A waveform recognition method based on deep neural network under small sample conditions

CN114036983BInactive Publication Date: 2024-11-05NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111310138.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-11-05
Publication Date
2024-11-05
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

同时,已知的基于生成对抗网络(Generative Adversarial Network,GAN)的数据扩增方法复杂性高,无法满足AMC场景的需求

Benefits of technology

[0062] The waveform identification method proposed by the present invention can achieve effective data amplification for signal waveforms with different modulation modes and different probability distributions. The simulation results show that this method can be used to determine the minimum number of sampling samples for BPSK, QPSK and 16QAM waveforms. At the same time, the waveform recognition performance after training the neural network using the data set expanded with the minimum number of sampling samples can reach the level of using sufficient samples to train the neural network. detection performance

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114036983B_ABST
    Figure CN114036983B_ABST
Patent Text Reader

Abstract

The present invention discloses a waveform recognition method based on a deep neural network under small sample conditions, including two parts: amplification of small samples and waveform recognition. Small sample amplification includes determining the minimum number of sampling samples required for the amplified signal and realizing small sample amplification using a probability density function estimation method. The present invention first determines the minimum number of sampling samples required for the signal to be amplified by the KLIEP algorithm; then estimates the probability density function of the signal to be amplified by the minimum sampling sample, and completes the amplification of the small sample according to the estimated probability density function; finally, the amplified signal and the deep neural network are used to realize the recognition of the waveform. This method can achieve the waveform recognition performance under a sufficient sample data set by using the detection performance of the small sample amplification data set.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of digital communication technology, and in particular to a waveform recognition method based on a deep neural network under small sample conditions. Background Art

[0002] Automatic Modulation Classification (AMC) of communication signals is a core technology in military electronic warfare and civilian spectrum monitoring and management. AMC refers to the classification of intercepted signals of unknown modulation using methods in pattern recognition or decision theory. Traditional modulation recognition algorithms rely on expert knowledge and known channel models and parameters. However, in actual communication scenarios, there is often very little prior knowledge about the intercepted signals. Therefore, the application of traditional modulation recognition methods is greatly limited.

[0003] Signal recognition methods based on deep neural networks (DNN) provide a new way to solve the shortcomings of traditional modulation recognition methods. In recent years, convolutional neural network (CNN) modulation recognition classifiers have been widely studied. The classifier can complete signal feature extraction and modulation mode discrimination based on the input data, thereby reducing the dependence on expert knowledge and avoiding complex artificial signal feature design. In addition, the structure based on the long short-term memory network (LSTM) is also used to design various modulation recognition classifiers. The classifier can use its memory function to extract signal features in the time dimension.

[0004] However, AMC based on deep neural network model is a "data-driven" model, and a large amount of training data is required to obtain good modulation recognition performance. However, in actual communication environments such as battlefields, prior data is extremely precious, and only a small number of valid samples can usually be obtained in actual applications. At this time, due to insufficient training data, the DNN model will have poor classification performance, or even fail to achieve modulation recognition classification.

[0005] Most of the existing data sample augmentation methods are aimed at the field of image processing, such as rotation, translation and cropping of pictures. At the same time, the known data augmentation methods based on Generative Adversarial Network (GAN) are highly complex and cannot meet the needs of AMC scenarios. In addition, how to determine the minimum number of samples required in the data set to be augmented is an unknown problem (the number of samples in the data set to be augmented is too small, resulting in inconsistent data distribution between the augmented data set and the sample data set, thereby affecting the detection performance of the augmented data set; too many samples will bring computational burden). Therefore, to achieve waveform recognition under small samples, it is first necessary to determine the minimum number of sampling samples required in the data set to be augmented; secondly, a method is needed to effectively amplify waveform samples with different modulation signals and different probability distributions. In addition, the complexity of the data augmentation method should be low and not rely on large-scale cluster servers in the cloud to meet the needs of battlefield environments. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention provides a waveform recognition method based on a deep neural network under small sample conditions, including determining the minimum number of sampling samples required for amplifying the signal and realizing the amplification of small samples using a probability density function estimation method. The method can achieve effective data amplification for signal waveforms with different modulation modes and different probability distributions. In addition, the method has low computational complexity, and the detection performance of the neural network trained using the amplified data set can reach the detection performance of the neural network trained using sufficient samples.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] A waveform recognition method based on a deep neural network under small sample conditions, the waveform recognition method comprising the following steps:

[0009] S1, use the detection receiver to sample the unknown signal waveforms of different modulation modes respectively to obtain small sample sampling signals Where i represents the type of modulation, and n represents the number of finite sampling samples;

[0010] S2, using the KLIEP algorithm, the minimum number of samples N required for effective amplification of the I-channel signal and the Q-channel signal is determined by the probability density function approximation method. I and N Q ;

[0011] S3, from the sampled signal s (i) Select N I I-channel samples are used as the signal set to be amplified right Estimating the probability density function, we get Estimated probability density function Using the estimated For small sample sets Perform data amplification to obtain the amplified signal waveform set in, represents the small sample data set before I-way amplification, N I represents the number of samples in the dataset before amplification, represents the dataset after I-way amplification, N p represents the number of samples in the amplified data set, and N p >>N I ;

[0012] S4, from the sampled signal s (i) Select N Q Q-path samples are used as the signal set to be amplified right Estimating the probability density function, we get Estimated probability density function Using the estimated For small sample sets Perform data amplification to obtain the amplified Q-channel signal waveform data set in, represents the data set after Q-path amplification, N q represents the number of samples in the amplified data set, and N q >>N Q ;

[0013] S5, respectively, the amplified data sets in step S3 and step S4 and Perform cumulative amount-based preprocessing to extract signal features and obtain a preprocessed amplification data set and in, represents the preprocessing function of the signal;

[0014] S6, repeat steps S1 to S5 to obtain the expanded deep neural network training data set And use Tra Aug Train the neural network to realize signal waveform recognition, where 1, 2, ...m represents the modulation mode sequence number.

[0015] To optimize the above technical solutions, the specific measures taken also include:

[0016] Furthermore, in step S1, the process of using a detection receiver to sample the unknown signal waveforms of different modulation modes includes the following steps:

[0017] S11, let the time domain received signal be expressed as:

[0018]

[0019] Where y(t) is the received signal in the time domain, represents the nth symbol of the i-th modulation mode sent, where represents the modulation type; g(t) = h(t)*p(t), where h(t) is the channel impulse response, p(t) is the pulse shaping function, T s is the symbol period; w(t) is independent and identically distributed Gaussian white noise;

[0020] S12, sampling the above time domain received signal, each sampling segment contains n symbols, and obtaining a sampled signal s (i) : It represents the nth waveform sample of the i-th modulation mode, and n represents the number of finite sampling samples.

[0021] Furthermore, in step S2, the KLIEP algorithm is used to determine the minimum number of samples N required for effective amplification of the I-channel signal and the Q-channel signal by using the probability density function approximation method. I and N Q The process includes the following steps:

[0022] S21, sampling the detection waveform of the unknown modulation mode at a certain signal-to-noise ratio to obtain a sampling data set in, and Respectively I and Q sampling signals;

[0023] S22, order:

[0024]

[0025] st N num1 ≤N1,N num2 ≤N2

[0026] Among them, ω(s) represents the sampled data s (i) The ratio of the probability density estimate to the true probability density value under a small sample data set, which is a function of the sampling data s; and Respectively and The true probability density function value of and By sampling the dataset and The estimation is obtained, N1 and N2 represent the number of I-channel and Q-channel sampling points corresponding to the true probability density function estimation; and It means that the and The probability density function of num1 ≤N1,N num2 ≤N2; by changing the size of the sampling sample of the small sample data set N num1 and N num2 , find the minimum number of sampling points N num1 and N num2 , so that

[0027] S23, using sampled data sets of length N1 and N2 and express and Use length N num1 and N num2 The sample data set and express and Where N num1 and N num2 according to and Increasing way, and is the set initial sample number, k is the set scaling factor;

[0028] S24, at a given N num1 and N num2 Next, we can find the corresponding small sample data set and

[0029]

[0030]

[0031] Find within the defined interval such that and The one with the smallest statistical distance to ω(s) = 1 and The number of samples N at this time num1 and N num2 They are defined as the minimum number of samples N required for effective amplification of small sample data sets. I and N Q ; Among them, α l is the parameter to be determined, which can be calculated using the iterative algorithm; φ l(·) is the selected kernel function; l is the independent variable related to the type of kernel function, b represents the number of kernel functions, j, and k are the independent variables related to the number of samples of I-way and Q-way small sample data sets, respectively.

