Multi-system biorthogonal keying signal demodulation method based on STFT-CNN

Abstract

The invention discloses a multi-system biorthogonal keying signal demodulation method based on STFT-CNN, and the method comprises the steps: converting an information source into I-path information bits and Q-path information bits, and respectively mapping the information bits in the I-path information bits and the Q-path information bits to an I-path spreading code set and a Q-path spreading code set, repeated coding is added behind each spreading code in the I-path spreading code set and the Q-path spreading code set, that is, after each spreading code is repeated for multiple times, QPSK modulation is carried out, and an MBOK modulation signal is generated; and extracting time-frequency characteristics in the MBOK modulation signal, inputting the time-frequency characteristics into a machine learning model, outputting a corresponding likelihood probability, the likelihood probability being used for indicating a category to which a current time-frequency characteristic sample belongs or a state of a spreading code to which the current time-frequency characteristic sample belongs, and recovering information source information according to a one-to-one mapping relationship between the category or the state of the spreading code to which the current time-frequency characteristic sample belongs and information bits of an information source. According to the invention, the demodulation performance is improved.

Description

Technical Field

[0001] The invention relates to the technical field of communication, and in particular to a multi-level bi-orthogonal keying signal demodulation method based on STFT-CNN. Background Art

[0002] M-ary Bi-Orthogonal Keying (MBOK) is a soft spread spectrum technology for multi-level spread spectrum. Compared with direct sequence spread spectrum, MBOK increases the spread spectrum gain but does not increase the system bandwidth. Therefore, this spread spectrum method reduces the transmission bandwidth while maintaining the excellent anti-interference performance of traditional direct sequence spread spectrum, and can be used in some occasions with limited spectrum or variable transmission rate.

[0003] In the MBOK spread spectrum system model, the I / Q paths of the M-ary dual orthogonal keying signal each require M channels of length N. c (Spreading code length N c ) to represent the M states of log2M-bit source information, and map the I / Q two-way source information log2M-bit to two orthogonal spreading code sets C respectively. I ={C I0 ,C I1 …,C I(M-1)} and C Q ={C Q0 ,C Q1 …,C Q(M-1)},in is the spreading code of channel I, where is the I-channel spread spectrum code chip; is the spreading code of Q channel, where is the Q-channel spread spectrum code; after mapping the log2M-bit source information to the spread spectrum code, the baseband signal obtained after interpolation and root raised cosine filtering can be expressed as where s I [n],s Q [n] represents the code element state of the source information, c I [i] = {±1}, c Q [i] = {±1} represents the chips of the I / Q two-way spreading code respectively; g(t-iT c -nT s ) represents a root raised cosine pulse shaped pulse. N c is the spreading code length. c is the duration of the spreading code chip, T s =(N c ·T c) is the symbol duration of the source information mapping.

[0004] The demodulation method commonly used in the MBOK spread spectrum system is to use a matched filter to achieve demodulation based on the principle of maximizing the signal-to-noise ratio. MBOK spread spectrum signals generally use non-coherent demodulation, and the received signal is correlated with the M spread spectrum codes of the I and Q channels respectively. Since there are M spread spectrum codes in the I / Q spread spectrum code set, there are M·M states when the two spread spectrum codes are combined, and each state corresponds to two spread spectrum codes in the I / Q spread spectrum code set. The received signal is correlated with the M·M spread spectrum code combination, and the maximum value among the M·M correlation values ​​is searched. Then, according to the one-to-one mapping relationship between the spread spectrum code and the information bit sent by the transmitter, the 2m-bit information sent by the transmitter is determined (where m=log2M). This method requires the received signal to be correlated with the M·M combination states of the local spread spectrum code, so the amount of calculation is large. Summary of the invention

[0005] In view of this, the present invention provides a multi-level bi-orthogonal keying signal demodulation method based on STFT-CNN.

[0006] The invention discloses a multi-level bi-orthogonal keying signal demodulation method based on STFT-CNN, which comprises:

[0007] Convert the information source into I-channel information bits and Q-channel information bits, map the information bits in the I-channel information bits and Q-channel information bits to the I-channel spreading code set and the Q-channel spreading code set, respectively, add repetition coding after each spreading code in the I-channel spreading code set and the Q-channel spreading code set, that is, after each spreading code is repeated several times, QPSK modulation is performed to generate an MBOK modulated signal;

[0008] The time-frequency features in the MBOK modulated signal are extracted and input into the machine learning model, and the corresponding likelihood probability is output. The likelihood probability is used to indicate the category to which the current time-frequency feature sample belongs or the state of the spread spectrum code to which it belongs. According to the one-to-one mapping relationship between the category or the state of the spread spectrum code and the information bits of the source, the source information is restored.

