A weak signal detection method, system, device and medium based on deep learning

Abstract

A weak signal detection method, system, device and medium based on deep learning, the method is, first, constructing a weak signal time-frequency matrix dataset; then, preprocessing the time-frequency matrix dataset; constructing a deep learning signal detection network, and finally detecting the weak signal by using the trained network; the system, device and medium can detect the weak signal based on the weak signal detection method based on deep learning; the application has better detection performance, can greatly improve the accuracy of weak signal detection, has high real-time performance and good reliability; the information of the signal in the time domain and the frequency domain can be obtained at the same time, which provides a premise for subsequent signal analysis and processing; the algorithm has better robustness.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wireless communication signal blind detection, and particularly relates to a weak signal detection method, system, device and medium based on deep learning. BACKGROUND

[0002] In the civil aspect, with the rise of software radio and cognitive radio, wireless communication technology has been rapidly developed and widely applied, and the spectrum resource is in short supply and the electromagnetic environment is increasingly complex. Unlicensed illegal equipment occupies the spectrum resource at will, which affects public broadcasting, television and other equipment, so it is urgent to increase the monitoring of radio spectrum resources. Spectrum resource monitoring is to use a receiver to detect radio signals for a long time, to obtain the activity of each frequency band through signal blind detection technology, and to identify illegal signals to implement shielding, suppression and other measures to ensure that authorized legal equipment can safely use spectrum resources. In actual scenarios, due to the small transmission power and coverage range of illegal equipment, as well as the loss caused by path propagation, the illegal signals received by the monitoring equipment are very weak.

[0003] In the military aspect, electronic warfare has become a key factor affecting the victory of the battlefield, and mastering electromagnetic spectrum information is the premise of mastering intelligence, which plays an important role in battlefield decision-making. With the evolution of war patterns, changes in combat environment and the development of new technologies, electromagnetic spectrum has gradually evolved from an important carrier of battlefield information acquisition and transmission into the sixth combat dimension parallel to land, sea, air, space and network. Compared with other dimensions of threats, electromagnetic threats have high dynamics, so there are higher requirements for the accuracy and timeliness of electronic reconnaissance. Electronic reconnaissance is to use electronic equipment to listen to the activities of enemy electronic equipment, and to obtain the electromagnetic information of the enemy by collecting and processing the electromagnetic waves of the enemy. The first thing to do for the collected electromagnetic waves is signal blind detection, then signal processing to obtain the characteristic parameters of the signal, and finally inferring the purpose and threat level according to the characteristic parameters. In actual scenarios, due to the use of low probability of interception technology and the interference of multipath effect in the signal propagation process, the useful signals received are very weak.

[0004] In a cooperative communication system, the receiving end obtains the signal parameter information of the transmitting end in advance, and the information exchange between the transmitting and receiving ends is relatively simple. However, in actual application scenarios, whether it is the monitoring of illegal equipment in the civil aspect or the listening to enemy electronic equipment in the military aspect, the receiving end cannot obtain the prior information of the transmitting end, which is a non-cooperative communication process. Therefore, blind processing of the received signal is needed, and the first thing is signal detection to determine whether there is information of interest in time and reliably, and if there is, to obtain its information in the time and frequency domains, which provides a necessary prerequisite for subsequent signal processing and analysis.

[0005] The existing conventional communication signal detection algorithm is usually based on the characteristic parameters of the signal, and whether the signal exists is determined by judging whether the characteristic parameters are located in a reasonable value range interval, mainly including energy detection method, correlation detection method and cyclostationary detection method.

[0006] The core idea of the energy detection method is that when the signal exists, the calculated signal energy is greater than that of pure noise, and whether the signal exists is determined by setting an appropriate energy threshold. The key of this method is to select a suitable energy threshold. When the noise floor is relatively stable, the threshold value can be easily determined, but when the noise fluctuation is large, it is difficult to determine the threshold value. Moreover, only the information of a single dimension in time domain or frequency domain can be obtained, and the information of two dimensions cannot be obtained at the same time.

[0007] The core principle of the correlation detection method is to use the different correlation of the signal and the noise to realize detection. Under non-cooperative conditions, the correlation of the to-be-detected signal and the noise is poor, and the improvement of the signal is limited under low signal-to-noise ratio conditions. Moreover, it is only used to determine whether the signal exists, and cannot determine the signal occurrence time and the frequency bandwidth.

[0008] The core theory of the cyclostationary detection method is to use the cyclostationary characteristics of the signal to distinguish it from the noise. Generally, it has high computational complexity and high computational cost. Moreover, it is only used to determine whether the signal exists, and cannot determine the signal occurrence time and the frequency bandwidth.

