Vehicle communication channel estimation algorithm based on convolutional neural network

Abstract

The invention belongs to the field of channel estimation in vehicle communication, provides a vehicle communication channel estimation algorithm based on a convolutional neural network, and aims to solve the problems of multipath effect, noise interference, high calculation complexity and the like of a traditional channel estimation method in a vehicle-mounted environment. According to the method, a deep super-resolution convolutional neural network VDSR and a blind denoising convolutional neural network RIDNet are combined, three channel attention mechanisms of SENet, ECANet and CBAM are introduced into the networks, and the adaptability and accuracy of the model in a complex environment are improved by enhancing the feature expression ability. The method specifically comprises the following steps: performing preliminary channel estimation by adopting radial basis function (RBF) interpolation, performing further processing through an improved deep residual convolutional neural network, and finally optimizing a channel estimation result through super-resolution and denoising operation. The method effectively improves the channel estimation precision in low-mobility and high-mobility scenes, has a wide application prospect, and is particularly suitable for channel estimation in a high-speed Internet of Vehicles environment.

Description

Technical Field

[0001] The invention belongs to the field of channel estimation in vehicle communication and proposes a vehicle communication channel estimation algorithm based on convolutional neural network. Background Art

[0002] The IEEE 802.11p standard is a network protocol for vehicle-to-vehicle communication. It defines the physical layer specifications of vehicle-to-vehicle communication based on the orthogonal frequency division multiplexing (OFDM) scheme. Channel estimation is supported by pilot subcarriers and is widely used in wireless communications between vehicles and road facilities. It has become one of the key technologies for many real-time traffic applications. Since the wireless channel in the vehicle environment has a high Doppler frequency shift and a large delay spread, obtaining accurate channel mapping relationships in the vehicle communication system is the key to detection and demodulation. Traditional channel estimation methods include least squares (LS) and minimum mean square error (MMSE). Generally speaking, traditional channel estimation methods have problems such as multipath effects and noise interference in vehicle communication scenarios, lack high accuracy and robustness, and face significant performance degradation and high computational complexity in high mobility scenarios. Compared with traditional methods, deep learning (DL) algorithms have stronger nonlinear modeling capabilities and adaptability, and can effectively reduce the impact of various noise interferences on signals. Deep learning algorithms have been integrated into wireless communication physical layer applications and have achieved great success in improving overall system performance. They can be well applied to low-complexity estimators. In recent years, scholars have begun to study vehicle communication channel estimation based on deep learning.For example, based on feedforward neural network [Gizzini AK, Chafii MA survey on deep learningbased channel estimation in doubly dispersive environments[J].IEEE Access, 2022, 10:70595-70619] model, artificial neural network [Sattiraju R, Weinand A, Schotten HD. Channel estimation in C-V2X using deep learning[C] / / 2019 IEEE International Conference on Advanced Networks and Telecommunications Systems(ANTS).IEEE, 2019:1-5] model, long short-term memory network [Dos Reis AF, Medjahdi Y, Chang BS, et al. Low Complexity LSTM-NN-Based Receiver for Vehicular Communications in thePresence of High-Power Amplifier Distortions[J].IEEE Access, 2022, 10:121985-122000] model, deep neural network [Moon S, Kim H, Hwang I. Deep learning-based channel estimation and tracking for millimeter-wave vehicular communications[J].Journal of Communications and Networks, 2020, 22(3):177-184] model.

