1. Introduction
Multiple-input multiple-output (MIMO) radar uses waveform diversity techniques to improve power efficiency, clutter suppression ability, and other performances. A monostatic MIMO radar with
M transmitting/receiving antennas is shown in
Figure 1. The
M signals transmitted by the
M antennas are different. The echo of all transmitted signals are collected by each receiving antenna. MIMO diversity gain is achieved by an ideal separation of the echo. The filter
hm,
m = 1, 2, …,
M is matched to the
m-th transmit signal. The echo from the other
M-1 transmitted signals causes correlation interference at the output of
hm. A MIMO waveform set is orthogonal when the cross-correlation function between any pair of
M transmit signals is zero. Although orthogonality is not achievable within a limited time-bandwidth product, a low cross-correlation is achievable. Furthermore, the sidelobe of the auto-correlation of the waveform set should be lowered to improve pulse compression performance. The research of the MIMO radar waveform design focuses on reducing the peak of auto-correlation function sidelobe and cross-correlation function within a limited time-bandwidth product [
1]. Orthogonal frequency division multiplexing (OFDM) signal is suitable for application in MIMO radar systems because of its high range resolution, high Doppler resolution, and high design freedom [
2,
3,
4,
5,
6,
7].
There has been much research on the OFDM waveform for single-input single-output (SISO) radar systems. Initially, OFDM waveforms were used in multi-carrier FMCW radar systems [
8]. Subsequent studies of the OFDM waveform have adopted this modulation strategy of sub-carriers in their framework. However, most existing radar OFDM waveforms have a non-constant time-domain modulus. As the power amplifiers of real radar systems are generally non-linear, it is critical to reduce the peak-to-mean envelope power ratio (PMEPR) and peak-to-average power ratio (PAPR) of OFDM waveforms. At the same time, it is also important to optimize the performance of the auto-correlation function and ambiguity function [
9,
10,
11]. To solve this problem, practical radar systems can use power amplifiers with higher dynamic range and linearity, but the trade-offs are higher system cost and complexity. Several researchers have evaluated the impact of directly limiting the amplitude of the OFDM waveform, and showed that the magnitude of noise and interference increases [
12]. Besides, the direct design of the constant envelope OFDM waveform is another way. R. Mohseni et al. [
13] and Wen-Qin Wang et al. [
14] have proposed several constant envelope OFDM waveforms with good correlation performance and flexible parameter design.
In recent years, with the development of MIMO radar, there has been an increasing number of studies on designing OFDM waveform sets for MIMO radar systems. According to the different sub-carriers, the MIMO radar OFDM waveform sets are divided into two categories: LFM-OFDM [
15,
16,
17,
18,
19,
20] and PC-OFDM [
21,
22,
23,
24,
25]. The sub-carrier of the LFM-OFDM waveform set is a chirp signal, and thus, the LFM-OFDM waveform set has better Doppler tolerance. The PC-OFDM waveform set consists of multi-carrier phase-coded signals. Comprehensive and in-depth studies on single-carrier phase-coded waveform set for MIMO radar can be found in [
26,
27,
28,
29], and not many problems are left in the multi-carrier case. LFM-OFDM waveform sets have better prospects in MIMO-SAR systems due to the advantages of chirp signals [
30,
31,
32]. Wen-Qin Wang [
33] has designed a constant envelope MIMO radar LFM-OFDM waveform set, termed OFDM chirp, modulated by random sub-chirp codes. OFDM chirp waveform set has relatively high cross-correlation function peaks [
34]. If its cross-correlation peaks are reduced to a certain extent, the OFDM chirp waveform can be applied to not only MIMO synthetic aperture radar (SAR) systems [
35,
36], but also other MIMO radar systems [
37]. Compared with LFM signals, non-linear frequency modulation (NLFM) signals have a lower auto-correlation sidelobe and a larger degree of freedom. Based on the time-frequency structure of NLFM signals, a possible approach to suppress the correlation peaks of the OFDM chirp waveform set is to use different sub-chirp rates, e.g., Gao et al. [
38] use the time-frequency structure of the NLFM signals to construct the sub-chirps of the OFDM chirp waveform set with lower cross-correlation peaks. The obtained new waveform set is referred to as the OFDM-NLFM waveform set in this paper. Since then, there have been several studies on the OFDM-NLFM waveform set [
39,
40]. Currently, the IN-OFDM designed by Xiang Lan et al. [
41] is the OFDM-NLFM waveform set with the lowest peak of auto-correlation sidelobe ratio (PASR) and cross-correlation ratio (PCCR).