[0009] Further, the converting of the information source into I-channel information bits and Q-channel information bits, and mapping the information bits in the I-channel information bits and the Q-channel information bits to the I-channel spreading code set and the Q-channel spreading code set, respectively, includes:

[0010] The source is converted into I-channel information bits and Q-channel information bits through serial-to-parallel conversion. Each m=log2M information bits in the I-channel information bits are mapped to the I-channel spreading code set C. I ={C I0 ,C I1 …,C I(M-1)}, each m = log2M information bits in the Q-path information bits are mapped to the Q-path spreading code set C Q ={C Q0 ,C Q1 …,C Q(M-1)}A spreading code in .

[0011] Furthermore, the spread spectrum code adopts Gold code or Walsh code as orthogonal spread spectrum code; the Gold code is composed of two m-sequence preferred pairs, the two m-sequence preferred pairs generate multiple orthogonal Gold codes, M code sequences are selected from the multiple orthogonal Gold codes to form an orthogonal spread spectrum code set for the M-ary dual orthogonal keying signal I path, and then M code sequences are selected from the remaining code sequences in the multiple orthogonal Gold codes to form an orthogonal spread spectrum code set for the M-ary dual orthogonal keying signal Q path.

[0012] Furthermore, the training sample acquisition process of the machine learning model includes:

[0013] After obtaining the MBOK modulated signal, N sampling signal samples are intercepted according to the synchronization position of the spread spectrum code, and the sampling signal samples are short-time Fourier transformed to obtain the complex matrix R STFT , for the complex matrix R STFT Take the modulus of each element of and get R STFT The module value matrix R modulus ; The matrix R modulus Normalize to get the input sample matrix R′ of the convolutional neural network modulus , the sample matrix R′ modulus As training samples for convolutional neural networks.

[0014] Furthermore, a convolutional neural network is adopted as a machine learning model, and the constructed convolutional neural network includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a first fully connected layer, a second fully connected layer and an output layer which are connected in sequence.

[0015] Furthermore, the cost function of the machine learning model is:

[0016] Suppose there is a training data set X containing m samples:

[0017] X={(x0,y0),…,(x i ,y i ),…(x m ,y m )}

[0018] Where X∈R n×n ,y i∈{0,1,…,M·M}, M·M is the total classification category, M is the number of I / Q two-way spreading codes, and M·M is the number of combined states of I / Q two-way spreading codes; x i is the i-th sample in the sample set; y i For sample x i The corresponding sample label;

[0019] Create a parameter model of the convolutional neural network to estimate the distribution probability of the training data set X. According to the principle of maximum likelihood estimation, maximizing the likelihood probability of the parameter model is equivalent to minimizing the cross entropy between the empirical distribution on the training data set and the probability distribution on the parameter model.

[0020] Furthermore, maximizing the following formula is equivalent to minimizing the cost function θ Loss :

[0021]

[0022] Among them, logp(y i |x i ; θ) is the likelihood probability, θ is the network parameter of the convolutional neural network;

[0023]

[0024] Wherein, j = {0, 1, ..., k-1}, j represents the jth category, and k represents the total number of categories.

[0025] The second fully connected layer of the convolutional neural network outputs the unnormalized log probability z j ,

[0026] z j =logp(y i =j|x i ;θ)

[0027] Use the Softmax function in the output layer to j Exponentiation and normalization yields:

[0028]

[0029] Among them, 1{y i =j} is an indicative function, that is, when the value in the brackets is true, the function result is 1, otherwise the result is 0; It represents the parameter penalty regularization term added to prevent overfitting, and λ represents the penalty coefficient.

[0030] Furthermore, the use of a stochastic gradient descent algorithm to train a convolutional neural network includes:

[0031] Initialize the convolutional neural network parameters and select the sample matrix R′ from the training set modulus It forms the input layer, passes through the first pooling layer, the second convolution layer, the second pooling layer, and then passes through the first fully connected layer and the second fully connected layer to enter the output layer. The output layer calculates the error between the prediction result of the current convolutional neural network model and the actual result based on the annotation information of the training sample matrix;

[0032] Based on the error between the predicted value and the true value obtained by forward propagation, back propagation is performed in the order from the output layer to the input layer, using the stochastic gradient descent algorithm. Each iteration traverses the entire training set. When the cost function decreases with the increase in the number of iterations until it finally converges to the specified accuracy, the training process of the convolutional neural network is terminated and the parameter model of the convolutional neural network is saved.

