OFDM (Orthogonal Frequency Division Multiplexing) system pilot frequency design and channel estimation method and system

Abstract

The invention provides a pilot frequency design and channel estimation method for an OFDM (Orthogonal Frequency Division Multiplexing) system, which is used for improving the performance of the system in a dynamic channel scene and specifically comprises a pilot frequency design module and an up-sampling module, firstly, in a pilot frequency design module, a neural network responsible for pilot frequency design selects corresponding elements from LS estimation of a complete channel according to maximum Doppler frequency shift and time delay spread which reflect time and frequency change speed of the channel as available information; designing a pilot frequency format most suitable for the current channel environment and obtaining an LS estimation result of a designed pilot frequency position; then, the LS estimation result of the pilot frequency position is sent to an up-sampling module formed by full connection for interpolation, and an estimation value of a complete channel is obtained. Compared with the existing channel estimation algorithm, the method provided by the invention can adapt to different channel scenes, the time required for designing the pilot frequency is short, the method has good robustness for a dynamic scene, and the channel estimation performance is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communications, and in particular relates to an OFDM system pilot design and channel estimation method. Background Art

[0002] The OFDM system can effectively counteract frequency selective fading, thereby ensuring communication accuracy in complex and dynamically changing communication scenarios. Channel estimation is a key component of the OFDM system, and its accuracy will profoundly affect the performance of the entire system. The traditional LS estimation algorithm is widely used in actual communication systems, but its performance is limited because it ignores the impact of noise.

[0003] In recent years, deep learning has achieved great success in fields such as computer vision and natural language processing, and a large number of research works have combined deep learning with wireless communications, such as resource allocation, channel state information feedback, signal detection, etc. At the same time, there are also a lot of works that use neural networks in deep learning, such as DNN and CNN, for channel estimation, achieving excellent performance. These channel estimation methods often use a fixed pilot format. However, the use of a fixed pilot format in different channel environments cannot effectively track channel changes and is difficult to adapt to increasingly complex and dynamically changing channel scenarios. Therefore, how to use the current deep learning technology to design an OFDM system channel estimation method that can adapt to dynamic channel environments, is highly robust and has excellent performance is a problem that needs to be solved urgently. Summary of the invention

[0004] Purpose of the invention: The present invention provides an OFDM system pilot design and channel estimation method, aiming to utilize the powerful learning ability of neural networks to design appropriate pilot formats for different channel environments, and to solve the problems of poor robustness and limited channel estimation performance of existing channel estimation methods in dynamic scenarios.

[0005] Technical solution: The technical solution adopted by the present invention specifically includes the following steps:

[0006] Step 1: Consider having N f subcarriers, N t OFDM system with 100 symbols. Assume that the transmitted signal It is known that LS channel estimation is performed on the entire channel to obtain the estimation matrix of the complete channel The matrix Convert to vector form Where d = N f ×N t , represents the total number of elements in the entire OFDM system; a size of And each row has only one element value 1, and the other elements are 0. PD With vector Multiply them, and then select the location of the k pilots that can most effectively track the current channel environment from the positions of the d elements of the entire channel matrix, and obtain the LS estimate at that position

[0007] Step 2: The maximum Doppler shift D1 and delay spread D2, which reflect the speed of channel changes in time and frequency, are used as available information and input into the Concrete DNN network composed of fully connected layers, and the vector is output. This vector will be used to construct the pilot design matrix M PD ;

[0008] Step 3: A PD Perform shape change and get the matrix And perform discretization operation on the matrix to convert it into a matrix The number of rows and columns of this matrix are the number of pilots k and the total number of OFDM elements d, respectively, and each row is close to a one-hot vector, with only one element value of 1 and the values ​​of the rest of the elements are 0;

[0009] Step 4: Transform the matrix M PD With vector Multiply by to obtain the LS channel estimate of the designed pilot position Then it is sent to the upsampling module composed of upsampling DNN for interpolation to obtain the complete channel estimation result. Then the whole process of pilot design and channel estimation is completed;

[0010] Step 5: Use a data set containing channels with different maximum Doppler shifts and delay spreads to jointly train the Concrete DNN and the upsampling DNN, so that it can eventually design the most appropriate pilot format according to different channel environments and complete the corresponding channel estimation process.

[0011] As a further technical solution of the present invention, in step 1, during the training of the network, assuming that the entire transmission channel matrix X is known, the LS channel estimation process of the complete channel can be expressed as:

[0012]

[0013] Among them, y, h, are the vector forms of Y, H, and X respectively, ||·||2 is the l2 distance, (·) ⊙-1 is the inverse Hadamard product.

