Deep learning-based otfs joint channel estimation and data detection method and device

Abstract

The application discloses a satellite scene-based OTFS joint channel estimation and data detection method and model based on deep learning. The method comprises the following steps: constructing a data set, designing a pilot design scheme, constructing a joint channel estimation and data detection deep learning model, repeatedly training the deep learning model, and obtaining channel state information and data symbols through prediction data; comparing the network prediction result with corresponding channel state information and data symbol labels, evaluating the output of the deep learning model through the least mean square error and cross-entropy loss; and obtaining the optimized deep learning model to obtain the channel state information and data symbols through the verification data. The application focuses on learning the OTFS received signal through the deep learning network to simultaneously complete the channel estimation and data detection tasks, and the pilot design scheme and data slicing processing scheme can significantly improve the accuracy of channel estimation and data detection and reduce the overhead of pilot resources while maintaining high performance.

Description

Technical Field

[0001] This invention belongs to the field of OTFS modulated communication channel estimation technology, specifically relating to a deep learning-based OTFS joint channel estimation and data detection method and apparatus. Background Technology

[0002] Satellite communication is a crucial component of integrated air-space-ground networks, providing broad coverage and reliability under diverse geographical conditions. Traditional OFDM (Orthogonal Frequency Division Multiplexing) modulation achieves high communication rates in terrestrial communications through subcarrier multiplexing. However, in satellite scenarios, especially with low-Earth orbit satellites, significant Doppler frequency offset leads to a loss of orthogonality in OFDM subcarriers, resulting in reduced computational efficiency and accuracy.

[0003] Orthogonal Time-Frequency Space Modulation (OTFS), as an evolution of OFDM modulation, transforms time-varying channels in the time-frequency domain into time-delay-Doppler domain time-independent channels through Inverse Symplectic Finite Fourier Transform (ISFFT) and Symplectic Finite Fourier Transform (SFFT). This ensures that all transmitted data receives the same channel gain, thereby improving performance in high-speed mobile scenarios. OTFS modulation has been proven to provide communication rates and bit error rates exceeding those of OFDM modulation in high-mobility scenarios.

[0004] In the field of wireless communication, efficient and accurate channel estimation and data detection are crucial for improving communication quality and data transmission rates. However, traditional OFDM channel estimation and data detection methods are not applicable to Over-the-Air (OTFS). Therefore, some researchers have designed a pilot design scheme based on guard intervals based on the characteristics of OTFS, using a threshold-based method for channel estimation and a message passing algorithm for data detection. However, the introduction of guard intervals leads to a significant problem of low spectral efficiency in this pilot design scheme, making it unsuitable for power-constrained satellite scenarios. Meanwhile, some researchers have proposed a superimposed pilot design scheme, which superimposes pilot symbols and data symbols and eliminates the guard interval, improving spectral efficiency and reducing pilot resource overhead. However, this scheme, by eliminating the guard interval, increases the complexity of receiver channel estimation and signal demodulation algorithms and degrades performance. As communication technology demands higher performance and lower resource overhead, the limitations of traditional methods become increasingly apparent. Therefore, it is necessary to design novel pilot design schemes and channel estimation and data detection methods to improve performance indicators while maximizing spectral efficiency.

[0005] In the field of computer science, artificial intelligence, represented by deep learning networks, has demonstrated enormous potential in areas such as image processing and data prediction. However, applying it to channel estimation and data detection, utilizing its nonlinear characteristics to learn from received signals for channel estimation and data detection, represents a completely new research approach. Currently, there are some studies utilizing deep learning networks for OTFS communication, such as patent application CN117424782A, which uses convolutional neural networks for noise reduction in the guard interval region and then performs channel estimation on the denoised data. This approach primarily focuses on noise reduction using convolutional networks, concentrating more on the noise reduction process. Furthermore, after obtaining the denoised data, additional channel estimation algorithms are needed to obtain channel state information, resulting in high computational complexity and the inability to perform data detection tasks. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method and apparatus for joint channel estimation and data detection based on deep learning and OTFS, which solves the problems of low spectrum utilization and high computational complexity of channel estimation in the prior art.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] In satellite scenarios, the pilot design scheme adds a guard interval at the pilot symbol location, specifically expressed by the following formula:

[0009]

[0010] in, This represents the transmitted signal in the time-delay-Doppler domain. These represent Doppler and time delay indexes, respectively. These represent the pilot symbols and the symbols located at the Doppler index, respectively. The delay index is Data symbols, These represent the Doppler and time delay indices of the pilot symbols, respectively.

