An OFDM Channel Estimation Method Based on Spatial Specific Neural Network

Abstract

The present invention belongs to the field of communication technologies, and particularly relates to an OFDM channel estimation method based on a spatial specific neural network. The present invention constructs a channel estimation neural network for a fast time-varying OFDM channel, solves the problem of limited ICI suppression performance of OFDM channel estimation by a deep learning method based on implicit ICI suppression. The constructed network includes a preprocessing network and a channel refinement network composed of spatial specific neural network operators. The preprocessing network eliminates ICI from the input initial channel estimate and its auxiliary data, and the channel refinement network further suppresses the influence of residual noise to improve the accuracy of the channel estimate value. At the same time, a loss weighting mechanism is applied during the training process to significantly improve the adaptability of the network.

Description

Technical Field

[0001] The present invention belongs to the field of communication technologies, and particularly relates to an OFDM channel estimation method based on a spatial specific neural network. Background Art

[0002] An Orthogonal Frequency Division Multiplexing (OFDM) system has good anti-multipath fading characteristics and high spectral efficiency, and is widely used in broadband wireless communication systems. In a communication scenario where the transceiver moves relatively fast, the huge Doppler frequency shift generated by the high-speed movement of the terminal platform will cause serious inter-carrier interference (ICI), thus seriously destroying the orthogonality between OFDM subcarriers and leading to a decrease in the accuracy of channel estimation results. In such a high-speed mobile channel, traditional least square (LS) channel estimation based on statistical signal processing, linear minimum mean square error (LMMSE) channel estimation, and noise reduction methods based on Discrete Fourier Transform (DFT) are difficult to obtain channel estimation performance that meets the system requirements. Given the successful application of deep learning methods in physical layer communication, a data-driven deep learning method can be used to improve the OFDM channel estimation performance in a high-speed mobile environment.

[0003] The OFDM channel estimation method based on deep learning uses the initial channel estimation result obtained by the traditional method as the input of the neural network, and the output after the non-linear processing of the neural network is used as the final estimation result of the channel coefficients. In the early OFDM system receivers based on deep learning, channel estimation and signal detection were implemented using the same neural network (Ye H, Li G Y, Juang B H. Power of deep learning for channel estimation and signal detection in OFDM systems[J]. IEEE Wireless Communications Letters, 2017, 7(1): 114-117.). Subsequent research work found that using an independent neural network for channel estimation can achieve better system performance, such as ComNet (Gao X, Jin S, Wen C K, et al. ComNet: Combination of deep learning and expert knowledge in OFDM receivers[J]. IEEE Communications Letters, 2018, 22(12): 2627-2630.), ChannelNet (Soltani M, Pourahmadi V, Mirzaei A, et al. Deep learning-based channel estimation[J]. IEEE Communications Letters, 2019, 23(4): 652-655.), ReEsNet (Li L, Chen H, Chang H H, et al. Deep residual learning meets OFDM channel estimation[J]. IEEE Wireless Communications Letters, 2019, 9(5): 615-618.) and CRCENet (Peng Q, Li J, Shi H. Deep Learning Based Channel Estimation for OFDM Systems With Doubly Selective Channel[J]. IEEE Communications Letters, 2022, 26(9): 2067-2071.), etc. However, the above methods do not explicitly process the ICI caused by high-speed movement, and the accuracy of channel estimation still needs to be further improved.Therefore, Sun Yi et al. used two fully connected layers to first process the initial channel estimation to eliminate the impact of ICI (Sun Y, Shen H, Du Z, et al. ICINet: ICI-Aware Neural Network Based Channel Estimation for Rapidly Time-Varying OFDM Systems[J]. IEEE Communications Letters, 2021, 25(9): 2973-2977.), which greatly reduced the impact of ICI on channel estimation. However, most of these deep learning-based OFDM channel estimation methods implicitly handle the ICI impact as part of the channel, or explicitly handle ICI with a simple neural network, rather than designing a proprietary neural network for explicit processing according to the characteristics of ICI, which restricts the performance of channel estimation. Summary of the Invention

[0004] Aiming at the deficiencies of the existing traditional frequency-domain channel estimation methods based on statistical signal processing in OFDM systems, the present invention provides an OFDM channel estimation method based on explicitly eliminating the impact of ICI using a spatial-specific neural network operator and further suppressing the impact of residual noise based on the ResNet structure.

