Channel estimation method for MIMO scenarios based on fully connected deep learning

Abstract

The present invention provides a MIMO scenario channel estimation method based on fully connected deep learning. The method comprises the steps of: constructing a data set; constructing a deep learning model based on fully connected deep learning; repeatedly training the deep learning model, and recovering an equivalent channel matrix through prediction data, comparing the matrix with its corresponding channel state information label, and evaluating the output of the deep learning model through minimum mean square error; obtaining an optimized deep learning model for obtaining channel state information through verification data. The present invention focuses on learning OTFS pilot information through a fully connected network to complete the channel estimation task, significantly improving the accuracy of channel estimation and reducing the overhead of pilot resources while maintaining high performance.

Description

Technical Field

[0001] The present invention relates to the technical field of OTFS modulation communication channel estimation, and in particular to a MIMO scenario channel estimation method based on fully connected deep learning. Background Art

[0002] With the rapid development of wireless communication technology, satellite communication scenarios have become an indispensable part of the future 6G communication architecture. Traditional OFDM (Orthogonal Frequency Division Multiplexing) modulation can achieve higher communication rates through subcarrier multiplexing in ground-side communications. However, when facing satellite scenarios, especially low-orbit satellites (a MIMO (Multiple-Input Multiple-Output) scenario), due to the high Doppler frequency deviation, the orthogonality of the subcarriers in OFDM modulation will be lost, resulting in reduced performance in terms of computational efficiency and accuracy.

[0003] OTFS (Orthogonal Time Frequency Space Modulation), as an OFDM evolution modulation technology, can transform the time-frequency domain time-varying channel into a sparse non-time-varying channel in the delay-Doppler domain through ISFFT (Inverse Symplectic Finite Fourier Transform) and SFFT (Inverse Symplectic Finite Fourier Transform) transformations to reduce the complexity of subsequent processing. OTFS modulation has been proven to provide communication rates and bit error rate performance that exceed OFDM modulation in high-mobility scenarios.

[0004] In the field of wireless communications, efficient and accurate channel estimation has become the key to improving communication quality and data transmission rate. However, the traditional OFDM channel estimation method is not suitable for OTFS channel estimation. Therefore, some scholars have designed a pilot design scheme based on the protection interval according to the modulation characteristics of OTFS, and used a threshold-based method to complete the channel estimation of OTFS. Under high signal-to-noise ratio conditions, this method can complete the channel estimation task very well. As the signal-to-noise ratio decreases, the performance of this method will drop sharply. Generally, the pilot resource overhead is increased to reduce the occurrence of this situation. As communication technology has higher requirements for communication performance and resource overhead, the limitations of traditional methods are becoming increasingly prominent.

[0005] In the field of artificial intelligence, the fully connected deep learning network, as a classic network model, has shown great potential in the fields of image and data prediction. However, applying it to the field of channel estimation and using its nonlinear characteristics to learn pilot information for channel estimation is a new research method. The fully connected deep learning network has nonlinear characteristics and can theoretically provide better noise resistance. There are also some related studies using other deep learning networks for signal estimation, such as patent application CN117424782A, which uses convolutional neural networks to perform noise reduction in the protection interval area and then performs channel estimation on the noise-reduced data. It mainly focuses on using convolutional networks for noise reduction, and focuses more on the noise reduction process. After obtaining the noise-reduced data, additional channel estimation results are required, and the computational complexity is relatively high. Summary of the invention

[0006] Due to the above-mentioned defects in the prior art, the present invention provides a MIMO scenario channel estimation method and model based on fully connected deep learning, which focuses on learning OTFS pilot information through a fully connected network to complete the channel estimation task, while maintaining high performance, significantly improving the accuracy of channel estimation and reducing the overhead of pilot resources.

