Low-spectral-density signal detection method based on time domain cooperative enhancement type artificial intelligence

Abstract

The invention discloses a low-spectral-density signal detection method based on time domain cooperative enhancement type artificial intelligence, and belongs to the technical field of safety communication. According to the method, the problems of low detection and classification precision and poor real-time performance of a traditional method are solved. According to the invention, cyclic self-correlation physical feature extraction is carried out on input short window and long window collaborative samples in a time domain, and detection and identification are carried out on low-spectral-density burst signals in combination with an enhanced network learning strategy. The implicit periodic statistical characteristics of the signals are mined through a cyclic autocorrelation function, I and Q signals are converted into a time delay-cyclic frequency matrix, and the characteristic separability of the signals under the condition of low spectral density is remarkably improved; through an enhanced network learning strategy, the increase of the complexity of an observation sample and the staged optimization of a training process are carried out, the internal relevance of signal features is further mined, the perception ability of a model to the existence of signals is enhanced, and then the fine recognition of signal types is gradually integrated. The method can be applied to low-spectral-density signal detection.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of secure communication, and particularly relates to a time-domain cooperative enhancement type artificial intelligence low spectral density signal detection method. BACKGROUND

[0002] With the rapid evolution of wireless communication technology and the increasing complexity of electromagnetic spectrum environment, the field of physical layer security is facing unprecedented challenges. In the scenarios of spectrum detection and Internet of Things, low spectral density burst covert signals have become the blind area of traditional signal detection technology due to their extremely low power spectral density, short burst characteristics and high dynamic modulation mode. Such signals show extremely low interception probability in time domain, frequency domain and energy domain, so that traditional detection methods such as energy detection, matched filtering, high-order statistics or time-frequency analysis are difficult to effectively deal with low spectral density burst covert signals, and there are problems such as insufficient detection sensitivity, high false alarm rate, and difficulty in balancing computational complexity and real-time performance. Under this background, it is urgent to develop a detection method based on the physical characteristics of the signal essence that can effectively suppress noise interference. The cyclic autocorrelation function can mine the periodic statistical characteristics of the signal, so as to detect weak signals with cyclic frequency in strong noise background, which is not possessed by other detection methods. However, the traditional cyclic autocorrelation method still faces the problem of low detection and classification accuracy caused by insufficient feature extraction when dealing with low spectral density burst covert signals, and the real-time performance of detection is also poor. SUMMARY

[0003] The purpose of the application is to solve the problems of low detection and classification accuracy and poor real-time performance of detection existing in traditional detection methods, and a time-domain cooperative enhancement type artificial intelligence low spectral density signal detection method is proposed.

[0004] The technical scheme adopted by the application to solve the above technical problems is: a time-domain cooperative enhancement type artificial intelligence low spectral density signal detection method, which specifically comprises the following steps:

[0005] Step one, setting sampling rate f s , signal type, signal time window length T1 and signal time window length T2;

[0006] Step two, generating a noisy signal sample according to the sampling rate f s , signal type, signal time window length T1 and T2;

[0007] Step three, adding detection labels and classification labels to the noisy signal sample generated in step two to obtain a first data set and a second data set;

[0008] The first data set is divided into two parts of training set and test set, to obtain the first training set and the first test set; the second data set is divided into two parts of training set and test set, to obtain the second training set and the second test set;

[0009] Step four, constructing an incremental network model, training the constructed network model with the first training set and the second training set, testing the constructed network model with the first test set and the second test set, stopping training when the loss of the network model on the first test set and the second test set no longer decreases, and obtaining the trained network model;

[0010] Step five, cutting the data to be analyzed into signal segments each with a time window length of T1, and recording each obtained signal segment as K is the number of obtained signal segments; each signal segment is converted into a time-delay-cyclic frequency matrix, and each time-delay-cyclic frequency matrix is taken as an input of the trained network model, and the probability of the existence of a signal in each signal segment is output by the detection head of the trained network model, and the probability of the existence of a signal in the k'th signal segment is recorded as

[0011] wherein, If then the k'th signal segment contains a signal, otherwise the k'th signal segment does not contain a signal;

[0012] If the k'th signal segment contains a signal, a signal segment composed of the latter half of the k'-1'th signal segment, the k'th signal segment and the former half of the k'+1'th signal segment is cut out from the data to be analyzed. The cut-out signal segment is converted into a time-delay-cyclic frequency matrix, and the time-delay-cyclic frequency matrix is taken as an input of the trained network model, and the probability of the signal in the signal segment belonging to each type is output by the classification head of the trained network model. The signal type in the signal segment is determined according to the output result of the classification head.

[0013] Further, the sampling rate f s = 1 kHz, the signal types include BPSK signal, QPSK signal, QAM16 signal, AM-DSB signal, FSK signal, OFDM signal, LoRa signal and pure noise signal, and each signal contains I and Q two channels.

[0014] Further, the specific process of step two is as follows:

[0015] Step two one, generate noiseless signal samples with time window length T1 and T2 for each signal type respectively in python platform;

[0016] Step two two, mix each signal sample generated in step two one with Gaussian white noise respectively to generate each noisy signal sample meeting the set signal-to-noise ratio;

[0017] The signal-to-noise ratio ranges from -20dB to 18dB, and a signal-to-noise ratio is set every 2dB interval.

[0018] Further, in step two one, noiseless signals with time window length T1 and T2 are generated for each signal type respectively, and the specific process is as follows:

[0019] Taking time window length T1 as an example:

[0020] (1) BPSK signal generation: randomly generate a binary sequence with length N1, map 0 in the binary sequence to -1, map 1 in the binary sequence to 1, and take the mapping result of the binary sequence as the data of I channel, and take 0 as the data of Q channel;

[0021] Wherein, N1=f s ×T1;

[0022] (2) QPSK signal generation: randomly generate a quaternary sequence with length N1, map 00 in the quaternary sequence to Map 01 in the quaternary sequence to Map 10 in the quaternary sequence to Map 11 in the quaternary sequence to Then the data in I channel are the cosine values of the mapping results in turn, and the data in Q channel are the sine values of the mapping results in turn;

[0023] (3) QAM16 signal generation: randomly generate a 16-ary sequence with length N1, map 0000 in the 16-ary sequence to -3-3i, map 0001 in the 16-ary sequence to -3-i, map 0010 in the 16-ary sequence to -3+3i, map 0011 in the 16-ary sequence to -3+i, map 0100 in the 16-ary sequence to -1-3i, map 0101 in the 16-ary sequence to -1-i, map 0110 in the 16-ary sequence to -1+3i, map 0111 in the 16-ary sequence to -1+i, map 1000 in the 16-ary sequence to 3-3i, map 1001 in the 16-ary sequence to 3-i, map 1010 in the 16-ary sequence to 3+3i, map 1011 in the 16-ary sequence to 3+i, map 1100 in the 16-ary sequence to 1-3i, map 1101 in the 16-ary sequence to 1-i, map 1110 in the 16-ary sequence to 1+3i, and map 1111 in the 16-ary sequence to 1+i;

[0024] Then, each data in the I channel is the real part of each mapping result in turn, and each data in the Q channel is the imaginary part of each mapping result in turn.

[0025] (4) AM-DSB signal generation: take a sine wave with frequency f c as a baseband signal, modulate the baseband signal with a carrier wave with frequency f m to obtain a modulated signal s(t) = A(1 + mcos(2pft m ))cos(2pft c ), take the modulated signal as the I channel data, and take 0 as the Q channel data.

