A target parameter estimation method of an OTFS-ISAC system based on a deep unfolding network

By combining OTFS modulation and the deep unfolded network ADMM-Net, the problem of subcarrier orthogonality destruction in OFDM systems under high-speed scenarios is solved, achieving high-precision and efficient target parameter estimation, which is suitable for intelligent scenarios such as intelligent transportation and drone monitoring.

CN117938608BActive Publication Date: 2026-11-10FUZHOU UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410153422.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-02-04
Publication Date
2026-11-10
Estimated Expiration
2044-02-04

AI Technical Summary

Technical Problem

In high-speed scenarios, the Doppler effect in traditional OFDM systems can disrupt the orthogonality of subcarriers, leading to a decline in communication and sensing performance. Existing technologies struggle to achieve high-performance target parameter estimation in integrated sensing and communication systems.

Method used

OTFS modulation is used to modulate the data in the time-delay-Doppler domain, and combined with the deep unfolded network ADMM-Net, the target parameter estimation performance is improved through deep learning.

Benefits of technology

It improves the accuracy and efficiency of target parameter estimation, reduces computation time, is more adaptable, and can maintain high-performance communication and sensing capabilities in dynamic scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117938608B_ABST
    Figure CN117938608B_ABST
Patent Text Reader

Abstract

The application relates to a target parameter estimation method of an OTFS-ISAC system based on a deep unfolding network, and the method comprises the following steps: 1, a modulator maps information symbols in a time-delay-doppler domain to samples in a time-frequency domain by using ISFFT, and then creates a time-domain signal transmitted through a wireless channel through a Heisenberg transformation; 2, on a receiver, the time-domain signal is mapped to a time-frequency domain through a Wigner transformation, and then is mapped to a time-delay-doppler domain through SFFT for symbol demodulation; 3, OTFS matrixization and vectorization are carried out, so that subsequent problem establishment and network training are facilitated; 4, according to the sparsity of a radar channel in a time-delay-doppler domain, a target parameter problem is converted into a compressed sensing problem, and a target function is established; 5, a deep unfolding network ADMM-Net used for solving the CS problem is constructed and trained, and then demodulated symbols are input into the trained network, so that target parameters are estimated. The method is beneficial to improving the target parameter estimation performance of the ISAC system.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of communication technology, and specifically to a target parameter estimation method for an OTFS-ISAC system based on a deep unfolded network. Background Technology

[0002] The Internet of Vehicles (IoV) is a key tool for realizing intelligent transportation, providing a golden opportunity for its advancement. With the deep integration of new technologies such as 5G, edge computing, and integrated communication and sensing, the application of intelligent connected vehicles is becoming increasingly widespread, and my country's IoV industry has made significant progress in many areas. Observations reveal that the IoV primarily possesses the following characteristics: First, it ensures safe distances between vehicles, thereby reducing the risk of collisions. Second, through real-time navigation, the IoV helps drivers plan their trips more effectively, improving traffic efficiency. All these functions rely on integrated perception and communication technology as their foundation. Therefore, it can be determined that integrated perception and communication will be one of the important research directions in the future field of intelligent transportation.

[0003] For a long time, communication and sensing technologies have developed and evolved independently. Currently, communication and sensing networks are two separate networks. This means that in scenarios requiring both communication and sensing (such as autonomous driving and drone management), application systems need to access two independent networks to meet practical needs. This approach has two disadvantages: high cost and the inability to strictly synchronize information. Radar and communication, however, have exhibited a developmental pattern of mutual interweaving, inspiration, and borrowing. Currently, their technological development shows a high degree of consistency—pursuing wider bandwidth, larger arrays, greater focus on line-of-sight channel signal processing, and miniaturization, resulting in increasingly similar hardware architectures, channel characteristics, and signal processing. This provides a very suitable foundation for utilizing wireless communication infrastructure to achieve sensing, enabling future networks to transcend traditional communication and provide ubiquitous sensing services for measuring and even imaging the surrounding environment. As one of the key technologies potentially enabling 6G, ISAC can empower mobile cellular networks with new sensing capabilities such as speed measurement, ranging, positioning, target imaging, and recognition, meeting the new demands of intelligent scenarios such as smart transportation, drone monitoring, autonomous driving environmental perception, and robot interaction.