[0003] The literature [M. Soltani, V. Pourahmadi, A. Mirzaei, and H. Sheikhzadeh, "Deep Learning-Based Channel Estimation," IEEE Communications Letters, vol. 23, no. 4, pp. 652-655, 2019] proposed a CNN-based channel estimation model ChannelNet, which uses radial basis function (RBF) interpolation as the initial channel estimation, then regards the RBF estimated channel as a low-resolution image, and finally integrates super resolution CNN (SRCNN, super reso-lution CNN) and denoising CNN (DNCNN, denoising CNN) on the RBF estimated channel. ChannelNet uses two CNNs, which makes the computational burden heavy in real-time applications. Although it overcomes the limitations of traditional methods to a certain extent, it still has the problems of high latency and insufficient performance under rapidly changing channel conditions. Deep learning technology has a significant effect on channel estimation in vehicle communications. The channel is regarded as a low-resolution image. The super-resolution network can recover a higher-precision channel state from the low-resolution channel information, effectively reducing the error in channel estimation and making the channel estimation stable under low mobility conditions. As the vehicle speed increases, the channel is changing rapidly in high and ultra-high mobility scenarios, and there is more noise and interference. The use of a deep denoising network can effectively remove the noise and interference in the received signal, thereby improving the accuracy of channel estimation in high-speed vehicle movement scenarios. In actual communication environments, the presence of noise and interference will have a negative impact on channel estimation. Therefore, it is particularly important for application scenarios such as vehicle communications to have accurate learning capabilities and reduce channel estimation errors in the rapid changes of the channel. Summary of the invention

[0004] In order to further improve the denoising ability of CNN in channel estimation and solve the problem of insignificant performance improvement in high-speed vehicle scenarios, this paper proposes an improved channel estimation method that combines the deep super-resolution convolutional neural network VDSR and the blind denoising convolutional neural network RIDNet, and introduces three attention mechanisms, SENet, ECANet and CBAM, into these two networks. The introduction of channel attention can effectively enhance the expressiveness and selectivity of features, thereby improving the adaptability and accuracy of the model in complex environments. The contributions of this paper can be summarized as follows:

[0005] This paper proposes a channel estimation algorithm based on convolutional neural network. The technical solution adopted by this method is to combine the deep super-resolution convolutional neural network VDSR and the blind denoising convolutional neural network RIDNet, and introduce three attention mechanisms, SENet, ECANet and CBAM, into these two networks. The introduction of channel attention can effectively enhance the expressiveness and selectivity of features, thereby improving the adaptability and accuracy of the model in complex environments. The main process of this method is as follows:

[0006] We first use the radial basis function (RBF) interpolation method to perform preliminary channel estimation on the received signal. The channel feature is regarded as a two-dimensional low-resolution noise image, where each pixel represents the channel state information. Based on the preliminary estimation, the channel features are further processed using an improved deep residual convolutional neural network, which extracts the time-frequency features of the channel through convolutional layers, activation layers, and pooling layers. Then, we introduce three channel attention mechanisms based on the deep residual CNN network VDSR and RIDNet to enhance the feature representation ability. A channel attention module is added after multiple important convolutional layers to dynamically adjust the weights of key features to enhance the prominence of important features. We perform super-resolution reconstruction or denoising on the model to convert the low-resolution channel image into a high-resolution image to reduce the impact of noise. After multiple layers of convolution and processing, the model outputs a high-resolution channel estimation result, which is directly used for subsequent signal demodulation and data transmission. During the training process, a supervised learning method is used to optimize the network parameters by minimizing the error between the predicted channel estimate and the true channel, thereby improving the generalization ability of the model. Finally, according to the needs of vehicle applications, the transmission parameters are dynamically adjusted to adapt to different mobility conditions and data rate requirements.

[0007] In general, we use different network architectures for different mobility scenarios: In low mobility environments, the channel estimation results are fed to the VDSR network to take advantage of its deep feature extraction capabilities under relatively stable channel characteristics. At the same time, in order to improve model performance, three attention mechanisms, SENet, ECANet, and CBAM, are introduced into the VDSR network to enhance the effectiveness of feature expression and selection. In high and ultra-high mobility scenarios, the RIDNet network is used to cope with the challenge of rapid channel changes. SENet, ECANet, and CBAM attention mechanisms are also introduced to optimize feature fusion and improve estimation accuracy.

