OFDM system channel estimation method based on deep neural network

Abstract

The invention provides an OFDM (Orthogonal Frequency Division Multiplexing) system channel estimation method based on a deep neural network, which comprises the following steps of: generating simulation channel data as a test sample, including a receiving signal and a pilot signal; a deep neural network architecture is built, optimizer training parameters are configured, and model training is carried out; estimating the received signal and the pilot signal by using a traditional algorithm to obtain a coarse estimation value; inputting the signal into the trained deep neural network to obtain a fine estimation value, and combining the fine estimation value with the coarse estimation value to obtain a final channel estimation value; and performing multiple simulation tests in different channel environments, evaluating whether the performance meets requirements, if not, adjusting the traditional algorithm and deep neural network parameters, performing data processing, estimation and evaluation again, and ending the process until the requirements are met. According to the invention, a traditional algorithm and deep learning are combined, a low-complexity and high-precision hybrid channel estimation method is designed, and the channel estimation performance of an OFDM system in a complex channel environment is improved.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication, and specifically to a channel estimation method for OFDM systems based on a deep neural network. Background Art

[0002] Due to its high spectral efficiency and strong anti-multipath fading ability, OFDM technology is widely used in fields such as 4G / 5G mobile communication and optical communication. However, OFDM systems are very sensitive to channel distortion, and the accuracy of channel estimation directly affects the signal recovery quality at the receiving end. Traditional channel estimation methods, such as the least squares method (LS), although simple in calculation process, have poor anti-noise performance and are difficult to meet the high-precision estimation requirements in low signal-to-noise ratio scenarios. The minimum mean square error (MMSE) algorithm can optimize the estimation accuracy by using channel statistical information and has relatively better performance, but it depends on the known channel covariance matrix. In actual communication processes, due to the difficulty of obtaining an accurate channel covariance matrix, the practical application of this algorithm is limited.

[0003] In recent years, deep learning technology has shown excellent performance in non-linear modeling and data-driven optimization, especially in fields such as image processing and natural language processing. Applying deep learning technology to channel estimation can learn channel characteristics through neural networks, reduce pilot overhead, improve estimation accuracy, and adapt to complex channel environments. However, existing deep learning-based channel estimation methods, such as in high-speed mobile scenarios, the mean square error (MSE) increases significantly, and still have problems such as insufficient generalization ability, high network complexity, and difficulty in being deployed in real-time systems. Summary of the Invention

[0004] The present invention proposes a channel estimation method for OFDM systems based on a deep neural network, aiming to design a low-complexity and high-precision hybrid channel estimation method by combining the theoretical advantages of traditional algorithms and the non-linear modeling ability of deep learning, so as to improve the channel estimation performance of OFDM systems in complex channel environments and solve the problems proposed in the above background art. The technical solution provided by the present invention is as follows: A channel estimation method for OFDM systems based on a deep neural network, comprising the following steps: Step 1, generating simulation channel data of the OFDM system as test samples, including received signals and pilot signals; Step 2, building a deep neural network DNN architecture, configuring optimizer training parameters, and performing model training; Step 3, estimating the received signal and the pilot signal using the LS algorithm or the MMSE algorithm to obtain a rough estimate value of the channel; Step 4: Input the received signal and the pilot signal into the trained deep neural network (DNN) for estimation to obtain a fine estimate of the channel. Combine the fine estimate with the coarse estimate to obtain the final channel estimate value. Step 5: Conduct multiple simulation experiments under different channel environments. Evaluate whether the performance of Steps 1 - 4 meets the requirements by calculating the mean square error (MSE), bit error rate (BER), and computational complexity. If not, adjust the parameters of the traditional algorithm and the deep neural network, and re - perform data processing, estimation, and evaluation until the requirements are met and then end the process.

[0005] Preferably, the Min - Max normalization method is used to normalize the generated channel data to the interval.

[0006] Preferably, the optimizer selection principle is the Adam optimizer, and its training parameters are as follows: the learning rate is set to 0.0001 - 0.01 and adjusted according to the convergence situation during training; the batch size is set to 32 - 128 to balance memory occupancy and training efficiency; the number of training epochs is set to 200 - 500. Determine whether the model converges by the loss function J value of the validation data to avoid overfitting.