[0007] To achieve the above object, on the one hand, the present invention provides a MIMO scenario channel estimation method based on fully connected deep learning, characterized in that it includes the following steps:

[0008] S1. Construct a data set and divide it into a training set and a test set: Based on the standard protocol provided by 3GPP, simulate the OTFS modulated communication data, pre-process the complete OTFS data frame to obtain the guard interval area data containing valid pilot information and its corresponding channel state information label, and map these data into the training set and test set of the deep neural network after processing;

[0009] S2. Build a deep learning model based on fully connected deep learning;

[0010] S3, inputting the guard interval region data containing valid pilot information in the training set into the deep learning model based on fully connected deep learning, training the feature extraction function of the deep learning model so that the deep learning model can predict the channel state information from the input data, and at the same time, making the model train the training set with low signal-to-noise ratio, so that the model has good noise resistance performance;

[0011] S4. After obtaining the prediction output of the deep learning model, the equivalent channel matrix is ​​restored through the prediction data, the matrix is ​​compared with its corresponding channel state information label, and the output of the deep learning model is evaluated through the minimum mean square error (MSE);

[0012] S5. After repeating steps S3-S4 for a set number of times, an optimized deep learning model based on fully connected deep learning is obtained;

[0013] S6. After inputting the verification data into the trained model, the channel state information is obtained.

[0014] Furthermore, the S1 is specifically:

[0015] Assume that during the communication process, the OTFS transmitter first maps the delay-Doppler symbol x[k,l] to the time-frequency domain signal X[n,m] through ISFFT transformation as follows:

[0016]

[0017] where n = 0, …, N-1, m = 0, …, M-1;

[0018] The time-frequency domain signal X[n,m] is sampled into a time domain signal s(t) through the Heisenberg transform so that the transmitter can send:

[0019]

[0020] where g tx (t) is the transmission waveform, T is the symbol duration; Δf is the subcarrier spacing;

[0021] The time domain signal s(t) is characterized by a time delay and Doppler Time-varying channel , the received signal r(t) is expressed as:

[0022]

[0023] The channel It can be expressed as:

[0024]

[0025] Where P is the number of propagation paths, , and are the gain, delay and Doppler frequency deviation of the i-th propagation path respectively;

[0026] in and They can be expressed as:

[0027]

[0028] in , , represent the delay tap and Doppler tap respectively; and They are the lengths exceeding one subcarrier interval and symbol interval respectively.

[0029] After the receiver receives the signal, it calculates the cross-mixing function through matched filtering, which is expressed as follows:

[0030]

[0031]

[0032] in for( t'-t ) receives the conjugate of the pulse-shaped signal at the moment, for t' Receive signals at all times, is the frequency;

[0033] Finally, the time-frequency domain signal received by the receiver is converted into a delay-Doppler domain signal through SFFT, which is expressed as follows:

[0034]

[0035] After obtaining the received delay-Doppler domain signal, the guard interval region data is collected as the input data of the training set and the test set, and the channel state information during the simulation process is collected as the label data of the training set and the test set.

[0036] Furthermore, in S2, the deep learning model based on fully connected deep learning includes an input layer, a hidden layer and an output layer; the training data is input into the model through the input layer of the deep learning model, and is output by the model output layer after being calculated through the hidden layer. The output result will be put into the minimum mean square error loss device for comparison with the label data, and then the comparison result will be back-propagated to achieve the purpose of training the network.

[0037] Furthermore, in S3, the complete OTFS data frame is not used as the training set of the deep learning model, but the guard interval area with more effective information is selected as the training set of the deep learning model through data feature enhancement; for the high signal-to-noise ratio channel estimation task, the task is modeled as: Y = HX; for the low signal-to-noise ratio channel estimation task, the task is modeled as Y = HX+N, where H is the equivalent channel matrix, X is the transmitted signal matrix, and N is additive white Gaussian noise.

[0038] Furthermore, in S4, the output of the deep learning model and the label data are input into a loss function for comparison, and the loss function is defined as follows:

[0039]

[0040] in, Refers to the equivalent channel matrix prediction output of a fully connected deep learning network, refers to the label equivalent channel matrix.

[0041] Furthermore, in S5, the deep learning model performs back propagation through the loss function calculation value, and continuously updates the deep learning model parameters, and the update process is specifically as follows:

[0042]

[0043] in is the parameter of the nth fully connected deep learning model, is the training learning rate, is the first-order derivative of the fully connected deep learning model parameters, is the output of the loss function;

[0044] Repeatedly update the deep learning model parameters until all data in the training set are input;

[0045] Repeat the above process until the maximum number of iterations is reached.

