CN-122019963-A - Modulation signal identification method based on time multi-scale feature attention weighting

CN122019963ACN 122019963 ACN122019963 ACN 122019963ACN-122019963-A

Abstract

The invention discloses a modulating signal identification method based on time multi-scale characteristic attention weighting, which relates to the technical field of wireless communication, and comprises the following steps of carrying out normalized pretreatment on a received signal to obtain a pretreated signal; the method comprises the steps of respectively carrying out long-time feature extraction, short-time feature extraction and instantaneous feature extraction on a preprocessed signal to obtain three feature vectors of the long-time feature vector, the short-time feature vector and the instantaneous feature vector, carrying out signal-to-noise ratio estimation on the preprocessed signal to obtain a signal-to-noise ratio estimated value, determining a signal-to-noise ratio interval according to the signal-to-noise ratio estimated value, carrying out attention weighted fusion on the three feature vectors based on the signal-to-noise ratio interval to obtain a comprehensive feature vector, and inputting the comprehensive feature vector into a classifier to carry out modulation recognition to obtain a modulation recognition result. The invention improves the modulation recognition accuracy and stability under the conditions of complex electromagnetic environment and different signal to noise ratios through multi-scale feature extraction and signal to noise ratio self-adaptive weighted fusion.

Inventors

FANG CHENHAO
WANG RUOHUI
DUAN YAOHUI
FANG SHIZHE

Assignees

平湖空间感知实验室科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251215

Claims (10)

1. A method for identifying modulated signals based on attention weighting of time multiscale features, comprising: S1, carrying out normalized pretreatment on a received signal to obtain a pretreated signal; S2, respectively carrying out long-time feature extraction, short-time feature extraction and instantaneous feature extraction on the preprocessed signals to obtain long-time feature vectors, short-time feature vectors and instantaneous feature vectors; S3, carrying out signal-to-noise ratio estimation on the preprocessed signals to obtain signal-to-noise ratio estimation values; s4, determining a signal-to-noise ratio interval according to the signal-to-noise ratio estimated value, and carrying out attention weighted fusion on the long-time feature vector, the short-time feature vector and the instantaneous feature vector based on the signal-to-noise ratio interval to obtain a comprehensive feature vector; s5, inputting the comprehensive feature vector into a classifier for modulation recognition, and obtaining a modulation recognition result.
2. The method for identifying a modulated signal based on attention weighting of time multiscale features according to claim 1, wherein the normalized preprocessing of the received signal specifically comprises: performing down-conversion processing on the received signal to obtain a baseband in-phase quadrature signal; Performing timing synchronization processing on the baseband in-phase quadrature signal, and eliminating carrier frequency offset and timing synchronization errors; And carrying out power normalization processing on the signals subjected to the timing synchronization processing to obtain the preprocessing signals.
3. The method for identifying modulated signals based on time multi-scale feature attention weighting according to claim 1, wherein said performing long-time feature extraction, short-time feature extraction and instantaneous feature extraction on said preprocessed signals respectively comprises: In a long-time feature extraction channel, calculating the fourth-order accumulation amount, the sixth-order accumulation amount, the spectral density and the spectral peak feature of the preprocessed signal to obtain the long-time feature vector; In a short-time feature extraction channel, acquiring time-frequency distribution features of the preprocessing signals to obtain the short-time feature vector; In the instantaneous feature extraction channel, calculating the instantaneous feature change of the preprocessing signal to obtain the instantaneous feature vector.
4. The method for identifying a modulated signal based on attention weighting of time multiscale features according to claim 1, wherein said estimating the signal-to-noise ratio of the preprocessed signal comprises: calculating an autocovariance matrix of the preprocessed signals; Performing eigenvalue decomposition on the auto-covariance matrix to obtain a decomposition result, and dividing eigenvectors in the decomposition result into a signal subspace and a noise subspace based on an eigenvalue threshold; estimating signal power from larger eigenvalues in the signal subspace, and estimating noise power from an average of smaller eigenvalues in the noise subspace; and taking the ratio of the signal power to the noise power as the signal-to-noise ratio estimated value.
5. A modulating signal recognition system based on time multi-scale feature attention weighting is characterized by comprising a signal preprocessing module, a multi-scale extraction module, a signal-to-noise ratio estimation module, an attention weighting module and a modulating recognition module; the signal preprocessing module is used for carrying out normalized preprocessing on the received signals to obtain preprocessed signals; The multi-scale extraction module is used for respectively carrying out long-time feature extraction, short-time feature extraction and instantaneous feature extraction on the preprocessed signals to obtain long-time feature vectors, short-time feature vectors and instantaneous feature vectors; the signal-to-noise ratio estimation module is used for carrying out signal-to-noise ratio estimation on the preprocessed signal to obtain a signal-to-noise ratio estimation value; the attention weighting module is used for determining a signal-to-noise ratio interval according to the signal-to-noise ratio estimation value, and carrying out attention weighted fusion on the long-time feature vector, the short-time feature vector and the instantaneous feature vector based on the signal-to-noise ratio interval to obtain a comprehensive feature vector; The modulation recognition module is used for inputting the comprehensive feature vector into a classifier for modulation recognition to obtain a modulation recognition result.
6. A modulated signal recognition system based on temporal multi-scale feature attention weighting as defined in claim 5, wherein said normalized preprocessing of the received signal comprises: performing down-conversion processing on the received signal to obtain a baseband in-phase quadrature signal; Performing timing synchronization processing on the baseband in-phase quadrature signal, and eliminating carrier frequency offset and timing synchronization errors; And carrying out power normalization processing on the signals subjected to the timing synchronization processing to obtain the preprocessing signals.
7. The modulated signal recognition system based on time multi-scale feature attention weighting of claim 5, wherein said performing long-term feature extraction, short-term feature extraction, and instantaneous feature extraction on said preprocessed signal, respectively, comprises: In a long-time feature extraction channel, calculating the fourth-order accumulation amount, the sixth-order accumulation amount, the spectral density and the spectral peak feature of the preprocessed signal to obtain the long-time feature vector; In a short-time feature extraction channel, acquiring time-frequency distribution features of the preprocessing signals to obtain the short-time feature vector; In the instantaneous feature extraction channel, calculating the instantaneous feature change of the preprocessing signal to obtain the instantaneous feature vector.
8. A modulated signal recognition system based on temporal multi-scale feature attention weighting as defined in claim 5, wherein said signal-to-noise ratio estimation of said preprocessed signal comprises: calculating an autocovariance matrix of the preprocessed signals; Performing eigenvalue decomposition on the auto-covariance matrix to obtain a decomposition result, and dividing eigenvectors in the decomposition result into a signal subspace and a noise subspace based on an eigenvalue threshold; estimating signal power from larger eigenvalues in the signal subspace, and estimating noise power from an average of smaller eigenvalues in the noise subspace; and taking the ratio of the signal power to the noise power as the signal-to-noise ratio estimated value.
9. A computer device comprising a processor coupled to a memory, the memory having stored therein at least one computer program that is loaded and executed by the processor to cause the computer device to implement a method of modulating signal recognition based on temporal multi-scale feature attention weighting as claimed in any one of claims 1 to 4.
10. A computer readable storage medium, wherein at least one computer program is stored in the computer readable storage medium, the at least one computer program being loaded and executed by a processor to cause the computer to implement a modulated signal recognition method based on a time multiscale feature attention weighting as claimed in any one of claims 1 to 4.

