OFDM channel estimation method based on shallow neural network
By using a channel estimation method based on shallow neural networks and employing DCT technology for channel compression and interpolation, the high computational complexity of traditional methods and the large computational resource consumption of deep learning methods are solved, thus achieving efficient channel estimation in OFDM systems.
Patent Information
- Application Number
- CN202511414440.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-29
- Publication Date
- 2026-01-23
AI Technical Summary
In OFDM wireless communication, traditional channel estimation methods require complex mathematical modeling to ensure accuracy, while deep learning methods require a lot of computational resources, making it difficult to find a balance between computational complexity and estimation accuracy.
A channel estimation method based on shallow neural networks is adopted. Image processing is used to remove redundant channel information, DCT technology is used for channel compression, and shallow neural networks are combined for channel interpolation and restoration to reduce computational complexity while maintaining high estimation accuracy.
While reducing computational complexity, it achieves high accuracy in channel estimation, combining the advantages of low complexity of traditional methods and high estimation accuracy of deep learning methods, and is suitable for OFDM systems.
Smart Images
Figure CN121396705A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of wireless communication, in particular to an OFDM channel estimation method based on a shallow neural network. BACKGROUND
[0002] In the wireless communication scenario based on orthogonal frequency division multiplexing (OFDM) technology, the wireless channel as the carrier of the wireless signal, the bridge between the transmitting end and the receiving end, plays a crucial role. Due to the existence of fading phenomenon, the receiving end needs to estimate the model parameters in the assumed channel model, i.e. channel estimation, in order to correctly recover the transmitting end signal from the signal, so as to equalize the channel influence.
[0003] In the traditional channel estimation scheme, the channel estimation method based on pilot signals uses pilot signals to estimate the channel, which has the advantages of stability and reliability, but needs to model the channel complexly to ensure the estimation accuracy. On the other hand, the channel estimation method combined with deep learning avoids complex mathematical modeling through data-oriented algorithms, but generally needs to consume a large amount of computing resources for network training and reasoning. SUMMARY
[0004] The present application studies the channel estimation method under the OFDM system, and proposes an OFDM channel estimation method based on a shallow neural network, which uses image processing method to eliminate channel redundant information and reduce the amount of neural network input data, so as to play the advantages of deep learning-based channel estimation method in estimation accuracy in the case of reducing the computational complexity, and solve the problems in the above background technology.
[0005] In order to achieve the above purpose, the present application adopts the following technical scheme:
[0006] An OFDM channel estimation method based on a shallow neural network, comprising the following steps:
[0007] Step 1, preprocessing: obtaining the most basic channel least squares (LS) estimation, using DCT technology to transform the estimation value to the image feature domain, and compressing in the dimension;
[0008] Step 2, interpolation: inputting the compressed version of the channel estimation output by step 1 into a shallow neural network, using the neural network to complete the function of channel estimation interpolation, and outputting the corresponding estimation value;
[0009] Step 3, restoration: performing inverse DCT transformation on the channel estimation value output by step 2, transforming to the signal frequency domain, and obtaining the full-dimension channel estimation value.
[0010] Preferably, the signal is preprocessed in step 1, and the specific steps are as follows:
[0011] (1-1) First, perform least square (LS) channel estimation on the received signal at the pilot position as follows:
[0012]
[0013] where Y denotes the received frequency domain signal, S is the pilot signal, n e [n0, n0+ ΔN c , n0+ 2ΔN c ,..., N c -1] denotes the subcarrier index of the pilot signal within the OFDM symbol, n0 is the starting subcarrier of the pilot, ΔN c is the interval of the pilot in the subcarrier, N c is the total number of subcarriers within a single OFDM symbol; k e [k0, k0+ ΔN s , k0+ 2ΔN s ,..., N s -1] denotes the index of the OFDM symbol where the pilot is located, k0 is the starting index of the OFDM symbol where the pilot is located, ΔN s is the interval of the OFDM symbol where the pilot is located, N s denotes the total number of OFDM symbols.
[0014] (1-2) Resize the channel estimation value obtained in (1-1) to retain only the valid estimation value as follows:
[0015]
[0016] where
[0017] n = n0+ iΔN c , k = k0+ jΔn s ,
[0018]
[0019] (1-3) Perform a discrete cosine transform (DCT) on the H p estimated in (1-2) to obtain its representation in the image feature domain H p,dec as follows:
[0020]
[0021] where p and q satisfy
[0022]
[0023] a p and a q are defined as follows:
[0024]
[0025] in
[0026] (1-4) Take the H obtained in (1-3) p,dec To perform compression, specifically, define a compression function. as follows:
[0027] V in =f in,cp (H p,dec )
[0028] Then the compression function f in,cp The operation is as follows: scan from H using zig-zag. p,dec Place the corresponding values in the top left corner of the complex vector V. in In the middle, V in The length is L in .