[0046] On the other hand, the present invention provides a MIMO scenario channel estimation model based on fully connected deep learning, characterized in that the channel state information can be obtained by inputting the received signal pilot information into the MIMO scenario channel estimation model using steps S1 to S4 of the above-mentioned MIMO scenario channel estimation method.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] (1) The method of the present invention can more effectively obtain channel state information from pilot information by using the guard interval region data as a training set; it not only improves the rate of network training and reduces the complexity, but also provides richer and more accurate input data for the deep learning network, thereby enhancing the accuracy of channel estimation;

[0049] (2) The method of the present invention targets low signal-to-noise ratio samples. During the training process, the deep learning model not only predicts the channel state information, but also learns how to resist the influence of noise, making it more comprehensive in simulating and understanding the channel estimation behavior, making the method more generalizable and robust.

[0050] (3) The present invention effectively improves the performance of channel estimation, reduces the channel estimation time, reduces the overall computational complexity, and reduces the overhead of pilot resources through a method based on fully connected deep learning, making it applicable in power-constrained satellite scenarios, making deep learning technology more practical and efficient in channel estimation in the field of satellite communications. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The present invention and its features and advantages will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings, which are not drawn entirely to scale, emphasis being placed on illustrating the subject matter of the present invention.

[0052] Figure 1 It is a flow chart of a MIMO scenario channel estimation method based on fully connected deep learning in an embodiment of the present invention;

[0053] Figure 2 is a flowchart of simulating communication data that complies with the 3GPP standard and dividing a training set and a test set in an embodiment of the present invention;

[0054] Figure 3 Schematic diagram of the design scheme of embedded pilot placement in MIMO scenario in an embodiment of the present invention;

[0055] Figure 4 is a schematic diagram illustrating a deep learning model based on fully connected deep learning in an embodiment of the present invention;

[0056] Figure 5 It is the result analysis of the normalized minimum mean square error between the channel state information predicted on the test set and the actual channel state information under different signal-to-noise ratios in the embodiment of the present invention;

[0057] Figure 6 It is an analysis of the bit error rate results under different signal-to-noise ratios when data detection is performed based on the predicted channel state information in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] To further help understand the technical solution of the present invention, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. It should be understood that all of these exemplary embodiments described are only partial embodiments and examples of the present invention, rather than all of them. On the contrary, these exemplary embodiments are provided so that those skilled in the art can more thoroughly understand the present invention and can more completely convey the technical content of the present invention to those skilled in the art.

[0059] In the following detailed description, many specific details are set forth to provide a more thorough understanding of the present invention. However, it is obvious to those skilled in the art that well-known algorithms and models are not shown in detail to avoid obscuring the subject matter of the present invention. Example

[0060] This embodiment provides a MIMO scenario channel estimation method based on fully connected deep learning, such as Figure 1 As shown, the following steps are included:

[0061] S1. Build a data set and divide it into a training set and a test set. The specific flow chart is as follows Figure 2 As shown;

[0062] In this embodiment, firstly, multiple M*N two-dimensional matrices are generated to simulate the OTFS data frame in the delay-Doppler domain. This data frame represents the data symbols that need to be sent by different transmitting antennas in the actual communication system. Next, pilots and their guard intervals are inserted in different areas of this two-dimensional matrix. The OTFS pilot design in the MIMO scenario is as follows: Figure 3 After completing pilot placement, guard interval setting, and data symbol mapping, the matrix is ​​subjected to ISFFT transformation and Heisenberg transformation to convert the delay-Doppler domain signal into the time-frequency domain signal and the time domain signal in turn;

[0063] After constructing the transmit antenna data, the channel is modeled according to the NTN-TDL scenario in the 3GPP standard. The channel parameters are initialized by setting the number of transmit antennas, the number of receive antennas, the number of channel paths, the channel path gain, the channel path delay, and the channel path Doppler to construct the channel model. The transmit antenna transmit signal is input into the channel model to further obtain the receive antenna receive signal.

