A deep learning-based pre-training carrier frequency offset calibration method

Abstract

The application discloses a kind of pre-training carrier frequency offset calibration methods based on deep learning, belong to wireless communication technical field.Method includes: step 1, by simulating drive pre-training, pre-train backbone deep neural network on synthetic OFDM signal, learn generalized feature under hardware distortion;Step 2, for target equipment, use a small amount of real data lightweight fine-tuning regression layer, realize device-specific adaptation;Step 3, utilize convolution-full connection hybrid architecture to extract CP base feature, and carry out CFO estimation;Step 4, by Sim2Real transfer learning framework, bridge simulation and reality gap, improve CFO estimation robustness;Step 5, according to demodulation result optimization compensation parameter, reduce bit error rate.The application realizes 30 times bit error rate reduction on heterogeneous SDR platform, processing delay is low, cost-effective is high.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wireless communication, and particularly relates to a pre-training carrier frequency offset calibration method based on deep learning, and is especially suitable for CFO estimation of an OFDM system on a heterogeneous software defined radio platform. BACKGROUND

[0002] As a core technology of modern wireless communication systems, orthogonal frequency division multiplexing (OFDM) is widely used in high-speed data transmission scenarios such as 5G and Wi-Fi. However, OFDM systems are extremely sensitive to synchronization errors, and carrier frequency offset (CFO) is one of the key factors leading to performance degradation. CFO is mainly caused by the frequency mismatch of local oscillators at the transmitting end and the receiving end. On a software defined radio (SDR) platform, due to hardware heterogeneity (such as software defined radio devices from different manufacturers), CFO estimation faces severe challenges. Traditional CFO estimation methods can be mainly divided into two categories: pilot-based methods and blind estimation methods. Pilot-based methods insert known sequences for frequency offset calibration, but this will reduce spectral efficiency; blind estimation methods such as cyclic prefix (CP) based methods use the redundancy of the CP and the symbol tail for correlation analysis, but this method is easily affected by noise amplification in a low signal-to-noise ratio (SNR) environment, resulting in a dramatic increase in estimation error. Specifically, the CP-based method calculates the frequency offset through maximum likelihood estimation, but its nonlinear parameter operation introduces significant square loss when the SNR is lower than 6 dB, and single-point estimation ignores the phase information of each sampling point, which cannot adapt to hardware distortion.

[0003] In recent years, deep learning technology has provided a new approach to CFO estimation. Deep neural networks (DNNs) can learn hardware-specific distortion patterns, but existing DNN-based CFO estimators lack device-level adaptability. For example, the CNN-Attention-DNN architecture can improve estimation accuracy, but it ignores the performance variations caused by heterogeneous hardware. This variation is caused by uncalibrated radio frequency impairments such as I / Q two-way sampling sequence imbalance, oscillator drift, and phase noise. These impairments exhibit unique fingerprints on different hardware devices, causing the performance of traditional estimators to degrade severely when deployed across platforms. Experiments show that under indoor multipath conditions, the bit error rate (BER) of traditional cyclic prefix-based frequency offset estimation methods on common software defined radio platforms can reach 10 -2 orders of magnitude, while the existing DNN methods have limited optimization of BER without adaptation, still in the order of 10 -2 orders of magnitude, which cannot meet the needs of high-reliability applications such as the Internet of Things or industrial automation.

[0004] In addition, the theoretical limit of CFO estimation is defined by the Cramer-Rao lower bound (CRLB), which is negatively related to SNR, but the traditional method is far from reaching this limit in a dynamic environment. The estimation variance based on the CP strategy is more than 3 times the CRLB, and the DNN method can approach the CRLB, but the simulation and reality gap needs to be addressed. In the prior art, simulation-driven training can reduce costs, but imperfect hardware distortion models lead to deployment failures in reality; and device-specific calibration requires a large amount of real data, increasing deployment overhead. Therefore, there is an urgent need for a CFO calibration scheme that balances accuracy, efficiency, and generalization ability to address the challenges of heterogeneous SDR platforms. SUMMARY

[0005] In order to make up for the shortcomings of the prior art, the present application aims to provide a pre-training carrier frequency offset calibration method based on deep learning, based on the deep collaboration of software defined radio (SDR) hardware platform and deep learning algorithm, for the carrier frequency offset (CFO) estimation problem of heterogeneous devices, a full-link solution from simulation pre-training, device fine-tuning to real-time compensation is proposed.

