Data model dual-drive iterative receiving method for non-orthogonal superposition pilot frequency transmission

Abstract

The invention discloses a data model dual-drive iterative receiving method for non-orthogonal superposition pilot frequency transmission, and belongs to the technical field of wireless communication. According to the method, an iterative receiving architecture comprising a channel estimation sub-module, a signal detection sub-module and a decoding sub-module and a channel estimation network based on residual training is constructed by fusing traditional model driving and data driving deep learning technologies. The method specifically comprises the following steps: performing pilot interference elimination and initial data detection by utilizing initial channel estimation; pilot frequency and data interference elimination is carried out through reliable data estimation fed back by decoding, and least square channel estimation is generated; channel estimation is denoised by adopting a residual convolutional network, and a channel coefficient is optimized; and iteratively updating a detection and decoding result, and finally outputting original bit estimation. According to the method, error propagation is suppressed in combination with a self-supervision mechanism, the channel estimation precision and the data detection performance are remarkably improved in a non-orthogonal superposition pilot frequency scene, and the method has the advantages of being low in complexity, high in robustness and suitable for a dynamic scene and meets the high-throughput transmission requirement of a 6G communication system.

Description

Technical Field

[0001] The present invention relates to a data model dual-driven iterative receiving method for non-orthogonal superimposed pilot transmission, belonging to the field of wireless communication technology. Background Art

[0002] The wireless transmission scheme based on non-orthogonal superimposed pilots realizes the non-orthogonal coexistence of pilots and data by superimposing pilots and data on the same resource block according to a certain power ratio, which is different from the orthogonal (competing) relationship between pilots and data in traditional wireless transmission systems. It can greatly improve the system throughput under limited transmission bandwidth, providing a new idea for the development of 6G systems. On the other hand, in the non-orthogonal superimposed pilot transmission system, the superposition of pilots and data will cause serious coupling interference, posing a major challenge to channel estimation and signal detection. Therefore, an iterative receiver architecture that combines channel estimation, signal detection, and decoding is generally adopted in receiver design to reduce the coupling interference between pilots and signals. However, existing iterative receiving schemes face challenges such as high receiving complexity and large delay overhead in practical applications, and related designs highly depend on assumptions that are difficult to meet in actual systems, with limited practical application and generalization capabilities in non-orthogonal superimposed pilot transmission for 6G.

[0003] In recent years, the development of deep learning-related technologies in the field of artificial intelligence has brought new research ideas and design paradigms to the field of mobile communication. For intelligent receiving schemes for non-orthogonal superimposed pilot transmission, typical work uses a data-driven deep learning method to design a black box network, effectively breaking through the limitations of traditional modular design. However, due to the lack of interpretability, it is difficult for black box receiving schemes to balance performance and complexity, which requires integrating expert knowledge in the design to develop a more reasonable network structure and joint optimization method. Therefore, considering introducing a data model dual-driven adaptive enhancement mechanism on the basis of the iterative receiving architecture to form an iterative receiving scheme with high performance, moderate complexity, and strong robustness to scene changes. Summary of the Invention

[0004] Object of the Invention: The present invention provides a data model dual-driven iterative receiving method for non-orthogonal superimposed pilot transmission, aiming to use a neural network-assisted joint optimization design driven by data models to form an intelligent receiving scheme with high performance, moderate complexity, and strong robustness to scene changes.