[0011] A deep learning-based joint channel estimation and data detection method for OTFS in satellite scenarios includes the following steps:

[0012] S1. Based on the above pilot design scheme, construct a dataset including a training set and a test set;

[0013] S2. Construct a deep learning model for joint channel estimation and data detection, including a coarse channel estimation network model, an iterative data detection network model, and a data-assisted channel estimation network model.

[0014] S3. Input the training data of different networks in the training set into the corresponding network models to train the feature extraction function of the deep learning model, so that the deep learning model can predict the channel state information and data symbol output from the input data.

[0015] S4. Recover the equivalent channel matrix from the output channel state information, compare the equivalent channel matrix with its corresponding channel state information label, and evaluate the output of the coarse channel estimation network model and the data-assisted channel estimation network model.

[0016] S5. Compare the output data symbols with the labels of their transmitted signals to evaluate the output of the iterative data detection network model;

[0017] S6. Repeat steps S3 to S5 a set number of times to obtain the optimized joint channel estimation and data detection deep learning model.

[0018] S7. After inputting the received signal acquired in real time by the receiving antenna into the trained model, channel state information and data detection results are obtained.

[0019] The dataset is constructed as follows:

[0020] First, a two-dimensional matrix of size M×N is used to simulate the OTFS data frame in the time-delay-Doppler domain, which represents the data symbols sent from the transmitting antenna to the receiving antenna in the actual communication system.

[0021] Next, data symbols and pilot symbols are inserted according to the pilot design scheme; the inserted matrix is ​​then subjected to ISFFT and Heisenberg transform in sequence to convert the time-delay-Doppler domain signal into time-frequency domain signal and time domain signal in sequence.

[0022] After constructing the transmit antenna data, set the channel parameters according to the NTN-TDL scenario in the 3GPP standard, initialize them, and complete the channel model construction;

[0023] The transmitted signal from the transmitting antenna is input into the channel model to further obtain the received signal at the receiving antenna affected by channel and environmental noise;

[0024] After the receiving antenna obtains the channel model output, it sequentially transforms the time-domain signal into a time-frequency domain signal and a time-delay-Doppler domain signal through Wegener transform and SFFT transform.

[0025] Based on the obtained time-delay-Doppler domain signal, the received signal data is stored to construct the training and test sets of the deep learning model.

[0026] The specific process of step S2 is as follows:

[0027] The coarse channel estimation network is constructed by using a pilot channel extraction module and a noise reduction sub-network.

[0028] The iterative data detection network is constructed by using an interference cancellation module, a channel feature extraction subnetwork, a slice processing module, and a data detection subnetwork.

[0029] The data-aided channel estimation network is constructed through a data preprocessing module and an iterative channel estimation subnetwork.

[0030] By connecting the coarse channel estimation network, the iterative data detection network, and the data-aided channel estimation network in series, a complete joint channel estimation and data detection network is obtained.

[0031] The specific method for training the deep learning model in step S3 is as follows:

[0032] First, the received signal is preprocessed by splitting the real and imaginary parts of the received signal and then concatenating them in the data channel dimension to form tensor data.

[0033] For the coarse channel estimation network, the preprocessed received signal is directly selected as the input training set data for the coarse channel estimation network;

[0034] For the iterative data detection network, the preprocessed received signal and channel state information label data are selected as the input training set data for the iterative data detection network.

[0035] For data-assisted channel estimation networks, preprocessed received signals and data symbol label data are selected as the input training set data for the data-assisted channel estimation network.

[0036] Simultaneously, training data from different networks are input into the corresponding network models to train the three different networks in parallel.

[0037] The minimum mean square error is used as the loss function to train the channel estimation network model, as shown in the following formula:

[0038]

[0039] in, This represents the loss value of the minimum mean square error loss function. The equivalent channel matrix prediction output of a fully connected deep learning network. This refers to the tag-equivalent channel matrix;

[0040] The cross-entropy loss is used as the loss function to train the data detection model, as shown in the following formula:

[0041]

[0042] in This represents the loss value of the cross-entropy loss function. This represents the total number of all data symbols. For the category of data symbols, For symbolic functions, Representative observation sample Category The predicted probability.

[0043] The calculated value of the loss function is backpropagated to update the parameters of the deep learning model. The update process is specifically represented by the following formula:

[0044]

[0045] in, For the first Parameters of the deep learning model To train the learning rate, The first derivative of the parameters of the deep learning model. It is the output of the loss function, including and .