[0029] Preferably, step 2 outputs V from (1-4). in The input is interpolated into a shallow neural network to obtain a complex vector V. out Its length is L out The neural network was trained using simulated data, and the specific steps are as follows:
[0030] (2-1) Generating effective full-dimensional channel data through computer simulation Add additive noise at different levels, and calculate its LS estimate at the pilot location using the steps described in (1-1).
[0031] (2-2) For the result obtained in (2-1) Execute (1-2), (1-3), and (1-4) sequentially to obtain the corresponding V. in ;
[0032] (2-3) For H obtained in (2-1), directly perform a full-dimensional DCT transformation to obtain its representation H in the image feature domain. dec ;
[0033] (2-4) The H obtained in (2-3) dec Perform compression operations and define compression functions. as follows:
[0034] V target =f out,cp (H dec )
[0035] Then the compression function f out,cp The operation is as follows: scan from H using zig-zag. decthe upper left corner of V target , where V target is a complex vector of length L out ;
[0036] (2-5) define the parameters of the neural network as Θ, based on Θ, the output V in of (2-2) as the input of the neural network, the output of the neural network can be represented as:
[0037] V out = NN(V in , Θ)
[0038] The training process of the neural network is to find the optimal parameters so that the loss function L(V out , V target ) is minimized:
[0039]
[0040] where the loss function L is defined as:
[0041]
[0042] (2-6) input the output V in of (1-4) into the neural network with parameters , and get the output V out .
[0043] Preferably, the step 3 restores the output V out of the neural network to full-dimensional channel estimation values, and the specific steps are as follows:
[0044] (3-1) define the decompression function fill the values of V out into the matrix for L out <N c × N s , after filling the values of V out , the remaining positions of can be filled with zeros;
[0045] (3-2) for the obtained in (3-2), the full-dimensional channel estimation values in the frequency domain can be obtained by inverse DCT as follows:
[0046]
[0047] a p and a q are defined as follows:
[0048]
[0049]
[0050] wherein P=N c Q=N s .
[0051] The beneficial effects of the present application are: the channel estimation method based on deep neural network proposed in the present application has more excellent channel estimation performance compared with the traditional channel estimation method by using data-oriented optimization algorithm to learn channel characteristics; since in the OFDM system, the channel is a complex matrix with two dimensions of frequency domain and time domain, the channel estimation method based on deep neural network often needs to use a complex neural network structure for channel estimation; the present application considers the image features of the channel, uses the DCT-based image compression algorithm to compress and reduce the dimension of the channel, extracts the key information of the channel for subsequent neural network processing, greatly reduces the data amount and computational complexity of the neural network to be processed under the condition of ensuring the final channel estimation accuracy, thereby simplifying the neural network structure, and combining the low complexity of the traditional channel estimation method and the high estimation accuracy of the channel estimation method based on deep neural network. BRIEF DESCRIPTION OF DRAWINGS
[0052] Figure 1 is the overall flowchart of the present application.
[0053] Figure 2 is the distribution diagram of two-dimensional region energy before and after DCT transformation of the present application.
[0054] Figure 3 is the graph of the change of channel estimation MSE of different methods with SNR under the EVA channel model of the present application.
[0055] Figure 4 is the graph of the change of channel estimation MSE of different methods with SNR under the ETU channel model of the present application. DETAILED DESCRIPTION
[0056] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all embodiments.
[0057] Referring to Figures 1-4 , the present application provides an OFDM channel estimation method based on a shallow neural network, comprising the following steps:
[0058] Step 1, preprocessing: get the basic channel least square (LS) estimation, transform the estimation to image feature domain using DCT technology, and compress in dimension;
[0059] Step 2, interpolation: input the compressed version of channel estimation output by step 1 into a shallow neural network, use the neural network to complete the function of channel estimation interpolation, and output the corresponding estimation value;
[0060] Step 3, restoration: perform inverse DCT on the channel estimation value output by step 2, transform to signal frequency domain, and get the full-dimensional channel estimation value.
