A modular OFDM receiving method and system based on deep learning, a computer device and a medium

Abstract

The application discloses a kind of based on deep learning's modular OFDM receiving method, system, computer equipment and medium, method includes: input data is organized as frequency domain complex matrix, frequency domain complex matrix includes seqLTF, demodLTF, demodData and seqPilot;SeqLTF and demodLTF are input to channel estimation module, and output channel estimation result;Channel estimation result, demodData and seqPilot are input to pilot phase tracking module, and output demodulated data field feature after phase correction;Channel estimation result and demodulated data field feature after phase correction are input to equalization demapping module, and output the log-likelihood ratio of each coding bit.The OFDM receiving procedure of the application is explicitly mapped as three neural network modules, tensor transferred between modules has explicit physical meaning, and enhances interpretability.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a modular OFDM receiving method, system, computer device, and medium based on deep learning. Background Technology

[0002] Orthogonal Frequency Division Multiplexing (OFDM) is a core physical layer transmission technology in modern wireless communication systems such as WiFi, 5G, and LTE. At the OFDM receiver, the signal undergoes key processing steps sequentially, including channel estimation, channel equalization, and demapping, ultimately outputting a log-likelihood ratio (LLR) to support subsequent channel decoding. The quality of decoding performance hinges on the accurate acquisition of Channel State Information (CSI) and the accurate calculation of the LLR value.

[0003] Traditional OFDM receiver schemes are based on classical signal processing theory and pilot-aided design, mainly employing algorithms such as Least Squares (LS), Minimum Mean Square Error (MMSE), or Linear Minimum Mean Square Error (LMMSE) to estimate CSI. However, traditional channel estimation algorithms have significant technical bottlenecks: the LS algorithm suffers a significant degradation in estimation accuracy under low signal-to-noise ratio (SNR) scenarios; while high-precision algorithms such as MMSE can improve estimation performance, they rely on complex prior channel statistics and have high computational complexity, making them difficult to match the actual needs of modern mobile communication terminals for low resource overhead.

[0004] Therefore, existing technologies still need improvement. Summary of the Invention

[0005] To address the aforementioned shortcomings of existing technologies, this invention provides a modular OFDM receiving method, system, computer device, and medium based on deep learning. The technical solution adopted by this invention is as follows: In a first aspect, the present invention provides a modular OFDM receiving method based on deep learning, the method comprising: The input data is organized into a frequency domain complex matrix, which includes seqLTF, demodLTF, demodData, and seqPilot. The seqLTF reflects the channel estimation standard LTF reference sequence specified by the protocol; the demodLTF reflects the frequency domain symbols obtained by demodulation at the receiver after the standard LTF reference sequence has been transmitted through the channel; the demodData reflects the frequency domain symbols obtained by demodulation at the receiver after the data field has been transmitted through the channel; and the seqPilot reflects the pilot symbol reference sequence defined by the protocol in the data field. The seqLTF and the demodLTF are input into the channel estimation module, and the channel estimation result is output. The channel estimation results, the demodData, and the seqPilot are input into the pilot phase tracking module, and the demodulated data field characteristics after phase correction are output. The channel estimation result and the demodulated data field features after phase correction are input into the equalization demapping module, which outputs the log-likelihood ratio of each coded bit to complete OFDM reception based on the log-likelihood ratio of each coded bit.

[0006] In one implementation, the channel estimation module includes an upsampling unit and three cascaded inverted residual blocks.

[0007] In one implementation, the seqLTF and the demodLTF are input to the channel estimation module, and the channel estimation result is output, including: The feature spaces of the seqLTF and the demodLTF are expanded by upsampling; Based on the inverted residual block, features are first extracted by increasing the dimensionality, then the correlation between subcarriers is captured by 3x3 convolution, and then fused by dimensionality reduction. The frequency domain statistical characteristics and noise distribution of the channel are implicitly learned, and the channel estimation results are output.

[0008] In one implementation, the pilot phase tracking module includes: a feature fusion block, a channel attention residual block, and a projection residual block.

[0009] In one implementation, the channel estimation result, the demodData, and the seqPilot are input to the pilot phase tracking module, and the demodulated data field characteristics after phase correction are output, including: Based on the feature fusion block, the channel estimation result, the demoData and the features of seqPilot are fused through the Concat operation, and the channel estimation result is used to assist phase tracking. By using channel attention residual blocks, all channel features are adaptively weighted to enhance effective features; By projecting residual blocks, high-dimensional fused features are projected onto low-dimensional features, optimizing feature representation and outputting demodulated data field features after phase correction.

[0010] In one implementation, the equalization demapping module includes: an inverted residual block, a projected residual block, and a fully connected layer.

[0011] In one implementation, the channel estimation result and the phase-corrected demodulated data field features are input to an equalization demapping module, which outputs the log-likelihood ratio for each coded bit, including: The Concat operation fuses the characteristics of the channel estimation results with those of the demodulated data field after phase correction, and uses the channel information from the channel estimation results to assist in equalization calculation. The inverted residual block is used to learn the equalization and demapping functions, and then the projected residual block is used for feature fusion and dimensionality reduction. Feature fitting is performed through a fully connected layer, and the log-likelihood ratio of each encoded bit is output.

[0012] In one implementation, the method further includes: Within a preset signal-to-noise ratio range, the complete channel transmission process is simulated to generate receiver frequency domain data, which is then paired with the transmitter coded bit sequence as coded bit labels to obtain a training dataset. Each training sample in the training dataset includes the frequency domain complex matrix. Using a random number seed independent of the training dataset, and following system parameters and channel configurations completely identical to those of the training dataset, test frequency domain data is generated to obtain the test dataset.

[0013] In one implementation, the method further includes: The model is trained based on the training dataset, and in each training iteration, the corresponding frequency domain complex matrix is ​​extracted from the training dataset and input into the channel estimation module, pilot phase tracking module and equalization demapping module in the model. The log-likelihood ratio of each encoded bit in the final output is compared with the encoded bit label in the training dataset to calculate the binary cross-entropy loss. By using the backpropagation algorithm, the parameters of the neural network model are updated, enabling the training and optimization of the channel estimation module, pilot phase tracking module, and equalization demapping module.

[0014] In one implementation, the method further includes: The test dataset is input into the trained and optimized channel estimation module, pilot phase tracking module, and equalization demapping module to obtain the test results; Performance evaluation is performed based on the test results to verify the performance of the channel estimation module, pilot phase tracking module, and equalization demapping module in the model.

[0015] In one implementation, performance evaluation based on the test results includes: Based on the test results, the bit error rate of the communication system under different signal-to-noise ratios during the test phase was statistically analyzed. Based on the test results, the model performance when transmitting data frames of different lengths is determined, and the model's adaptive processing capability for variable-length data frames in the WiFi protocol is verified.

[0016] Secondly, embodiments of the present invention also provide a modular OFDM receiving system based on deep learning, wherein the system is used to implement the steps of the modular OFDM receiving method based on deep learning described in any of the above-mentioned schemes, and the system includes: The frequency domain complex matrix determination module is used to organize the input data into a frequency domain complex matrix, which includes seqLTF, demodLTF, demodData, and seqPilot. The seqLTF reflects the channel estimation standard LTF reference sequence specified by the protocol; the demodLTF reflects the frequency domain symbols obtained by demodulation at the receiving end after the standard LTF reference sequence is transmitted through the channel; the demodData reflects the frequency domain symbols obtained by demodulation at the receiving end after the data field is transmitted through the channel; and the seqPilot reflects the pilot symbol reference sequence defined by the protocol in the data field. The channel estimation module is used to output channel estimation results based on the seqLTF and the demodLTF; The pilot phase tracking module is used to output the characteristics of the demodulated data field after phase correction based on the channel estimation results, the demodData, and seqPilot. The equalization demapping module is used to output the log-likelihood ratio of each coded bit based on the channel estimation result and the characteristics of the demodulated data field after phase correction, so as to complete OFDM reception based on the log-likelihood ratio of each coded bit.

