A Deep Learning-Based Channel Estimation Method for AmBC Systems

Abstract

This invention discloses a channel estimation method for an AmBC system based on deep learning, comprising the following steps: S1, building an AmBC communication system according to design requirements, setting specific inputs, and building a simulation platform; S2, at the transmitting end, the frequency domain transmission symbol with pilots is first modulated by IDFT, then inserted with a cyclic prefix and a parallel-to-serial transform before being transmitted; S3, the time domain signal received at the receiving end is first subjected to a serial-to-parallel transform and cyclic prefix removal, and then demodulated by DFT to obtain the frequency domain symbol; S4, performing LS estimation on the CSI at the pilot; the LS estimation value, the transmitted and received pilot symbols are fed together into a k-layer neural network to obtain the optimized CSI; S5, performing high-resolution reconstruction on the CSI; S6, multiple LSTM units are horizontally cascaded to form a single LSTM network, and then two LSTM networks are vertically cascaded together; S7, the predicted value H is... LSTM (n) is updated. This invention improves transmission efficiency while ensuring transmission quality under high signal-to-noise ratio conditions.

Description

Technical Field

[0001] This invention relates to the field of channel estimation, and in particular to a channel estimation method for AmBC systems based on deep learning. Background Technology

[0002] The Internet of Things (IoT) has been developing for nearly 30 years and has remained a hot topic. Currently, countries around the world have formulated their own unique policies and plans, actively promoting and applying IoT technology in fields such as industry, agriculture, and healthcare. The 3GPP (3rd Generation Partnership Project) adopted the Narrowband Internet of Things (NB-IoT) technical standard in Release 13, marking NB-IoT's entry into the commercial stage. While current cellular networks and mobile internet enable convenient and fast communication between people, the concept of the Internet of Everything has not yet been truly realized. The IoT still requires more technological support and further research. Backscattering technology was proposed by Stockman in 1948, subsequently leading to corresponding Radio Frequency Identification (RFID) products and applications. Because RFID technology can automatically identify radio frequency tags and has low power consumption, it has become an important technology in the Internet of Things (IoT). The basic principle of RFID is that a reader sends a dedicated radio frequency signal to the RFID tag. Upon receiving the signal, the RFID tag loads its own information onto the signal and reflects it back to the reader. The reader then processes the signal to obtain the information sent by the RFID tag. However, with the development of IoT technology, RFID technology also faces some challenges, such as double path loss, short communication distance, and self-interference from the reader's radio frequency source signal at the reader's location, affecting signal transmission performance. To address these issues, several novel backscattering technologies have been proposed, including Bistatic Backscatter Communication (BiBC) and Ambient Backscatter Communication (AmBC).

[0003] In BiBC technology, the transmitter and receiver of the reader are separated and no longer integrated into a single device. The transmitter generates a carrier wave and sends it to a nearby reflecting device (such as a sensor or RFID tag). The reflecting device then loads its own information onto the carrier wave and reflects it back to the receiver. This technology can improve the transmission distance of backscatter communication systems and has a wider range of applications.

[0004] Compared to BiBC technology, AmBC technology uses ambient wireless signals as both a power source and a carrier source to transmit its information, eliminating the need for a dedicated carrier transmitter and resulting in lower power consumption. Furthermore, AmBC technology utilizes ambient wireless signals for information transmission, requiring no additional spectrum and alleviating the current scarcity of spectrum resources.

[0005] However, AmBC technology also has some problems. Since the reflector reflects ambient wireless signals, the reflected signal at the receiver is subject to strong interference from these signals, making it difficult for the receiver to recover the reflector's symbol. In many applications, the number of reflectors is very large, and existing multiple access technologies, such as Frequency Division Multiple Access (FDMA) and Time Division Multiple Access (TDMA), are insufficient to meet the multiple access requirements of AmBC technology. Furthermore, the circuit design of reflectors is relatively simple and cannot handle complex signal and information processing, which can lead to security issues.