[0005] In a typical communication scenario, the channel frequency response (CFR) at different spatial positions in the time-frequency grid is a sample of a random process, and its specific characteristics vary with the spatial position of the time-frequency grid. However, the convolutional neural network (CNN) shares the same set of kernels for feature extraction at all spatial positions, and this operation has a poor coupling with the characteristics of the above CFR time-frequency grid. Contrary to the spatial invariance of CNN, the idea of the spatial-specific neural network operator Involution (Li D, Hu J, Wang C, et al. Involution: Inverting the inherence of convolution for visual recognition [C] / / Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition. 2021: 12321-12330.) is to generate different kernels at specific spatial positions. This characteristic fits well with the characteristics of the CFR time-frequency grid. Therefore, the present invention designs a spatial-specific neural network operator for fast time-varying OFDM system CFR estimation based on this idea. The preprocessing network InvoPreNet (Modified Involution-Based Preprocessing Subnetwork) composed of this operator first performs ICI cancellation on the input initial channel estimate and its auxiliary data, and then sends the output result to the channel refinement network ResCSINet (Residual Channel State Information Subnetwork) to further suppress the influence of residual noise and improve the accuracy of the channel estimate value. The present invention uses a neural network operator designed based on the spatial-specific idea to overcome the defects that traditional frequency-domain channel estimation methods and the aforementioned deep learning-based channel estimation methods are difficult to efficiently handle ICI and affect the accuracy of OFDM channel estimation.

[0006] The technical solution of the present invention is as follows:

[0007] An OFDM channel estimation method based on a spatial-specific neural network, comprising the following steps:

[0008] S1. Construct a training data set:

[0009] According to the received pilot signal and the transmitted pilot symbol, obtain the initial value of the channel estimate Among them, the OFDM subframe consists of N subcarriers and T OFDM symbols, where the symbol index t = 1,..., T, and the subcarrier index n = 1,..., N; from the pilot signal and the initial value of channel estimation the transmitted signal is estimated The initial value of channel estimation within an OFDM subframe the received signal and the estimated transmitted signal The real and imaginary parts form a multi-channel tensor as a training data:

[0010]

[0011] where denotes concatenation in the third dimension, Re(·) denotes taking the real part, and Im(·) denotes taking the imaginary part; multiple training data are obtained to form a training dataset;

[0012] S2. Construct a channel estimation network model, including a preprocessing network and a channel refinement network composed of spatial specific neural network operators;

[0013] The processing method of the preprocessing network for the training data is to first pad zeros to Z:

[0014]

[0015] where denotes concatenation in the first dimension, and N ICI is the set number of unilateral subcarriers;

[0016] According to N ICI the kernel size K = 2N ICI +1 is determined, and slices are taken from Z according to the determined kernel size Then, according to the defined a kernel is generated

[0017]

[0018] where denotes the flatten() operation of the neural network to flatten it into a vector form, W 0 and W 1 denote the linear transformations of two fully connected layers, σ(·) denotes the non-linear activation function ReLU(·), denotes the reshape operation;

[0019] The kernel is combined with the corresponding slice Perform a dot product summation operation to obtain the channel estimation result that eliminates the influence of ICI at this position.

[0020]

[0021] Among them, is the channel estimation result output by the preprocessing network, and ∑(·) represents summing all elements of the tensor;

[0022] The channel refinement network is a residual structure neural network composed of two one-dimensional CNNs and a non-linear activation function ReLU(·). The channel refinement network is used to recombine the real and imaginary parts of to obtain the final channel estimation value

[0023] S3. Use the training data set constructed in S1 to train the channel estimation network model constructed in S2, and train the neural network using a loss function with a weighted mechanism:

[0024]

[0025] Among them, f X (·) and Θ X represent the transformation formula of the neural network and the corresponding set of network parameters, and represent the training sample set of the neural network and the corresponding set size, H represents the true channel frequency domain response, and λ is a coefficient related to the signal-to-noise ratio; thus, a trained channel estimation network model is obtained;

[0026] S4. After processing the acquired signal using the method in S1, input it into the trained channel estimation network model to obtain the channel estimation result.

