Research And Design Of Automatic Modulation Recognition Technology Scheme Based On Machine Learning

Automatic modulation identification is a very critical technology in non-cooperative communication systems.It is widely used in many fields,such as military and civilian fields.Specifically,when some or all communication parameters are unknown,it can identify the modulation type of the received signal.It is helpful for related signal processing work such as subsequent signal demodulation,and finally improves the performance of the communication system.With the increasingly complex and changeable communication scenarios,additive white gaussian noise assumed in many previous studies no longer satisfies the accurate description of the actual channel environment.In practice,due to the high-speed movement of modern vehicles such as planes and trains,the signals will be affected by the Doppler effect when users communicate in them,and the channel environment presents a characteristic that changes regularly with time.Moreover,due to practical reasons,such as lightning,thunderstorms,multi-user interference,and equipment failures,there are random and varying degrees of sharp impulse noise in the channel environment.Therefore,this paper studies the automatic modulation identification problem under time-varying channel and impulse noise respectively.The main contents are as follows:(1)For the Rayleigh fading time-varying channel,this paper proposes the Grassmannian manifold feature extraction method for the first time.First,it is completed by modeling a set of time-correlated slot constellations on the Grassmannian manifold to extract manifold features,secondly,using a manifold learning network with a three-layer structure of a full-rank mapping layer,an orthogonal layer,and a projection mapping layer to reduce the dimensionality of the manifold features and complete the mapping from the manifold space to the euclidean space.Using a simple convolutional neural network consisting of one convolutional layer to complete the classification task.Compared with traditional convolutional neural networks and slotted feature fusion networks in other literatures,the specific performance of the proposed method has been effectively verified by experiments.(2)For Alpha stable distribution impulse noise,this paper proposes a light-weight network-based identification method.The first part uses a signal preprocessing operation including logarithmic domain mapping and threshold processing to effectively suppress Alpha stable distribution noise and limit the signal points to a reasonable range.The latter part is constructed by a three-layer Ghost unit structure with output channels 16,32,and 8 based on the Ghost Net idea combined with a common convolution layer to complete the classification task.The experimental results show that,compared with the existing convolutional neural network and residual network,the method proposed in this paper has higher accuracy and lower computational complexity.