[0004] The mainstream approach of ISAC (Integrated Sensing and Communication Acoustics) is based on integrated waveform design using multi-carrier communication signals, primarily OFDM signals and modified OFDM-based signals. Besides communication signals, the application of multi-carrier signals in radar systems has also been extensively studied, promoting the development of integrated waveform design for sensing and communication. Since the integrated waveform based on OFDM signals carries communication information, the uncertainty of this information leads to randomness in the integrated waveform with each transmission, thus affecting the radar's ambiguity function performance. OFDM-based integrated sensing and communication schemes rely on the orthogonality of subcarriers at both the communication and radar ends. In high-speed scenarios, severe Doppler modulation can disrupt the orthogonality of subcarriers, affecting communication and radar target image reconstruction performance. This disruption of subcarrier orthogonality results in reconstructed target images with high sidelobes or high noise floor.

[0005] To address the challenge of high-performance sensing communication in dual-selective channels in high-speed scenarios, the novel OTFS modulation technique has been proposed in recent years. Unlike previous frequency-time domain-based modulation methods, OTFS modulation embeds modulation symbols in the time-delay-Doppler domain, allowing each symbol to traverse the entire frequency-time channel. This leverages the dual-selective characteristics of high-speed scenarios to achieve frequency-time diversity gain, thereby improving communication performance. Furthermore, in practice, the number of reflectors is finite, resulting in a finite number of multipath paths. The channel exhibits sparse characteristics in the time-delay-Doppler domain, making it more suitable than the frequency-time domain for representing dual-selective channels in high-speed scenarios. This can benefit target parameter estimation in high-speed environments. Additionally, OTFS modulation can be viewed as a two-dimensional precoding of OFDM modulation, allowing for integration with existing frequency-time domain-based transmit / receive systems. OTFS modulation holds promise as a crucial technology for high-performance wireless communication in future high-speed scenarios. On the other hand, with the development of artificial intelligence, intelligent sensing communication has become one of the future directions for ISAC (Intelligent Sensing Communication). The advantages of combining deep learning and parameter estimation are mainly reflected in the following aspects: (1) Improved parameter accuracy: Deep learning has a powerful data learning capability. By analyzing a large amount of data, a more accurate channel model can be established, thereby estimating the target parameters more accurately. (2) Enhanced adaptability of parameter estimation: Parameter estimation technology based on deep learning can extract and classify signals through deep learning networks. This adaptability makes the method of combining deep learning and parameter estimation have a wide range of application prospects in various scenarios. (3) Improved efficiency of parameter estimation: Parameter estimation technology based on deep learning can achieve fast training and real-time prediction through large-scale parallel computing and deep learning platforms, thereby improving the efficiency of parameter estimation. (4) Enhanced interpretability and transparency of parameter estimation: Deep learning models can provide interpretability and transparency, which makes it easier for people to understand the model's output and prediction results. Therefore, this invention utilizes the sparsity of the OTFS channel and the deep unfolded network to apply the deep unfolded network to ISAC, aiming to further improve the target parameter estimation performance of the ISAC system. Summary of the Invention

[0006] The purpose of this invention is to provide a target parameter estimation method for the OTFS-ISAC system based on deep unfolded networks, which is beneficial to improving the target parameter estimation performance of the ISAC system.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: a target parameter estimation method for an OTFS-ISAC system based on deep unfolded networks, characterized by comprising the following steps:

[0008] Step S1: The modulator first uses the inverse symplectic finite Fourier transform (ISFFT) to map the information symbol x[k,l] in the time-delay-Doppler domain to the sample X[n,m] in the time-frequency domain, and then uses the Heisenberg transform to create the time-domain signal s(t) transmitted through the wireless channel.

[0009] Step S2: On the receiver, the time-domain signal r(t) is mapped to the time-frequency domain through Wigner transform, and then mapped to the time-delay-Doppler domain using symplectic finite Fourier transform (SFFT) to obtain y[k,l] through symbol demodulation.