[0007] Preferably, L2 regularization and Dropout techniques are used to prevent model overfitting: Add a Dropout layer in the DNN, and set the Dropout probability to 0.2 - 0.5 to randomly discard some neurons and reduce the co - adaptation phenomenon between neurons; apply L2 regularization to the model parameters to punish overly large parameter values. The loss function of L2 regularization adds a regularization term on the basis of the original loss function J : , where λ is the regularization coefficient, W is the set of model parameters.

[0008] Preferably, for the MMSE algorithm used in Step 3, an adaptive pilot insertion method is adopted as an improvement strategy, and its formula is: , where is the number of pilots, α is the coefficient adjusted according to the channel change situation, B is the signal bandwidth, is the maximum Doppler shift.

[0009] Preferably, for the MMSE algorithm used in Step 3, a method based on eigenvalue decomposition is adopted as an improvement strategy to perform eigenvalue decomposition on the channel's autocovariance matrix : , where U is the eigenvector matrix, Λis an eigenvalue diagonal matrix.

[0010] Preferably, an online learning mechanism is added to the deep neural network DNN. During the online learning process, according to the newly received signal and the pilot signal , the neural network parameters are updated using the stochastic gradient descent method θ , and the update formula is: , where is the parameter at the t-th update, η is the learning rate, is the loss function J with respect to the parameter gradient.

[0011] Preferably, network pruning and quantization techniques are introduced to reduce the complexity of the deep neural network DNN, specifically expressed as: Set a threshold τ , for the elements of the model parameter set W , the network pruning operation is expressed as:

[0012] Let the original parameter be z , and the quantized parameter be , the quantization formula is:

[0013] where s is the quantization step size and round is the rounding function.

[0014] Preferably, a generative adversarial network GAN is introduced to enhance the diversity of training data.

[0015] Preferably, a multi-task learning framework is introduced to simultaneously optimize the two tasks of channel estimation and signal detection by sharing the feature extraction layer.

[0016] Compared with the prior art, the beneficial effects achieved by the present invention are: First, by combining the coarse estimation of traditional algorithms and the fine estimation of deep neural networks, as well as optimizing the model, under the same computational complexity, the channel estimation accuracy of the present invention is significantly better than that of traditional LS and MMSE algorithms, which can effectively improve the signal recovery quality at the receiving end and enhance the estimation accuracy; Second, by adopting a low-complexity coarse estimation algorithm, model pruning and quantization techniques, the computational complexity of the present invention is significantly reduced, meeting the strict requirements of real-time communication for computing resources, making the present invention more practical; ​III. The powerful adaptability of the online learning mechanism and the deep neural network enables the present invention to maintain stable performance in complex channel environments such as high mobility and multipath fading, effectively suppressing the influence of noise and nonlinear distortion on channel estimation and showing stronger robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, but do not constitute a limitation to the present invention. In the drawings: Figure 1 is the overall flowchart of the whole process of implementing the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0019] To make the above objects, features, and effects of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0020] Embodiment 1: A channel estimation method for an OFDM system based on a deep neural network, as Figure 1 shown, includes the following steps: Step 1, using the communication system toolbox of Matlab, generate simulation channel data of the OFDM system as test samples, including received signals and pilot signals .

[0021] Among them, the range of the signal-to-noise ratio SNR is taken from -5 dB to 20 dB, with a step of 3 dB, so as to simulate different intensities of noise interference scenarios; for the multipath fading channel, using the Rayleigh fading model and the Rician fading model, set different fading parameters, such as the number of multipaths, fading factors, etc., so as to simulate typical wireless communication environments such as cities and suburbs; for the time-varying channel, considering the influence of Doppler frequency shift, according to different moving speeds (such as 50 km / h, 100 km / h, 200 km / h, etc.), generate corresponding time-varying channel models, so as to simulate the channel changes in high-speed moving scenarios.

[0022] Further, adopt the Min-Max normalization method to normalize the generated channel data to the interval. For the amplitude value A of the received signal, calculate its minimum value and the maximum value , through the formula Perform normalization; for the channel state information CSI, perform a similar normalization operation; normalization can ensure faster convergence of the data during subsequent neural network training and improve training efficiency.

[0023] Step 2, build a deep neural network DNN architecture, configure the optimizer training parameters, and perform model training.