This paper develops the signal model and the optimization model of the OFDM-NLFM waveform set. The time-frequency structure of the OFDM-NLFM waveform set is constructed by sub-chirp signals with different time-width, center frequency, and chirp rates. After determining the number of transmitted signals of the OFDM-NLFM waveform set, the number of sub-bands, and other system parameters, an OFDM-NLFM waveform set is generated by the following steps. The first step is to initialize the NLFM signal’s time-frequency structure (NLFM parameters). Then, the time-frequency structure of the NLFM signal is divided into several sub-bands, or sub-pulses. Next, we replace the nonlinear frequency modulation curve in each sub-band with a linear frequency modulation curve and generate the piecewise chirp signal to approximate the NLFM signal. Finally, the sequence order of sub-chirp signals is permutated according to the sub-chirp sequence code matrix, and the sign of chirp rates is changed according to the sub-chirp rate plus and minus (PM) code matrix. Especially, each row of the sub-chirp sequence code matrix is a permutation that determines the order of the sub-chirp sequence for each transmit signal of the OFDM-NLFM waveform set, and each element of the sub-chirp rate PM code matrix is a 1-bit binary number that determines whether the sub-chirp signal is an up-chirp signal or a down-chirp signal. This paper defines the objective function as the sum of the cross-correlation functions considering that the correlation interferences from the other transmit signals are additive.
To design the OFDM-NLFM waveform set with the lowest possible PASR and PCCR, this paper proposes an OFDM-NLFM waveform set design algorithm based on alternating optimization. There are both discrete code matrices and continuous NLFM parameters of the OFDM-NLFM waveform set. The expression between the objective function and the waveform parameters consists of time-frequency structure transformation and correlation function calculation. Hence, the OFDM-NLFM waveform set design problem is a high-dimensional optimization problem with a complex objective function and complicated optimization variables. Dimensionality reduction and decomposition are efficient solutions to a high-dimensional optimization problem with fast execution [
42,
43,
44]. The method proposed in this paper adopts the idea of the alternating direction multiplier method (ADMM) [
45], which is also known as alternating optimization. The parameters of the OFDM-NLFM waveform set to be optimized include continuous and discrete variables. For continuous variables, using the particle swarm optimization (PSO) algorithm [
46] to find the minimum value of the complex objective function is simple and efficient. For discrete variables, it is difficult to find the optimal solution analytically since the sub-chirp sequence code matrix and the sub-chirp rate PM code matrix are unconstrained. The coordinate descent (CD) algorithm [
47] is efficient to solve the unconstrained optimization problem. Inspired by the CD algorithm, a new block coordinate descent (BCD) algorithm is proposed to find the optimal solution of the sub-chirp sequence code matrix and the sub-chirp rate PM code matrix. In this method, the optimization parameter matrices are divided into blocks according to different transmit signals that they determine. In short, the OFDM-NLFM waveform set design algorithm based on alternating optimization is composed of two sub-algorithms; the PSO algorithm optimizes the continuous NLFM parameters, and the BCD algorithm optimizes the sub-chirp sequence code matrix and the sub-chirp rate PM code matrix. After either of the two sub-algorithms has been executed, the optimal result is passed to the other sub-algorithm, ensuring that the objective function decreases monotonically until it converges.
This paper develops the signal model and parameter optimization model of the OFDM-NLFM waveform set. A novel OFDM-NLFM waveform set design algorithm based on alternating optimization is proposed. With the same system parameters as the current optimal IN-OFDM waveform, the PASR and PCCR of the obtained waveform set are about 5 dB lower than that of the IN-OFDM. The rest of this paper is organized as follows.
Section 2 introduces the signal model of the OFDM-NLFM waveform set and establishes the mathematical model of MIMO radar OFDM-NLFM waveform set optimization design problem.
Section 3 presents the details of the OFDM-NLFM waveform set design algorithm based on alternating optimization.
Section 4 performs numerical simulations under different system parameters. The results are compared with the current optimal OFDM-NLFM waveform set, IN-OFDM. The influences of different parameters on the algorithm design results are analyzed.
Section 5 is the conclusion.
