Method for improving wireless channel estimation precision in machine learning through noise reduction processing

Abstract

The invention provides a method for improving wireless channel estimation precision in machine learning through noise reduction processing, and the method comprises the steps: building a deep neural network, and taking a training set after DFT denoising as the input of the neural network; in the training stage, the mean square error of the estimated value and the actual channel response serves as a loss function, and optimized model parameters are obtained by minimizing the loss function. The invention aims to solve the problem that the channel estimation result error of an OFDM system in a low signal-to-noise ratio environment is large, for example, the noise suppression performance of a model is poor due to the fact that the amplification effect of a least square method on noise is not considered in the previous deep learning training process. The method has the advantages that channel estimation in a high-noise environment can be completed by adopting fewer pilot frequency subcarrier resources, and the estimation precision is obviously superior to that of a traditional channel estimation method. The method is suitable for a wireless communication system in a high-noise environment, and the communication efficiency is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communications, and in particular relates to a method for realizing channel estimation in a wireless communication process through a deep neural network after noise reduction processing. Background Art

[0002] Orthogonal frequency division multiplexing (OFDM) technology is widely used in the field of communications due to its excellent resistance to multipath fading and high spectrum utilization. In this system, the number of subcarriers increases significantly with the demand for communication throughput, and how to obtain accurate frequency response becomes a prominent issue. The present invention aims to improve the communication efficiency based on OFDM communication systems. As the number of user terminals increases, the noise between users increases, and the present invention proposes an effective response plan. The existing technology is not robust enough under low signal-to-noise ratio conditions because the impact of noise in channel data on the estimation results is ignored during the training process of the deep learning model. Summary of the invention

[0003] Purpose of the invention: In order to solve the above technical problems, the present invention proposes a method for improving the accuracy of wireless channel estimation in machine learning through noise reduction processing, which can suppress the interference of noise in the training set on training and ensure the performance of the DNN channel estimation model in a high noise environment.

[0004] Technical solution: A channel estimation solution that combines DFT denoising technology with deep neural networks. Before training the DNN model, the data set is preprocessed, the interpolated channel response is converted to the time domain through inverse Fourier transform, and then the sampling point channels after the maximum delay are set to zero, thereby removing noise interference in the training set, and then converted into frequency points through Fourier transform as the input of the deep neural network model, thereby suppressing the impact of noise on the model training effect.

[0005] The specific steps include:

[0006] Step 1: Convert the received baseband signal into the frequency domain and extract the known signal Y from it;

[0007] Step 2: Based on the pilot signal, the receiving end estimates the channel response at the pilot position;

[0008] Step 3: Estimate the channel responses at all positions based on the channel responses at the pilot positions;

[0009] Step 4: Divide the time domain channel into two parts, the effective channel and the noise, according to the maximum path delay. The sampling points within the maximum path delay are the effective channel, and the other sampling points are the noise. Set the noise sampling points to zero, and then use Fourier transform to convert the effective channel to the original frequency domain to achieve noise suppression.

[0010] Step 5: Build a deep neural network, define the loss function, use the estimated channel data after discrete Fourier processing as input, train the model, and save the optimal parameters.

[0011] Preferably, the signal received by the receiving end is expressed as Y=HX+Z, wherein H is the channel response, X is the transmitted signal, H is the channel, and Z is Gaussian noise.

[0012] Preferably, in step 2, the receiving end estimates the channel of the pilot signal using a least squares method.

[0013] Preferably, in step 3, an interpolation method is used to estimate the channel response at position n.

[0014] Preferably, the input of the deep neural network is the channel estimate after noise reduction processing

[0015] Preferably, when performing the discrete Fourier denoising process, the sampling point τ corresponding to the maximum path delay is known in advance m .

[0016] Preferably, the loss function L is defined as Where N is the number of training samples, Estimating channels for deep neural networks.

[0017] Beneficial effects:

[0018] The present invention proposes a method for improving the accuracy of wireless channel estimation in machine learning through noise reduction processing. Traditional deep channel estimation uses the least squares channel estimation as input, ignoring the impact of noise in the data set on the channel estimation results. This solution proposes to improve the accuracy of channel estimation of the learning model in low signal-to-noise ratio areas by suppressing the impact of noise in the training set data. Under the condition of a signal-to-noise ratio of 20dB, the normalized mean square error of the channel estimated by the deep neural network model without discrete Fourier processing is 0.012, while the estimation accuracy of the model after optimization processing by this solution reaches 0.008, which is a 33% improvement in performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The channel response estimated by the least squares method is transformed into the time domain through inverse Fourier transform, and the maximum channel delay τ m The channel responses after the corresponding sampling points are all set to 0, and then the remaining channel responses are transformed into the frequency domain through Fourier transform to achieve the effect of noise suppression.

[0020] Figure 2 In a typical urban environment, considering the multipath effect, such as Figure 2,This scheme shows the power delay spectrum of a typical urban environment. There are twelve signal propagation paths, and each path will experience different path losses and produce corresponding delays.

