Research On Wireless Communication Interference Identification Technology Based On Deep Learning

Wireless communication systems are vulnerable to electromagnetic attacks and interference.Improving the security and stability of wireless communication systems is a growing concern in modern wireless communications.The wireless interference signal identification technology is a key technology for effective understanding and sensing of the spectrum environment and situation by identifying the type of interference signal without any a priori information.In civil communications,wireless communication interference signal identification technology is widely used in spectrum monitoring,cognitive radio,interference management,etc.In addition,in military communications,wireless communication interference signal identification technology is used in electronic countermeasures,secure communications and analysis of intercepted signals.Traditional techniques for identifying interference signals include Maximum Likelihood(ML)judgement methods and Feature Extraction(FE)based methods,however,ML methods require complete and accurate channel state information,which is not only difficult to implement,but also computationally complex;while FE method requires tedious feature engineering and expert knowledge,and the recognition accuracy cannot meet the requirements of high-precision recognition.To this end,this dissertation investigates the key technology of wireless communication interference recognition based on deep learning.First,the technique of constructing wireless communication interference recognition network based on deep learning is investigated,and the global convolutional neural network architecture is proposed to solve the limitation problem of local extraction of convolutional neural network.Then,the interference recognition method of multi-domain network is proposed,and the multi-domain information of interference signal is adopted as the input of the network,which effectively improves the recognition accuracy.Next,a low-complexity interference recognition technique based on deep learning is investigated,and an adaptive forward propagation algorithm,network global pruning method and time-frequency-aware convolutional neural network are proposed to reduce the computational cost of the neural network.Finally,the quantization technique of interference recognition network based on deep learning is studied,and the low-bit quantization and binary network quantization algorithms of interference recognition network are proposed,which enable the network to complete the recognition task when quantizing the weights and activation functions simultaneously,and the recognition accuracy has almost no degradation compared with the full precision network.The main research contributions and innovations of this dissertation are as follows:1.To address the problem that the interference recognition performance of convolutional neural networks is limited by the local extraction,we propose a wireless interference signal recognition algorithm based on Bring Globality into Convolutional Neural Networks(BGCNN).The BGCNN incorporates the Transformer while retaining the convolutional layer,allowing the network to capture both local and global features,and the proposed network can learn highly expressive features.In addition,to address the huge computational overhead caused by the self-attentive module in the Transformer,this dissertation proposes the Wireless Interference Recognition Transformer(WIR-Transfer),the WIR-Transformer introduces a window division mechanism to effectively reduce the complexity of multi-headed self-attention and introduces inter-window information interaction and block aggregation mechanisms to improve recognition performance.2.In order to effectively improve the recognition accuracy of interference recognition networks based on deep learning,a method of interference recognition based on Multi-domain Networks(MDN)is proposed to overcome the problem of high computational complexity and low recognition performance caused by most existing deep learning algorithms that only use a single domain information as the input of the network and neglect to make full use of the information of multiple transformation domains.In this dissertation,we propose to use the multi-domain information of the interference signal as the input of the network to make full use of its complementary,so that the network can use different domain information to give more reliable recognition results.Specifically,the time-frequency image(TFI)and the frequency sequence(FS)of the interference signal are extracted from the time-frequency and frequency domains and used as input to the network.For these multi-domain information,this dissertation uses MDN to capture both the TFI and FS features of the interference signal.In addition,in order to merge the extracted features of the two domains,three new fusion mechanisms are proposed,which greatly improve the performance of interference signal recognition.3.To address the problem of high computational complexity based on deep learning,a low-complexity interference recognition algorithm based on deep learning is proposed to overcome the problem of significant degradation in recognition performance when reducing the computational complexity.Firstly,starting from the redundancy of the network in deep learning-based interference recognition,the Adaptive Forward Propagation(AFP)algorithm is proposed,which enables the deployed network to allocate appropriate computational resources according to the difficulty of the input signal and allows samples to stop early in the forward propagation process of the network.This allows the network to achieve a better balance of accuracy and complexity.For the AFP algorithm,three adaptive propagation control mechanisms are proposed,including a confidence mechanism,an Interference to Noise Ratio(INR)estimation mechanism and a reinforcement-based learning mechanism.Secondly,a method for measuring redundant neurons is proposed,and a global pruning network is constructed based on this method to reduce the computational complexity of the deep learning-based interference recognition method.Finally,the Time Frequency Component-aware Convolutional Neural Network(TFCCNN)is proposed.The network is built with a time-frequency component-aware module,so that the network only calculates the location of the time-frequency components in the input time-frequency image and ignores the unimportant regions,thus further reducing redundant computations.4.To address the problem of quantization and hardware deployment of interference recognition based on deep learning,a quantization algorithm for interference recognition network based on deep learning is proposed,which enables the network to complete the recognition task when quantizing both weights and activation functions,and the recognition accuracy has almost no degradation compared with that of the full-precision network.For low-bit quantization networks,this dissertation proposes the method of Guide Training with Full-precision Network(GTFN),which enables the quantization network to learn from the full-precision network.To further reduce the training cost,this dissertation proposes Guide Training without Full-precision Network(GTWFN)to assist in training the quantization network,which provides a virtual full-precision network using an artificially designed probability distribution to guide the training.In addition,to make gradient back-propagation easier,this dissertation proposes an Auxiliary Output Module(AOM)to help train the quantile network,which creates multiple paths to update the parameters of the network and effectively improve the recognition accuracy.The binary quantization is also investigated in this dissertation.In order to overcome the difficulty of gradient propagation in the backpropagation,a new Approximation of Gradients(AG)method is proposed,which bridges the accuracy gap between binary networks and full-accuracy networks.In addition,to address the bottleneck problem of severe performance degradation in binary networks,this dissertation proposes a Minimizing Quantization Noise(MQN)loss function and an AOM-assisted training method to further improve performance.The key technology of wireless communication interference identification based on deep learning proposed in this dissertation can provide a solid theoretical foundation and strong technical support for spectrum sensing and electromagnetic interference identification,and it has high guiding significance and research value.