OFDM system frequency offset and channel joint estimation algorithm based on multi-task neural network

Abstract

The application discloses an OFDM system frequency offset and channel joint estimation algorithm based on a multi-task neural network, and comprises the following steps: pre-processing a pilot signal at a receiving end to obtain the input of a joint estimation neural network; constructing a frequency offset and channel joint estimation multi-task neural network architecture; in an offline stage, a large number of differentiated data with different frequency offsets under different channels are constructed to train the network in an end-to-end mode; in an online stage, a symbol containing the influence of the frequency offset and the fading channel is input into the trained network, and the frequency offset value and the channel estimation value can be obtained simultaneously. The application proposes a frequency offset and channel joint estimation algorithm based on multi-task learning, and introduces an attention mechanism to eliminate the influence of the frequency offset estimation error on the channel estimation, so that the system calculation complexity is reduced, and the precision of the frequency offset estimation and the channel estimation is greatly improved compared with a traditional algorithm.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology and proposes a joint estimation algorithm for frequency offset and channel of OFDM system based on multi-task neural network. Background Technology

[0002] Orthogonal Frequency Division Multiplexing (OFDM) is a multi-carrier modulation technique that uses orthogonal subcarriers to transmit data in parallel. It boasts advantages such as high spectral efficiency and strong resistance to frequency-selective fading, and has been widely used in 4G LTE and 5G NR systems. However, OFDM systems are highly sensitive to carrier frequency offset, especially in complex future communication scenarios. Doppler shifts caused by relative motion and frequency offsets resulting from oscillation asynchrony between transceivers can disrupt the orthogonality between subcarriers, leading to inter-symbol interference and a sharp decline in system performance. Furthermore, wireless channels in high-speed mobile scenarios change drastically, exhibiting rapid, variable, and non-stationary characteristics, rendering traditional channel estimation schemes inapplicable. Therefore, accurate carrier synchronization and channel estimation techniques are of great significance for the practical application of communication systems in complex scenarios.

[0003] Traditional carrier frequency synchronization algorithms are mainly divided into frequency offset estimation algorithms based on cyclic prefix (CP) and frequency offset estimation algorithms based on training sequences. CP-based algorithms calculate the frequency offset based on the correlation of repeated data; training sequence-based algorithms estimate the carrier frequency offset by performing correlation operations between the received sequence and a locally known sequence, based on the location of the peak value. However, under low signal-to-noise ratio conditions, the estimation performance of traditional algorithms deteriorates sharply due to severe noise interference.

[0004] Traditional channel estimation algorithms typically employ pilot-assisted estimation methods. LS (Least Squares) is the most commonly used channel estimation method, which is computationally simple and has low complexity, but its performance is poor at low signal-to-noise ratios. The LMMSE (Linear Minimum Mean Squared Error) algorithm has significantly improved performance compared to the LS algorithm, but it cannot be applied in practice due to its high computational complexity and severe lack of prior knowledge.

[0005] In recent years, with the rapid development of deep learning, it has been introduced into wireless communication in pursuit of better algorithm performance. However, existing deep learning algorithms typically consider only one task—carrier synchronization or channel estimation—in isolation. For example, they might perform carrier frequency offset estimation in a Gaussian white noise or stationary channel, or perform channel estimation after frequency offset synchronization is complete, without considering the correlation between the two. However, in real-world systems, the effects of carrier frequency offset and fading channels coexist, making it necessary to study frequency offset synchronization and channel estimation simultaneously. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes a joint estimation algorithm for frequency offset and channel in OFDM systems based on a multi-task neural network, which effectively improves the estimation accuracy of frequency offset and channel.

[0007] The technical solution adopted in this invention includes the following steps:

[0008] (1) The pilot signal is preprocessed at the receiving end to obtain the input of the joint estimation neural network;

[0009] (2) Construct a multi-task neural network architecture for joint estimation of frequency offset and channel;

[0010] (3) In the offline stage, a large amount of differentiated data with different frequency offsets under different channels is constructed to train the network in an end-to-end manner;

[0011] (4) In the online phase, symbols that simultaneously include the effects of frequency offset and fading channel are input into the trained network, which can simultaneously obtain the frequency offset value and the channel estimate value.