[0021] Figure 3 Constructing a deep neural network structure such as Figure 3 , the real and imaginary parts of the channel complex number after discrete Fourier denoising are used as the input of the model, and the model is iteratively trained and saved. When the model performs channel estimation, the output result is the real and imaginary parts of the channel data, which are integrated to obtain the complete channel response.

[0022] Figure 4 Under the conditions of a signal-to-noise ratio of 20dB and a pilot density of 1 / 4 (i.e., the ratio of the number of carriers used for pilots to the number of carriers carrying data), the channel response estimated by this scheme best fits the actual channel, verifying that this scheme can effectively improve the accuracy of the deep neural network channel estimation model.

[0023] Figure 5 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0024] The present invention will be further described below in conjunction with the accompanying drawings:

[0025] In the OFDM system, the received signal after multipath fading can be expressed as:

[0026]

[0027] Where N D is the number of taps of the filter, h d is the channel gain of the corresponding path, τ d is the signal delay of the dth path, and t is the sampling time from the first path. Usually, channel estimation is performed at the receiving end. Based on the received signal containing pilot information, the least squares method is used to estimate the channel, and the cost function is defined as

[0028]

[0029] Where E{·} is the expected value. From the cost function, we can see that the least squares channel estimation scheme and the signal-to-noise ratio This indicates that the least squares channel estimation will amplify the impact of noise. In order to solve the impact of noise enhancement caused by the least squares method, the channel data is preprocessed before training the model. The discrete Fourier transform technology is used to suppress noise and ensure the adaptability of the channel estimation model in a high noise environment.

[0030] In order to study the beneficial effects of the proposed method, an OFDM system with a subcarrier number N of 64 and a subcarrier spacing of 15KHz is simulated as an example.

[0031] Step 1: Convert the received baseband signal into frequency domain and extract the known signal Y.

[0032] Step 2: Based on the pilot signal, use the least squares method to estimate the channel response of the pilot position:

[0033]

[0034] where x p is the transmitted signal used for channel estimation, y p is the corresponding received signal of the pilot position allocated in the OFDM frame.

[0035] Step 3: To estimate the channel impulse response at all locations The channel response at the nth position is calculated by interpolation:

[0036]

[0037] Wherein l, m (l<m) are the two pilot positions closest to position n.

[0038] Step 4: Suppress the noise part in the data through discrete Fourier transform, and interpolate the channel gain of the j-th subcarrier Perform an inverse discrete Fourier transform.

[0039]

[0040] Where z[p] represents the time domain noise. Assume that the sampling point τ corresponding to the maximum path delay m It is known in advance that multipath components greater than this threshold can be considered as noise paths and set to zero. Through this operation, the channel response becomes:

[0041]

[0042] Then the time domain response is transformed into Convert to frequency domain. The denoised channel response is obtained as follows:

[0043]

[0044] Step 5: Multiple channel data The real and imaginary parts are separated and used as the input of the deep neural network model for training. Build a neural network model and define the loss function as the mean square error between the estimated channel and the true channel:

[0045]

[0046] N train is the number of training samples. The weight w and weight b of each iteration neuron are updated according to the following formula:

[0047]

[0048] η is the set learning rate.

Claims

1. A method for improving the accuracy of wireless channel estimation in machine learning by noise reduction processing, characterized in that: The following steps are involved: Step 1: Convert the received baseband signal into the frequency domain and extract the known signal Y from it; Step 2: Based on the pilot signal, the receiving end estimates the channel response at the pilot position; Step 3: Estimate the channel responses at all positions based on the channel responses at the pilot positions; Step 4: Divide the time domain channel into two parts, the effective channel and the noise, according to the maximum path delay. The sampling points within the maximum path delay are the effective channel, and the other sampling points are the noise. Set the noise sampling points to zero, and then use Fourier transform to convert the effective channel to the original frequency domain to achieve noise suppression. Step 5: Build a deep neural network, define the loss function, use the estimated channel data after discrete Fourier processing as input, train the model, and save the optimal parameters.

2. According to claim 1, a method for improving the accuracy of wireless channel estimation in machine learning by noise reduction processing is characterized in that: The signal received by the receiving end is expressed as Y=HX+Z, where H is the channel response, X is the transmitted signal, H is the channel, and Z is Gaussian noise.

3. According to claim 1, a method for improving the accuracy of wireless channel estimation in machine learning by noise reduction processing is characterized in that: In step 2, the receiving end estimates the channel of the pilot signal using the least square method.

4. According to claim 1, a method for improving the accuracy of wireless channel estimation in machine learning by noise reduction processing is characterized in that: In step 3, an interpolation method is used to estimate the channel response at position n.

5. According to claim 1, a method for improving the accuracy of wireless channel estimation in machine learning by noise reduction processing is characterized in that: The input of the deep neural network is the channel estimate after noise reduction 6. The method of improving the accuracy of wireless channel estimation in machine learning by noise reduction processing according to claim 1, characterized in that: When performing the discrete Fourier denoising process, the sampling point τ corresponding to the maximum path delay is known in advance m .

7. The method of improving the accuracy of wireless channel estimation in machine learning by noise reduction processing according to claim 1, characterized in that: The loss function L is defined as Where N is the number of training samples, Estimating channels for deep neural networks.

Images