[0024] The deep neural network DNN is particularly suitable for the OFDM channel estimation algorithm due to its powerful non-linear modeling ability, data-driven learning method, and adaptability to complex channel environments. The deep neural network mentioned above can adopt a multi-layer perceptron MLP or a convolutional neural network CNN. The convolutional neural network CNN is suitable for extracting the characteristics of the channel in the frequency domain and spatial domain, and is particularly suitable for dealing with the multipath effect and frequency selective fading in the OFDM system.

[0025] If MLP is adopted, set the number of hidden layers to 3 - 5 layers, and the number of neurons in each layer is adjusted according to the input data dimension and model complexity requirements, generally between 64 - 256; if CNN is adopted, set 2 - 3 convolutional layers, the convolutional kernel size is 3×3 or 5×5, the stride is 1, the pooling layer adopts max pooling, and the pooling kernel size is 2×2.

[0026] Use the Deep Learning Toolbox of Matlab and select a suitable optimizer to perform model training. The optimizer can be the Adam optimizer, and its training parameters are: the learning rate is set to 0.0001 - 0.01, adjusted according to the convergence situation during training, the batch size is set to 32 - 128 to balance memory occupancy and training efficiency, the number of training epochs is set to 200 - 500 epochs, and the convergence of the model is judged by the loss function J value of the validation data to avoid overfitting.

[0027] Furthermore, adopt L2 regularization and Dropout technology to prevent model overfitting. Add a Dropout layer in the DNN, and set the Dropout probability to 0.2 - 0.5 to randomly discard some neurons and reduce the co-adaptation phenomenon between neurons; at the same time, apply L2 regularization to the model parameters to punish too large parameter values and make the model more generalized. The loss function of L2 regularization Add a regularization term to the original loss function J : , where λ is the regularization coefficient, W is the set of model parameters.

[0028] In addition, deep neural networks can also incorporate attention mechanisms (such as Transformer) to enhance the network's ability to extract key channel features.

[0029] Step 3: Estimate the received signal and the pilot signal using the LS algorithm or the MMSE algorithm to obtain a rough estimate of the channel.

[0030] For the LS algorithm, based on the received signal and the pilot signal , calculate the channel estimate according to the least squares criterion. The formula is: where X is the matrix composed of pilot signals, and T represents the matrix transpose. This algorithm can quickly generate preliminary channel state information with relatively low computational complexity.

[0031] For the MMSE algorithm, the formula for calculating the MMSE estimate is: where is the autocovariance matrix of the channel, H is the channel matrix, is the noise variance, and I is the identity matrix. Since it is difficult to obtain an accurate channel covariance matrix, a low-complexity approximation method is introduced, including adaptive pilot insertion and approximate calculation based on eigenvalue decomposition to simplify the calculation process. Retain the first 80% of the eigenvalues to reduce the computational complexity.

[0032] Adaptive pilot insertion formula: where is the number of pilots, α is the coefficient adjusted according to the channel variation, B is the signal bandwidth, and is the maximum Doppler shift. The traditional fixed pilot insertion method may lead to low spectral efficiency in complex channel environments. Through the adaptive pilot insertion strategy of this formula, the number and position of pilots can be dynamically adjusted according to the channel environment, reducing the pilot overhead while ensuring the estimation accuracy. This strategy is particularly suitable for high-frequency band communications (such as millimeter wave and terahertz bands), and can significantly improve the spectral utilization rate, providing support for 5G / 6G communication systems.

[0033] For the method based on eigenvalue decomposition, perform eigenvalue decomposition on the autocovariance matrix of the channel: where U is the eigenvector matrix, and Λ is the eigenvalue diagonal matrix. By retaining the main eigenvalues, the calculation process is simplified and the computational complexity is reduced.

[0034] Step 4: Input the received signal and the pilot signal into the trained deep neural network for estimation to obtain a fine estimate of the channel, and combine the fine estimate with the coarse estimate to obtain the final channel estimate value.

[0035] The core process of the deep neural network DNN is as follows: Step 41: Load the corresponding pre-trained neural network model according to different SNR values; Step 42: Separate the complex received signal into real and imaginary parts, and interleave the real and imaginary parts to form a neural network input vector; Step 43: Normalize the input data; Step 44: Use the loaded neural network model for prediction; Step 45: Denormalize the output of the neural network model; Step 46: Combine the output real and imaginary parts into a complex channel fine estimate value.

