Lightweight deep receiver construction method and system, receiver, device, medium

Abstract

The application discloses a lightweight deep receiver construction method and system, a receiver, equipment and a medium, wherein the construction method comprises the following steps: constructing a MU-MIMO OFDM uplink transmission system based on non-orthogonal superimposed pilot transmission to obtain an original training data set; constructing a lightweight deep receiver, wherein the lightweight deep receiver is based on time domain shift and frequency domain shift operations; and performing multiple iteration model training and convergence on the lightweight deep receiver by using the original training data set. By designing an uplink transmission system based on non-orthogonal superimposed pilot transmission to enhance the frequency spectrum usage rate, and then constructing a lightweight deep receiver based on time domain shift and frequency domain shift to perform model training, since the pilot structure is stable and the data symbols are random, the shift operation effectively enhances the separability of the two, so that the lightweight deep receiver can realize the decoupling of data and pilot in the non-orthogonal superimposed pilot scene, and the decoding accuracy of the data stream is significantly improved.

Description

Technical Field

[0001] This application relates to the field of information and communication technology, and in particular to a lightweight depth receiver method and system, receiver, device and medium based on non-orthogonal superposition pilot transmission. Background Technology

[0002] In future wireless communication systems designed for massive connectivity, traditional orthogonal pilot schemes cannot simultaneously balance pilot overhead and data transmission throughput. They consume a lot of pilot resources, resulting in low transmission efficiency and making it difficult to efficiently support the massive access demands of the Internet of Things scenario.

[0003] Meanwhile, under complex or non-ideal channel conditions, traditional modular receivers typically rely on independent signal processing modules such as channel estimation, equalization, and demodulation. Because these modules operate serially, errors from the preceding stages are continuously propagated and amplified to the following stages, resulting in a significant decrease in decoding performance and limited overall system robustness.

[0004] Currently, deep learning-based receivers have become a new research direction under the aforementioned background. However, the inventors have found that existing network models proposed in the industry have a large number of design parameters and high computational costs, making it difficult to meet the low power consumption and low latency requirements of access terminals. Therefore, there is an urgent need to provide a more optimized and lightweight deep learning model with lower complexity. Summary of the Invention

[0005] To address the problem, this application proposes a lightweight deep receiver construction method and system, receiver, device, and medium based on non-orthogonal superimposed pilots. The technical solution is as follows:

[0006] On the one hand, a lightweight method for constructing a deep receiver is provided, including the following steps:

[0007] A MU-MIMO OFDM uplink transmission system based on non-orthogonal superimposed pilots was constructed to obtain the original training dataset;

[0008] Construct a lightweight depth receiver, wherein the lightweight depth receiver is based on time-domain shifting and frequency-domain shifting operations;

[0009] The lightweight deep receiver is iterated multiple times using the original training dataset until the model converges.

[0010] The original training dataset includes sample quadruples. ;in, Indicates the total received signal. Represents the total pilot matrix, Indicates noise intensity. Represents the original information bit tensor. Indicates the modulation order.

[0011] The construction of a lightweight deep receiver includes the following steps: concatenating and reshaping the dimensions of the real and imaginary parts of the original training dataset to adjust the input shape;

[0012] The first convolutional layer maps the channel dimension to a high-dimensional feature space and performs layer normalization to obtain feature data.

[0013] The feature data is processed by a multi-layer block shift decoder to achieve implicit mapping for channel estimation and channel equalization.

[0014] The output processed by the multi-layer block shift decoder is mapped to the log-likelihood ratio of each resource unit through layer normalization and a second convolutional layer.

[0015] The block shift decoder includes the following steps: grouping the feature data along the channel dimension; performing time-domain shifting on the first and second groups of feature data, frequency-domain shifting on the third and fourth groups of feature data, while keeping the other groups of feature data unchanged; reorganizing and refining the shifted feature data through layer normalization and multilayer perceptron, and fusing it through residual connections.

[0016] The original training set was constructed into a comprehensive training dataset covering multiple different channel scenarios.

[0017] ,

[0018] in, Indicates the first An aggregated dataset of sample quadruplets corresponding to each channel scenario.

[0019] The optimization objective for the lightweight depth receiver is as follows:

[0020] ;

[0021] in, Representing the channel scenario Loss function of depth receiver This indicates the batch size during the metamodel update phase. Represents the dataset A small batch of channel sample quadruples is sampled. This represents the expected loss of the sampled data. Learnable parameters for lightweight depth receivers;

[0022] Iterative training of a lightweight deep receiver includes:

[0023] Phase 1: Update task parameters for each channel scenario:

[0024] ;

[0025] in, Indicates the task learning rate; This represents the model parameters at the start of the i-th task. This represents the loss for the i-th task with respect to the parameters. gradient, This represents the parameters after a gradient update for the i-th task;

[0026] Phase Two: Meta-model parameter updates for all channel scenarios:

[0027] ;

[0028] in, This represents the learning rate of the meta-model.

[0029] On the other hand, a lightweight depth receiver constructed using any of the above methods is provided, which is used to achieve accurate decoding of data streams in non-orthogonal superposition pilot transmission schemes.

[0030] Furthermore, a lightweight depth receiver system is provided, the system comprising:

[0031] A transmitter is used to encode and transmit data streams in a non-orthogonal superposition pilot transmission scheme.

[0032] The aforementioned lightweight depth receiver is used to receive data transmitted by the transmitter and accurately decode the transmitted data.

[0033] In another aspect, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the above methods.

[0034] On the other hand, a readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements any of the above-mentioned lightweight deep receiver construction methods.

[0035] The beneficial effects of this application are as follows: A lightweight deep receiver construction method is proposed, which involves constructing a MU-MIMO OFDM uplink transmission system based on non-orthogonal superimposed pilot transmission to obtain the original training dataset; constructing a lightweight deep receiver, wherein the lightweight deep receiver is based on time-domain shifting and frequency-domain shifting operations; and using the original training dataset to perform multiple iterative model training and convergence on the lightweight deep receiver. By designing a MU-MIMO OFDM uplink transmission system based on non-orthogonal superimposed pilot transmission, the spectrum utilization is enhanced, and a reliable training dataset is obtained. Then, by constructing a lightweight deep receiver based on time-domain shifting and frequency-domain shifting for model training, the shifting operation effectively strengthens the separability of the two due to the stable pilot structure and the randomness of the data symbols. This enables the lightweight deep receiver to achieve decoupling of data and pilots in non-orthogonal superimposed pilot scenarios, significantly improving the decoding accuracy of the data stream. Meanwhile, this lightweight depth receiver does not introduce any new learnable parameters in spatial interaction, effectively reducing computational complexity and storage overhead, and avoiding the complex collaborative design of multiple independent modules in traditional solutions. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating an embodiment of the lightweight depth receiver construction method of this application;

[0037] Figure 2 This is a flowchart illustrating yet another embodiment of the lightweight depth receiver construction method of this application.

[0038] Figure 3 This is a schematic diagram of the architecture of an embodiment of the MU-MIMO OFDM uplink transmission system based on non-orthogonal superimposed pilot transmission according to this application;

[0039] Figure 4 This is a schematic diagram of the training architecture of an embodiment of the lightweight depth receiver of this application;

[0040] Figure 5 This is a schematic diagram comparing the BER performance of the lightweight depth receiver of this application with that of different receiver solutions.

[0041] Figure 6 This is a schematic diagram comparing the throughput performance of the lightweight depth receiver of this application with that of different receiver solutions.

[0042] Figure 7 A schematic diagram showing the BER performance comparison of the lightweight depth receiver of this application under different parameter settings in one embodiment;

[0043] Figure 8 This is a schematic diagram of the architecture of an embodiment of the adaptive depth receiver communication system for diverse channel scenarios according to this application.

[0044] Figure 9 This is a schematic diagram illustrating the BER performance of the channel adaptive deep receiver metamodel in this application under the RMa channel scenario;

[0045] Figure 10 This is a schematic diagram of the training parameter relationship of the adaptive depth receiver under diverse channel scenarios in this application;

[0046] Figure 11 This is a schematic diagram of the framework of an embodiment of the electronic device provided in this application;

[0047] Figure 12 This is a schematic diagram of a framework of an embodiment of the computer-readable storage medium provided in this application. Detailed Implementation

[0048] To facilitate understanding of this application, a more detailed description is provided below with reference to the accompanying drawings and specific embodiments. Preferred embodiments of this application are shown in the drawings. However, this application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of this application.

[0049] It should be noted that, unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. For example, the term "a plurality of" includes two or more.

[0050] The following are explanations of some terms used in this application:

[0051] 3GPP: Third Generation Partnership Project

[0052] UMa: Urban Macrocell

[0053] MU: Multi-user

[0054] MU-MIMO: Multiple User Multiple Input Multiple Output

[0055] OFDM: Orthogonal Frequency Division Multiplexing

[0056] MIMO: Multiple Input Multiple Output

[0057] NOSIP: Non-orthogonal superposition pilot

[0058] OP: Orthogonal Pilot

[0059] Goodput: Actual throughput

[0060] Sionna: An open-source simulation library for deep learning in physical layer communication.

[0061] BER: Bit Error Rate

[0062] QAM: Quadrature Amplitude Modulation

[0063] BPSK: Binary Phase Shift Keying

[0064] AWGN: Additive White Gaussian Noise

[0065] As described in the background section, traditional receivers are modular and have their own limitations.

[0066] In recent years, some deep learning models have emerged that directly learn the entire OFDM receiver processing flow from input to output, no longer relying on manually designed modular processes.

[0067] Specifically, in this field, existing research has demonstrated the feasibility of this approach:

[0068] Method 1: The application of deep learning in OFDM systems was studied. The channel estimation and signal detection performance was improved by training deep neural networks end-to-end. It performed particularly well in complex nonlinear channels and showed broad communication potential.

[0069] Method 2 is a fully convolutional deep learning receiver that eliminates the need for explicit channel estimation, performing end-to-end optimization directly from the received signal to the decoded data. This method adapts to complex channel conditions, significantly improving communication performance and robustness.

[0070] Method 3 is an end-to-end deep learning method that directly replaces the traditional receiver and channel estimation module with a neural network, reducing pilot overhead and improving communication efficiency and system robustness.

[0071] Method 4: A lightweight deep receiver method based on multi-head self-attention mechanism is proposed to achieve accurate mapping of received signal to log-likelihood ratio in multi-layer non-orthogonal superimposed pilot transmission scenarios, so as to improve communication efficiency by reducing pilot overhead.

[0072] While the lightweight deep receiver (SANet) in Method 4 achieves excellent performance in feature representation and semantic modeling through the multi-head self-attention (MHSA) mechanism, the computational and memory requirements of MHSA increase quadratically with the scale of time-frequency resources, making it difficult to deploy in large-scale MIMO, broadband systems, or hardware-constrained scenarios. Therefore, existing technologies have significant shortcomings in terms of decoding performance, transmission efficiency, and computational complexity.

[0073] Based on the aforementioned problems, the inventors recognized the urgent need for a lightweight, robust, and applicable method for a lightweight deep receiver suitable for non-orthogonal superimposed pilot scenarios. To this end, this application proposes a multi-user multiple-input multiple-output (MU-MIMO) orthogonal frequency division multiplexing (OFDM) system based on a non-orthogonal superimposed pilot transmission scheme, and a lightweight deep receiver construction method based on shift operation (hereinafter referred to as ShiftRx). This method replaces the highly complex attention module with an efficient spatial feature interaction mechanism, achieving accurate decoding of the superimposed data stream while effectively reducing computational overhead and improving system transmission efficiency.

[0074] The following are specific examples:

[0075] On the one hand, refer to Figure 1 As shown, this application provides a lightweight deep receiver construction method, including the following steps:

[0076] S11: Construct a MU-MIMO OFDM uplink transmission system based on non-orthogonal superimposed pilots to obtain the original training dataset;

[0077] S12: Construct a lightweight depth receiver, wherein the lightweight depth receiver is based on time-domain shifting and frequency-domain shifting operations;

[0078] S13: Iterate the lightweight deep receiver multiple times using the original training dataset until the model converges.

[0079] The construction of a MU-MIMO OFDM uplink transmission system based on non-orthogonal superimposed pilots includes:

[0080] The system's transmission resources are configured for use in the frequency domain. Subcarriers, used in the time domain A symbol, to achieve Data transmission for each user;

[0081] At the transmitting end, the data symbols and pilot symbols for each user on each resource unit are non-orthogonally superimposed to obtain superimposed symbols, i.e.:

[0082] ,

[0083] in, Indicates the first Data symbol matrix of each user, Indicates the first Pilot symbol matrix for each user, Indicates the first Weight matrix of data symbols for each user, Index for users;

[0084] Receiver, base station deployment Root receiving antenna, the first The received signals from each receiving antenna are:

[0085] ,

[0086] in, For the first The user to the The channel matrix of the root receiving antenna, It is additive white Gaussian noise with element variance of... ;

[0087] Will Received signal from root receiving antenna After cascading, the total received signal of the base station is formed. .

[0088] For details, please refer to Figure 3 The diagram shows the architecture of a MU-MIMO OFDM uplink transmission system. The Channel represents the communication channel, with the transmitter on the left and the receiver on the right. The diagram illustrates a typical... Each User Equipment (UE) transmitter, to Send information bit streams separately to After modulation and other processing, and with the addition of pilot signals, the signal is transmitted to the air interface. At the base station receiving end, it is then transmitted through the pilot symbols. Received signal and noise intensity Decode to achieve deep reception and acquire After performing a hard decision, the estimated bitstream is obtained. The process.

[0089] Specifically, before wireless channel transmission, for each transmitting user... ( ), which in each data resource unit ( , Transmit a bit vector on ) . The bit vectors of each user are mapped to OFDM resource units and rearranged into... Next, each bit vector By using Gray markers -QAM modulation is mapped to complex baseband symbols Therefore, we obtain the first... Baseband symbol OFDM resource raster for each user :

[0090]

[0091] To achieve accurate bit-level data reception, the transmitter generates a pilot matrix. It typically consists of a zero-mean pseudo-random BPSK sequence. In the uplink transmission system based on non-orthogonal superimposed pilots of this application, to further improve spectrum utilization efficiency, instead of using the general block-style superimposed pilot design method, the pilot and data signals are multiplexed on the same resource unit, that is, superimposed transmission is performed on each time-domain symbol and each frequency-domain subcarrier, which can be expressed as...

[0092]

[0093] in, Indicates the first The data weights for each user. Let the average power allocation ratio be... Then the normalized power constraint is

[0094]

[0095] This ensures that the total transmit power is normalized, that is .

[0096] During uplink transmission, the superimposed signals undergo multipath fading before reaching the base station. The received signal at each receiving antenna is

[0097]

[0098] Among them, symbols Represents the Hadama product. For the first The user to the The composite channel gain of the root receiving antenna, It is additive white Gaussian noise (AWGN) with element variance of . All received signals After cascading, the base stations form a total received signal. This is used for subsequent signal decoding.

[0099] To ensure the quality of each user's transmitted signal, the peak power of its time-domain waveform was further limited to avoid power saturation. The design also needs to ensure that the transmit power of each user is normalized in each time domain symbol, that is...

[0100] .

[0101] Ensuring the total transmit power is normalized reduces interference between different users. Furthermore, it ensures that the transmit power of each user is normalized in each time domain symbol, thereby optimizing and balancing the transmit signal quality of individual users and the relative received channel strength among multiple users, thus improving the capacity of the entire uplink transmission system.

[0102] The original training dataset includes the total received signal. Total pilot matrix and noise intensity Lightweight depth receivers use raw information bit tensors Used as labeled data for training, with log-likelihood ratio For output; where, Indicates the modulation order.

[0103] The construction of a lightweight deep receiver includes the following steps: concatenating and reshaping the dimensions of the real and imaginary parts of the original training dataset to adjust the input shape;

[0104] The first convolutional layer maps the channel dimension to a high-dimensional feature space and performs layer normalization to obtain feature data.

[0105] The feature data is processed through a multi-layer block shift decoder to achieve implicit mapping for channel estimation and channel equalization.

[0106] The output of the multi-layer block shift decoder is processed by layer normalization and a second convolutional layer to map the log-likelihood ratio of each resource unit.

[0107] Among them, reference Figure 2 As shown, the block shift decoder includes the following steps:

[0108] S121: Group the feature data along the channel dimension;

[0109] S122: Shift the first and second groups of feature data in the time domain, shift the third and fourth groups of feature data in the frequency domain, and keep the other groups of feature data unchanged;

[0110] S123: The shifted feature data is reorganized and refined through layer normalization and multilayer perceptron, and then fused through residual connections.

[0111] For details, please refer to Figure 4 First, at the base station, the total received signal from the original dataset... Total pilot matrix The real and imaginary dimensions are concatenated to form the received signal tensor. and pilot tensor .

[0112] Furthermore, a lightweight depth receiver is employed. To process three types of inputs: received signal tensor pilot tensor and noise intensity .in, These are learnable parameters for the lightweight deep receiver. By using noise intensity as an input feature, the network model can acquire prior knowledge of the signal-to-noise ratio (SNR) and dynamically adjust its internal processing strategy. This enhances the generalization ability and robustness of the lightweight deep receiver, while also simplifying the training process and improving convergence.

[0113] The three inputs are concatenated and reshaped along the "channel" dimension, and then rearranged into a unified feature tensor. Among them, the temporal symbol dimension With frequency domain subcarrier dimension The "height" and "width" are considered as a two-dimensional feature map. This representation allows us to use a two-dimensional shift mechanism to jointly model the correlation between the time and frequency domains, thereby efficiently extracting multipath and interference structures.

[0114] rearranged feature tensor First, after a first The convolutional layer (Conv2D) maps the channel dimension to a higher-dimensional embedding space, with the mapped dimension being of size . Then, layer normalization is performed, and the output feature assumption is as follows. .

[0115] Then a set of block shift decoders (Shift Blocks) are stacked. The layered Shift Block decoder implicitly maps channel estimation and channel equalization. The Shift Block decoder is the core of the lightweight deep receiver in this application. By introducing a Shift Block decoder based on time-domain and frequency-domain shift operations into the lightweight deep receiver scheme, the model can observe superimposed signals under different time-frequency offset conditions, thereby learning the time-frequency invariant structural characteristics of the pilots. Since the pilot structure is stable while the data symbols are random, the shift operation effectively enhances their separability, enabling the lightweight deep receiver to decouple data and pilots in non-orthogonal superimposed pilot scenarios.

[0116] Each Shift Block decoder includes:

[0117] Channel grouping: Input feature tensor It is uniformly divided along the channel dimension into The unit tensor, denoted as The dimensions of each group satisfy .

[0118] Time-domain shift: By appropriately shifting the signal in time, the receiver becomes more robust to different timing deviations.

[0119]

[0120] Frequency domain shift: By appropriately shifting the subcarriers, the model can learn richer features under different frequency domain conditions, thereby improving decoding robustness and mitigating the impact of frequency-selective fading.

[0121]

[0122] The other groups remain unchanged to preserve the original spatial structure, thereby ensuring the integrity of the information of the entire feature.

[0123] By shifting data in the time and frequency domains, we can not only improve the diversity of data, but also simulate the characteristics of Doppler shift and multipath fading during the transmission process of communication systems.

[0124] Layer normalization and multilayer perceptrons: These are used to reshape and refine shifted features. Layer normalization improves feature stability, and nonlinear mapping enhances feature representation, thus fully leveraging the spatial interaction effects brought about by the shift operation. The expansion factor of a multilayer perceptron is defined as... A multilayer perceptron consists of two linear layers: an expansion layer and a contraction layer. The expansion layer expands the feature dimension from... Mapped to The contraction layer then reduces the feature dimension from Mapped to The outputs of these components are then fused through residual connections, which preserves low-level spatiotemporal information while enhancing high-order dependencies.

[0125] Finally, the aggregated feature map is subjected to layer normalization again, and then through a second... The convolutional layer (Conv2D) is processed, and then rearranged to obtain the decoded data for each user on each resource unit. The log-likelihood ratios (LLRs) of bits are denoted as .

[0126] Again Apply hard decision function Obtain the estimated bit stream :

[0127]

[0128] in,

[0129]

[0130] The lightweight depth receiver uses binary cross-entropy loss to measure the prediction. The error between the actual bits B is represented by the loss function:

[0131]

[0132] in, For batch sizes of lightweight depth receivers, This is the Sigmoid function.

[0133] Compared with multi-head self-attention-based architectures, the lightweight deep receiver ShiftRx proposed in this application does not introduce any new learnable parameters in spatial interactions, thus achieving comparable modeling capabilities with significantly lower computational and memory overhead, providing a practical and efficient solution.

[0134] Meanwhile, in order to improve spectrum utilization, the non-orthogonal superposition pilot scenario in this application does not adopt the general block superposition pilot design method, but superimposes data symbols and pilot symbols on each resource unit; in addition, in terms of power design, it ensures that the power of each user on each time domain symbol is normalized, so that the data characteristics are not disrupted during the convolution process in the lightweight receiver.

[0135] In summary, in the aforementioned non-orthogonal superposition pilot scenario, this lightweight receiver not only significantly improves the decoding accuracy of the data stream but also effectively reduces computational complexity and storage overhead, avoiding the complex collaborative design of multiple independent modules in traditional solutions. Furthermore, its bit error rate performance has a clear advantage over traditional receiver solutions.

[0136] Furthermore, in order to enable the aforementioned lightweight deep receiver to adapt to diverse channel environments, the training dataset was constructed to cover multiple channel scenarios.

[0137] like Figure 8 As shown, let the first... The dataset corresponding to each channel scenario is as follows: This includes the total received signal generated under that specific channel condition. Total pilot matrix Noise variance and the original information bit tensor as the label data The comprehensive training dataset for multiple channel scenarios is represented as follows:

[0138]

[0139] This dataset enables receivers to learn from the total received signal in different wireless environments. Nonlinear mapping strategy to LLR.

[0140] Taking the Urban Macro (UMA) channel scenario as an example, the construction process of the comprehensive training dataset is as follows: First, data symbols, pilot symbols, and superimposed symbols for each user are generated according to the system configuration; then, the superimposed symbol data is propagated through the UMA channel to obtain the corresponding total received signal at the base station. Simultaneously, the corresponding total pilot signal matrix is ​​collected. Precise bit tag data, i.e., the original information bit tensor. and noise variance This constitutes the sample quadruple. and aggregated into a dataset By repeating the above process for other channel scenarios (such as...) Ultimately, a comprehensive training dataset covering various wireless environments can be obtained, enabling the lightweight deep receiver of this application to have adaptive operation capabilities under different channel conditions.

[0141] In modern wireless communication systems, enabling deep receivers to efficiently adapt to time-varying channel conditions is crucial for ensuring communication robustness and reliability. This paper addresses several representative channel scenarios. (e.g., urban macrocells (UMa), urban microcells (UMi), and rural macrocells (RMa)) The optimization objective of a lightweight deep receiver with channel adaptation capabilities can be expressed as:

[0142] ;

[0143] in, Representing the channel scenario Loss function of depth receiver This indicates the batch size during the metamodel update phase. Represents the dataset A small batch of channel sample quadruples is sampled. This represents the expected loss of the sampled data. The optimization objective is to learn a global set of initial values ​​for meta-parameters, enabling the model to generalize well in various channel environments and quickly adapt to new or unknown channel scenarios.

[0144] Traditional transfer learning methods typically assume that the source and target domains (e.g., different channel scenarios) have similar data distributions. However, in real-world wireless communication systems, different channel models (such as RMa, UMa, and UMi) exhibit significant differences in delay spread, path loss, and angular spread. For instance, in the UMi scenario, delay spread is large, and the Doppler effect may be high, while in the RMa scenario, delay spread is small, and the Doppler effect is generally low. These differences cause the statistical characteristics of the received signals in different channel models to no longer be independent and identically distributed. This distribution shift often leads to feature shift and negative transfer problems. Furthermore, although meta-learning methods based on second-order gradients (such as Model-Independent Meta-Learning (MAML)) can facilitate rapid adaptation between tasks, their dependence on higher-order derivatives introduces huge computational and memory overhead, making them difficult to implement in resource-constrained semantic receiver architectures.

[0145] Furthermore, to address the aforementioned issues, this application employs a first-order meta-learning strategy based on the Reptile algorithm, which significantly reduces training complexity while achieving efficient and rapid adaptation. Within this framework, each channel scenario... Treated as an independent "task," the global meta-model parameters are continuously updated through an iterative process across multiple tasks. .

[0146] The overall training process includes two main phases:

[0147] 1) Specific task updates in single-channel scenarios:

[0148] For each task First, copy the current metamodel parameters to initialize the model corresponding to this task. Then, from the corresponding dataset A small batch of channel samples were sampled. Next, use the task-specific learning rate. The Adam optimizer is used to update the model for this specific channel scenario in several steps:

[0149]

[0150] in, This represents the model parameters at the start of the i-th task. This represents the loss for the i-th task with respect to the parameters. gradient, This represents the parameters after a gradient update for the i-th task (task-adaptive parameters).

[0151] 2) Meta-model update in multi-channel scenarios:

[0152] After completing specific updates to the models for each channel scenario, the meta-model parameters will be adjusted in the direction of parameters adapted to the task, as follows:

[0153]

[0154] in, The learning rate is the value of the meta-model.

[0155] like Figure 10 As shown, for example, there are three channel scenarios, corresponding to three tasks, and the initial parameters of the meta-model are... After several updates, Updated to At this point, the gradient direction of the meta-model parameters is closest to the gradient direction of each task parameter, meaning an optimal solution for meta-model parameters that can adapt to three channel scenarios has been obtained. The final updated parameters are still denoted as... ,Right now .

[0156] In summary, by iteratively training on multiple channel tasks, the resulting deep receiver can be trained with only minor fine-tuning (using the learning rate). This allows for rapid adaptation to new channel conditions, thereby significantly reducing model retraining costs and enhancing the system's portability and practical application value.

[0157] The detailed training process is shown in Algorithm 1 below:

[0158]

[0159] In summary, by constructing a comprehensive training dataset covering multiple channel scenarios and further optimizing the training process of the aforementioned lightweight deep receiver through two stages of specific task updates and meta-model updates, the lightweight deep receiver gains channel adaptability, thereby exhibiting excellent decoding performance under various channel scenarios.

[0160] Specifically, this application implements a MU-MIMO OFDM system using the UMa channel model from the Sionna open-source library to evaluate the throughput gain of a lightweight deep receiver and a non-orthogonal superposition pilot transmission scheme in a communication system.

[0161] The detailed parameter settings of the communication system are shown in Table 1.

[0162] Table 1 Communication System Parameter Settings

[0163]

[0164] The settings for network parameters and training parameters are shown in Table 2.

[0165] Table 2 Network and training parameter settings

[0166]

[0167] Figure 5 The bit error rate (BER) performance of the proposed lightweight deep receiver (Proposed ShiftRx - NOSIP) was compared with that of various traditional and deep learning receivers, including the "Perfect Channel State Information (Perfect CSI) + Linear Minimum Mean Square Error (LMMSE) - OP" receiver, the "Least Squares (LS) + Linear Minimum Mean Square Error (LMMSE) - OP" receiver, the receiver in Method 2 (MLP - NOSIP), the receiver based on residual networks in Method 3 (ResNet - NOSIP), and the receiver based on Transformers in Method 4 (SANet - NOSIP). Some traditional receiving algorithms employ the OP scheme because they cannot decode signals under the NOSIP scheme.

[0168] In contrast, while the ShiftRx proposed in this application operates under the more spectrally efficient NOSIP scheme, its BER is still comparable to that of a "Perfect CSI + LMMSE" receiver. Furthermore, existing lightweight depth receivers based on MLP and ResNet struggle with effective decoding under NOSIP conditions, resulting in a higher BER across the entire SNR range. SANet, which utilizes the multi-head self-attention (MHSA) mechanism, improves decoding accuracy compared to MLP and ResNet, but still lags significantly behind ShiftRx proposed in this application.

[0169] To evaluate the performance improvement of NOSIP and the proposed lightweight depth receiver, in Figure 6 The paper presents the effective throughput performance of various receiver schemes under different SNR conditions. The effective system throughput is used as the metric, defined as follows:

[0170]

[0171] Among them, factor This indicates the proportion of data symbols on a resource grid. Specifically, this applies to NOSIP and pilotless scenarios. In the orthogonal pilot (OP) configuration .

[0172] It is evident that the traditional "LS + LMMSE - OP" falls below the baseline of "perfect CSI + LMMSE - OP" due to channel estimation errors and pilot overhead. The curve labeled "perfect CSI + LMMSE - w / o pilots" can be considered an ideal upper bound under the assumption of perfect channel knowledge and no consumption of pilot resources.

[0173] In comparison, the ShiftRx using NOSIP proposed in this application achieves the highest effective throughput among all receivers, even surpassing the "perfect CSI + LMMSE - OP" receiver using orthogonal pilots. This gain is mainly due to the NOSIP transmission mechanism: superimposing pilot and data signals within the same resource unit, thereby improving spectral efficiency. In the low SNR range ( The proposed ShiftRx improves the average effective throughput by approximately 1.16% compared to SANet (2 to 3 dB); and improves it by approximately 25.14% compared to the traditional "LS + LMMSE - OP" receiver across the entire SNR range. The results demonstrate that the proposed ShiftRx not only maintains reliable decoding under NOSIP but also significantly improves throughput and spectral efficiency compared to traditional designs.

[0174] Figure 7 The BER performance of the proposed ShiftRx under different parameter configurations is presented. "w / o TD shift" and "w / o FD shift" represent the cases where the time-domain shift and frequency-domain shift mechanisms are removed, respectively. As can be seen from the figure, increasing... or Both can effectively enhance feature extraction capabilities and improve BER performance. Especially when , At that time, ShiftRx achieved the lowest BER under full SNR conditions, demonstrating superior learning and generalization capabilities. Removing either the time-domain or frequency-domain shift mechanism resulted in significant performance degradation, indicating that the two mechanisms complement each other in suppressing multipath fading and frequency-selective interference. These results validate the effectiveness of the proposed dual-domain shift mechanism in improving signal decoding robustness.

[0175] Table 3 in , , Under certain configurations, the model complexity and memory overhead of SANet based on MHSA and ShiftRx proposed in this application were compared. The results show that ShiftRx reduces FLOPs by 25.4%, parameter count by 16.5%, GPU memory usage by 82.6%, and model size by 16.4%. SANet based on MHSA exhibits quadratic complexity in terms of computation and memory with respect to sequence length, making it highly sensitive to increases in time-frequency resources. In contrast, ShiftRx replaces global attention with a lightweight time-domain and frequency-domain shifting mechanism, making the complexity approximately linear. This design effectively alleviates the "exponential" growth in memory usage, providing better scalability and practicality in large-scale or real-time communication systems.

[0176] Table 3 Comparison of parameter count and complexity between ShiftRx and SANet in this application

[0177]

[0178] FLOPs measures the computational complexity of an algorithm or model. Params measures the number of parameters in a model. GPU Memory measures the amount of video memory used during model training or testing. Model Size represents the storage space of the model file.

[0179] Figure 9 illustrates the bit error rate performance of the proposed ShiftRx-based channel adaptive deep receiver under different channel environments. Here, "ShiftRx - NOSIP" indicates a model trained directly on a single task without meta-learning. This represents the number of iterations of meta-training during the outer-layer optimization process. The "ShiftRx - NOSIP - fine-tuned" curve in the figure represents the performance after further adapting the model to the target channel through a small number of inner-layer updates, based on the meta-learned model. It can be seen that under certain SNR conditions, the performance is better after meta-training and a small number of inner-layer updates to adapt to the target channel.

[0180] On the other hand, a lightweight depth receiver constructed using any of the above methods is provided, which is used to achieve accurate decoding of data streams in non-orthogonal superposition pilot transmission schemes.

[0181] Furthermore, a lightweight depth receiver system is provided, the system comprising:

[0182] A transmitter is used to encode and transmit data streams in a non-orthogonal superposition pilot transmission scheme.

[0183] The aforementioned lightweight depth receiver is used to receive data transmitted by the transmitter and accurately decode the transmitted data.

[0184] In another aspect, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the above methods.

[0185] Specifically, please refer to Figure 11 The electronic device described in this application embodiment may specifically include a processor 210 and a memory 220. The memory 220 is coupled to the processor 210.

[0186] Processor 210 is used to control the operation of electronic devices. Processor 210 can also be referred to as a CPU (Central Processing Unit). Processor 210 may be an integrated circuit chip with signal processing capabilities. Processor 210 can also be a general-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor, or processor 210 can be any conventional processor.

[0187] The memory 220 is used to store computer programs and may be RAM, ROM, or other types of storage devices. Specifically, the memory may include one or more computer-readable storage media, which may be non-transitory. The memory may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory is used to store at least one line of program code.

[0188] The processor 210 is used to execute computer programs stored in the memory 220 to implement the methods described in the various method embodiments of this application.

[0189] In some embodiments, the electronic device may further include a peripheral device interface 230 and at least one peripheral device. The processor 210, memory 220, and peripheral device interface 230 can be connected via a bus or signal line. Each peripheral device can be connected to the peripheral device interface 230 via a bus, signal line, or circuit board. Specifically, the peripheral device includes at least one of a radio frequency circuit 240, a display screen 250, an audio circuit 260, and a power supply 270.

[0190] Peripheral device interface 230 can be used to connect at least one I / O (Input / output) related peripheral device to processor 210 and memory 220. In some embodiments, processor 210, memory 220 and peripheral device interface 230 are integrated on the same chip or circuit board; in some other embodiments, any one or two of processor 210, memory 220 and peripheral device interface 230 can be implemented on separate chips or circuit boards, which is not limited in this embodiment.

[0191] The radio frequency (RF) circuit 240 is used to receive and transmit RF (Radio Frequency) signals, also known as electromagnetic signals. The RF circuit 240 communicates with communication networks and other communication devices via electromagnetic signals; therefore, the RF circuit 240 is the communication circuit of the electronic device. The RF circuit 240 converts electrical signals into electromagnetic signals for transmission, or converts received electromagnetic signals back into electrical signals. Optionally, the RF circuit 240 includes: an antenna system, an RF transceiver, one or more amplifiers, a tuner, an oscillator, a digital signal processor, a codec chipset, a user identity module card, etc. The RF circuit 240 can communicate with other terminals through at least one wireless communication protocol. This wireless communication protocol includes, but is not limited to: the World Wide Web, metropolitan area networks, intranets, various generations of mobile communication networks (2G, 3G, 4G, and 5G), wireless local area networks, and / or WiFi (Wireless Fidelity) networks. In some embodiments, the RF circuit 240 may also include circuitry related to NFC (Near Field Communication), which is not limited in this application.

[0192] Display screen 250 is used to display a UI (User Interface). This UI may include graphics, text, icons, videos, and any combination thereof. When display screen 250 is a touch display screen, it also has the ability to collect touch signals on or above its surface. These touch signals can be input as control signals to processor 210 for processing. In this case, display screen 250 can also be used to provide virtual buttons and / or a virtual keyboard, also known as soft buttons and / or a soft keyboard. In some embodiments, there may be one display screen 250, located on the front panel of the electronic device; in other embodiments, there may be at least two display screens, located on different surfaces of the electronic device or in a folded design; in still other embodiments, display screen 250 may be a flexible display screen, located on a curved or folded surface of the electronic device. Furthermore, display screen 250 may be configured as a non-rectangular, irregular shape, i.e., a non-rectangular screen. Display screen 250 may employ LCD (Liquid Crystal Display) or OLED (Organic Light Source) displays. Materials such as organic light-emitting diodes (OLEDs) are used to fabricate these materials.

[0193] The audio circuit 260 may include a microphone and a speaker. The microphone is used to collect sound waves from the user and the environment, converting the sound waves into electrical signals that are input to the processor 210 for processing, or input to the radio frequency circuit 240 for voice communication. For stereo sound acquisition or noise reduction purposes, multiple microphones may be used, each located in a different part of the electronic device. The microphone may also be an array microphone or an omnidirectional microphone. The speaker is used to convert electrical signals from the processor 210 or the radio frequency circuit 240 into sound waves. The speaker may be a conventional diaphragm speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, it can convert electrical signals not only into audible sound waves but also into inaudible sound waves for purposes such as distance measurement. In some embodiments, the audio circuit 260 may also include a headphone jack.

[0194] Power supply 270 is used to supply power to various components in an electronic device. Power supply 270 can be alternating current, direct current, a disposable battery, or a rechargeable battery. When power supply 270 includes a rechargeable battery, the rechargeable battery can be a wired rechargeable battery or a wireless rechargeable battery. A wired rechargeable battery is a battery that is charged via a wired line, while a wireless rechargeable battery is a battery that is charged via a wireless coil. The rechargeable battery can also be used to support fast charging technology.

[0195] For a detailed description of the functions and execution processes of each functional module or component in the embodiments of the intelligent control platform device of this application, please refer to the descriptions in the above-mentioned method embodiments of this application, which will not be repeated here.

[0196] In the embodiments provided in this application, it should be understood that the disclosed intelligent control platform device and method can be implemented in other ways. For example, the embodiments of the intelligent control platform device described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0197] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0198] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0199] On the other hand, a readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements any of the above-mentioned lightweight deep receiver construction methods.

[0200] Please see Figure 12If the integrated units described above are implemented as software functional units and sold or used as independent products, they can be stored in computer-readable storage medium 300. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions / computer programs to cause an intelligent control platform device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of this application. The aforementioned storage medium includes various media such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, as well as electronic devices such as computers, mobile phones, laptops, tablets, and cameras that have the aforementioned storage media.

[0201] The execution process of program data in a computer-readable storage medium can be described with reference to the above-described method embodiments of this application, and will not be repeated here.

[0202] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0203] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0204] In summary, this application has the following beneficial effects:

[0205] First, a multi-user multiple-input multiple-output orthogonal frequency division multiplexing uplink system based on a non-orthogonal superposition pilot transmission scheme is designed and constructed to generate the original training and test datasets.

[0206] Then, a lightweight depth receiver based on time-domain and frequency-domain shift operations is constructed. Time-domain shifting, by appropriately shifting the signal in time, makes the receiver more robust to different timing deviations; while frequency-domain shifting, by appropriately shifting the subcarrier, allows the model to learn richer features under different frequency domain conditions, thereby improving decoding robustness and mitigating the impact of frequency-selective fading. By introducing time-domain and frequency-domain shifting into the lightweight depth receiver scheme, the model can observe superimposed signals under different time-frequency offset conditions, thereby learning the time-frequency invariant structural features of the pilot. Since the pilot structure is stable while the data symbols are random, the shift operation effectively enhances their separability, enabling the lightweight depth receiver to decouple data and pilots in non-orthogonal superimposed pilot scenarios.

[0207] Finally, the lightweight deep receiver is trained iteratively multiple times using the generated training dataset until the model converges. In non-orthogonal pilot signal scenarios, this lightweight receiver not only significantly improves the decoding accuracy of the data stream but also effectively reduces computational complexity and storage overhead, avoiding the complex collaborative design of multiple independent modules in traditional schemes. Furthermore, its bit error rate performance has a significant advantage over traditional receiver schemes.

[0208] Furthermore, by constructing a comprehensive training dataset covering multiple channel scenarios and optimizing the training process of the aforementioned lightweight deep receiver through two stages of specific task updates and meta-model updates, the lightweight deep receiver gains channel adaptability, thereby exhibiting excellent decoding performance under various channel scenarios.

Claims

1. A lightweight depth receiver construction method, characterized in that, Including the following steps: A MU-MIMO OFDM uplink transmission system based on non-orthogonal superimposed pilots was constructed to obtain the original training dataset; The original training dataset includes sample quadruples. ;in, Indicates the total received signal. Represents the total pilot matrix, Indicates noise intensity. Represents the original information bit tensor. The modulation order is represented by Nr, the number of receiving antennas is represented by T, the number of OFDM symbols is represented by S, the number of subcarriers is represented by K, and the number of users is represented by K. Build a lightweight depth receiver; The construction of the lightweight depth receiver includes the following steps: The original training dataset is subjected to dimension concatenation and dimension reshaping of its real and imaginary parts to adjust the input shape. The first convolutional layer maps the channel dimension to a high-dimensional feature space and performs layer normalization to obtain feature data. The feature data is processed by a multi-layer block shift decoder to achieve implicit mapping of channel estimation and channel equalization. The block shift decoder is based on time-domain shift and frequency-domain shift operations. The output of the multi-layer block shift decoder is processed by layer normalization and a second convolutional layer to map the log-likelihood ratio of each resource unit. The lightweight deep receiver is iterated multiple times using the original training dataset until the model converges.

2. The lightweight depth receiver construction method according to claim 1, characterized in that, The block shift decoder includes the following steps: Group the feature data along the channel dimension; The first and second groups of feature data are shifted in the time domain, the third and fourth groups of feature data are shifted in the frequency domain, and the feature data of other groups remain unchanged. The shifted feature data is reorganized and refined through layer normalization and multilayer perceptron, and then fused through residual connections.

3. The lightweight depth receiver construction method according to claim 1, characterized in that, The original training dataset was constructed into a comprehensive training dataset covering multiple different channel scenarios. ， in, Indicates the first The aggregated dataset of the sample quadruplets corresponding to each channel scenario This refers to the comprehensive training dataset, which is the set of aggregated datasets.

4. The lightweight depth receiver construction method according to claim 3, characterized in that, The optimization objective of the lightweight depth receiver is: ； in, Representing the channel scenario Loss function of depth receiver This indicates the batch size during the metamodel update phase. Represents the dataset A small batch of channel sample quadruples is sampled. This represents the expected loss of the sampled data. This refers to the lightweight depth receiver. These are the learnable parameters of the lightweight depth receiver. All of these are input data from the lightweight depth receiver, where B is the input data of the receiver. The corresponding original information bit tensor is used as the tag data of the lightweight depth receiver; The iterative training of the lightweight depth receiver includes: In the first phase, the task parameters for each of the aforementioned channel scenarios are updated: ； in, Indicates the task learning rate; This represents the model parameters at the start of the i-th task. This represents the loss for the i-th task with respect to the parameters. gradient, This represents the parameters after a gradient update for the i-th task; In the second stage, the meta-model parameters for all the aforementioned channel scenarios are updated: ； in, This represents the learning rate of the meta-model.

5. A lightweight depth receiver system, characterized in that, The system includes: A transmitter is used to encode and transmit data streams in a non-orthogonal superposition pilot transmission scheme. A lightweight depth receiver constructed by any one of the methods described in claims 1-4 is used to receive the transmitted data from the transmitter and to accurately decode the transmitted data.

6. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the method as described in any one of claims 1-4.

7. A readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed by a processor, implements the lightweight depth receiver construction method as described in any one of claims 1-4.

Images