A Data-Driven and Knowledge-Constrained Signal Carrier Detection and Recognition Method

Abstract

This invention discloses a signal carrier detection and recognition method based on data-driven and knowledge-constrained approaches. The method includes: first, constructing a real dataset and an enhanced simulation dataset comprising time-frequency data processed by logarithmic amplitude spectrum normalization; then, training a deep learning-based neural network using both the simulation and real datasets through transfer learning; next, using the trained neural network to detect the target signal; and finally, verifying the detection results based on knowledge constraints and hierarchical prediction information, correcting the detection results according to the verification results, thereby achieving signal carrier detection and recognition. The signal carrier detection and recognition method based on data-driven and knowledge-constrained approaches proposed in this application utilizes a deep learning neural network with high learning efficiency, low learning difficulty, small training data requirements, low human intervention, and high signal recognition accuracy, making this method more applicable to real-world scenarios and more practical.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of signal detection, and particularly relates to a signal carrier detection and identification method based on data driving and knowledge constraint. BACKGROUND

[0002] Short wave communication, also known as high frequency (HF) communication, has a working frequency of 3-30 MHz, and is widely applied in various fields due to its long communication distance, strong anti-destroying ability, high flexibility, strong independent communication ability and low operation cost. Since the short wave channel is an open channel, it is greatly affected by the sferics noise and human noise, especially the time delay, Doppler frequency shift, spectrum expansion, channel fluctuation and fading, and a large number of signals of different types and using various communication protocols exist in the short wave frequency band, and various new burst signals and frequency hopping signals constantly appear, so it is a very complex and challenging work to perform electromagnetic spectrum monitoring and analysis on the signals in the frequency band.

[0003] The traditional short wave spectrum detection method mainly includes an energy detection method, a matching filter (correlation) detection method, a cyclic stationary feature detection method and a feature extraction method based on a time-frequency image. The energy detection method is a detection method based on signal energy, the detection performance of which is greatly affected by the channel environment and the threshold, and the carrier time-frequency detection and classification cannot be completed at the same time. The matching filter (correlation) detection method is a method based on correlation matching, which needs to complete frequency point detection and filtering and carrier priori knowledge, and the application scene is limited to the time domain start-stop detection of known signal categories. The cyclic stationary feature detection method analyzes the spectrum correlation characteristics to detect signals, and the calculation amount of the method is large, the real-time requirement cannot be met, and the signal time-frequency detection cannot be accurately realized. The image feature detection mainly uses the image detection method to complete signal detection, and the traditional short wave spectrum detection methods have certain deficiencies.

[0004] In recent years, with the development of deep learning technology, deep learning-based methods for shortwave signal detection and recognition have been proposed, mainly including: deep learning-based sequence detection methods, deep learning-based image recognition methods, and deep learning-based target detection methods. Among these, deep learning-based sequence detection methods transform the detection problem into a signal recognition problem and a sequence labeling problem. Their main advantages are avoiding manual feature determination of signal thresholds and improving the signal detection probability of the detection algorithm. Their main disadvantage is the inability to simultaneously determine the signal start and end times, start and end frequencies, and carrier type. Deep learning-based image recognition methods transform the detection problem into a signal recognition problem, and their main... The advantages of this approach are high accuracy in classification and detection, and the elimination of the need for manual feature extraction and threshold parameter determination. The main drawback is the inability to accurately determine the signal start and end frequencies and times. Deep learning-based target detection methods use deep learning image target detection and recognition methods to complete signal time-frequency localization and classification. Their main advantages are high detection and recognition rates, the ability to simultaneously provide time-frequency localization and carrier category information, and the elimination of the need for manual feature extraction and threshold parameter determination. However, their main drawbacks are the requirement for a large amount of real data for training and insufficient performance in detecting target signals with special time-frequency characteristics. All of the above deep learning-based shortwave signal detection and recognition methods also have certain shortcomings. Summary of the Invention

[0005] In view of this, the present invention proposes a signal carrier detection and identification method based on data-driven and knowledge-constrained methods. This method uses deep learning to perform hierarchical localization and identification of various shortwave signals of different systems, verifies and identifies the detection results through prior knowledge, and trains and learns the neural network through a transfer learning method that enhances the coupling of simulation data and actual channels.