[0027] The beneficial effects of the present invention are as follows: The present invention constructs a channel estimation neural network for fast time-varying OFDM channels, and solves the problem of limited ICI suppression performance in OFDM channel estimation based on implicit ICI suppression deep learning methods. At the same time, the application of the loss weighting mechanism during training significantly improves the adaptability of the network. Brief Description of the Drawings

[0028] Figure 1 is a schematic diagram of an OFDM pilot symbol provided by an embodiment of the present invention;

[0029] Figure 2 is the OFDM channel estimation network structure based on a spatial specific neural network operator provided by an embodiment of the present invention;

[0030] Figure 3 is different N provided by an embodiment of the present invention ICINMSE performance comparison of the lower InvoPreNet (28 pilot symbols in total within an OFDM subframe, N p = 2, T p = 14);

[0031] Figure 4 is the NMSE performance comparison of different training strategies provided by the embodiments of the present invention (28 pilot symbols in total within an OFDM subframe, N p = 2, T p = 14);

[0032] Figure 5 is the comparison between InvoEsNet provided by the embodiments of the present invention and classical statistical signal processing methods;

[0033] Figure 6 is the robust performance of InvoEsNet provided by the embodiments of the present invention at different moving speeds. Detailed implementation manners

[0034] The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0035] Embodiment

[0036] In this example, a single-antenna OFDM system based on the 5G standard protocol is adopted, and its configuration is shown in Table 1. Among them, according to the 5G standard protocol, the number of active subcarriers is set to N active = 288. The pilot data uses the Zadoff-Chu (ZC) sequence, and the pilot pattern adopts the common comb-shaped pilot mode, that is, the pilot symbols are continuously inserted in time for T p = 14, and are equally spaced in the frequency domain for N p = 2 pilot symbols, as Figure 1 shown. The channel adopts a two-path water-air-ground channel model, and the normalized maximum multi-frequency offset of the Jakes Doppler spectrum is

[0037] Table 1 Parameter configuration of the single-antenna OFDM system

[0038] Parameter Configuration Carrier frequency C band Subcarrier spacing (△f) 240 kHz Number of subcarriers (N) 512 Number of OFDM subframe symbols (T) 14 Modulation method 16QAM

[0039] The implementation steps of this example are as follows:

[0040] Step 1: Use the LS estimation algorithm to calculate the initial channel estimation value at the pilot position for the received pilot signal and the transmitted pilot symbol . Then, apply the linear interpolation algorithm to obtain all the initial channel estimation values including the data symbol positions as shown in the following formula:

[0041]

[0042] in, represents the Hadamard product (i.e., element-wise multiplication), represents Hadamard inversion (i.e., element-wise inversion), and interpolation(·) represents linear interpolation.

[0043] Step 2: Based on the received signal Y t and the initial value of channel estimation The estimated transmitted signal is calculated using the Zero-Forcing (ZF) equalization algorithm. As shown below:

[0044]

[0045] Wherein, Demod(·) represents the demodulation operation of the communication system.

[0046] Step 3: Reassemble the initial channel estimation value within the OFDM subframe Estimated transmitted signal and the received signal Y t , where t = 1, ..., 14. Specifically, all channel estimation initial values ​​in the OFDM subframe are Receiving Signals and the estimated transmitted signal The real and imaginary parts of like Figure 2 shown.

[0047] Step 4: Construct a preprocessing network InvoPreNet consisting of space-specific neural network operators.

[0048] InvoPreNet consists of a specially designed space-specific neural network operator, such as Figure 2 First, according to the set number of single-side subcarriers N ICI , perform a circular zero padding operation. Then, given a coordinate (n, t) that needs to be calculated, take a slice from Z with that position as the center Where K = 2N ICI +1. Next, By kernel generation function Generate the kernel of InvoPreNet The obtained core With the corresponding slice Perform the dot product summation operation and output the final result of InvoPreNet. The network parameters of InvoPreNet are summarized in Table 2.