[0010] Step S3: Convert OTFS into matrices and vectors to facilitate subsequent problem setup and network training;

[0011] Step S4: Based on the sparsity of the radar channel in the time delay-Doppler domain, the target parameter problem is transformed into a compressed sensing problem, and the objective function is established.

[0012] Step S5: Construct and train the deep unfolded network ADMM-Net for solving the CS problem, and then input the demodulation symbol y[k,l] into the trained network to estimate the target parameters.

[0013] Furthermore, the specific implementation method of step S1 is as follows:

[0014] For OTFS modulation, consider a grid on the time-delay-Doppler plane. Modulate an M×N data symbol block {X DD [k,l], k=0,1,...,N-1, l=0,1,...,M-1}, where T is the symbol length and Δf is the subcarrier spacing; assuming the sampling interval is T / M and the total duration of the transmission frame is NT; the OTFS transmitter uses ISFFT to convert the time-delay-Doppler domain symbols to the time-frequency domain:

[0015]

[0016] Where n = 0, 1, 2, ..., N-1, m = 0, 1, 2, ..., M-1;

[0017] Then, in X FT Applying the Heisenberg transform, the time-domain transmission signal s(t) is generated:

[0018]

[0019] Among them, g tx (t) is the pulse shaping waveform in the transmitter.

[0020] Furthermore, the specific implementation method of step S2 is as follows:

[0021] By sampling the time-domain received signal r(t) in the dual-select fading channel, the discrete received signal in the time domain is represented as:

[0022]

[0023] Where n = 0, 1, 2, ..., MN-1, ω(n) is the additional noise, P is the number of propagation paths, and l p =τ p MΔf,k p =ν p NT is the delay and Doppler grid number on the delay-Doppler domain lattice Λ, h p τ p and ν p These represent the path gain, delay, and Doppler shift associated with the p-th path, respectively.

[0024] The received signal r(t) is converted into the received data symbol Y in the time-frequency domain through the Wigner transform. FT :

[0025]

[0026] Among them, g rx It is the pulse shaping waveform in the receiver;

[0027] Finally, regarding Y FT Perform a symplectic finite Fourier transform (SFFT) to generate the received symbol Y in the delayed Doppler domain. DD [kl]:

[0028]

[0029] Where l = 0, 1, 2, ..., M-1, k = 0, 1, 2, ..., N-1.

[0030] Furthermore, the specific implementation method of step S3 is as follows:

[0031] F M Defined as a Fast Fourier Transform matrix with M points, Defined as a discrete Fourier inverse transform matrix with N points; then the time-frequency domain symbols are written in matrix form:

[0032]

[0033] Among them, X DD ∈C M×N X FT ∈C M×N The time-domain transmission signal can then be written as:

[0034]

[0035] Among them, G tx =diag[g tx (0),g tx (T / M),...,g tx ((M-1)T / M)]∈C M×M Use a rectangular pulse waveform, i.e., G tx =I M , where I M It is an M-order identity matrix; then, the received signal in the time domain is represented as:

[0036] r=Hs+ω

[0037] in, It is the vectorized form of matrix S, where x is X. DD The vectorized form of H∈C MN×MN It is the time-delay-Doppler domain channel matrix:

[0038]

[0039] Where, Π∈R MN×MN For the permutation matrix:

[0040]

[0041] And Δ is an MN×MN diagonal matrix:

[0042] Δ=diag[z 0 ,z 1 ,...,z MN-1 ]

[0043] in, Matrices Π and Δ describe the time delay and Doppler effect, respectively; Y DD [k,l] is written as:

[0044]

[0045] Where R∈C M×N It constitutes the received signal sample r∈C MN×1 The time-frequency symbol is then used to obtain the received signal in the DD domain:

[0046]

[0047] in

[0048] Furthermore, the specific implementation method of step S4 is as follows:

[0049] Estimate the unknown parameters l of each path p from the noisy received signal y. p and kp With these estimates, the delay and Doppler shift of targets within the monitored area can be calculated, transforming the target parameter estimation problem into a compressed sensing problem. The LASSO problem is constructed as follows:

[0050]

[0051] Where λ is a weighting factor, α∈R I×1 (I≥MN) is a sparse vector, and A is an overcomplete dictionary; the l1 norm is introduced into the regularization term to obtain a sparse solution; the overcomplete dictionary A is denoted as:

[0052] A = [h'1, h'2, ... h'] I ]∈C MN×I

[0053] Where h' i Let A represent the i-th basis, i.e.:

[0054]

[0055] in, It is the channel matrix of the i-th path; when considering an open space propagation environment with few moving scatterers, the DD domain channel exhibits a sparse response; the target parameters are obtained by solving for the non-zero components of α, which contains delay and Doppler information in the channel.