[0006] To achieve this objective, the present invention adopts the following technical solution: a signal carrier detection and identification method based on data-driven and knowledge-constrained methods, the method comprising:

[0007] S1: Construct the training dataset;

[0008] S11: Read the actual acquired signal, perform a short-time Fourier transform on it, and obtain the corresponding real signal time-frequency data;

[0009] S12: Generate a simulation signal, perform a short-time Fourier transform on the simulation signal, and obtain the corresponding simulation signal time-frequency data;

[0010] S13: Enhance the time-frequency data of the simulated signal to obtain enhanced time-frequency data of the simulated signal;

[0011] S14: Perform frame header, frame tail, and frame body segmentation annotations on the real signal time-frequency data and the enhanced simulated signal time-frequency data, respectively;

[0012] S15: performing log amplitude spectrum normalization processing on the segmented labeled data of the real signal time-frequency data and the enhanced simulation signal time-frequency data, to obtain a real data set and a simulation data set for training;

[0013] S2: training a neural network based on deep learning using the real data set and the simulation data set;

[0014] S21: training a neural network based on deep learning using the simulation data set to obtain pre-training weights;

[0015] S22: retraining the neural network trained by the simulation data set using the real data set to optimize the training weights, to obtain a trained neural network based on deep learning;

[0016] S3: detecting and identifying a signal carrier using the trained neural network based on deep learning;

[0017] S31: collecting a to-be-tested signal, and performing short-time Fourier transform and log amplitude spectrum normalization processing on the to-be-tested signal to obtain a to-be-tested data set;

[0018] S32: inputting the to-be-tested data set into the neural network based on deep learning for detection, and outputting a detection result;

[0019] S33: checking and identifying the detection result;

[0020] First, the detection result is classified: the frame header, frame tail and frame body of the detection result are respectively judged, and the detection result is classified according to the judgment, if the detection result is classified, the detection result is assigned to a classifiable signal of the same frequency band, and the same frequency accumulation is performed; if the detection result is not classified, the detection result is further classified using a signal specification judgment method, if the detection result is further classified, the detection result is assigned to a classifiable signal of the same frequency band, and the same frequency accumulation is performed; if the detection result is still not classified, the same frequency accumulation of the detection result is directly performed;

[0021] Then, the above classification process is repeated until the judgment of all detection results is completed

[0022] Finally, the detection result accumulated for multiple times in the same frequency is checked, and the neural network detection result is corrected according to the checking result, to complete the detection and identification of the to-be-tested signal.

[0023] Preferably, the simulation signal in S12 is obtained by joint simulation of a simulation signal and a Watterson model.

[0024] Preferably, the method of superimposing the time-frequency data of the simulation signal and the background data in the real acquisition signal in S13 is any one of Gaussian noise background enhancement, random spectrum shift enhancement, random interference signal addition enhancement, real background noise superposition enhancement, random gain enhancement and random time delay enhancement.

[0025] Preferably, the neural network based on deep learning takes the CenterNet framework as a basic framework and takes DLA34 or ResNet18 as a backbone network.

[0026] Preferably, in S33, the signal specification decision mode is: extracting classification and parameter extraction of the detection result, making a decision based on prior knowledge or extracting a single carrier signal from the signal.