[0006] In AmBC systems, the average power of the reflected link signal is several orders of magnitude lower than that of the direct link signal. Therefore, the reflected link can be treated as interference and demodulated to extract the RF source symbol. Then, the direct link signal can be recovered by subtracting this signal from the received signal. Finally, the remaining signal is processed to recover the reflected device symbol. To recover the reflected device symbol, channel estimation of the direct link must be performed first.

[0007] With the development of the Internet of Vehicles (IoV), the increasing number of high-speed mobile scenarios places increasingly higher demands on wireless communication systems. High-speed movement and multipath transmission can lead to time-selective and frequency-selective fading in the channel. Estimating Channel State Information (CSI) from the affected received signal is one of the crucial steps for the receiver to recover transmitted symbols. Traditional pilot-based channel estimation methods require inserting numerous pilot symbols into the transmitted symbols when dealing with rapidly changing channels, resulting in reduced system transmission efficiency. To reduce pilot overhead, channel prediction techniques can be used to predict continuously changing CSI, but the commonly used autoregressive (AR) linear prediction method has insufficient prediction performance. In recent years, deep learning (DL) technology, with its powerful data processing capabilities, has brought entirely new design methods to communication systems. Therefore, by analyzing the existing research results of deep learning in channel estimation and prediction, and following the model-driven approach, we use deep learning technology to solve some problems in traditional channel estimation and prediction methods. We have conducted an in-depth study on the channel estimation and prediction method of Orthogonal Frequency Division Multiplexing (OFDM) system under Rayleigh dual-select fading channel. Summary of the Invention

[0008] To address the existing problems, this invention provides a channel estimation method for AmBC systems based on deep learning, the specific scheme of which is as follows:

[0009] A deep learning-based channel estimation method for AmBC systems includes the following steps:

[0010] S1. According to the design requirements, build an AmBC communication system other than the neural network using the Matlab R2023a simulation platform, set specific inputs including the number of subcarriers N, channel length L, pilot number P, and non-zero number T experimental conditions, and build the simulation platform.

[0011] S2, at the transmitting end, the frequency domain transmission symbol with pilot is first modulated by IDFT, and then transmitted after inserting a cyclic prefix and parallel-to-serial conversion;

[0012] S3, after passing through the Rayleigh dual-select fading channel, the time-domain signal received by the receiver first undergoes serial-to-parallel conversion and cyclic prefix removal, and then DFT demodulation to obtain the frequency-domain symbol.

[0013] S4, perform LS estimation on the channel state information (CSI) at the pilot; feed the LS estimate, the transmitted and received pilot symbols together into the k-layer neural network to obtain the optimized CSI;

[0014] S5 introduces a two-dimensional CNN to extract the correlation information of the channel along the time and frequency axes, and performs high-resolution reconstruction of CSI;

[0015] S6, multiple LSTM units are cascaded horizontally to form an LSTM network, and then two LSTM networks are cascaded vertically together. The LSTM units in the same layer have the same parameters. Finally, the fully connected layer is used to extract features and output the results. The number of units in each LSTM layer represents the memory length of the network, determines the temporal correlation of the network, and the number of units in each layer is at least the number of pilot OFDM symbols Sp.

[0016] S7, firstly, ZF equalization is used to bring the channel prediction value closer to its theoretical LS estimate, and then a K-layer neural network is used to bring the channel prediction value closer to its actual value, thereby eliminating errors within a certain range; after the LSTM network completes the prediction process of the CSI of the nth data OFDM symbol, the predicted value H is... LSTM (n) is updated.

[0017] Preferably, the time-domain signal received in step S3 , can be represented as Where, ∗ represents convolution, N is the number of subcarriers, x(n) and h(n) are the time-domain symbols transmitted on the nth subcarrier and the time-domain channel response that passes through, and w(n) is Gaussian white noise with a mean of 0 and a variance of δ2.

[0018] Corresponding frequency domain received symbols , can be represented as , where Y(k), H(k), X(k) and W(k) are the DFT results of y(n), h(n), x(n) and w(n) respectively;

[0019] For a dual-select fading channel, pilots are placed at the beginning of each transmitted symbol block of length Sb, where the number of pilot OFDM symbols is Sp and the number of pure data OFDM symbols is Sd = Sb − Sp.