[0033] Furthermore, each iteration traverses the entire training set, including:

[0034] The stochastic gradient descent algorithm is used to calculate the gradient of each layer of network parameters in turn, and the gradient is used to update the network parameters of each layer. MBOK signal samples are continuously selected from the training set, and the above operation is repeated until the entire training set is traversed, that is, one iteration is completed.

[0035] Furthermore, when training the machine learning model, training samples with different signal-to-noise ratios are generated as input for training the convolutional neural network; when one iteration is completed, the test error of the validation set samples is used to estimate the generalization error of the convolutional neural network obtained in this iteration.

[0036] Due to the adoption of the above technical solution, the present invention has the following advantages: the present invention has a small amount of calculation and improves the demodulation performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0038] Figure 1 It is a block diagram of the MBOK modulation system model according to an embodiment of the present invention;

[0039] Figure 2 It is a block diagram of the MBOK demodulation system model according to an embodiment of the present invention;

[0040] FIG3(a), FIG3(b), FIG3(c) and FIG3(d) are schematic diagrams of STFT samples of MBOK signals generated by different spreading codes according to an embodiment of the present invention;

[0041] Figure 4 This is a structural diagram of a convolutional neural network model according to an embodiment of the present invention;

[0042] Figure 5 It is a learning model training error curve diagram of an embodiment of the present invention;

[0043] Figure 6 Schematic diagram of demodulation performance of Gold codes with two different point lengths according to an embodiment of the present invention;

[0044] Figure 7 It is a schematic diagram of the demodulation performance of Walsh codes with two different point lengths according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The present invention is further described in conjunction with the accompanying drawings and embodiments, and the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field should fall within the scope of protection of the embodiments of the present invention.

[0046] See also Figure 1 The present invention provides an embodiment of a multi-level bi-orthogonal keying signal demodulation method based on STFT-CNN, which comprises:

[0047] Convert the information source into I-channel information bits and Q-channel information bits, map the information bits in the I-channel information bits and Q-channel information bits to the I-channel spreading code set and the Q-channel spreading code set, respectively, add repetition coding after each spreading code in the I-channel spreading code set and the Q-channel spreading code set, that is, after each spreading code is repeated several times, QPSK modulation is performed to generate an MBOK modulated signal;

[0048] The time-frequency features in the MBOK modulated signal are extracted and input into the machine learning model, and the corresponding likelihood probability is output. The likelihood probability is used to indicate the category to which the current time-frequency feature sample belongs or the state of the spread spectrum code to which it belongs. According to the one-to-one mapping relationship between the category or the state of the spread spectrum code and the information bits of the source, the source information is restored.

[0049] Regarding the spreading code status, for example:

[0050] Spreading code set C I ={C I0 ,C I1 …,C I(M-1)} and C Q ={C Q0 ,C Q1 …,C Q(M-1) ] are 2*M spreading codes for I / Q channels respectively; assuming M=32, the spreading code for I channel is C I =[C I0 ,C I1 …,CI(M-1)} respectively correspond to 32 binary codes (source information bits) of “00000, 00001, 00010, 00011, 00100, …, 11110, 11111”; similarly, the Q-channel spread spectrum code C Q ={C Q0 ,C Q1 …,C Q(M-1)} respectively correspond to 32 binary codes (source information bits) of "00000, 00001, 00010, 00011, 00100, ..., 11110, 11111", and the I / Q two-way spread spectrum code is combined together (corresponding to 10-bit binary code), for a total of 2 10 =1024 combinations, one of which is called a spreading code state; the spreading code state corresponds one to one with the binary code (source information bit).

[0051] In one embodiment, generating the MBOK modulation signal includes:

[0052] The source is converted into I / Q information bits through serial-to-parallel conversion. Each m=log2M information bits in the I path are mapped to the spread spectrum code set C. I ={C I0 ,C I1 …,C I(M-1)}, each m = log2M information bits in the Q path are mapped to a spreading code set C Q ={C Q0 ,C Q1 …,C Q(M-1)}A spreading code in .