[0056] To avoid coupling interference, constraints are introduced to obtain:

[0057]

[0058] stα-z=0.

[0059] Furthermore, the specific implementation method of step S5 is as follows:

[0060] The iterative process is expanded into a deep neural network consisting of a fixed number of stages to construct ADMM-Net. The structure of ADMM-Net has four modules: reconstruction module, relaxation module, nonlinear transformation module, and scaling parameter module.

[0061] 1) Reconstruction Module: The reconstruction module updates α; network input y∈C MN×1 and from the k-1 stage and As input to the refactoring module; the output of the refactoring module in stage k is:

[0062]

[0063] in, and These are learnable parameters; the function p(·) creates a real-valued vector by vertically concatenating the real and imaginary components of the input.

[0064]

[0065] all and This represents the vertical connection between the real and imaginary parts of α, ξ, z, and u; the input values ​​of the reconstruction module in stage 1 are respectively and

[0066] 2) Relaxation Module: Relaxation Module Ξ k The output is:

[0067]

[0068] Where, μ k It is a learnable parameter used to speed up the convergence process;

[0069] 3) Nonlinear Transformation Module: Nonlinear Transformation Module Z k The operation in this context is equivalent to the soft thresholding process in the ADMM algorithm; the output of the soft thresholding operation is:

[0070]

[0071] Where, λ k It is a learnable L1 regularization parameter;

[0072] 4) Parameter Update Module: Parameter Update Module U k The output is:

[0073]

[0074] Where, η k It is a learnable parameter corresponding to the gradient step size; the network output comes from the output of the last stage.

[0075]

[0076] The k-stage learning of the corresponding network requires training the following parameters.

[0077] Compared with existing technologies, this invention has the following advantages: This invention proposes a target parameter estimation method for an OTFS-ISAC system based on a deep unfolded network. To address the problem that traditional OFDM systems are sensitive to the Doppler effect and, under dynamic conditions, disrupt the orthogonality between subcarriers, leading to a decline in communication and sensing performance, this invention uses OTFS signals instead of OFDM signals. This is because OTFS signals modulate data in the time-delay-Doppler domain, allowing each modulation symbol to traverse the entire frequency-time channel. This leverages the dual-selection characteristics of dynamic scenarios to obtain frequency-time diversity gain, thereby improving communication performance. Furthermore, the time-delay-Doppler domain is more suitable than the frequency-time domain for representing dual-selection channels in dynamic scenarios, which benefits parameter estimation in dynamic situations. Moreover, to improve estimation accuracy and reduce computation time, a deep unfolded network, ADMM-Net, is proposed. This network maps the iterations of the traditional iterative algorithm ADMM to several stages of a deep unfolded network with learnable parameters, thus exhibiting stronger adaptability and lower computation time. Attached Figure Description

[0078] Figure 1 This is a diagram of the deep unfolded network structure for target parameter estimation in the OTFS-ISAC system in this embodiment of the invention.

[0079] Figure 2 This is a performance comparison between ADMM and ADMM-Net using different numbers of stages / iterations in the embodiments of the present invention.

[0080] Figure 3 This is a performance comparison of different algorithms under different signal-to-noise ratio conditions in the embodiments of this invention.

[0081] Figure 4 This is a performance comparison of different algorithms with different numbers of targets when the signal-to-noise ratio is 15dB in this embodiment of the invention.

[0082] Figure 5 In the embodiments of this invention, different algorithms have different time delays (when k p =7 o'clock) and different Doppler (when l p Performance comparison at 7 (=7).

[0083] Figure 6 These are images restored using different algorithms in the embodiments of this invention.

