A discrete constant modulus quadrature phase encoding design method for multiple targets

Abstract

The present invention discloses a method for designing discrete constant modulus quadrature phase coding for multiple targets. The method comprises the following steps: S1: determining the number of waveform signals M and the number of code elements N of each waveform of a multi-input multi-output radar, and randomly generating training samples of a size of M×N; S2: calculating the distance of a range gate and the waveform correlation function, and setting a loss function based on the distance and correlation function; S3: inputting the training samples into an established phase coding network for training, performing modulo-1 and discretization processing on the network output to obtain discrete constant modulus phase coding; S4: calculating the coding loss value based on the loss function, and updating the network parameters based on the loss value; S5: proceeding to S6 upon completion of training, otherwise returning to S3; S6: obtaining a trained network, inputting the data to be coded into the trained network, and obtaining a discrete constant modulus quadrature phase coded signal. This method enables the multi-input multi-output radar to operate normally in multi-target scenarios, and the output is a discrete code, which is more suitable for practical applications.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar signal processing, and more specifically, relates to a discrete constant modulus orthogonal phase coding design method for multiple targets. Background Art

[0002] Faced with increasingly complex combat environments and missions, accurate acquisition of war information is a pressing requirement for future weaponry. Traditional phased array radars form narrow beams in the transmitting airspace, exhibit strong directivity, and possess high power, making them susceptible to enemy detection and interception. During target detection, the inherent limitations of phased arrays can lead to detection issues due to the radar cross-section flickering of some targets. Furthermore, because phased array radars require scanning of the airspace, they cannot resolve the conflict between real-time signal processing and bandwidth. Furthermore, the signals emitted by each transmitting channel of a phased array are uniformly separated, resulting in limited flexibility in subsequent signal processing.

[0003] MIMO radars, like MIMO communication systems, utilize multiple antennas for both transmission and reception, with independent transmission signals and receiving channels. MIMO radars are categorized into distributed and centralized MIMO radars. Centralized MIMO radars have a similar array layout to traditional radars, utilizing waveform diversity technology to transmit orthogonal waveform signals at the transmitter. Matched filtering separates these orthogonal signals at the receiver, increasing the radar's virtual aperture and the spatial degrees of freedom of the radar array. This results in higher degrees of freedom, range and Doppler resolution, improved anti-interception performance, and enhanced target detection capabilities.

[0004] The key to waveform diversity technology is orthogonal waveform design. Therefore, orthogonal waveform design is a key research direction for MIMO radars. Its performance is typically reflected in several aspects, including the mainlobe autocorrelation ratio (MCSR), orthogonality, and range resolution. Range resolution is determined by the width of the mainlobe output after matched filtering; a narrower mainlobe indicates better range resolution. The MCSR refers to the ratio of the mainlobe to sidelobe peaks after matched filtering. A high MSR can mitigate the problem of weak target obscuration. Orthogonality refers to the cross-correlation amplitude between different transmitted signals. The orthogonality of waveforms directly impacts signal separation at the receiver, thereby impacting the performance of MIMO radars. Excessively large cross-correlation coefficients can disrupt the radar's omnidirectional pattern. Excessively large and inconsistent imaginary parts of the cross-correlation coefficients or inconsistent autocorrelation peaks across waveforms can also reduce the radar's signal-to-noise ratio. Therefore, waveforms with good orthogonality are essential for the proper operation and full realization of the advantages of MIMO radars.

[0005] There are four main types of orthogonal waveform design: code division, frequency division, time division, and Doppler division. Phase coding offers greater freedom and flexibility in waveform autocorrelation and cross-correlation design. When the target velocity is within the waveform's Doppler tolerance, it can form a narrowband waveform with excellent orthogonal performance. Consequently, numerous researchers at home and abroad have studied orthogonal phase coding. Among them, there are waveform design methods based on convex optimization of signal correlation matrix, representative ones include new cycle algorithm, weighted new cycle algorithm and phase coding design algorithm based on alternating direction multiplier method, etc. These methods can design waveforms with good performance, but the algorithm iteration time is long. At the same time, these algorithms all produce continuous phases, and in reality, due to hardware limitations, continuous phases are limited in practical application, so discrete phases are a better choice; in addition, there are waveform design methods based on non-convex optimization methods, representative ones include waveform design methods based on binary region coordinate descent method, genetic algorithm, simulated annealing algorithm and residual fully connected network, etc. These methods directly optimize the waveform by establishing a loss function related to the correlation coefficient, and have also achieved good performance, but these methods only optimize the overall autocorrelation and cross-correlation performance, and cannot optimize partial areas in the autocorrelation and cross-correlation, and have limitations.

[0006] Furthermore, existing orthogonal phase-coded waveform design methods fail to consider waveform performance in multi-target scenarios, a common practice in real-world target detection. In these scenarios, the matched filter output of strong targets severely impacts the orthogonality of waveforms at weaker targets, preventing the range gate signals at these weak targets from separating the orthogonal waveforms, severely degrading the target detection performance of multi-input, multi-output radars. Consequently, orthogonal phase-coded waveforms designed using existing methods suffer varying degrees of performance loss when performing multi-target detection, particularly when detecting densely packed targets. Summary of the Invention

[0007] In order to avoid the problems of existing methods such as the limited application of continuous phase in practice and varying degrees of loss when performing multi-target detection, the present invention provides a discrete constant modulus orthogonal phase coding design method for multiple targets, improves the orthogonality performance of multi-input multi-output orthogonal phase coding waveforms in multi-target scenarios, and enables multi-input multi-output radars to operate normally in multi-target scenarios; and discretizes the output of the phase coding network so that the output phase coding is discrete phase coding, which is more suitable for practical applications.

[0008] In order to achieve the above-mentioned purpose of the present invention, the technical solutions adopted are as follows:

[0009] A discrete constant modulus quadrature phase encoding design method for multiple objectives, the method comprising the following steps:

[0010] S1: Determine the number M of orthogonal waveform signals emitted by the MIMO radar and the number N of code elements in each waveform, and randomly generate T data ranging from 0 to 1 with a size of M×N as training samples;

[0011] S2: Calculate the distance of the range gate and the waveform correlation function, and then set the loss function based on the distance of the range gate and the waveform correlation function;

[0012] S3: Input the training samples into the pre-established phase encoding network for training, and perform modulo 1 processing and discretization processing on the output of the phase encoding network in turn to obtain discrete constant modulus phase encoding;

[0013] S4: Calculate the loss value of discrete constant modulus phase encoding according to the set loss function, and update the parameters of the phase encoding network according to the loss value;

[0014] S5: Determine whether the training is finished, if so, go to S6, otherwise return to S3;

[0015] S6: Obtain a trained phase encoding network, input the M×N to-be-encoded data into the trained phase encoding network, and obtain a discrete constant modulus orthogonal phase encoding signal.

[0016] Preferably, the distance of the range gate is calculated based on the minimum detectable distance of the MIMO radar and a fixed distance interval.

[0017] Preferably, the calculation formula of the waveform correlation function is:

[0018]

[0019] in is the cross-correlation function between the m1th waveform and the m2th waveform with a delay of k. When m1=m2, is the autocorrelation function; represents the nth symbol of the m1th waveform, is the conjugate complex number representing the n+kth codeword of the m2th waveform.

[0020] Preferably, the specific calculation formula of the loss function is as follows:

[0021]

[0022] Where N is the number of symbols, M is the number of waveforms, APSL mid is the maximum value of the side lobe modulus of the waveform's autocorrelation coefficient, CPSL is the maximum value of the modulus of the cross-correlation coefficient, and AISL mid is the sum of the moduli of all autocorrelation coefficient side lobes, CISL is the sum of the moduli of all cross-correlation coefficients, AISL isumis the adjacent sum of the imaginary parts of all autocorrelation coefficient side lobes, CISL isum is the sum of the imaginary parts of all correlation coefficients, CISL rsum is the consecutive sum of the real parts of all correlation coefficients, It is the cumulative sum of all autocorrelation sidelobe coefficients whose real part is less than 0; l1~l8 are used to adjust APSL mid , CPSL, AISL mid , CISL, AISL isum , CISL isum , CISL rsum and The weight of , and l1+l2+l3+l4+l5+l6+l7+l8=1;

[0023] The APSL mid , CPSL, AISL mid , CISL, AISL isum , CISL isum , CISL rsum and The calculation formulas are as follows:

[0024]

[0025] in is the cross-correlation function between the m1th waveform and the m2th waveform with a delay of k, r mm (k) is the autocorrelation function with a time delay of k, real(·) and imag(·) refer to the real and imaginary parts respectively. round(·) refers to the rounding operator, R i is the distance of the i-th range gate, R i The fourth power of (x) neg Refers to taking the part of x that is less than 0.

[0026] Preferably, the distance R of the i-th range gate i The specific calculation formula is:

[0027] R i =R min +deltaR×(i-1),i=1,…,N

[0028] where R min is the distance to the first range gate, deltaR is the interval between each range gate, and N is the number of code elements.

[0029] Preferably, the APSL mid , CPSL, AISL midand CISL are the evaluation coefficients of waveform orthogonality performance in the case of single target; AISL isum , CISL isum , CISL rsum and It is the evaluation coefficient of the waveform orthogonality performance in the case of dense multi-target.

[0030] Preferably, the phase encoding network includes a long short-term memory network and a fully connected network, and is trained using an unsupervised training method; the parameter update method of the phase encoding network is to use the Adam method to calculate the gradient and update through back propagation.

[0031] Preferably, the output phase encoding of the discretized phase encoding network is specifically performed by first performing modulo 1 processing and normalization on the output of the phase encoding network, and then discretizing the phase encoding of the output of the phase encoding network using a nonlinear function;

[0032] The specific calculation formula for the modulo 1 processing is:

[0033]

[0034] Where xy is the phase code output by the phase coding network, x is the integer part of the phase code, y is the fractional part of the phase code, and Z mod The output of the phase encoding network is normalized by modulo 1 processing.

[0035] The specific calculation formula for the discretization process is:

[0036]

[0037] Where L is the number of binary digits, Z mod Z is the output of the phase encoding network after modulo 1 processing and normalization. dis is the output after discretization of the phase encoding of the phase encoding network,

[0038]

[0039] Preferably, the specific calculation formula of the discrete constant modulus phase encoding is:

[0040] Y=exp(j2πZ dis )∈R batch×N×M

[0041] Where Y is discrete constant modulus phase encoding, Z dis is the output after discretization of the phase encoding of the phase encoding network, batch is the number of samples in the mini-batch group for initializing phase encoding, and N is the number of code elements.

[0042] The beneficial effects of the present invention are as follows:

[0043] 1. Using a loss function calculated from the range gate distance and waveform correlation function to calculate the discrete constant modulus phase encoding loss value to update the phase encoding network, this effectively improves the problem of waveform orthogonality degradation in MIMO radars caused by the superposition of different target signals in multi-target situations. This allows the MIMO radar to maintain good waveform orthogonality and operate normally even when facing multiple targets, even densely packed targets, leveraging the unique advantages of MIMO radars.

[0044] 2. Discretizing the output of the phase coding network can prevent the data gradient of the phase coding network from being lost while maintaining the training of the phase coding network, and make the output sequence accurately represented by a finite binary number, so that the designed waveform can be easily quantized in practical applications while maintaining the original orthogonal performance effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 The present invention is a flowchart of a method for designing discrete constant modulus orthogonal phase coding for multiple targets in an embodiment of the present invention.

[0046] Figure 2 The figure is an overall block diagram of a phase coding network for a discrete constant modulus orthogonal phase coding design method for multiple targets in an embodiment of the present invention.

[0047] Figure 3 Schematic diagram of the internal structure of the long short-term memory network in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The following describes the embodiments of the present invention with reference to the accompanying drawings and preferred embodiments. Those skilled in the art will readily appreciate the other advantages and benefits of the present invention from the disclosure herein. The present invention may also be implemented or applied through various other specific embodiments, and the various details in this specification may be modified or altered based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are intended only to illustrate the present invention and are not intended to limit the scope of protection of the present invention.

[0049] It should be noted that the illustrations provided in the following embodiments are merely schematic illustrations of the basic concept of the present invention. Therefore, the illustrations only show components related to the present invention and are not drawn according to the number, shape, and size of components in actual implementation. In actual implementation, the type, quantity, and proportion of each component may be changed arbitrarily, and the component layout may also be more complex.

[0050] Example 1

[0051] A flowchart of a discrete constant modulus orthogonal phase encoding design method for multiple targets in an embodiment of the present invention is shown in FIG. Figure 1 The specific steps are as follows:

[0052] S1: Determine the number M of orthogonal waveform signals emitted by the MIMO radar and the number N of code elements in each waveform, and randomly generate T data ranging from 0 to 1 with a size of M×N as training samples;

[0053] S2: Calculate the distance of the range gate and the waveform correlation function, and then set the loss function based on the distance of the range gate and the waveform correlation function;

[0054] S3: Input the training samples into the pre-established phase encoding network for training, and perform modulo 1 processing and discretization processing on the output of the phase encoding network in turn to obtain discrete constant modulus phase encoding;

[0055] S4: Calculate the loss value of discrete constant modulus phase encoding according to the set loss function, and update the parameters of the phase encoding network according to the loss value;

[0056] S5: Determine whether the training is finished, if so, go to S6, otherwise return to S3;

[0057] S6: Obtain a trained phase encoding network, input the M×N to-be-encoded data into the trained phase encoding network, and obtain a discrete constant modulus orthogonal phase encoding signal.

[0058] In this embodiment, the number of waveforms M is set to 3, the number of symbols N is set to 128, and 1000 initialization phase codes X = [x1, x2, x3] with a scale of 3×128 are randomly generated. T ∈R M×N As training samples, where x m ∈R 1×N is the code of the mth group of waveforms. The value range of the elements in the sample matrix X is [0,1]. The data x in the sample matrix is ​​just the phase in the phase encoding. The actual phase encoding waveform is y = e j2πx , where every 10 are a small batch group, and there are 100 small batch groups in total, that is, the number of samples in a small batch group is 10. The input value range of the phase encoding network is a real number sequence in the range of [0,1], that is, the input matrix of a small batch group is where X n refers to the nth X in the mini-batch.

[0059] The overall block diagram of the phase coding network of a discrete constant modulus orthogonal phase coding design method for multiple targets in an embodiment of the present invention is as follows: Figure 2As shown, the phase encoding network includes a long short-term memory network and a fully connected network. The long short-term memory network has 7 hidden layers, each hidden layer has M units, and the cell parameter C of the mth cell in the nth layer is nm H nm ∈R batch×1024 All are initialized to 0. The internal diagram of the long short-term memory network is as follows Figure 3 As shown in the figure, there are two hidden layers and two cells. Figure 3 In the figure, each green block is a cell, the yellow circular block in each cell represents an operator, and the yellow rectangular block represents a fully connected network. Each cell contains four fully connected networks, where the main parameters have the weight W. f ,W i ,W c ,W o ∈R (q+1024)×1024 With the bias value b f ,b i ,b c ,b o ∈R 1×1024 , where q is the feature dimension of each layer input (the input feature dimension of the first layer is the number of code elements N). The internal structure of the cell can be divided into three parts: the input gate, the forget gate, and the output gate. The input gate determines which information is added to the historical memory; the forget gate determines which historical memories need to be discarded based on the output of the previous cell and the input of the current cell; the output gate determines the output information dimension and the size of their values. The long short-term memory network is connected to a fully connected network with 1024 neurons. The fully connected network also has weights W∈R 1024×N With b∈R 1×N The specific calculation process of the first hidden layer in the long short-term memory network is as follows:

[0060] Forget Gate:

[0061] Input Gate:

[0062] Output gate: H 12 =O t tanh(C 12 )

[0063] in refers to the mth waveform of the bth group of data in a mini-batch group, σ(·) and tanh(·) are two activation functions,

[0064] The output of the first cell in the first layer of the long short-term memory network is H 12 . H12 The operation continues as the input of the first cell in the second hidden layer. It means that H and The data of the corresponding rows of the two matrices are concatenated to form a new matrix. Then the output matrix Z of the LSTM network is:

[0065] Z={H 72 ,H 73 ,H 74}∈R batch×L×M .

[0066] The calculation process of the fully connected network is: where Z∈R batch×L×M is the output matrix of the fully connected network.

[0067] In order to make the phase coding network better able to train and optimize, the output needs to be normalized. Since the phase coding is in the form of complex exponentials, and complex exponentials have periodic properties. Therefore, the phase coding output of the phase coding network is also periodic, with a period of 1. Based on this property of period 1, we normalize the output of the phase coding network through modulo 1 processing. The modulo 1 function is as follows,

[0068]

[0069] Where xy refers to a real number with an integer part of x and a decimal part of y.

[0070] Perform modulo 1 processing on the network output, and the output is Z mod =mod1(Z), and then use a nonlinear function to discretize the network output; this can ensure that the data gradient is not lost during training while the output sequence can be accurately represented by a limited number of binary numbers, making the designed waveform easy to quantize in practical applications while maintaining the original orthogonal performance. The specific steps are as follows:

[0071] First, set the number of binary digits to L. Here we take L = 10, that is, the discretized phase can be represented by a ten-bit binary number, and then use the nonlinear function Z mod The specific calculation formula for integer processing is as follows:

[0072]

[0073] The final discrete constant modulus phase encoding is Y = exp(j2πZ dis )∈R batch×N×M , so the output phase-coded signal is also constant modulus.

[0074] Calculate the waveform correlation function. Generally speaking, the phase-coded signal is in the form of a complex exponential of samples, i.e. y = exp(2πjx), where j is an imaginary unit. The cross-correlation function between the m1th waveform and the m2th waveform with a delay of k is is defined as:

[0075]

[0076] represents the nth symbol of the m1th waveform, is the conjugate complex number representing the n+kth symbol of the m2th waveform. When m1=m2, is the autocorrelation function. Since real number operations are commonly used in current neural networks, it is necessary to convert the waveform correlation function operations into real number operations. The phase-coded signal y can be expressed as:

[0077] y=exp(2πjx)=cos(x)+jsin(x)

[0078] The waveform correlation function operation can be expressed as:

[0079]

[0080] The modulus, real part and imaginary part of the waveform correlation function can be calculated separately.

[0081] In the case of multiple targets, the main reason for the deterioration of the orthogonality performance of the signal waveform at the range gate where the weak target is located is the interference of the strong target signal on the weak target signal, and the main difference between strong and weak targets is the difference in signal power. Therefore, it is necessary to reduce the interference caused by the strong target signal on the weak target signal, that is, to design the waveform according to the data power of each range gate. The power difference between targets with similar characteristics located at different range gates mainly comes from the difference in distance. However, since the power of the superimposed signal of the data at different range gates is different, it is impossible to accurately calculate the signal superposition power of all range gates. The waveform can only be roughly designed by using the ratio of the distances of different range gates to approximate the power ratio. The distance of the range gate is calculated based on the minimum detectable distance of the multi-input multi-output radar and the fixed distance interval.

[0082] Based on the parameters set by the MIMO radar itself, calculate the distance information at the first N range gates. Let the distance of the first range gate be Rmin, and the interval between each range gate be deltaR. According to the radar equation, the distance of the i-th range gate is:

[0083] R i =R min +deltaR×(i-1),i=1,…,N

[0084] Calculate the loss function. The network's loss function is related to the maximum and cumulative sum of the waveform autocorrelation sidelobe coefficients and the maximum and cumulative sum of the cross-correlation coefficients in the single-target case, as well as the waveform autocorrelation coefficients and cross-correlation coefficients at different targets in the multi-target case. The loss function is designed to reduce the waveform autocorrelation sidelobes and cross-correlation coefficients while ensuring that the real part of the autocorrelation sidelobes is as positive as possible and that the amplitudes of the imaginary parts are canceled out as much as possible when they are superimposed. Furthermore, the amplitude of the superposition of the waveform cross-correlation coefficients of different range gates divided by the fourth power of the distance of the corresponding range gate is minimized, thereby minimizing the impact of the waveform orthogonality degradation caused by the superposition of different target signals.

[0085] Defining APSL mid is the maximum value of the side lobe modulus of the waveform's autocorrelation coefficient, CPSL is the maximum value of the modulus of the cross-correlation coefficient, and AISL mid is the sum of the moduli of all autocorrelation coefficient side lobes, CISL is the sum of the moduli of all cross-correlation coefficients, AISL isum is the adjacent sum of the imaginary parts of all autocorrelation coefficient side lobes, CISL isum is the sum of the imaginary parts of all correlation coefficients, CISL rsum is the consecutive sum of the real parts of all correlation coefficients, is the cumulative sum of all autocorrelation sidelobe coefficients whose real part is less than 0, which can be expressed as:

[0086]

[0087] in is the cross-correlation function between the m1th waveform and the m2th waveform with a delay of k, r mm (k) is the autocorrelation function with a time delay of k, real(·) and imag(·) refer to the real and imaginary parts respectively. round(·) refers to the rounding operator, R i is the distance of the i-th range gate, R i The fourth power of (x) neg Refers to taking the part of x that is less than 0. Note that APSL mid , CPSL, AISL mid and CISL are the evaluation coefficients of waveform orthogonality performance in the case of single target; AISL isum , CISL isum , CISL rsum and AISL rsum is the evaluation coefficient of the waveform orthogonality performance in the case of dense multi-target. Therefore, the loss function can be set as:

[0088]

[0089] Where N is the number of symbols, M is the number of waveforms, l1+l2+l3+l4+l5+l6+l7+l8=1, where l1=0.14, l2=0.14, l3=0.14, l4=0.18, l5=0.18, l6=0.09, l7=0.09, l8=0.09. These coefficients are used to adjust APSL. mid , CPSL, AISL mid , CISL, AISL isum , CISL isum , CISL rsum and The weights of these eight indicators are determined by the coefficient. The larger the coefficient, the greater the weight of the corresponding indicator. The phase encoding output by the trained network is more consistent with the standard of the indicator with the larger weight. However, it is impossible for the phase encoding output by the network to simultaneously make all eight of the above indicators zero. Therefore, it is necessary to adjust the weights between the different indicators to ensure that the phase encoding output by the network has good orthogonal performance in the multi-target case.

[0090] The loss is calculated by the loss function, and the Adam method is used to calculate the gradient and update the network parameters of the phase encoding network through back propagation. The specific formula is:

[0091] (W f ,W c ,W o ,W i ,W,b c ,b o ,b i ,b f ,b)=min(Loss)

[0092] Here, the learning rate lr is initialized to 0.0005 and the training cycle is 2000. The Adam optimizer is a variant of the gradient descent algorithm used to update the weights of the neural network. The process of updating the weight W using the Adam optimizer is as follows,

[0093]

[0094] in, is the gradient of W in the loss function, M t The first moment estimate of the parameter gradient, V t is the second-order moment estimate of the parameter gradient, β1, β2 are the attenuation coefficients, which we set to β1 = 0.9, β2 = 0.99, M0 and V0 are initialized to 0, due to the large attenuation coefficient, M t and V t The update is mainly affected by M t-1 and V t-1 In order to avoid the influence of M t and Vt In the early stage of training, it is biased towards 0 to reduce the impact of the deviation on the early stage of training. t and V t Correct the deviation. and It is best to update the weight W. In the formula, ε is a very small number, which is used to prevent the denominator from being 0. Here we set ε = 0.00000001.

[0095] Similarly, the update process of bias b is the same as that of W, so we will not go into details here.

[0096] Using this loss function to update the network parameters of the phase coding network can effectively improve the problem of reduced orthogonality in MIMO radar waveforms caused by the superposition of different target signals in multi-target situations. This allows the MIMO radar to maintain good orthogonality in the waveforms even when facing multiple targets, even densely packed targets, and function normally, leveraging the unique advantages of MIMO radars.

[0097] The phase coding network is trained and terminated after a given number of training times. The phase coding network obtained at this time can input any random sequence for encoding and output discrete constant modulus orthogonal phase coding with good orthogonal performance. The method of this embodiment is faster than the traditional design method. At the same time, the phase coding network uses an unsupervised training method. No labeled data is required during training. Only a randomly generated sequence that meets the requirements can be provided to the phase coding network for training.

[0098] Obviously, the above embodiments of the present invention are merely examples for the purpose of illustrating the present invention, and are not intended to limit the embodiments of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A discrete constant modulus quadrature phase encoding design method for multiple targets, characterized by: The method comprises the following steps: S1: Determine the number M of orthogonal waveform signals emitted by the MIMO radar and the number N of code elements in each waveform, and randomly generate T data ranging from 0 to 1 with a size of M×N as training samples; S2: Calculate the distance of the range gate and the waveform correlation function, and then set the loss function based on the distance of the range gate and the waveform correlation function; S3: Input the training samples into the pre-established phase encoding network for training, and perform modulo 1 processing and discretization processing on the output of the phase encoding network in turn to obtain discrete constant modulus phase encoding; S4: Calculate the loss value of discrete constant modulus phase encoding according to the set loss function, and update the parameters of the phase encoding network according to the loss value; S5: Determine whether the training is finished, if so, go to S6, otherwise return to S3; S6: Obtain a trained phase encoding network, input the M×N to-be-encoded data into the trained phase encoding network, and obtain a discrete constant modulus orthogonal phase encoding signal; The calculation formula of the waveform correlation function is: in is the cross-correlation function between the m1th waveform and the m2th waveform with a delay of k. When m1=m2, is the autocorrelation function; represents the nth symbol of the m1th waveform, is the conjugate complex number representing the n+kth symbol of the m2th waveform; And, the specific calculation formula of the loss function is as follows: Where N is the number of symbols, M is the number of waveforms, APSL mid is the maximum value of the side lobe modulus of the waveform's autocorrelation coefficient, CPSL is the maximum value of the modulus of the cross-correlation coefficient, and AISL mid is the sum of the moduli of all autocorrelation coefficient side lobes, CISL is the sum of the moduli of all cross-correlation coefficients, AISL isum is the adjacent sum of the imaginary parts of all autocorrelation coefficient side lobes, CISL isum is the sum of the imaginary parts of all correlation coefficients, CISL rsum is the consecutive sum of the real parts of all correlation coefficients, It is the cumulative sum of all autocorrelation sidelobe coefficients whose real part is less than 0; l1~l8 are used to adjust APSL mid , CPSL, AISL mid , CISL, AISL isum , CISL isum , CISL rsum and , and l1+l2+l3+l4+l5+l6+l7+l8=1.

2. The method for designing discrete constant modulus quadrature phase encoding for multiple objectives according to claim 1, characterized in that: The distance of the range gate is calculated based on the minimum detectable distance of the MIMO radar and a fixed distance interval.

3. The discrete constant modulus quadrature phase encoding design method for multiple targets according to claim 1, characterized in that: The APSL mid , CPSL, AISL mid , CISL, AISL isum , CISL isum , CISL rsum and The calculation formulas are as follows: in is the cross-correlation function between the m1th waveform and the m2th waveform with a delay of k, r mm (k) is the autocorrelation function with a time delay of k, real(·) and imag(·) refer to the real and imaginary parts respectively. round(·) refers to the rounding operator, R i is the distance of the i-th range gate, R i The fourth power of (x) neg Refers to taking the part of x that is less than 0.

4. The method for designing discrete constant modulus quadrature phase encoding for multiple objectives according to claim 3, wherein: The distance R of the i-th range gate i The specific calculation formula is: R i =R min +deltaR×(i-1),i=1,…,N where R min is the distance to the first range gate, and deltaR is the interval between each range gate.

5. The method for designing discrete constant modulus quadrature phase encoding for multiple objectives according to claim 3, wherein: The APSL mid , CPSL, AISL mid and CISL are the evaluation coefficients of waveform orthogonality performance in the case of single target; AISL isum , CISL isum , CISL rsum and AISL rsum It is the evaluation coefficient of the waveform orthogonality performance in the case of dense multi-target.

6. The method for designing discrete constant modulus quadrature phase encoding for multiple objectives according to claim 1, characterized in that: The phase encoding network includes a long short-term memory network and a fully connected network, and is trained using an unsupervised training method; the parameter update method of the phase encoding network is to use the Adam method to calculate the gradient and update through back propagation.

7. The method for designing discrete constant modulus quadrature phase encoding for multiple objectives according to claim 1, characterized in that: The specific calculation formula for the modulo 1 processing is: Where xy is the phase code output by the phase coding network, x is the integer part of the phase code, y is the fractional part of the phase code, and Z mod The output of the phase encoding network is normalized by modulo 1 processing.

8. The method for designing discrete constant modulus quadrature phase encoding for multiple objectives according to claim 7, characterized in that: The specific calculation formula for the discretization process is: Where L is the number of binary digits, Z mod Z is the output of the phase encoding network after modulo 1 processing and normalization. dis is the output after discretization of the phase encoding of the phase encoding network, 9. The method for designing discrete constant modulus quadrature phase encoding for multiple objectives according to claim 8, characterized in that: The specific calculation formula of the discrete constant modulus phase encoding is: Y=exp(j2πZ dis )∈R batch×N×M Where Y is discrete constant modulus phase encoding, Z dis is the output after discretization of the phase encoding of the phase encoding network, batch is the number of samples in the mini-batch group for initializing phase encoding, and N is the number of code elements.

Images