3. OFDM-NLFM Waveform Set Design Algorithm Based on Alternating Optimization
In the above-mentioned MIMO radar OFDM-NLFM waveform set optimization problem, the solution space is of high dimension, and its objective function is complex. This paper proposes an OFDM-NLFM waveform set design algorithm based on alternating optimization. Especially, we divide the solution space into sub-spaces according to its characteristics, and the optimization problem in sub-spaces is viable to solve.
For an OFDM-NLFM waveform set with M transmit signals, each of which has N sub-chirp signals, the optimization variable X includes three parts. X1 is NLFM parameters consisting of 2M continuous variables. X2 is a matrix consisting of MN binary bits, with 2MN different values. X3 consists of M permutations, with a total of (N!)M different values. Thus the dimension of the solution space Ω is very high, and the structure of the optimization variable is complex containing continuous, discrete variables and permutations. To map from the solution space to the objective function domain, we first map the solution space to the waveform, and then calculate the performance indicators of the MIMO waveform set. A difficulty is that the simple auto- and cross-correlation functions are not elementary functions, but instead the functionals of the waveform set. To conclude, the objective function does not possess good properties such as continuity and derivability. The OFDM-NLFM waveform set optimization problem defined in Equation (25) is an NP-hard constrained optimization problem with mixed discrete and continuous optimization variables. Its objective function is nonlinear, nonconvex, and non-differentiable. In this case, the common linear programming, continuous optimization, and convex optimization methods fail, and the other optimization design algorithms for the MIMO radar waveform set cannot apply to the OFDM-NLFM waveform set as they are limited to its specific waveform set.
As the objective function is difficult to be simplified, the most direct approach is to separate discrete variables from continuous variables. Alternating optimization is a computational framework to solve high-dimensional optimization problems [
45]. Alternating optimization decomposes the large global problem into smaller and easier-solved sub-problems by coordinate decomposition and obtains the solution of the global problem by combining the solutions of the sub-problems. The OFDM-NLFM waveform set design algorithm proposed by this paper is based on the idea of alternating optimization. If we keep the continuous variables unchanged and optimize the discrete variables, the problem is a combinatorial optimization problem. Instead, if we keep the discrete variables unchanged, the problem is a continuous optimization problem. This paper selects appropriate optimization sub-algorithms to solve each of the two problems. After one sub-algorithm has been executed, the current optimal solution is passed to the input of the other sub-algorithm.
The optimization of the discrete variable of the OFDM-NLFM waveform set is an unconstrained combinatorial optimization problem. Since the dimension of the discrete variable is very large, it is not feasible to find the global optimal solution accurately. Thus, the CD algorithm can be used to solve the unconstrained optimization problem by conducting approximate minimization along the coordinate direction or in the coordinate hyperplane [
47]. This paper optimizes the sub-chirp sequence code matrix and the sub-chirp rate PM code matrix based on the idea of the CD algorithm. The CD algorithm is fast in execution, but its optimization result could be unstable. To balance the efficiency and stability of the CD algorithm, the parameter matrices are divided into blocks according to the different transmit signals that they determine. Each row of the sub-chirp sequence code matrix and the sub-chirp rate PM code matrix corresponds to each transmit signal of the OFDM-NLFM waveform set. The BCD algorithm optimizes one row of the matrix with the other rows remaining unchanged. In this way, it can greatly reduce the computational complexity of the algorithm when optimizing the high dimensional matrices.
The intelligent optimization algorithm is a simple and direct solution for the continuous parameters of the OFDM-NLFM waveform set, considering that the objective function is very complex and difficult to be simplified and transformed. The PSO algorithm is a random search algorithm originally inspired by the foraging behavior of birds [
46]. It is found to be suitable for dealing with high-dimensional continuous variable optimization problems with the complex objective function. Hence, in this paper, the PSO algorithm is used to optimize NLFM parameters.
The overall block diagram of the proposed OFDM-NLFM waveform set design algorithm is shown in
Figure 3, and the specific steps of the algorithm are as follows.
Step 1: Select the initial NLFM parameters and construct M time-frequency curves according to Equation (1). The sub-chirp rate PM code matrix with MN elements is generated by random binary numbers. The sub-chirp sequence code matrix is generated by M random permutations. M is the number of transmit signals, and N is the number of the sub-chirp signals of each transmit signal.
Step 2: After the parameter initialization step, the time domain segmentation parameters in Equation (13) should be set properly.