[0053] This embodiment will use Gold code or Walsh code as orthogonal spread spectrum coding. Gold code is composed of two m-sequence preferred pairs through modulo 2 addition, and each group of m-sequence preferred pairs can generate 2n+1 Gold sequences (where n is the highest order of the m-sequence primitive polynomial). Adding a 0 at the end of all Gold codes generated by the same preferred pair will form an even-position Gold code, also known as an orthogonal (or quasi-orthogonal) Gold code, whose code sequences are mutually orthogonal (or quasi-orthogonal).

[0054] Assume that the selected two m-sequences are optimally paired as follows:

[0055]

[0056] The above m-sequence preferably generates 65 orthogonal Gold codes (code sequence length is 64), from which M (M=32) code sequences are selected to form an orthogonal spread spectrum code set for the M-ary dual orthogonal keying signal I path, and then M code sequences are selected from the remaining code sequences to form an orthogonal spread spectrum code set for the M-ary dual orthogonal keying signal Q path. The spread spectrum code chip length is 64.

[0057] Walsh codes are constructed by the rows or columns of the Hadamard matrix, and a new Walsh orthogonal spread spectrum code is generated by combining the M-sequence and the Walsh sequence. The method of constructing the orthogonal spread spectrum code set for the I / Q two-way Walsh dual orthogonal keying signal is the same as the Gold code, and the spread spectrum code chip length is 64.

[0058] Because short-time Fourier transform (STFT) calculation is required in demodulation, repetition coding is added after the spread spectrum code. Repetition coding is to repeat the current spread spectrum code k times (the number of repetitions is determined according to the channel conditions. In order to obtain better demodulation performance, the more repetitions, the better the demodulation performance, but the spectrum efficiency is sacrificed), and then QPSK modulation is performed. MBOK modulation system model is as follows Figure 1 shown.

[0059] In one embodiment, a machine learning model is built:

[0060] This embodiment uses the time-frequency features of the extracted MBOK signal to be input into a convolutional neural network (CNN) for demodulation. Convolutional neural network is a commonly used deep learning model that has good performance in processing abstract high-dimensional features, has strong generalization ability and is good at processing complex data. It can automatically extract signal features without preprocessing and perform classification demodulation. The MBOK demodulation system model is as follows: Figure 2 shown.

[0061] According to the one-to-one mapping relationship between the spread spectrum code and the information bits sent by the transmitter, a machine learning model can be constructed. The input of the model is the modulated signal generated by the spread spectrum code through the transmission channel and the signal received at the receiving end. The signal contains the time-frequency characteristics of the corresponding spread spectrum code. The signal features are passed through the learning model and the corresponding likelihood probability is output. The meaning of this probability indicates which category the current feature sample belongs to, or which spread spectrum code state it belongs to.

[0062] Suppose there is a data set containing m feature samples:

[0063] X={(x0,y0),…,(x i ,y i ),…(x m ,y m )}

[0064] By learning the training set, a mapping from the input space to the output space is established:

[0065] f:X→Y

[0066] The input space is the feature sample set space, and the output space is the likelihood probability. When a new sample is input, the model can give the likelihood probability of which category the input sample belongs to.

[0067] In this way, the demodulation problem of MBOK spread spectrum signal is transformed into a multi-classification problem in machine learning. The following will build a machine learning model for MBOK signal demodulation from four aspects. These four aspects are generating a specific data set, machine learning model, cost function, and optimization algorithm.

[0068] Optionally, generate specific datasets, including:

[0069] After acquiring the MBOK wireless transmission signal, N (N = 2N) c 、4N c , the interception length is consistent with the repetition encoding length of the sending end, N c The signal samples are sampled (the signal sampling rate is consistent with the chip rate) and then the sampled signal samples are subjected to short-time Fourier transform (STFT) calculation to obtain input feature samples for training and inference of the convolutional neural network.

[0070] Short-time Fourier transform (STFT) is to use a window function w(k) to divide a time domain signal x(n) of length N into many segments for Fourier transform processing. The window function w(k) slides on the time axis, intercepts the signal x(n) in segments (the length of the window function is L), and then multiplies the window function w(k) with the signal x(n). The result of the multiplication is then calculated by DFT to obtain the time-frequency matrix (here the length of DFT is selected to be consistent with the length of the window function L).

[0071] The discrete short-time Fourier transform is given by:

[0072]

[0073] The sample lengths of the MBOK modulation signal obtained in this embodiment are N=128 and 256, respectively, which are consistent with the repetitive coding length of the modulation end. When performing STFT calculation, the window function w(k) slides on the time axis in an overlapping mode, and its overlapping lengths are Overlap_Len=63 and 61, respectively. The length of DFT, the length of the window function and the length of the spread spectrum code N c =64 are the same.