[0020] Preferably, step S4 specifically involves: when the number of subcarriers is N, receiving a frequency domain OFDM symbol. and the corresponding transmit frequency domain OFDM symbols First, it passes through the LS channel estimator to obtain... The input is fed into a K-layer neural network, and after optimization by the K-layer neural network, a real-valued CSI estimate with alternating real and imaginary parts is obtained. The CSI estimates obtained by taking different values ​​of K are compared, and the channel estimation performance varies under different channel conditions. Under the conditions of signal-to-noise ratio gradients of {5, 10, 15, 20} dB and a maximum Doppler frequency of {583} Hz, 10,000 sets of training symbols are randomly generated for each condition to form a training dataset, which is used to train a k-layer neural network. One of the training samples can be represented as... It is input; It is a label that indicates the actual CSI of the current OFDM symbol, where the real and imaginary parts are arranged alternately.

[0021] Preferably, step S5 specifically involves: two identical CNNs, one for the real part and one for the imaginary part, being used for channel interpolation. Their inputs are two TF grids composed of the real and imaginary parts of Sp pilot OFDM symbols, including the CSI estimates at the pilot subcarriers obtained by the channel estimation network in S4 and the zero values ​​at other data locations. After convolution processing by the two CNNs, two complete TF grids HpCNN, each containing the real and imaginary parts of the CSI at all locations, are obtained. The two CNNs are trained offline and then deployed in the OFDM system to implement the interpolation function.

[0022] Preferably, step S6 specifically involves: assuming the number of subcarriers is N, the number of LSTM units in each layer is d, and the current time is n, then the input to the first layer LSTM is the CSI of each OFDM symbol containing all subcarriers from the previous 1 time to the previous d time: The input to the second LSTM layer is the output of the first LSTM layer, which is the short-term memory of the first LSTM layer from the previous 1 time step to the previous d time steps. Where L 1 This indicates the number of neurons in the first LSTM layer. The input to the fully connected layer is the short-term memory output from the last unit of the second LSTM layer. Where L 2 This indicates the number of neurons in the second LSTM layer, and the output is the predicted CSI value at the current time step. The prediction process can be briefly represented as follows: Where P represents the prediction process of the LSTM network; the LSTM network is first trained offline and then deployed in the system to realize the prediction function.

[0023] Preferably, step S7, which updates the predicted value HLSTM(n), includes the following five steps:

[0024] S71, recover the nth OFDM transmitted symbol through ZF equalization, i.e.

[0025] S72 performs a hard decision on the recovery of soft symbols. ;

[0026] S73, perform LS estimation on the hard symbol, i.e. ;

[0027] S74, the optimized CSI is obtained through the K-layer neural network proposed in S3: H DNN (n);

[0028] S75, H LSTM The value of (n) is determined by H DNN (n) substitution.

[0029] The present invention also discloses a computer-readable storage medium storing a computer program, which, when executed, performs the method described in any of the above-mentioned embodiments.

[0030] The present invention also discloses a computer system including a processor, a storage medium storing a computer program, and the processor reading from the storage medium and running the computer program to perform the method described in any of the preceding claims.

[0031] The beneficial effects of this invention are as follows:

[0032] (1) The threshold iteration algorithm with strong curve fitting capability of this invention is expanded into a K-layer neural network and used to optimize the LS estimation value, which solves the problems of noise affecting the LS method and the need for channel prior information in the LMMSE method. CNN with strong two-dimensional image processing capability is used for channel interpolation, which solves the problems of edge blurring and unclear details in spline interpolation. LSTM network with strong time series processing capability is used for pilot-free channel prediction, and data decision feedback mechanism is used to reduce the impact of error propagation. The structure of this channel estimation and prediction network is very flexible. The three networks can be trained independently or replaced by other more advanced networks to obtain better performance. Simulation results show that the channel estimation and prediction network proposed in this invention solves the problem of poor performance of traditional channel estimation and prediction methods, and improves transmission efficiency while ensuring transmission quality under high signal-to-noise ratio conditions.