[0005] Technical Solution: The present invention adopts the following technical solutions:

[0006] The data model dual-driven iterative receiving method for non-orthogonal superimposed pilot transmission described in the present invention includes channel estimation, signal detection, and decoding sub-modules, as well as a channel estimation network based on residual training, and includes the following steps:

[0007] (1) Calculate the initial channel coefficient estimate using the linear minimum mean square error algorithm, and input it into the signal detection sub-module. Combine the transmitted pilot sequence to perform pilot interference cancellation on the received signal vector, give the initial data estimate, and input it into the decoding sub-module for error correction;

[0008] (2) The decoding sub-module outputs more accurate a priori likelihood information, remaps it as the second data estimate and feeds it back to the channel estimation sub-module. Combine the initial channel coefficient estimate, the transmitted pilot sequence, and the second data estimate fed back by the decoding sub-module to perform pilot and data interference cancellation on the received signal vector, and calculate the least squares channel estimate to input into the channel estimation network. The channel estimation network is designed by a convolutional neural network based on residual training, performs denoising processing on the input least squares estimate, and outputs a more accurate channel coefficient estimate;

[0009] (3) Input the more accurate channel coefficient estimate into the signal detection sub-module to update the pilot interference cancellation process and output the data estimate. The decoding sub-module performs error correction decoding on the data estimate and obtains the estimate of the original transmitted bits.

[0010] Further, step (1) specifically includes:

[0011] (1.1) Combine the received signal vectors y r of the m = 1, 2,..., N m th receiving antennas and the pilot sequences p t transmitted by the n = 1, 2,..., N n th transmitting antennas to calculate the initial channel coefficient estimate of the channel coefficient between the mth receiving antenna and the nth transmitting antenna

[0012] (1.2) The channel coefficient estimate is input into the signal detection sub-module. At the w = 1, 2,..., Wth time-frequency resources, combine the transmitted pilot vector and the channel estimation matrix to perform pilot interference cancellation on the received signal vector ; According to the data reception component after pilot interference cancellation and the channel estimation matrix give the initial first data estimate and calculate the extrinsic log-likelihood information of the data symbol corresponding to the q = 1, 2,..., Q bits of the nth transmitting antenna and the wth time-frequency resource and input it into the decoding sub-module for error correction, where Q = log2M is the number of bits corresponding to the M-ary quadrature amplitude modulation constellation symbol; the subscript E is used to label the extrinsic information.

[0013] Further, step (2) specifically includes:

[0014] (2.1) The decoding sub-module outputs more accurate prior likelihood information for remapping to data estimation and feeds it back to the channel estimation sub-module; subscript A is used to label the prior;

[0015] (2.2) The channel estimation sub-module combines the initial channel coefficient estimation to transmit the pilot sequence p n and the second data estimation fed back by the decoding sub-module to perform pilot and data interference cancellation on the received signal vector y m ; then, based on the pilot received component after pilot and data interference cancellation and the transmitted pilot sequence p n it gives the least squares channel estimation and inputs it into the channel estimation network; subscript p is used to label the pilot;

[0016] (2.3) The denoising network based on residual training is used as the channel estimation network, and this denoising network is designed by a convolutional neural network structure; for the least squares channel estimation vectors corresponding to n = 1, 2,..., N t root transmitting antennas as the input of this channel estimation network, the input vector of length W is first mapped to a K×T-dimensional matrix in time-frequency resources where K is the number of subcarriers in the frequency domain and T is the number of time-domain symbols, and W = KT; then the real and imaginary parts of the matrix are separated, and is input into the convolutional neural network in the form of 2N t feature maps; in addition, to suppress error propagation, a self-supervised mechanism is introduced, and the reliability of the second data estimation fed back by the decoding sub-module is used to affect the calculation of the output features of the convolutional neural network; finally, the output feature map of the channel estimation network corresponds to a more accurate channel coefficient estimation

[0017] Furthermore, in step (3), the channel coefficient estimation output by the channel estimation network is input into the signal detection sub-module, and on the w-th time-frequency resource, in combination with the transmitted pilot vector p [w] and the updated channel estimation matrix it performs pilot interference cancellation on the received signal vector y [w] to update the data received component and the first data estimation and correspondingly update the calculation of the log-likelihood information and input it into the decoding sub-module for error correction decoding to obtain an estimate of the original transmitted bits

[0018] Furthermore, the channel estimation network extracts potential correlation features of the input feature map using convolutional layers, and learns the mapping rule from the least squares channel estimation to the residual. The specific steps are as follows: The input feature map is first input into convolutional layer 1, and processed using C1 convolutional kernels of size F×F×2N t ; subsequently, it passes through 4 residual modules, each module containing 2 convolutional layers, and the feature map is processed using C1 convolutional kernels of size F×F×C1; after continuing to process the feature map using C1 convolutional kernels of size F×F×C1 in convolutional layer 2, convolutional layer 3 finally outputs 2N t feature maps of size K×T, corresponding to the real and imaginary parts of the channel coefficient vector group . The convolutional layer parameter is set as C1 = 32 and F = 3; in addition, the confidence information is used to measure the reliability of the data estimation fed back by the decoding sub-module, and the confidence matrix is input into a deep neural network composed of three fully connected layers, which respectively contain {32, 32, N } neurons. The deep neural network outputs confidence features t which are still mapped into a matrix of size K×T according to time-frequency resources, forming another N feature maps input into convolutional layer 3, thereby affecting the calculation of the output features and suppressing error propagation. t

[0019] Furthermore, the mini-batch gradient descent algorithm is used to perform end-to-end supervised learning training on the complete receiving module composed of the channel estimation, signal detection and decoding sub-module, and the channel estimation network trained based on residuals, and trained for multiple rounds; a batch of the training set is represented as a set composed of S randomly generated samples, where the true channel coefficient h (i) is used as the label, and the received signal vector y (i) and the transmitted pilot sequence p (i) are used as the input features of the complete receiving module. An adaptive momentum estimation optimizer is selected to optimize the trainable parameters in the channel estimation network; the mean square error loss function is used for training, and the learning rate is set to 0.001.

[0020] The present invention also provides a data model dual-driven iterative receiving system for non-orthogonal superimposed pilot transmission, used to implement the above method, including:

[0021] A channel estimation sub-module, used for initial channel coefficient calculation and interference cancellation;

[0022] A signal detection sub-module, which generates data estimation based on channel estimation;

[0023] A decoding sub-module, which corrects the data estimation and feeds back prior information;​

[0024] A channel estimation network based on residual training optimizes channel coefficient estimation;

[0025] A control module coordinates the iterative execution and parameter update of each sub-module.

[0026] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory. When the processor executes the program, the method described above is implemented.

[0027] The present invention also provides a non-transitory computer-readable storage medium storing computer instructions for causing a computer to execute the method described above.

[0028] The memory, the processor, and the non-transitory computer-readable storage medium all adopt common devices, having wide applicability and accessibility. Users can independently select appropriate electronic devices and storage media according to factors such as their actual needs, budgets, and application scenarios to implement the method of the present invention.

[0029] Advantageous effects: Compared with the prior art, the present invention has the following advantageous effects:

[0030] Based on an iterative architecture of joint channel estimation, detection, and decoding, the present invention designs a receiver joint optimization module. The potential correlation characteristics of input features are extracted by a channel estimation network driven by a data model, thereby realizing the adaptive enhancement of the iterative receiver and improving the accuracy of iterative information exchange between sub-modules. Compared with existing solutions, the present invention has the advantages of high performance, moderate complexity, and strong robustness to scene changes. Brief Description of the Drawings

[0031] Figure 1 is a system block diagram of an embodiment of the present invention;

[0032] Figure 2 is a schematic diagram of the architecture of the channel estimation network used in the present invention. Detailed Embodiments

[0033] The present invention will be specifically described below in conjunction with the drawings and embodiments of the MIMO-OFDM system.

[0034] I. System Model Adopted in this Embodiment

[0035] Consider a MIMO-OFDM system using spatial multiplexing technology. There are N t antennas at the transmitting end and N r antennas at the receiving end. The OFDM transmission frame contains T consecutive symbols, and the number of subcarriers is K. In this embodiment, N t is set to N r= 2, K = 72, T = 14. In this embodiment, 16-QAM constellation modulation and LDPC channel coding are adopted. To obtain channel state information, a pilot sequence of length W is mapped to different time-frequency resources, where W = KT, and the pilot sequence transmitted by the nth transmit antenna is denoted as p n , where n = 1, 2,..., N t . For data transmission, the channel encoder encodes an information bit stream of length N b = 2016 at a code rate of r = 1 / 2, and the codeword is modulated to a data vector d n ∈ A W , where N c = 4032, A is the 16-QAM constellation symbol set, and the number of bits Q corresponding to the 16-QAM constellation symbol is 4.

[0036] Under the SIP transmission architecture, the transmission block of the nth transmit antenna can be expressed as:

[0037]

[0038] where ρ = 0.3 is the power allocation coefficient. Corresponding to the m = 1, 2,..., N r th receive antenna, the received signal vector y m can be expressed as:

[0039]

[0040] where, P n = Diag(p n ) represents a diagonal matrix with the main diagonal elements being p n , and similarly, D n = Diag(d n ), h m,n is the channel coefficient, w m is a complex Gaussian white noise vector, and the noise variance is

[0041] On the wth time-frequency resource, the corresponding received components represented by Equation (2) can form an N r -dimensional vector, and there is Based on this, an equivalent MIMO subsystem can be constructed:

[0042]

[0043] where represents the channel matrix, is the pilot vector, w [w] is the noise.

[0044] In this embodiment, considering the uplink scenario, a MIMO-OFDM channel is generated using a cluster delay line model. In the test environment, the mobile speed v of the user terminal is 360 km / h, and the time-frequency correlation information required for the initial linear minimum mean square error channel estimation is calculated based on the second-order statistical average of 10 5 sets of MIMO-OFDM channel samples.

[0045] II. Specific steps of this embodiment

[0046] As Figure 1 shown, the embodiment of the present invention provides a system block diagram of a data model dual-driven iterative receiving method for non-orthogonal superimposed pilot transmission, which includes a channel estimation, signal detection and decoding sub-module, and a channel estimation network based on residual training. In this embodiment, the signal detection sub-module first gives an initial data estimate according to the linear minimum mean square error channel estimation, and inputs it into the decoding sub-module for error correction, and feeds back a more accurate data estimate to the channel estimation sub-module for pilot and data interference cancellation of the received components, and gives the least squares channel estimation as the input to the channel estimation network. The structure of the channel estimation network is as Figure 2 shown. The convolutional layer is used to extract the potential correlation characteristics of the input features, and the input features are denoised through the residual training mechanism, and a more accurate channel coefficient estimation is output. Subsequently, the data estimation calculation of the signal detection sub-module and the error correction decoding process of the decoding sub-module are updated, and an estimation of the original transmitted bits is given.

[0047] Offline, end-to-end supervised learning training is performed on the complete receiving module composed of the channel estimation, signal detection and decoding sub-modules and the channel estimation network based on residual training to optimize the trainable parameters in the channel estimation network. When deployed online, the trained network is used to complete the estimation of the original transmitted bits. One forward estimation is divided into four steps: initial signal detection and decoding, data estimation feedback and interference cancellation, channel estimation network denoising, and update of signal detection and decoding:

[0048] (1) Initial signal detection and decoding

[0049] The initial channel coefficient estimation of the channel coefficient between the m-th receiving antenna and the n-th transmitting antenna is calculated using the linear minimum mean square error algorithm and input into the signal detection sub-module. At the w-th time-frequency resource, combined with the transmitted pilot vector p [w] and the channel estimation matrix the received signal vector y [w] is subjected to pilot interference cancellation; according to the data received component after pilot interference cancellation and the channel estimation matrix the initial first data estimate And calculate the extrinsic log-likelihood information of the q-th bit corresponding to the data symbol of the n-th transmitting antenna and the w-th time-frequency resource. The input decoding sub-module performs error correction.

[0050] (2) Data estimation feedback and interference cancellation

[0051] The decoding sub-module outputs more accurate a priori likelihood information. A priori information Through remapping calculation, the data symbol estimation of the n-th transmitting antenna and the w-th time-frequency resource can be restored. Form the second data estimation vector. And feedback it to the channel estimation sub-module, combined with the initial channel coefficient estimation. Transmit the pilot sequence p n And the second data estimation vector For the received signal vector y m Perform pilot and data interference cancellation, and give the received pilot component after interference cancellation.

[0052] (3) Channel estimation network denoising

[0053] According to the received pilot component And the transmitted pilot sequence p n Give the least squares channel estimation. An input vector of length W Is mapped to a K×T-dimensional matrix according to time-frequency resources. And separate the matrix Of the real and imaginary parts, and Is input into the channel estimation network in the form of 2N t Feature maps; The input feature maps are first input into Convolutional Layer 1 (Conv1), and processed using C1 convolutional kernels of size F×F×2N t Subsequently, through 4 residual modules (ResBlock), each module contains 2 convolutional layers, and the feature maps are processed using C1 convolutional kernels of size F×F×C1; After continuing to process the feature maps using C1 convolutional kernels of size F×F×C1 in Convolutional Layer 2 (Conv2), Convolutional Layer 3 (Conv3) finally outputs 2N t K×T-dimensional feature maps, corresponding to the real and imaginary parts of the channel coefficient vector group The convolutional layer parameters are set as C1 = 32, F = 3; To suppress error propagation, a self-supervised mechanism is introduced, and the confidence information Measures the reliability of the second data estimation feedback by the decoding sub-module, and The confidence matrix Is input into a deep neural network composed of three fully connected layers, respectively containing {32, 32, Nt} neurons, and the deep neural network outputs confidence features It is still mapped to a K×T-dimensional matrix according to time-frequency resources, forming another N t feature maps, thus affecting the calculation of the output features.

[0054] Use the mini-batch gradient descent algorithm to perform end-to-end supervised learning training on the complete receiving module composed of the above channel estimation, signal detection and decoding sub-modules and the channel estimation network based on residual training. In this embodiment, a total of 100 rounds of training are performed, and each round contains 128 batches. A batch of the training set is represented as a set composed of S randomly generated samples, where the true channel coefficient h (i) is used as the label, and the received signal vector y (i) and the transmitted pilot sequence p (i) are used as the input features of the complete receiving module. i is the sample serial number. In this embodiment, S is taken as 64. The moving speed v of the user terminal corresponding to the channel samples in the training set is 108 km / h. The time-frequency related information required for the initial linear minimum mean square error channel estimation is also calculated based on the second-order statistical average of the channel samples in the training set. Select the adaptive momentum estimation optimizer to optimize the trainable parameters in the neural network. The training uses the mean square error loss function, and the learning rate is set to 0.001. In each training, 64 samples are sent into the network for forward propagation. After calculating the loss function, backpropagation is performed to optimize the parameters of the channel estimation network.

[0055] After completing the above training offline, the network is deployed online. In this embodiment, the moving speed v of the user terminal corresponding to the test channel samples for evaluating the network performance is 360 km / h. Since the convolutional neural network structure effectively extracts the potential correlation characteristics of the input features during training, it can still effectively denoise the input features for new scenarios during testing, compensating for the performance loss caused by the insufficient initial estimation based on the linear minimum mean square error method; in addition, on the basis of the iterative receiving architecture, a lightweight channel estimation network auxiliary module is introduced for joint optimization design between modules, and the method of data model dual-driven deep learning effectively balances performance and complexity. Thus, an intelligent receiving solution with high performance, moderate complexity, and strong robustness to scene changes is formed.

[0056] (4) Update signal detection and decoding

[0057] Input the channel coefficient estimation output by the channel estimation network into the signal detection sub-module, and combine the transmitted pilot vector p [w] and the updated channel estimation matrix to perform pilot interference cancellation on the received signal vector y [w] and update the data reception component and the first data estimation Update the log-likelihood information of the q-th bit corresponding to the data symbol of the n-th transmitting antenna and the w-th time-frequency resource Calculate, and the input decoding sub-module performs error correction decoding to obtain an estimation of the original transmitted bits

[0058] The above-described only represents a preferred embodiment of the present invention in conjunction with the accompanying drawings, and cannot be used to limit the scope of rights included in the present invention. It should be understood that any equivalent changes made without departing from the spirit of the present invention fall within the protection scope covered by the claims of the present invention.

Claims

1. A data model dual-driven iterative reception method for non-orthogonal superposition pilot transmission, characterized in that: The following steps are involved: (1) Based on the received signal vector and the pilot sequence, the initial channel coefficient estimate is calculated by the linear minimum mean square error algorithm, input into the signal detection submodule to eliminate the pilot interference, generate the initial data estimate and input into the decoding submodule for error correction; (2) The decoding submodule outputs the prior likelihood information, which is remapped into a second data estimate and fed back to the channel estimation submodule. The pilot and data interference of the received signal are eliminated by combining the initial channel estimate and the feedback data to generate a least squares channel estimate. (3) inputting the least squares channel estimate into a channel estimation network based on residual training for denoising, and outputting an optimized channel coefficient estimate; (4) The signal detection submodule and the decoding submodule are updated using the optimized channel coefficients, and pilot interference elimination and detection decoding are iteratively performed to finally obtain the original transmitted bit estimate.

2. The method according to claim 1, characterized in that The pilot and data interference elimination in step (2) is specifically as follows: according to the second data estimation fed back by the decoding submodule, the pilot and data interference components are eliminated from the received signal, the pilot reception component is separated and the least squares channel estimation is calculated.

3. The method according to claim 1, characterized in that The channel estimation network based on residual training in step (3) includes: The input layer maps the channel estimation vector into a time-frequency matrix and separates the real and imaginary parts as feature maps; Multiple residual modules extract the spatial correlation of feature maps and optimize the denoising process through convolutional layer design; The self-supervision mechanism introduces the confidence information of decoding feedback and generates additional feature maps through deep neural networks to suppress error propagation.

4. The method according to claim 3, characterized in that The residual module contains two convolutional layers, each of which uses C1 convolutional layers of size F×F×2N. t The convolution kernel processes the feature map and finally outputs a K×T dimensional feature map to reconstruct the real and imaginary parts of the channel coefficients, where F is the convolution kernel size and N t is the number of transmitting antennas, K is the number of subcarriers in the frequency domain, and T is the number of symbols in the time domain.

5. The method according to claim 1, characterized in that The method performs end-to-end supervised training through a small batch gradient descent algorithm, adopts a mean square error loss function and an adaptive momentum optimizer, and the training set includes randomly generated received signals, pilot sequences, and true channel coefficient labels.

6. The method according to claim 1, characterized in that The method is applicable to MIMO-OFDM systems, where pilot signals and data are transmitted in a non-orthogonal manner, and the accuracy of channel estimation and signal detection in high mobility scenarios is improved through iterative optimization.

7. A data model dual-driven iterative receiving system for non-orthogonal superposition pilot transmission, used to implement the method according to any one of claims 1 to 6, characterized in that: include: Channel estimation submodule, used for initial channel coefficient calculation and interference elimination; A signal detection submodule that generates data estimates based on the channel estimates; The decoding submodule estimates and corrects data errors and feeds back prior information; A channel estimation network based on residual training to optimize channel coefficient estimation; The control module coordinates the iterative execution and parameter update of each sub-module.

8. An electronic device comprising a memory, a processor and a computer program stored in the memory, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium storing computer instructions, characterized in that: The instructions are used to enable a computer to execute the method according to any one of claims 1 to 6.

Images