Radio Signal Detection And Recognition Based On Semi-supervised Learning

With the rapid development of wireless communication technology,the modulation modes and channel environment of signals have become increasingly complex.The effective detection and recognition of radio signals attracted intensive research interest in the field of wireless communication.However,most of current radio signal recognition techniques are based on time-domain or frequency-domain characteristics of signal,and signal parameters are extracted one by one.Moreover,the recognition performance is affected by signal-to-noise ratio seriously.A radio signal detection and recognition framework based on signal timefrequency spectrum characteristics is designed in this work.The object detection method with deep learning is introduced in order to overcome the time-consuming and laborious problem of signal sample tagging.In addition,semi-supervised learning is adopted to reduce the cost of manual tagging.Several research achievements and contributions are summarized as the followings:Firstly,the task of object detection is introduced into the radio signal detection and recognition,and the signal is converted into the form of time-frequency images through shorttime Fourier transform.Secondly,an improved Center Net-based object detection network is proposed to solve the problem of slow detection.A learnable residual structure is added to adjust the branch weights,and multi-scale pyramid convolution is utilized to improve signal detection performance without increasing time overhead.Besides,the deformable convolution is leveraged in the decoding network to expand the network's receptive field.Finally,the pyramid feature fusion structure is employed in the detection stage,while the multiple fused features of different scales are adopted for the final prediction.According to the experiments results,the time-frequency graph signal detection based on the improved Center Net can increase the detection speed by more than 4 times compared with the original network,while the recognition accuracy is reduced by only 0.3 %.Semi-supervised signal detection and recognition approach significantly reduces the demand of manual labeling.Due to the scarcity of labeled signal data,three semi-supervised signal detection methods are adopted to generate unlabeled training data for better accuracy.Among them,the pseudo-label-based method improves the accuracy when the accuracy of the reference model is high,but it has certain limitations.The Teacher-Student framework slightly improves the accuracy of signal detection under the same amount of signal data.The collaborative consistency training method also reduces the amount of signal annotations with maintaining accuracy.Therefore,the semi-supervised signal detection method can be leveraged to reduce the cost of manual labeling and bring significant improvement to signal detection and recognition tasks in practical applications.