[0012] Furthermore, this invention treats frequency offset estimation and channel estimation as two different but related tasks, and constructs a joint estimation neural network architecture based on multi-task learning.

[0013] Furthermore, the generation of input data in step (1) of the present invention involves performing least-squares estimation on the pilot signal to obtain the channel state information at the pilot, and then converting the complex data into real numbers as the input to the network.

[0014] Furthermore, the joint estimation multi-task neural network architecture described in step (2) of the present invention consists of a shared layer, a frequency offset estimation network, a channel estimation network, and a frequency offset cancellation module.

[0015] Furthermore, the frequency offset elimination module in the joint estimation multi-task neural network architecture described in step (2) of the present invention introduces an attention mechanism to reduce the impact of frequency offset error on channel estimation.

[0016] Furthermore, in step (3) of the present invention, the offline stage is to train the network in an end-to-end manner by constructing a large number of symbols with different frequency deviations under different channel states, and to construct a suitable loss function to balance the two sub-tasks of frequency offset estimation and channel estimation.

[0017] Furthermore, the online stage described in step (4) of the present invention can output the frequency offset estimate and the channel estimate simultaneously, which greatly reduces the algorithm complexity and improves the estimation accuracy.

[0018] Compared with the prior art, the present invention has the following advantages:

[0019] (1) This invention introduces multi-task learning into the joint estimation process of frequency offset and channel, and can output the frequency offset value and the channel estimation value at the same time.

[0020] (2) This invention proposes a frequency offset elimination method based on attention mechanism, which increases the channel’s receptive field to frequency offset by differentiated frequency offset data, effectively reducing the impact of frequency offset estimation error on channel estimation.

[0021] (3) The frequency offset and channel joint estimation algorithm based on multi-task neural network proposed in this invention has greatly improved the accuracy of frequency offset and channel estimation. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the implementation process of the present invention.

[0023] Figure 2 This is a diagram of the neural network structure constructed in this invention.

[0024] Figure 3 This section compares the RMSE estimation performance of the present invention with that of traditional frequency offset estimation algorithms.

[0025] Figure 4 This paper compares the RMSE estimation performance of the present invention with that of traditional channel estimation algorithms. Detailed Implementation

[0026] The method described in this invention will be explained in detail with reference to the accompanying drawings and specific embodiments.

[0027] Reference Figure 1 The implementation steps for this example are as follows:

[0028] (1) The pilot signal is preprocessed at the receiving end to obtain the input of the joint estimation neural network;

[0029] After the transmitted signal passes through the multipath fading channel, least squares (LS) estimation is first performed at the receiver to obtain the channel frequency domain response at the pilot, which is expressed as (1).

[0030] H p，LS =Y p / X p (1)

[0031] Where X p Given the known pilot signal, Y p The pilot signal at the receiving end is affected not only by carrier frequency offset but also by multipath fading channels. The impact of both factors is reflected in the differences between the transmitting and receiving signals, and both are further affected by Gaussian white noise. Therefore, we consider using the pilot signal to jointly estimate both frequency offset and channel characteristics.

[0032] However, since neural networks cannot handle complex data, we need to modify H. p，LS Preprocessing is performed by extracting H p，LS The real and imaginary parts are separated and then concatenated in the second dimension to obtain real data, which is used as the input data for the joint estimation neural network.

[0033] (2) Construct a multi-task network architecture for joint estimation of frequency offset and channel;

[0034] Considering that frequency offset estimation and channel estimation are two coupled tasks, this invention proposes a joint frequency offset and channel estimation algorithm based on multi-task learning. The constructed neural network is as follows: Figure 2 As shown, it mainly consists of a shared layer, a frequency offset estimation network, a channel estimation network, and a frequency offset cancellation module. The shared layer connects the frequency offset estimation network and the channel estimation network in parallel, and the frequency offset estimation network and the channel estimation network are connected through the frequency offset cancellation module.