[0033] (2) This invention maps the parameters of the communication system to the elements of reinforcement learning to build a reinforcement learning system. Then, for the state acquisition step, it proposes a noise intensity estimation method based on a threshold iteration algorithm unfolded into a K-layer neural network and a future channel state information prediction method based on multi-step LSTM. Finally, it introduces the BDQN algorithm to divide the action into simple sub-actions in multiple dimensions, solving the problem of a large number of actions, and learning how to adjust the pilot scheme according to the current system state. Simulation results prove the effectiveness of the noise intensity estimation method based on a threshold iteration algorithm unfolded into a K-layer neural network and the future channel state information prediction method based on multi-step LSTM proposed in this invention. It shows that the adaptive pilot position and spacing adjustment mechanism based on BDQN can focus on transmission quality or transmission efficiency, or it can be a compromise between the two. It also shows that the joint OFDM system based on deep learning channel estimation and prediction network and BDQN can improve transmission efficiency while ensuring transmission quality. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 It is a K-layer neural network model;

[0036] Figure 2 It is a convolutional neural network model;

[0037] Figure 3 A schematic diagram of an LSTM unit;

[0038] Figure 4 BDQN model flowchart;

[0039] Figure 5 Here is a diagram of the LSTM network structure;

[0040] Figure 6 This is a flowchart of the present invention;

[0041] Figure 7 This is a diagram of a reflection model. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] To address the poor performance of traditional channel estimation and prediction methods, this invention designs a deep learning-based channel estimation and prediction network. It consists of a K-layer neural network using a threshold iteration method for strong curve fitting capabilities, a two-dimensional convolutional neural network (CNN) for strong image reconstruction capabilities, and a Long Short-Term Memory (LSTM) network for strong time series processing capabilities. These three neural networks are used to optimize traditional LS channel estimation results, channel interpolation, and channel prediction, respectively. Furthermore, this invention introduces a data decision feedback mechanism to reduce the impact of error propagation. Simulation results show that all three neural networks outperform their respective traditional methods, and the proposed channel estimation and prediction network improves the system's transmission quality and efficiency.

[0044] Building upon existing research in channel estimation and prediction networks, this invention addresses the limitations of fixed pilot modes by proposing an adaptive adjustment mechanism for pilot position and spacing based on Branching Dueling Q-Network (BDQN). This mechanism, combined with a deep learning channel estimation and prediction network, further explores the possibility of reducing pilot overhead. In the BDQN algorithm, the environment state consists of the current noise intensity obtained with the assistance of a K-layer neural network using a threshold iteration method and the future CSI predicted by a multi-step LSTM network. The agent's action consists of the length of the transmitted symbol block and the position of the pilot subcarriers. The reward value is determined by the pilot overhead and the accuracy of the receiver's channel prediction. Simulation results show that BDQN allows the system's pilot overhead to be dynamically adjusted according to the current system state. BDQN can focus on transmission quality or transmission efficiency, or a trade-off between the two. Compared to fixed pilot modes, BDQN ensures system transmission quality while allowing system transmission efficiency to gradually improve with increasing signal-to-noise ratio (SNR).

[0045] like Figures 1 to 7 A deep learning-based channel estimation method for AmBC systems includes the following steps:

[0046] S1. According to the design requirements, an AmBC communication system other than the neural network was built using the Matlab R2023a simulation platform. Specific inputs were set, including experimental conditions such as the number of subcarriers N, channel length L, number of pilots P, and number of non-zero numbers T, and the simulation platform was built.

[0047] S2, at the transmitting end, the frequency domain transmission symbol with pilot is first modulated by IDFT, and then transmitted after inserting a cyclic prefix and parallel-to-serial conversion.

[0048] S3, after passing through the Rayleigh dual-select fading channel, the time-domain signal received by the receiver first undergoes serial-to-parallel conversion and cyclic prefix removal, and then DFT demodulation to obtain the frequency-domain symbol.