[0014] As a further technical solution of the present invention, in step 2, the output vector A of Concrete DNN PD The expression is:

[0015] A PD =f PD (D1,D2,Θ PD ) (2)

[0016] Among them, f PD (·) is the function of Concrete DNN, Θ PD are all the trainable parameters contained in the structure. Since vector A PD is the output of the neural network. Its internal values ​​are all floating point numbers, so it is difficult to transform it into the required pilot design matrix M only by changing its shape. PD .

[0017] As a further technical solution of the present invention, in step 3, the matrix α is calculated by the following formula: PD Perform discretization operation and then convert it into matrix M PD :

[0018]

[0019] in, is the matrix M PD The element in row l and column i, g i is a vector sampled from a Gumbel distribution The i-th element of is the matrix α PD The element at row l and column i, is the matrix α PD The element in the lth row and the jth column, T represents temperature, which is used to control the degree of discretization.

[0020] As a further technical solution of the present invention, in step 4, the pilot frequency is designed The process is expressed as:

[0021]

[0022] LS channel estimation for designed pilot positions The interpolation process is expressed as:

[0023]

[0024] Among them, f UP (·) is the function of upsampling DNN, Θ UP are all the trainable parameters contained in the structure, and are the operations of taking the real part and the imaginary part respectively.

[0025] As a further technical solution of the present invention, in step 5, the loss function of training the entire network structure is as follows:

[0026]

[0027] Among them, H i is the i-th channel sample in the training data set, is the estimate of the ith channel sample obtained by the network after pilot design and channel estimation, and N is the total number of samples in the training set.

[0028] Beneficial effects: Compared with the prior art, the present invention adopts the above technical solution and has the following beneficial effects:

[0029] The OFDM system pilot design and channel estimation method designed by the present invention inputs the maximum Doppler shift and delay spread that reflect the speed of channel changes in time and frequency into the neural network as available information, and assists the network to design a pilot format that best meets the current channel environment, thereby improving the network's tracking ability for fading channels and increasing the system's channel estimation performance. The pilot design method of the present invention can dynamically design the most suitable pilot structure according to the current channel environment. Compared with the traditional, fixed pilot format, it can adapt to different channel scenarios, and the time required to design the pilot is short, and it has good robustness to dynamic scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A schematic diagram of the principle of the pilot design and channel estimation method of the OFDM system designed for the present invention;

[0031] Figure 2 It is a schematic diagram of the specific structure of the pilot design network in the present invention. DETAILED DESCRIPTION

[0032] The embodiments of the present invention will be described below with reference to the accompanying drawings.

[0033] like Figure 1As shown, the present invention designs a pilot design and channel estimation method for an OFDM system, specifically, it includes two modules: pilot design and upsampling; the input of the pilot design module includes the LS estimation of the complete channel, the maximum Doppler frequency shift and delay spread of the channel, and it selects corresponding elements from the LS estimation of the complete channel according to the maximum Doppler frequency shift and delay spread, designs the pilot format that best suits the current channel environment, and outputs the LS estimation result of the designed pilot position; the input of the upsampling module is the LS estimation result of the designed pilot position, and it interpolates the input and outputs the estimated value of the complete channel. This method specifically includes the following steps:

[0034] Step 1: Consider having N f subcarriers, N t OFDM system with 100 symbols. Assume that the transmitted signal It is known that LS channel estimation is performed on the entire channel to obtain the estimation matrix of the complete channel The matrix Convert to vector form Where d = N f ×N t , represents the total number of elements in the entire OFDM system; by And each row has only one element value 1, and the other elements are 0. PD With vector Multiply them, and then select the location of the k pilots that can most effectively track the current channel environment from the positions of the d elements of the entire channel matrix, and obtain the LS estimate at that position

[0035] Specifically, at the receiver, after removing the cyclic prefix and performing a fast Fourier transform, the frequency domain representation of the received signal is It can be expressed as:

[0036] Y=H⊙X+W (1)

[0037] Among them, ⊙ is the Hadamard product, H, They are the channel matrices in frequency domain and have a mean of 0 and a variance of Additive white Gaussian noise.

[0038] Because in the process of training the network, it is assumed that the entire transmission channel matrix X is known, the process of LS channel estimation of the complete channel can be expressed as:

[0039]

[0040] Among them, y, h, are the vector forms of Y, H, and X respectively, ||·||2 is the l2 distance, (·)⊙-1 is the inverse Hadamard product.