[0035] The main function of the shared layer is to extract some common features from the two tasks of frequency offset estimation and channel estimation. The network consists of two convolutional layers. The first layer has a kernel size of (6, 2) and 16 channels, and uses the LeakyReLU activation function. The second layer also has a kernel size of (6, 2) and 16 channels, and uses the LeakyReLU activation function. The expression for the LeakyReLU function is shown in (2).

[0036]

[0037] The frequency offset estimation network is based on a fully connected layer network. The first two fully connected layers have 400 and 200 neurons, respectively, and use the ReLU activation function. Both layers employ Dropout technology. The last fully connected layer has 100 neurons.

[0038] The channel estimation network uses a convolutional neural network as its core architecture. The first layer has a kernel size of (6, 2) and 32 channels, and uses the LeakyReLU activation function. The second layer has a kernel size of (6, 2) and 64 channels, and also uses the LeakyReLU activation function.

[0039] The frequency offset cancellation module consists of three fully connected layers. The first fully connected layer is connected to the output of the third fully connected layer of the frequency offset estimation network, the second fully connected layer is connected to the input of the first convolutional layer of the channel estimation network, and the third fully connected layer is connected to the input of the second convolutional layer of the channel estimation network. By multiplying, weighted values ​​are assigned, and the channel's receptive field to frequency offset is increased by the differentiated frequency offset data, so as to guide the attention mechanism.

[0040] (3) In the offline stage, a large amount of differentiated data with different frequency offsets under different channels is constructed to train the network in an end-to-end manner;

[0041] In the offline phase, a binary bitstream is randomly configured, with cell IDs ranging from 0 to 2. The channel is a randomly selected multipath fading channel, and the normalized carrier frequency offset ranges from -0.5 to 0.5. A minimum resolution of 0.01 is used to generate 50,000 sets of sample data, which are then divided into training and validation sets in an 8:2 ratio. Each sample is a pilot signal containing frequency offset and fading channel information. The corresponding normalized carrier frequency offset and the actual channel value serve as the labels for the frequency offset estimation and channel estimation networks, respectively.

[0042] The Adam optimizer was selected as the neural network optimizer, with an initial learning rate of 0.001, a batch size of 256, and a maximum training iteration count of 40. A joint estimation loss function was constructed to balance the weights between the frequency offset estimation and channel estimation tasks.

[0043] The loss function for frequency offset estimation is the cross-entropy function, as shown in expression (3).

[0044]

[0045] Where N is the number of samples, K is the number of categories, and t ij y is an indicator that the i-th sample belongs to the j-th class. ij It is the output of sample i of category j.

[0046] The loss function for channel estimation is the mean square error function, as shown in expression (4).

[0047]

[0048] Where N is the number of samples, x i y is the actual channel value.i This is the channel estimate. The network is trained and optimized using training and validation set data.

[0049] (4) In the online phase, symbols that simultaneously include the effects of frequency offset and fading channel are input into the trained network, which can simultaneously obtain the frequency offset value and the channel estimate value.

[0050] After network training is completed, a certain number of symbols containing both frequency offset and fading channel effects are generated as a test set. During the online phase, the test set data is input into the trained network, and the corresponding frequency offset value and channel estimate value can be obtained simultaneously through the current input. The effectiveness of both frequency offset estimation and channel estimation is measured by the root mean square error, expressed as equations (5) to (6).

[0051]

[0052] Where ε i For the true normalized frequency offset value, H is the frequency offset estimate. i This is the actual channel value. R is the channel estimate, and R is the number of samples in the test set.

[0053] The frequency offset and channel estimation algorithm based on a multi-task neural network proposed in this invention reduces the computational complexity in the online stage and improves the accuracy of frequency offset estimation and channel estimation.

[0054] The effects of this invention can be further illustrated by the following simulations:

[0055] 1. Simulation conditions

[0056] The simulation uses an FDD-LTE downlink transmission system with a transmission bandwidth of 10MHz, an FFT point count of 1024, cyclic prefix lengths of 80 and 72, a subcarrier spacing of 15kHz, a system sampling rate of 15.36MHz, and a modulation scheme of 16QAM. The channel model is an extended vehicle channel model, with multipath delays of 0, 30, 150, 310, 370, 710, 1090, 1730, and 2510 nanoseconds, and power attenuations of 0, -1.5, -1.4, -3.6, -0.6, -9.1, -7.0, -12.0, and -16.9dB, respectively.