[0006] The specific technical solutions of the present application are as follows:

[0007] On the one hand, the present application provides a pre-training carrier frequency offset calibration method based on deep learning, comprising:

[0008] Step 1, pre-training backbone deep neural network on synthetic OFDM signal through simulation-driven pre-training, learning generalized features under hardware distortion;

[0009] Step 2, for the target device, use real data to fine-tune the regression layer, realize device-specific adaptation;

[0010] Step 3, use convolutional-full connection hybrid architecture to extract cyclic prefix (CP) base features for CFO estimation;

[0011] Step 4, bridge the simulation and reality gap through the Sim2Real transfer learning framework, improve the robustness of CFO estimation;

[0012] Step 5, optimize compensation parameters according to demodulation results to reduce bit error rate.

[0013] Further, in step 1, simulation-driven training includes:

[0014] Step 1.1, construct a synthetic OFDM signal model, where OFDM symbol length K, cyclic prefix length G, and time domain signal is represented as:

[0015] ,

[0016] where s is the symbol vector, denotes the part of the cyclic prefix, denotes the part where the tail of the symbol vector is used as a cyclic prefix;

[0017] Step 1.2, incorporate parameterized hardware impairment models, including phase noise, I / Q imbalance, oscillator drift, to simulate real SDR device impairments;

[0018] Step 1.3, use a backbone deep neural network for pre-training, network input is the I / Q sample sequence or phase sequence of the received signal, output is the CFO estimate;

[0019] Step 1.4, optimize network parameters by minimizing the mean square error loss function, to avoid the cost of collecting real data across devices.

[0020] Further, in the step 2, the lightweight device adaptation includes freezing the convolutional layers, only optimizing the fully connected head, using 1000 real OFDM frames for fine-tuning, reducing the risk of overfitting.

[0021] Specifically, the specific content of the lightweight device adaptation includes:

[0022] Step 2.1, collect a small number of real OFDM frames for each target SDR device, typically 1000 frames;

[0023] Step 2.2, freeze the convolutional layer weights of the pre-trained network, only fine-tune the fully connected regression layer, and retain hardware-independent feature knowledge;

[0024] Step 2.3, fine-tune using real data, loss function combined with OFDM demodulation results, in the form of:

[0025] ,

[0026] where, denotes the loss function, is the OFDM demodulation function, is the compensated signal, is the known transmission symbol;

[0027] Step 2.4, optimize by gradient descent, adapt to device-specific radio frequency (RF) impairments, and reduce the risk of overfitting.

[0028] Further, in the step 3, feature extraction and CFO estimation includes:

[0029] Step 3.1, use CP structure for conjugate multiplication: for the CP segment of the received signal and the symbol tail , calculate ;

[0030] Step 3.2, Deriving phase sequence , containing CFO-induced phase rotation;

[0031] Step 3.3, Convolve front-end processing I / Q two-way sampling sequence sample phase information, including feature extraction (64-channel 1D convolution, batch normalization, GELU activation), down-sampling (step 2 convolution) and depth separation;

[0032] Step 3.4, CFO estimation using global pooling and regression head (fully connected layer 512→256→128→1).

[0033] Further, in the step 4, the Sim2Real framework correlates the simulation pre-training accuracy with the real performance, and uses the loss function combined with the OFDM demodulation function for optimization.

[0034] Specifically, the specific content of the Sim2Real transfer learning framework includes:

[0035] Step 4.1, learn generalized features on synthetic data through simulation pre-training, and the pre-training loss function is to minimize CFO estimation error;

[0036] Step 4.2, use real device data for light fine-tuning to bridge the gap between simulation and reality;

[0037] Step 4.3, optimize through loss function, adapt to device-specific damage, such as unobserved disturbance caused by hardware ;

[0038] Step 4.4, verify the performance of the framework on heterogeneous SDR platforms, and improve the robustness of CFO estimation.

[0039] Further, in the step 5, the optimization compensation parameters include:

[0040] Step 5.1, adjust the compensation waveform amplitude parameter according to the demodulation symbol length, so that the symbol length is uniform;

[0041] Step 5.2, if the overall phase rotation is observed, modify the compensation scheme, add an initial phase to the compensation signal, so that the final symbol sequence matches the QPSK constellation;

[0042] Step 5.3, dynamically optimize parameters through feedback mechanism to reduce bit error rate.

[0043] On the other hand, the present application provides an electronic device comprising one or more processors and a memory, the memory storing one or more programs, the programs comprising instructions for executing a deep learning-based pre-training carrier frequency offset calibration method as described above.

[0044] The application also provides a storage medium, which stores an executable program, and the executable program includes instructions for executing a deep learning-based pre-training carrier frequency offset calibration method.