[0074] After STFT calculation, the complex matrix R is obtained STFT(the matrix dimension is 64x64), and then the complex matrix R STFT Take the modulus of each element of and get R STFT The module value matrix R modulus . The matrix R modulus Normalize to get the input sample matrix R′ of the convolutional neural network modulus .

[0075] The normalization of the input sample is calculated as follows:

[0076]

[0077] Where E(·) represents the mean and Var(·) represents the variance.

[0078] The I / Q channels of the M(32)-ary dual orthogonal keying signal each contain 5-bit information, totaling 32 states. The I / Q channels together contain a total of 1024 states. The sample matrix R′ under each of the 1024 states can be obtained separately. modulus , as the training sample of the convolutional neural network. The inference sample of the convolutional neural network can be obtained in the same way.

[0079] The time-frequency samples generated by the modulation signals generated by different spread spectrum code combinations are shown in Figure 3(a), Figure 3(b), Figure 3(c) and Figure 3(d) respectively. It can be seen from the figure that different sequences have different center frequencies and bandwidths, and the main lobe width is concentrated in a certain part of the spectrum of the entire frequency band. Such time-frequency characteristics are very suitable for convolutional neural networks to extract features.

[0080] In the data simulation, the MBOK signal with M=32 is used to generate the modulation signal. The spreading code adopts the above-mentioned Gold code or Walsh code, and the spreading code chip length is 64.

[0081] Each m = log2M = 5 information bits in path I are mapped to the spreading code set C I ={C I0 ,C I1 …,C I31}, each m = log2M = 5 information bits in the Q path are mapped to the spreading code set C Q ={C Q0 ,C Q1 …,C Q31 Therefore, the I / Q two-way spreading code is combined to generate a total of 1024 combination states. When training the neural network, 1024 training samples and sample labels need to be generated to form a training sample set.

[0082] This embodiment is limited to discussing the demodulation performance of MBOK signals under Gaussian white noise channel conditions, and assumes that the symbol synchronization of the MBOK spread spectrum signal has been completed at the receiving end. At the receiving end, through down-conversion, low-pass filtering and extraction, a sample signal synchronized with the MBOK symbol signal and with the same symbol rate is obtained. At the receiving end, samples of 2 times the symbol length and 4 times the symbol length are obtained respectively, and short-time Fourier transform is performed to obtain a time-frequency matrix. The short-time Fourier transform is calculated for the two sample lengths (128, 256) according to the parameters in Table 1 to obtain training samples, as shown in Table 1.

[0083] Table 1 Sampling point STFT transformation parameters

[0084]

[0085]

[0086] Optional, regarding machine learning models:

[0087] The convolutional neural network is used as the machine learning model. The layers of the constructed convolutional neural network are set as follows. The convolutional neural network model structure is as follows: Figure 4 shown.

[0088] Input layer: dimension is 64×64;

[0089] Convolutional layer C1: contains 6 feature layers, each with a dimension of 60×60 and a convolution kernel size of 5×5;

[0090] Pooling layer S2: contains 6 feature layers, each with a dimension of 30×30, using maximum pooling for extraction with an extraction ratio of 2;

[0091] Convolutional layer C3: contains 16 feature layers, each with a dimension of 26×26 and a convolution kernel size of 5×5;

[0092] Pooling layer S4: contains 16 feature layers, each with a dimension of 13×13, using maximum pooling for extraction with an extraction ratio of 2;

[0093] Fully connected layer F5: contains 120 neuron nodes;

[0094] Fully connected layer F6: contains 84 neuron nodes;

[0095] Output layer: Softmax fully connected.

[0096] Optional, optional, about the cost function:

[0097] Suppose there is a training data set containing m samples.

[0098] X={(x0,y0),…,(xi ,y i ),…(x m ,y m )}

[0099] where X∈R n×n ,y i ∈{0,1,…,M·M}, M·M is the total classification category, M is the number of I / Q two-way spreading codes, and M·M is the number of combined states of I / Q two-way spreading codes. i is the i-th sample in the sample set; y i For sample x i The corresponding sample label.

[0100] A parameter model needs to be created to estimate the distribution probability of the training data set X. According to the principle of maximum likelihood estimation, maximizing the likelihood probability of the parameter model is equivalent to minimizing the cross entropy between the empirical distribution on the training data set and the probability distribution on the parameter model.