[0084] Figure 7 This is a comparison of the running time under different algorithms in the embodiments of the present invention. Detailed Implementation

[0085] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0086] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0087] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0088] This embodiment provides a target parameter estimation method for an Orthogonal Time Frequency Space-Integrated Sensing and Communications (OTFS-ISAC) system based on deep unfolded networks, characterized by the following steps:

[0089] Step S1: The modulator first uses the Inverse Symplectic Fast Fourier Transform (ISFFT) to map the information symbol x[k,l] in the time-delay-Doppler domain to the sample X[n,m] in the time-frequency domain, and then uses the Heisenberg Transform to create the time-domain signal s(t) transmitted through the wireless channel.

[0090] Step S2: At the receiver, the time-domain signal r(t) is mapped to the time-frequency domain through Wigner transform, and then mapped to the time-delay-Doppler domain using symplectic finite Fourier transform (SFFT) for symbol demodulation to obtain y[k,l].

[0091] Step S3: Convert the OTFS matrix and vectorize it to facilitate subsequent problem setup and network training.

[0092] Step S4: Based on the sparsity of the radar channel in the time delay-Doppler domain, the target parameter problem is transformed into a compressed sensing (CS) problem, and the objective function is established.

[0093] Step S5: Construct and train the deep unfolding network ADMM-Net (Alternating Direction Method of Multipliers Network) for solving the CS problem, and then input the demodulation symbol y[k,l] into the trained network to estimate the target parameters.

[0094] In this embodiment, the specific implementation method of step S1 is as follows:

[0095] For OTFS modulation, consider a grid on the time-delay-Doppler plane. Modulate an M×N data symbol block {X DD [k,l], k=0,1,...,N-1, l=0,1,...,M-1}, where T is the symbol length and Δf is the subcarrier spacing; assuming the sampling interval is T / M and the total duration of the transmission frame is NT; the OTFS transmitter uses ISFFT to convert the time-delay-Doppler domain symbols to the time-frequency domain:

[0096]

[0097] Where n = 0, 1, 2, ..., N-1, m = 0, 1, 2, ..., M-1.

[0098] Then, in X FT Applying the Heisenberg transform, the time-domain transmission signal s(t) is generated:

[0099]

[0100] Among them, g tx (t) is the pulse shaping waveform in the transmitter. Non-rectangular pulses cause non-orthogonality, thus reducing bit error rate performance. Therefore, we consider using rectangular pulses.

[0101] In this embodiment, the specific implementation method of step S2 is as follows:

[0102] By sampling the time-domain received signal r(t) in a dual-select fading channel, the discrete received signal in the time domain can be expressed as:

[0103]

[0104] Where n = 0, 1, 2, ..., MN-1, ω(n) is the additional noise, P is the number of propagation paths, and l p =τ p MΔf,k p =ν p NT is the delay and Doppler grid number on the delay-Doppler domain lattice Λ, h p τ p and ν p These represent the path gain, delay, and Doppler shift associated with the p-th path, respectively.

[0105] The received signal r(t) is converted into the received data symbol Y in the time-frequency domain through the Wigner transform. FT :

[0106]

[0107] Among them, g rx It is the pulse shaping waveform in the receiver.

[0108] Finally, regarding Y FT Perform a Symptotic Fast Fourier Transform (SFFT) to generate the received symbol Y in the delayed Doppler domain. DD [kl]:

[0109]

[0110] Where l = 0, 1, 2, ..., M-1, k = 0, 1, 2, ..., N-1.

[0111] In this embodiment, the specific implementation method of step S3 is as follows:

[0112] F M Defined as a Fast Fourier Transform (FFT) matrix with M points, Defined as an N-point Inverse Discrete Fourier Transform (IDFT) matrix; then the time-frequency domain symbols can be written in matrix form:

[0113]

[0114] Among them, X DD ∈C M×N X FT ∈C M×N The time-domain transmission signal can then be written as:

[0115]

[0116] Among them, G tx =diag[g tx (0),g tx (T / M),...,g tx ((M-1)T / M)]∈C M×M Here, a rectangular pulse waveform, namely G, is used. tx =I M , where I M It is an M-order identity matrix; then, the received signal in the time domain can be represented as:

[0117] r=Hs+ω

[0118] in, It is the vectorized form of matrix S, where x is X. DD The vectorized form of H∈C MN×MN It is the time-delay-Doppler domain channel matrix:

[0119]

[0120] Where, Π∈R MN×MN For the permutation matrix (positive cyclic shift):

[0121]

[0122] And Δ is an MN×MN diagonal matrix:

[0123] Δ=diag[z 0 ,z 1 ,...,z MN-1 ]

[0124] in, Matrices Π and Δ describe the time delay and Doppler effect, respectively; Y DD [k,l] can be written as:

[0125]

[0126] Where R∈C M×N It constitutes the received signal sample r∈C MN×1 The time-frequency symbol; then the received signal in the DD (Delay-Doppler) domain can be obtained:

[0127]

[0128] in

[0129] In this embodiment, the specific implementation method of step S4 is as follows:

[0130] The purpose of this step is to estimate the unknown parameters l of each path p from the noisy received signal y. p and k p With these estimates, the delay and Doppler shift of targets within the monitored area can be calculated. We transform the target parameter estimation problem into a compressed sensing problem. The LASSO (Least Absolute Shrinkage and Selection Operator) problem is constructed as follows:

[0131]

[0132] Where λ is a weighting factor, α∈R I×1(I≥MN) is a sparse vector, and A is an overcomplete dictionary; we introduce the l1 norm in the regularization term to obtain a sparse solution; the overcomplete dictionary A can be denoted as:

[0133] A = [h'1, h'2, ... h'] I ]∈C MN×I

[0134] Where h' i Let A represent the i-th basis, i.e.:

[0135]

[0136] in, It is the channel matrix of the i-th path; when considering an open space propagation environment with few moving scatterers, the DD domain channel exhibits a sparse response; the target parameters can be obtained by solving the non-zero components of Δ, which contains delay and Doppler information in the channel.

[0137] To avoid coupling interference, we introduce constraints to obtain:

[0138]

[0139] stα-z=0.

[0140] In this embodiment, the specific implementation method of step S5 is as follows:

[0141] The idea behind this algorithm is to expand the iterative process into a deep neural network consisting of a fixed number of stages, thus constructing ADMM-Net. Figure 1 This is the structure of the ADMM-Net proposed in this embodiment. Each module corresponds to an iteration, and an arrow indicates the functional interdependence of the iterations. The ADMM-Net structure has four types of modules: reconstruction module, relaxation module, nonlinear transformation module, and scaling parameter module.

[0142] 1) Reconstruction Module: The reconstruction module updates α; network input y∈C MN×1 and from the k-1 stage and As input to the refactoring module; the output of the refactoring module in stage k is:

[0143]

[0144] in, and These are learnable parameters; the function p(·) creates a real-valued vector by vertically concatenating the real and imaginary components of the input.

[0145]

[0146] all and This represents the vertical connection between the real and imaginary parts of α, ξ, z, and u; the input values ​​of the reconstruction module in stage 1 are respectively and

[0147] 2) Relaxation Module: Relaxation Module Ξ k The output is:

[0148]

[0149] Where, μ k It is a learnable parameter that can speed up the convergence process.

[0150] 3) Nonlinear Transformation Module: Nonlinear Transformation Module Z k The operation in this context is equivalent to the soft thresholding process in the ADMM algorithm; the output of the soft thresholding operation is:

[0151]

[0152] Where, λ k It is a learnable L1 regularization parameter.

[0153] 4) Parameter Update Module: Parameter Update Module U k The output is:

[0154]

[0155] Where, η k It is a learnable parameter corresponding to the gradient step size; the network output comes from the output of the last stage.

[0156]

[0157] The k-stage learning of the corresponding network requires training the following parameters. The value of η is initialized based on theoretically or empirically proven values, and all parameters are learnable, which greatly improves the adaptability of the network.

[0158] The method of this invention improves estimation accuracy and computation time. Compared with other parameter estimation methods, it is found that this scheme can achieve the best real-time performance while maintaining the highest estimation accuracy.