[0009] In the existing weak signal detection algorithm: the high-order spectrum analysis has a large amount of calculation and is not suitable for scenes with high real-time requirements; the chaos theory needs to know the frequency of the to-be-detected periodic signal, and is not suitable for blind detection scenes; the stochastic resonance machine resonance adiabatic approximation theory has strict restrictions on the frequency, amplitude and input noise intensity of the to-be-detected signal, and requires that these values must be small, which is not suitable for blind detection scenes of high-frequency communication signals; some deep learning networks take the time domain data or characteristic parameters of the signal as input, and although good detection performance is achieved, the information of the signal in time domain and frequency domain cannot be obtained, which increases the difficulty of subsequent signal processing.

[0010] In summary, the existing technology has the following problems:

[0011] The conventional detection algorithm with good detection performance under high signal-to-noise ratio conditions deteriorates under low signal-to-noise ratio conditions;

[0012] The method with good detection performance under low signal-to-noise ratio conditions often needs prior information, and has large calculation amount and poor real-time performance;

[0013] The existing technology can only determine whether the signal exists or obtain the information of the signal in time or frequency domain, and cannot obtain the information of two dimensions in time and frequency domain at the same time. SUMMARY

[0014] In order to overcome the above-mentioned deficiencies of the prior art, the purpose of the present application is to provide a weak signal detection method, system, device and medium based on deep learning, which realizes timely and reliable detection of small signals in strong noise background; a weak signal detection network is constructed by using the enhanced signal time-frequency transformation matrix as the input in the form of combining signal time-frequency transformation, low-light image enhancement algorithm and deep learning target detection network, after the network training is completed, without prior information, timely and reliable detection of weak signals is realized, and two-dimensional information of the signal in time domain and frequency domain can be obtained.

[0015] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0016] A weak signal detection method based on deep learning, specifically comprising the following steps:

[0017] Step one, constructing a weak signal time-frequency matrix dataset: obtaining a weak signal time-frequency matrix dataset through signal time-frequency transformation and time-frequency information labeling;

[0018] Step two, time-frequency matrix preprocessing: obtaining an enhanced time-frequency matrix dataset through time-frequency matrix enhancement and time-frequency matrix cutting;

[0019] Step three, constructing a deep learning signal detection network and training: inputting the enhanced time-frequency matrix dataset into the detection network for training, optimizing the network parameters through the loss function combined by DFL, CIoU Loss and BCE, and obtaining the trained detection network;

[0020] Step four, detecting the weak signal by using the trained network: inputting the time-frequency matrix of the weak signal to be detected into the detection network, completing the signal existence judgment, and obtaining the time-frequency information of the signal.

[0021] The specific method of step one is:

[0022] 1.1) weak signal time-frequency transformation;

[0023] The time-frequency matrix dataset of the weak signal is obtained through spectrum transformation, and the calculation method is as follows:

[0024]

[0025] In the formula, h(n) represents a window function, and N represents the length of a discrete window function;

[0026] 1.2) time-frequency matrix information labeling

[0027] The labeling information of the time-frequency matrix is the coordinates (x l ,yd ) and the upper right corner coordinate (x r ,y u ) is expressed as (x l ,y d ,x r ,y u ); through the analysis of the conversion relationship of the signal parameters of the known start time t start , the end time t end , the carrier frequency fc and the bandwidth B, the conversion formula of the remaining part and the priori parameter is obtained, as follows:

[0028]

[0029] wherein, Fs is the sampling frequency of the signal, nfft is the number of points of the signal spectrum diagram transformation, part_len is the window function length in the time-frequency transformation; through the above formula, the labeling information file of the time-frequency matrix is automatically generated while the time domain data of the weak signal is simulated.

[0030] The specific method of the step two is:

[0031] 2.1) Time-frequency matrix enhancement

[0032] The mean filtering and gamma transformation in the image enhancement field are combined to enhance the time-frequency matrix; first, the mean smoothing operation is performed on the time-frequency matrix of the weak signal in step one to reduce the high volatility of the data in the time-frequency matrix of the weak signal, and the calculation process is as follows:

[0033]

[0034] wherein, m and n are the spatial filter sizes, generally, m, n, a and b are all odd integers, and a=(m-1) / 2 and b=(n-1) / 2;

[0035] Then, the normalized time-frequency matrix is obtained, and the formula is as follows:

[0036]

[0037] wherein, TFR n and TFR m represent the normalized and mean smoothed time-frequency matrix respectively, and the value range of TFR n is in the interval [0, 1];

[0038] Finally, the nonlinear operation of the power of the value in the normalized time-frequency matrix is performed to increase the value of the signal part and relatively compress the value of the noise part, and the calculation process is as follows:

[0039] TFRe (x,y) = TFR n (x,y) γ

[0040] where TFR e and TFR n denote the enhanced and normalized time-frequency matrix respectively, and γ is an integer greater than 1;