[0026] Wherein, A is the amplitude, and m is the modulation index.

[0027] (5) FSK signal generation: randomly generate a binary number, map the generated binary number to a cosine signal cos(2pft), f is the frequency, sample K samples from the mapped cosine signal cos(2pft), K = f s / R, and R is the bit rate.

[0028] Similarly, each binary number is generated in turn and processed respectively, the sampled samples are used to form the I channel data of the FSK signal, and the length of the I channel data is N1; the Q channel data of the FSK signal is all 0.

[0029] When the generated binary number is 0, f = 50 Hz; when the generated binary number is 1, f = 100 Hz.

[0030] (6) OFDM signal generation: define the number of subcarriers as 116, define the number of FFT points as 128, then the data of the 59th subcarrier to the 116th subcarrier are sequentially taken as the 1st to 58th data, and 12 zeros are added after the 58th data, then the data of the 1st subcarrier to the 58th subcarrier are sequentially taken as the 71st to 128th data, to obtain data with a length of 128;

[0031] Perform inverse fast Fourier transform on the data with a length of 128 to generate a time domain signal, then take the last 32 bits of the generated time domain signal as a cyclic prefix, copy the cyclic prefix to the front of the generated time domain signal, that is, concatenate the cyclic prefix and the generated time domain signal as an OFDM symbol, and the number of points of each OFDM symbol is 160;

[0032] Take N1 / 160 OFDM symbols, then the I channel data of the OFDM signal is the real part of the OFDM symbol, and the Q channel data of the OFDM signal is the imaginary part of the OFDM symbol;

[0033] (7) LoRa signal generation: generate a linear chirp signal with a frequency from f0 to f'0, then the I channel data is cos(2πf(t)t), and the Q channel data is sin(2πf(t)t), wherein f(t) represents a frequency varying with time, and f(t)∈[f0,f'0];

[0034] (8) Pure noise signal generation: the I channel is Gaussian white noise with a mean of 0 and a variance of 1, and the Q channel is Gaussian white noise with a mean of 0 and a variance of 1.

[0035] Further, in the step two, each generated signal sample is mixed with Gaussian white noise, and the calculation method of the Gaussian white noise power is:

[0036]

[0037] Wherein, P signal is the signal power, P noise is the noise power, and SNR represents the signal-to-noise ratio of the noisy signal sample;

[0038] Then, according to the calculated noise power P noise , Gaussian white noise with a mean of 0 and a variance of P noise is generated, and the generated Gaussian white noise is added to the I channel and the Q channel of the noise-free signal respectively to obtain a noisy signal sample.

[0039] Further, the specific process of the step three is:

[0040] Step three one, adding labels and obtaining a data set

[0041] Adding detection labels to signal samples with length N1 and length N2: detection label y of pure noise signal d = 0; detection label y of BPSK signal, QPSK signal, QAM16 signal, AM-DSB signal, FSK signal, OFDM signal and LoRa signal d = 1;

[0042] Adding classification labels to signal samples with length N2 only: classification label y of BPSK signal c = 0, classification label y of QPSK signal c = 1, classification label y of QAM16 signal c = 2, classification label y of AM-DSB signal c = 3, classification label y of FSK signal c = 4, classification label y of OFDM signal c = 5, classification label y of LoRa signal c = 6;

[0043] Using signal samples with length N1 with detection labels to form a first data set dataset_1, and using signal samples with length N2 with detection labels and classification labels to form a second data set dataset_2;

[0044] Step three two, record the first data set as denotes the i-th signal sample in the first data set, M1 represents the number of signal samples with length N1, denotes the detection label of the i-th signal sample, and the first data set is divided into a first training set and a first test set, and the number of signal samples in the first training set and the first test set is in the ratio of 7:3;

[0045] Record the second data set as denotes the j-th signal sample in the second data set, M2 represents the number of signal samples with length N2, denotes the detection label of the j-th signal sample, denotes the classification label of the j-th signal sample, and the second data set is divided into a second training set and a second test set, and the number of signal samples in the second training set and the second test set is in the ratio of 7:3.

[0046] Further, the working process of the improved network model is:

[0047] Taking the input of the improved network model as the input of the first convolutional layer, and taking the output of the first convolutional layer as the input of the first relu activation function layer;

[0048] The output of the first relu activation function layer is taken as the input of the first max pooling layer, the output of the first max pooling layer is taken as the input of the second convolutional layer, and the output of the second convolutional layer is taken as the input of the second relu activation function layer;

[0049] The output of the second relu activation function layer is taken as the input of the second max pooling layer, the output of the second max pooling layer is taken as the input of the third convolutional layer, and the output of the third convolutional layer is taken as the input of the third relu activation function layer;

[0050] The output of the third relu activation function layer is taken as the input of the global average pooling layer, and the output of the global average pooling layer is taken as the input of the detection head and the classification head, respectively; wherein:

[0051] The detection head sequentially includes a first full connection layer, a fourth relu activation function layer, a second full connection layer and a sigmoid activation function layer, and outputs the probability of the signal sample existing signal through the sigmoid activation function layer;

[0052] The classification head sequentially includes a third full connection layer, a fifth relu activation function layer, a fourth full connection layer and a softmax activation function layer, and outputs the probability of the signal sample belonging to various signal types through the softmax activation function layer.

[0053] Further, the training process of the improved network model is:

[0054] Step four one, randomly extract 80% of the signal samples from the first training set and 20% of the signal samples from the second training set;

[0055] Step four two, each signal sample is respectively preprocessed by cyclic autocorrelation, and each signal sample is respectively converted into a time delay-cyclic frequency matrix through the cyclic autocorrelation preprocessing;

[0056] Step four three, the time delay-cyclic frequency matrix corresponding to each signal sample is taken as the input of the improved network model; the time delay-cyclic frequency matrix corresponding to each signal sample is used to train the improved network model in the first stage;

[0057] The parameters of the improved network model are adjusted according to the loss function during training, and the current training stage is stopped when the loss of the model on the first test set and the second test set does not decrease continuously for several rounds;

[0058] Step four four, 50% of the signal samples are randomly extracted from the first training set and 50% of the signal samples are randomly extracted from the second training set, and each signal sample is respectively converted into a time delay-cyclic frequency matrix through the cyclic autocorrelation preprocessing;

[0059] Step four five, using the time-delay-cyclic frequency matrix obtained in step four four, the model is trained in the second stage, and the parameters of the progressive network model are adjusted according to the loss function during training, until the loss of the model on the first test set and the second test set no longer decreases for several rounds, and the training of the current training stage is stopped;

[0060] Step four six, 20% of the signal samples are randomly extracted from the first training set, and 80% of the signal samples are randomly extracted from the second training set, and each randomly extracted signal sample is converted into a time-delay-cyclic frequency matrix through cyclic autocorrelation preprocessing;

[0061] Step four seven, using the time-delay-cyclic frequency matrix obtained in step four six, the model is trained in the third stage, and the parameters of the progressive network model are adjusted according to the loss function during training, until the loss of the model on the first test set and the second test set no longer decreases for several rounds, and the training of the current training stage is stopped, that is, the finally trained model is obtained.