[0045] Compared with the prior art, the application has the following advantages:

[0046] 1) High precision and low complexity: through Sim2Real transfer learning, the CFO estimation error variance is close to the Cramér-Rao lower bound, the bit error rate is reduced by 30 times at low signal-to-noise ratio, and the computational complexity is only 5-7% of that of traditional methods.

[0047] 2) Device adaptive capability: on various SDR platforms, device-specific calibration is realized, and the bit error rate is reduced from 10 -2 to 10 -3 levels of the traditional scheme.

[0048] 3) High cost-effectiveness: analog pre-training avoids a large amount of real data collection, fine-tuning only requires 1000 frames, and the hardware resource occupancy rate is less than 20%.

[0049] 4) Real-time processing: supporting 280,000 symbols per second, the latency is less than 1ms, and it is suitable for dynamic environments. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is a workflow diagram of the uplink transceiver of the application;

[0051] Figure 2 is a direct demodulation constellation diagram of the receiver end of the application;

[0052] Figure 3 is a constellation diagram of the neural network demodulation after fine-tuning of the application. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the application clearer, the application will be further described in detail below in combination with the drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and do not limit the application.

[0054] Embodiment 1

[0055] As shown in the figure, the embodiment provides a deep learning-based pre-training carrier frequency offset calibration method, based on the deep collaboration of software-defined radio (SDR) hardware platform and protocol-aware algorithm, aiming at the constellation distortion problem caused by NB-IoT uplink dynamic phase rotation, a full-link unwinding scheme from radio frequency reception, protocol analysis to dynamic compensation is proposed, and the specific implementation process is as follows: Figure 1

[0056] ​Step 1. Pre-train the backbone DNN on synthetic OFDM signals to learn generalized features under hardware distortions.

[0057] This step pre-trains the backbone DNN on synthetic OFDM signals to learn generalized hardware distortion features, avoiding real data collection costs.

[0058] First, construct a synthetic OFDM signal model: let the symbol length K = 256, the cyclic prefix length G = 8, and the time-domain signal be represented as where s is a randomly generated QPSK modulated symbol vector, denotes the part where the cyclic prefix is, denotes the part where the tail of the symbol vector is used as a cycle, and the sampling frequency is set to 1.92 MHz to simulate a typical wireless scenario. Next, incorporate a parameterized hardware distortion model, including phase noise (using a Wiener process model with a variance of 10 -4 rad²), I / Q two-way sampling sequence imbalance (amplitude mismatch 0.1 dB, phase mismatch 1°), and oscillator drift (±2.5 ppm), which are based on the data manual of a specific SDR device to ensure simulation authenticity. The DNN input is the I / Q two-way sampling sequence sample sequence of the received signal (length K+G = 264), and the output is the CFO estimate value. The pre-training dataset contains 100,000 synthetic samples, with a signal-to-noise ratio (SNR) range covering 0-30 dB, and the Adam optimizer (learning rate 10 -3 ) is used during training, and the loss function uses mean squared error (MSE): where is the network-estimated CFO, is the true value. The pre-training goal is to minimize the MSE to enable the network to learn a robust mapping under distortion. The training period is set to 100 rounds, with a batch size of 64, and early stopping is used to prevent overfitting. After pre-training, the model is verified on an independent test set, requiring the CFO estimation error variance to be close to the Cramér-Rao lower bound (CRLB), such as an error variance of less than 1 Hz² at SNR = 6 dB.

[0059] Step 2. For the target device, use real data to lightly fine-tune the regression layer to achieve device-specific adaptation.

[0060] For the target SDR device, this step achieves device-specific adaptation through light fine-tuning.

[0061] First, a small amount of real OFDM frames (typical number 1,000 frames) are collected for each device, frame structure reference 3GPP standard: subcarrier number 12, symbol number 14, no pilot design, using differential QPSK (DQPSK) modulation to eliminate phase ambiguity. Data is collected in an indoor multipath environment (ETU70 channel model), ensuring diversity. Fine-tune the pre-trained network's convolutional layer weights (retain hardware-independent knowledge), and only optimize the fully connected regression head (layer dimension 512→256→128→1). The fine-tuning loss function combines the OFDM demodulation results, defined as:

[0062] ,

[0063] where, represents the loss function, is the OFDM demodulation function (dependent on symbol length and cyclic prefix), is the compensated signal, is the known transmission symbol. This design avoids the problem of unobservable real CFO, and indirectly optimizes the network through demodulation accuracy. Optimization uses gradient descent method, learning rate set to 10 -4 , batch size is 32, data is divided into 800 frames for training and 200 frames for verification. To suppress overfitting, neuron random closing and regularization (weight decay 10 -4 ) techniques are introduced to improve generalization, and the training patience is set to 10 rounds. After fine-tuning, the model's bit error rate (BER) on the target device should be less than 10 -3 , and the parameter increment accounts for only 5%-7% of the overall, achieving low-cost adaptation.

