Research On Blind Detection And Recognition Of Underwater Acoustic Communication Signals Based On Deep Learning

The blind detection and modulation recognition play an important role in the information recovery of underwater acoustic(UWA)communication signals in non-cooperative reception scenarios.They are of great significance to the exploitation and utilization of marine resources,as well as the improvement of underwater reconnaissance and early warning ability.Conventional test statistics-based methods are always not robust enough under the environment of UWA multipath channels and complex-distribution noise.Deep learning(DL)-based methods can automatically extract deep features,and are less dependent on artificial domain knowledge.However,they require large amounts of data from the target testing channel for training,which are unavailable in non-cooperative reception scenarios.Aiming at the above problems,this paper focuses on the blind detection problem of weak UWA communication signals,as well as the modulation recognition problem under the condition of a small and zero sample set.This work has broken through the key technologies of noise reduction in complex marine environment and DL network construction and learning strategy design under the condition of insufficient or missing training data.It also realized the neural network-based blind detector and modulation recognizer suitable for complex marine environment,which has been verified by the practical signals.The main work of the paper are as follows:1.To conduct the blind detection under the condition of complex marine ambient noise,this paper proposes a joint noise reduction-based method with an impulsive noise preprocessor(INP)and a relativistic conditional generative adversarial network(RCGAN).In this method,the INP with adaptive threshold and the RCGAN are firstly adopted to mitigate the noise in the received signals,which has improved the signal-to-noise ratio(SNR)of the signals.Then a convolutional neural network(CNN)-based binary classification network is constructed to automatically extract detection features and recognize between the UWA communication signals and noise.In order to improve the detection performance of the network under the condition of unknown channel and insufficient training data,a transfer data model is built to effectively reduce the adverse impact of the UWA channel on signal detection.The results of simulation experiments and practical signal tests both prove the effectiveness of the proposed algorithm,which is robust to ambient noise with complex distributions.Its detection performance under low SNR conditions is better than that of existing algorithms.2.Aiming at the scenario where only a small amount of training data from the target testing channel is available,a sample-sample modulation recognition method based on neural network late fusion is proposed.This method first builds an attention aided convolutional neural network(Att-CNN)and a sparse auto-encoder,which are light-weight and efficient to extract features from the temporal waveforms and square spectra of the received signals,respectively.Then a late fusion is made to combine the prediction results of the two networks,and a good recognition result is obtained for several typical UWA communication signals.Moreover,a transfer learning-based two-step training strategy is adopted to resolve the issue of insufficient training data from the target testing channel.The results of simulation experiments and practical signal tests both demonstrate that the proposed method requires less training data from the target channel,and is more robust against different UWA channels.3.Aiming at the scenario where no training data form the target testing channel is available,a zero-sample modulation recognition method based on multi-spectrum fusion is proposed.This method feeds neural networks with multi-category spectrum estimation data,which are more robust to the UWA channel,to reduce the adverse influence of the target channel characteristics on the signal modulation recognition.This makes the network model trained on the simulation data can work effectively even in the unknown channels.Meanwhile,a self-attention mechanism is adopted to evaluate the validity of different categories of spectrum and to perform adaptive weighting.This technique largely alleviates the problem that the overall recognition performance declines when a single spectrum is serious damaged by the UWA channel.The results of simulation experiments and practical signal tests preliminarily prove the feasibility of this method in actual UWA channel environment.