[0041] 2.2) Time-frequency matrix cutting

[0042] According to the input size w i × h e required by the detection network, which corresponds to the number of columns and rows of the time-frequency matrix; the enhanced time-frequency matrix with size w e × h e is cut into multiple sub-matrices according to step 2.1);

[0043] First, the size of TFR p is increased by padding zeros, and the size of the padded time-frequency matrix TFR p is w p × h e , where w p is an integer multiple of w e , and h p is an integer multiple of h e , as shown in the following formula:

[0044]

[0045] where denotes rounding up, m is the number of cuttings per row, and n is the number of cuttings per column; the overlapping point of TFR p and TFR p is the top left corner of the matrix to facilitate subsequent cutting and label information conversion;

[0046] Next, the padded time-frequency matrix TFR c is cut, and the size of the cut time-frequency sub-matrix TFR i is w i × h c , which is cut into m × n, according to the position (i, j) of TFR e in TFR p , the absolute coordinates of the lower left corner and the upper right corner in TFR p are obtained as follows: The cutting of TFR c is completed according to the absolute coordinates;

[0047] Then, the original label information from step 1.2) is cut, and the cut label information is obtained. c With the time-frequency submatrix TFR c Correspondence; first based on the time-frequency submatrix TFR c In TFR p The absolute coordinates of the lower left and upper right corners are used to determine whether the area contains a signal component. and If the signal is included, the annotation information is cut; otherwise, the annotation information is not cut. The cutting formula is as follows:

[0048]

[0049] Finally, the segmented annotation information is converted into a deep learning object detection label file format, including three parts: the object category c, the normalized center coordinates (x, y) of the object, and the normalized length and width values ​​(w, h) of the object, represented as (c, x, y, w, h). The deep learning signal detection network contains one category, namely signal, so c = 0. The conversion formulas for the remaining parts are as follows:

[0050]

[0051] All the cut time-frequency submatrices TFR c The time-frequency matrix dataset of the detection network is obtained by integrating and naming the corresponding label files.

[0052] The specific method for step three is as follows:

[0053] By using a loss function that combines DFL, CIoU Loss, and BCE to guide the optimization algorithm in updating network parameters, the network can better fit the data.

[0054] First, DFL is used to enable the network to quickly focus the predicted location on the vicinity of the target location. The calculation method is as follows:

[0055] L DFL (S i ,S i+1 )=-((y i+1 -y)log(S i )+(yy i )log(S i+1 ))

[0056] Among them, S i and S i+1 By y i y i+1 y is determined, where y is the label value of the actual target;

[0057] Then, the CIoU Loss is used to calculate the IoU of the predicted box and the target box to further accurately predict the position, and the calculation method is as follows:

[0058]

[0059] Wherein, a is a weight function, and υ is a parameter for measuring the consistency of the length-width ratio of the predicted box and the real box, and the definitions are as follows:

[0060]

[0061]

[0062] Wherein, ρ is the Euclidean distance between the center points of the predicted box and the real box, and c is the diagonal distance of the smallest rectangle covering the predicted box and the real box simultaneously.

[0063] Finally, the Binary Cross Entropy (BCE) is used to calculate, and the calculation formula is as follows:

[0064]

[0065] Wherein, y i and p i are the class probability of the real frame and the predicted frame respectively.

[0066] The specific method of step four is:

[0067] First, cut the time-frequency sub-matrix TFR c of the signal to be detected after step two 2.2) into the time-frequency sub-matrix TFR c ;

[0068] Then, the detection label result of the time-frequency sub-matrix TFR c is converted into the form of the detection frame as follows:

[0069]

[0070] Wherein, (x lc ,y dc ) and (x rc ,y uc ) are the left lower corner coordinates and the right upper corner coordinates of the detection frame respectively, w i and h i are the input length and width dimensions required by the detection network, that is, the column number and row number of the time-frequency sub-matrix;

[0071] Then, according to the time-frequency sub-matrix TFR cAt the position (i, j) in the time-frequency matrix TFR, the relative coordinates are converted to absolute coordinates by the following formula:

[0072]

[0073] When there is an overlap between the detection boxes, x l and y u Take the minimum value of several overlapping boxes, x r and y d Take the maximum value of several overlapping boxes, get the detection box coordinates of the signal complete time-frequency matrix TFR;

[0074] Finally, the start and end time, center frequency and bandwidth of the weak signal are obtained according to the following formula, and the detection of the weak signal is completed:

[0075]

[0076] In the formula, Fs is the sampling frequency of the signal, nfft is the number of points of the signal spectrum diagram transformation, and part_len is the window function length in time-frequency transformation.