[0064] Step 3, use a convolutional-full connection hybrid architecture to extract CP-based features for CFO estimation.

[0065] This step uses the cyclic prefix (CP) structure and DNN architecture for robust feature extraction.

[0066] First, extract the CP segment and the corresponding symbol tail from the received signal r (distorted by the multipath channel h and CFO effect): , and derive the phase sequence , which contains CFO-induced phase rotation and noise components. The DNN adopts a convolution-full connection hybrid architecture: the convolution front-end consists of three stages—feature extraction (64-channel 1D convolution, kernel size 7, batch normalization, and GELU activation), down-sampling (stride 2 convolution, kernel size 5), and depthwise separable convolution (spatial convolution kernel 3 followed by pointwise convolution); a global pooling layer compresses the feature map to 512 dimensions; the regression head consists of fully connected layers (512→256→128→1) with GELU and SiLU activation functions alternately used. The network input is the phase sequence Φ, and the output is the CFO estimate . In training, the adaptive average pooling operation is denoted as:

[0067] ,

[0068] where F is the convolution feature map and T is the time dimension. This architecture optimizes feature learning under hardware distortion, and when deployed on FPGA, achieves a processing speed of 280,000 symbols per second through pipeline design, with a latency of less than 1 ms.

[0069] Step 4, bridge the gap between simulation and reality through the Sim2Real transfer learning framework, and improve the robustness of CFO estimation.

[0070] Verify performance through the simulation and reality transfer framework, bridge the gap. After simulation pre-training, fine-tune and evaluate on the real SDR platform. In the experimental setup, the transmitting end is an arbitrary SDR device, and the receiving end is the target device, simulating a heterogeneous deployment scenario; the channel condition is indoor multipath (delay spread 1 μs), and the signal-to-noise ratio (SNR) dynamic range is 0-15 dB. After fine-tuning, evaluate performance by compensating CFO and demodulating: use DQPSK modulation, differential encoding to eliminate residual phase rotation, and power normalization of demodulated symbols. The result analysis shows that the Sim2Real framework significantly improves robustness, for example, at SNR=6 dB, the CFO estimation variance of the fine-tuned model is 1.1 times that of CRLB, while the traditional CP-based method is more than 3 times. Error distribution comparison shows that the DNN method has high error concentration at low SNR, with no significant outliers.

[0071] Bit error rate analysis further verifies the advantage, and the fine-tuned model optimizes BER from 10 -2 orders of magnitude to 10 -3 of the traditional scheme on all platforms, especially on low-end SDR devices, with a 30-fold improvement. As shown in Figure 2 and Figure 3 , the demodulation constellation diagrams after calibration by the traditional CP scheme and the present scheme are recorded, showing that under large-scale CFO conditions, the traditional scheme calibration is basically ineffective, while the present scheme compensates and the symbols converge to the ideal DQPSK distribution.

[0072] Step 5, optimize compensation parameters according to demodulation results, reduce bit error rate.

[0073] According to the demodulation results, the size of the final symbol modulus is observed, and if the modulus is not uniform, the amplitude parameter in step 2 is flexibly adjusted when generating the compensation waveform, so that the final demodulation symbol modulus is as uniform as possible.

[0074] According to the demodulation results, the parameters are dynamically adjusted to ensure the accuracy of symbol recovery. First, based on the output symbol modulus, the amplitude is controlled: let the average value of the demodulation symbol sequence modulus be A, if the standard deviation exceeds the threshold (such as 0.1), adjust the compensation waveform amplitude, the correction scheme is , wherein is the received signal, is the compensation sequence conjugate. If the overall phase rotation is observed (detect by constellation fitting), further add the initial phase, and correct it to , so that the symbol sequence matches the ideal QPSK constellation. The optimization process is iterated every 100 frames, and the parameters are adjusted in real time through the feedback mechanism. At the same time, use the DQPSK modulation transmission model, and its phase encoding rule is , wherein , to ensure that the phase ambiguity is not sensitive.

[0075] Embodiment 2

[0076] This embodiment proposes an electronic device comprising one or more processors and a memory, the memory storing one or more programs comprising instructions for executing a deep learning-based pre-trained carrier frequency offset calibration method as described in Embodiment 1.