[0049] Specifically, the received time-domain signal , can be represented as Where, ∗ represents convolution, N is the number of subcarriers, x(n) and h(n) are the time-domain symbols transmitted on the nth subcarrier and the time-domain channel response that passes through, and w(n) is Gaussian white noise with a mean of 0 and a variance of δ2.

[0050] Corresponding frequency domain received symbols , can be represented as , where Y(k), H(k), X(k) and W(k) are the DFT results of y(n), h(n), x(n) and w(n), respectively.

[0051] For a dual-select fading channel, pilots are placed at the beginning of each transmitted symbol block of length Sb, where the number of pilot OFDM symbols is Sp and the number of pure data OFDM symbols is Sd = Sb − Sp.

[0052] S4. Perform LS estimation on the channel state information (CSI) at the pilot; feed the LS estimate, the transmitted and received pilot symbols together into a k-layer neural network (choose 1, 2, 3, 4, 5, 6) to obtain the optimized CSI.

[0053] Specifically: when the number of subcarriers is N, a received frequency domain OFDM symbol and the corresponding transmit frequency domain OFDM symbols First, it passes through the LS channel estimator to obtain... The input is fed into a K-layer neural network, and after optimization by the K-layer neural network, a real-valued CSI estimate with alternating real and imaginary parts is obtained. The CSI estimates obtained by taking different values ​​of K are compared, and the channel estimation performance varies under different channel conditions. Under the conditions of signal-to-noise ratio gradients of {5, 10, 15, 20} dB and a maximum Doppler frequency of {583} Hz, 10,000 sets of training symbols are randomly generated for each condition to form a training dataset, which is used to train a k-layer neural network. One of the training samples can be represented as... It is input; The label represents the actual CSI of the current OFDM symbol, where the real and imaginary parts alternate. The training dataset consists of training samples from multiple environments with different signal-to-noise ratios (SNRs), which can improve the generalization ability of FC-DNN to noise, enabling the k-layer neural network to have good optimization performance in both low and high SNR environments.

[0054] The K-layer neural network is first trained offline and then deployed in the OFDM system to implement the channel estimation function. After training, the K-layer neural network has a strong generalization ability. In practical applications, it does not need to know the specific channel and noise related statistical information, thus getting rid of the dependence on prior information to a certain extent.

[0055] S5, the K-layer neural network can only process one-dimensional information and cannot perform two-dimensional processing on the two-dimensional TF grid. Therefore, a two-dimensional CNN is introduced to extract the correlation information of the channel along the time and frequency axes to reconstruct the CSI at high resolution. Specifically, two CNNs with the same structure, one for the real part and one for the imaginary part, are used for channel interpolation. Their inputs are two TF grids composed of the real and imaginary parts of Sp pilot OFDM symbols, including the CSI estimates at the pilot subcarriers obtained by the channel estimation network in S4 and the zero values ​​at other data locations. After convolution processing by the two CNNs, two complete TF grids HpCNNs containing the real and imaginary parts of the CSI at all locations are obtained. The two CNNs are trained offline and then deployed in the OFDM system to implement the interpolation function.

[0056] S6, multiple LSTM units are cascaded horizontally to form an LSTM network, and then two LSTM networks are cascaded vertically together. The LSTM units in the same layer have the same parameters. Finally, the fully connected layer is used to extract features and output the results. The number of horizontal units in each LSTM layer represents the memory length of the network, determines the temporal correlation of the network, and the number of units in each layer is at least the number of pilot OFDM symbols Sp.

[0057] Specifically, assuming the number of subcarriers is N, the number of LSTM units in each layer is d, and the current time is n, then the input of the first layer LSTM is the CSI of each OFDM symbol containing all subcarriers from the previous 1 time to the previous d time: The input to the second LSTM layer is the output of the first LSTM layer, which is the short-term memory of the first LSTM layer from the previous 1 time step to the previous d time steps. Where L1 represents the number of neurons in the first LSTM layer. The input to the fully connected layer is the short-term memory output of the last unit of the second LSTM layer. Where L2 represents the number of neurons in the second LSTM layer, and the output is the predicted CSI value at the current time. The prediction process can be briefly represented as follows: Where P represents the prediction process of the LSTM network; the LSTM network is first trained offline and then deployed in the system to realize the prediction function.