[0101]

[0102] Among them, θ is the network parameter of the convolutional neural network;

[0103] Maximizing the above formula is equivalent to minimizing the cost function

[0104]

[0105] j={0,1,…,k-1}, j represents the jth category, and k represents the total number of categories.

[0106] The output of the last fully connected layer (F6) of the convolutional neural network predicts the unnormalized log probability z j ,

[0107] z j =logp(y i =j|x i ;θ)

[0108] Then use the Softmax function to j Exponentiation and normalization yield

[0109]

[0110] Where 1{y i =j} is an indicative function, that is, when the value in the brackets is true, the function result is 1, otherwise the result is 0. It represents the parameter penalty regularization term added to prevent overfitting, and λ represents the penalty coefficient.

[0111] Optional, regarding the optimization algorithm:

[0112] The time-frequency matrix obtained by STFT transformation of MBOK signal is selected as the training set sample, and the above convolutional neural network is trained by supervised learning. Since the number of training set samples is large, the amount of calculation for calculating the gradient descent is large, so the stochastic gradient descent algorithm is used to train the convolutional neural network: in each gradient descent calculation of the training algorithm, only one batch (50 sample matrices) is selected from the training set to complete the forward propagation and error back propagation calculation processes of the convolutional neural network.

[0113] (1) Network parameter initialization:

[0114] First, initialize the network parameters. The training algorithm of convolutional neural network is usually iterative (non-convex optimization), so it is necessary to set the starting point of the iteration. First, initialize the parameters of the convolutional neural network. The parameters of each layer of the network are initialized according to the standard method.

[0115]

[0116] Where U(·) represents uniform distribution, m represents the m inputs of the current layer, and n represents the n outputs of the current layer.

[0117] (2) Forward propagation calculation process:

[0118] Select R' from the training set modulus The sample matrix constitutes the input layer, passes through the convolution layer and the pooling layer, and then passes through the fully connected layer. Finally, according to the annotation information of the training sample matrix, the error between the prediction result of the current convolutional neural network model and the actual result (sample label) is calculated; the forward propagation convolution layer is calculated according to the following formula, where (*) represents convolution and the activation function uses the ReLU function.

[0119]

[0120] In the above formula Perform convolution operation for two variables;

[0121] The fully connected layer of the forward propagation is calculated as follows:

[0122]

[0123] In the above formula Multiply two variables.

[0124] (3) Error back propagation calculation process:

[0125] Then, based on the error between the predicted value and the true value obtained by forward propagation, the back propagation algorithm uses the stochastic gradient descent algorithm to calculate the gradient of each layer of network parameters in the order from the output layer to the input layer, and uses the gradient to update the network parameters of each layer, the purpose of which is to make the prediction results of this set of samples through the neural network model closer to the true value. Continue to select MBOK signal samples from the training set and repeat the above operation until the entire training set is traversed, that is, one iteration is completed.

[0126] When using the stochastic gradient descent algorithm, each iteration needs to traverse the entire training set. When the cost function decreases with the increase in the number of iterations until it finally converges to a suitable accuracy, the training process of the convolutional neural network is ended and the convolutional neural network parameter model is saved.

[0127] Algorithm 1:

[0128] Input: training set, learning rate η (learning rate decays linearly until the end of the iteration);

[0129] Output: multi-layer convolutional neural network with connection weight coefficients and offset values;

[0130] Initialization parameters: Initialize all parameters in the network within the range;

[0131] When the number of iterations of the operation is less than or equal to the set number of iterations, the following operation is performed:

[0132] Select k samples from the training set and calculate the output y of the current sample based on the current network parameters i ;

[0133] Calculate gradient estimates (calculate the gradient of output layer neurons, calculate the gradient of hidden layer neurons);

[0134] Update the model parameters and adjust the parameters in the negative gradient direction of the target:

[0135]

[0136] The iteration ends when the number of iterations executed by the operation is greater than the set number of iterations.

[0137] In one embodiment, in order to prevent overfitting and enhance the generalization ability of the convolutional neural network, training samples with signal-to-noise ratios of SNR=-5dB, -3dB, 0dB, and +3dB are generated during training as inputs for training the neural network. During training, validation set samples are added (the validation set and the training set are independent and identically distributed). When one iteration is completed, the test error of the validation set samples is used to approximate the generalization error, as shown in Table 2.