[0077] A weak signal detection method based on deep learning system, comprising:

[0078] A time-frequency transformation module is used in step one to obtain a time-frequency matrix dataset of the weak signal through spectrum diagram transformation;

[0079] An information labeling module is used in step one to obtain a labeling information file corresponding to the time-frequency matrix through a conversion formula of prior parameters;

[0080] A time-frequency matrix enhancement module is used in step two to enhance the time-frequency matrix of the weak signal and reduce the high volatility of noise by combining mean filtering and gamma transformation in the image enhancement field;

[0081] A time-frequency matrix cutting module is used in step two to cut the time-frequency matrix into time-frequency sub-matrices of the required size of the detection network without losing the characteristics of the signals in the time-frequency matrix;

[0082] A detection network training module is used in step three to optimize the network parameters by combining DFL, CIoU Loss and BCE loss functions to obtain a trained detection network;

[0083] A time-frequency sub-matrix detection module is used in step four to obtain the detection label results of the time-frequency sub-matrix TFR c by detecting all time-frequency sub-matrices TFR c ;

[0084] A time-frequency sub-matrix merging module is used in step four to merge the time-frequency sub-matrices TFRc The position in the time-frequency matrix TFR gets the absolute coordinates of the label information, the detection frame coordinates of the time-frequency matrix are obtained by merging, and the center frequency, bandwidth, start and end time information of the signal are obtained by conversion according to the formula.

[0085] A device of a weak signal detection method based on deep learning, specifically comprising:

[0086] A memory for storing a computer program;

[0087] A processor for implementing the weak signal detection method based on deep learning in steps one to four when executing the computer program.

[0088] A computer readable storage medium, the computer readable storage medium stores a computer program, the computer program is executed by the processor to be able to detect the weak signal based on the weak signal detection method based on deep learning in any one of claims 1 to 5.

[0089] Compared with the prior art, the present application has the following advantages:

[0090] For weak signal blind detection, the traditional signal detection method in the prior art scheme, including energy detection method, power spectrum detection method, correlation detection method and cyclostationary detection method, has low detection probability for non-cooperative signal under low signal-to-noise ratio condition, and can only obtain signal existence information and single dimension information in time domain or frequency domain, and cannot obtain information in both dimensions. The weak signal detection based on deep learning of the present application can greatly improve the accuracy of non-cooperative weak signal detection compared with the traditional signal detection method, and can obtain information of the signal in both time domain and frequency domain in addition to the existence information of the signal, which provides a prerequisite for subsequent signal analysis and processing.

[0091] The weak signal detection method in the prior art scheme has a large calculation amount of high-order spectrum analysis and is not suitable for scenes with high real-time requirements, the chaos theory needs to know the frequency of the periodic signal to be detected, and is not suitable for blind detection scenes, the adiabatic approximation theory of stochastic resonance machine resonance has strict restrictions on the frequency, amplitude and intensity of the input noise of the signal to be detected, and requires that these values must be small. This stringent condition makes the method not suitable for blind detection scenes of high-frequency communication signals. Some deep learning networks take time domain data or feature parameters of signals as input, although good detection performance is achieved, but the information of the signal in both time domain and frequency domain cannot be obtained. The weak signal detection based on deep learning of the present application has relatively less calculation amount compared with the existing weak signal detection method, and can process non-cooperative high-frequency weak communication signals. In addition to the existence information, information of the signal in both time domain and frequency domain can be obtained, which provides a prerequisite for subsequent signal analysis and processing. BRIEF DESCRIPTION OF DRAWINGS

[0092] Figure 1 is the flow chart of the weak signal detection algorithm based on deep learning of the present application.

[0093] Figure 2 is the experimental background chart of the weak signal detection algorithm based on deep learning of the present application.

[0094] Figure 3 is the deep learning detection network model chart of the present application.

[0095] Figure 4 is the detection performance comparison chart of the present application. DETAILED DESCRIPTION

[0096] The present application will be further described in detail below with reference to the accompanying drawings.

[0097] A weak signal detection method based on deep learning, see Figure 1 , specifically comprising the following steps:

[0098] Step one, constructing a weak signal time-frequency matrix dataset: obtaining a weak signal time-frequency matrix dataset through signal time-frequency transformation and time-frequency information labeling;

[0099] 1.1) Weak signal time-frequency transformation

[0100] The time-frequency matrix dataset of the weak signal is obtained through the spectral transformation, and the calculation method is as follows:

[0101]

[0102] In the formula, h(n) represents a window function, and N represents the length of a discrete window function;

[0103] 1.2) Time-frequency matrix information labeling

[0104] The labeling information of the time-frequency matrix is represented as (x l ,y d ,x r ,y u ) by the lower left corner coordinates (x l ,y d ) and the upper right corner coordinates (x r ,y u ) of the detection frame; through the analysis of the conversion relationship of the signal with the known start time t start , the cutoff time t end , the carrier frequency fc and the bandwidth B parameters converted into the time-frequency matrix, the conversion formula of the remaining part and the prior parameters is obtained, as shown below;

[0105]

[0106] Fs, nfft, part len where Fs is the sampling frequency of the signal, nfft is the number of points of the signal spectrogram transform, part len is the length of the window function in the time-frequency transform; comprehensive signal characteristics and window length selection, the sliding window overlap degree in this paper is 0, which refers to the proportion of the overlapping part between adjacent windows in the process of segmenting the signal using the window function. Through the above formula, the time-frequency matrix annotation file is automatically generated while the time-domain data of the weak signal is simulated, which reduces the workload and difficulty of data set construction, and also ensures the accuracy of data set annotation. Figure 3 The time-frequency graph of the high signal-to-noise ratio signal and the annotation information thereof.

