Search

CN-121980195-A - Signal waveform detection and identification method based on TCN network model

CN121980195ACN 121980195 ACN121980195 ACN 121980195ACN-121980195-A

Abstract

The invention provides a signal waveform detection and identification method based on a TCN network model, and belongs to the technical field of information perception and identification. According to the invention, the deep learning model is utilized to complete signal detection and identification, the original signal is preprocessed, the training learning process of the deep learning model parameters is completed, and the trained model is utilized to realize rapid detection and identification of the signal. The method has the advantages of high calculation speed, strong adaptability and high detection accuracy, and can be applied to novel communication reconnaissance and communication signal processing systems.

Inventors

  • ZHANG GUANJIE
  • HU YANG
  • RONG QIANG
  • CHEN TAOYI
  • Jiang menglan
  • LI YANBIN
  • CHEN JINYONG
  • SHI TUO
  • LIU CHUNRAN
  • LI CHUNZE
  • LU NINGNING

Assignees

  • 中国电子科技集团公司第五十四研究所
  • 中电网络空间研究院有限公司

Dates

Publication Date
20260505
Application Date
20260403

Claims (5)

  1. 1. The signal waveform detection and identification method based on the TCN network model is characterized by comprising the following steps of: Step 1, acquiring an original signal waveform, marking a signal starting position, a signal ending position and a signal type to form basic sample data, carrying out sample data enhancement in a mode of transverse scaling, longitudinal scaling and noise addition, and constructing an overlapped type signal sample; step 2, constructing a TCN time convolution neural network model; Step 3, establishing a type prediction cross entropy loss function, and training a TCN time convolution neural network model by adopting an Adam optimization algorithm; And 4, acquiring a real signal waveform, carrying out normalization pretreatment, and predicting the bandwidth, frequency and type of the real signal waveform by using the trained TCN time convolution neural network model to finish detection and identification of the waveform signal.
  2. 2. The method for detecting and identifying signal waveforms based on TCN network model according to claim 1, wherein the specific manner of step 1 is as follows: Step 101, collecting an original signal, performing FFT conversion, and marking the starting position, the ending position and the signal type information of the signal on spectrum data to form a basic sample data set; 102, for sample data in a basic sample data set, completing sample data enhancement by means of transverse scaling, longitudinal scaling, noise addition and sample part superposition, so that the basic sample data set is expanded; and 103, carrying out normalization processing on the expanded sample data to enable each sample to be subjected to independent same distribution.
  3. 3. The signal waveform detection and recognition method based on the TCN network model according to claim 1, wherein the TCN time convolution neural network model is formed by sequentially connecting 2 one-dimensional convolution layers, 4 residual modules and 2 full connection layers, and the residual modules are formed by 2 convolution layers, 1 normalization layer, reLu activation functions and a Dropout layer.
  4. 4. The method for detecting and identifying signal waveforms based on a TCN network model according to claim 1, wherein in step 3, the type prediction cross entropy loss function uses mean square error loss MSELoss, the training uses Adam optimizer to perform gradient update on the TCN time convolution neural network model, and the training is stopped when the mean square error loss is no longer reduced.
  5. 5. The method for detecting and identifying signal waveforms based on TCN network model according to claim 1, wherein the specific manner of step 4 is as follows: step 401, performing FFT conversion on a signal to be predicted, and performing normalization preprocessing; step 402, predicting the position and shape parameters of the real signal waveform by using the trained TCN time convolution neural network model, converting the data, outputting the bandwidth, frequency and type parameters of the signal waveform, and completing the detection and identification of the signal waveform.

Description

Signal waveform detection and identification method based on TCN network model Technical Field The invention belongs to the technical field of information sensing and recognition, and particularly relates to a signal waveform detection and recognition method based on a TCN network model, which can rapidly and accurately detect signals. Background The traditional signal detection and recognition method is mainly carried out step by utilizing detection and decision number classification, and the method has poor effect and insufficient adaptability under the condition of low signal-to-noise ratio, needs to rely on a great deal of expertise experience, and cannot well meet the modern electromagnetic spectrum sensing requirement. The existing signal detection and identification method based on frequency spectrum is mainly a detection method for identifying the existence of a target signal from an actual received signal and extracting parameters such as signal frequency by analyzing time-frequency domain characteristics (amplitude, phase, power, bandwidth, spectrum shape and the like) of the signal. The frequency spectrum-based signal detection algorithm mainly surrounds how to extract frequency domain features of signals, suppresses signal noise, is developed in two aspects, adapts different detection scenes such as stable signals, non-stable signals and weak signals, and is divided into a classical frequency domain detection algorithm and an improved frequency domain detection algorithm, wherein the classical frequency domain detection algorithm is directly based on frequency spectrum/power spectrum obtained by FFT to carry out judgment and mainly adapts signals with stable and known frequency features, the improved frequency domain detection algorithm improves detection performance in the modes of frequency spectrum optimization, feature fusion, self-adaptive threshold and the like, and overcomes the limitations of weak noise resistance and low detection precision on the non-stable signals of the classical algorithm, but the algorithms mostly need manual preset threshold and features. Whether the classical frequency domain detection algorithm or the improved frequency domain detection algorithm, the adaptive capacity of the detection scene for complex and unknown interference signals is weak. Disclosure of Invention Aiming at the defects of the existing detection and recognition method, a deep learning network is fused, the algorithm automatically recognizes the frequency domain characteristics of signals/noise/interference through data training, self-adaptive characteristic extraction and judgment are realized, and a signal waveform detection and recognition method based on a TCN network model is provided. The invention is based on the TCN network model, utilizes the sampling data to learn model parameters, can realize the efficient detection and identification of different types of signal waveforms, and has the advantages of strong adaptability and accurate detection. The invention adopts the technical scheme that: A signal waveform detection and identification method based on a TCN network model comprises the following steps: Step 1, acquiring an original signal waveform, marking a signal starting position, a signal ending position and a signal type to form basic sample data, carrying out sample data enhancement in a mode of transverse scaling, longitudinal scaling and noise addition, and constructing an overlapped type signal sample; step 2, constructing a TCN time convolution neural network model; Step 3, establishing a type prediction cross entropy loss function, and training a TCN time convolution neural network model by adopting an Adam optimization algorithm; And 4, acquiring a real signal waveform, carrying out normalization pretreatment, and predicting the bandwidth, frequency and type of the real signal waveform by using the trained TCN time convolution neural network model to finish detection and identification of the waveform signal. Further, the specific mode of the step 1 is as follows: Step 101, collecting an original signal, performing FFT conversion, and marking the starting position, the ending position and the signal type information of the signal on spectrum data to form a basic sample data set; 102, for sample data in a basic sample data set, completing sample data enhancement by means of transverse scaling, longitudinal scaling, noise addition and sample part superposition, so that the basic sample data set is expanded; and 103, carrying out normalization processing on the expanded sample data to enable each sample to be subjected to independent same distribution. Further, the TCN time convolution neural network model is formed by sequentially connecting 2 one-dimensional convolution layers, 4 residual modules and 2 full-connection layers, wherein each residual module is formed by 2 convolution layers, 1 normalization layer, reLu activation functions and Dropout layers. In step 3, the t