[0049] Table 2 Network Parameters of InvoPreNet

[0050] Parameter Configuration Zero padding <![CDATA[N ICI > Kernel size <![CDATA[K×1 (K = 2N ICI + 1)]]> Moving step size 1 Activation function ReLU(·) <![CDATA[Fully connected layer 1W 0 > (K×6×2)×(K×6) <![CDATA[Fully connected layer 2W 1 > (K×6×2)×(K×6×2)

[0051] InvoPreNet is trained using the Adam optimizer, with the initial learning rate set to 0.001 and the batch size set to 128. The weighted mean squared error (WMSE) loss function used to train the InvoPreNet neural network is as follows:

[0052]

[0053] where λ is defined as shown below:

[0054]

[0055] Step 5: Construct the channel refinement network ResCSINet based on the residual network.

[0056] To focus on the network extracting high-frequency features along the time axis, a simple-structured channel refinement network ResCSINet is constructed based on one-dimensional CNN. Specifically, the main structure of ResCSINet consists of two one-dimensional CNNs with a kernel size of 3 and both the number of input and output channels being 512, and a non-linear ReLU activation layer is inserted between the two convolutional layers. The input of ResCSINet is the channel estimation result output by InvoPreNet The output result is the sum of the input data and the output data of the last convolutional layer, as Figure 2 shown. This output result is recombined according to the real and imaginary parts of the data to obtain the final channel estimation value The network parameters of ResCSINet are summarized in Table 3.

[0057] Table 3 Network Parameters of ResCSINet

[0058]

[0059] ResCSINet is trained using the Adam optimizer, with the initial learning rate set to 0.001 and the batch size set to 128. The WMSE loss function used to train the ResCSINet neural network is defined as follows:

[0060]

[0061] where f Res (·) and Θ Res represent the transformation formula of ResCSINet and the corresponding set of network parameters, and the definition of λ is as shown in Equation (8).

[0062] Step 6: Generate sufficient data according to the OFDM system parameters set in this embodiment, and execute Steps 1 to 3 to convert the generated data into a data set required for neural network training. Then, divide this data set into a training set, a validation set, and a test set according to the ratio of 6:2:2, and input them into two neural networks, InvoPreNet and ResCSINet, for training. For the sake of easy expression, the combined structure of the two networks, InvoPreNet and ResCSINet, is called "InvoEsNet".

[0063] Step 7: Perform channel estimation using the obtained trained network.

[0064] The example of the present invention first compares the NMSE performance of InvoPreNet under different N ICI as shown below. It can be seen that the NMSE performance of the channel estimation after eliminating the ICI influence by InvoPreNet is significantly improved compared with the traditional LS estimation algorithm, and as N Figure 3 increases from 0 to 4, the performance gradually converges, and finally the performance of N ICI = 3 is almost the same as that of N ICI = 4. Therefore, considering the trade-off between performance improvement and computational complexity, this embodiment selects N ICI = 3 as the appropriate parameter setting. In addition, the figure also shows the effect of CNN with the same kernel size on ICI cancellation when N ICI = 3. It can be seen that, consistent with the analysis in the invention content, due to the poor coupling with the characteristics of the CFR time-frequency grid, CNN has almost no effect on eliminating ICI. ICI The figure shows the comparison of NMSE performance of different training strategies. It can be seen that the performance of the network trained under fixed high / low signal-to-noise ratio or simple mixing of multiple signal-to-noise ratios is poor. However, the network based on the weighted training strategy (i.e., using the WMSE loss function, denoted as Weighted InvoEsNet) can almost achieve the performance of testing under the training signal-to-noise ratio (for example: the 15dB test performance of InvoEsNet uses the network optimized by the training data with a signal-to-noise ratio of 15dB).

[0065] Figure 4 The figure shows the comparison of NMSE performance of different training strategies. It can be seen that the performance of the network trained under fixed high / low signal-to-noise ratio or simple mixing of multiple signal-to-noise ratios is poor. However, the network based on the weighted training strategy (i.e., using the WMSE loss function, denoted as Weighted InvoEsNet) can almost achieve the performance of testing under the training signal-to-noise ratio (for example: the 15dB test performance of InvoEsNet uses the network optimized by the training data with a signal-to-noise ratio of 15dB).