[0077] Embodiment 3

[0078] This embodiment provides a computer-readable storage medium comprising one or more programs for execution by one or more processors of an electronic device. These programs include instructions for executing a deep learning-based pre-trained carrier frequency offset calibration method as described in Embodiment 1.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not limited thereto; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions described in the foregoing embodiments can still be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A pre-trained carrier frequency offset calibration method based on deep learning, characterized in that, include: Step 1: Pre-train the backbone deep neural network on the synthesized OFDM signal through simulation-driven pre-training to learn the generalized features under hardware distortion. Simulation-driven training includes: Step 1.1: Construct a synthetic OFDM signal model, where the OFDM symbol length K, the cyclic prefix length G, and the time-domain signal representation is: ， Where s is the symbol vector, This indicates the part with the cyclic prefix. The end of the symbol vector is used as the loop portion; Step 1.2: Incorporate a parametric hardware distortion model, including phase noise, I / Q sampling sequence imbalance, and oscillator drift, to simulate damage to real SDR equipment; Step 1.3: Use a backbone deep neural network for pre-training. The network input is the I / Q sampling sequence or phase sequence of the received signal, and the output is the CFO estimate. Step 1.4: Optimize network parameters by minimizing the mean squared error loss function to avoid the cost of collecting real data across devices; Step 2: For the target device, use real data to lightweight fine-tune the regression layer to achieve device-specific adaptation; lightweight device adaptation includes freezing convolutional layers and optimizing only fully connected heads; the specific content of lightweight device adaptation includes: Step 2.1: For each target SDR device, collect 1000 real OFDM frames. Step 2.2: Freeze the convolutional layer weights of the pre-trained network, and only fine-tune the fully connected regression layers to retain hardware-independent feature knowledge; Step 2.3: Fine-tune using real data. The loss function is combined with the OFDM demodulation results, and takes the following form: ， in, Represents the loss function. For OFDM demodulation function, To compensate for the signal, The transmission symbol is known; Step 2.4: Through gradient descent optimization, adapt to device-specific RF impairments and reduce the risk of overfitting; Step 3: Extract cyclic prefix (CP) basis features using a convolutional-fully connected hybrid architecture to perform CFO estimation; feature extraction and CFO estimation include: Step 3.1: Perform conjugate multiplication using the CP structure: For the received signal r, extract the cyclic prefix (CP) segment. and the corresponding symbol tail Calculated using conjugate multiplication: ; Step 3.2: Derive the phase sequence The phase sequence Includes CFO-induced phase rotation and noise components; Step 3.3: The DNN adopts a convolutional-fully connected hybrid architecture: the convolutional front end includes three stages—feature extraction, downsampling, and depthwise separating convolution; the global pooling layer compresses the feature map to 512 dimensions; the regression layer consists of fully connected layers; GELU and SiLU activation functions are used alternately; the network input is the phase sequence Φ, and the output is the CFO estimate. ; Step 4: Using the Sim2Real transfer learning framework, bridge the gap between simulation and reality to improve the robustness of CFO estimation; Step 5: Optimize the compensation parameters based on the demodulation results to reduce the bit error rate.

2. The deep learning-based pre-trained carrier frequency offset calibration method according to claim 1, characterized in that, In step 4, the Sim2Real framework combines simulation pre-training with fine-tuning based on real acquired data, and uses a loss function combined with an OFDM demodulation function for optimization.

3. The deep learning-based pre-trained carrier frequency offset calibration method according to claim 2, characterized in that, The Sim2Real transfer learning framework includes the following: Step 4.1: Learn generalized features on synthetic data through simulation pre-training. The pre-training loss function is to minimize the CFO estimation error. Step 4.2: Use real device data for lightweight fine-tuning to bridge the gap between simulation and reality; Step 4.3: Optimize the loss function to adapt to specific equipment damage; Step 4.4: Verify the performance of the framework on the heterogeneous SDR platform and improve the robustness of CFO estimation.

4. The deep learning-based pre-trained carrier frequency offset calibration method according to claim 1, characterized in that, In step 5, the optimization compensation parameters include: Step 5.1: Adjust the compensation waveform amplitude parameters according to the demodulation symbol magnitude to make the symbol magnitude uniform; Step 5.2: If an overall phase rotation is observed If necessary, the compensation scheme is modified by adding an initial phase to the compensation signal so that the final symbol sequence matches the QPSK constellation. Step 5.3: Optimize parameters dynamically through a feedback mechanism to reduce the bit error rate.

5. An electronic device, characterized in that, It includes one or more processors and a memory, the memory storing one or more programs containing instructions for performing a deep learning-based pre-trained carrier frequency offset calibration method as described in claim 1.

6. A storage medium storing an executable program, characterized in that, The executable program includes instructions for performing a deep learning-based pre-trained carrier frequency offset calibration method as described in claim 1.

Images