[0032] Furthermore, the parameter α to be determined l The solution process includes the following steps:

[0033] S241, Order:

[0034]

[0035] in, and Small sample data sets The true probability density function and estimated density function of Is a sufficient sample data set The probability density function of , D represents the integration interval;

[0036] S242, since the first term of the formula in step S241 and the parameter α to be determined l After discarding this term, the new function J is defined as:

[0037]

[0038] S243, yes After normalization, we get Right now:

[0039]

[0040] Among them, p represents The independent variables related to the sample size N1;

[0041] S244, α l The problem of solving is simplified to:

[0042]

[0043] st

[0044] S245, solving the problem formula in step S244 by an iterative algorithm to obtain α l ;

[0045] S246, will α l Substitution The expression is The corresponding minimum number of samples N num1 Similarly, we can get The corresponding minimum number of samples N num2 .

[0046] Furthermore, the following formula is used to estimate the probability density function of the I-channel and Q-channel sample data:

[0047]

[0048] In the formula, represents the probability density estimate obtained at s, S i represents the sample of the i-th sampling, I(·) is the indicator function, n represents the number of sampling points, k(u) represents the kernel function used in density estimation, and the Gaussian kernel function is used in the calculation h represents the window width in density estimation;

[0049] According to the estimated probability density function Any number of data that obey this distribution can be generated, that is, the amplification of small sample data can be achieved.

[0050] Further, in step S5, the amplified data sets in step S3 and step S4 are respectively and The process of performing cumulant-based preprocessing includes the following steps:

[0051] Assume {x(t)} is a stationary random process, and its P-order k-th conjugate cumulant is defined as:

[0052] C pk (t1, t2, ..., t k )=Cum{x(t),x(t+t1),...,x * (t+t k-1 )}

[0053] Where t represents the initial sampling time, t1, t2, ...t k is the sampling delay variable, p represents the order of the accumulant, * represents the conjugate operation, and k represents the number of conjugate operations; the accumulant selected in the preprocessing is After preprocessing, the I- and Q-channel amplification data are expressed as:

[0054]

[0055] In the formula, C 2k , ...C 6k They represent the k-th conjugate cumulants of different orders, and represents the amplified I-path and Q-path data sets;

[0056] Perform data preprocessing on the amplified signals of different modulation modes i to obtain the amplified deep neural network training data set i=1, 2, ..., m, where m represents the number of modulation modes.

[0057] Further, in step S6, using Tra Aug The neural network used to realize signal waveform recognition is iteratively trained to obtain the optimal parameters of the neural network; the parameter update process during the neural network training process is as follows:

[0058]

[0059] In the formula, θ new is the updated parameter, θ old is the parameter before updating, γ>0 is the learning rate, is the gradient value in each update process.

[0060] Furthermore, the neural network is used to recognize BPSK, QPSK and 16QAM waveforms.

[0061] The beneficial effects of the present invention are:

[0062] The waveform recognition method proposed in the present invention can achieve effective data amplification for signal waveforms with different modulation modes and different probability distributions. The simulation results show that the method can be used to determine the minimum number of sampling samples for BPSK, QPSK and 16QAM waveforms. At the same time, the waveform recognition performance of the neural network trained with the data set amplified with the minimum number of sampling samples can reach the detection performance of the neural network trained with sufficient samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a flow chart of a waveform recognition method based on a deep neural network under small sample conditions in an embodiment of the present invention.

[0064] Figure 2 KLIEP performance curve of 16QAM I-channel signal under different sampling numbers.

[0065] Figure 3 KLIEP performance curve of 16QAM Q-channel signal under different sampling numbers.

[0066] Figure 4 It is a waveform recognition performance comparison curve between the amplified data set and the sufficient sample data set in the present invention. DETAILED DESCRIPTION

[0067] The present invention will now be described in further detail with reference to the accompanying drawings.

[0068] It should be noted that the terms such as "upper", "lower", "left", "right", "front", "back", etc. cited in the invention are only for the convenience of description and are not used to limit the scope of implementation of the present invention. Changes or adjustments in their relative relationships should be regarded as the scope of implementation of the present invention without substantially changing the technical content.

[0069] Figure 1 Flowchart of the waveform recognition method based on deep neural network under small sample conditions of an embodiment of the present invention. Figure 1 , the waveform recognition method comprises the following steps:

[0070] S1, use the detection receiver to sample the unknown signal waveforms of different modulation modes respectively to obtain small sample sampling signals Wherein, i represents the type of modulation, and n represents the number of finite sampling samples.

