CNN convolutional neural network-based digital signal automatic modulation identification method

Abstract

The invention belongs to the technical field of digital communication signal modulation, and discloses a CNN convolutional neural network-based digital signal automatic modulation identification method. The method comprises the following steps of performing cyclic spectrum analysis on the received digital signal, extracting an axial projection profile map of an amplitude normalization cyclic spectrum, and obtaining a one-dimensional feature vector x belonging to Rn * 1; Carrying out dimension reduction processing on the feature vector by utilizing an improved Fisher algorithm to obtain a low-dimensional feature vector y belongs to xm * 1; And finally, designing a deep CNN network structure, and determining network initialization parameters. According to the method, a Keras deep learning framework is utilized, an existing network layer function is directly called, and a deep network structure is built; An early stop strategy is adopted in the network training process, the network overfitting phenomenon is effectively prevented, after network training is completed, the training effect is verified through a test data set, and automatic signal modulation recognition is completed. For an MQAM signal, when the signal to noise ratio is greater than 0dB, the recognition rate reaches 97% or above; For an MPSK signal, the signal to noise ratio is greater than or equal to-. And when the identification rate is 4dB, the identification rate reaches 95%.

Description

technical field [0001] The invention belongs to the technical field of digital communication signal modulation, and in particular relates to a digital signal automatic modulation recognition method based on a CNN convolutional neural network. Background technique [0002] At present, the existing technologies commonly used in the industry are as follows: Aiming at the modulation recognition problem in the AWGN environment, the existing methods can be roughly divided into three categories: based on maximum likelihood hypothesis testing, based on feature extraction pattern recognition, and based on depth study method. Among them, the identification method based on the maximum likelihood hypothesis test involves a lot of prior information such as signal mean, variance, covariance, etc., which are difficult to obtain accurately in non-cooperative communication. Therefore, many scholars now focus on the modulation recognition method based on feature extraction and deep learning....

AI Technical Summary

Problems solved by technology

[0004] (1) The method based on the maximum likelihood hypothesis test involves a lot of prior information such as the signal mean, variance, covariance, etc., and these in non-cooperative communication, because the sending and receiving ends do not know in advance, after complex channel effects Afterwards, the prior information has changed a lot, so it is difficult to obtain it accurately; and the hypothesis testing method is generally based on the statistical characteristic analysis of the modulated signal under noise interference, and the judgment criteria are derived, and these judgment criteria are usually only applicable to a specific Identification of a class of modulated signals with a narrow identification range

[0005] (2) The method of pattern recognition based on feature extraction has problems such as few types of recognition and poor robustness under low signal-to-noise ratio, and the classification ability of traditional classifiers is limited, and the requirements for the proposed features are extremely high, and the recognition effect is good or bad. In addition to being affected by the type of recognition signal, it largely depends on the discrimination of deep features

[0008] For the hypothesis testing method, the difficulty lies in that the channel effect has a great influence on the analysis of statistical characteristics, which makes it difficult to obtain effective decision criteria; for the pattern recognition method of feature extraction, the difficulty lies in that the extracted features are greatly affected by noise, low signal-to-noise It is difficult to obtain effective classification features; for deep learning methods, the difficulty lies in the lack of a complete theoretical system for the design and optimization of deep neural networks, and more relying on experience and trial

Images