[0008] In summary, we improve the channel estimation network that is difficult to adapt to the fast-moving vehicle scenario by using a residual network and a channel attention mechanism to further enhance the network’s feature extraction capability. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1It is a schematic diagram of the overall structure of the method of the present invention.

[0010] Figure 2 It is the VDSR model network structure.

[0011] Figure 3 It is the network structure of the RIDNet model.

[0012] Figure 4 It is the network structure of the EAM model.

[0013] Figure 5 It is a structure composed of enhanced attention modules SENet, ECANet and CBAM.

[0014] Figure 6 These are the model structures of SENet, ECANet and CBAM introduced in VDSR.

[0015] Figure 7 These are the model structures of SENet, ECANet and CBAM introduced into RIDNet.

[0016] Figure 8 It is the loss function convergence curve of VDSR and RIDNet models.

[0017] Fig. 9 It is a comparison of BER performance under different mobility rates using QPSK modulation method.

[0018] Fig.10 It is a comparison of NMSE performance under different mobility rates using QPSK modulation method. DETAILED DESCRIPTION

[0019] The following further illustrates a vehicle communication channel estimation algorithm based on convolutional neural network proposed by the present invention in conjunction with the accompanying drawings.

[0020] We used the NVIDIA GeForce RTX 3060 hardware device, and all models were developed under the tensorflow2.2.0 deep learning development framework, using the official pre-trained network implementation.

[0021] We conducted comparative experiments on the improved VDSR and RIDNet networks at different vehicle mobility rates: According to the IEEE 802.11p standard, 10,000 experimental data set indexes were randomly generated, of which 8,000 were used for training and 2,000 were used for testing. When generating the data set, the experimental simulation parameters for different mobility scenarios were specified, as shown in Table 1: including the number of OFDM symbols, channel model, modulation method, and signal-to-noise ratio range. In terms of training models, the VDSR network is used in low mobility scenarios, while the RIDNet network is used in high and ultra-high mobility scenarios. We use normalized mean square error (NMSE) and bit error rate (BER) as evaluation indicators to compare the performance of different models.

[0022] Step 1: Use the radial basis function (RBF) interpolation method to perform preliminary channel estimation on the received signal

[0023] The radial basis function (RBF) interpolation method is used to perform preliminary channel estimation on the received signal. Specifically, RBF interpolation assigns weights to each data point by calculating the distance between each data subcarrier and its adjacent pilot subcarrier. This process can be viewed as mapping the channel state information into a two-dimensional space, where the channel characteristics of each data subcarrier are regarded as a pixel in the image. In this way, the channel is represented as a two-dimensional low-resolution image, and each pixel value in the image corresponds to the channel gain at a specific time and frequency. Since convolutional neural networks (CNNs) can effectively extract spatial features and patterns when processing images, this preliminary channel estimation method provides basic data for subsequent deep learning processing. Specifically, the low-resolution image generated by RBF interpolation contains the spatial and temporal correlations of the channel, and CNN can use this information for super-resolution reconstruction and denoising, thereby improving the accuracy and robustness of channel estimation.

[0024] Step 1 starts with the OFDM signal obtained from the receiving end. The signal can be expressed as in is the received signal, is the channel response, is the signal sent, is noise. Then, the received pilot symbol is estimated by least squares (LS) to obtain the preliminary channel response

[0025]

[0026] in is the received pilot symbol, is the transmitted pilot symbol. The pilot symbol is then interpolated using the RBF interpolation method to estimate the channel response for each data subcarrier:

[0027]

[0028] where ω j is the RBF weight, Φ is the RBF function (such as a Gaussian function): K f and K t are vectors of frequency and time indices, respectively, K pI is the number of pilot symbols. Through RBF interpolation, the channel estimate generated It can be viewed as a two-dimensional low-resolution image, where each pixel value corresponds to the channel gain at a specific time and frequency. In this way, the generated channel estimate It is regarded as a two-dimensional low-resolution image, and each pixel value in the image corresponds to the channel gain at a specific time and frequency. Finally, the generated low-resolution image is input into the deep learning model (VDSR and RIDNet) for further processing to improve the accuracy and denoising ability of channel estimation, thereby improving the overall channel estimation performance.