[0027] The beneficial effects of the present application are: the signal carrier detection and recognition method based on data driving and knowledge constraint proposed in the present application has the following advantages: (1) for the problem of large-scale data demand, a neural network training method combining simulation data and real data is constructed, which can achieve the training effect of a large amount of real data by using only a small amount of real data, improve the learning efficiency and reduce the data demand; (2) a variety of enhanced simulation data set construction methods are proposed, which improve the fidelity of simulation data, realize simulation data enhancement and reduce the learning difficulty of the migration application from simulation data to real data; (3) a learning method of frame header, frame tail and frame body labeling is proposed, which effectively improves the short-wave signal detection and recognition ability and avoids the performance degradation caused by standard labeling; (4) the log amplitude spectrum is used as the input of the deep learning target detection network, which enhances the time-frequency signal feature discrimination ability and improves the signal detection and recognition performance; (5) the simulation data transfer learning method is used, which reduces the algorithm complexity and data demand, improves the algorithm learning efficiency and enhances the algorithm application practicability; (6) a verification method based on knowledge constraint and hierarchical prediction information is proposed, which realizes the verification and confirmation of signals through the connection characteristics, parameter rules and category information, and improves the recognition ability of complex system signals.

[0028] In summary, the signal carrier detection and recognition method based on data driving and knowledge constraint proposed in the present application has high learning efficiency, low learning difficulty, small training data demand, low degree of human intervention, high signal recognition precision, and is more suitable for practical scenarios and more practical. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The flowchart of the signal carrier detection and recognition method based on data driving and knowledge constraint in the embodiments of the present application is shown in the figure.

[0030] Figures 2-7Fig. 1 is an effect schematic diagram of the method for enhancing the time-frequency data of the simulation signal according to an embodiment of the present application;

[0031] Figure 2 Fig. 2 is an effect schematic diagram of the method for enhancing the time-frequency data of the simulation signal according to an embodiment of the present application;

[0032] Figure 3 Fig. 3 is an effect schematic diagram of the method for enhancing the time-frequency data of the simulation signal according to an embodiment of the present application;

[0033] Figure 4 Fig. 4 is an effect schematic diagram of the method for enhancing the time-frequency data of the simulation signal according to an embodiment of the present application;

[0034] Figure 5 Fig. 5 is an effect schematic diagram of the method for enhancing the time-frequency data of the simulation signal according to an embodiment of the present application;

[0035] Figure 6 Fig. 6 is an effect schematic diagram of the method for enhancing the time-frequency data of the simulation signal according to an embodiment of the present application;

[0036] Figure 7 Fig. 7 is an effect schematic diagram of the method for enhancing the time-frequency data of the simulation signal according to an embodiment of the present application;

[0037] Figure 8 Fig. 8 is a comparison diagram of the time-frequency diagram of the simulation signal before and after passing through the Watterson channel according to an embodiment of the present application;

[0038] Figure 9 Fig. 9 is a comparison diagram of the frame header / frame / frame body annotation and the standard annotation of different types of signals according to an embodiment of the present application;

[0039] Figure 10 Fig. 10 is a signal calibration process schematic diagram according to an embodiment of the present application;

[0040] Figure 11 Fig. 11 is a comparison diagram of the signal correction after the signal calibration according to an embodiment of the present application;

[0041] Figure 12 Fig. 12 is a signal detection enhancement schematic diagram based on the simulation transfer learning according to an embodiment of the present application. DETAILED DESCRIPTION

[0042] Those skilled in the art will appreciate that the embodiments described herein are presented for the purpose of understanding the principles of the present application and should be understood as not limiting the scope of protection of the present application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations according to the technical inspiration of the present application without departing from the essence of the present application, and these modifications and combinations are still within the scope of protection of the present application.

[0043] The present application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] The application provides a signal carrier detection and identification method based on data driving and knowledge constraint. Figure 1 As shown in the flowchart, the method comprises the following steps:

[0045] First step: constructing a training data set, the specific process is as follows:

[0046] Firstly, the real collected signal is read, and the real collected signal is subjected to short-time Fourier transform to obtain corresponding real signal time-frequency data.

[0047] Then, the simulation signal data is generated by simulation, and the simulation signal data is subjected to short-time Fourier transform to obtain corresponding simulation signal time-frequency data.