[0058] S7, firstly, ZF equalization is used to bring the channel prediction value closer to its theoretical LS estimate, and then a K-layer neural network is used to bring the channel prediction value closer to its actual value, thereby eliminating errors within a certain range. After the LSTM network completes the prediction process for the CSI of the nth data OFDM symbol, the predicted value H is... LSTM (n) is updated.

[0059] Specifically, for the predicted value H LSTM (n) The update process includes the following 5 steps:

[0060] S71, recover the nth OFDM transmitted symbol through ZF equalization, i.e.

[0061] S72 performs a hard decision on the recovery of soft symbols. ;

[0062] S73, perform LS estimation on the hard symbol, i.e. ;

[0063] S74, the optimized CSI is obtained through the K-layer neural network proposed in S3: H DNN (n);

[0064] S75, H LSTM The value of (n) is determined by H DNN (n) substitution.

[0065] If the hard decision value X obtained in step S72 above is ZFH If (n) is sufficiently accurate, then the H obtained in step S74 DNN (n) will be very close to the real CSI, thus mitigating the error propagation problem.

[0066] This invention addresses the channel estimation and prediction problems in OFDM communication systems by proposing a deep learning-based channel estimation and prediction network. The network consists of a threshold iteration algorithm decomposed into a K-layer neural network, a CNN, and an LSTM network. The K-layer neural network and CNN handle channel estimation and interpolation, while the LSTM network handles channel prediction. The threshold iteration algorithm, with its powerful curve fitting capabilities, is decomposed into a K-layer neural network to optimize the LS estimation, addressing the noise-dependent issues of the LS method and the requirement for prior channel information in the LMMSE method. The CNN, with its powerful 2D image processing capabilities, is used for channel interpolation, resolving the edge blurring and detail issues of spline interpolation. The LSTM network, with its powerful time series processing capabilities, is used for pilot-free channel prediction, and a data decision feedback mechanism is used to reduce the impact of error propagation. This channel estimation and prediction network structure is highly flexible; the three networks can be trained independently or replaced by more advanced networks to achieve better performance. Simulation results demonstrate that the proposed channel estimation and prediction network overcomes the poor performance of traditional channel estimation and prediction methods, and improves transmission efficiency while maintaining transmission quality under high signal-to-noise ratio conditions.

[0067] To further explore the relationship between system transmission quality and transmission efficiency, this invention proposes an adaptive pilot position and spacing adjustment mechanism based on reinforcement learning, and jointly constructs a continuous transmission OFDM communication system with a channel estimation and prediction network based on deep learning. This invention maps the parameters of the communication system to the elements of reinforcement learning, constructing a reinforcement learning system. Then, for the state acquisition step, it proposes a noise intensity estimation method based on a threshold iteration algorithm expanded into a K-layer neural network and a future channel state information prediction method based on multi-step LSTM. Finally, it introduces the BDQN algorithm to divide actions into simple sub-actions in multiple dimensions, solving the problem of a large number of actions, and learns how to adjust the pilot scheme according to the current system state. Simulation results demonstrate the effectiveness of the proposed noise intensity estimation method based on a threshold iteration algorithm expanded into a K-layer neural network and the future channel state information prediction method based on multi-step LSTM. This shows that the adaptive pilot position and spacing adjustment mechanism based on BDQN can focus on transmission quality or transmission efficiency, or a compromise between the two, and that the joint OFDM system based on deep learning channel estimation and prediction network and BDQN can improve transmission efficiency while ensuring transmission quality.

[0068] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, performs the method described above.

[0069] A computer system includes a processor and a storage medium on which a computer program is stored, wherein the processor reads from the storage medium and runs the computer program to perform the method described above.