[0046] The deep learning-based OTFS joint channel estimation and data detection device includes a coarse channel estimation network, an iterative data detection network, and a data-aided channel estimation network connected in series.

[0047] The coarse channel estimation network includes a pilot signal extraction module and a noise reduction sub-network;

[0048] The iterative data detection network includes an interference cancellation module, a channel feature extraction subnetwork, a feature fusion module, a data slicing module, and a data detection subnetwork;

[0049] The data-assisted channel estimation network includes a data preprocessing module and an iterative channel estimation subnetwork.

[0050] The noise reduction subnetwork, data detection subnetwork, and iterative channel estimation subnetwork are all composed of U-Net networks; the feature extraction subnetwork is composed of multi-layer convolutional units.

[0051] A computer-readable storage medium storing computer-readable instructions that, when executed by a processor, invoke all or part of the steps of the method.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] 1. This invention presents a novel pilot design scheme, which significantly improves pilot resource utilization compared to traditional embedded pilot designs. Compared to traditional superimposed pilot designs, this scheme significantly reduces interference between pilot and data symbols, substantially improving system performance. Furthermore, this pilot design scheme can be applied as a module to other channel estimation algorithms.

[0054] 2. The method described in this invention models the channel estimation task and the data detection task in OTFS as image denoising and pixel segmentation tasks respectively, and processes the two tasks together, so that the original channel estimation and data detection tasks complement each other, improving the performance index of the method and making the method more generalizable and robust.

[0055] 3. The pilot design scheme and data slicing operation methods proposed in this invention significantly improve system performance compared with traditional algorithms. At the same time, they are universal and can be widely applied to other channel estimation and data detection methods, helping to improve performance indicators. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of the pilot design scheme for joint channel estimation and data detection in an embodiment of the present invention.

[0057] Figure 2 This is a flowchart illustrating the framework of the deep learning network model for joint channel estimation and data detection in this embodiment of the invention.

[0058] Figure 3This is a flowchart of the data slicing operation in an embodiment of the present invention.

[0059] Figure 4 The simulation diagram of this invention compares and tests the normalized minimum mean square error performance of different pilot design schemes under different signal-to-noise ratios and different pilot-to-data power ratios in the channel estimation task.

[0060] Figure 5 The simulation diagram of this invention compares and tests the normalized minimum mean square error performance index of different channel estimation algorithms under different signal-to-noise ratios in the channel estimation task.

[0061] Figure 6 The simulation diagram of this invention compares and tests the bit error rate performance of different data detection algorithms under different signal-to-noise ratios in the data detection task.

[0062] Figure 7 The simulation diagram of this invention compares and tests the bit error rate performance of different deep learning networks in data detection tasks, whether or not they use data slicing operations. Detailed Implementation

[0063] To further aid in understanding the technical solutions of the present invention, exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. It should be understood that all described exemplary embodiments are merely some embodiments and examples of the present invention, and not all of them. Rather, these exemplary embodiments are provided so that those skilled in the art can more thoroughly understand the present disclosure and more completely convey its technical content to them.

[0064] In the following detailed description, numerous specific details are set forth to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that well-known algorithms and models are not shown in detail to avoid obscuring the spirit of the invention.

[0065] Specific embodiments, such as Figures 1 to 7 As shown,

[0066] Example 1,

[0067] The pilot design scheme provided in this example incorporates a guard interval at the pilot symbol location, such as... Figure 1 As shown, the mathematical representation of this pilot design scheme is as follows:

[0068]

[0069] in, This represents the transmitted signal in the time-delay-Doppler domain. These represent Doppler and time delay indexes, respectively. These represent the pilot symbols and the symbols located at the Doppler index, respectively. The delay index is Data symbols, These represent the Doppler and time delay indices of the pilot symbols, respectively.

[0070] The pilot design scheme provided in this embodiment can achieve better output results than other pilot design schemes while using fewer pilot resources. Furthermore, the proposed pilot design scheme also shows significant performance improvements when applied to other channel estimation algorithms. The pilot design scheme involved in this invention can help the OTFS channel estimation algorithm improve performance metrics while using fewer pilot resources.

[0071] Example 2,

[0072] This embodiment provides a method and apparatus for joint OTFS channel estimation and data detection based on deep learning in satellite scenarios, specifically including the following steps:

[0073] S1. According to the pilot design scheme described in Example 1, a dataset is constructed and divided into a training set and a test set: Based on the standard protocol provided by 3GPP, the simulated OTFS modulated communication data is processed to obtain the training set data and its corresponding validation set label data for model training. After processing, these data are mapped to the training set and test set of the deep neural network.