[0159] Figure 2The comparison of NMSE between ADMM-Net and ADMM with different network layers / iteration numbers is shown when SNR = 15dB and the target number is 2. Clearly, the NMSE of a 4-level network is -7.21dB. However, over 100 iterations, the average convergence NMSE of ADMM is only -4.52dB. When the number of network layers exceeds 20, the NMSE of ADMM-Net changes slowly.

[0160] Figure 3 This section compares the performance of different algorithms under varying signal-to-noise ratios (SNRs). Specifically, it compares the NMSE performance of several methods with a target number of 2 at different SNRs. Figure 3 In the results, ADMM-Net exhibits significantly lower NMSE compared to other algorithms under various conditions. At SNR = 14 dB, ADMM-Net outperforms CVX, OMP, and matched filters by approximately 9 dB, and outperforms ADMM and ISTA by approximately 4 dB. The NMSE of almost all methods decreases with increasing SNR, while ADMM-Net's performance remains stable. These results demonstrate that ADMM-Net possesses strong robustness under diverse noise conditions.

[0161] Figure 4 This section compares the performance of several methods with different numbers of targets at a signal-to-noise ratio of 15 dB. Figure 4 In the simulations, the NMSE of almost all methods increases with the number of targets. However, ADMM-Net consistently outperforms other algorithms. This is because errors in the estimation of each target accumulate as the number of targets increases, leading to performance degradation. In simulations, ADMM-Net outperforms ADMM and ISTA by an average of approximately 2.4 dB, and OMP and CVX by an average of approximately 6 dB. Furthermore, except in the case of a single target, the matching filter consistently performs the worst.

[0162] Figure 5 The effects of time delay and Doppler shift on the NMSE of a single target are shown. Figure 5 In the NMSE comparison, ADMM, ISTA, and OMP have relatively similar results. ADMM-Net's NMSE is lower than that of CVX and matched filter algorithms with arbitrary delays and Doppler shifts. In terms of NMSE performance, ADMM-Net outperforms ADMM, ISTA, and OMP by approximately 6 dB overall. When l and k increase, ADMM-Net's NMSE fluctuations are similar to those of the other algorithms.

[0163] Figure 6 Different algorithms were used for image processing after restoration. Figure 6In the results, we can see that the reconstructed image from ADMM-Net is very close to the true value, clearer and more accurate than ADMM, while the reconstructed image from ADMM contains a lot of low-amplitude background noise. This indicates that ADMM-Net can achieve better estimation accuracy than ADMM.

[0164] Figure 7 The table shows the single-run time of each algorithm. We can see that ADMM-Net outperforms ADMM and ISTA by about 4dB, and OMP and CVX by about 9dB. ADMM-Net's runtime is about 3% less than CVX and OMP, and about 25% less than ADMM and ISTA. This is because iterative algorithms require a large number of iterations to converge, leading to high computational complexity, while deep unfolding-based algorithms significantly reduce computation time by approximating iterative algorithms with deep neural networks.

[0165] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0166] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0167] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0168] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0169] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for estimating target parameters in an OTFS-ISAC system based on deep unfolded networks, characterized in that, Includes the following steps: Step S1: The modulator first uses the inverse symplectic finite Fourier transform (ISFFT) to extract the information symbols in the time-delay-Doppler domain. Samples mapped to the time-frequency domain Then, a time-domain signal for transmission over a wireless channel is created using the Heisenberg transform. ; Step S2: On the receiver, the time domain signal The signal is obtained by mapping to the time-frequency domain using the Wigner transform, and then mapping to the time-delay-Doppler domain using the symplectic finite Fourier transform (SFFT) for symbol demodulation. ; Step S3: Convert OTFS into a matrix and then into a vector. Step S4: Based on the sparsity of the radar channel in the time delay-Doppler domain, the target parameter problem is transformed into a compressed sensing problem, and the objective function is established. Step S5: Construct and train the deep unfolded network ADMM-Net for solving the CS problem, and then demodulate the symbols. Input the trained network into the network to estimate the target parameters; The specific implementation method of step S4 is as follows: From the noisy received signal Mid-estimation of each path Unknown parameters and ; Calculate the delay and Doppler shift of targets within the monitored area; transform the target parameter estimation problem into a compressed sensing problem. + The LASSO problem is constructed as follows: in, It is a weighting factor. It is a sparse vector. , It is an overcomplete dictionary; introduced in the regularization terms. Norms are used to obtain sparse solutions; overly complete dictionaries Recorded as: in express The One base, namely: in, This is the channel matrix for the i-th path; when considering an open-space propagation environment with few moving scatterers, the DD domain channel exhibits a sparse response; this is achieved by solving for the delay and Doppler information contained in the channel. The non-zero components are used to obtain the target parameters; To avoid coupling interference, constraints are introduced to obtain: s.t. In step S5, the iterative process is expanded into a deep neural network consisting of a fixed number of stages to construct ADMM-Net. The structure of ADMM-Net includes four modules: reconstruction module, relaxation module, nonlinear transformation module, and parameter update module.