[0062] Further, the specific process of step four two is:

[0063] Step four two one, for any signal sample randomly extracted in step four one Convert the signal sample into a complex number form x(n):

[0064] x(n) = I(n) + jQ(n

[0065] Wherein, I(n) represents the n-th bit of the I channel data of the signal sample , Q(n) represents the n-th bit of the Q channel data of the signal sample , n = 0, 1, …, N1-1, and j represents an imaginary unit;

[0066] Calculate the cyclic autocorrelation function value of x(n)

[0067]

[0068] Wherein, τ represents the time delay, the unit is millisecond, τ = 0, 1, …, D1, D1 represents the time delay dimension, α represents the cyclic frequency, the unit is Hz, α = 0, 1, …, F1, x(n+τ) = I(n+τ) + jQ(n+τ), I(n+τ) represents the n+τ-th bit of the I channel data of the signal sample , Q(n+τ) represents the n+τ-th bit of the Q channel data of the signal sample , x * (n+τ) represents the conjugate of x(n+τ), e represents the base number of natural logarithm, and j represents an imaginary unit;

[0069] the cyclic autocorrelation function value of the signal sample modulo, the modulo result is taken as an element in a time-delay-cyclic frequency matrix .

[0070] Step four two, for any signal sample randomly extracted in step four one convert the signal sample into a complex number form x(n):

[0071] x(n) = I(n) + jQ(n

[0072] wherein I(n) represents the n-th bit in the I channel data of the signal sample , Q(n) represents the n-th bit in the Q channel data of the signal sample , n = 0, 1, …, N2-1;

[0073] calculate the cyclic autocorrelation function value of x(n)

[0074]

[0075] wherein τ represents time delay, unit: millisecond, τ = 0, 1, …, D2, D2 represents time delay dimension, α represents cyclic frequency, unit: Hz, α = 0, 0.5, 1, …, F2, x(n+τ) = I(n+τ) + jQ(n+τ), I(n+τ) represents the n+τ-th bit in the I channel data of the signal sample , Q(n+τ) represents the n+τ-th bit in the Q channel data of the signal sample , x * (n+τ) represents the conjugate of x(n+τ), e represents the base number of natural logarithm, j represents the imaginary unit;

[0076] take the cyclic autocorrelation function value of the signal sample modulo, the modulo result is taken as an element in a time-delay-cyclic frequency matrix .

[0077] Further, the loss function adopted in the first stage training is :

[0078]

[0079] the loss function adopted in the second stage training is :

[0080] ​​

[0081] Loss function adopted in the third stage training is:

[0082]

[0083] wherein, denotes the detection task loss function, denotes the classification task loss function;

[0084]

[0085] wherein, y d denotes the detection label, p d denotes the probability of the existing signal detected by the model, y c,k denotes the classification label, p c,k denotes the probability of the classification result of the model belonging to the kth signal.

[0086] The beneficial effects of the present application are:

[0087] The present application performs cyclic autocorrelation physical feature extraction on the short window and long window of the input time domain in cooperation with the sample, and combines with the progressive network learning strategy to detect and identify the low spectral density burst signal. The cyclic autocorrelation function is used to mine the periodic statistical characteristics of the signal, and the I and Q signals are converted into a time delay-cyclic frequency matrix, which significantly improves the feature separability of the signal under low spectral density conditions; through the progressive network learning strategy, the increase in the complexity of the observation sample is combined with the stage optimization of the training process, which further mines the internal correlation of the signal features, first strengthens the perception ability of the model to the signal existence, and then gradually integrates the fine recognition of the signal type, ensuring the synchronous improvement of the detection and classification performance. Through the cooperative sample input of the short time window and the long time window in the time domain, it provides the possibility for the network to mine more information quantity; the present application forms a complete closed-loop optimization path from physical feature extraction to model training, and realizes efficient detection and accurate classification of low spectral density burst signals under low SNR environment, which can significantly improve the sensitivity, accuracy and real-time performance of the detection system, and provides a new solution for the signal processing technology in the physical layer security field. BRIEF DESCRIPTION OF DRAWINGS

[0088] Figure 1 is a flowchart of a time domain cooperative progressive artificial intelligence low spectral density signal detection method according to the present application;

[0089] Figure 2 is a flowchart of generating data samples according to the present application. DETAILED DESCRIPTION

[0090] Detailed implementation one, combined with Figure 1 and Figure 2The embodiment is described. The time domain cooperative enhancement artificial intelligence low spectral density signal detection method described in the embodiment specifically comprises the following steps:

[0091] Step one, setting a sampling rate f s (representing the number of sampled points per second), a signal type, a signal time window length T1 and a signal time window length T2 (in the present application, the values of T1 and T2 are 1 second and 2 seconds respectively, the sample point number of the signal with the time window length T1 is N1, N1=f s ×1, and the sample point number of the signal with the time window length T2 is N2, N2=f s ×2);

[0092] Step two, generating noisy signal samples according to the sampling rate f s , the signal type, the signal time window length T1 and T2;

[0093] Step three, adding detection labels and classification labels to the noisy signal samples generated in step two to obtain a first data set and a second data set;

[0094] The first data set is divided into a training set and a test set to obtain a first training set and a first test set, and the second data set is divided into a training set and a test set to obtain a second training set and a second test set;

[0095] Step four, constructing an enhancement network model, training the constructed network model by using the first training set and the second training set, testing the constructed network model by using the first test set and the second test set, stopping training when the loss of the network model on the first test set and the second test set no longer decreases, and obtaining a trained network model;

[0096] Step five, dividing the data to be analyzed into segments with a signal time window length T1, and recording each obtained signal segment as K is the number of obtained signal segments; each signal segment is converted into a time delay-cyclic frequency matrix, each time delay-cyclic frequency matrix is taken as an input of the trained network model, the probability of the existence of a signal in each signal segment is output by a detection head of the trained network model, and the probability of the existence of a signal in the k'th signal segment is recorded as

[0097] wherein, If , the k'th signal segment contains a signal, otherwise the k'th signal segment does not contain a signal.

[0098] If there is a signal in the k'th signal segment, a signal consisting of the second half of the k'-1th signal segment, the k'th signal segment, and the first half of the k'+1th signal segment is extracted from the data to be analyzed. (If the boundary is exceeded, fill it with zeros), the intercepted signal Convert it into a time delay-cycle frequency matrix, use the time delay-cycle frequency matrix as the input of the trained network model, and output the signal fragment through the classification head of the trained network model The probability that the signal in belongs to each type Determine the signal segment based on the output of the classification head The signal type in .

[0099] The signal type is determined as follows:

[0100] (1) Obtain the maximum probability and the second maximum probability output by the classification head. If the maximum probability does not exceed 0.5 and the difference between the maximum probability and the second maximum probability does not exceed 0.2, the signal type is a type other than BPSK signal, QPSK signal, QAM16 signal, AM-DSB signal, FSK signal, OFDM signal, and LoRa signal;

[0101] (2) If the maximum probability exceeds 0.5, or if the maximum probability does not exceed 0.5 but the difference between the maximum probability and the second highest probability exceeds 0.2, the signal type is the signal type corresponding to the maximum probability.

[0102] The core design idea of ​​the method of the present invention is derived from the simulation of the human cognitive process, that is, through the fusion of experience accumulation and information increment, the neural network can gradually transition from simple signal existence judgment to accurate identification of signal type when facing a low spectral density burst covert signal. Specifically, the method of the present invention aims to enable the network to perceive the presence or absence of signals only through short-time window (such as 1 second) samples in the initial stage, which is similar to the initial judgment of humans on unfamiliar things; as the time window is extended (such as 2 seconds) and the complexity of the observed samples increases, the network can be exposed to richer information, thereby further distinguishing the specific type of signal based on the previous detection experience. The method of the present invention can also be further understood from the perspective of information increment. The short-time window sample has a limited amount of information and is suitable for rapid detection of signal existence. As the observation time increases, the statistical laws of the signal gradually emerge, providing a reliable basis for classification, that is, the long-time window sample contains more detailed features, which makes it possible to finely classify the signal type.

[0103] By combining the cyclic autocorrelation physical feature extraction and the incremental network learning strategy, the application can significantly improve the classification accuracy while ensuring the detection sensitivity, and realize the synchronous optimization of detection and recognition performance. The application can quickly respond to the real-time detection demand of hidden signals in the Internet of Things scene, and meet the accurate identification requirement of signal types, which has significant value.

[0104] Specific implementation method two: different from the specific implementation method one, the sampling rate f s = 1 kHz, the signal type contains BPSK signal, QPSK signal, QAM16 signal, AM-DSB signal, FSK signal, OFDM signal, LoRa signal and pure noise signal, and each signal contains I (in-phase) and Q (quadrature) two channels.

[0105] The other steps and parameters are the same as those in the specific implementation method one.

[0106] In the application, the sample dimension of the signal with a time window length of 1 second is (N1, 2), and the sample dimension of the signal with a time window length of 2 seconds is (N2, 2).

[0107] Specific implementation method three: different from the specific implementation method two, the specific process of step two is:

[0108] Step two one, generating noise-free signal samples with time window lengths of T1 and T2 for each signal type in the python platform respectively;

[0109] Step two two, mixing each signal sample generated in step two one with Gaussian white noise respectively to generate each noisy signal sample meeting the set signal-to-noise ratio;

[0110] The signal-to-noise ratio (SNR) ranges from -20 dB to 18 dB, and a signal-to-noise ratio is set every 2 dB interval.

[0111] The other steps and parameters are the same as those in the specific implementation method two.

[0112] Specific implementation method four: different from the specific implementation method three, in step two one, noise-free signals with time window lengths of T1 and T2 are generated for each signal type, and the specific process is:

[0113] Taking the time window length T1 as an example:

[0114] (1) BPSK signal generation: randomly generating a binary sequence with a length of N1, mapping 0 in the binary sequence to -1, mapping 1 in the binary sequence to 1, taking the mapping result of the binary sequence as the data of the I channel, and taking 0 as the data of the Q channel;

[0115] Wherein, N1=f s ×T1;

[0116] (2) QPSK signal generation: randomly generate a quaternary sequence with length N1, map 00 in the quaternary sequence to Map 01 in the quaternary sequence to Map 10 in the quaternary sequence to Map 11 in the quaternary sequence to Then, each data in the I channel is the cosine value of each mapping result in turn, and each data in the Q channel is the sine value of each mapping result in turn.

[0117] (3) QAM16 signal generation: randomly generate a 16-ary sequence with length N1, map 0000 in the 16-ary sequence to -3-3i, map 0001 in the 16-ary sequence to -3-i, map 0010 in the 16-ary sequence to -3+3i, map 0011 in the 16-ary sequence to -3+i, map 0100 in the 16-ary sequence to -1-3i, map 0101 in the 16-ary sequence to -1-i, map 0110 in the 16-ary sequence to -1+3i, map 0111 in the 16-ary sequence to -1+i, map 1000 in the 16-ary sequence to 3-3i, map 1001 in the 16-ary sequence to 3-i, map 1010 in the 16-ary sequence to 3+3i, map 1011 in the 16-ary sequence to 3+i, map 1100 in the 16-ary sequence to 1-3i, map 1101 in the 16-ary sequence to 1-i, map 1110 in the 16-ary sequence to 1+3i, and map 1111 in the 16-ary sequence to 1+i.

[0118] Then, each data in the I channel is the real part of each mapping result in turn, and each data in the Q channel is the imaginary part of each mapping result in turn.

[0119] (4) AM-DSB signal generation: take a sine wave with frequency f c (taking 100 Hz) as a baseband signal, modulate the baseband signal by a carrier wave with frequency f m (taking 10 Hz), to obtain a modulated signal s(t)=A(1+mcos(2πf m t))cos(2πf c t), take the modulated signal as the I channel data, and take 0 as the Q channel data.

[0120] Wherein, A is an amplitude (in the present application, the value of A is 1), and m is a modulation index (in the present application, the value of m is 0.5).

[0121] (5) FSK signal generation: randomly generate a binary number, map the generated binary number to a cosine signal cos(2πft), f is the frequency, sample K samples from the mapped cosine signal cos(2πft), K = f s R, R is the bit rate (in the present application, R = 10 bps, i.e. 10 bits per second);

[0122] Similarly, each binary number is generated in turn and processed, and the sampled samples are used to form the I channel data of the FSK signal, and the length of the I channel data is N1; the Q channel data of the FSK signal is all 0;

[0123] When the generated binary number is 0, f = 50 Hz; when the generated binary number is 1, f = 100 Hz;

[0124] (7) OFDM signal generation: define the number of subcarriers as 116 (i.e. 116 actual data), and define the FFT point number as 128, then the data of the 59th subcarrier to the 116th subcarrier are sequentially taken as the 1st to 58th data, and 12 zeros are appended after the 58th data, and then the data of the 1st subcarrier to the 58th subcarrier are sequentially taken as the 71st to 128th data, to obtain data with a length of 128;

[0125] Perform inverse fast Fourier transform (IFFT) on the data with a length of 128 to generate a time domain signal, and then take the last 32 bits of the generated time domain signal as a cyclic prefix, and copy the cyclic prefix to the front of the generated time domain signal, i.e. concatenate the cyclic prefix and the generated time domain signal as an OFDM symbol, and the point number of each OFDM symbol is 160;

[0126] Take N1 / 160 OFDM symbols, then the I channel data of the OFDM signal is the real part of the OFDM symbol, and the Q channel data of the OFDM signal is the imaginary part of the OFDM symbol;

[0127] (7) LoRa signal generation: generate a linear chirp signal with a frequency from f0 (e.g. -50 Hz) to f'0 (e.g. 50 Hz), then the I channel data is cos(2πf(t)t), and the Q channel data is sin(2πf(t)t), where f(t) represents the frequency varying with time, f(t) ∈ [f0, f'0];

[0128] (8) Pure noise signal generation: the I channel is Gaussian white noise with a mean of 0 and a variance of 1, and the Q channel is Gaussian white noise with a mean of 0 and a variance of 1.

[0129] The other steps and parameters are the same as in the third embodiment.

[0130] Similarly, the method of this embodiment can generate a noise-free signal with a time window length of T2 for each signal type. Furthermore, multiple signal samples with a time window length of T1 and multiple signal samples with a time window length of T2 can be generated for each signal type to ensure that the training set data has a certain size.