[0071] S2, using the KLIEP algorithm, the minimum number of samples N required for effective amplification of the I-channel signal and the Q-channel signal is determined by the probability density function approximation method. I and N Q .

[0072] S3, from the sampled signal s (i) Select N I I-channel samples are used as the signal set to be amplified right Estimating the probability density function, we get Estimated probability density function Using the estimated For small sample sets Perform data amplification to obtain the amplified signal waveform set in, represents the small sample data set before I-way amplification, N I represents the number of samples in the dataset before amplification, represents the dataset after I-way amplification, N p represents the number of samples in the amplified data set, and N p >>N I .

[0073] S4, from the sampled signal s (i) Select N Q Q-path samples are used as the signal set to be amplified right Estimating the probability density function, we get Estimated probability density function Using the estimated For small sample sets Perform data amplification to obtain the amplified Q-channel signal waveform data set in, represents the data set after Q-path amplification, N q represents the number of samples in the amplified data set, and N q >>N Q .

[0074] S5, respectively, the amplified data sets in step S3 and step S4 and Perform cumulative amount-based preprocessing to extract signal features and obtain a preprocessed amplification data set and in, Represents the preprocessing function of the signal.

[0075] S6, repeat steps S1 to S5 to obtain the expanded deep neural network training data set And use Tra Aug Train the neural network to realize signal waveform recognition, where 1, 2, ...m represents the modulation mode sequence number.

[0076] Step 1: Waveform sampling

[0077] The detection receiver samples the unknown signal waveform arriving at different times, and the time domain received signal can be expressed as:

[0078]

[0079] Where y(t) is the received signal in the time domain, represents the nth symbol of the i-th modulation mode sent, where represents the modulation type; g(t) = h(t)*p(t), where h(t) is the channel impulse response, p(t) is the pulse shaping function, T s is the symbol period; w(t) is independent and identically distributed Gaussian white noise. After sampling the above time domain received signal, it can be expressed as: Represents the nth waveform sample of the i-th modulation method.

[0080] A small sample data set can be obtained by sampling the received signal in the time domain.

[0081] Step 2: Determine the minimum number of samples in the data set to be amplified

[0082] In this embodiment, the minimum number of sampling samples in the data set to be amplified is first determined based on the KLIEP algorithm, and then the small sample data set is amplified by the minimum number of sampling samples. The method for determining the minimum number of sampling samples based on probability density function approximation is as follows:

[0083]

[0084] In the formula, and They represent the estimated and The probability density function of and Respectively and The true probability density function of num represents the number of small samples required to approximate the probability density function. However, there is a large calculation error in the above method of estimating the probability density function of the numerator and denominator, and it is difficult to calculate in high-dimensional processing. In this embodiment, the KLIEP algorithm is used to directly find the minimum number of samples N I and N Q , without having to estimate the probability density functions of the numerator and denominator separately.

[0085] Specifically, the method for determining the minimum sampling sample based on the KLIEP algorithm is as follows:

[0086] Sample the detection waveform of the unknown modulation mode at a certain signal-to-noise ratio to obtain a sampled data set in, and Respectively The I and Q sampling signals of . Order:

[0087]

[0088] st N num1 ≤N1,N num2 ≤N2

[0089] Among them, ω(s) represents the sampled data s (i) The ratio of the probability density estimate to the true probability density value under a small sample data set, which is a function of the sampling data s; and Respectively and The true probability density function of and By sampling the dataset and Therefore, the number of sampling points N1 and N2 is usually large; and It means that the and The probability density function of num1 ≤N1,Nnum2 ≤N2. By changing the size N of the sample of the small sample data set num1 and N num2 , so as to find the minimum number of sampling points num1 and num2, so that

[0090] In the formula, and Use sampled datasets of length N1 and N2 and express, and Use length N num1 and N num2 The sample data set and Indicates that N num1 and N num2 according to and The method increases in increments, among which, and is the initial number of samples set, and k is the set scaling factor.

[0091] At a given N num1 and N num2 Under this condition, we can obtain the corresponding small sample data set and and The expression is as follows:

[0092]

[0093]

[0094] Find within the defined interval such that and The one with the smallest statistical distance to ω(s) = 1 and At this time, the number of samples N num1 and N num2 This is the minimum number of samples required to achieve effective expansion of a small sample data set. l is the parameter to be determined, which can be calculated using the iterative algorithm; φ l (·) is the selected kernel function; l is the independent variable related to the type of kernel function, b represents the number of kernel functions, j, and k are the independent variables related to the number of samples of the I-way and Q-way small sample data sets, respectively. and With a similar process, in this embodiment, The analysis of α is used as an example to illustrate l The calculation process:

[0095]

[0096] in, and Small sample data sets The true probability density function and estimated density function of Is a sufficient sample data set The probability density function of is, and D represents the integral interval. Since the first term in the above formula is related to the parameter α to be determined l It is irrelevant and can be discarded. Therefore, the new function J is defined as follows:

[0097]

[0098] In addition, due to is a probability density function, which satisfies the sum of the integral interval to 1, so After normalization, we can get

[0099]

[0100] Among them, p represents The independent variable related to the sample size N1.

[0101] By J expression and α l The problem of solving is simplified to:

[0102]

[0103] st

[0104] The above formula can be solved by iterative algorithm to obtain α l . l Substitution The expression can be obtained The corresponding minimum number of samples N num1 , that is, N num1 =N I .

[0105] Step 3: Data augmentation

[0106] The minimum number of samples N for I and Q paths of the data set to be augmented is obtained by the KLIEP algorithm. I and N Q After that, the probability density function estimation method is used to estimate the probability density function and perform data amplification on the I-channel and Q-channel sample data. The method for estimating the probability density function and data amplification is as follows:

[0107]

[0108] In the formula, represents the probability density estimate obtained at s, S i represents the sample of the i-th sampling, I(·) is the indicator function, n represents the number of sampling points, k(u) represents the kernel function used in density estimation, and the Gaussian kernel function is used in the calculation h represents the window width in density estimation. At the same time, according to the estimated probability density function Any number of data that obey this distribution can be generated, that is, the amplification of small sample data can be achieved.

[0109] After the above steps, the expanded data sets can be expressed as:

[0110]

[0111]

[0112] Among them, N p >>N I And N q >>N Q At the same time, the number of samples in the expanded data set meets the data volume requirements of neural network training.

[0113] Step 4: Preprocess the amplification data

[0114] In this example, in order to improve the detection performance of DNN in signal waveform recognition, the amplified data is preprocessed based on cumulants. The cumulant preprocessing process is as follows. Let {x(t)} be a stationary random process, and its P-order k-th conjugate cumulant is defined as:

[0115] C pk (t1, t2, ..., t k )=Cum{x(t),x(t+t1),...,x * (t+t k-1 )}

[0116] Where t represents the initial sampling time, t1, t2, ...t k is the sampling delay variable, p represents the order of the accumulant, * represents the conjugate operation, and k represents the number of conjugate operations. The accumulant selected in the preprocessing is After preprocessing, the I-channel and Q-channel amplification data can be expressed as:

[0117]

[0118] Data preprocessing is performed on the amplified signals of different modulation modes to obtain the amplified deep neural network training data set Among them, 1, 2, ...m represent some unknown modulation method.

[0119] Step 5: Neural Network Training

[0120] The training data set Tra formed by the amplified samples Aug By training the neural network classifier, the optimal parameters of the neural network can be found through iterative training, thereby achieving classification and recognition of different signal waveforms. The parameter update process during neural network training is as follows:

[0121]

[0122] In the formula, θ new is the updated parameter, θ old is the parameter before updating, γ>0 is the learning rate, is the gradient value in each update process.

[0123] Finally, the method in this example is simulated and verified, that is, firstly, the small sample waveform data of BPSK, QPSK and 16QAM in AWGN channel is amplified, and then the amplified data set is used for waveform recognition and detection.

[0124] Figure 2 and Figure 3 The KLIEP performance curves of the I and Q waveform sampling signals of 16QAM under different sample numbers are shown in the figure. When the sample data is small, such as the number of samples is 1K and 5K, the KLIEP algorithm can obtain There is a large error between the expected value of ω(s) = 1, that is, there is a large error between the probability density function of the small sample data set and the probability density function of the sufficient sample data set. At this time, expanding the data set will cause the distribution of the expanded data set to be inconsistent with the distribution of the sufficient sample data set. When the number of samples is 16K, the KLIEP algorithm obtains The mean is 0.944, the variance is 0.0035, and if we increase the number of samples, The mean and variance of remain almost unchanged, so 16K is the minimum number of sampling samples for data amplification of 16QAM signal in AWGN channel determined by KLIEP algorithm.

[0125] Figure 4The waveform recognition performance of BPSK, QPSK and 16QAM amplified signals with different numbers of sampling samples under AWGN channel. As can be seen from the figure, when the number of sampling samples contained in the data set to be amplified is small, the distribution of the amplified data set is inconsistent with the original data set, resulting in poor waveform recognition performance. Specifically, when the number of amplified samples of BPSK, QPSK and 16QAM is selected as 2K or 5K, there is a large performance loss in the waveform recognition performance of the amplified data set compared with the sufficient sample data set. Based on the KLIEP algorithm in this example, the minimum number of waveform sampling samples of BPSK, QPSK and 16QAM is first determined to be 12K, 14K and 16K respectively, and then the data set is amplified using the minimum number of sampling samples. From the waveform recognition performance of the amplified data set, it can be seen that the maximum performance loss of the recognition performance of the amplified data set and the sufficient sample sampling data set in the entire SNR range is only about 0.3dB, which illustrates the effectiveness of the waveform recognition method based on deep neural network under small sample conditions in this example.

[0126] By verifying a waveform recognition method based on a deep neural network under a small sample condition in this embodiment on the MATLAB platform, it can be seen from the simulation results that the method can be used to determine the minimum number of sampling samples of unknown signal waveforms with different modulation modes and different probability distributions. At the same time, the minimum number of sampling samples can be used to effectively amplify the sampling signal, and the waveform recognition performance of the amplified data set can reach the performance of training the neural network using a sufficient sample data set.

[0127] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A waveform recognition method based on deep neural network under small sample conditions, characterized in that: The waveform recognition method comprises the following steps: S1, use the detection receiver to sample the unknown signal waveforms of different modulation modes respectively to obtain small sample sampling signals Where i represents the type of modulation, and n represents the number of finite sampling samples; S2, using the KLIEP algorithm, the minimum number of samples N required for effective amplification of the I-channel signal and the Q-channel signal is determined by the probability density function approximation method. I and N Q ; S3, from the sampled signal s (i) Select N I I-channel samples are used as the signal set to be amplified right Estimating the probability density function, we get Estimated probability density function Using the estimated For small sample sets Perform data amplification to obtain the amplified signal waveform set in, represents the small sample data set before I-way amplification, N I represents the number of samples in the dataset before amplification, represents the dataset after I-way amplification, N p represents the number of samples in the amplified data set, and N p >>N I ; S4, from the sampled signal s (i) Select N Q Q-path samples are used as the signal set to be amplified right Estimating the probability density function, we get Estimated probability density function Using the estimated For small sample sets Perform data amplification to obtain the amplified Q-channel signal waveform data set in, represents the data set after Q-path amplification, N q represents the number of samples in the amplified data set, and N q >>N Q ; S5, respectively, the amplified data sets in step S3 and step S4 and Perform cumulative amount-based preprocessing to extract signal features and obtain a preprocessed amplification data set and in, represents the preprocessing function of the signal; S6, repeat steps S1 to S5 to obtain the expanded deep neural network training data set And use Tra Aug Train the neural network to realize signal waveform recognition, where 1, 2, …m represents the modulation mode sequence number.

2. The waveform recognition method based on deep neural network under small sample conditions according to claim 1 is characterized in that: In step S1, the process of using a detection receiver to sample the unknown signal waveforms of different modulation modes includes the following steps: S11, let the time domain received signal be expressed as: Where y(t) is the received signal in the time domain, represents the nth symbol of the i-th modulation mode sent, where represents the modulation type; g(t) = h(t)*p(t), where h(t) is the channel impulse response, p(t) is the pulse shaping function, T s is the symbol period; w(t) is independent and identically distributed Gaussian white noise; S12, sampling the above time domain received signal, each sampling segment contains n symbols, and obtaining a sampled signal s( i ): It represents the nth waveform sample of the i-th modulation mode, and n represents the number of finite sampling samples.

3. The waveform recognition method based on deep neural network under small sample conditions according to claim 1 is characterized in that: In step S2, the KLIEP algorithm is used to determine the minimum number of samples N required for effective amplification of the I-channel signal and the Q-channel signal by using the probability density function approximation method. I and N Q The process includes the following steps: S21, sampling the detection waveform of the unknown modulation mode at a certain signal-to-noise ratio to obtain a sampling data set in, and Respectively I and Q sampling signals; S22, order: s.t.N num1 ≤N1,N num2 ≤N2 Among them, ω(s) represents the sampled data s (i) The ratio of the probability density estimate to the true probability density value under a small sample data set, which is a function of the sampling data s; and Respectively and The true probability density function value of and By sampling the dataset and The estimation is obtained, N1 and N2 represent the number of I-channel and Q-channel sampling points corresponding to the true probability density function estimation; and It means that the and The probability density function of num1 ≤N1,N num2 ≤N2; by changing the size of the sampling sample of the small sample data set N num1 and N num2 , find the minimum number of sampling points N num1 and N num2 , so that S23, using sampled data sets of length N1 and N2 and express and Use length N num1 and N num2 The sample data set and express and Where N num1 and N num2 according to and Increasing way, and is the set initial sample number, k is the set scaling factor; S24, at a given N num1 and N num2 Next, we can find the corresponding small sample data set and Find within the defined interval such that and The one with the smallest statistical distance to ω(s) = 1 and The number of samples N at this time num1 and N num2 They are defined as the minimum number of samples N required for effective amplification of small sample data sets. I and N Q ; Among them, α l is the parameter to be determined, which is calculated according to the iterative algorithm; φ l (·) is the selected kernel function; l is the independent variable related to the type of kernel function, b represents the number of kernel functions, and j and k are the independent variables related to the number of samples of I-way and Q-way small sample data sets, respectively.

4. The waveform recognition method based on deep neural network under small sample conditions according to claim 3 is characterized in that: Parameter α to be determined l The solution process includes the following steps: S241, Order: in, and Small sample data sets The true probability density function and estimated density function of Is a sufficient sample data set The probability density function of , D represents the integration interval; S242, since the first term of the formula in step S241 and the parameter α to be determined l It is irrelevant. After discarding this term, the new function J is defined as: S243, yes After normalization, we get Right now: Among them, p represents The independent variables related to the sample size N1; S244, α l The problem of solving is simplified to: S245, solving the problem formula in step S244 by an iterative algorithm to obtain α l ; S246, α l Substitution The expression is The corresponding minimum number of samples N num1 .

5. The waveform recognition method based on deep neural network under small sample conditions according to claim 1 is characterized in that: The following formula is used to estimate the probability density function of the I-channel and Q-channel sample data: In the formula, represents the probability density estimate obtained at s, S i represents the sample of the i-th sampling, I(·) is the indicator function, n represents the number of sampling points, k(u) represents the kernel function used in density estimation, and the Gaussian kernel function is used in the calculation h represents the window width in density estimation; According to the estimated probability density function Generate any number of data that obey the distribution to achieve the amplification of small sample data.

6. The waveform recognition method based on deep neural network under small sample conditions according to claim 1 is characterized in that: In step S5, the amplified data sets in steps S3 and S4 are respectively and The process of performing cumulant-based preprocessing includes the following steps: Assume {x(t)} is a stationary random process, and its P-order k-th conjugate cumulant is defined as: C pk (t1,t2,...,t k )=Cum{x(t),x(t+t1),...,x * (t+t k-1 )} Where t represents the initial sampling time, t1, t2, ...t k is the sampling delay variable, p represents the order of the accumulant, * represents the conjugate operation, and k represents the number of conjugate operations; the accumulant selected in the preprocessing is After preprocessing, the I- and Q-channel amplification data are expressed as: In the formula, C 2k ,...C 6k They represent the k-th conjugate cumulants of different orders, and represents the amplified I-path and Q-path data sets; Perform data preprocessing on the amplified signals of different modulation modes i to obtain the amplified deep neural network training data set m represents the number of modulation modes.

7. The waveform recognition method based on deep neural network under small sample conditions according to claim 1 is characterized in that: In step S6, using Tra Aug The neural network used to realize signal waveform recognition is iteratively trained to obtain the optimal parameters of the neural network; the parameter update process during the neural network training process is as follows: In the formula, θ new is the updated parameter, θ old is the parameter before updating, γ>0 is the learning rate, is the gradient value in each update process.

8. The waveform recognition method based on deep neural network under small sample conditions according to claim 1 is characterized in that: The neural network is used to recognize BPSK, QPSK and 16QAM waveforms.