[0036] Step 5: Conduct multiple simulation experiments under different channel environments. Each experiment runs independently 100 - 1000 times. Evaluate whether the performance of Steps 1 - 5 meets the requirements by calculating the mean square error MSE, bit error rate BER, and computational complexity. If not, adjust the parameters of the traditional algorithm and the deep neural network, and re - conduct data processing, estimation, and evaluation until the requirements are met and then end the process.

[0037] When calculating the mean square error MSE, compare the channel estimate value obtained from each experiment with the true channel state, and calculate through the formula where N is the number of samples, is the channel estimate value, is the true channel value.

[0038] When calculating the bit error rate BER, use the channel estimate value for signal demodulation at the receiving end, and count the ratio of the number of error symbols after demodulation to the total number of symbols.

[0039] When calculating the computational complexity, record the CPU time and memory occupancy required for one run of the present invention, and take the average value through multiple experiments.

[0040] Embodiment 2: Design an online learning mechanism for the deep neural network DNN so that the model can track channel changes and adapt to the dynamic channel environment.

[0041] During the online learning process, according to the newly received signal and the pilot signal , use the stochastic gradient descent method to update the neural network parameters θ , and the update formula is: , where is the parameter at the t-th update, η is the learning rate, is the loss function J with respect to the parameter .

[0042] In a dynamic channel environment, at regular time intervals (such as 10 - 100 symbol periods), new received data is selected as training samples, and the DNN model is fine-tuned in an incremental learning manner, enabling the present invention to track channel changes in real time.

[0043] Example 3: Introduce network pruning and quantization techniques to reduce the complexity of the deep neural network DNN and reduce the number of model parameters.

[0044] Set a threshold τ , for the elements of the model parameter set W , the network pruning operation can be expressed as:

[0045] z The quantization technique quantizes the parameters and calculations in the model. For example, floating-point parameters are quantized into fixed-point numbers. Let the original parameter be z , and the quantized parameter be , and the quantization formula is:

[0046] where s is the quantization step and round is the rounding function.

[0047] Network pruning reduces the number of model parameters by removing redundant neurons and connections; the quantization technique converts floating-point weights into low-bit representations, further reducing the computational and storage overhead. These optimization measures enable the present invention to operate efficiently on resource-constrained devices (such as mobile terminals or embedded devices), meeting the requirements of real-time communication systems. In addition, this lightweight design can also reduce the power consumption of the present invention, conforming to the development trend of green communication.

[0048] Example 4: The present invention can also introduce a generative adversarial network GAN to enhance the diversity of training data and improve the generalization ability of the model. GAN can generate simulation data close to the real channel environment. By calculating the KL divergence between the generated data and the real data, the effectiveness of the generated data is ensured, making up for the problem of insufficient actual data and improving the generalization ability of the deep learning model. Through the multi-scenario data generated by GAN (such as high-speed movement, dense multipath, non-linear distortion, etc.), the present invention can maintain high-precision estimation performance in a wider range of channel environments. In addition, GAN can also be used for data augmentation to further improve the robustness of the model. The core code is as follows: function h_est = DNN_est_gan_augmented(y, SNR) persistent gan_net dnn_net; if isempty(gan_net) % Load the pre-trained model gan_data = load('gan_model.mat'); dnn_data = load('dnn_estimator_gan_augmented.mat'); gan_net = gan_data.gan; dnn_net = dnn_data.dnn_network; end % Input preprocessing yy = [real(y); imag(y)]; yy = (yy - mean_vec). / std_vec; % Normalize % Use GAN to generate augmented data noise = randn(gan_net.generator.Layers(1).InputSize,1); augmented_input = predict(gan_net.generator, dlarray(noise,'CB')); augmented_input = [yy, extractdata(augmented_input)]; % DNN estimation h_est = predict(dnn_net, augmented_input); h_est = h_est(1) + 1i*h_est(2); % Convert to complex number end Example 5: Traditional OFDM systems usually process channel estimation and signal detection as independent modules. The present invention can also introduce a multi-task learning framework. By sharing the feature extraction layer, it can optimize both the channel estimation and signal detection tasks simultaneously, improving the overall performance of the system. This end-to-end optimization method not only simplifies the system design but also significantly reduces the bit error rate (BER) and improves the communication quality. The core code is as follows: % Modified DNN_est function to support joint inference function [h_estDNN, detected_symbol] = DNN_est_joint(y, SNR) persistent mt_net; if isempty(mt_net) % Load the pre-trained multi-task network loaded = load('multi_task_wireless_net.mat'); mt_net = loaded.mt_net; end % Perform joint inference [h_estDNN, detected_symbol] = joint_inference(mt_net, y, SNR); end The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A channel estimation method for an OFDM system based on a deep neural network, characterized in that It includes the following steps: Step 1: Generate the simulation channel data of the OFDM system as test samples, including received signals and pilot signals; Step 2: Build the deep neural network DNN architecture, configure the training parameters of the optimizer, and conduct model training; Step 3: Estimate the received signals and pilot signals using the LS algorithm or MMSE algorithm to obtain the rough estimated value of the channel; Step 4: Input the received signals and pilot signals into the trained deep neural network DNN for estimation to obtain the fine estimated value of the channel, and combine the fine estimated value with the rough estimated value to obtain the final channel estimated value; Step 5: Conduct multiple simulation experiments in different channel environments, and evaluate whether the performance of Steps 1 - 4 meets the requirements by calculating the mean square error MSE, bit error rate BER, and computational complexity. If not, adjust the parameters of the traditional algorithm and the deep neural network, and re - conduct data processing, estimation, and evaluation until the requirements are met and then end the process.