[0061] The whole process is divided into 3 steps, a total of 12 processes:
[0062] In step 1, the signal is preprocessed, and the specific operation is as follows:
[0063] (1-1) first, the least square (LS) channel estimation of the signal at the received pilot position is as follows:
[0064]
[0065] Where Y represents the received frequency domain signal, and S is the pilot signal. Let the pilot start subcarrier be n0=0, and the interval ΔN c =32, the total number of subcarriers in a single OFDM symbol N c =128, then the subcarrier index where the pilot is located is: n∈[0,32,64,96]; let the pilot start symbol index k0=k0=0, and the interval ΔN s =4, the total number of OFDM symbols N s =14, then the OFDM symbol index where the pilot is located is k∈[0,4,8,12].
[0066] (1-2) resize the channel estimation value obtained in (1-1) to only keep the valid estimation value as follows:
[0067]
[0068] Where
[0069] n=32*i=[0,32,64,96], k=4*j=[0,4,8,12],
[0070] i=[0,1,2,3], j=[0,1,2,3]
[0071] (1-3) perform discrete cosine transform (DCT) on the H p estimated in (1-2) to get its representation Hp,dec As follows:
[0072]
[0073] where p and q satisfy
[0074] p = [0, 1, 2, 3], q = [0, 1, 2, 3]
[0075] a p and a p are defined as follows:
[0076]
[0077]
[0078] (1-4) Compress the H p,dec obtained in (1-3), specifically, define the compression function As follows:
[0079] V in = f in,cp (H p,dec )
[0080] The operation of the compression function f in,cp is: through zig-zag scanning, the corresponding value from the top left corner of H p,dec is sequentially placed into the complex vector V in , where the length of V in is L in = 16.
[0081] The step 2 inputs the V in output by (1-4) into a shallow neural network for interpolation to obtain a complex vector V out , whose length is L out . The specific architecture of the neural network is shown in Table 1:
[0082] Table 1, shallow neural network structure
[0083] Layer Type Activation Function Input Layer None Fully Connected Layer Relu Fully Connected Layer Relu Fully Connected Layer None Output Layer None
[0084] The training of the neural network is carried out through simulation data, and the specific steps are as follows:
[0085] (2-1) Generate effective full-dimensional channel data through computer simulation Add different levels of additive noise, and calculate the LS estimation value at the pilot position through the steps described in (1-1)
[0086] (2-2) For the H Execute (1-2), (1-3), and (1-4) sequentially to obtain the corresponding V. in .
[0087] (2-3) For H obtained in (2-1), directly perform a full-dimensional DCT transformation to obtain its representation H in the image feature domain. dec .
[0088] (2-4) The H obtained in (2-3) dec Perform compression operations and define compression functions. as follows:
[0089] V target =f out,cp (H dec )
[0090] Then the compression function f out,cp The operation is as follows: scan from H using zig-zag. dec Place the corresponding values in the top left corner of the complex vector V. target In the middle, V target The length is L out =64.
[0091] (2-5) Define the parameters of the neural network as Θ. Based on Θ, (2-2) the output V in As the input to a neural network, the output of the neural network can be expressed as:
[0092] V out =NN(V in ,Θ)
[0093] The training process of a neural network is to find the optimal parameters. Make the loss function L(V) out V target ) number is the smallest:
[0094]
[0095] The loss function L is defined as follows:
[0096]
[0097] (2-6) The V output from (1-4) in Input parameters are In the neural network, the output V is obtained. out .
[0098] Step 3 will output V of the neural network out The specific steps to restore the channel estimate to full dimensions are as follows:
[0099] (3-1) Definition of the decompression function The values of V out are filled into the matrix by zig-zag manner Since 64<128x14, after the filling of V out is completed, the remaining positions of V can be filled with zeros.
[0100] (3-2) For the V obtained in (3-2), the full-dimension channel estimation value in the frequency domain can be obtained by inverse DCT as follows:
[0101]
[0102] a p and a q are defined as follows:
[0103]
[0104] Figure 3 and Figure 4 respectively show the channel estimation MSE of different methods under the EVA and ETU channel models with the change of SNR. It can be seen that the DCT-SNN method proposed in the present application is extremely close to the performance of the Ideal MMSE method under the condition of low signal-to-noise ratio, which shows that the DCT-SNN method has good noise suppression, and this characteristic is derived from the construction of the data set in the training process of the interpolation network. Under the condition of high signal-to-noise ratio, the performance of the DCT-SNN method tends to be close to the LS method and the MMSE method. At this time, the bottleneck of the channel estimation accuracy is the prediction ability of the interpolation network, that is, the degree of learning of the characteristics of the channel itself, and we can see that the performance of the DCT-SNN method is slightly better than that of the LS method and the MMSE method, which shows that the interpolation network effectively learns the characteristics of the channel.
[0105] The above description is only the preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto, any skilled person in the art can make equivalent replacement or change according to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.