[0107] Step two, time-frequency matrix preprocessing: through time-frequency matrix enhancement and time-frequency matrix cutting, an enhanced time-frequency matrix data set is obtained;

[0108] 2.1) Time-frequency matrix enhancement

[0109] The mean filtering and gamma transformation in the image enhancement field are combined to enhance the time-frequency matrix; first, the mean smoothing operation is performed on the time-frequency matrix data to reduce the high volatility of the data in the time-frequency matrix of the weak signal, and the calculation process is as follows:

[0110]

[0111] Where m and n are the spatial filter sizes, generally, m, n, a and b are all odd integers, and a = (m-1) / 2, b = (n-1) / 2.

[0112] Then the mean smoothed time-frequency matrix is normalized, and the formula is as follows:

[0113]

[0114] Where TFR n and TFR m represent the normalized and mean smoothed time-frequency matrix respectively, and the numerical range of TFR n is in the interval [0, 1].

[0115] Finally, the values in the time-frequency matrix are subjected to nonlinear operations of power to increase the values of the signal part and relatively compress the values of the noise part, and the calculation process is as follows:

[0116] TFR e (x,y) = TFR n (x,y) γ

[0117] Where TFR e and TFR nγ represents the normalized time-frequency matrix after enhancement and mean smoothing, respectively, where γ is an integer greater than 1.

[0118] 2.2) Time-frequency matrix cutting

[0119] According to the input length and width dimensions w required by the detection network i ×h i , corresponding to the number of columns and rows of the time-frequency matrix; with dimensions w e ×h e The enhanced time-frequency matrix after step 2.1) is divided into multiple sub-matrices;

[0120] First, fill zero to increase TFR e The size of the filled time-frequency matrix TFR p The length and width dimensions are w p ×h p , make w e For w p Integer multiples of h e for h p For multiples of integers, the formula is as follows:

[0121]

[0122] In the formula, This indicates rounding up, where m is the number of segments that can be divided into in each row, and n is the number of segments that can be divided into in each column; TFR e and TFR p The point of overlap is the top left corner of the matrix, to facilitate subsequent segmentation and conversion of label information;

[0123] Next, the time-frequency matrix TFR after cutting and filling is... p The time-frequency submatrix TFR after cutting c The size is equal to the required input dimensions w of the network. i ×h i Cut into m×n pieces, according to TFR c In TFR e Given the position (i,j) in the TFR, we can obtain its position in the TFR. p The absolute coordinates of the bottom left and top right corners are represented as follows: Complete the TFR according to absolute coordinates p The cutting;

[0124] Then, the original label information from step 1.2) is cut, and the cut label information is obtained. c With the time-frequency submatrix TFR c Correspondence; first based on the time-frequency submatrix TFR c In TFR pThe absolute coordinates of the lower left corner and the upper right corner are determined to determine whether they contain the signal part. If And If the signal is contained, the label information is cut, otherwise the label information is not cut, and the cutting formula is as follows:

[0125]

[0126] Finally, the cut label information is converted into a deep learning target detection label file format, including target class c, target normalized center coordinates (x, y), and target normalized length value (w, h) three parts, represented as (c, x, y, w, h); The signal detection network of deep learning contains one class, i.e. signal, so c=0, and the remaining part is converted as follows:

[0127]

[0128] All the cut time-frequency sub-matrices TFR c And the corresponding label file is integrated and named to obtain the time-frequency matrix data set of the detection network.

[0129] Step three, build a deep learning signal detection network and train: by inputting the enhanced time-frequency matrix set into the detection network for training, the network parameters are optimized through the loss function combining DFL, CIoU Loss and BCE, and the trained detection network is obtained;

[0130] The loss function combining DFL, CIoU Loss and BCE guides the optimization algorithm to update the network parameters, so that the network can better fit the data;

[0131] The weak signal detection network structure is composed of backbone network and head network. The main task of the backbone network is feature extraction, which extracts features from the input time-frequency graph to obtain feature maps of the time-frequency graph at different scales; The main task of the head network is feature fusion and prediction regression. First, the feature maps from the backbone network are fused, then the signal is detected on the fused feature maps, and the boundary of the detection frame is predicted and regressed to obtain the position information of the signal in the time-frequency graph. The network structure is shown in Figure 2 .