[0064] After the receiving antenna obtains the output result of the channel model, it first converts the time domain signal into the time-frequency domain signal and the delay-Doppler domain signal in sequence through the Wigner transformation and SFFT transformation. Based on the obtained delay-Doppler domain signal, the two-dimensional matrix data of the guard interval area on all received signals is stored to construct the training set and test set of the deep learning model;

[0065] Furthermore, step S1 is specifically as follows:

[0066] The OTFS data frame x[k,l] in the MIMO scenario is constructed as follows:

[0067]

[0068] in k pi , l pi They are i The placement position of the pilot signal, k v , are the maximum Doppler frequency deviation and the maximum delay respectively. and They are pilot signals and Data resources of delay-Doppler domain resource blocks.

[0069] Assume that during the communication process, the OTFS transmitter first maps the delay-Doppler symbol x[k,l] to the time-frequency domain signal X[n,m] through ISFFT transformation as follows:

[0070]

[0071] Where n = 0, …, N-1, m = 0, …, M-1.

[0072] The time-frequency domain signal X[n,m] is sampled into a time domain signal s(t) through the Heisenberg transform so that the transmitter can send:

[0073]

[0074] where g tx (t) is the transmission waveform, T is the symbol duration, and Δf is the subcarrier spacing.

[0075] The time domain signal s(t) is characterized by a time delay and Doppler Time-varying channel , the received signal r(t) is expressed as:

[0076]

[0077] The channel It can be expressed as:

[0078]

[0079] Where P is the number of propagation paths, , and are the gain, delay and Doppler frequency deviation of the i-th propagation path respectively;

[0080] in and They can be expressed as:

[0081]

[0082] in , , represent the delay tap and Doppler tap respectively; and They are the lengths exceeding one subcarrier interval and symbol interval respectively.

[0083] After the receiver receives the signal, it calculates the cross-mixing function through matched filtering, which is expressed as follows:

[0084]

[0085]

[0086] in for( t'-t ) receives the conjugate of the pulse-shaped signal at the moment, for t' Receive signals at all times, is the frequency.

[0087] Finally, the time-frequency domain signal received by the receiver is converted into a delay-Doppler domain signal through SFFT, which is expressed as follows:

[0088]

[0089] After receiving the delay Doppler domain signal, collect the guard interval area data and convert the size of (g m , g n ) is reshaped and straightened to a size of (g m *g n ,1), and then split the complex number into real and imaginary parts and concatenate them into a one-dimensional complex number set of size (2*g m *g n , 1)’s one-dimensional matrix is ​​used as the input data of the training set and the test set, and the channel state information in the simulation process is collected as the label data of the training set and the test set.

[0090] S2. Build a MIMO scenario deep learning model based on fully connected deep learning;

[0091] The overall framework of the deep learning model described in this embodiment is as follows Figure 4 As shown in the figure, it consists of an input layer, a hidden layer, and an output layer.

[0092] The input layer refers to the fully connected linear layer, through which the input data is input into the subsequent network layers after changing the dimension and the sensor leaf. In the initial layer, the real and imaginary parts of the received signal are directly combined and input together, and all data are processed at the same time. The subsequent hidden layers are similar to the input layer, but each hidden layer is connected to an activation function. Through the activation function ReLU function, the linear result is nonlinearized, and the channel features learned by the network are nonlinearly transformed. The output layer converts the output of the hidden layer into channel state information to obtain the prediction result of channel estimation.

[0093] S3. The specific process of training the network is as follows:

[0094] The signal in the guard interval area is selected instead of the complete OTFS data frame as the training set of the deep learning model. Through data feature enhancement, the guard interval area with more effective information is selected as the training set of the deep learning model.

[0095] The influence of noise can be ignored in the high signal-to-noise ratio channel estimation task, and the task can be modeled as:

[0096] Y = HX,

[0097] Where H is the equivalent channel matrix and X is the transmit signal matrix. The high signal-to-noise ratio training set allows the model to focus on channel estimation of H through Y and X. However, when the signal-to-noise ratio is low, the task model is:

[0098] Y = HX+N, where N is additive white Gaussian noise,

[0099] Noise will have a great impact on channel estimation. Therefore, by training with low signal-to-noise ratio data, the deep learning model can better resist the influence of noise.