[0041] Step 2: The maximum Doppler shift D1 and delay spread D2, which reflect the speed of channel changes in time and frequency, are used as available information and input into the Concrete DNN network composed of fully connected layers, and the vector is output. This vector will be used to construct the pilot design matrix M PD ;

[0042] Specifically, the maximum Doppler shift reflects how fast the channel changes in time, while the delay spread reflects how fast the channel changes in frequency. By inputting these parameters into the network, the network can learn to determine the number and position of pilots in time and frequency. The output vector A of the Concrete DNN is PD The expression is:

[0043] A PD =f PD (D1,D2,Θ PD ) (3)

[0044] Among them, f PD (·) is the function of Concrete DNN, Θ PD are all the trainable parameters contained in the structure. Since vector A PD is the output of the neural network. Its internal values ​​are all floating point numbers, so it is difficult to transform it into the required pilot design matrix M only by changing its shape. PD .

[0045] Step 3: A PD Perform shape change and get the matrix And perform discretization operation on the matrix to convert it into a matrix The number of rows and columns of this matrix are the number of pilots k and the total number of OFDM elements d, respectively, and each row is close to a one-hot vector, with only one element value of 1 and the values ​​of the rest of the elements are 0;

[0046] Specifically, the matrix α is given by the following formula PD Perform discretization operation and then convert it into matrix M PD :

[0047]

[0048] in, is the matrix M PD The element in row l and column i, g i is a vector sampled from a Gumbel distribution The i-th element of is the matrix α PDThe element at row l and column i, is the matrix α PD The element in the lth row and jth column; T represents the temperature, which is used to control the degree of discretization. It is set to a large value in the initial stage of Concrete DNN network training and gradually decreases and tends to 0 as the network is trained. When T tends to 0, the matrix M calculated by formula (4) is PD Each row of will approach the form of a one-hot vector.

[0049] Step 4: Transform the matrix M PD With vector Multiply by to obtain the LS channel estimate of the designed pilot position Then it is sent to the upsampling module composed of upsampling DNN for interpolation to obtain the complete channel estimation result. Then the whole process of pilot design and channel estimation is completed;

[0050] Specifically, design the pilot The process is expressed as:

[0051]

[0052] Among them, the matrix M PD In the initial stage of training, the elements of each row are relatively random, so that the network can learn as many channel-related features as possible; as the training progresses, each row will approach a one-hot vector. The matrix M PD With size Vector When multiplied, it is equivalent to selecting k pilot positions that can best track channel changes from d channel elements, thereby completing the design of the pilot format and obtaining the LS channel estimate of the designed pilot position.

[0053] LS channel estimation for designed pilot positions The interpolation process is expressed as:

[0054]

[0055] Among them, f UP (·) is the function of upsampling DNN, Θ UP are all the trainable parameters contained in the structure, and are the operations of taking the real part and the imaginary part respectively.

[0056] Step 5: Use a data set containing channels with different maximum Doppler shifts and delay spreads to jointly train the Concrete DNN and the upsampling DNN, so that it can eventually design the most appropriate pilot format according to different channel environments and complete the corresponding channel estimation process.

[0057] Specifically, the Concrete DNN and upsampling DNN in the network structure can be one or more layers of full connection. The channel model adopts the TDL-A channel model in 3GPP. The SNR of the channel is randomly selected from the set [0,5,...,25]dB, the maximum Doppler shift is randomly selected from the three sets of [0,20,...,100]Hz, [150,200,...,550]Hz, and [600,700,...,1000]Hz, and the delay spread is randomly selected from the three sets of [0,20,...,100]ns, [150,200,...,550]ns, and [600,800,...,1200]ns. The dataset contains 102400 independent channels, 80% of which are used for training and 20% for validation. The entire network is trained using the Adam optimizer for 1000 epochs, with the batch size set to 512, the initial learning rate being 0.001, and decaying by a factor of 10 every 500 epochs.

[0058] The loss function for training the entire network structure is as follows:

[0059]

[0060] Among them, H i is the i-th channel sample in the training data set, is the estimate of the ith channel sample obtained by the network after pilot design and channel estimation, and N is the total number of samples in the training set. Since the network is trained using the above mixed data set, the network can learn the channel characteristics of different channel environments and design the pilot format that can best track the changes in the current channel environment based on the maximum Doppler frequency shift and delay expansion of the channel.

[0061] like Figure 2 As shown, first, the LS estimate of the complete channel matrix Converted into vector form At the same time, the maximum Doppler shift and delay spread that reflect the speed of the channel change in time and frequency are sent to ConcreteDNN to generate vector A PD , and after shape change and discretization operations, it is transformed into a matrix M in which each row is in one-hot form PD ; Finally, the matrix M PD With vector Multiply to get the designed pilot frequency The pilot designed under the current channel environment can accurately track the changes in the channel, thereby improving the channel estimation performance of the overall system.

