Research On Intelligent Recognition Technology For Modulation Patterns In Underwater Acoustic Communication

Underwater acoustic communication is crucial in various applications,such as ocean resource exploration,marine environment monitoring,and underwater robot communication.However,underwater acoustic communication faces numerous challenges due to the unique characteristics of the underwater acoustic channel,including high noise,strong multipath effects,and limited bandwidth.Modulation scheme recognition is one of the key technologies in underwater acoustic communication systems.Adaptive modulation,interference suppression,and signal detection can be achieved by accurately identifying the modulation scheme,thereby improving the reliability and robustness of the underwater acoustic communication system.Traditional modulation scheme recognition methods mainly include feature extraction and statistical model-based methods.These methods have some drawbacks when applied to underwater acoustic communication modulation scheme recognition,such as difficulties in feature extraction,high model complexity,and poor adaptability.Moreover,traditional methods struggle to effectively deal with strong noise and multipath interference in the underwater acoustic channel,resulting in limited recognition performance.Therefore,it is imperative to explore new underwater acoustic communication modulation scheme recognition methods to overcome the shortcomings of traditional approaches.In recent years,deep learning methods have achieved remarkable results in the field of communication modulation scheme recognition.Compared with traditional methods,deep learning methods have advantages such as automatic feature learning,strong adaptability,and high robustness.Deep learning methods can automatically learn and extract high-level features of underwater acoustic signals by constructing deep neural network models.Furthermore,deep learning methods possess strong nonlinear fitting capabilities,enabling them to effectively handle complex noise and multipath interference in the underwater acoustic channel,thus improving recognition accuracy.Consequently,applying deep learning methods to underwater acoustic communication modulation scheme recognition has broad research prospects.This thesis conducts in-depth research on the various problems mentioned above and proposes a modulation recognition algorithm for underwater acoustic communication based on deep learning,which effectively improves the classification accuracy,robustness,and computational efficiency of communication modulation recognition.The main research work of this paper includes the following three aspects:Firstly,a multi feature joint recognition algorithm based on graph convolutional neural network,GCNN-MFS,is proposed to address the problem of similar modulation modes with similar modulation orders in underwater communication signal modulation recognition tasks,which leads to a decrease in recognition performance of deep neural networks.This algorithm first takes the image representation of the signal as the input of the neural network,and uses transfer learning to train a Caffe Net network suitable for modulation recognition of underwater acoustic communication signals.At the same time,two-stage expert features combined with support vector machines are used to further classify modulation modes with autocorrelation spectral peak properties.Subsequently,the performance of different image representations and the generalization ability of neural networks were validated on the Chi Sou Sea and QALake datasets,respectively.The generalization ability of this dataset for underwater acoustic signal recognition scenarios was demonstrated on the UASimul dataset,and the performance of the GCNN-MFS algorithm was analyzed.The experimental results show that among the three image representation methods,the input based on equipotential constellation diagram has the best modulation representation ability.The network trained on the simulated underwater acoustic channel dataset can perform well in classifying the measured underwater acoustic signals,and GCNN-MFS shows significant performance improvement compared to pure Caffe Net networks.Secondly,in response to the high computational complexity of deep neural networks and the difficulty in deploying them on end-to-end underwater acoustic signal recognition edge devices,lightweight research has been carried out in the framework of deep learning algorithms for underwater acoustic signal recognition.Firstly,for convolutional neural network models based on equipotential constellations,the Taylor expansion pruning criterion is introduced to measure the contribution of each convolutional kernel to recognition.Then,the smaller contribution convolutional kernel is removed to complete the compression of the recognition model.Finally,in conjunction with ensemble learning strategies,compensate for the performance loss caused by network pruning.The experimental results based on the UASimu dataset show that the proposed lightweight network framework based on ensemble learning effectively reduces the number of network parameters and computational complexity while maintaining the recognition ability of the original network.In addition,the lightweight model was deployed on edge devices for testing,and the experimental results fully demonstrated the lightweight ability of the proposed method for large neural networks and the effectiveness of underwater acoustic signal recognition.Thirdly,two lightweight convolutional neural networks for underwater acoustic signal recognition tasks are proposed to address the complex structure and difficulty in parallelization of long short-term memory models.They are the time series convolutional neural network GIQNet based on gated channel mixing and the reparameterized causal convolutional network Rep CCNet.The two proposed frameworks combine the advantages of easy parallelization of non causal convolutional neural networks,as well as the temporal dependencies of long short-term memory models.The two networks,based on the characteristics of underwater acoustic signal recognition tasks,adopted different functional modules to achieve high-performance recognition algorithms with lower network complexity.The experimental results show that the two lightweight networks not only achieve good recognition performance on complex QDLake underwater acoustic signal datasets,but also achieve higher recognition accuracy on the publicly available radio reference dataset RML2016.10 A compared to existing methods.At the same time,the number of model parameters is lower and the computational complexity is smaller.Rep CCNet has an advantage in parameter quantity,while GIQNet outperforms the former in FLOPs.