[0029] Step 2: Construct an improved deep residual convolutional neural network at different movement rates

[0030] Based on the preliminary estimation, an improved deep residual convolutional neural network is constructed: a VDSR model is constructed at low mobility rates, and a RIDNet model is constructed at high and ultra-high mobility rates.

[0031] The VDSR model consists of multiple convolution + ReLu layers and residual blocks, which deepens the network structure and uses a deeper network to use a larger receptive field to obtain context information. VDSR selects a 3×3 convolution kernel and a network with a depth of d, which has a receptive field of (2d+1)(2d+1). Its structure is as follows Figure 2 As shown. The first convolution layer of VDSR first extracts features from the input signal, while the last layer is used for signal reconstruction. Both convolution layers have convolution kernels of size 3×3×64, where the convolution kernels operate on a 3×3 spatial region across 64 channels (feature maps). In traditional SRCNN models, the input low-resolution image must pass through all layers until it reaches the output layer. For many weight layers, this is an end-to-end relationship that requires long-term memory. Therefore, it is easy to encounter problems of vanishing / exploding gradients. Residual learning is used in the VDSR model to suppress the occurrence of this gradient problem. Let x a represents the signal obtained after the interpolation estimation model, y aRepresents the output signal after convolution training. Given a training data set, the predicted value of the input signal is y a =f a (x a ), the formula for minimizing the mean square error is:

[0032]

[0033] Since the input and output signals are largely similar, a residual image is defined in which most values ​​are likely to be zero or small. The loss function of the VDSR model is calculated as

[0034]

[0035] Among them, f a (x a ) represents the network prediction result.

[0036] The RIDNet model adopts the structure of multi-level residual connection and enhanced attention module, which can capture and process motion blur and changes in the signal more quickly, and better extract channel features while reducing processing time delay. The RIDNet model provides a wide receptive field by performing kernel expansion in the initial two branch convolutions of each enhanced attention module, which can better restore signal details and edges, effectively process the lost detail information in the signal, and improve the quality of signal restoration. This paper selects the RIDNet model for training under high and ultra-high mobility rates. The RIDNet model consists of three main modules: feature extraction, feature learning residual on the residual module, and reconstruction. Its structure is as follows: Figure 3 shown.

[0037] The feature extraction module of the RIDNet model consists of only one convolutional layer, which can extract the initial features from the input noise signal. Its calculation formula is:

[0038] f0=M e (x″) (5)

[0039] Among them, represents the convolution of the noisy input image. Then the feature learning residual passed to the residual module is calculated as

[0040] f r =M fl (f0) (6)

[0041] Among them, f r represents the learned features, M fl (·) represents the residual component on the main feature learning residual, which is composed of cascaded enhancement attention modules (EAM), and its structure is as follows Figure 4 The EAM block is repeated four times in total, and the number of channels of each convolutional layer is fixed to 64. Except for the last convolutional layer in the enhanced residual block and the convolutional layer of the feature attention unit, the convolutional kernel size of each remaining convolutional layer is 3×3, and zero padding is used for smaller convolutional layers to achieve the same size of output feature maps.

[0042] Step 3: Introduce three channel attention mechanisms into the improved deep residual convolutional neural network to enhance the feature selection ability of the model.