[0066] Figure 5Shows the comparison between InvoEsNet and classical statistical signal processing methods. Under the condition of using the same number of pilots (i.e., a total of 2×14 = 28 pilots set in the embodiment), the performance of InvoEsNet is far superior to that of traditional LS and LMMSE algorithms. In addition, when the number of pilots is small, the performance of the traditional LMMSE method is only slightly better than that of the LS method at low signal-to-noise ratios and almost approaches at high signal-to-noise ratios. However, after increasing the number of pilots to a certain extent (specifically, a total of 24×14 = 336 pilots as shown in the figure), the performance of the LMMSE algorithm is far superior to that of the LS algorithm and is also slightly better than InvoEsNet at low signal-to-noise ratios (SNR < 15 dB). However, it can be seen that there is still a huge gap between the performance of the traditional LS algorithm and InvoEsNet, and the performance of the LMMSE algorithm is almost the same as that of InvoEsNet at high signal-to-noise ratios (SNR > 15 dB). However, implementing the LMMSE algorithm requires prior information of the channel correlation matrix, which is difficult to obtain in the actual implementation process. In summary, although the LMMSE algorithm is slightly better than InvoEsNet at low signal-to-noise ratios (SNR < 15 dB) and almost the same as InvoEsNet at high signal-to-noise ratios (SNR > 15 dB) by increasing the pilot overhead, it also pays a high additional cost (12 times the pilots compared to InvoEsNet) and requires channel prior information. Therefore, compared with the traditional LS and LMMSE algorithms, the proposed InvoEsNet has obvious performance advantages with a small number of pilots and does not require difficult-to-implement conditions such as channel prior information.

[0067] Figure 6 Shows the robust performance of InvoEsNet and the existing ICINet at different mobile speeds. Among them, InvoEsNet and ICINet are trained under the condition of f d_n = 0.225 and tested at f d_n = 0.25, 0.2, 0.1, 0.05 respectively. The results show that the performance of InvoEsNet improves significantly as the mobile speed decreases (i.e., the decrease of f d_n ) and can maintain good channel estimation performance within a wide range of mobile speeds. In addition, the performance improvement of InvoEsNet is greater than that of the existing ICINet, and the performance of InvoEsNet is always superior to that of the existing ICINet at equal mobile speeds (i.e., equal f d_n ).

Claims

1. An OFDM channel estimation method based on a spatial specific neural network, characterized in that, it includes the following steps: S1. Construct a training data set: Based on the received pilot signal and the transmitted pilot symbol, obtain the initial value of channel estimation where an OFDM subframe consists of N subcarriers and T OFDM symbols, symbol index t = 1,..., T, subcarrier index n = 1,..., N; from the pilot signal and the initial value of channel estimation estimate the transmitted signal For the initial value of channel estimation within one OFDM subframe the received signal and the estimated transmitted signal compose a multi-channel tensor with the real and imaginary parts as a training data: wherein represents cascading in the third dimension, Re(·) represents taking the real part, Im(·) represents taking the imaginary part; obtaining multiple training data to form a training data set; S2. Construct a channel estimation network model, including a preprocessing network and a channel refinement network composed of spatial specific neural network operators; The processing method of the preprocessing network for the training data is to first pad zeros to Z: Among them, indicates cascading in the first dimension, and N ICI is the set number of single-sided subcarriers; According to N ICI Determine that the kernel size is K = 2N ICI + 1, and take slices from Z according to the determined kernel size Then, according to the defined Generate the kernel Among them, indicates that the flatten() operation of the neural network will be flattened into a vector form, W 0 and W 1 represent the linear transformations of two fully connected layers, and σ(·) represents the non-linear activation function ReLU(·), represents the reshape operation; Multiply the kernel by the corresponding slice and perform a dot product summation operation to obtain the channel estimation result that eliminates the ICI effect at the (n, t) position Among them, is the channel estimation result output by the preprocessing network, and ∑(·) represents summing over all elements of the tensor; The channel refinement network is a residual structure neural network composed of two one-dimensional CNNs and a non-linear activation function ReLU(·). The channel refinement network is used to recombine the real and imaginary parts of S3. Use the training data set constructed in S1 to train the channel estimation network model constructed in S2, and train the neural network with a loss function using a weighted mechanism: Among them, f X (·) and Θ X represent the transformation formula of the neural network and the corresponding set of network parameters, and represent the training sample set of the neural network and the corresponding set size, H represents the true channel frequency response, and λ is a coefficient related to the signal-to-noise ratio; thus, a trained channel estimation network model is obtained; S4. After processing the obtained signal using the method in S1, input it into the trained channel estimation network model to obtain the channel estimation result.

Images