[0074] The specific process is as follows:

[0075] In this embodiment, firstly, multiple units of size M are generated. A two-dimensional matrix of N is used to simulate the OTFS data frame in the time-delay-Doppler domain, which represents the data symbols that the transmitting antenna needs to send in a real communication system.

[0076] Next, pilot symbols, data symbols, and guard intervals are inserted according to the designed pilot design scheme, and constellation diagram mapping is completed according to the modulation method. After the symbol insertion is completed, ISFFT and Heisenberg transforms are performed on the two-dimensional matrix in sequence to convert the time-delay-Doppler domain signal into time-frequency domain signal and time domain signal in sequence.

[0077] Configure the channel parameters according to the NTN-TDL scenario in the 3GPP standard, mainly including 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, and complete the construction of the channel model.

[0078] The time-domain transmitted signal is input into the channel model through the transmitting antenna, and then transmitted through the channel to obtain the time-domain received signal of the receiving antenna.

[0079] After the receiving antenna receives the channel model output, it transforms the time-domain signal into a time-frequency domain signal and a time-delay-Doppler domain signal through Wegener transform and SFFT transform. The receiving antenna data is collected, and the channel model parameters and the original transmitting antenna data are stored to construct the training and testing sets for the deep learning model.

[0080] Furthermore, the specific steps are as follows:

[0081] OTFS data frames in satellite scenarios (i.e., transmitted signals in the delay-Doppler domain) The structure is as follows:

[0082]

[0083] Assume that during the communication process, the OTFS transmitter first transmits the time-delay-Doppler symbol. Mapped to the time-frequency domain signal via ISFFT transform The process is as follows:

[0084]

[0085] in {0, 1, … , -1}, {0, 1, … , -1} represent time and frequency indices, respectively.

[0086] Time-frequency domain signal The sampled signal is obtained by Heisenberg transform and then converted to a time-domain signal. So that the transmitter can send:

[0087]

[0088] in For the transmitted pulse waveform, in The duration of an OTFS data frame. The subcarrier spacing.

[0089] Time domain signal By characteristic of time delay and Doppler Time-varying channel Receive signal Represented as:

[0090]

[0091] in This represents additive Gaussian noise, and based on this, the time-varying channel... It can be represented as:

[0092]

[0093] in, It is the number of transmission paths. , and They are the first Gain, delay, and Doppler frequency offset of each propagation path;

[0094] in, and They can be represented as:

[0095]

[0096] in , }, representing the first, second, third, fourth, fifth, sixth, seventh, eighth, ninth, tenth, twentieth Delay taps and Doppler taps for each channel path.

[0097] After receiving the signal, the receiver calculates the cross-mixing function using matched filtering. The processed time-domain received signal is expressed as follows:

[0098]

[0099]

[0100] in To receive pulse waveforms, and These represent time and frequency indices, respectively.

[0101] Finally, the time-frequency domain received signal is converted into a signal by the receiver. The signal is converted to a time-delay-Doppler domain signal via SFFT. The expression is as follows:

[0102]

[0103] After obtaining the time-delay-Doppler domain received signal, the original transmitted signal, channel parameters, and received signal are collected and stored to construct training and test datasets.

[0104] S2. Construct a deep learning model for joint channel estimation and data detection, including a coarse channel estimation network model, an iterative data detection network model, and a data-assisted channel estimation network model.

[0105] The deep learning model framework process described in this embodiment is as follows: Figure 2 As shown, it mainly includes a coarse channel estimation network, an iterative data detection network, and a data-assisted channel estimation network;

[0106] Furthermore, the specific details of the deep learning network model for the joint channel estimation and data detection network are as follows:

[0107] The coarse channel estimation network is constructed from a pilot signal extraction module and a noise reduction sub-network.

[0108] The iterative data detection network is constructed from an interference cancellation module, a channel feature extraction subnetwork, a feature fusion module, a data slicing module, and a data detection subnetwork.

[0109] The data-assisted channel estimation network is constructed from a data preprocessing module and an iterative channel estimation subnetwork;

[0110] By connecting the coarse channel estimation network, the iterative data detection network, and the data-aided channel estimation network in series, a complete joint channel estimation and data detection network is obtained.