[0043] The channel attention mechanism weights the features of each channel to enhance the network's attention to important features while suppressing the influence of unimportant features, thereby improving overall performance. SENet is a network architecture that introduces a channel attention mechanism to enhance the expressiveness of features. The model structure is as follows: Figure 5 (a) shows. The network obtains channel-level features through global average pooling operations and uses fully connected layers to generate corresponding channel weights, thereby achieving dynamic recalibration. This mechanism enables the network to automatically learn which feature channels are more important to the final task, thereby effectively improving performance. ECANet is optimized based on SENet, mainly focusing on improving computational efficiency. The model structure is shown in Figure 5 (b) It introduces a one-dimensional convolution to replace the fully connected layer, reducing the computational complexity while maintaining good performance. By modeling the local channel relationship, ECANet can capture important features more efficiently and further improve the performance of the network. CBAM proposes a module that combines channel and spatial attention at the same time, which can enhance feature representation more comprehensively. The model structure is shown in Figure 5 (c) shows that the channel attention module can focus on important feature channels, while the spatial attention module emphasizes the key areas in the feature map. This dual mechanism effectively captures important information in the image to improve the performance of the model in various visual tasks, ensuring its wide adaptability.

[0044] Although VDSR has achieved good results in achieving image super-resolution under the condition of low vehicle movement rate, its limitations are still obvious. First, although the model has a large depth, it lacks the ability to select important features and is easily disturbed by noise and irrelevant information. Secondly, VDSR fails to fully utilize the differences in spatial and channel features of the image, resulting in limitations in detail recovery and feature representation. To this end, the introduction of the channel attention mechanism can effectively emphasize important features, while suppressing the influence of irrelevant channels and enhancing the model's sensitivity to key areas. In addition, the introduction of spatial attention enables the model to focus on important areas in the vehicle communication channel, thereby better restoring details and better adapting to the changing channel environment. This paper proposes three improved models SEVDSR, ECVDSR and CBVDSR under low mobility, that is, channel attention SENet, ECANet and CBAM are introduced in VDSR respectively, and the model structures are as follows Figure 6 shown.

[0045] Similarly, although RIDNet performs well in image super-resolution, it still has some limitations when dealing with rapidly changing vehicle communication channel scenarios. RIDNet has limited capabilities in feature fusion, especially when dealing with complex features, which can easily lead to the loss of detail information. The network's selection bias for different channel features may affect the utilization efficiency of global information. The introduction of the channel attention mechanism can significantly enhance the feature selection ability of the model. Channel attention can automatically identify and emphasize important features and suppress irrelevant information; while spatial attention further enhances the focus on key areas to ensure that details are retained during the reconstruction process. This paper proposes three improved models SERIDNet, ECRIDNet and CBRIDNet under high mobility and ultra-high mobility, that is, channel attention SENet, ECANet and CBAM are introduced in RIDNet respectively, and the model structure is as follows Figure 7 shown.

[0046] In summary, we improve the channel estimation network for different vehicle movement rates and represent the channel as a two-dimensional low-resolution image, where each pixel value corresponds to the channel gain at a specific time and frequency. A large number of information pairs (low-resolution and high-resolution) are trained using a deep residual-based convolutional neural network to capture more channel feature details, thereby generating clearer high-resolution channel images. Through super-resolution reconstruction and denoising, the low-resolution channel image is converted into a high-resolution channel estimation result.