[0048] Then, the simulation signal is subjected to enhancement processing to obtain enhanced simulation signal time-frequency data, and the enhancement methods mainly include: a) Gaussian noise background enhancement; b) random frequency spectrum shift enhancement; c) random interference signal addition enhancement; d) real background noise superposition enhancement; e) random gain enhancement; and f) random time delay enhancement, and the effects of different enhancement methods are shown in Figures 2-7 In this embodiment, the simulation signal time-frequency data is superimposed and enhanced with the real signal background data.

[0049] The time-frequency data of the real signal is obtained by dividing the real signal by STFT (short-time Fourier transform), the simulation signal is obtained by joint simulation of the simulation signal and the Watterson model, and the simulation signal time-frequency comparison chart before and after the Watterson model is shown in Figure 8 Through the enhancement operation of the simulation signal, simulation data and measured data in various complex scenarios are obtained, and sample mixed enhancement is realized.

[0050] Then, the real signal time-frequency data and the enhanced simulation signal time-frequency data are respectively subjected to frame header, frame tail and frame body segmentation labeling. Since the shortwave signals suitable for the method proposed in the application have large differences in different parts of the frame header and the frame tail, in order to improve the performance of the deep learning target detection and identification algorithm, the application proposes a method of segmenting and labeling the frame header, the frame tail and the frame body of the signal to construct the training data set. Especially for PSK signals and multi-tone signals, the remaining load in the time-frequency data is highly similar except for the special frame header and frame tail, therefore, for this kind of signal, the method of segmenting and labeling the frame header, the frame tail and the frame body is adopted. For the same signal, the standard labeling method and the frame header, frame tail and frame body segmentation labeling method are used for labeling, as shown in Figure 9

[0051] ​The application will train and test the neural network respectively by using the data labeled by the standard labeling method and the data labeled by the frame header, frame tail and frame body segmentation, and the simulation performance is shown in Table 1, and the simulation results show that the frame header, frame tail and frame body segmentation labeling method proposed in the application can more effectively distinguish similar signals and improve the detection performance, and the mAP can be improved by 26% compared with the existing standard labeling method, the mAP (Mean average precision) is one of the most important evaluation indexes in the target detection algorithm, and the mAP is the average value of multi-class AP, and the AP (Average precision) represents the average accuracy.

[0052] Table 1 Performance comparison of different labeling methods

[0053] Network structure Input size Labeling method Dataset mAP Centernet-DLA34 (512,512,1) Standard labeling Real dataset 0.72358 Centernet-DLA34 (512,512,1) Frame header frame body frame tail labeling Real dataset 0.98343

[0054] Finally, in order to enhance the detection performance of weak signals and offset the influence of channel fading characteristics, the real signal time-frequency data and the enhanced simulation signal time-frequency data labeled by segmentation are subjected to log amplitude spectrum normalization processing to obtain the real data set and the simulation data set for training.

[0055] The log amplitude spectrum normalization processing process is as follows:

[0056]

[0057]

[0058] Wherein, S i is the time-frequency spectrum obtained after the signal is subjected to STFT; i represents the i-th signal, and is a positive integer; and ε is a minimum value to avoid the case that the logarithm is infinite, and finally the log amplitude spectrum S i_log and the normalized log amplitude spectrum

[0059] Table 2 Performance comparison of different data preprocessing

[0060]

[0061]

[0062] The short wave detection and recognition performance of the trained deep learning-based neural network under different input data types such as log amplitude spectrum, amplitude spectrum, amplitude / phase and IQ data is compared, and the results are shown in Table 2, and the results show that the normalized log amplitude spectrum proposed in the application as the input of the deep learning-based neural network can more effectively extract signal features, and the mAP can be improved by 1.5% to 2.8% compared with the existing general amplitude spectrum method.

[0063] Second step: training the deep learning based neural network using the dataset;

[0064] In this embodiment, the CenterNet framework is used as the basic framework of the neural network, and DLA or ResNet18 is used as the backbone network of the neural network, but the application is not limited to the above framework and network structure, and other target detection network framework and network structure can also be used for replacement. The neural network uses normalized log amplitude spectrum as data input, and respectively uses simulation dataset and real dataset for training test.