[0070] Those skilled in the art will further appreciate that the various illustrative logic blocks, modules, circuits, and algorithm steps described in conjunction with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or a combination of both. To clearly illustrate this hardware-software interchangeability, the various illustrative components, blocks, modules, circuits, and steps are described above in a generalized manner in terms of their functionality. Whether such functionality is implemented as hardware or software depends on the specific application and the design constraints imposed on the overall system. Those skilled in the art may implement the described functionality in different ways for each specific application, but such implementation decisions should not be construed as departing from the scope of this invention.

[0071] The various illustrative logic blocks, modules, and circuits described in conjunction with the embodiments disclosed in this invention can be implemented or performed using a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described in this invention. The general-purpose processor may be a microprocessor, but in alternatives, it may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors cooperating with a DSP core, or any other such configuration.

[0072] The steps of the methods or algorithms described in conjunction with the embodiments disclosed in this invention may be embodied directly in hardware, in a software module executed by a processor, or in a combination of both. The software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor so that the processor can read and write information to / from the storage medium. In an alternative, the storage medium may be integrated into the processor. The processor and storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and storage medium may reside as discrete components in the user terminal.

[0073] In one or more exemplary embodiments, the described functionality may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software as a computer program product, the functionality may be stored or transmitted as one or more instructions or code on or through a computer-readable medium. A computer-readable medium includes both computer storage media and communication media, encompassing any medium that facilitates the transfer of a computer program from one location to another. A storage medium may be any available medium accessible to a computer. By way of example and not limitation, such a computer-readable medium may include RAM, ROM, EEPROM, CD-ROM or other optical disc storage, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and is accessible to a computer. Any connection is also legitimately referred to as a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of a medium. As used in this invention, disks and discs include compact discs (CDs), laser discs, optical discs, digital multi-purpose discs (DVDs), floppy disks, and Blu-ray discs, wherein disks typically reproduce data magnetically, while discs reproduce data optically using lasers. Combinations of the above should also be included within the scope of computer-readable media.

[0074] The prior description of this disclosure is provided to enable any person skilled in the art to make or use it. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles defined in this invention can be applied to other variations without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not intended to be limited to the examples and designs described herein, but should be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0075] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A channel estimation method for an AmBC system based on deep learning, characterized in that, Includes the following steps: S1. According to the design requirements, build an AmBC communication system other than the neural network using the Matlab R2023a simulation platform, set specific inputs including the number of subcarriers N, channel length L, pilot number P, and non-zero number T experimental conditions, and build the simulation platform. S2, at the transmitting end, the frequency domain transmission symbol with pilot is first modulated by IDFT, and then transmitted after inserting a cyclic prefix and parallel-to-serial conversion; S3, after passing through the Rayleigh dual-select fading channel, the time-domain signal received by the receiver first undergoes serial-to-parallel conversion and cyclic prefix removal, and then DFT demodulation to obtain the frequency-domain symbol. S4, perform LS estimation on the channel state information (CSI) at the pilot; feed the LS estimate, the transmitted and received pilot symbols together into the k-layer neural network to obtain the optimized CSI; S5 introduces a two-dimensional CNN to extract the correlation information of the channel along the time and frequency axes, and performs high-resolution reconstruction of CSI; S6, multiple LSTM units are cascaded horizontally to form an LSTM network, and then two LSTM networks are cascaded vertically together. The LSTM units in the same layer have the same parameters. Finally, the fully connected layer is used to extract features and output the results. The number of units in each LSTM layer represents the memory length of the network, determines the temporal correlation of the network, and the number of units in each layer is at least the number of pilot OFDM symbols Sp. S7, firstly, ZF equalization is used to bring the channel prediction value closer to its theoretical LS estimate, and then a K-layer neural network is used to bring the channel prediction value closer to its actual value, thereby eliminating errors within a certain range. After the LSTM network completes the prediction process for the CSI of the nth data OFDM symbol, the predicted value H is... LSTM (n) is updated.