[0131] Specific embodiment 5: This embodiment differs from specific embodiment 4 in that, in step 22, each generated signal sample is mixed with Gaussian white noise respectively, and the calculation method of Gaussian white noise power is:

[0132]

[0133] Among them, P signal is the signal power (the average of the signal I channel power and Q channel power), P noise is the noise power, SNR represents the signal-to-noise ratio of the noisy signal sample;

[0134] Then according to the calculated noise power P noise , generating a mean of 0 and a variance of P noise The generated Gaussian white noise is added to the I channel and Q channel of the noise-free signal respectively to obtain a noisy signal sample (it should be noted here that in order to be similar to the noise-free signal, noise also needs to be added to the pure noise signal generated in the fourth specific implementation method).

[0135] Other steps and parameters are the same as those in the fourth embodiment.

[0136] For any signal type, a certain number of samples of that signal type are generated at each SNR value and each time window length (for example, 100 noisy signal samples are generated for each combination of signal type, SNR value, and time window length), and the dimension of the noisy signal samples is still (N1, 2) or (N2, 2).

[0137] Specific embodiment 6: This embodiment differs from specific embodiment 5 in that the specific process of step 3 is as follows:

[0138] Step 3.1. Add labels and obtain datasets

[0139] Add detection labels to signal samples of length N1 and length N2: detection label y of pure noise signal d =0; detection tag y for BPSK signal, QPSK signal, QAM16 signal, AM-DSB signal, FSK signal, OFDM signal and LoRa signal d =1;

[0140] Only add classification labels to signal samples of length N2: classification label y of BPSK signalc = 0, classification label y of QPSK signal c = 1, classification label y of QAM16 signal c = 2, classification label y of AM-DSB signal c = 3, classification label y of FSK signal c = 4, classification label y of OFDM signal c = 5, classification label y of LoRa signal c = 6

[0141] The first data set dataset_1 is composed of signal samples with detection label and length n1, and the second data set dataset_2 is composed of signal samples with detection label and classification label and length N2;

[0142] Step three two, the first data set is recorded as represents the i-th signal sample in the first data set, M1 represents the number of signal samples with length N1, represents the detection label of the i-th signal sample, and the first data set is divided into a first training set and a first test set, and the number of signal samples in the first training set and the first test set is 7:3;

[0143] The second data set is recorded as represents the j-th signal sample in the second data set, M2 represents the number of signal samples with length N2, represents the detection label of the j-th signal sample, represents the classification label of the j-th signal sample, and the second data set is divided into a second training set and a second test set, and the number of signal samples in the second training set and the second test set is 7:3.

[0144] The other steps and parameters are the same as those in the fifth embodiment.

[0145] In the present application, M1 = M2 = 16000 (8 types of signal types x 20 SNRs x 100 samples), the sample matrix composed of the first data set has a dimension of The dimension of the detection label vector is The sample matrix composed of the second data set has a dimension of The dimension of the detection label vector is The dimension of the classification label vector is The first dataset dataset_1 and the second dataset dataset_2 are respectively divided into a training set (70%) and a test set (30%), and for the first dataset dataset_1, the number of training set samples is 0.7M1, denoted as a first training set The number of test set samples is 0.3M1, denoted as a first test set For the second dataset dataset_2, the number of training set samples is 0.7M2, denoted as a second training set The number of test set samples is 0.3M2, denoted as a second test set The obtained training set and test set are used for subsequent training and evaluation of the model.

[0146] Specific embodiment seven: the difference between this embodiment and specific embodiment six is that the working process of the enhanced network model is:

[0147] The input of the enhanced network model is taken as the input of the first convolutional layer, and the output of the first convolutional layer is taken as the input of the first relu activation function layer;

[0148] The output of the first relu activation function layer is taken as the input of the first max pooling layer, and the output of the first max pooling layer is taken as the input of the second convolutional layer, and the output of the second convolutional layer is taken as the input of the second relu activation function layer;

[0149] The output of the second relu activation function layer is taken as the input of the second max pooling layer, and the output of the second max pooling layer is taken as the input of the third convolutional layer, and the output of the third convolutional layer is taken as the input of the third relu activation function layer;

[0150] The output of the third relu activation function layer is taken as the input of the global average pooling layer, and the output of the global average pooling layer is taken as the input of the detection head and the classification head respectively; wherein:

[0151] The detection head sequentially includes a first full connection layer, a fourth relu activation function layer, a second full connection layer and a sigmoid activation function layer, and outputs the probability of the signal sample existing signal through the sigmoid activation function layer;

[0152] The classification head sequentially includes a third full connection layer, a fifth relu activation function layer, a fourth full connection layer and a softmax activation function layer, and outputs the probability of the signal sample belonging to various signal types through the softmax activation function layer.

[0153] The other steps and parameters are the same as those of specific embodiment six.

[0154] Specific embodiment eight: the difference between this embodiment and specific embodiment seven is that the training process of the enhanced network model is:

[0155] Step four one, randomly extract 80% of the signal samples from the first training set and 20% of the signal samples from the second training set;

[0156] Step four two, respectively perform cyclic autocorrelation preprocessing on each of the randomly extracted signal samples, and respectively convert each of the signal samples into a time-delay-cyclic frequency matrix through the cyclic autocorrelation preprocessing;

[0157] Step four three, use the time-delay-cyclic frequency matrix corresponding to each of the signal samples as the input of the progressive network model, and perform first-stage training of the progressive network model using the time-delay-cyclic frequency matrix corresponding to each of the signal samples;

[0158] Adjust the parameters of the progressive network model according to the loss function during training until the loss of the model on the first test set and the second test set no longer decreases for a plurality of rounds, and stop the training of the current training stage;

[0159] Step four four, randomly extract 50% of the signal samples from the first training set and 50% of the signal samples from the second training set, and respectively convert each of the randomly extracted signal samples into a time-delay-cyclic frequency matrix through the cyclic autocorrelation preprocessing;

[0160] Step four five, perform second-stage training of the model using the time-delay-cyclic frequency matrix obtained in step four four, and adjust the parameters of the progressive network model according to the loss function during training until the loss of the model on the first test set and the second test set no longer decreases for a plurality of rounds, and stop the training of the current training stage;

[0161] Step four six, randomly extract 20% of the signal samples from the first training set and 80% of the signal samples from the second training set, and respectively convert each of the randomly extracted signal samples into a time-delay-cyclic frequency matrix through the cyclic autocorrelation preprocessing;

[0162] Step four seven, perform third-stage training of the model using the time-delay-cyclic frequency matrix obtained in step four six, and adjust the parameters of the progressive network model according to the loss function during training until the loss of the model on the first test set and the second test set no longer decreases for a plurality of rounds, and stop the training of the current training stage, i.e., obtain the finally trained model.

[0163] The other steps and parameters are the same as those in the seventh embodiment.