[0065] In this step, in view of the large-scale data requirement faced by the deep learning based short wave signal detection algorithm, a transfer learning method based on simulation data is proposed, and the process is as shown in Figure 12 First, the deep learning based neural network is trained using the simulation dataset to obtain the pre-training weight, and then the deep learning based neural network is retrained using the real dataset to optimize the training weight, and the trained deep learning neural network is obtained, so as to improve the data utilization rate and the performance of the deep learning algorithm.

[0066] Third step: detecting the signal carrier using the trained deep learning neural network;

[0067] First, the short-time Fourier transform and log amplitude spectrum normalization processing are performed on the to-be-detected signal to obtain the to-be-tested dataset; then the to-be-tested dataset is input into the deep learning neural network for detection, and the detection result is output, including signal frequency point, bandwidth, start and end time and category;

[0068] Fourth step: checking the detection result. Since the time-frequency characteristics of the short wave signal itself or the frame body part are highly similar, the frame header and the frame tail part are quite different from the frame body part, so it is necessary to use the frame header, frame tail and frame body detection and identification method, and for the result predicted by the deep learning based neural network in the above step, this embodiment proposes a checking method based on knowledge constraint and hierarchical prediction information, and the process is as shown in Figure 10 The basic principle of this method is to realize the rechecking and confirmation of the easy aliasing signal by using the intercommunication characteristics of the same frequency band signal, the constraint of the signal carrier parameter specification and the time domain signal category information, and the process is as shown in Figure 10

[0069] ​First, the frame header, frame tail and frame body of the detection result are respectively judged, and the signal is classified according to the judgment result. If the detection result can be classified, the detection result and the same frequency signal thereof are assigned to the classifiable category, and the same frequency signal accumulation is performed. If the detection result cannot be classified, the signal is judged by using the signal specification judgment mode. If the detection result can be classified, the detection result and the same frequency signal thereof are assigned to the classifiable category, and the same frequency signal accumulation is performed. If the signal cannot be classified, the same frequency signal accumulation is directly performed.

[0070] The above signal specification judgment mode is to extract the classification and parameters of the signal, and to judge based on prior knowledge or to identify the signal by extracting the single carrier signal of the signal.

[0071] Since the information obtained by a single sample may have errors, the classified signal needs to be subjected to time-frequency multi-segment accumulation correction operation. Therefore, the same frequency band detection results of multiple accumulations need to be verified next. The detection results are corrected according to the verification results, and finally the signal carrier detection and identification are completed.

[0072] In addition to improving the classification and identification performance, the above verification method based on knowledge constraints and hierarchical prediction information can also eliminate false alarm targets and targets not of interest, thereby improving the overall recall rate of the algorithm, as shown in Figure 11 .

[0073] The present application aims at large-scale data demand problems. The training results of the neural network are compared by using simulation data sets and real data sets respectively. The training performance of the same neural network under different learning methods is compared by using the same data set to train different depth neural networks (backbone network uses DLA and ResNet18 respectively) and using different data sets to train the same neural network. The results are shown in Table 3.

[0074] Table 3 Performance comparison of different network structures and data sources

[0075] Network structure Input size Data type Dataset mAP Centernet-DLA34 (512,512,1) Log amplitude spectrum Real dataset 0.98343 Centernet-Res18 (512,512,1) Log amplitude spectrum Real dataset 0.91408 Centernet-Res18 (512,512,1) Log amplitude spectrum Simulated dataset 0.55832 Centernet-Res18 (512,512,1) Log amplitude spectrum Simulated -> Real 0.98098

[0076] The above results show that when only real data is used to train the neural network, reducing the complexity of the neural network will result in a decrease in performance (from DLA to ResNet18), but if the transfer learning neural network training method proposed in the present application is used to pre-train the neural network with simulation data and then optimize the training with real data, the performance of the neural network with lower complexity can almost reach the performance of the complex network trained with real data, which indicates that the complexity of the neural network has little effect on the transfer learning neural network training method proposed in the present application, and the transfer learning neural network training method proposed in the present application can improve the performance of the deep learning algorithm; the performance difference is large when the same neural network is trained with simulation data and real data respectively, and the neural network trained only with simulation data cannot be applied to real data.