Description

Modulation signal identification method based on time multi-scale feature attention weighting Technical Field The invention relates to the technical field of wireless communication, in particular to a modulating signal identification method based on time multi-scale feature attention weighting. Background With rapid development of wireless communication technology and increasing complexity of electromagnetic environment, automatic identification of modulation modes of unknown signals in non-cooperative communication scenes such as electronic inspection, spectrum monitoring and adaptive communication is urgently required. The modulated signal identification is used as a key link of a cognitive radio and intelligent communication system, and the performance of the modulated signal identification directly influences the effectiveness of subsequent signal demodulation, information acquisition and network management. However, the existing method has obvious defects in terms of feature utilization rate and environmental adaptability, and particularly, distinguishing features of signals on different time scales are difficult to fully mine, so that the identification stability is poor and the accuracy is low under complex channel conditions and different signal-to-noise ratio changes, and the practical application efficiency is seriously restricted. Aiming at the technical problems, the current mainstream modulation recognition technology is mainly divided into two types, namely a traditional method based on artificial feature engineering and an automatic feature learning method based on machine learning. The traditional method realizes modulation classification by extracting statistical parameters such as high-order accumulation quantity, spectral characteristics, time analysis results and the like of signals and combining with classifiers such as decision trees or support vector machines. The deep learning method automatically learns the hierarchical feature representation by using a convolutional neural network, a cyclic neural network and other models. Part of the improvement scheme introduces the steps of generating an equipotential star map and carrying out secondary identification to improve the accuracy of modulation identification. Despite the advances made in the above-described techniques, significant drawbacks remain. The traditional artificial characteristic method depends on expert experience design, has limited characteristic expression capability, is difficult to adapt to complex channel environments such as multipath fading, frequency selective fading and the like, and has insufficient expansibility on a novel digital modulation mode. The existing deep learning method reduces the dependence on artificial feature design, but most of the deep learning method adopts a feature extraction structure with a single time scale or a fixed scale, and cannot effectively utilize multi-scale complementary information of signals in the time dimension, so that feature differentiation is insufficient. The method for generating the equipotential star map has complex processing flow and poor real-time performance. In general, most methods in the prior art analyze signal features only from a single scale or a single domain, and fail to effectively fuse multi-scale information of a time domain, a frequency domain and a transform domain, so that the recognition accuracy is not high in a complex channel environment, and the overall robustness is not good. In summary, the existing modulation signal recognition technology has shortcomings in terms of signal feature comprehensiveness, signal-to-noise ratio suitability and complex environment robustness, and is difficult to meet the recognition requirements of high precision and high stability in a non-cooperative communication scene. Therefore, development of a modulation recognition method capable of fully utilizing signal time multi-scale characteristics and dynamically optimizing characteristic weight distribution according to signal-to-noise ratio strength is needed to break through the bottleneck of the prior art and realize accurate and reliable recognition of multiple modulation modes under different signal-to-noise ratio conditions. Disclosure of Invention Aiming at the defects of the prior art, the invention specifically provides a modulating signal recognition method based on time multiscale feature attention weighting, which aims at solving the problems of low recognition accuracy and poor stability caused by the fact that the existing modulating recognition method cannot adaptively optimize time multiscale feature weight according to signal-to-noise ratio change, and specifically comprises the following steps: 1) In a first aspect, the present invention provides a method for identifying a modulated signal based on attention weighting of time multi-scale features, and the specific technical scheme is as follows: S1, carrying out normalized pretreatment on a recei