[0017] In one implementation, the channel estimation module includes: an upsampling unit and three cascaded inverted residual blocks; The upsampling unit is used to expand the feature space of the seqLTF and the demodLTF; The inverted residual block of the channel estimation module is used to first extract features by increasing the dimensionality, then capture the correlation between subcarriers by 3x3 convolution, and then fuse them by reducing the dimensionality to implicitly learn the frequency domain statistical characteristics and noise distribution of the channel.

[0018] In one implementation, the pilot phase tracking module includes: a feature fusion block, a channel attention residual block, and a projection residual block.

[0019] The feature fusion block is used to fuse the channel estimation result, the demoData, and the features of seqPilot through the Concat operation, and to use the channel estimation result to assist phase tracking. The channel attention residual block is used to adaptively weight all channel features to enhance effective features; The projection residual block is used to project high-dimensional fused features to low dimensions, thereby optimizing feature representation.

[0020] In one implementation, the equalization demapping module includes an inverted residual block, a projected residual block, and a fully connected layer; The inverted residual block of the equalization demapping module is used to learn the equalization and demapping functions; The projection residual blocks of the equalization demapping module are used for feature fusion and dimensionality reduction; The fully connected layer is used for feature fitting and outputs the log-likelihood ratio for each encoded bit.

[0021] In one implementation, the system further includes: The training dataset construction module is used to simulate the complete channel transmission process within a preset signal-to-noise ratio range, generate receiver frequency domain data, and use the transmitter coded bit sequence as coded bit label to obtain the training dataset. Each training sample in the training dataset includes the frequency domain complex matrix. The test dataset building module is used to generate test frequency domain data using a random number seed that is independent of the training dataset, while following the same system parameters and channel configuration as the training dataset, thus obtaining the test dataset.

[0022] In one implementation, the system further includes a model training and optimization module, comprising: The training data input unit is used to train the model based on the training dataset, and in each training iteration, extracts the corresponding frequency domain complex matrix from the training dataset and inputs it into the channel estimation module, pilot phase tracking module and equalization demapping module in the model. The binary cross-entropy loss calculation unit is used to compare the log-likelihood ratio of each encoded bit in the final output with the encoded bit label in the training dataset to calculate the binary cross-entropy loss. The back gradient propagation unit is used to update the parameters of the neural network model through the back gradient propagation algorithm, thereby enabling the training and optimization of the channel estimation module, pilot phase tracking module, and equalization demapping module.

[0023] In one implementation, the method further includes a testing and performance evaluation module, comprising: The testing unit is used to input the test dataset into the trained and optimized channel estimation module, pilot phase tracking module, and equalization demapping module to obtain test results; The performance evaluation unit is used to evaluate the performance based on the test results to verify the performance of the channel estimation module, pilot phase tracking module and equalization demapping module in the model.

[0024] Thirdly, embodiments of the present invention also provide a computer device, wherein the computer device includes a memory, a processor, and a deep learning-based modular OFDM receiving program stored in the memory and executable on the processor. When the processor executes the deep learning-based modular OFDM receiving program, it implements the steps of the deep learning-based modular OFDM receiving method of any of the above-described schemes.

[0025] Fourthly, embodiments of the present invention also provide a computer-readable storage medium, wherein a deep learning-based modular OFDM receiver program is stored on the computer-readable storage medium, the deep learning-based modular OFDM receiver program implementing the steps of the deep learning-based modular OFDM receiver method described in any of the above schemes on the computer-readable storage medium.

[0026] Beneficial Effects: Compared with existing technologies, this invention provides a modular OFDM receiving method based on deep learning. First, the input data is organized into a frequency-domain complex matrix, including seqLTF, demodLTF, demodData, and seqPilot. The seqLTF reflects the channel estimation standard LTF reference sequence specified by the protocol; the demodLTF reflects the frequency-domain symbols obtained by demodulation at the receiver after the standard LTF reference sequence is transmitted through the channel; the demodData reflects the frequency-domain symbols obtained by demodulation at the receiver after the data field is transmitted through the channel; and the seqPilot reflects the pilot symbol reference sequence defined by the protocol in the data field. Then, the seqLTF and demodLTF are input to the channel estimation module, which outputs the channel estimation result. Next, the channel estimation result, the demodData, and the seqPilot are input to the pilot phase tracking module, which outputs the demodulated data field features after phase correction. Then, the channel estimation result and the demodulated data field features after phase correction are input into the equalization demapping module, and the log-likelihood ratio of each coded bit is output to complete OFDM reception based on the log-likelihood ratio of each coded bit.

[0027] This invention explicitly maps the OFDM receiving process to three neural network modules. The tensors passed between modules have clear physical meanings, enhancing interpretability. Furthermore, based on the characteristics of the WiFi protocol, this invention proposes an inference optimization strategy where the output of the channel estimation module can be reused within a frame. Channel estimation is performed only once, and the result is reused for all symbols within the frame, significantly reducing computational load. Attached Figure Description

[0028] Figure 1 This diagram illustrates a technical comparison between the deep learning-based modular OFDM receiving method provided in this embodiment of the invention and traditional methods.

[0029] Figure 2 This diagram illustrates the application environment of the modular OFDM receiving method based on deep learning provided in this embodiment of the invention.

[0030] Figure 3 This is a flowchart illustrating a preferred embodiment of the modular OFDM receiving method based on deep learning provided in this invention.

[0031] Figure 4 This is a schematic diagram of the neural network structure of each module in the modular OFDM receiving method based on deep learning provided in the embodiments of the present invention.

[0032] Figure 5 This is a schematic diagram of the inverted residual block in a modular OFDM receiver based on deep learning, provided in an embodiment of the present invention.

[0033] Figure 6 This is a schematic diagram of the channel attention residual block in a modular OFDM receiver based on deep learning, provided in an embodiment of the present invention.

[0034] Figure 7 for Figure 6 A schematic diagram of the SE block structure.

[0035] Figure 8 This is a schematic diagram of the projection residual block in the modular OFDM receiving method based on deep learning provided in an embodiment of the present invention.

[0036] Figure 9 A schematic diagram comparing the BER performance of transmitting a 4-symbol data frame under different signal-to-noise ratio conditions.

[0037] Figure 10 A schematic diagram comparing the BER performance of transmitting a 64-symbol data frame under different signal-to-noise ratio conditions.

[0038] Figure 11 This is a system block diagram of a modular OFDM receiver system based on deep learning, provided for an embodiment of the present invention.

[0039] Figure 12 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0041] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content, operations, or steps, nor does it require execution in the described order. For example, some operations or steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0042] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0043] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. For example, the first control information and the second control information are only used to distinguish different control information and do not limit their order.

[0044] Those skilled in the art will understand that the words "first" and "second" do not limit the quantity or the order of execution, and that the words "first" and "second" do not necessarily imply that they are different.

[0045] It should also be understood that the terms "and / or" as used in this specification and the appended claims refer to any combination of one or more of the associated listed items and all possible combinations, and include such combinations.

[0046] In recent years, deep learning (DL) has made groundbreaking progress in fields such as computer vision and natural language processing, and its application in wireless communication has also demonstrated significant technological potential. Related research mainly focuses on replacing some traditional processing modules in OFDM receivers with deep learning technology. Experiments have shown that its bit error rate (BER) performance is improved to varying degrees compared to traditional OFDM receiving schemes. Existing deep learning-based OFDM receiving technology schemes are as follows: End-to-end deep learning receiver: This scheme constructs an end-to-end architecture through a convolutional neural network (CNN) and residual connections, completing the mapping from the frequency domain received signal to the LLR output in one go, realizing joint learning of channel estimation, equalization and demapping operations, and achieving transmission performance superior to traditional methods in 5G application scenarios.

[0047] The OFDM receiving method based on Transformer: This scheme uses the multi-head attention mechanism of Transformer to learn the correlation between subcarriers. The baseband signal of the receiver is input into the OFDM receiver based on Transformer to complete channel estimation and signal detection, and the output binary bit sequence is used as the recovered transmission data.

[0048] A hybrid approach combining traditional methods and deep learning: This approach employs a serially cascaded architecture. First, an improved LMMSE algorithm is used to perform initial channel estimation on the received pilot signal. Then, a lightweight residual network model is used to perform global feature modeling on the channel estimation matrix. Finally, the output of the residual network is upsampled using bilinear interpolation to recover the high-resolution channel estimation results for all subcarriers. This approach introduces deep learning enhancement only for the single channel estimation module.

[0049] Existing deep learning-based OFDM receiver methods can be mainly divided into two categories. The first category replaces a single module in the OFDM receiver with a learnable module based on deep learning. The second category adopts an end-to-end architecture, using neural networks to replace the functions of multiple consecutive modules such as channel estimation, equalization, and demapping in traditional OFDM receivers. To improve performance, existing end-to-end solutions generally employ complex network structures such as self-attention and multi-head attention mechanisms. The high computational complexity and global memory access patterns of these solutions result in slow model inference speeds, failing to meet the real-time data processing requirements of mobile devices. Furthermore, the high energy consumption resulting from high computational complexity poses a severe challenge to the battery life of mobile terminals, making effective deployment difficult in resource-constrained scenarios (such as edge devices and IoT nodes). While hybrid solutions that replace single modules can reduce the number of parameters and computational complexity, because each processing module is designed independently, estimation errors from previous modules propagate directly to subsequent modules, making it impossible to eliminate error accumulation through joint optimization. This may ultimately lead to the system performance getting stuck in local optima. On the other hand, although end-to-end deep learning solutions can achieve global joint optimization, they rely entirely on data-driven feature learning. The feature representations of each layer within the network lack a clear correspondence with the physical processing of the communication system, exhibiting "black box" characteristics. This results in a serious lack of interpretability and makes it impossible to perform module-level optimization for specific physical functions.

[0050] Existing end-to-end solutions employ a multi-symbol training and inference model, repeatedly executing the entire inference process for each OFDM symbol in the data frame. In WiFi communication scenarios, a physical layer data frame typically contains multiple OFDM symbols that are sequentially transmitted, and the intra-frame channel can be considered quasi-static. However, existing solutions fail to utilize this characteristic, repeatedly calculating common information such as channel estimation for each symbol, resulting in computational redundancy and impacting system inference efficiency.

[0051] To address the problems of existing technologies, this embodiment provides a modular OFDM receiving method based on deep learning. It fully utilizes the powerful feature extraction and nonlinear fitting capabilities of deep learning while also considering the practical requirements of high precision, low latency, and low power consumption in WiFi signal processing scenarios. Through modular architecture design and an intra-frame channel estimation result multiplexing strategy, it improves bit error rate performance while maintaining good interpretability and inference efficiency. Specifically, this embodiment first organizes the input data into a frequency domain complex matrix, which includes seqLTF, demodLTF, demodData, and seqPilot. The seqLTF reflects the standard LTF reference sequence for channel estimation as specified by the protocol; the demodLTF reflects the frequency domain symbols obtained by demodulation at the receiver after the standard LTF reference sequence is transmitted through the channel; the demodData reflects the frequency domain symbols obtained by demodulation at the receiver after the data field is transmitted through the channel; and the seqPilot reflects the pilot symbol reference sequence defined by the protocol in the data field. Then, the seqLTF and demodLTF are input to the channel estimation module, and the channel estimation result is output. Next, the channel estimation result, the demodData, and the seqPilot are input to the pilot phase tracking module, which outputs the demodulated data field features after phase correction. Then, the channel estimation result and the demodulated data field features after phase correction are input to the equalization demapping module, which outputs the log-likelihood ratio of each coded bit, thereby completing OFDM reception based on the log-likelihood ratio of each coded bit.

[0052] In practical applications, the deep learning-based modular OFDM receiving method of this embodiment can be applied to, for example... Figure 2 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be placed in the cloud or on other network servers. The modular OFDM receiving method based on deep learning can be executed by terminal 102 or server 104, or it can be executed collaboratively by terminal 102 and server 104.

[0053] The terminal 102 can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, IoT device, or portable wearable device. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices, etc. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices, etc.

[0054] Server 104 can be an independent physical server, a cloud server, or a service node in a blockchain system, where the service nodes in the blockchain system form a peer-to-peer network.

[0055] In addition, server 104 can also be a server cluster consisting of multiple physical servers, which can be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0056] Terminal 102 and server 104 can be connected via Bluetooth, USB (Universal Serial Bus) or network, etc., and this application does not impose any restrictions.

[0057] The core idea of ​​this invention is to explicitly map the physical processing flow of an OFDM receiver into a modular structure of a neural network. Each module corresponds to a specific physical function, and the modules are connected through physically meaningful intermediate features, forming a modular end-to-end computing architecture. Compared to traditional methods, this invention achieves cross-module error adaptive compensation through end-to-end joint training; compared to black-box end-to-end methods, this invention enhances interpretability and supports computational optimization through modular design. Figure 1 As shown, traditional receivers have independently designed modules with fixed parameters. This invention replaces these modules with a learnable neural network structure, specifically including a channel estimation module, a pilot phase tracking module, and an equalization demapping module. The optimal processing strategy is automatically learned through a data-driven approach while maintaining the physical logical relationships between modules. Compared to traditional OFDM receiving methods, this invention achieves collaborative optimization of the parameters of the three modules through end-to-end joint training, solving the error accumulation problem caused by independent module design in traditional methods. Compared to black-box end-to-end methods, the modular design and physically interpretable intermediate feature transfer of this invention enhance the interpretability of the technical solution, support targeted optimization for individual modules (such as changing residual block types or adjusting convolution parameters), and reduce computational complexity.

[0058] In some embodiments, such as Figure 3As shown, a modular OFDM receiving method based on deep learning is provided. This method can be applied to a terminal, such as a computer or other intelligent product terminal. Specifically, the method includes the following steps: Step S100: Organize the input data into a frequency domain complex matrix, which includes seqLTF, demodLTF, demodData, and seqPilot. The seqLTF reflects the channel estimation standard LTF reference sequence specified by the protocol; the demodLTF reflects the frequency domain symbols obtained by demodulation at the receiving end after the standard LTF reference sequence is transmitted through the channel; the demodData reflects the frequency domain symbols obtained by demodulation at the receiving end after the data field is transmitted through the channel; and the seqPilot reflects the pilot symbol reference sequence defined by the protocol in the data field.

[0059] In practical applications, the technical solution of this invention is applicable to various OFDM-based wireless communication systems, especially WiFi 6 (IEEE 802.11ax) systems. It can support transmission scenarios with single-user / multi-user, different channel bandwidths, and different modulation and coding schemes. The following uses the WiFi 6 single-user downlink transmission scenario as an example to explain the specific implementation of the technical solution in detail.

[0060] According to the IEEE 802.11ax standard, the structure of a HE SU PPDU (High Throughput Single User Physical Layer Protocol Data Unit) includes a preamble, HE-LTF (High Throughput Long Training Field), and HE-Data (High Throughput Data Field). This invention first organizes the input data into a frequency-domain complex matrix, which includes seqLTF, demodLTF, demodData, and seqPilot. Specifically, seqLTF is a predefined reference sequence for channel estimation in the WiFi protocol. Its length and subcarrier distribution strictly follow the IEEE 802.11ax standard and are pre-stored at the receiver as a known sequence. The seqLTF is generated as follows: based on the WiFi protocol version (e.g., IEEE 802.11ax) and channel bandwidth corresponding to the target work order number, the corresponding HE-LTF sequence is retrieved from a pre-defined reference sequence library and converted into a frequency-domain complex matrix. For example, with a 20MHz channel bandwidth, the HE-LTF sequence contains 242 effective subcarriers (excluding DC subcarriers and guard subcarriers), and the dimension of seqLTF is [1, 242, 2] (1 represents the number of symbols, 242 represents the number of subcarriers, and 2 represents the real and imaginary parts of the complex number). The data format of seqLTF is a complex matrix, where each element consists of a real and an imaginary part.

[0061] The demodLTF is a frequency domain symbol obtained by demodulating the time-domain signal received by the receiver through an antenna after the seqLTF has been transmitted through the channel. It undergoes synchronization processing (symbol synchronization, carrier synchronization) and FFT transformation, and contains channel fading and noise information. The demodLTF is generated as follows: the receiver performs time synchronization and frequency offset correction on the received time-domain signal, extracts the time-domain interval of the HE-LTF, and performs an FFT transformation to obtain the frequency domain symbol demodLTF. Its dimensions are consistent with seqLTF, for example, [1, 242, 2], ensuring a one-to-one correspondence with the subcarriers of seqLTF. The data characteristics of demodLTF: demodLTF and seqLTF satisfy the relationship: demodLTF = seqLTF × H + N, where H is the true channel response and N is the noise vector. Therefore, demodLTF is the core input data for channel estimation.

[0062] `demodData` is the frequency domain symbol obtained by demodulation at the receiving end after the HE-Data field has been transmitted through the channel. It carries the user's original data (such as text, images, commands, etc.). The generation method of `demodData` is similar to `demodLTF`. The receiving end extracts the time domain interval corresponding to HE-Data in the PPDU and obtains `demodData` after FFT transformation. Its dimensions are [M, 1, 242, 2], where M is the number of OFDM symbols contained in the HE-Data field. The value of M is determined according to the data frame length to adapt to the transmission requirements of variable-length data frames in the WiFi protocol, and 242 is the number of subcarriers, including data subcarriers and pilot subcarriers. The data characteristics of `demodData` are affected by a combination of channel fading, noise, and phase distortion, and it is the object of subsequent phase correction and equalization demapping processing.

[0063] `seqPilot` is a predefined sequence of pilot symbols in the HE-Data field, used to estimate the common phase error (CPE). Its position, spacing, and values ​​strictly adhere to the IEEE 802.11ax standard. `seqPilot` is generated by inserting pilot symbols into pilot subcarriers based on parameters such as the number of symbols and bandwidth, forming a frequency-domain complex matrix of dimension [M,1,K,2], where K is the number of pilot subcarriers. The data characteristics of `seqPilot` are: `seqPilot` is a known sequence at the receiver, corresponding one-to-one with the pilot positions in `demodData`, used to assist in the estimation and compensation of phase errors.

[0064] The neural network model in this embodiment consists of three core modules connected in series. Figure 1 As can be seen from the diagram, the neural network model includes: a channel estimation module, a pilot phase tracking module, and an equalization demapping module.

[0065] Step S200: Input the seqLTF and the demodLTF into the channel estimation module and output the channel estimation result.

[0066] In this embodiment, combined with Figure 4 As shown, the channel estimation module in this embodiment includes an upsampling unit (conv2D1×1) and three cascaded inverted residual blocks (Inverted RBs). The inputs to the channel estimation module are seqLTF and demodLTF, and the output is the channel estimation result H. The parameter settings for each component are as follows: Upsampling unit: Implemented using a 1×1 convolutional layer, with 2 input channels (real and imaginary parts), 32 output channels, and a kernel stride of 1 to ensure that the output feature map size is consistent with the input. Inverted Residual Block (RB): Three residual blocks are connected in series, and each residual block adopts an inverted residual structure. The structure of the inverted residual block is as follows. Figure 5 As shown, it specifically includes: Dimensional convolutional layer: 1×1 convolution, 32 input channels, 64 output channels, stride 1; Deep convolutional layer: 3×3 depth convolution, 64 input channels, 64 output channels, stride 1, used to capture the correlation between subcarriers; Dimensionality reduction convolutional layer: 1×1 convolution, 64 input channels, 32 output channels, stride 1; Residual connection: The output of the dimensionality-reduced convolutional layer is added to the input of the residual block to alleviate the gradient vanishing problem and improve training stability.

[0067] In practical applications, this embodiment first concatenates the seqLTF (dimension [1,242,2]) and demodLTF (dimension [1,242,2]) to obtain fused input data with dimensions [2,242,2]. After the fused input data obtained by fusing the seqLTF and demodLTF is input into the channel estimation module, the channel estimation module expands the feature spaces of the seqLTF and demodLTF by upsampling. For example, specifically, the number of channels is expanded from 4 to 32, resulting in a feature map with dimensions [2,242,32], thus expanding the feature space to extract richer channel information.

[0068] Next, the upsampled feature maps are sequentially input into three cascaded inverted residual blocks. Based on these inverted residual blocks, features are first extracted by upscaling, then the correlation between subcarriers is captured by 3x3 convolution, and finally fused by downscaling. The frequency domain statistical characteristics and noise distribution of the channel are implicitly learned, and the channel estimation result is output. Specifically, the processing procedure for each residual block is as follows: a. Dimensionality Upgrading: The 32-channel features are upgraded to 64 channels using 1×1 convolution, enhancing the feature representation capability; b. Correlation capture: Local feature extraction is performed on the 242 subcarriers of each channel through 3×3 depth convolution to capture the spatial correlation between subcarriers, such as the channel fading correlation between adjacent subcarriers; c. Dimensionality reduction and fusion: The 64-channel features are reduced to 32 channels by 1×1 convolution, reducing the number of parameters and computational cost; d. Residual connection: The dimensionality-reduced features are added to the input features of the residual block to preserve the original feature information and alleviate gradient vanishing.

[0069] After the three inverted residual blocks are processed, the output channel estimation result H is obtained, with dimensions [2,242,32]. This result implicitly includes the frequency domain statistical characteristics of the channel (such as the delay spread and fading amplitude of multipath propagation) and noise distribution information. Compared with the traditional LS method (H=demodLTF / seqLTF), it has stronger noise resistance and interpolation accuracy.

[0070] In this embodiment, a 3×3 depthwise convolution is used for carrier correlation capture. The receptive field of the convolution kernel covers adjacent subcarriers, which can effectively capture the channel fading correlation between subcarriers and improve the smoothness of channel estimation. This embodiment implicitly learns the distribution patterns of the channel and noise through a deep learning model, eliminating the need to manually design denoising algorithms and interpolation functions, and can adapt to noise intensity and fading modes under different channel scenarios. The "dimensionality increase-depth convolution-dimensionality reduction" structure of the inverted residual block significantly reduces the number of parameters and lowers the computational complexity while ensuring feature extraction capabilities.

[0071] Step S300: Input the channel estimation result, the demodData, and the seqPilot into the pilot phase tracking module, and output the demodulated data field features after phase correction.

[0072] like Figure 4 As can be seen from the diagram, the pilot phase tracking module in this embodiment includes a feature fusion block, two cascaded channel attention residual blocks (SE-RB), and a projection residual block (Projection RB). The inputs to the pilot phase tracking module are the channel estimation result H, demodData, and seqPilot, and the output is the demodulated data field features after phase correction. The parameter settings for each component of the pilot phase tracking module are as follows: Feature fusion block: Uses Concat (channel splicing) operation, with no additional parameters; Channel Attention Residual Block (SE-RB): Two residual blocks are cascaded, and each residual block has the structure of "inverted residual block + channel attention mechanism". Figure 6 and Figure 7 As shown, Figure 7 for Figure 6 The structural diagram of the SE block in the diagram includes: Inverted residual block section: The inverted residual block structure is consistent with that in the channel estimation module (1×1 convolution with increased dimensionality → 3×3 depth convolution → 1×1 convolution with decreased dimensionality → residual connection), with 64 input channels and 64 output channels; The channel attention mechanism consists of two sub-modules: Squeeze and Excitation. a. Squeeze sub-module: Global average pooling converts the feature map [M, 2, 242, 64] into a vector [M, 2, 242, 1], capturing global information for each channel. b. Excitation sub-module: Two fully connected layers. The first layer has 64 input channels and 16 output channels (compression factor 4), with ReLU activation. The second layer has 16 input channels and 64 output channels, with Sigmoid activation, outputting weight coefficients for each channel. c. Feature weighting: The weight coefficients output from the excitation sub-module are multiplied channel-by-channel with the output features of the inverted residual block, enhancing effective features and suppressing ineffective features. Projection Residual Block (RB), the structure of the projection residual block is as follows Figure 8 As shown: It includes a 1×1 convolutional layer with residual connections, 64 input channels, 32 output channels, and a convolutional kernel stride of 1.

[0073] In practical applications, after inputting the dimensions of the channel estimation results H, demodData, and seqPilot into the pilot phase tracking module, this embodiment first adjusts the dimensions of the channel estimation results H, demodData, and seqPilot. Specifically, this includes: expanding the dimensions of the channel estimation results H (dimension [2,242,32]) output by the channel estimation module to fit the number of symbols M of demodData, obtaining the expanded features of H with dimensions [M,2,242,32] (achieved by copying the symbol dimension of H). The demodData dimension is [M,1,242,2]. In this embodiment, the number of channels is expanded from 2 to 32 through a 1×1 convolutional layer, obtaining the features of demodData with dimensions [M,1,242,32]. The seqPilot dimension is [M, 1, K, 2] (K is the number of pilot subcarriers). In this example, it is expanded to [M, 1, 242, 2] based on the position of the pilot subcarriers (consistent with the dimension of demodData). The data subcarrier positions are filled with zeros. Then, the number of channels is expanded to 32 through a 1×1 convolutional layer to obtain the features of the seqPilot with dimension [M, 1, 242, 32].

[0074] Next, based on the feature fusion block, the channel estimation result, the demodData, and the seqPilot features are fused through a Concat operation, using the channel estimation result to assist phase tracking. Through a channel attention residual block, all channel features are adaptively weighted to enhance effective features. Then, through a projection residual block, the high-dimensional fused features are projected to a low dimension, optimizing feature representation and outputting the demodulated data field features after phase correction. Specifically, in this embodiment, through the Concat operation of the feature fusion block, the expanded features of the channel estimation result H ([M,2,242,32]), the features of demodData ([M,1,242,32]), and the features of seqPilot ([M,1,242,32]) are concatenated to obtain fused features with dimensions [M,2,242,64]. Then, combined with... Figure 4 As shown, the fused features are input into the first channel attention residual block (SE-RB), and the processing procedure is as follows: a. Inverted residual block feature extraction: Local features are extracted through "dimensionality increase-depth convolution-dimensionality reduction-residual connection", and the output dimension is [M, 2, 242, 64]; b. Channel attention weighting: By capturing global information through the squeezing submodule, the submodule is stimulated to generate weight coefficients, and the features of 64 channels are adaptively weighted to strengthen the features related to phase error.

[0075] Then, a second feature optimization is performed: the output of the first SE-RB is input into the second SE-RB, and the processing of the channel attention residual block above is repeated to further optimize the feature representation and suppress noise interference.

[0076] Finally, feature dimensionality reduction and fusion are performed: the output of the second SE-RB ([M,2,242,64]) is input into the projected residual block, and the number of channels is reduced from 64 to 32 through 1×1 convolution. At the same time, the core features are preserved through residual connections, and the demodulated data field features after phase correction are output (denoted as demodData', dimension [M,2,242,32]). In this embodiment, the output demodData' is the feature representation after CPE estimation and compensation, eliminating the influence of time-varying phase noise on demodData.

[0077] In this embodiment, the channel estimation result H is fused, and the channel fading information contained in H is used to distinguish the effects of phase error and channel fading, thereby improving the accuracy of phase estimation. Furthermore, the channel attention mechanism, through adaptive weighting, strengthens the phase correlation characteristics between pilot symbols and data symbols, suppressing invalid features caused by noise and interference, significantly improving phase correction performance, especially in low signal-to-noise ratio scenarios. In addition, this embodiment learns the phase compensation pattern from pilot to data symbols using a deep learning model, eliminating the need for manually designing a phase tracking algorithm and enabling adaptive handling of time-varying phase noise.

[0078] Step S400: Input the channel estimation result and the demodulated data field features after phase correction into the equalization demapping module, and output the log-likelihood ratio of each coded bit to complete OFDM reception based on the log-likelihood ratio of each coded bit.

[0079] The core function of the equalization and demapping module in this embodiment is to jointly perform channel equalization and demapping, eliminate residual channel fading and noise interference, and output the log-likelihood ratio (LLR) for each coded bit. The input to the equalization and demapping module is the channel estimation result H and the demodulated data field feature demodData' after phase correction, and the output is the LLR. Figure 4 As can be seen, the equalization demapping module includes two cascaded inverted residual blocks (RBs), one projection residual block (RB), and a fully connected layer. The parameter settings for each component of the equalization demapping module are as follows: Inverted Residual Block (RB): Two residual blocks are connected in series. The structure is the same as the inverted residual block in the channel estimation module. There are 64 input channels and 64 output channels. Projection Residual Block (RB): 64 input channels, 32 output channels, dimensionality reduction is achieved through 1×1 convolution, and residual connections preserve core features; Fully connected layer: The input dimension is [M,2,242,32], and the output dimension is [M,(242-K)*B], where B is the number of coded bits corresponding to each subcarrier symbol.

[0080] In this embodiment, after inputting the channel estimation result and the features of the demodulated data field after phase correction into the equalization and demapping module, the channel estimation result and the features of the demodulated data field after phase correction are fused through a Concat operation, and the channel information of the channel estimation result is used to assist the equalization calculation. Then, the equalization and demapping functions are learned through the inverted residual block, and feature fusion and dimensionality reduction are performed through the projected residual block. Finally, feature fitting is performed through a fully connected layer, and the log-likelihood ratio of each coded bit is output.

[0081] In practical applications, combined with Figure 4 As shown in the diagram, this embodiment first expands the channel estimation result H([2,242,32]) into an H-expanded feature of [M,2,242,32]. Next, through a Concat operation, the H-expanded feature ([M,2,242,32]) is concatenated with demodData'([M,2,242,32]) to obtain a fused feature with dimensions [M,242,64]. Then, the fused feature is input into the first inverted residual block, and through a structure of "dimensionality increase-depth convolution-dimensionality reduction-residual connection," the channel equalization function is learned to eliminate residual channel fading interference. Next, the output of the first inverted residual block is input into the second inverted residual block to further learn the demapping function, converting the frequency domain symbol features into bit-level features. Then, the output of the second inverted residual block ([M,2,242,64]) is input into the projected residual block, and through a 1×1 convolution, the number of channels is reduced from 64 to 32, reducing the number of parameters in the fully connected layer. Finally, the output of the projection residual block ([M,2,242,32]) is input into the fully connected layer, and the LLR of each coded bit is output through linear fitting, with dimensions [M,(242-K)*B]. For example, in 16-QAM modulation, B=4, and each subcarrier symbol corresponds to a 4-bit LLR value. The sign of the LLR indicates the probability tendency of the bit to be "1" or "0", and the absolute value represents the confidence level of the probability.

[0082] In traditional methods, equalization and demapping are two independent steps. This invention uses a deep learning model to jointly learn the two functions, avoiding error accumulation caused by step-by-step processing and improving the accuracy of bit estimation. Furthermore, this embodiment integrates the channel estimation result H, enabling the equalization and demapping process to adapt to dynamic channel changes and maintain good performance even in complex channel scenarios. This embodiment directly outputs the LLR through a fully connected layer, eliminating the need for an additional soft-decision module, simplifying the reception process. Simultaneously, the deep learning model can learn a more accurate bit probability distribution, providing high-quality input for subsequent LDPC decoding.

[0083] In other implementations, to achieve end-to-end training and performance evaluation of the model, it is necessary to construct training and testing datasets, maintaining statistical independence between the two datasets to ensure the objectivity of the test results. This embodiment can simulate the complete channel transmission process within a preset signal-to-noise ratio range, generating receiver frequency domain data, and combining it with the transmitter's coded bit sequence as coded bit labels to obtain a training dataset. Each training sample in the training dataset includes the aforementioned frequency domain complex matrix.

[0084] Specifically, the process of generating the training dataset is as follows: 1. System Parameter Configuration: Based on a WiFi 6 single-user downlink transmission scenario, the core system parameters are configured as follows: Physical layer frame format: HE SU PPDU; Channel bandwidth: 20MHz; Modulation method: 16-QAM; Coding method: LDPC, code rate 1 / 2; Channel model: TGax Channel Model-B (a channel model conforming to the IEEE 802.11ax standard, supporting multipath propagation, path loss, and shadowing fading); Carrier frequency: 5.25GHz; Transceiver antenna: SISO (single-input single-output); Transceiver distance: 5m; Noise type: Additive white Gaussian noise (AWGN).

[0085] 2. Signal Generation and Transmission Simulation: a. Transmitter Signal Generation: Based on the above parameters, a random raw bit sequence is generated. A complete HESU PPDU physical frame is constructed through encoding, modulation, and other operations, with the encoded bit sequence serving as tag data; b. Channel Transmission Simulation: The PPDU is transmitted using the TGax Channel Model-B, with AWGN added. c. Receiver signal processing: The receiver performs synchronization processing and FFT transformation on the transmitted signal, demodulates it to obtain demodLTF and demodData, and calls the preset seqLTF and seqPilot to form four frequency domain complex matrices. 3. Training Dataset Construction: Each training sample contains "input data + label data". The input data consists of four frequency domain complex matrices: seqLTF, demodLTF, demodData, and seqPilot. The dataset size is 3000 samples generated for each signal-to-noise ratio (SNR), with an SNR range of [0, 30] dB and a step size of 0.2 dB, for a total of 61 SNR points. The total number of samples in the training dataset is 183,000.

[0086] This embodiment uses a random number seed independent of the training dataset, and follows the same system parameters and channel configuration as the training dataset to generate test frequency domain data, thus obtaining the test dataset.

[0087] Specifically, the generation process of the test dataset is as follows: The system parameters of the test dataset are completely consistent with those of the training dataset to ensure the matching of the test scenario and the training scenario. A random number seed independent of the training dataset is used (e.g., the random number seed for the training set is 12345, and for the test set it is 67890) to avoid data overlap between the training and test sets. 500 samples are generated for each signal-to-noise ratio (SNR), with an SNR range of [0, 30] dB and a step size of 0.2 dB, for a total of 61 SNR points. The total number of samples in the test dataset is 30,500.

[0088] During training, this embodiment trains the model based on the training dataset. In each training iteration, a corresponding frequency-domain complex matrix is ​​extracted from the training dataset to form input data. This input data is then fed into the channel estimation module, pilot phase tracking module, and equalization demapping module within the model. Next, the log-likelihood ratio of each coded bit in the final output is compared with the coded bit label in the training dataset to calculate the binary cross-entropy (BCE) loss, using the following formula: The formula is: Loss = - (y×log (σ(LLR)) + (1 - y)×log (1 -σ(LLR))) where y is the original encoded bit label (0 or 1), and σ is the Sigmoid function used to convert LLR into a probability value in the interval [0,1].

[0089] Finally, the parameters of the channel estimation module, pilot phase tracking module, and equalization demapping module are updated using the backpropagation gradient algorithm. Specifically, this embodiment uses the Adam optimizer, calculates the loss value based on BCE, and performs backpropagation gradients to update the parameters of all modules. After training, the complete model is saved for subsequent inference and testing.

[0090] Furthermore, this embodiment also inputs the test dataset into the trained and optimized channel estimation module, pilot phase tracking module, and equalization demapping module to obtain test results. Then, performance evaluation is performed based on the test results to verify the model's performance. Specifically, this embodiment can, based on the test results, statistically analyze the bit error rate (BER) of the communication system under different signal-to-noise ratios during the testing phase. The formula for BER is: BER = Number of error bits / Total number of bits. Furthermore, based on the test results, the model performance when transmitting data frames of different lengths can be determined, verifying the model's adaptive processing capability for variable-length data frames in the WiFi protocol.

[0091] In summary, existing end-to-end deep learning OFDM receiver solutions generally adopt a "black box" architecture, lacking a clear correspondence between network internal features and physical processing flows, resulting in insufficient interpretability and an inability to be modularly optimized. This invention, however, explicitly maps channel estimation, phase error tracking, and equalization demapping into independent neural network modules, with intermediate feature tensors with clear physical meaning passed between modules, achieving physical interpretability while maintaining the advantages of end-to-end joint training. Existing solutions based on Transformer or multi-head attention mechanisms suffer from high computational complexity, slow inference speed, and high resource consumption, making it difficult to meet real-time processing requirements. This invention, however, uses local convolution operations to significantly reduce computational complexity and greatly improve inference speed, meeting the real-time and low-power requirements of WiFi scenarios. Existing solutions employ a multi-symbol inference mode, resulting in a large amount of redundant computation during data frame processing. This invention innovatively proposes a flexible architecture design based on single-symbol training and supporting multi-symbol batch inference, capable of adaptively handling the actual transmission scenarios of variable-length data frames in the WiFi protocol, without requiring retraining the model for different frame lengths, and exhibiting good generalization ability. The channel estimation result intra-frame multiplexing strategy proposed in this invention utilizes the quasi-static characteristics of WiFi intra-frame channels. The channel estimation module performs only one calculation, and its output is multiplexed for subsequent processing of all symbols within the frame, effectively eliminating computational redundancy. This invention achieves cross-module error adaptive compensation through end-to-end joint training, and compared with traditional methods, it can significantly improve bit error rate performance under various signal-to-noise ratio conditions.

[0092] The technical feasibility and performance advantages of this invention have been verified through complete simulation experiments, with simulation parameters set as described in the above embodiments. Using the method of this patent, the BER performance of traditional OFDM receivers can be effectively improved under different signal-to-noise ratio conditions, showing a certain performance improvement compared to the classic end-to-end OFDM receiver DeepRx. Furthermore, the model exhibits good generalization ability in scenarios involving the transmission of WiFi data frames of different lengths.

[0093] like Figure 9 As shown, Figure 9 To compare the BER performance of transmitting 4-symbol data frames under different signal-to-noise ratio conditions, for example, under a 0dB signal-to-noise ratio condition, the BER of the traditional method, DeepRx and the method proposed in this invention are 0.4933901, 0.4812371 and 0.4454547, respectively. The method proposed in this invention can reduce the BER by 9.72% compared to the traditional method and by 7.44% compared to DeepRx.

[0094] like Figure 10 As shown, Figure 10To compare the BER performance of transmitting 64-symbol data frames under different signal-to-noise ratio (SNR) conditions, for example, under a 25 dB SNR condition, the BERs of the traditional method, DeepRx, and the method proposed in this invention are 0.001922435, 0.00003260289, and 0.000002605559, respectively. The method proposed in this invention can reduce the BER by 99.86% compared to the traditional method and by 92.01% compared to DeepRx.

[0095] Based on the above embodiments, the present invention also provides a modular OFDM receiving system based on deep learning, the system being used to implement the steps in the above method embodiments. Specifically, as Figure 11 As shown, the system in this embodiment includes: a frequency domain complex matrix determination module 10, a channel estimation module 20, a pilot phase tracking module 30, and an equalization demapping module 40. The frequency domain complex matrix determination module 10 organizes the input data into a frequency domain complex matrix, which includes seqLTF, demodLTF, demodData, and seqPilot. The seqLTF reflects the channel estimation standard LTF reference sequence specified by the protocol; the demodLTF reflects the frequency domain symbols obtained by demodulation at the receiving end after the standard LTF reference sequence is transmitted through the channel; the demodData reflects the frequency domain symbols obtained by demodulation at the receiving end after the data field is transmitted through the channel; and the seqPilot reflects the pilot symbol reference sequence defined by the protocol in the data field. The channel estimation module 20 outputs a channel estimation result based on the seqLTF and the demodLTF. The pilot phase tracking module 30 outputs the characteristics of the demodulated data field after phase correction based on the channel estimation result, the demodData, and the seqPilot. The equalization demapping module 40 is used to output the log-likelihood ratio of each coded bit based on the channel estimation result and the characteristics of the demodulated data field after phase correction, so as to complete OFDM reception based on the log-likelihood ratio of each coded bit.

[0096] In one implementation, the channel estimation module 20 includes: an upsampling unit and three cascaded inverted residual blocks; The upsampling unit is used to expand the feature space of the seqLTF and the demodLTF; The inverted residual block of the channel estimation module is used to first extract features by increasing the dimensionality, then capture the correlation between subcarriers by 3x3 convolution, and then fuse them by reducing the dimensionality to implicitly learn the frequency domain statistical characteristics and noise distribution of the channel.

[0097] In one implementation, the pilot phase tracking module 30 includes: a feature fusion block, a channel attention residual block, and a projection residual block.

[0098] The feature fusion block is used to fuse the channel estimation result, the demoData, and the features of seqPilot through the Concat operation, and to use the channel estimation result to assist phase tracking. The channel attention residual block is used to adaptively weight all channel features to enhance effective features; The projection residual block is used to project high-dimensional fused features to low dimensions, thereby optimizing feature representation.

[0099] In one implementation, the equalization demapping module 40 includes an inverted residual block, a projected residual block, and a fully connected layer; The inverted residual block of the equalization demapping module is used to learn the equalization and demapping functions; The projection residual blocks of the equalization demapping module are used for feature fusion and dimensionality reduction; The fully connected layer is used for feature fitting and outputs the log-likelihood ratio for each encoded bit.

[0100] In one implementation, the system further includes: The training dataset construction module is used to simulate the complete channel transmission process within a preset signal-to-noise ratio range, generate receiver frequency domain data, and use the transmitter coded bit sequence as coded bit label to obtain the training dataset. Each training sample in the training dataset includes the frequency domain complex matrix. The test dataset building module is used to generate test frequency domain data using a random number seed that is independent of the training dataset, while following the same system parameters and channel configuration as the training dataset, thus obtaining the test dataset.

[0101] In one implementation, the system further includes a model training and optimization module, comprising: The training data input unit is used to train the model based on the training dataset, and in each training iteration, extracts the corresponding frequency domain complex matrix from the training dataset and inputs it into the channel estimation module, pilot phase tracking module and equalization demapping module in the model. The binary cross-entropy loss calculation unit is used to compare the log-likelihood ratio of each encoded bit in the final output with the encoded bit label in the training dataset to calculate the binary cross-entropy loss. The back gradient propagation unit is used to update the parameters of the neural network model through the back gradient propagation algorithm, thereby enabling the training and optimization of the channel estimation module, pilot phase tracking module, and equalization demapping module.

[0102] In one implementation, the method further includes a testing and performance evaluation module, comprising: The testing unit is used to input the test dataset into the trained and optimized channel estimation module, pilot phase tracking module, and equalization demapping module to obtain test results; The performance evaluation unit is used to evaluate the performance based on the test results to verify the performance of the channel estimation module, pilot phase tracking module and equalization demapping module in the model.

[0103] The principles and functions of each module and unit in the system of this embodiment are the same as those of each step in the above method embodiment, and will not be repeated here.

[0104] In some embodiments, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 12 As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through WiFi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a modular OFDM receiving method based on deep learning. The display unit of the computer device is used to form a visually visible image. It can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0105] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0106] In some embodiments, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps: The input data is organized into a frequency domain complex matrix, which includes seqLTF, demodLTF, demodData, and seqPilot. The seqLTF reflects the channel estimation standard LTF reference sequence specified by the protocol; the demodLTF reflects the frequency domain symbols obtained by demodulation at the receiver after the standard LTF reference sequence has been transmitted through the channel; the demodData reflects the frequency domain symbols obtained by demodulation at the receiver after the data field has been transmitted through the channel; and the seqPilot reflects the pilot symbol reference sequence defined by the protocol in the data field. The seqLTF and the demodLTF are input into the channel estimation module, and the channel estimation result is output. The channel estimation results, the demodData, and the seqPilot are input into the pilot phase tracking module, and the demodulated data field characteristics after phase correction are output. The channel estimation result and the demodulated data field features after phase correction are input into the equalization demapping module, which outputs the log-likelihood ratio of each coded bit to complete OFDM reception based on the log-likelihood ratio of each coded bit.

[0107] It should be noted that the user information (including but not limited to user device function information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0108] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0109] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0110] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A modular OFDM receiving method based on deep learning, characterized in that, The method includes: The input data is organized into a frequency domain complex matrix, which includes seqLTF, demodLTF, demodData, and seqPilot. The seqLTF reflects the channel estimation standard LTF reference sequence specified by the protocol; the demodLTF reflects the frequency domain symbols obtained by demodulation at the receiver after the standard LTF reference sequence has been transmitted through the channel; the demodData reflects the frequency domain symbols obtained by demodulation at the receiver after the data field has been transmitted through the channel; and the seqPilot reflects the pilot symbol reference sequence defined by the protocol in the data field. The seqLTF and the demodLTF are input into the channel estimation module, and the channel estimation result is output. The channel estimation results, the demodData, and the seqPilot are input into the pilot phase tracking module, and the demodulated data field characteristics after phase correction are output. The channel estimation result and the demodulated data field features after phase correction are input into the equalization demapping module, which outputs the log-likelihood ratio of each coded bit to complete OFDM reception based on the log-likelihood ratio of each coded bit.

2. The modular OFDM receiving method based on deep learning according to claim 1, characterized in that, The channel estimation module includes an upsampling unit and three cascaded inverted residual blocks.

3. The modular OFDM receiving method based on deep learning according to claim 2, characterized in that, The seqLTF and demodLTF are input to the channel estimation module, and the channel estimation results are output, including: The feature spaces of the seqLTF and the demodLTF are expanded by upsampling; Based on the inverted residual block, features are first extracted by increasing the dimensionality, then the correlation between subcarriers is captured by 3x3 convolution, and then fused by dimensionality reduction. The frequency domain statistical characteristics and noise distribution of the channel are implicitly learned, and the channel estimation results are output.

4. The modular OFDM receiving method based on deep learning according to claim 3, characterized in that, The pilot phase tracking module includes: a feature fusion block, a channel attention residual block, and a projection residual block.

5. The modular OFDM receiving method based on deep learning according to claim 4, characterized in that, The channel estimation results, the demodData, and the seqPilot are input to the pilot phase tracking module, which outputs the demodulated data field characteristics after phase correction, including: Based on the feature fusion block, the channel estimation result, the demoData and the features of seqPilot are fused through the Concat operation, and the channel estimation result is used to assist phase tracking. By using channel attention residual blocks, all channel features are adaptively weighted to enhance effective features; By projecting residual blocks, high-dimensional fused features are projected onto low-dimensional features, optimizing feature representation and outputting demodulated data field features after phase correction.

6. The modular OFDM receiving method based on deep learning according to claim 5, characterized in that, The equalization demapping module includes: inverted residual blocks, projected residual blocks, and fully connected layers.

7. The modular OFDM receiving method based on deep learning according to claim 6, characterized in that, The channel estimation result and the phase-corrected demodulated data field features are input into the equalization demapping module, which outputs the log-likelihood ratio for each coded bit, including: The Concat operation fuses the characteristics of the channel estimation results with those of the demodulated data field after phase correction, and uses the channel information from the channel estimation results to assist in equalization calculation. The inverted residual block is used to learn the equalization and demapping functions, and then the projected residual block is used for feature fusion and dimensionality reduction. Feature fitting is performed through a fully connected layer, and the log-likelihood ratio of each encoded bit is output.

8. The modular OFDM receiving method based on deep learning according to claim 7, characterized in that, The method further includes: Within a preset signal-to-noise ratio range, the complete channel transmission process is simulated to generate receiver frequency domain data, which is then paired with the transmitter coded bit sequence as coded bit labels to obtain a training dataset. Each training sample in the training dataset includes the frequency domain complex matrix. Using a random number seed independent of the training dataset, and following system parameters and channel configurations completely identical to those of the training dataset, test frequency domain data is generated to obtain the test dataset.

9. The modular OFDM receiving method based on deep learning according to claim 8, characterized in that, The method further includes: The model is trained based on the training dataset, and in each training iteration, the corresponding frequency domain complex matrix is ​​extracted from the training dataset and input into the channel estimation module, pilot phase tracking module and equalization demapping module in the model. The log-likelihood ratio of each encoded bit in the final output is compared with the encoded bit label in the training dataset to calculate the binary cross-entropy loss. By using the backpropagation algorithm, the parameters of the neural network model are updated, enabling the training and optimization of the channel estimation module, pilot phase tracking module, and equalization demapping module.

10. The modular OFDM receiving method based on deep learning according to claim 9, characterized in that, The method further includes: The test dataset is input into the trained and optimized channel estimation module, pilot phase tracking module, and equalization demapping module to obtain the test results; Performance evaluation is performed based on the test results to verify the performance of the channel estimation module, pilot phase tracking module, and equalization demapping module in the model.

11. The modular OFDM receiving method based on deep learning according to claim 10, characterized in that, Performance evaluation is performed based on the test results, including: Based on the test results, the bit error rate of the communication system under different signal-to-noise ratios during the test phase was statistically analyzed. Based on the test results, the model performance when transmitting data frames of different lengths is determined, and the model's adaptive processing capability for variable-length data frames in the WiFi protocol is verified.

12. A modular OFDM receiving system based on deep learning, characterized in that, The system is used to implement the steps of the deep learning-based modular OFDM receiving method according to any one of claims 1-11, the system comprising: The frequency domain complex matrix determination module is used to organize the input data into a frequency domain complex matrix, which includes seqLTF, demodLTF, demodData, and seqPilot. The seqLTF reflects the channel estimation standard LTF reference sequence specified by the protocol; the demodLTF reflects the frequency domain symbols obtained by demodulation at the receiving end after the standard LTF reference sequence is transmitted through the channel; the demodData reflects the frequency domain symbols obtained by demodulation at the receiving end after the data field is transmitted through the channel; and the seqPilot reflects the pilot symbol reference sequence defined by the protocol in the data field. The channel estimation module is used to output channel estimation results based on the seqLTF and the demodLTF; The pilot phase tracking module is used to output the characteristics of the demodulated data field after phase correction based on the channel estimation results, the demodData, and seqPilot. The equalization demapping module is used to output the log-likelihood ratio of each coded bit based on the channel estimation result and the characteristics of the demodulated data field after phase correction, so as to complete OFDM reception based on the log-likelihood ratio of each coded bit.

13. The modular OFDM receiver system based on deep learning according to claim 12, characterized in that, The channel estimation module includes: an upsampling unit and three cascaded inverted residual blocks; The upsampling unit is used to expand the feature space of the seqLTF and the demodLTF; The inverted residual block of the channel estimation module is used to first extract features by increasing the dimensionality, then capture the correlation between subcarriers by 3x3 convolution, and then fuse them by reducing the dimensionality to implicitly learn the frequency domain statistical characteristics and noise distribution of the channel.

14. The modular OFDM receiver system based on deep learning according to claim 12, characterized in that, The pilot phase tracking module includes: a feature fusion block, a channel attention residual block, and a projection residual block; The feature fusion block is used to fuse the channel estimation result, the demoData, and the features of seqPilot through the Concat operation, and to use the channel estimation result to assist phase tracking. The channel attention residual block is used to adaptively weight all channel features to enhance effective features; The projection residual block is used to project high-dimensional fused features to low dimensions, thereby optimizing feature representation.

15. The modular OFDM receiver system based on deep learning according to claim 12, characterized in that, The equalization demapping module includes an inverted residual block, a projected residual block, and a fully connected layer; The inverted residual block of the equalization demapping module is used to learn the equalization and demapping functions; The projection residual blocks of the equalization demapping module are used for feature fusion and dimensionality reduction; The fully connected layer is used for feature fitting and outputs the log-likelihood ratio for each encoded bit.

16. The modular OFDM receiver system based on deep learning according to claim 12, characterized in that, The system also includes: The training dataset construction module is used to simulate the complete channel transmission process within a preset signal-to-noise ratio range, generate receiver frequency domain data, and use the transmitter coded bit sequence as coded bit label to obtain the training dataset. Each training sample in the training dataset includes the frequency domain complex matrix. The test dataset building module is used to generate test frequency domain data using a random number seed that is independent of the training dataset, while following the same system parameters and channel configuration as the training dataset, thus obtaining the test dataset.

17. The modular OFDM receiver system based on deep learning according to claim 16, characterized in that, The system also includes a model training and optimization module, comprising: The training data input unit is used to train the model based on the training dataset, and in each training iteration, extracts the corresponding frequency domain complex matrix from the training dataset and inputs it into the channel estimation module, pilot phase tracking module and equalization demapping module in the model. The binary cross-entropy loss calculation unit is used to compare the log-likelihood ratio of each encoded bit in the final output with the encoded bit label in the training dataset to calculate the binary cross-entropy loss. The back gradient propagation unit is used to update the parameters of the neural network model through the back gradient propagation algorithm, thereby enabling the training and optimization of the channel estimation module, pilot phase tracking module, and equalization demapping module.

18. The modular OFDM receiving method based on deep learning according to claim 17, characterized in that, The method also includes a testing and performance evaluation module, comprising: The testing unit is used to input the test dataset into the trained and optimized channel estimation module, pilot phase tracking module, and equalization demapping module to obtain test results; The performance evaluation unit is used to evaluate the performance based on the test results to verify the performance of the channel estimation module, pilot phase tracking module and equalization demapping module in the model.

19. A computer device, characterized in that, The computer device includes a memory, a processor, and a deep learning-based modular OFDM receiver program stored in the memory and executable on the processor. When the processor executes the deep learning-based modular OFDM receiver program, it implements the steps of the deep learning-based modular OFDM receiver method as described in any one of claims 1-12.

20. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a deep learning-based modular OFDM receiver program, which implements the steps of the deep learning-based modular OFDM receiver method as described in any one of claims 1-12 on the computer-readable storage medium.

Images