2. The method according to claim 1, characterized in that: The time-domain signal received in step S3 It can be expressed as y(n)=h(n)*x(n)+w(n); where * denotes convolution, N is the number of subcarriers, x(n) and h(n) are the time-domain symbols transmitted on the nth subcarrier and the time-domain channel response that passes through, and w(n) is Gaussian white noise with a mean of 0 and a variance of δ2; Corresponding frequency domain received symbols It can be expressed as Y(k)=H(k)X(k)+W(k), where Y(k), H(k), X(k) and W(k) are the DFT results of y(n), h(n), x(n) and w(n) respectively; For a dual-select fading channel, pilots are placed at the beginning of each transmitted symbol block of length Sb, where the number of pilot OFDM symbols is Sp and the number of pure data OFDM symbols is Sd = Sb - Sp.

3. The method according to claim 1, characterized in that, Step S4 specifically involves: when the number of subcarriers is N, receiving a frequency domain OFDM symbol. and the corresponding transmit frequency domain OFDM symbols First, it passes through the LS channel estimator to obtain... The input is fed into a K-layer neural network, and after optimization by the K-layer neural network, a real-valued CSI estimate with alternating real and imaginary parts is obtained. The CSI estimates obtained by taking different values ​​of K are compared, and the channel estimation performance varies under different channel conditions. Under the conditions of signal-to-noise ratio gradients of {5, 10, 15, 20} dB and a maximum Doppler frequency of {583} Hz, 10,000 sets of training symbols are randomly generated for each condition to form a training dataset, which is used to train a k-layer neural network. One of the training samples can be represented as... It is input; It is a label that indicates the actual CSI of the current OFDM symbol, where the real and imaginary parts are arranged alternately.

4. The method according to claim 1, characterized in that, Step S5 specifically involves: two identical CNNs, one for the real part and one for the imaginary part, being used for channel interpolation. Their inputs are two TF grids composed of the real and imaginary parts of Sp pilot OFDM symbols, respectively, including the CSI estimate at the pilot subcarrier obtained by the channel estimation network in S4 and the zero values ​​at other data locations. After two CNN convolutions, two complete TF meshes H are obtained, each containing the real and imaginary parts of the CSI at all locations. pCNN The two CNNs were trained offline and then deployed in the OFDM system to perform interpolation.

5. The method according to claim 1, characterized in that, Step S6 is as follows: Assuming the number of subcarriers is N, the number of LSTM units in each layer is d, and the current time is n, then the input of the first layer LSTM is the CSI of each OFDM symbol containing all subcarriers from the previous 1 time to the previous d time: The input to the second LSTM layer is the output of the first LSTM layer, that is, the short-term memories from the previous 1 time step to the previous d time steps of the first LSTM layer. Where L 1 This indicates the number of neurons in the first LSTM layer; the input to the fully connected layer is the short-term memory output of the last unit of the second LSTM layer. Where L 2 This indicates the number of neurons in the second LSTM layer, and the output is the predicted CSI value at the current time step. The prediction process can be briefly represented as follows: Where P represents the prediction process of the LSTM network; the LSTM network is first trained offline and then deployed in the system to realize the prediction function.

6. The method according to claim 1, characterized in that, Step S7: Predict the value H LSTM (n) The update process includes the following 5 steps: S71, recover the nth OFDM transmitted symbol through ZF equalization, i.e. S72, perform a hard decision on the recovery of soft symbols, X ZFH (n) = sign(X) ZFS (n)); S73, perform LS estimation on the hard symbol, i.e. S74, the optimized CSI is obtained through the K-layer neural network proposed in S3: H DNN (n); S75, H LSTM The value of (n) is determined by H DNN (n) substitution.

7. A computer-readable storage medium, characterized in that: The medium contains a computer program, which, when run, performs the method as described in any one of claims 1 to 6.

8. A computer system, characterized in that: It includes a processor and a storage medium, on which a computer program is stored, and the processor reads from the storage medium and runs the computer program to perform the method as described in any one of claims 1 to 6.

Images