[0132] The loss function of the detection network mainly includes regression loss and classification loss.

[0133] The regression loss is composed of DFL and CIoU Loss: first, DFL is used to make the network quickly focus on the target position, and the calculation method is as follows:

[0134] L DFL (S i ,S i+1) = -((y i+1 -y)log(S i +(y-y i )log(S i+1 ))

[0135] wherein S i and S i+1 are determined by y i , y i+1 , y i , and y is the label value of the real target;

[0136] Then, the CIoU Loss is used to calculate the IoU of the predicted frame and the target frame, to further accurately predict the position, and the calculation method is as shown below:

[0137]

[0138] wherein a is a weight function, and υ is a parameter for measuring the consistency of the length-width ratio of the predicted frame and the real frame, and the definitions are as shown below.

[0139]

[0140]

[0141] wherein p is the Euclidean distance between the center points of the predicted frame and the real frame, and c is the diagonal distance of the smallest rectangle covering the predicted frame and the real frame.

[0142] The classification loss is used to calculate the category probability and the category loss, and the Binary Cross Entropy (BCE) is used for calculation, and the calculation formula is as shown below:

[0143]

[0144] wherein y i and p i are the category probabilities of the real frame and the predicted frame, respectively;

[0145] Through the above processing, the trained detection network is obtained.

[0146] Step four, using the trained network to detect the weak signal: by inputting the time-frequency matrix of the to-be-detected weak signal into the detection network, the signal existence judgment is completed, and the time-frequency information of the signal is obtained.

[0147] First, the time-frequency sub-matrix TFR c of the to-be-detected signal after cutting according to step two 2.2) is input into the trained detection network of step three, to obtain the detection label result of the time-frequency sub-matrix TFR c ;

[0148] Then, the detection label result of the time-frequency sub-matrix TFR is converted into the form of a detection frame according to the following formula: c

[0149]

[0150] Where (x lc ,y dc ) and (x rc ,y uc ) are the left lower corner coordinates and the right upper corner coordinates of the detection frame respectively, w i and h i are the input length and width dimensions required by the detection network, that is, the column number and the row number of the time-frequency sub-matrix;

[0151] Then, the relative coordinates are converted into absolute coordinates according to the following formula at the position (i,j) of the time-frequency sub-matrix TFR c in the time-frequency matrix TFR;

[0152]

[0153] When there is a detection frame overlap, x l and y u take the minimum value in the several overlapping frames, x r and y d take the maximum value in the several overlapping frames, to obtain the detection frame coordinates of the complete time-frequency matrix TFR of the signal;

[0154] Finally, the start and end time, the center frequency and the bandwidth of the weak signal are obtained according to the following formula, and the detection of the weak signal is completed:

[0155]

[0156] Where Fs is the sampling frequency of the signal, nfft is the point number of the signal spectrum graph transformation, and part_len is the window function length in the time-frequency transformation.

[0157] Simulation experiment:

[0158] The effects of the present application can be further proved by the following simulation experiment:

[0159] (1) Experimental environment and experimental data

[0160] The actual application background considered by the present algorithm is shown in Figure 3 In the absence of interference, due to the use of low probability of intercept communication technology by the enemy, the transmitted signal has small transmission power, so that the signals in the received data of the monitoring device are weak.

[0161] The simulation parameters of the weak signal collection data in the experiment are as follows: ​

[0162] Weak signal simulation parameters

[0163]

[0164] In order to increase the robustness and reliability of the simulation experiment, the symbol rate is a random value in the preset four values, the modulation mode is a random value in the preset common 18 modulation modes, the noise is Gaussian noise, the bit signal-to-noise ratio changes in the range of-5dB to 11dB with an increment of 2dB, 2000 sampling data are generated under each signal-to-noise ratio, the signal duration in each sampling data is also a random value, the starting time is a random value in 5ms-15ms, and the cutoff time is a random value in 30ms-50ms. The time resolution of the time-frequency matrix obtained by time-frequency transformation of the sampling data is 0.5ms in time and 305.1758Hz in frequency. After the time-frequency matrix preprocessing step, there are about 28000 time-frequency sub-matrices. Among them, the data amount ratio of the training set and the validation set is 9:1. In addition, the sampling data with the bit signal-to-noise ratio ranging from-8dB to 12dB with an increment of 2dB and the same parameters as the above table are generated as the test set of the algorithm performance, and 1000 sampling data are generated under each signal-to-noise ratio. In addition, the weak signal detection network model is constructed using Python language on the PyTorch2.1.2 framework, the GPU is NVIDIA Quadro T2000, the CUDA version is 12.2, and the CPU is i7-10750H.