2. The target parameter estimation method for the OTFS-ISAC system based on deep unfolded networks according to claim 1, characterized in that, The specific implementation method of step S1 is as follows: For OTFS modulation, consider a grid on the time-delay-Doppler plane. Modulate one Data symbol blocks ,in For the symbol length, Let the subcarrier spacing be ; assuming the sampling interval is . The total duration of the transmission frame is The OTFS transmitter uses ISFFT to convert the time-delay-Doppler domain symbols to the time-frequency domain: in, ; Then, in Applying the Heisenberg transform, a time-domain transmission signal is generated. : in, It is the pulse shaping waveform in the transmitter.

3. The target parameter estimation method for the OTFS-ISAC system based on deep unfolded networks according to claim 2, characterized in that, The specific implementation method of step S2 is as follows: By analyzing the time-domain received signal in the dual-select fading channel Sampling is performed, and the discrete received signal in the time domain is represented as: in, , To add noise, For the number of propagation paths, , It is a time-delay-Doppler domain lattice The delay and Doppler grid number on the surface, , and Respectively represent the first Path gain, delay, and Doppler shift associated with each path; Received signal It is converted into received data symbols in the time-frequency domain through Wigner transform. : in, It is the pulse shaping waveform in the receiver; Finally, for Perform a symplectic finite Fourier transform (SFFT) to generate the received symbol in the delayed Doppler domain. : in, .

4. The target parameter estimation method for the OTFS-ISAC system based on deep unfolded networks according to claim 3, characterized in that, The specific implementation method of step S3 is as follows: Will Defined as having The fast Fourier transform matrix of each point will Defined as having The discrete Fourier inverse transform matrix of each point; then the time-frequency domain symbols are written in matrix form: in, , The time-domain transmission signal can then be written as: in, Use a rectangular pulse waveform, i.e. ,in yes An identity matrix of order 1; then, the received signal in the time domain is represented as: in, It is a matrix The vectorized form, yes The vectorized form; It is the time-delay-Doppler domain channel matrix: in, For the permutation matrix: and yes Diagonal matrix: in, ;matrix and The time delay and Doppler effect are described respectively; Written as: in, It constitutes the received signal sample The time-frequency symbol is then used to obtain the received signal in the DD domain: in .

5. The target parameter estimation method for the OTFS-ISAC system based on deep unfolded networks according to claim 1, characterized in that, The specific implementation of step S5 includes: 1) Refactoring Module: The refactoring module... Update; network input and from the k-1 stage and As input to the refactoring module; the output of the refactoring module in stage k is: in, , and These are learnable parameters, functions. A real-valued vector is created by vertically concatenating the real and imaginary components of the input: all , , and express , and and The vertical connection between the real and imaginary parts; the input values ​​of the reconstruction module in stage 1 are respectively and ; 2) Relaxation Module: The output of the relaxation module is: in, It is a learnable parameter used to speed up the convergence process; 3) Nonlinear Transformation Module: Nonlinear Transformation Module The operation in this context is equivalent to the soft thresholding process in the ADMM algorithm; the output of the soft thresholding operation is: in, It is a learnable resource. Regularization parameters; 4) Parameter Update Module: The output of the parameter update module is: in, It is a learnable parameter corresponding to the gradient step size; the network output comes from the output of the last stage. : The k-th stage of learning requires training the following parameters. .