[0164] The model training of the application is divided into three stages, each stage adjusts the proportion of short window (1 second) and long window (2 seconds) data, stage 1 (detection priority): 80% short window + 20% long window, that is, 80% x 0.7M1 samples are randomly extracted from the first training set, 20% x 0.7M2 samples are randomly extracted from the second training set, and the total sample number M is 11200; stage 2 (progressive learning): 50% short window + 50% long window, that is, 50% x 0.7M1 samples are randomly extracted from the first training set, 50% x 0.7M2 samples are randomly extracted from the second training set, and the total sample number is 11200; stage 3 (fine classification): 20% short window + 80% long window, that is, 20% x 0.7M1 samples are randomly extracted from the first training set, 80% x 0.7M2 samples are randomly extracted from the second training set, and the total sample number is 11200. Stage 1: short window Long window Stage 2: short window Long window Stage 3: short window Long window 1 second time window sample corresponding time delay-cyclic frequency matrix 2 second time window sample corresponding time delay-cyclic frequency matrix The input time delay-cyclic frequency matrix first passes through the first convolutional layer Conv2D(32,(3,3)), the output of the first convolutional layer is taken as the input of the first relu activation function layer, the dimension of the output of the first relu activation function layer is (D-2,F-2,32), and then the output of the first relu activation function layer is taken as the input of the first maximum pooling layer MaxPooling2D(2,2), the dimension of the output of the first maximum pooling layer is The output of the first maximum pooling layer is taken as the input of the second convolutional layer Conv2D(64,(3,3)), and the output of the second convolutional layer is taken as the input of the second relu activation function layer. The dimension of the output of the second relu activation function layer is The output of the second relu activation function layer is taken as the input of the second maximum pooling layer MaxPooling2D(2,2), and the dimension of the output of the second maximum pooling layer is The output of the second maximum pooling layer is taken as the input of the third convolutional layer Conv2D(128,(3,3)), and the dimension of the output of the third convolutional layer is The output of the third convolutional layer is taken as the input of the third relu activation function layer, and the output of the third relu activation function layer is taken as the input of the global pooling layer GlobalAveragePooling2D. The number of elements output by the global pooling layer is 128;

[0165] The output of the global pooling layer is taken as the input of the detection head. In the detection head, the output of the global pooling layer passes through a first fully connected layer, and the output of the first fully connected layer is taken as the input of a fourth ReLU activation function layer. The number of elements output by the fourth ReLU activation function layer is 64. The output of the fourth ReLU activation function layer is taken as the input of a second fully connected layer. The output of the second fully connected layer is taken as the input of a sigmoid activation function. The number of elements output by the sigmoid activation function is 1 (the probability p of the sample existing signal output by the sigmoid activation function d ). In the classification head, the output of the global pooling layer passes through a third fully connected layer, and the output of the third fully connected layer is taken as the input of a fifth ReLU activation function layer. The number of elements output by the fifth ReLU activation function layer is 64. The output of the fifth ReLU activation function layer is taken as the input of a fourth fully connected layer. The output of the fourth fully connected layer is taken as the input of a softmax function. The number of elements output by the softmax function is 7 (the probability p of the sample belonging to various signal types output by the softmax function c ). L2 regularization is applied in the convolutional layer and the fully connected layer to prevent overfitting. A regularization parameter (such as a weight penalty coefficient λ of 0.01) is set in each layer. When the loss on the test set no longer decreases for a plurality of consecutive rounds (for example, 10 epochs), the training is stopped. The batch size is set to 32. The Adam optimizer is used. The initial learning rate is set to 0.001. The defined training data and loss function are used to train each stage. After each round of training, the first test set and the second test set are used to evaluate the model. The detection accuracy (Accy d ) and the classification accuracy (Accyy c ) are recorded. During the training process, the weighted average value Score = 0.5Accy d + 0.5Accy c of the detection accuracy and the classification accuracy on the test set is calculated. When the third stage of training is completed, the weighted average value reaches the highest value on the test set. The model weight trained in the third stage is stored.

[0166] Specific implementation method nine: the difference between this implementation method and specific implementation method eight is that the specific process of step four two is as follows:

[0167] Step four two one, for any signal sample randomly extracted in step four one ( , indicating that the signal sample comes from the first data set), the signal sample is converted into a complex number form x(n):

[0168] x(n) = I(n) + jQ(n

[0169] Where I(n) represents the signal sample The nth bit in the I channel data, Q(n) represents the signal sample The nth bit in the Q channel data, n = 0, 1, ..., N1-1, j represents the imaginary unit;

[0170] Calculate the cyclic autocorrelation function value of x(n)

[0171]

[0172] Where τ represents the delay in milliseconds, τ = 0, 1, ..., D1, D1 represents the delay dimension, α represents the cycle frequency in Hz, α = 0, 1, ..., F1, x(n+τ) = I(n+τ) + jQ(n+τ), I(n+τ) represents the signal sample The n+τth bit in the I channel data, Q(n+τ) represents the signal sample The n+τth bit in the Q channel data, x * (n+τ) represents the conjugate of x(n+τ), e represents the base of natural logarithm, and j represents the imaginary unit;

[0173] Cyclic autocorrelation function value Modulo, the modulo obtained As a delay-cycle frequency matrix The elements in ; when τ = 0 and α = 0, the corresponding modulo results are used as the elements of the first row and first column of the delay-cyclic frequency matrix, and the corresponding modulo results when τ = 0 and α = 0, 1, ..., F1 are used as the elements of the first row of the delay-cyclic frequency matrix;

[0174] Step 422: For any signal sample randomly selected in step 41 ( Indicates that the signal sample comes from the second data set), the signal sample Convert to complex form x(n):

[0175] x(n)=I(n)+jQ(n)

[0176] Where I(n) represents the signal sample The nth bit in the I channel data, Q(n) represents the signal sample The nth bit in the Q channel data, n = 0, 1, ..., N2-1;

[0177] Calculate the cyclic autocorrelation function value of x(n)

[0178]

[0179] Where τ represents the delay in milliseconds, τ = 0, 1, ..., D2, D2 represents the delay dimension, α represents the cycle frequency in Hz, α = 0, 0.5, 1, ..., F2, x(n+τ) = I(n+τ) + jQ(n+τ), I(n+τ) represents the signal sample The n+τth bit in the I channel data, Q(n+τ) represents the signal sample The n+τth bit in the Q channel data, x * (n+τ) represents the conjugate of x(n+τ), e represents the base of natural logarithm, and j represents the imaginary unit;

[0180] Cyclic autocorrelation function value Modulo, the modulo obtained As a delay-cycle frequency matrix The elements in ; when τ = 0 and α = 0, the corresponding modulo results are used as the elements of the first row and first column of the delay-cyclic frequency matrix, and the corresponding modulo results when τ = 0 and α = 0, 1, ..., F2 are used as the elements of the first row of the delay-cyclic frequency matrix.

[0181] Other steps and parameters are the same as those in the eighth embodiment.

[0182] It should be noted that for the signal samples taken from the first data set D1=501, F1=201; for the signal samples taken from the second data set D2=1001, F2=401.

[0183] Specific embodiment 10: This embodiment differs from specific embodiment 9 in that the loss function used in the first stage training is for:

[0184]

[0185] The loss function used in the second stage of training for:

[0186]

[0187] The loss function used in the third stage of training for:

[0188]

[0189] in, represents the loss function of the detection task, Represents the classification task loss function;

[0190]

[0191] wherein y d represents a detection label, p d represents a probability of a model detecting a presence signal, y c,k represents a classification label, p c,k represents a probability of a model classifying a result as belonging to a kth signal. The base of the logarithm in the present application can be 10, but is not limited to 10.

[0192] The other steps and parameters are the same as those in Embodiment 9.

[0193] The above calculation examples of the present application are only to illustrate the calculation model and calculation process of the present application, and are not a limitation on the embodiments of the present application. For those skilled in the art, other different forms of changes or variations can be made on the basis of the above description, and it is impossible to enumerate all the embodiments here. Any obvious changes or variations derived from the technical solutions of the present application are still within the protection scope of the present application.