[0138] Table 2 Training samples and test samples

[0139]

[0140]

[0141] Training parameters:

[0142] Learning rate: The initial value is 0.05, and the learning rate η = η * 0.99 (the learning rate is updated after each iteration, and the learning rate decays linearly until the iteration ends);

[0143] Number of batch samples: 50;

[0144] Total number of iterations: 50;

[0145] Training: Each time, 50 samples are taken from the training sample set (total number of samples is 204,800), and a forward propagation calculation and an error back propagation calculation are completed. Then different samples are taken, and the process is repeated 4096 times to complete one iteration. A total of 50 iterations are required to complete the training.

[0146] The training error and verification accuracy are as follows Figure 5 As shown. Figure 5 It can be seen that as the training error decreases, the accuracy of the validation set samples gradually increases. This shows that as the number of iterations increases, the network parameters gradually converge to a suitable accuracy.

[0147] The present invention adopts Gold code or Walsh code as orthogonal spread spectrum coding to generate modulated signals, and obtains four network models according to the above training process, as shown in Table 3.

[0148] Table 3 Four training network models

[0149] Model Spread Spectrum Code Sampling point length Network Type Network Model 1 Gold Code 128 CNN Network Model 2 Gold Code 256 CNN Network Model 3 Walsh Code 128 CNN Network Model 4 Walsh Code 256 CNN

[0150] Then, test samples of the four network models are collected respectively, and the performance of the trained network models is tested. The test samples are input into the network, and the likelihood probability of the network output is calculated through forward propagation. A hard decision is made based on the likelihood probability to obtain the spread spectrum code state. Finally, the source information is restored based on the one-to-one mapping relationship between the spread spectrum code and the information bits sent by the transmitter. The demodulation performance simulation test of the four trained network models is carried out using test samples. The symbol error rate based on the Gold code is shown in Figure 1. Figure 6 As shown, the symbol error rate based on Walsh code Figure 7 shown.

[0151] The simulation results are compared with the performance of the MBOK demodulation method under ideal conditions (demodulation is achieved by using a matched filter based on the principle of maximizing the signal-to-noise ratio). As can be seen from the figure, the demodulation performance of the Gold code with a sampling point length of 256 is closest to the ideal performance. The demodulation performance of the Gold code and Walsh code with a sampling point length of 128 is poor. This is because the longer the sampling point, the more time averaging can be obtained through overlap when performing STFT calculations, which improves the demodulation performance at the expense of spectrum efficiency.

[0152] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A multi-level biorthogonal keying signal demodulation method based on STFT-CNN, characterized in that: include: Convert the information source into I-channel information bits and Q-channel information bits, map the information bits in the I-channel information bits and Q-channel information bits to the I-channel spreading code set and the Q-channel spreading code set, respectively, add repetition coding after each spreading code in the I-channel spreading code set and the Q-channel spreading code set, that is, after each spreading code is repeated several times, QPSK modulation is performed to generate an MBOK modulated signal; The time-frequency features in the MBOK modulated signal are extracted and input into the machine learning model, and the corresponding likelihood probability is output. The likelihood probability is used to indicate the category to which the current time-frequency feature sample belongs or the state of the spread spectrum code to which it belongs. According to the one-to-one mapping relationship between the category or the state of the spread spectrum code and the information bits of the source, the source information is restored.

2. The multi-level biorthogonal keying signal demodulation method based on STFT-CNN according to claim 1, characterized in that: The converting of the information source into I-channel information bits and Q-channel information bits, and mapping the information bits in the I-channel information bits and the Q-channel information bits to the I-channel spread spectrum code set and the Q-channel spread spectrum code set respectively, comprises: The source is converted into I-channel information bits and Q-channel information bits through serial-to-parallel conversion. Each m=log2M information bits in the I-channel information bits are mapped to the I-channel spreading code set C. I ={C I0 ,C I1 …,C I(M-1) }, each m = log2M information bits in the Q-path information bits are mapped to the Q-path spreading code set C Q ={C Q0 ,C Q1 …,C Q(M-1) }A spreading code in .

3. The multi-level biorthogonal keying signal demodulation method based on STFT-CNN according to claim 1, characterized in that: The spread spectrum code adopts Gold code or Walsh code as orthogonal spread spectrum code; the Gold code is composed of two m-sequence preferred pairs, the two m-sequence preferred pairs generate multiple orthogonal Gold codes, M code sequences are selected from the multiple orthogonal Gold codes to form an orthogonal spread spectrum code set for an M-ary dual orthogonal keying signal I path, and then M code sequences are selected from the remaining code sequences in the multiple orthogonal Gold codes to form an orthogonal spread spectrum code set for an M-ary dual orthogonal keying signal Q path.