Step 3: After the initial waveform has been obtained, keep the sub-chirp rate PM code matrix and sub-chirp sequence code matrix unchanged. Optimize NLFM parameters using the PSO algorithm and obtain the current optimal NLFM parameters. Before executing, the PSO algorithm, population size, the maximum number of iterations, and other parameters should be set properly.
Step 4: Keep the current optimal NLFM parameters obtained in step 3 unchanged, sub-chirp rate PM code matrix and sub-chirp sequence code matrix are optimized based on the BCD algorithm. Firstly, the matrix to be optimized is divided into several small blocks. Secondly, each block is optimized in turns with other blocks unchanged. Finally, the current optimal sub-chirp rate PM code matrix and sub-chirp sequence code matrix are obtained.
Step 5: If the difference between the objective function values before and after step 3 and step 4 is lower than the threshold, the optimization algorithm converges. Output the current optimal NLFM parameters, sub-chirp rate PM code matrix, and sub-chirp sequence code matrix. Otherwise, jump to step 3 and continue.
The specific implementation of the above steps is described in the following.
Section 3.1 introduces the initialization of OFDM-NLFM waveform set parameters.
Section 3.2 introduces the optimization of NLFM parameters based on the PSO algorithm.
Section 3.3 introduces the optimization of sub-chirp rate PM code matrix and sub-chirp sequence code matrix based on the BCD algorithm.
3.1. OFDM-NLFM Waveform Set Parameter Initialization
The principle of the parameter initialization for the OFDM-NLFM waveform set is to make the peak of the auto-correlation function sidelobe low. The NLFM parameters in Equation (12) affect the peak of the auto-correlation function sidelobe, measured by the PASR in Equation (21). In order to help select the appropriate initial values of the NLFM parameters, this section traverses the parameter value in a certain range to test its impact on the PASR. The PASR is also affected by the sub-chirp rate PM code matrix and sub-chirp sequence code matrix. Therefore, in this section, the PASR of M transmit signals is tested under several groups of random initialed sub-chirp rate PM code matrices and sub-chirp sequence code matrices. For each group of the matrices, the test goes through all the values of NLFM parameters with .
Set
M = 1 in Equations (12) and (21) because a bigger
M is not necessary when PCCR is not analyzed. The PASR under two different groups of sub-chirp rate PM code matrices and sub-chirp sequence code matrices are shown in
Figure 4 and
Figure 5. The results show that the PASR is lower when the NLFM parameters are in the interval of
. Thus, the initial values of NLFM parameters should be selected from the above interval. It can be seen from
Figure 4 and
Figure 5 that different sub-chirp rate PM code matrix and sub-chirp sequence code matrix have little impact on PASR. Therefore, we set the NLFM parameters to the random numbers in the above interval, and generate initial values of the sub-chirp rate PM code matrix and sub-chirp sequence code matrix using random numbers and permutations. The initial solution of the optimization
is defined as follows.
3.2. NLFM Parameters Optimization Based on PSO
Every time after initializing the OFDM-NLFM waveform set parameters or executing of one iteration of alternating optimization, the PSO algorithm is used to optimize the NLFM parameters. According to Equation (23), the number of NLFM parameters is 2
M. Thus the particle dimension is 2
M. The interval of the NLFM parameters is
. According to the global optimization model in Equation (25), the optimization model of the PSO algorithm can be expressed as
where
represents the objective function of the PSO.
and
represents the current optimal sub-chirp rate PM code matrix and sub-chirp sequence code matrix after the
k-th iteration of the alternating optimization. Note that they are the initial solution described in Equation (26) when
k = 0. Since the PSO algorithm is executed iteratively, the initial population of the
k+1-th iteration should include the optimal NLFM parameters in the
k-th iteration. Thus, the population of the
k+1-th iteration is initialized as
where
is the initial population.
q is the number of particles.
is assigned to the optimal solution after the
k-th execution of the PSO. The rest of the particles are initialized by random numbers. Initializing the population in this way ensures the current best solution to be passed to the next generation. After the
k+1-th PSO algorithm has been executed, the optimal solution can be expressed as
where
represents the
G-th generation population after the
k+1-th execution of PSO.
G is the maximum generation of the PSO algorithm.
is the optimal solution, and the current optimal NLFM parameter is
. Recall that before each execution of the PSO algorithm, the NLFM parameters, sub-chirp rate PM code matrix, and sub-chirp sequence code matrix are set to the current optimal values to ensure that the objective function decreases monotonously.