[0047] The above are the implementation steps of the vehicle communication channel estimation algorithm based on convolutional neural network of the present invention. Figure 8The loss function convergence curves of the VDSR and RIDNet models are described. As can be seen from the figure, the loss value of the model continues to decrease with the increase in the number of iterative training times, and the error value basically reaches the lowest value around 10 epochs, and tends to a relatively stable value after 25 epochs. This shows that the VDSR and RIDNet models in this paper have good convergence and can achieve stable results after a small number of training times. Fig. 9 , 10 The experimental results of our structure on randomly generated datasets are shown. Experiments show that in low mobility scenarios, the BER values ​​of the SEVDSR, ECVDSR, and CBVDSR models are 1 to 5 dB lower than those of the SRCNN model. The CBVDSR model shows the best performance, especially at 25 dB, when the bit error rate is 5.2 dB lower than that of SRCNN. In High and Very High mobility scenarios, the bit error rates of SERIDNet, ECRIDNet, and CBRIDNet are also reduced by 1 to 4 dB compared to the comparison model DNCNN. Similarly, the model we proposed still has a performance advantage in NMSE. In low mobility scenarios, the ECVDSR model is slightly better than SRCNN by about 1.4 dB when the signal-to-noise ratio SNR = 30 dB. On the other hand, in high mobility scenarios, the NMSE values ​​of the proposed SERIDNet, ECRIDNet, and CBRIDNet are about 1.44dB, 0.38dB, and 2.06dB lower than those of DNCNN, respectively, when SNR = 20dB, while in very high mobility scenarios, the NMSE of SERIDNet, ECRIDNet, and CBRIDNet are about -1.87dB, 0.74dB, and 1.81dB lower than those of DNCNN at the same SNR. Experiments have shown that our architecture can accurately and effectively improve the channel estimation network in a variable vehicle communication environment.

Claims

1. A vehicle communication channel estimation algorithm based on deep residual network, characterized in that: The steps are as follows: Step 1: Use the radial basis function (RBF) interpolation method to perform preliminary channel estimation on the received signal. Step 2: Construct an improved deep residual convolutional neural network at different movement rates. Step 3: Introduce three channel attention mechanisms into the improved deep residual convolutional neural network to enhance the feature selection ability of the model.

2. The vehicle communication channel estimation algorithm based on deep residual network according to claim 1 is characterized in that: An improved channel estimation method based on a deep residual CNN network is proposed, and different network architectures are used for different mobility scenarios. In a low mobility environment, the channel estimation results are fed to the VDSR network to utilize its deep feature extraction capabilities under relatively stable channel characteristics. At the same time, in order to improve the performance of the model, three attention mechanisms, SENet, ECANet and CBAM, are introduced into the VDSR network to enhance the effectiveness of feature expression and selection. In high and ultra-high mobility scenarios, the RIDNet network is used to cope with the challenge of rapid channel changes. SENet, ECANet and CBAM attention mechanisms are also introduced to optimize feature fusion and improve estimation accuracy.

3. The vehicle communication channel estimation algorithm based on deep residual network according to claim 1 is characterized in that: Due to the rapid change of the channel and noise interference. In a high-speed moving vehicle environment, the channel state changes rapidly, making channel estimation complex and difficult, especially in high mobility and ultra-high mobility scenarios. Traditional channel estimation methods are difficult to adapt, resulting in increased estimation errors, which in turn affects the overall performance. In our channel estimation network, the feature extraction network adopts an improved deep residual network that introduces three channel attention mechanisms, while the channel estimation module is newly established. The dataset in this paper is randomly generated according to the IEEE 802.11p standard and contains 10,000 experimental dataset indexes, of which 8,000 are used for training sets and 2,000 are used for test sets. When generating the dataset, experimental simulation parameters for different mobile scenarios are specified, such as the number of OFDM symbols, channel model, modulation order, and signal-to-noise ratio range. Based on this dataset, the channel estimation module is trained for 500 cycles to complete the training process using a deep residual network.

4. The channel estimation algorithm based on deep residual network according to claim 1, characterized in that: According to different vehicle movement conditions, the algorithm selects different network models for channel estimation, performs preliminary channel estimation through radial basis function interpolation method, converts the received signal into a low-resolution channel representation, and forms a two-dimensional time-frequency image. Then, the improved deep residual network is used to further process the preliminary estimation. Subsequently, the model uses the advantages of CNN to convert the low-resolution channel image into a high-resolution image and reduce the influence of noise through super-resolution reconstruction and denoising. Finally, after multiple layers of convolution and processing, the model outputs a high-resolution channel estimation result, which is directly used for subsequent signal demodulation and data transmission. The supervised learning method is used in the training process to optimize the network parameters by minimizing the error between the predicted channel estimate and the real channel, thereby improving the generalization ability of the model.

Images