[0165] (2) Simulation results and analysis

[0166] In this experiment, by inputting the experimental data into the energy detection method, the autocorrelation detection method and the method described in the present application respectively, the simulation results obtained are as shown in Figure 4 Figure 4 (a) is a detection probability curve, Figure 4 (b) is a false alarm probability curve. From the figure, it can be seen that whether the detection rate or the false alarm rate, the method described in the present application has better performance compared with the other two methods, and at 4dB, the detection rate is above 90% and the false alarm rate is below 5%.​

Claims

1. A weak signal detection method based on deep learning, characterized in that, Specifically, the following steps are included: Step 1: Construct a weak signal time-frequency matrix dataset: Obtain a weak signal time-frequency matrix dataset through signal time-frequency transformation and time-frequency information annotation; Step 2, Time-Frequency Matrix Preprocessing: The enhanced time-frequency matrix dataset is obtained through time-frequency matrix enhancement and time-frequency matrix segmentation; 2.1) Time-Frequency Matrix Enhancement The time-frequency matrix is ​​enhanced using a combination of mean filtering and gamma transform, techniques commonly used in image enhancement. First, mean smoothing is applied to the time-frequency matrix of the weak signal from step one to reduce the high volatility of the data. The calculation process is shown below: Where m and n are the sizes of the spatial filter, respectively. Generally, m, n, a, and b are all odd integers, and a = (m-1) / 2, b = (n-1) / 2; Then, the mean-smoothed time-frequency matrix is ​​normalized, as shown in the following formula: Among them, TFR n and TFR m Let TFR represent the normalized and mean-smoothed time-frequency matrices, respectively. n The numerical range is within the interval [0,1]. Finally, a nonlinear power operation is performed on the values ​​in the normalized time-frequency matrix to increase the signal value and compress the noise value. The calculation process is shown below: TFR e (x,y)=TFR n (x,y) γ Among them, TFR e and TFR n These represent the enhanced and normalized time-frequency matrices, respectively, where γ is an integer greater than 1; 2.2) Time-frequency matrix cutting According to the input length and width dimensions w required by the detection network i ×h i , corresponding to the number of columns and rows of the time-frequency matrix; with dimensions w e ×h e The enhanced time-frequency matrix after step 2.1) is divided into multiple sub-matrices; First, fill zero to increase TFR e The size of the filled time-frequency matrix TFR p The length and width dimensions are w p ×h p , make w e For w p Integer multiples of h e for h p For multiples of integers, the formula is as follows: In the formula, This indicates rounding up, where m is the number of segments that can be divided into in each row, and n is the number of segments that can be divided into in each column; TFR e and TFR p The point of overlap is the top left corner of the matrix, to facilitate subsequent segmentation and conversion of label information; Next, the time-frequency matrix TFR after cutting and filling is... p The time-frequency submatrix TFR after cutting c The size is equal to the required input dimensions w of the network. i ×h i Cut into m×n pieces, according to TFR c In TFR e Given the position (i,j) in the TFR, we can obtain its position in the TFR. p The absolute coordinates of the bottom left and top right corners are represented as follows: Complete the TFR according to absolute coordinates p The cutting; Then, the original label information from step 1.2) is cut, and the cut label information is obtained. c With the time-frequency submatrix TFR c Correspondence; first based on the time-frequency submatrix TFR c In TFR p The absolute coordinates of the lower left and upper right corners are used to determine whether the area contains a signal component. and If the signal is included, the annotation information is cut; otherwise, the annotation information is not cut. The cutting formula is as follows: Finally, the segmented annotation information is converted into a deep learning object detection label file format, including three parts: the object category c, the normalized center coordinates (x, y) of the object, and the normalized length and width values ​​(w, h) of the object, represented as (c, x, y, w, h). The deep learning signal detection network contains one category, namely signal, so c = 0. The conversion formulas for the remaining parts are as follows: All the cut time-frequency submatrices TFR c The time-frequency matrix dataset of the detection network is obtained by integrating and naming the corresponding label files. Step 3: Construct and train a deep learning signal detection network: The enhanced time-frequency matrix dataset is input into the detection network for training. The network parameters are optimized using a loss function that combines DFL, CIoU Loss and BCE to obtain a trained detection network. Step 4: Use the trained network to detect weak signals: By inputting the time-frequency matrix of the weak signal to be tested into the detection network, the existence of the signal is determined and the time-frequency information of the signal is obtained.

2. The weak signal detection method based on deep learning according to claim 1, characterized in that, The specific method for step one is as follows: 1.1) Time-frequency transformation of weak signals; The time-frequency matrix dataset of the weak signal is obtained through spectral transformation, and the calculation method is as follows: In the formula, h(n) represents the window function, and N represents the length of the discrete window function; 1.2) Time-frequency matrix information annotation The annotation information of the time-frequency matrix is ​​represented by the coordinates of the lower left corner of the detection box (x). l ,y d ) and the coordinates of the upper right corner (x r ,y u ) is represented as (x l ,y d ,x r ,y u ); By using the known start time t start Deadline t end The analysis of the transformation relationship between the signal with carrier frequency fc and bandwidth B as a time-frequency matrix yields the transformation formulas for the remaining parts and prior parameters, as shown below. Where Fs is the sampling frequency of the signal, nfft is the number of points in the signal spectrum transformation, and part_len is the length of the window function in the time-frequency transformation; using the above formula, while obtaining the time-domain data of the weak signal through simulation, the annotation information file of the time-frequency matrix is ​​automatically generated.

3. The weak signal detection method based on deep learning according to claim 1, characterized in that, The specific method for step three is as follows: By using a loss function that combines DFL, CIoU Loss, and BCE to guide the optimization algorithm in updating network parameters, the network can better fit the data. First, DFL is used to enable the network to quickly focus the predicted location on the vicinity of the target location. The calculation method is as follows: L DFL (S i ,S i+1 )=-((and i+1 -y)log(S i )+(yy i )log(S i+1 )) Among them, S i and S i+1 By y i y i+1 y is determined, where y is the label value of the actual target; Next, CIoU Loss is used to calculate the IoU between the predicted bounding box and the target bounding box to further refine the predicted location. The calculation method is as follows: Where α is the weighting function, and υ is a parameter that measures the consistency of the aspect ratio between the predicted bounding box and the ground truth bounding box, defined as follows: Where ρ is the Euclidean distance between the center points of the predicted box and the ground truth box, and c is the diagonal distance of the smallest rectangle that simultaneously covers the predicted box and the ground truth box. Finally, the binary cross entropy (BCE) is used for calculation, as shown in the following formula: Among them, y i and p i These represent the class probabilities of the ground truth bounding box and the predicted bounding box, respectively.

4. The weak signal detection method based on deep learning according to claim 1, characterized in that, The specific method for step four is as follows: First, the time-frequency submatrix TFR of the signal to be detected after being segmented in step 2.2) is... c Input the detection network trained in step three to obtain the time-frequency submatrix TFR c The results of the detection labels; Next, the time-frequency submatrix TFR is calculated using the following formula. c The detection label results are converted into detection bounding boxes: Among them, (x lc ,y dc ) and (x rc ,y uc ) represent the coordinates of the bottom left and top right corners of the detection box, respectively, w i and h i To determine the required input length and width dimensions for the detection network, i.e., the number of columns and rows of the frequency submatrix; Then, according to the time-frequency submatrix TFR c The relative coordinates of the position (i,j) in the time-frequency matrix TFR are converted to absolute coordinates according to the following formula; Iterate through all bounding boxes; when there are overlapping bounding boxes, x l and y u Take the smallest value from several overlapping boxes, x r and y d The maximum value in several overlapping frames is taken to obtain the detection frame coordinates of the complete time-frequency matrix (TFR) of the signal; Finally, the start and end times, center frequency, and bandwidth of the weak signal are obtained according to the following formula, thus completing the detection of the weak signal: In the formula, Fs is the sampling frequency of the signal, nfft is the number of points in the signal spectrum transformation, and part_len is the length of the window function in the time-frequency transformation.

5. A system for a weak signal detection method based on deep learning according to any one of claims 1 to 4, characterized in that, include: The time-frequency transformation module is used in step one to obtain the time-frequency matrix dataset of the weak signal through spectral transformation; The information annotation module is used in step one to obtain the annotation information file corresponding to the time-frequency matrix through the conversion formula of the prior parameters; The time-frequency matrix enhancement module is used in step two to enhance the time-frequency matrix of weak signals and reduce the high volatility of noise by combining mean filtering and gamma transformation in the field of image enhancement. The time-frequency matrix cutting module is used in step two to cut the time-frequency matrix into time-frequency sub-matrices of the required size for the detection network without losing the characteristics of the signal in the time-frequency matrix. The detection network training module is used in step three to optimize the network parameters using a loss function that combines DFL, CIoU Loss, and BCE to obtain a trained detection network. The time-frequency submatrix detection module is used in step four to detect all time-frequency submatrices (TFR). c The time-frequency submatrix TFR is obtained. c The results of the detection labels; The time-frequency submatrix merging module is used in step four to merge the time-frequency submatrix TFR. c The absolute coordinates of the tag information are obtained from its position in the time-frequency matrix (TFR). These coordinates are then combined to obtain the detection frame coordinates of the time-frequency matrix. The center frequency, bandwidth, and start and end time information of the signal are then obtained by converting the coordinates according to the formula.

6. The device for a weak signal detection method based on deep learning according to any one of claims 1 to 4, characterized in that: Specifically, it includes: Memory, used to store computer programs; A processor is used to implement the deep learning-based weak signal detection method described in steps one to four when executing the computer program.

7. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it can perform weak signal detection based on the deep learning-based weak signal detection method according to any one of claims 1 to 4.

Images