Claims
1. A method for OFDM channel estimation based on shallow neural networks, characterized in that, Includes the following steps: Step 1, Preprocessing: Obtain the most basic channel least squares LS estimate, use DCT technology to transform the estimate into the image feature domain, and perform dimensional compression; Step 2, Interpolation: Input the compressed version of the channel estimate output from Step 1 into a shallow neural network, and use the neural network to perform the channel estimate interpolation function and output the corresponding estimated value. Step 3, Restoration: Perform an inverse DCT transform on the channel estimate output in Step 2 to transform it into the signal frequency domain, and obtain the full-dimensional channel estimate.
2. The OFDM channel estimation method based on shallow neural networks according to claim 1, characterized in that, In step 1, the signal undergoes preprocessing, the specific steps of which are as follows: (1-1) The least-squares LS channel estimation of the signal at the received pilot position is as follows: Where: Y represents the received frequency domain signal, S is the pilot signal, and n∈[n0,n0+ΔN] c ,n0+2ΔN c ,…,N c -1] represents the subcarrier index of the pilot signal within the OFDM symbol, n0 is the pilot start subcarrier, and ΔN c N represents the spacing of the pilots on the subcarriers. c The total number of subcarriers within a single OFDM symbol; k∈[k0,k0+ΔN] s ,k0+2ΔN s ,…,N s -1] represents the index of the OFDM symbol where the pilot is located, k0 is the starting index of the OFDM symbol where the pilot is located, and ΔN s N represents the interval of the OFDM symbol containing the pilot signal. s This indicates the total number of OFDM symbols.
3. The OFDM channel estimation method based on shallow neural networks according to claim 2, characterized in that, The steps further include: (1-2) readjusting the channel estimates obtained in (1-1) and retaining only the valid estimates as follows: in: n=n0+iΔN c ,k=k0+jΔN s , 4. The OFDM channel estimation method based on shallow neural networks according to claim 3, characterized in that, The steps further include: (1-3) adjusting the H estimated in (1-2) p Perform Discrete Cosine Transform (DCT) to obtain its representation H in the image feature domain. p,dec as follows: Where p and q satisfy a p and a q The definition is as follows: in 5. The OFDM channel estimation method based on shallow neural networks according to claim 4, characterized in that, The steps further include: (1-4) taking the H obtained in (1-3) p,dec Perform compression and define a compression function. as follows: V in =f in,cp (H p,dec ) Compression function f in,cp The operation is as follows: scan from H using zig-zag. p,dec Place the corresponding values in the top left corner of the complex vector V. in In the middle, V in The length is L in .
6. The OFDM channel estimation method based on shallow neural networks according to claim 1, characterized in that, Step 2 will output V from (1-4) in The input is interpolated into a shallow neural network to obtain a complex vector V. out Its length is L out The neural network was trained using simulated data, and the specific steps are as follows: (2-1) Generating effective full-dimensional channel data through computer simulation Add additive noise at different levels and calculate its LS estimate at the pilot location using the steps described in (1-1). (2-2) For the result obtained in (2-1) Execute (1-2), (1-3), and (1-4) sequentially to obtain the corresponding V. in ; (2-3) For H obtained in (2-1), directly perform a full-dimensional DCT transformation to obtain its representation H in the image feature domain. dec ; (2-4) The H obtained in (2-3) dec Perform compression operations and define compression functions. as follows: V target =f out,cp (H dec ) Compression function f out,cp The operation is as follows: scan from H using zig-zag. dec Place the corresponding values in the top left corner of the complex vector V. target In the middle, V target The length is L out ; (2-5) Define the parameters of the neural network as Θ. Based on Θ, (2-2) the output V in As the input to a neural network, the output of the neural network can be represented as: V out =NN(V in ,Θ) The training process of a neural network is to find the optimal parameters. Make the loss function L(V) out V target ) number is the smallest: The loss function L is defined as follows: (2-6) The V output from (1-4) in Input parameters are In the neural network, the output V is obtained. out .
7. The OFDM channel estimation method based on shallow neural networks according to claim 1, characterized in that, Step 3 will output V of the neural network out The specific steps to restore the channel estimate to full dimensions are as follows: (3-1) Define the decompression function V out The values are filled into the matrix using the zig-zag method. For L out <N c ×N s In the case of V out After the value is filled in, Fill the remaining positions with zeros; (3-2) For the results obtained in (3-2) The full-dimensional channel estimate in the frequency domain obtained by inverse DCT transformation is as follows: a p and a q The definition is as follows: Where P = N c Q = N s .