Claims

1. A low spectral density signal detection method based on time domain collaborative enhancement artificial intelligence, characterized in that: The method specifically comprises the following steps: Step 1: Set the sampling rate f s , signal type, signal time window length T1 and signal time window length T2; Step 2: According to the sampling rate f s , signal type, signal time window length T1 and T2 generate noisy signal samples; Step 3: Add detection labels and classification labels to the noisy signal samples generated in step 2 to obtain a first data set and a second data set; Divide the first data set into a training set and a test set to obtain a first training set and a first test set; divide the second data set into a training set and a test set to obtain a second training set and a second test set; Step 4: construct an enhanced network model, train the constructed network model using the first training set and the second training set, and test the constructed network model using the first test set and the second test set until the loss of the network model on the first test set and the second test set no longer decreases, thereby stopping the training and obtaining a trained network model; Step 5: Divide the data to be analyzed into segments with a signal time window length of Y1, and record the obtained signal segments as k'=1,2,…,K, K is the number of signal segments obtained; each signal segment is converted into a time delay-cycle frequency matrix, and each time delay-cycle frequency matrix is ​​used as the input of the trained network model. The detection head of the trained network model outputs the probability of the presence of a signal in each signal segment, and the probability of the presence of a signal in the k'th signal segment is recorded as in, like Then there is a signal in the k'th signal segment, otherwise there is no signal in the k'th signal segment; If there is a signal in the k'th signal segment, a signal consisting of the second half of the k'-1th signal segment, the k'th signal segment, and the first half of the k'+1th signal segment is extracted from the data to be analyzed. The intercepted signal Convert it into a time delay-cycle frequency matrix, use the time delay-cycle frequency matrix as the input of the trained network model, and output the signal fragment through the classification head of the trained network model The probability that the signal in belongs to each type Determine the signal segment based on the output of the classification head The signal type in .

2. The method for detecting low spectral density signals based on time-domain collaborative enhancement artificial intelligence according to claim 1, characterized in that: The sampling rate f s =1kHz, the signal types include BPSK signal, QPSK signal, QAM16 signal, AM-DSB signal, FSK signal, OFDM signal, LoRa signal and pure noise signal, and each signal includes two channels, I and Q.

3. The method for detecting low spectral density signals based on time-domain collaborative enhancement artificial intelligence according to claim 2, characterized in that: The specific process of step 2 is: Step 21: Generate noise-free signal samples with time window lengths of T1 and T2 for each signal type in the Python platform; Step 22: Each signal sample generated in step 21 is mixed with Gaussian white noise to generate noisy signal samples that meet the set signal-to-noise ratio; The signal-to-noise ratio range is from -20dB to 18dB, and the signal-to-noise ratio can be set in 2dB intervals.

4. The method for detecting low spectral density signals based on time-domain collaborative enhancement artificial intelligence according to claim 3, characterized in that: In step 21, noise-free signals with time window lengths of T1 and T2 are generated for each signal type, respectively. The specific process is as follows: Take the time window length T1 as an example: (1) BPSK signal generation: Randomly generate a binary sequence of length N1, map the 0 in the binary sequence to -1, map the 1 in the binary sequence to 1, use the mapping result of the binary sequence as the data of the I channel, and set the data of the Q channel to all 0; Where N1 = f s ×T1; (2) QPSK signal generation: Randomly generate a quaternary sequence of length N1, and map the 00 in the quaternary sequence to Map 01 in the quaternary sequence to Map 10 in the quaternary sequence to Mapping 11 in the quaternary sequence to Then each data in the I channel is the cosine value of each mapping result, and each data in the Q channel is the sine value of each mapping result; (3) QAM16 signal generation: Randomly generate a hexadecimal sequence of length N1, map 0000 in the hexadecimal sequence to -3-3i, map 0001 in the hexadecimal sequence to -3-i, map 0010 in the hexadecimal sequence to -3+3i, map 0011 in the hexadecimal sequence to -3+i, map 0100 in the hexadecimal sequence to -1-3i, map 0101 in the hexadecimal sequence to -1-i, map 0110 in the hexadecimal sequence to -1+3i, and map 1 0111 in the hexadecimal sequence is mapped to -1+i, 1000 in the hexadecimal sequence is mapped to 3-3i, 1001 in the hexadecimal sequence is mapped to 3-i, 1010 in the hexadecimal sequence is mapped to 3+3i, 1011 in the hexadecimal sequence is mapped to 3+i, 1100 in the hexadecimal sequence is mapped to 1-3i, 1101 in the hexadecimal sequence is mapped to 1-i, 1110 in the hexadecimal sequence is mapped to 1+3i, and 1111 in the hexadecimal sequence is mapped to 1+i; Then each data in the I channel is the real part of each mapping result, and each data in the Q channel is the imaginary part of each mapping result; (4) AM-DSB signal generation: The frequency f c The sine wave is used as the baseband signal, and the frequency is f m The carrier modulates the baseband signal, and the modulated signal s(t)=A(1+mcos(2πf m t))cos(2πf c t), the modulated signal is used as the I channel data, and the Q channel data are all 0; Where A is the amplitude and m is the modulation index; (5) FSK signal generation: Generate a random binary number, map the generated binary number to a cosine signal cos(2πft), where f is the frequency, and sample K samples from the mapped cosine signal cos(2πft), where K = f s / R, R is the bit rate; Similarly, each binary number is generated in turn and processed respectively, and the sampled samples are used to form the I channel data of the FSK signal, and the length of the I channel data is N1; the Q channel data of the FSK signal are all 0; When the generated binary number is 0, f = 50 Hz; when the generated binary number is 1, f = 100 Hz; (6) OFDM signal generation: Define the number of subcarriers as 116 and the number of FFT points as 128. Then, the data from the 59th to the 116th subcarriers are sequentially used as the 1st to the 58th data, and 12 zeros are added after the 58th data. Then, the data from the 1st to the 58th subcarriers are sequentially used as the 71st to the 128th data, resulting in data with a length of 128. Perform inverse fast Fourier transform on the 128-bit data to generate a time domain signal. Then, extract the last 32 bits of the generated time domain signal as a cyclic prefix. Copy the cyclic prefix to the beginning of the generated time domain signal. Then, concatenate the cyclic prefix and the generated time domain signal to form an OFDM symbol. The number of points in each OFDM symbol is 160. Take N1 / 160 OFDM symbols, then the I channel data of the OFDM signal is the real part of the OFDM symbol, and the Q channel data of the OFDM signal is the imaginary part of the OFDM symbol; (7) LoRa signal generation: Generate frequency from f0 to f ' 0 linear chirp signal, the I channel data is cos(2πf(t)t), and the Q channel data is sin(2πf(t)t), where f(t) represents the frequency that changes with time, f(t)∈[f0,f ' 0]; (8) Pure noise signal generation: The I channel is Gaussian white noise with a mean of 0 and a variance of 1, and the Q channel is Gaussian white noise with a mean of 0 and a variance of 1.

5. The method for detecting low spectral density signals based on time-domain collaborative enhancement artificial intelligence according to claim 4, characterized in that: In step 22, each generated signal sample is mixed with Gaussian white noise. The power of Gaussian white noise is calculated as follows: Among them, P signal is the signal power, P noise is the noise power, SNR represents the signal-to-noise ratio of the noisy signal sample; Then according to the calculated noise power P noise , generating a mean of 0 and a variance of P noise The generated Gaussian white noise is added to the I channel and Q channel of the noise-free signal respectively to obtain the noisy signal samples.