[0111] Furthermore, the denoising subnetwork, data detection subnetwork, and iterative channel estimation subnetwork are all composed of a U-Net network. The U-Net network structure is as follows: Figure 2 As shown, a U-shaped structure is formed by a symmetrical encoder and decoder, and its specific structure is as follows:

[0112] The encoder consists of a series of Figure 2 The convolutional units shown consist of a convolutional layer, a batch normalization layer, and a rectified linear unit (ReLU). Following a series of convolutional units is a layer with a stride of 1 and a size of 2. A pooling operation of size 2 is used for downsampling. Additionally, with each downsampling operation, the feature channel dimension of the data is doubled to facilitate the network learning higher-dimensional features. In the decoder section, all operations mirror the encoder's, with a stride of 1 and a size of 2 following a series of convolutional endmembers. The transpose of the convolution operation is used for upsampling. At the same time, the feature channel dimension of the data will be halved.

[0113] Furthermore, the feature extraction subnetwork also consists of multiple layers of convolutional units, such as... Figure 2 As shown.

[0114] S3. Input the training data from different networks in the training set into the corresponding network models to train the feature extraction function of the deep learning models, enabling the deep learning models to predict channel state information and data symbol output from the input data; the specific process is as follows:

[0115] The input data to the coarse channel estimation network is the received signal. Receive signal First, the pilot signal is extracted using the pilot signal extraction module. Subsequently, the pilot signal is input into the denoising subnetwork to eliminate interference from data symbols and noise, resulting in a clean predicted pilot signal. .

[0116] The input data for the iterative data detection network is the first... Pilot signal obtained from the second prediction and received signals Receive signal First, the pilot signal is eliminated using the interference cancellation module. Interference; simultaneously, the channel feature extraction subnetwork extracts information from the pilot signal. After extracting the channel state information features, they are concatenated with the interference-canceled received signal along the channel dimension. This data is then input to the data slicing module, undergoing operations such as cyclic filling and data slicing before being fed into the data detection subnetwork. Finally, the data detection subnetwork predicts the... Secondary predicted transmission signal .

[0117] The input data for the data-assisted channel estimation network is the first... Secondary predicted transmission signal and received signals , and After data preprocessing and feature fusion, the data is input into the iterative channel estimation subnetwork to finally obtain the first... Pilot signal obtained from the second prediction .

[0118] Furthermore, the specific details of each module in the joint channel estimation and data detection network are as follows:

[0119] Pilot extraction module from size M N's received signal Extract the size as pilot signals, in which and Representing the maximum Doppler and time delay offsets of the channel path, respectively, the index of this region satisfies... .

[0120] The interference cancellation module will convert the pilot signal matrix From the received signal matrix The pure received signal is obtained by removing the impurities. To avoid interference between pilot symbols and data symbols during data detection tasks, the specific implementation is as follows:

[0121]

[0122] in .

[0123] The data slicing module mainly includes loop filling operations and data slicing operations, such as... Figure 3 As shown, the specific process is as follows:

[0124] The cyclic filling operation will receive signals. Cyclic padding is performed along the Doppler dimension to help the network better learn the features of data edges, as specifically represented below:

[0125]

[0126] in This represents the data after the cycle is filled. Representing the feature dimensions of the data, The Doppler index after cyclic filling. This is the size of the slice window.

[0127] The data slicing operation will cyclically fill the received signal. Divided into The data fragment, in which the first Data fragments It is expressed as follows:

[0128]

[0129] The data preprocessing module remaps the transmitted signal predicted by the network according to the modulation constellation diagram. Taking QAM modulation as an example, the specific process is as follows:

[0130]

[0131] in This represents the transmitted signal predicted by the network. This indicates the constellation type in QAM modulation.

[0132] S4. Recover the equivalent channel matrix from the output channel state information, compare the equivalent channel matrix with its corresponding channel state information label, and evaluate the outputs of the coarse channel estimation network model and the data-aided channel estimation network model; specifically,

[0133] After obtaining the prediction output of the channel estimation task, the prediction result is compared with its corresponding channel state information label, and the outputs of the coarse channel estimation network model and the data-aided channel estimation network model are evaluated using a loss function:

[0134] In channel estimation tasks, the minimum mean square error is used as the loss function, which is defined as follows:

[0135] in, The channel matrix obtained from network prediction. This refers to the tag channel matrix.

[0136] MSE measures the error between the predicted output of a deep learning model and the expected labeled data. If the network does not output the correct result, its value will be large. