Constant modulus MIMO radar waveform design based on iterative optimization network method
Through the unsupervised dual iteration optimization network method, the MIMO radar waveform design is optimized using the deep residual network, which solves the problems of low waveform design accuracy and high calculation cost in the existing technology, and realizes waveform generation with low side lobes and low calculation cost.
Patent Information
- Application Number
- CN202411177164.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-26
- Publication Date
- 2024-11-15
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing MIMO radar waveform design methods are low in design accuracy and high computational cost, and cannot generate long sequences or multiple waveform sets in an acceptable time when optimizing weighted peak side lobe levels and weighted integral side lobe levels.
Using the unsupervised dual iterative optimization network (ION) method based on the iterative optimization network method, the waveform design is optimized by the deep residual network, and the weighted comprehensive objective function is constructed to reduce the weighted maximum autocorrelation sidelobe and the weighted maximum cross-correlation sidelobe, and through The outer iteration and inner iteration processes achieve waveform optimization.
A lower weighted maximum autocorrelation sidelobe and weighted maximum cross-correlation sidelobe are achieved, reducing computational costs and generating high-quality waveform sets in acceptable times, especially for designs of long sequences and multiple waveforms.
Smart Images

Figure CN118962599A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of radar technology, and in particular to a constant modulus MIMO radar waveform design based on an iterative optimization network method. Background Art
[0002] In MIMO radar systems, constant modulus waveform design is considered a key technology. Waveform design is an effective way to achieve multiple performance improvements, such as improving target detection performance and enhancing parameter estimation. In addition, constant modulus constraints on waveforms help ensure that the radar power amplifier operates at maximum efficiency and avoid problems caused by nonlinear distortion. Therefore, constant modulus waveform design in MIMO radars has attracted much attention.
[0003] Existing research can be mainly divided into two categories: (1) relaxation methods, which relax the CM (constant modulus) constraint or objective function. However, due to the non-rigid CM solution, the design accuracy is reduced; (2) phase design methods, but they require huge computational effort.
[0004] In the first category of methods, the typical ones are: The paper "He H, Stoica P, Li J. Designing unimodular sequence sets with good correlations—Including an application to MIMO radar [J]. IEEE Transactions on Signal Processing, 2009, 57 (11): 4391-4405." proposed a weighted cyclic algorithm (WeCAN) to minimize the weighted integrated sidelobe level (WISL) by relaxing the objective function. The paper "Wang J, Wang Y. On the design of constant modulus probing waveforms with good correlation properties for MIMO radar via consensus-ADMM approach [J]. IEEE Transactions on Signal Processing, 2019, 67 (16): 4317-4332." proposed a nonlinear alternating direction multiplier method (ADMM) to optimize WISL by relaxing the CM constraint, but this will lead to a decrease in design accuracy.
[0005] In the second category of methods, the typical ones are: The paper "Bolhasani M, Mehrshahi E, Ghorashi SA, et al. Constant envelope waveform design to increase range resolution and SINR incorrelated MIMO radar [J]. Signal Processing, 2019, 163: 59-65." proposed a typical LBFGS method to design the phase by calculating the Hessian matrix. This method requires a lot of calculations. In order to reduce the computational cost, the paper "Song J, Babu P, Palomar D P. Sequence set design with good correlation properties via majorization-minimization [J]. IEEE Transactions on Signal Processing, 2016, 64 (11): 2866-2879." proposed an optimization-minimization (MM) framework to derive a more tractable alternative phase function. However, it is challenging to construct such an alternative function according to the shape of the objective function. To solve this problem, the paper "Cui G, Yu X, Piezzo M, et al. Constant modulus sequence set design with good correlation properties [J]. Signal processing, 2017, 139: 75-85." proposed a block coordinate descent (BCD) method to transform the problem into a decomposable problem with multiple one-dimensional sub-problems. However, long sequence design requires a huge amount of computation.
[0006] Existing methods mainly optimize the weighted peak sidelobe level without considering the impact of the weighted peak sidelobe level. Therefore, the WPSL value in the existing methods is too high. In addition, as the sequence length or the number of waveforms increases, the existing methods cannot generate the required waveform set within an acceptable time. Summary of the invention
[0007] The invention aims to provide a constant modulus MIMO radar waveform design method based on an iterative optimization network method, which realizes constant modulus MIMO radar waveform design based on an unsupervised dual iterative optimization network (ION) method. The MIMO radar waveform designed by the invention has lower weighted maximum autocorrelation sidelobe weighted maximum cross-correlation sidelobe and lower calculation cost.
[0008] The technical solution adopted by the present invention is a constant modulus MIMO radar waveform design based on an iterative optimization network method, comprising the following steps:
[0009] Step 1: Construct the optimization objective function of constant modulus MIMO radar waveform design:
[0010]
[0011] y m (n)∈[0,2π]
[0012] Among them, L is the optimization objective function, is the waveform matrix of MIMO radar, M is the number of transmitting antennas, N is the number of sub-pulses, is the sequence transmitted by the mth antenna, s m (n) represents the mth sequence s m The nth sub-pulse of m (n) is s m (n) phase sequence; l1 is the first weight coefficient, and the second to fifth weight coefficients l2~l5 are the weight of the weighted autocorrelation peak sidelobe WAPSL, the weight of the weighted cross-correlation peak sidelobe level WCPSL, the weight of the weighted autocorrelation integrated sidelobe level WAISL, and the weight of the weighted cross-correlation integrated sidelobe level WCISL, respectively;
[0013] Step 2, constructing an unsupervised dual-iteration optimization network algorithm to solve the optimization objective function constructed in step 1, and obtaining a constant modulus MIMO radar waveform optimization result based on the solution result;
[0014] Among them, the unsupervised dual-iteration optimization network algorithm includes external iterative processing and internal iterative processing;
[0015] The external iterative process is used to replace the initial value y of the phase sequence of the waveform matrix S 0 , when the preset outer iteration convergence condition is met, the constant modulus MIMO radar waveform optimization result is obtained based on the optimal phase sequence y′ output most recently by the inner iteration processing;
[0016] The inner iteration process is based on the deep residual network with the phase sequence initialization value y 0 As input, the incremental phase sequence is obtained through the output of the deep residual network Based on the fusion factor, the initial value y of the phase sequence 0 and the current incremental phase sequence Fusion is performed to obtain the optimized phase sequence y t ; The superscript t is used to identify the number of inner iterations; the loss function of the deep residual network is set based on the optimization objective function. Before the inner iteration convergence condition is met, the current optimization phase sequence y t Calculate the loss value c of the deep residual networkt , and based on the loss value c t Update the network parameters of the deep residual network, and then set the initial value y 0 Input the updated deep residual network to obtain the optimized phase sequence y for the next inner iteration t+1 ; When the inner iteration convergence condition is met, it will have the minimum loss value c t The optimized phase sequence y t The optimal phase sequence y′ is output as the iterative processing in this round.
[0017] Furthermore, WAPSL, WCPSL, WAISL and WCISL are specifically:
[0018]
[0019] Among them, the weighted correlation γ k represents the lag weight, lag time k∈[-N+1,N-1], r ml (k) is the sub-pulse delayed by k times and Aperiodic cross-correlation sidelobes or periodic cross-correlation sidelobes Transmit antenna index l=1,2,...,M.
[0020] Furthermore, the loss function of the deep residual network is:
[0021]
[0022] Among them, the lag weight
[0023] Weighted Autocorrelation Matrix
[0024] Weighted cross-correlation matrix
[0025] Weighted Relevance | Related vectors ⊙ is the Hadamard product operator.
[0026] Furthermore, based on the fusion factor, the initial value y of the phase sequence 0 and the current incremental phase sequence When fusion is performed, ξ1 is defined to represent y 0 The adaptive factor, ξ2 represents The adaptive factor, whose initial value is the preset value, is based on all the loss values c currently obtained in each inner iteration. t The minimum value c′ in is adaptively updated:
[0027]
[0028] Among them, c 0 represents the loss value corresponding to the initial value y of the phase sequence 0 .
[0029] Furthermore, the outer iteration convergence condition is set as: where θ1 is a preset outer iteration convergence factor.
[0030] Furthermore, the inner iteration convergence condition is specifically set as one of the following two conditions:
[0031] Condition 1: The number of inner iterations t reaches the preset maximum number of iterations N max ;
[0032] Condition 2:
[0033] Define the convergence amounts of two consecutive times within the convergence interval E as j pre , j now respectively, and their initial values are both set to 0; among them, the convergence interval E is the preset number of sub-iterations, and E is less than N max ;
[0034] Update the new convergence amount j now = j now + c t , where c t represents the loss value of the deep residual network calculated based on the current optimized phase sequence y t .
[0035] When the number of inner iterations t = E, then update the previous convergence amount j pre = j now , and reset the new convergence amount j now = 0;
[0036] When the number of inner iterations t = zE, detect whether it satisfies: |(j now - j pre ) / (j pre )| < θ2 and n > N min , if so, it is considered that the inner iteration convergence condition is satisfied, and the inner iteration process is exited; otherwise, update j pre = j now , j now = 0;
[0037] Among them, the integer N min represents the preset minimum number of iterations, and θ2 represents the preset inner iteration convergence factor.
[0038] Furthermore, the deep residual network includes a fully connected layer, a stacked structure of multiple residual blocks and a fully connected layer in sequence, wherein the first fully connected layer and the last fully connected layer both include an activation function sigmoid, and each residual block includes two fully connected layers. The output of the first fully connected layer of the residual block is activated by the sigmoid function to obtain the input of the second fully connected layer of the residual block; then the input of the residual block is added to the fully connected layer, the output of the first fully connected layer of the residual block is activated by the sigmoid function to obtain the output of the residual block, and then the output of the residual block is obtained by the activation function sigmoid.
[0039] The technical solution provided by the present invention brings at least the following beneficial effects:
[0040] The invention realizes constant modulus MIMO radar waveform design based on the proposed unsupervised dual iterative optimization network (ION) algorithm, which includes two sub-processes of external iteration and internal iteration. The external iteration updates the input waveform, and the internal iteration optimizes the waveform through a residual network. The function of the external iteration is to replace the initial value and judge the convergence condition of the ION algorithm; the function of the internal iteration is to optimize the waveform by using the residual network to fix the initial value. Compared with the existing method, the invention has lower side lobes on both weighted maximum autocorrelation side (weighted maximum autocorrelation side, WMAS) and weighted maximum crosscorrelation side (weighted maximum crosscorrelation side, WMCS), and the calculation amount is small. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] 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 of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0042] Figure 1 Schematic diagram of the processing process of the outer iteration of the unsupervised dual iterative optimization network (ION) according to an embodiment of the present invention;
[0043] Figure 2 is a schematic diagram of the structure of a deep residual network used in an embodiment of the present invention;
[0044] Figure 3 is a schematic diagram of the structure of the residual block used in an embodiment of the present invention;
[0045] Figure 4 is a schematic diagram of weighted maximum cross-correlation sidelobe (WMAS) when M=3 and N=128;
[0046] Figure 5 It is a schematic diagram of WMCS when M=3 and N=128;
[0047] Figure 6 It is a schematic diagram of WMAS when M=10 and N=128;
[0048] Figure 7 It is a schematic diagram of WMCS when M=10 and N=128;
[0049] Figure 8 It is a schematic diagram of WMAS under the lag weight diversity;
[0050] Fig. 9 It is a schematic diagram of WMCS under lagged weight diversity;
[0051] Fig.10 It is a comparison chart of the six indicators of the comprehensive waveform. (a) to (f) in the figure correspond to: WPSL, WISL, WAPSL, WAISL, WCPSL, and WCISL respectively;
[0052] Fig.11 It is a curve comparison chart of the impact of different inputs on ION;
[0053] Fig.12 It is a comparison chart of the six indicators of the comprehensive waveform under all lag conditions, and (a) to (f) in the figure correspond to: PSL, ISL, APSL, AISL, CPSL, and CISL respectively. DETAILED DESCRIPTION
[0054] To make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be described in detail and completely in conjunction with the drawings in the implementation of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the present application for protection, but merely represents selected embodiments of the present invention.
[0055] The embodiment of the present invention provides a constant modulus MIMO radar waveform design based on an iterative optimization network method. Considering that the weighted peak sidelobe (WPSL) and the weighted integrated sidelobe level (WISL) are both important indicators, the embodiment of the present invention proposes a weighted comprehensive objective function (WCOF) for constant modulus MIMO radar waveform design, which is a linear combination of WPSL and WISL. By changing the weight coefficient of WCOF, it can be converted into the optimization problem of interest. Minimizing the WCOF problem is a complex problem that cannot be directly solved by existing methods. The present invention utilizes the powerful nonlinear mapping capability of the deep learning network and proposes an unsupervised dual iterative optimization network (ION) algorithm. First, the input waveform is updated through external iteration; then the waveform is optimized through the residual network through internal iteration; finally, through continuous updating and optimization, the desired waveform sequence is obtained while meeting the convergence conditions.
[0056] Consider a MIMO radar system with M transmit antennas. Each antenna transmits a waveform with N subpulses. A CM constraint is imposed on the waveform to improve the efficiency of the transmitter. The nth subpulse of the mth waveform is defined as
[0057]
[0058] Where, m=1,...,M, n=1,...,N. represents the sequence transmitted by the mth antenna, then the waveform matrix is expressed as The corresponding phase matrix is in, represents the complex domain, represents the field of real numbers, (·) T Represents a transpose operation.
[0059] When the lag is k and The non-periodic cross-correlation sidelobes of
[0060]
[0061] Where, m,l=1,...,M,-N+1≤k≤N-1,(.) * represents the conjugate transpose.
[0062] Similarly, when the lag is k and The periodic cross-correlation of
[0063]
[0064] Where, m,l=1,...,M,-N+1≤k≤N-1. If m=l, then (2) and (3) represent the non-periodic and periodic autocorrelation sidelobes, respectively.
[0065] WPSL is defined as
[0066]
[0067] in
[0068]
[0069] Among them, r ml (k) or k∈[-N+1,N-1],m=1,2,...,M,l=1,2,...,M,γ=[γ -N+1 ,...,γ k ,...,γ N-1 ] is the set of lag weights.
[0070] In order to define the peak sidelobe more comprehensively, WPSL is divided into weighted autocorrelation peak sidelobe (WAPSL) and weighted cross-correlation peak sidelobe level (WCPSL), which are defined as
[0071]
[0072] WISL is defined as
[0073]
[0074] Similarly, WISL is split into weighted autocorrelation integrated sidelobe level (WAISL) and weighted cross-correlation integrated sidelobe level (WCISL), which are expressed as
[0075]
[0076] like The above formula represents the non-periodic situation; if The above formula represents the periodic situation.
[0077] Typically, existing methods design waveforms with good correlation by minimizing WISL, which is expressed as
[0078]
[0079] m=1,...,M,n=1,...,N
[0080] In the embodiment of the present invention, a WCOF including WISL and WPSL is proposed. It is noted that the magnitudes of WPSL, WAISL and WCISL are quite different (WCISL has the largest magnitude, while WPSL has the smallest magnitude), so WCISL is used as a benchmark to unify the magnitudes. The proposed WCOF is:
[0081]
[0082] y m (n)∈[0,2π].
[0083] Among them, [l1,l2,l3,l4,l5] are weight coefficients. Since the magnitude has been normalized, the value range of the weight coefficient should be [0,1]. By adjusting the weight coefficient, both methods can be optimized at the same time.
[0084] Note that equation (12) is a high-order complex nonlinear optimization problem, which is difficult to solve directly using traditional methods. The embodiment of the present invention uses a neural network model to solve it.
[0085] In one embodiment, the constant modulus MIMO radar waveform design based on the iterative optimization network method provided by the embodiment of the present invention includes the following steps:
[0086] In order to effectively solve the optimization problem of formula (12), an unsupervised dual ION algorithm is proposed by utilizing the powerful nonlinear mapping ability of deep learning network. Figure 1 ,The unsupervised dual ION algorithm consists of two sub-algorithms, the outer iteration and the inner iteration.,The role of the outer iteration is to replace the initial value and determine the,convergence conditions of the algorithm. The role of the inner iteration is to optimize the waveform,using the residual network to fix the initial value.
[0087] A. External Iteration Algorithm,ION’s external iteration continuously uses the optimal phase obtained,from the internal iteration to update the initial value and judge the,convergence. The external iteration consists of four modules.
[0088] (1) Input module: Define the input phase sequence as:
[0089]
[0090] Among them, y m (n)∈[0,1] is the nth sub-pulse phase sequence transmitted by the mth antenna, m=1,...,M, n=1,...,N.
[0091] The loss function for the input sequence is set as:
[0092] c 0 =L(y 0 ,γ) (14)
[0093] Among them, L(.) is the loss function, and τ is the lag weight vector. In addition, fusion factors ξ1 and ξ2 are constructed to update the input value. If it is the initial input (random sequence), let ξ1 = 0, ξ2 = 1; if it is the optimized input, let ξ1 = 0.9, ξ2 = 0.1. That is, the sum of the fusion factors ξ1 and ξ2 is 1. The initial value can be set based on the actual scenario. For the optimized input, the initial value of ξ1 is greater than ξ2.
[0094] (2) Internal optimization module: The optimized normalized phase sequence is obtained through internal iteration:
[0095] (y′,c′)=F(y 0 ,ξ1,ξ2) (15)
[0096] in, is the normalized phase sequence after optimization, and c′ is the corresponding loss value.
[0097] (3) Convergence judgment module: The convergence criterion is defined as:
[0098]
[0099] Among them, θ1 is the external iteration convergence factor, which is taken as a preset value.
[0100] (4) Output module: When the convergence conditions are met, The obtained normalized phase matrix is then reconstructed as
[0101]
[0102] Among them, mat(.) is the transformation from vector to matrix. Finally, the waveform matrix is reconstructed as:
[0103]
[0104] In summary, the implementation process of the outer iteration algorithm of ION in the embodiment of the present invention is shown in Table 1:
[0105] Table 1
[0106]
[0107] B. Internal Iteration Algorithm: The internal iterative algorithm optimizes the waveform by combining the deep residual network with the Adam algorithm. The Adam algorithm has strong non-convex optimization capabilities. Because the residual network can solve the problem of gradient disappearance / explosion, this feature can be used to support deeper network layers and further enhance non-convex optimization capabilities. The internal iterative algorithm includes a deep residual network module, a backward propagation module, and a convergence judgment module.
[0108] (1) Deep residual network module: Figure 2As shown in Figure 1, it consists of an input fully connected layer, multiple residual blocks, and an output fully connected layer. The fully connected layers for input and output are used for dimensionality conversion. Using multiple residual blocks, each residual block consists of two fully connected layers, enhances the non-convex optimization capability of the algorithm.
[0109] Input (phase sequence y obtained by external iteration 0 ) and the output is expressed as:
[0110]
[0111] Among them, P(.) is the transfer function of the deep residual network, is the optimal phase sequence. i |i=1,...,2R n +2}, b={b i |i=1,...,2R n +2} is the weight vector and offset, where W i and b i Represent the weight vector and offset of the i-th fully connected layer, R n is the number of remaining residual blocks.
[0112] Specifically, this embodiment uses R n = 20 residual blocks and 2 fully connected layers to form a deep residual network. Each residual layer has K = 128 neurons, and the input and output fully connected layers have M × N neurons.
[0113] Using the constructed deep residual network, the input normalized waveform sequence is optimized by optimizing the weight vector W and the offset b.
[0114] The mathematical expression of the fully connected layer in the network is expressed as:
[0115] z i =x i W i +b i (20)
[0116] Among them, x i and z i are the input and output of the ith fully connected layer, respectively. Subsequently, the activation function sigmoid(.) = 1 / (1+e -(.) ) to ensure that the output waveform elements are within the normalized interval [0,1], we can get x i+1 =sigmoid(z i ).
[0117] like Figure 3 As shown in Figure 1, the residual block is built on the fully connected layer and adds an identity mapping to the deep residual network to solve the problem of gradient disappearance / explosion.
[0118] set up is the input of the ith residual block, similar to (20), we can get Combined with the identity mapping, we can get:
[0119]
[0120] Assume the input normalized vector is The corresponding output produced by the deep residual network is In order to make full use of the optimization information of each step, two fusion factors are used to adaptively add the input phase sequence and the optimized phase sequence, which are recorded as:
[0121]
[0122] Among them, ξ1 is the adaptive factor of the input waveform, ξ2 is the adaptive factor of the optimized waveform, and y n It is the normalized phase waveform generated by the nth optimization of the internal iteration.
[0123] (2) Backward propagation module: The backpropagation module calculates the loss function and updates the parameters of the neural network through the Adam optimizer.
[0124] Since the output is in normalized form, it needs to be converted to the interval [0,2π]. To facilitate correlation calculation, stacked in the form of a phase matrix in represents the waveform transmitted by the mth antenna. The embodiment of the present invention adopts a real-valued network structure, and the pulse signal is in the form of Therefore, the complex matrix is transformed into a real matrix and an imaginary matrix, which are:
[0125]
[0126] For correlation calculation, it is necessary to construct a signal expansion matrix. In the non-periodic case, the real part expansion matrix and the imaginary part expansion matrix are:
[0127]
[0128] The extended matrix in the periodic case is:
[0129]
[0130] Among them, H -N is the matrix formed by deleting the nth row of H, H -1 is the matrix formed by deleting the first row of H. Similarly, J -N and J -1is constructed in the same way. 0 represents a vector of all zeros, and the subscript is used to indicate the dimension.
[0131] To simplify, define the auxiliary amount and The extended matrices used to represent the imaginary part and the real part are respectively, in the case of non-periodic In the case of cycles Formulas (2) and (3) show that correlation can be calculated through convolution.
[0132] The correlation can be split into an imaginary and a real part:
[0133]
[0134] in, They are m The real and imaginary parts of (n), They are The real and imaginary parts of .
[0135] The equivalent calculation of formula (27) in the convolutional network is expressed as:
[0136]
[0137] in, is the convolution operator. Using the same calculation method, we can get
[0138] Therefore, the correlation vector can be obtained as:
[0139]
[0140] According to the defined lag weights The weighted correlation can be obtained as:
[0141]
[0142] Where ⊙ is the Hadamard product operator. Then, the weighted autocorrelation matrix is constructed as
[0143]
[0144] The weighted cross-correlation matrix is constructed as:
[0145]
[0146] The equivalent optimization criterion constructed by equations (31) and (32) is:
[0147] WAISL=ΣC (33)
[0148] WCISL=2×∑V (34)
[0149] WAPSL=max(C) (35)
[0150] WAISL=ΣC (36)
[0151] Among them, Σ(.) and max(.) are the summation operator and the maximum value operator respectively.
[0152] According to formula (12), the equivalent loss function can be expressed as:
[0153]
[0154] Where n = 1, 2, ..., N, c n is the loss value obtained at the nth inner iteration.
[0155] The sequence with the minimum loss value can be obtained as:
[0156]
[0157] After calculating the loss value, the adaptive factor is updated to
[0158]
[0159] Among them, c 0 =L(y 0 ,γ),c′=L(y′,γ),L() is the loss function.
[0160] Finally, the Adam algorithm is used to optimize W and B in the deep residual network module.
[0161] (3) Convergence judgment module of internal iteration: define the maximum number of iterations as N max , the minimum number of iterations is N min , the inner iteration convergence factor is θ2, the convergence interval is E, and the convergence amount of two consecutive times in the convergence interval E is j pre =0 and j now =0, the inner iteration will exit if one of the following two conditions is met.
[0162] Condition 1: If n = N max , then exit the iteration;
[0163] Condition two:
[0164] (1)j now =j now +c n ;
[0165] (2) When n = E, j pre =j now ,j now =0;
[0166] (3) When it is the case that if |(j now - j pre ) / (j pre )| < θ2 and n > N min , then exit the inner iteration; otherwise j pre = j now , j now = 0.
[0167] In summary, the implementation process of the inner iteration algorithm of ION in the embodiments of the present invention is shown in Table 2:
[0168] Table 2
[0169]
[0170]
[0171] Embodiment
[0172] In order to verify the performance of the MIMO radar waveform obtained by the constant modulus MIMO radar waveform design based on the iterative optimization network method proposed in the embodiments of the present invention, the existing scheme and the proposed scheme are used to optimize the waveform.
[0173] Define WMAS and weighted WMCS to evaluate the peak performance of the waveform. First, calculate the weighted peak autocorrelation and the weighted peak cross-correlation with lag:
[0174]
[0175] Then it can be obtained that:
[0176]
[0177] The simulation experiment conditions are set as follows: the convergence factor of the outer iteration is set to θ1 = 0.001, and the default input phase y 0 is a random normalized phase sequence. The convergence factor of the inner iteration is set to θ2 = 0.001, the maximum number of iterations is N max = 5000, the minimum number of iterations is N min = 1000, and the convergence interval is E = 100. The number of residual blocks of the residual network is R n = 20, the number of neurons in each fully connected network layer is K = 128, and the learning rate of the Adam algorithm is set to κ = 0.0001. In the following simulation settings, if not otherwise specified, the weight coefficients of WCOF are set to l1 = l2 = l3 = l4 = l5 = 1.
[0178] (1) Comparison of small waveform sets and short sequences.
[0179] The number of waveforms M = 3, the sequence length of each waveform N = 128, and the hysteresis weight is set to:
[0180]
[0181] Figure 4 and Figure 5 The performance of WMAS and WMCS is shown when the number of waveforms M = 3 and the sequence length N = 128. It can be found that the WMAS and WMCS of the ION method are better than the other three methods.
[0182] (2) Comparison of large waveform sets and short sequences.
[0183] The number of waveforms M = 10, the sequence length of each waveform N = 128, and the lag weight is set to:
[0184]
[0185] Figure 6 and Figure 7 The performance of WMAS and WMCS is shown when the number of waveforms M = 10 and the sequence length N = 128. It can be seen that the correlation performance of the existing method is reduced because the existing method mainly optimizes the WISL metric without considering the impact of WPSL. The ION method of this embodiment uses the powerful nonlinear mapping ability of the residual network to jointly optimize WPSL and WISL, thereby keeping all maximum correlations at a low level within a certain delay range of lag n = 0.
[0186] (3) Analysis of lagged weight diversity.
[0187] The number of waveforms M = 10, the sequence length of each waveform N = 128, and the lag weight is set to:
[0188]
[0189] Figure 8 and Fig. 9 WMAS and WMCS under hysteresis weight diversity are shown respectively. It can be seen that the ION method of this embodiment has better autocorrelation performance than other methods while maintaining excellent cross-correlation.
[0190] (4) Weighted lagged correlation evaluation.
[0191] A comprehensive evaluation of waveforms with different sequence lengths is performed, including six indicators and optimization time. Note that it is difficult for nonlinear ADMM to maintain strict CM in the case of long sequences, so this simulation experiment only compares the ION method with the BCD and WeCAN methods. Set the maximum number of iterations of WeCAN to 10000, BCD to 500, the number of signals M = 10, and the signal sequence N = [64, 128, 256, 512, 1024]. Due to the long optimization time of BCD and WeCAN, some long sequence waveforms cannot be optimized within an acceptable time. Then set the lag weight to:
[0192]
[0193] from Fig.10 As can be seen from (a) and (b), the WPSL of the waveform generated by ION optimization is worse than that of BCD and WeCAN methods, but WISL is not as good as BCD and WeCAN methods in the case of short sequences. This is because the ION method is used to optimize all indicators, while the BCD and WeCAN methods are mainly used to optimize WISL. However, as the sequence length increases, the present invention will maintain the best state in both WPSL and WISL. Fig.10 (c) and (d) show that ION leads other methods in WAISL and WAPSL. This is because ION’s optimized WCOF normalizes all metrics by order of magnitude, resulting in the same state for autocorrelation and cross-correlation.
[0194] Fig.10 (f) and (e) show that ION is better than WeCAN and BCD methods in WCPSL, but worse than WeCAN and BCD methods in WCISL. However, as the number of sequences increases, the WCISL of the waveform optimized by ION does not decrease significantly.
[0195] Then the optimization time of different methods is analyzed, and the results are shown in Table 3. It can be seen that when the sequence length is greater than 128, the computational cost is more than one order of magnitude lower than that of other methods. In addition, it can be found that the optimization time of ION in this embodiment does not increase sharply with the increase of the waveform sequence length. The reason is that no matter how long the sequence is, the number of network parameters that need to be optimized is quite fixed.
[0196] Table 3 Optimization time of different algorithms
[0197]
[0198] (5) The impact of initial input on optimization.
[0199] Set the number of signals M = 10 and the sequence length N = 128. Use the random initialization sequence and the WeCAN optimization sequence as the initial input. Since WeCAN optimizes WISL and does not normalize it according to the order of magnitude, the weight vector is set to [0, 0, 0, 1 / (M-1), 1] to ensure the consistency of the optimization target.
[0200] from Fig.11 It can be seen that the initial loss of WeCAN is 244583, and the loss after optimization becomes 212373, which shows that the ION method of this embodiment has the ability to improve the performance of the conventional method to optimize the waveform. Moreover, the optimization speed is faster than the random input, with the number of iterations being 9909 and 17410 respectively.
[0201] (6) Correlation evaluation of all lags.
[0202] The comparison is mainly with BCD and CAN methods. The weight vector of BCD is set to [0.5, 0.25, 0.25], and the maximum number of iterations of the CAN method is 10000. The number of waveforms is set to M = 3, and the sequence length is N = [8, 16, 32, 256, 512, 1024, 2048]. The weight vector of ION is [1, 1, 1, 1, 1], and the lag weight is γ = 1.
[0203] from Fig.12 It can be found that the ION method of this embodiment has better performance than the existing methods in terms of six indicators. In particular, as the number of sequences increases, the advantages of ION will be further reflected. As for ISL, AISL and CISL, the other methods are close to the optimal performance with limited improvement. In addition, the performance of PSL, APSL and CPSL has been significantly improved, and the improvement is more obvious in the case of long sequences, which further proves that the residual network has a stronger nonlinear mapping ability than the traditional method.
[0204] Through the above simulation, the performance advantage of the constant modulus MIMO radar waveform design method based on the iterative optimization network method disclosed in the present invention is verified. The present invention uses the powerful nonlinear mapping ability of the deep learning network to solve the optimization problem of the constant modulus MIMO radar waveform design constructed by the present invention, and can provide better performance at a lower computing cost.
[0205] 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 aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
[0206] The above are only some embodiments of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the creative concept of the present invention, which all belong to the protection scope of the present invention.
Claims
1. Constant modulus MIMO radar waveform design based on iterative optimization network method, characterized in that: The following steps are involved: Step 1: Construct the optimization objective function of constant modulus MIMO radar waveform design: Among them, L is the optimization objective function, is the waveform matrix of MIMO radar, M is the number of transmitting antennas, N is the number of sub-pulses, is the sequence transmitted by the mth antenna, s m (n) represents the mth sequence s m The nth sub-pulse of m (n) is s m (n) phase sequence; l1 is the first weight coefficient, and the second to fifth weight coefficients l2~l5 are the weight of the weighted autocorrelation peak sidelobe WAPSL, the weight of the weighted cross-correlation peak sidelobe level WCPSL, the weight of the weighted autocorrelation integrated sidelobe level WAISL, and the weight of the weighted cross-correlation integrated sidelobe level WCISL respectively; Step 2, constructing an unsupervised dual-iteration optimization network algorithm to solve the optimization objective function constructed in step 1, and obtaining a constant modulus MIMO radar waveform optimization result based on the solution result; Among them, the unsupervised dual-iteration optimization network algorithm includes external iterative processing and internal iterative processing; The external iterative process is used to replace the initial value y of the phase sequence of the waveform matrix S 0 , when the preset outer iteration convergence condition is met, the constant modulus MIMO radar waveform optimization result is obtained based on the optimal phase sequence y′ output most recently by the inner iteration processing; The inner iteration process is based on the deep residual network with the phase sequence initialization value y 0 As input, the incremental phase sequence is obtained through the output of the deep residual network Based on the fusion factor, the initial value y of the phase sequence 0 and the current incremental phase sequence Fusion is performed to obtain the optimized phase sequence y t ; The superscript t is used to indicate the number of inner iterations; the loss function of the deep residual network is set based on the optimization objective function. Before the inner iteration convergence condition is met, the current optimization phase sequence y t Calculate the loss value c of the deep residual network t , and based on the loss value c t Update the network parameters of the deep residual network, and then set the initial value y 0 Input the updated deep residual network to obtain the optimized phase sequence y for the next inner iteration t+1 ; When the inner iteration convergence condition is met, it will have the minimum loss value c t The optimized phase sequence y t The optimal phase sequence y′ is output as the iterative processing in this round.
2. The constant modulus MIMO radar waveform design based on iterative optimization network method as claimed in claim 1, characterized in that: WAPSL, WCPSL, WAISL and WCISL are as follows: Among them, the weighted correlation γ k represents the lag weight, lag time k∈[-N+1,N-1], r ml (k) is the sub-pulse delayed by k times and Aperiodic cross-correlation sidelobes or periodic cross-correlation sidelobes Transmit antenna index l=1,2,...,M.
3. The constant modulus MIMO radar waveform design based on iterative optimization network method as claimed in claim 2, characterized in that: The loss function of the deep residual network is: Among them, the lag weight Weighted Autocorrelation Matrix Weighted cross-correlation matrix Weighted relevance Related vector ⊙ is the Hadamard product operator.
4. The constant modulus MIMO radar waveform design based on iterative optimization network method as claimed in claim 1, characterized in that: Based on the fusion factor, the initial value y of the phase sequence 0 and the current incremental phase sequence When fusion is performed, ξ1 is defined to represent y 0 The adaptive factor, ξ2 represents The adaptive factor, whose initial value is the preset value, is based on all the loss values c currently obtained in each inner iteration. t The minimum value c′ in is adaptively updated: Among them, c 0 Indicates the initial value of phase sequence y 0 The corresponding loss value.
5. The constant modulus MIMO radar waveform design based on iterative optimization network method as claimed in claim 4, characterized in that: For the initial input phase sequence initial value y 0 , set the initial values of ξ1 and ξ2 to: ξ1 = 0, ξ2 = 1; when the optimal phase sequence y′ obtained based on the inner iteration process is assigned the initial value y 0 When ξ1 and ξ2 are initialized to ξ1=0.9 and ξ2=0.
1.
6. The constant modulus MIMO radar waveform design based on iterative optimization network method as claimed in claim 4, characterized in that: The outer iteration convergence condition is set as: Among them, θ1 is the preset outer iteration convergence factor.
7. The constant modulus MIMO radar waveform design based on iterative optimization network method as claimed in claim 1, characterized in that: The inner iteration convergence condition is specifically set to one of the following two conditions: Condition 1: The number of internal iterations t reaches the preset maximum number of iterations N max ; Condition two: Define the convergence amount of two consecutive times in the convergence interval E as j pre 、j now , whose initial values are all set to 0; where the convergence interval E is the preset number of sub-iterations, and E is less than N max ; Update the new convergence amount j now =j now +c t , where c t Indicates that based on the current optimized phase sequence y t The calculated loss value of the deep residual network; When the number of inner iterations t = E, the convergence value j before updating pre =j now , and reset the new convergence amount j now =0; When the number of inner iterations \(t = zE\), check whether the following conditions are satisfied: \(|(j now -j pre ) / (j pre )| < \theta_2\) and \(n > N min . If so, it is considered that the inner iteration convergence condition is satisfied; otherwise, update \(j pre =j now ,j now = 0;\) Among them, the integer N min represents the preset minimum number of iterations, and θ2 represents the preset inner iteration convergence factor.
8. The constant modulus MIMO radar waveform design based on iterative optimization network method according to any one of claims 1 to 7, characterized in that: The deep residual network includes a fully connected layer, a stacked structure of multiple residual blocks and a fully connected layer in sequence, wherein the first fully connected layer and the last fully connected layer both include an activation function sigmoid, and each residual block includes two fully connected layers. The output of the first fully connected layer of the residual block is activated by the sigmoid function to obtain the input of the second fully connected layer of the residual block; then the input of the residual block is added to the fully connected layer, the output of the first fully connected layer of the residual block is activated by the sigmoid function to obtain the output of the residual block, and then the output of the residual block is obtained by the activation function sigmoid.
9. The constant modulus MIMO radar waveform design based on iterative optimization network method as claimed in claim 8, characterized in that: The number of residual blocks is 20, the number of neurons in the first fully connected layer of the deep residual network is 128, the number of neurons in the fully connected layer of the residual block is 128, and the number of neurons in the last fully connected layer of the deep residual network is M×N.
Citation Information
Patent Citations
MIMO radar waveform design method based on deep learning
CN111060902A
MIMO radar beam forming method based on residual network
CN111693976A
MIMO radar waveform optimization method based on iterative optimization network
CN113050077A
Cited By
Low-complexity cascaded generalized sidelobe cancellation beam forming method and device
CN119851679A