3.3. Sub-Chirp Rate PM and Sub-Chirp Sequence Code Matrix Optimization Based on BCD
After obtaining current optimal NLFM parameters by the PSO, the next step of the OFDM-NLFM waveform set design algorithm is to optimize the sub-chirp rate PM code matrix and sub-chirp sequence code matrix. This task is an unconstrained combinatorial optimization problem that can be expressed as
where
is the objective function of the BCD algorithm. The solution space of the optimization variables is a subspace of the solution space Ω. The goal of the above optimization is to obtain the optimal sub-chirp rate PM code matrix and sub-chirp sequence code matrix under the current optimal NLFM parameters. The optimal solutions are expressed as
,
which satisfy
where
and
are the optimal values of the sub-chirp rate PM code matrix and sub-chirp sequence code matrix after the
k-th execution of the sub-algorithm. These two current optimal values are also the initial solutions of the
k+1-th optimization. The optimization algorithm based on BCD proposed in this section is mainly built on the coordinate descent algorithm. The two code matrices are divided into smaller blocks. Each time we update one element or one block of the matrix, the values of the other elements are fixed. Considering that each row of the matrix corresponds to each transmit signal of the OFDM-NLFM waveform set, the above two matrices can be divided into
M blocks according to different rows.
The blocks of the above-mentioned matrix are divided as shown in Equations (32) and (33). In the BCD algorithm, one row of the matrix is updated with the rows corresponding to the other transmit signals unchanged. Each row of the matrices is updated alternately. Therefore, the optimization algorithm proposed in this paper is composed of the two layers of loops:
Outer loop: Update the m-th row vector of and of . The values of the other row vectors remain unchanged. The 1, 2,…, M-th row of the matrices are updated alternately until the objective function converges.
Inner loop: Update the n-th column of the row vector and , keeping the other columns unchanged. The 1, 2,…, N-th columns are updated alternately.
Therefore, the original high-dimensional optimization problem shown in Equation (30) is equal to solving several low-dimensional optimization problems. The optimization variable after the
i-th outer loop is
where
and
. The outer loop execution can also be summarized as a minimization problem as follows.
where
is the set of the binary row vectors whose length are
N.
GN is the set of all possible permutations of 1, 2, …,
N. The number of the element of the set is
N! and the optimization solution is expressed as
where
and
are the optimal row vectors,
and
is the optimal code matrices of the
i+1-th iteration. In the next iteration, we change the value of
m and optimize another row of the sub-chirp rate PM code matrix and sub-chirp sequence code matrix. The process of the outer loop can be summarized in Algorithm 1.
Algorithm 1: Block coordinate descent algorithm for X2 and X3 |
Input: Initial solution . |
Output: Optimal solution . |
Step 1: m = 1, i = 0, , . |
Step 2: Update the m-th row vector of and of . Input it to Algorithm 2 whose output is the optimal row vector and . |
Step 3: |
Step 4: i = i + 1. If is true, m = 1, otherwise m = m + 1. |
Step 5: Calculate the objective function using and . If the value of the objective function is not decreasing, and the optimal matrix is obtained, otherwise jump to Step 2 and continue. |
The inner loop calculates the optimal value of
,
. Similarly, denote
j be iteration counter, and the row vector after the
j+1-th iteration can be expressed as
When
j continues to increase until the value of the objective function converges, and the obtained optimal solution is
,
. In order to describe the update of the time-frequency code matrix, consisting of
M permutations, the exchange operation of the row vector or the permutation is defined as follows
where
represents the exchange of the
a-th and
b-th column of the row vector. According to the definition of the exchange operator, the update algorithm of the row vector can be summarized as in Algorithm 2.
Algorithm 2: The update algorithm for the m-th row of X2 and X3 |
Input: Current optimal row vectors , from Algorithm 1. |
Output: Optimal row vectors , . |
Step 1: n = 1, j = 0, , . |
Step 2: Exchange every columns with the n-th column of , and N new row vectors is obtained as . |
Step 3: Calculate N objective function using N new row vectors. Select the minimum objective function (if the l-th value is lowest) and its corresponding row vector is optimal solution expressed as . |
Step 4: The value of the n-th column of satisfies . Select the one making the objective function is the lowest and the optimal solution is . |
Step 5:j = j + 1. If is true, n = 1, otherwise n = n + 1. |
Step 6: Calculate the objective function using , . If the value of the objective function is not decreasing, the optimal row vectors , is obtained, otherwise jump to Step 2 and continue. |