[0077] The above results also show that the method of combining real data and simulation data in the present application can achieve the training effect of a large amount of real data with only a small amount of real data, and the method of combining real data and simulation data in the present application improves the learning efficiency and reduces the data requirement, making the method more suitable for practical scenarios and more practical.

Claims

1. A data-driven and knowledge-constrained based signal carrier detection and identification method, characterized in that, The method comprises: S1: constructing a training data set; S11: reading a real acquisition signal, performing short-time Fourier transform on the real acquisition signal, and obtaining corresponding real signal time-frequency data; S12: generating a simulation signal, performing short-time Fourier transform on the simulation signal, and obtaining corresponding simulation signal time-frequency data; S13: performing enhancement processing on the simulation signal time-frequency data to obtain enhanced simulation signal time-frequency data; S14: respectively performing frame header, frame tail and frame body segmentation labeling on the real signal time-frequency data and the enhanced simulation signal time-frequency data; S15: performing log amplitude spectrum normalization processing on the segmentation labeling data of the real signal time-frequency data and the enhanced simulation signal time-frequency data to obtain a real data set and a simulation data set for training; S2: training a neural network based on deep learning by using the real data set and the simulation data set; S21: training the neural network based on deep learning by using the simulation data set to obtain pre-training weights; S22: retraining the neural network trained by using the simulation data set by using the real data set to optimize the training weights and obtain a trained neural network based on deep learning; S3: detecting and identifying a signal carrier by using the trained neural network based on deep learning; S31: acquiring a to-be-tested signal, performing short-time Fourier transform and log amplitude spectrum normalization processing on the to-be-tested signal respectively to obtain a to-be-tested data set; S32: inputting the to-be-tested data set into the neural network based on deep learning for detection and outputting a detection result; S33: performing checking and identification on the detection result; Firstly, detection result decision classification is performed: the frame header, the frame tail and the frame body of the detection result are respectively decided, the detection result is classified according to the decision, if the detection result is classifiable, the detection result and the same frequency detection result are assigned to the classifiable type, and the same frequency accumulation is performed on the detection result; if the detection result is not classifiable, the detection result is further classified by using a signal specification decision mode, the signal specification decision mode is: the detection result is extracted and classified and parameters are extracted, decision is made based on prior knowledge or a single carrier signal is extracted from the signal, if the detection result is classifiable, the detection result and the same frequency detection result are further assigned to the classifiable type, and the same frequency accumulation is performed on the detection result; if the detection result is still not classifiable, the same frequency accumulation is directly performed on the detection result; then, the above decision classification process is repeated until the decision of all detection results is completed. Finally, the detection result accumulated for multiple times at the same frequency is checked, the detection result is corrected according to the checking, and the detection and identification of the to-be-tested signal are completed.

2. The data-driven and knowledge-constrained based signal carrier detection and identification method of claim 1, wherein, The simulation signal in S12 is obtained by joint simulation of a simulation signal and a Watterson model.

3. The data-driven and knowledge-constrained based signal carrier detection and identification method of claim 1, wherein, The simulation signal enhancement processing mode in S13 includes any one of a Gaussian noise background enhancement, a random spectrum shift enhancement, a random interference signal addition enhancement, a real background noise superposition enhancement, a random gain enhancement and a random time delay enhancement.

4. The data-driven and knowledge-constrained based signal carrier detection and identification method of claim 1, wherein, The neural network based on deep learning takes a CenterNet framework as a basic framework and takes a DLA34 or a ResNet18 as a backbone network.

Images