[0100] S4. Predict the channel estimation results and optimize the deep learning model parameters through the MSE loss:

[0101] The output of the deep learning model is compared with the label data input into the loss function, which is defined as follows:

[0102]

[0103] in, Refers to the equivalent channel matrix prediction output of a fully connected deep learning network, refers to the label equivalent channel matrix.

[0104] MSE measures the error between the predicted output of the deep learning model and the expected label data. If the network does not output the correct result, its value will be large. Therefore, this measurement indicator can be used to constrain the deep learning network to update the deep learning model parameters.

[0105] Furthermore, S5 is specifically:

[0106] The deep learning model performs back propagation through the loss function calculation value and continuously updates the deep learning model parameters. The specific update process is as follows:

[0107]

[0108] in is the parameter of the nth fully connected deep learning model, is the training learning rate, is the first-order derivative of the fully connected deep learning model parameters, is the output of the loss function.

[0109] By observing the channel prediction results output by the deep learning network on the training data set and comparing the prediction results with the actual channel, we can see the training status of the deep learning network;

[0110] Repeatedly update the deep learning model parameters until all data in the training set are input;

[0111] Repeat the above process until the maximum number of iterations is reached.

[0112] Furthermore, S6 specifically includes: inputting the input data in the test data set into the trained deep learning model to obtain the prediction result, and comparing the prediction result with the real channel to obtain the normalized mean square error between the two, continuously adjusting the channel parameters, and comparing with the channel estimation results of other algorithms. Figure 5 It can be seen that the deep learning model provided in this embodiment (corresponding to the proposed method in the figure) can output the result (normalized MSE) about 3dB higher than the best performance of other algorithms when using fewer pilot resources compared to other channel estimation algorithms. At the same time, the deep learning model can still achieve good channel estimation results under low signal-to-noise ratio conditions. This result shows that the present invention has better performance and noise resistance.

[0113] Based on the predicted output of the deep learning model, the received signal is data-checked and the results are compared with the ideal channel and other algorithm outputs. Figure 6 It can be seen that the results of the deep learning model (corresponding to the proposed method in the figure) are more consistent with the ideal channel results and can bring a higher transmission rate to the communication system.

[0114] The above describes the preferred embodiments of the present invention. It should be understood that the present invention is not limited to the above-mentioned specific embodiments, and the devices and structures that are not described in detail should be understood to be implemented in a common manner in the art; any technician familiar with the art can use the above-disclosed methods and technical contents to make many possible changes and modifications to the technical solutions of the present invention without departing from the scope of the technical solutions of the present invention, or modify them into equivalent embodiments of equivalent changes, which does not affect the essential content of the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention are still within the scope of protection of the technical solutions of the present invention.