4. The multi-level biorthogonal keying signal demodulation method based on STFT-CNN according to claim 3, characterized in that: The training sample acquisition process of the machine learning model includes: After obtaining the MBOK modulated signal, N sampling signal samples are intercepted according to the synchronization position of the spread spectrum code, and the sampling signal samples are short-time Fourier transformed to obtain the complex matrix R STFT , for the complex matrix R STFT Take the modulus of each element of and get R STFT The module value matrix R modulus ; The matrix R modulus Normalize to get the input sample matrix R of the convolutional neural network ′ modulus , the sample matrix R ′ modulus As training samples for convolutional neural networks.

5. The multi-level biorthogonal keying signal demodulation method based on STFT-CNN according to any one of claims 1 to 4, characterized in that: A convolutional neural network is used as a machine learning model. The constructed convolutional neural network includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a first fully connected layer, a second fully connected layer and an output layer, which are connected in sequence.

6. The multi-level biorthogonal keying signal demodulation method based on STFT-CNN according to claim 5, characterized in that: The cost function of the machine learning model is: Suppose there is a training data set X containing m samples: X={(x0,y0),…,(x i ,y i ),…(x Q ,y c )} Where X∈R n×n ,y i ∈{0,1,…,M·M}, M·M is the total classification category, M is the number of I / Q two-way spreading codes, and M·M is the number of combined states of I / Q two-way spreading codes; x i is the i-th sample in the sample set; y i For sample x i The corresponding sample label; Create a parameter model of the convolutional neural network to estimate the distribution probability of the training data set X. According to the principle of maximum likelihood estimation, maximizing the likelihood probability of the parameter model is equivalent to minimizing the cross entropy between the empirical distribution on the training data set and the probability distribution on the parameter model.

7. The multi-level biorthogonal keying signal demodulation method based on STFT-CNN according to claim 6, characterized in that: Maximizing the following formula is equivalent to minimizing the cost function θ Loss : Among them, logp(y i |x i ; θ) is the likelihood probability, θ is the network parameter of the convolutional neural network; Where j = {0, 1, ..., k-1}, j represents the jth category, and k represents the total number of categories; The second fully connected layer of the convolutional neural network outputs the unnormalized log probability z j , z j =logp(y i =j|x i ;θ) Use the Softmax function in the output layer to j Exponentiation and normalization yields: Among them, 1{y i =j} is an indicative function, that is, when the value in the brackets is true, the function result is 1, otherwise the result is 0; It represents the parameter penalty regularization term added to prevent overfitting, and λ represents the penalty coefficient.

8. The multi-level biorthogonal keying signal demodulation method based on STFT-CNN according to claim 1, characterized in that: The optimization algorithm of the machine learning model includes: Initialize the convolutional neural network parameters and select the sample matrix R from the training set ′ modulus It forms the input layer, passes through the first pooling layer, the second convolution layer, the second pooling layer, and then passes through the first fully connected layer and the second fully connected layer to enter the output layer. The output layer calculates the error between the prediction result of the current convolutional neural network model and the actual result based on the annotation information of the training sample matrix; Based on the error between the predicted value and the true value obtained by forward propagation, back propagation is performed in the order from the output layer to the input layer, using the stochastic gradient descent algorithm. Each iteration traverses the entire training set. When the cost function decreases with the increase in the number of iterations until it finally converges to the specified accuracy, the training process of the convolutional neural network is terminated and the parameter model of the convolutional neural network is saved.

9. The multi-level biorthogonal keying signal demodulation method based on STFT-CNN according to claim 8, characterized in that: Each iteration traverses the entire training set, including: The stochastic gradient descent algorithm is used to calculate the gradient of each layer of network parameters in turn, and the gradient is used to update the network parameters of each layer. MBOK signal samples are continuously selected from the training set until the entire training set is traversed, that is, one iteration is completed.

10. The multi-level biorthogonal keying signal demodulation method based on STFT-CNN according to claim 8 or 9, characterized in that: When training the machine learning model, training samples with different signal-to-noise ratios are generated as input for training the convolutional neural network. When one iteration is completed, the test error of the validation set samples is used to estimate the generalization error of the convolutional neural network obtained in this iteration.

Images