[0057] 2. Simulation Content

[0058] Figure 3This paper compares the root mean square error (RMSE) performance of frequency offset estimation between the present invention and traditional frequency offset estimation algorithms under different signal-to-noise ratios (SNRs). CP represents the CP-related frequency offset estimation algorithm, PSS represents the PSS-related frequency offset estimation algorithm, and DL represents the frequency offset estimation performance in the proposed multi-task neural network-based joint frequency offset and channel estimation algorithm. It can be seen that the RMSE of all algorithms decreases with increasing SNR. The algorithm proposed in this invention significantly outperforms the CP and PSS-related algorithms, resulting in more accurate frequency offset estimation.

[0059] Figure 4 This paper compares the root mean square error (RMSE) performance of the proposed channel estimation algorithm with that of traditional channel estimation algorithms under different signal-to-noise ratios. LS represents the LS channel estimation algorithm, LMMSE represents the LMMSE channel estimation algorithm, and DL represents the channel estimation performance of the proposed multi-task neural network-based joint frequency offset and channel estimation algorithm. It can be seen that the channel estimation algorithm of this invention significantly outperforms the LS and LMMSE algorithms.

[0060] The above is merely a further description of the present invention and is not intended to limit the implementation and application of this patent. All equivalent implementations of the present invention should be included within the scope of the claims of this patent.

Claims

1. A joint estimation algorithm for frequency offset and channel of OFDM system based on multi-task neural network, characterized in that, Includes the following steps: (1) The pilot signal is preprocessed at the receiving end to obtain the input of the joint estimation neural network; (2) Construct a multi-task neural network architecture for joint estimation of frequency offset and channel. The architecture consists of a shared layer, a frequency offset estimation network, a channel estimation network, and a frequency offset elimination module. The shared layer connects the frequency offset estimation network and the channel estimation network in parallel. At the same time, the frequency offset estimation network and the channel estimation network are connected through the frequency offset elimination module. The frequency offset elimination module consists of three fully connected layers. The first fully connected layer is connected to the output of the third fully connected layer of the frequency offset estimation network, the second fully connected layer is connected to the input of the first convolutional layer of the channel estimation network, and the third fully connected layer is connected to the input of the second convolutional layer of the channel estimation network. Weighted values ​​are assigned by multiplication. The differentiated frequency offset data increases the receptive field of the channel to frequency offset, so as to realize the guidance of the attention mechanism. (3) In the offline stage, a large amount of differentiated data with different frequency offsets under different channels is constructed to train the network in an end-to-end manner; (4) In the online phase, symbols that simultaneously include the effects of frequency offset and fading channel are input into the trained network, which can simultaneously obtain the frequency offset value and the channel estimate value.

2. The joint estimation algorithm for frequency offset and channel of OFDM system based on multi-task neural network as described in claim 1, characterized in that, By treating frequency offset estimation and channel estimation as two different but related tasks, a joint estimation neural network architecture based on multi-task learning is constructed.

3. The joint estimation algorithm for frequency offset and channel of OFDM system based on multi-task neural network as described in claim 1, characterized in that, The input data in step (1) is generated by performing least squares estimation on the pilot signal to obtain the channel state information at the pilot, and then converting the complex data into real numbers as the input to the network.

4. The joint frequency offset and channel estimation algorithm for OFDM systems based on multi-task neural networks as described in claim 1, characterized in that, The offline stage described in step (3) trains the network in an end-to-end manner by constructing a large number of symbols with different frequency deviations under different channel conditions, and constructs a suitable loss function to balance the two sub-tasks of frequency offset estimation and channel estimation.

5. The joint frequency offset and channel estimation algorithm for OFDM systems based on multi-task neural networks as described in claim 1, characterized in that, In the online stage described in step (4), the frequency offset estimate and the channel estimate can be output simultaneously, which greatly reduces the algorithm complexity and improves the estimation accuracy.

Images