[0062] The method of the present invention applies a deep learning network to an OFDM system, and uses the maximum Doppler shift and delay spread that reflect the speed of channel changes in time and frequency as available information to assist the network in designing a pilot format that best suits the current channel environment, thereby improving the channel estimation performance of the entire system. Compared with traditional pilot formats, the method proposed by the present invention can adapt to different channel scenarios and design the optimal pilot format for different channel scenarios, and has the advantages of strong robustness and good performance.

[0063] What is described above in conjunction with the accompanying drawings is only one of the preferred examples of the present invention and cannot be used to limit the scope of rights included in the present invention. It should be pointed out that any equivalent modifications made in accordance with the principles of the present invention are covered by the protection scope of the claims of the present invention.

Claims

1. A pilot design and channel estimation method for an OFDM system, characterized in that: The OFDM system has N f subcarriers, N t symbol, the method comprising the following steps: Step 1: Assume that the signal is sent Known, and perform LS channel estimation on the entire channel to obtain the estimation matrix of the complete channel The matrix Convert to vector form Where d = N f ×N t , represents the total number of elements in the entire OFDM system; a size of And each row has only one element value 1, and the other elements are 0. PD With vector Multiply them, and then select the location of the k pilots that can most effectively track the current channel environment from the positions of the d elements of the entire channel matrix, and obtain the LS estimate at that position Step 2: The maximum Doppler shift D1 and delay spread D2, which reflect the speed of channel changes in time and frequency, are used as available information and input into the Concrete DNN network composed of fully connected layers, and the vector is output. This vector will be used to construct the pilot design matrix M PD ; Step 3: A PD Perform shape change and get the matrix And perform discretization operation on the matrix to convert it into a matrix The number of rows and columns of this matrix are the number of pilots k and the total number of OFDM elements d, respectively, and each row is close to a one-hot vector, with only one element value of 1 and the values ​​of the rest of the elements are 0; Step 4: Transform the matrix M PD With vector Multiply by to obtain the LS channel estimate of the designed pilot position Then it is sent to the upsampling module composed of upsampling DNN for interpolation to obtain the complete channel estimation result. Then the whole process of pilot design and channel estimation is completed; Step 5: Use a data set containing channels with different maximum Doppler shifts and delay spreads to jointly train ConcreteDNN and upsampling DNN, so that it can eventually design the most appropriate pilot format according to different channel environments and complete the corresponding channel estimation process.

2. The OFDM system pilot design and channel estimation method according to claim 1, characterized in that: In step 2, the output vector A of Concrete DNN PD The expression is: A PD =f P ( D (D1,D2,Θ PD )(1) Among them, f PD () is the function of Concrete DNN, Θ PD are all the trainable parameters contained in this structure.

3. The OFDM system pilot design and channel estimation method according to claim 1, characterized in that: In step 3, the matrix α is calculated by the following formula: PD Perform discretization operation and then convert it into matrix M PD : in, is the matrix M PD The element in row l and column i, g i is a vector sampled from a Gumbel distribution The i-th element of is the matrix α PD The element at row l and column i, is the matrix α PD The element in the lth row and the jth column, T represents temperature, which is used to control the degree of discretization.

4. The OFDM system pilot design and channel estimation method according to claim 3, characterized in that: In step 3, the temperature T is set to a large value at the initial stage of Concrete DNN network training, and gradually decreases and approaches 0 as the network is trained; when T approaches 0, the matrix M calculated by formula (2) is PD Each row of will approach the form of a one-hot vector.

5. The OFDM system pilot design and channel estimation method according to claim 1, characterized in that: In step 4, the pilot frequency is designed The process is expressed as: Among them, the matrix M PD In the initial stage of training, the elements of each row are relatively random, so that the network can learn as many channel-related features as possible; as the training progresses, each row will approach a one-hot vector.

6. The OFDM system pilot design and channel estimation method according to claim 1, characterized in that: In step 4, the LS channel estimation of the designed pilot position The upsampling process is expressed as: Among them, f UP () is the function of upsampling DNN, Θ UP are all the trainable parameters contained in the structure, and are the operations of taking the real part and the imaginary part respectively.

7. The OFDM system pilot design and channel estimation method according to claim 1, characterized in that: In step 5, the loss function for training the entire network structure is as follows: Among them, H i is the i-th channel sample in the training data set, is the estimate of the ith channel sample obtained by the network after pilot design and channel estimation, and N is the total number of samples in the training set.

8. An OFDM system pilot design and channel estimation system, characterized in that: include: Pilot design module and up-sampling module; The pilot design module selects corresponding elements from the LS estimation of the complete channel according to the maximum Doppler shift and delay spread of the channel, designs the pilot format that best suits the current channel environment, and obtains the LS estimation result of the designed pilot position; The upsampling module is responsible for interpolating the LS estimates of the designed pilot positions to obtain an estimate of the complete channel.

Images