2. The channel estimation method for an OFDM system based on a deep neural network according to claim 1, wherein, Using the Min-Max normalization method, normalize the generated channel data to interval.

3. The channel estimation method for an OFDM system based on a deep neural network according to claim 2, wherein, The optimizer selection principle is the Adam optimizer, and its training parameters are as follows: the learning rate is set to 0.0001 - 0.01 and adjusted according to the convergence situation during the training process; the batch size is set to 32 - 128 to balance memory occupancy and training efficiency; the number of training epochs is set to 200 - 500 epochs, and the loss function of the validation data J is used to judge whether the model converges and avoid overfitting.

4. A channel estimation method for an OFDM system based on a deep neural network according to claim 3, characterized in that Use L2 regularization and Dropout techniques to prevent model overfitting: Add a Dropout layer in the DNN, set the Dropout probability to 0.2 - 0.5, randomly discard some neurons, and reduce the co-adaptation phenomenon between neurons; Apply L2 regularization to the model parameters to penalize overly large parameter values. The loss function of L2 regularization is added to the original loss function J with a regularization term: , where λ is the regularization coefficient, W is the set of model parameters.

5. The channel estimation method for an OFDM system based on a deep neural network according to claim 2, characterized in that, The MMSE algorithm used in step 3 adopts an adaptive pilot insertion method as an improvement strategy, and its formula is: , where is the number of pilots, α is the coefficient adjusted according to the channel change situation, B is the signal bandwidth, is the maximum Doppler shift.

6. The channel estimation method for an OFDM system based on a deep neural network according to claim 2, characterized in that, The MMSE algorithm used in step 3 adopts the method based on eigenvalue decomposition as the improvement strategy, and the autocovariance matrix of the channel is subjected to eigenvalue decomposition: , where U is the eigenvector matrix, Λ is the eigenvalue diagonal matrix.

7. A channel estimation method for an OFDM system based on a deep neural network according to any one of claims 1-6, characterized in that, Add an online learning mechanism to the deep neural network DNN. During the online learning process, according to the newly received signal and the pilot signal , update the neural network parameters using the stochastic gradient descent method θ . The update formula is: , where is the parameter at the t-th update, η is the learning rate, is the loss function J with respect to the parameter gradient.

8. A channel estimation method for an OFDM system based on a deep neural network according to any one of claims 1-6, characterized in that, Introduce network pruning and quantization techniques to reduce the complexity of the deep neural network DNN, which is specifically expressed as: Set a threshold τ For the elements of the model parameter set W the network pruning operation is expressed as: ​ ； Let the original parameter be z , and the quantized parameter be , and the quantization formula is: ； where s is the quantization step size and round is the rounding function.

9. A channel estimation method for an OFDM system based on a deep neural network according to any one of claims 1-6, characterized in that Introduce the generative adversarial network GAN to enhance the diversity of training data.

10. A channel estimation method for an OFDM system based on a deep neural network according to any one of claims 1-6, characterized in that, Introduce a multi - task learning framework to simultaneously optimize the two tasks of channel estimation and signal detection by sharing the feature extraction layer.

Images