6. The method for detecting low spectral density signals based on time-domain collaborative enhancement artificial intelligence according to claim 5, characterized in that: The specific process of step three is: Step 3.

1. Add labels and obtain datasets Add detection labels to signal samples of length N1 and length N2: detection label y of pure noise signal d =0; detection tag y for BPSK signal, QPSK signal, QAM16 signal, AM-DSB signal, FSK signal, OFDM signal and LoRa signal d =1; Only add classification labels to signal samples of length N2: classification label y of BPSK signal c = 0, the classification label y of the QPSK signal c =1, the classification label y of QAM16 signal c =2, the classification label y of AM-DSB signal c =3, classification label y of FSK signal c =4, the classification label y of the OFDM signal c =5, the classification label y of LoRa signal c =6; The first data set dataset_1 is composed of signal samples with a length of N1 and a detection label, and the second data set dataset_2 is composed of signal samples with a length of N2 and a classification label. Step 3.2: Record the first data set as represents the i-th signal sample in the first data set, M1 represents the number of signal samples of length N1, represents the detection label of the i-th signal sample, and divides the first data set into a first training set and a first test set. The ratio of the number of signal samples in the first training set to that in the first test set is 7:3; The second data set is denoted as represents the jth signal sample in the second data set, M2 represents the number of signal samples of length N2, represents the detection label of the j-th signal sample, represents the classification label of the j-th signal sample, and divides the second data set into a second training set and a second test set. The ratio of the number of signal samples in the second training set to that in the second test set is 7:

3.

7. The method for detecting low spectral density signals based on time-domain collaborative enhancement artificial intelligence according to claim 6, characterized in that: The working process of the enhanced network model is as follows: The input of the enhanced network model is used as the input of the first convolutional layer, and the output of the first convolutional layer is used as the input of the first relu activation function layer; The output of the first relu activation function layer is used as the input of the first maximum pooling layer, the output of the first maximum pooling layer is used as the input of the second convolutional layer, and the output of the second convolutional layer is used as the input of the second relu activation function layer; The output of the second relu activation function layer is used as the input of the second maximum pooling layer, and the output of the second maximum pooling layer is used as the input of the third convolutional layer, and the output of the third convolutional layer is used as the input of the third relu activation function layer; The output of the third relu activation function layer is used as the input of the global average pooling layer, and the output of the global average pooling layer is used as the input of the detection head and the classification head respectively; where: The detection head includes the first fully connected layer, the fourth relu activation function layer, the second fully connected layer and the sigmoid activation function layer in sequence. The sigmoid activation function layer outputs the probability of the signal sample existing. The classification head includes the third fully connected layer, the fifth relu activation function layer, the fourth fully connected layer and the softmax activation function layer in sequence. The softmax activation function layer outputs the probability that the signal sample belongs to various signal types.

8. The method for detecting low spectral density signals based on time-domain collaborative enhancement artificial intelligence according to claim 7, characterized in that: The training process of the enhanced network model is: Step 4:

1. Randomly sample 80% of the signal samples from the first training set and randomly sample 20% of the signal samples from the second training set; Step 42: Perform cyclic autocorrelation preprocessing on each randomly selected signal sample, and convert each signal sample into a time delay-cyclic frequency matrix through the cyclic autocorrelation preprocessing; Step 43: Using the delay-cycle frequency matrix corresponding to each signal sample as the input of the enhanced network model; using the delay-cycle frequency matrix corresponding to each signal sample to perform the first phase training on the enhanced network model; Adjust the parameters of the boosted network model according to the loss function during training until the loss of the model on the first test set and the second test set stops decreasing for several consecutive rounds, and then stop training in the current training phase. Step 44: randomly extract 50% of the signal samples from the first training set and randomly extract 50% of the signal samples from the second training set, and convert each of the randomly extracted signal samples into a time delay-cyclic frequency matrix through cyclic autocorrelation preprocessing; Step 45: Use the delay-cycle frequency matrix obtained in step 44 to perform the second phase of training on the model. Adjust the parameters of the boosted network model according to the loss function during training until the loss of the model on the first test set and the second test set stops decreasing for several consecutive rounds. Stop the training of the current training phase. Step 46: Randomly extract 20% of the signal samples from the first training set and randomly extract 80% of the signal samples from the second training set, and convert each randomly extracted signal sample into a time delay-cyclic frequency matrix through cyclic autocorrelation preprocessing; Step 47: Use the delay-cycle frequency matrix obtained in step 46 to perform the third stage of model training. Adjust the parameters of the boosted network model according to the loss function during training until the loss of the model on the first test set and the second test set no longer decreases for several consecutive rounds. Stop the training of the current training stage, and obtain the final trained model.

9. The method for detecting low spectral density signals based on time-domain collaborative enhancement artificial intelligence according to claim 8, characterized in that: The specific process of step 42 is as follows: Step 421: For any signal sample randomly selected in step 41 The signal samples Convert to complex form x(n): x(n)=I(n)+jQ(n) Where I(n) represents the signal sample The nth bit in the I channel data, Q(n) represents the signal sample The nth bit in the Q channel data, n = 0, 1, ..., N1-1, j represents the imaginary unit; Calculate the cyclic autocorrelation function value of x(n) Where τ represents the delay in milliseconds, τ = 0, 1, ..., D1, D1 represents the delay dimension, α represents the cycle frequency in Hz, α = 0, 1, ..., F1, x(n+τ) = I(n+τ) + jQ(n+τ), I(n+τ) represents the signal sample The n+τth bit in the I channel data, Q(n+τ) represents the signal sample The n+τth bit in the Q channel data, x * (n+τ) represents the conjugate of x(n+τ), e represents the base of natural logarithm, and j represents the imaginary unit; Cyclic autocorrelation function value Modulo, the modulo obtained As a delay-cycle frequency matrix Elements in Step 422: For any signal sample randomly selected in step 41 The signal samples Convert to complex form x(n): x(n)=I(n)+jQ(n) Where I(n) represents the signal sample The nth bit in the I channel data, Q(n) represents the signal sample The nth bit in the Q channel data, n = 0, 1, ..., N2-1; Calculate the cyclic autocorrelation function value of x(n) Where τ represents the delay in milliseconds, τ = 0, 1, ..., D2, D2 represents the delay dimension, α represents the cycle frequency in Hz, α = 0, 0.5, 1, ..., F2, x(n+τ) = I(n+τ) + jQ(n+τ), I(n+τ) represents the signal sample The n+τth bit in the I channel data, Q(n+τ) represents the signal sample The n+τth bit in the Q channel data, x * (n+τ) represents the conjugate of x(n+τ), e represents the base of natural logarithm, and j represents the imaginary unit; Cyclic autocorrelation function value Modulo, the modulo obtained As a delay-cycle frequency matrix Elements in .

10. The method for detecting low spectral density signals based on time-domain collaborative enhancement artificial intelligence according to claim 9, characterized in that: The loss function used in the first stage of training is for: The loss function used in the second stage of training for: The loss function used in the third stage of training for: in, represents the loss function of the detection task, Represents the classification task loss function; Among them, y d represents the detection label, p d represents the probability of the presence of a signal detected by the model, y c,k represents the classification label, p c,k It represents the probability that the model classification result belongs to the kth category signal.

Images