Claims

1. A MIMO scenario channel estimation method based on fully connected deep learning, characterized by: The following steps are involved: S1. Construct a data set and divide it into a training set and a test set: Based on the standard protocol provided by 3GPP, simulate the OTFS modulated communication data, pre-process the complete OTFS data frame to obtain the guard interval area data containing valid pilot information and its corresponding channel state information label, and map these data into the training set and test set of the deep neural network after processing; The S1 comprises the following sub-steps: First, multiple M*N two-dimensional matrices are generated to simulate the OTFS data frame in the delay-Doppler domain. This data frame represents the data symbols that need to be sent by different transmitting antennas in the actual communication system. Next, pilots and their guard intervals will be placed in different regions of this two-dimensional matrix; After completing pilot placement, guard interval setting, and data symbol mapping, the matrix is ​​subjected to ISFFT transformation and Heisenberg transformation to convert the delay-Doppler domain signal into the time-frequency domain signal and the time domain signal in turn; After constructing the transmit antenna data, the channel is modeled according to the NTN-TDL scenario in the 3GPP standard. The channel parameters are initialized by setting the number of transmit antennas, the number of receive antennas, the number of channel paths, the channel path gain, the channel path delay, and the channel path Doppler to construct the channel model. The signal transmitted by the transmitting antenna is input into the channel model to further obtain the received signal of the receiving antenna; After the receiving antenna obtains the output result of the channel model, it first converts the time domain signal into the time-frequency domain signal and the delay-Doppler domain signal in sequence through the Wigner transformation and SFFT transformation; Based on the obtained delay-Doppler domain signals, the two-dimensional matrix data of the guard interval area on all received signals are stored to construct the training set and test set of the deep learning model; S2. Construct a deep learning model based on fully connected deep learning; the deep learning model based on fully connected deep learning includes an input layer, a hidden layer and an output layer; the training data is input into the model through the input layer of the deep learning model, and is output by the model output layer after being calculated through the hidden layer, and the output result is put into the minimum mean square error loss device for comparison with the label data, and then the comparison result is back-propagated to achieve the purpose of training the network; S3, inputting the guard interval region data containing valid pilot information in the training set into the deep learning model based on fully connected deep learning, training the feature extraction function of the deep learning model so that the deep learning model can predict the channel state information from the input data, and at the same time, making the model train the training set with low signal-to-noise ratio, so that the model has good noise resistance performance; In S3, the complete OTFS data frame is not used as the training set of the deep learning model, but the guard interval area with more effective information is selected as the training set of the deep learning model through data feature enhancement; for the high signal-to-noise ratio channel estimation task, the task is modeled as: Y = HX; for the low signal-to-noise ratio channel estimation task, the task is modeled as Y = HX+N, where H is the equivalent channel matrix, X is the transmitted signal matrix, and N is the additive Gaussian white noise; S4. After obtaining the predicted output of the deep learning model, the equivalent channel matrix is ​​restored through the predicted data, the matrix is ​​compared with its corresponding channel state information label, and the output of the deep learning model is evaluated through the minimum mean square error loss function; S5. After repeating steps S3-S4 for a set number of times, an optimized deep learning model based on fully connected deep learning is obtained; In S5, the deep learning model performs back propagation through the loss function calculation value, and continuously updates the deep learning model parameters. The specific update process is as follows: in is the parameter of the nth fully connected deep learning model, is the training learning rate, is the first-order derivative of the fully connected deep learning model parameters, is the output of the loss function; Repeatedly update the deep learning model parameters until all data in the training set are input; Repeat the above process until the maximum number of iterations is reached; S6. After inputting the verification data into the trained model, the channel state information is obtained.

2. A MIMO scenario channel estimation method according to claim 1, characterized in that: The S1 is specifically: Assume that during the communication process, the OTFS transmitter first maps the delay-Doppler symbol x[k,l] to the time-frequency domain signal X[n,m] through ISFFT transformation as follows: where n = 0, …, N-1, m = 0, …, M-1; The time-frequency domain signal X[n,m] is sampled into a time domain signal s(t) through the Heisenberg transform so that the transmitter can send: where g tx (t) is the transmission waveform, T is the symbol duration; Δf is the subcarrier spacing; The time domain signal s(t) is characterized by a time delay and Doppler Time-varying channel , the received signal r(t) is expressed as: The channel It can be expressed as: Where P is the number of propagation paths, , and are the gain, delay and Doppler frequency deviation of the i-th propagation path respectively; in and They can be expressed as: in , , represent the delay tap and Doppler tap respectively; and They are the lengths exceeding one subcarrier interval and symbol interval respectively; After the receiver receives the signal, it calculates the cross-mixing function through matched filtering, which is expressed as follows: in for( t'-t ) receives the conjugate of the pulse-shaped signal at the moment, for t' Receive signals at all times, is the frequency; Finally, the time-frequency domain signal received by the receiver is converted into a delay-Doppler domain signal through SFFT, which is expressed as follows: After obtaining the received delay-Doppler domain signal, the guard interval region data is collected as the input data of the training set and the test set, and the channel state information during the simulation process is collected as the label data of the training set and the test set.

3. A MIMO scenario channel estimation method according to claim 1, characterized in that: In S4, the output of the deep learning model is compared with the label data input into the loss function, and the loss function is defined as follows: in, Refers to the equivalent channel matrix prediction output of a fully connected deep learning network, refers to the label equivalent channel matrix.

Images