point of the anomaly signal,the range of the frequency band that it affects,the intensity of interference,and the category of information.This paper uses the method based on image target detection for anomaly detection according to task requirements.For the specific scenarios of spectrum anomaly monitoring,we optimized and improved the detection model: preprocessing the spectrum image morphology;constructing a feature extraction network based on the concept of the receptive field for the scale of anomalous targets;using one-dimensional anomalous target prediction The module replaces the traditional grid target prediction.The redundancy of the model is reduced to a large extent,a nd excellent detection performance is achieved.The m AP value reaches 92.81%,and the detection probability is not less than 93.94% when the comprehensive false alarm probability is 7.64%.It can meet the needs of actual inspection tasks.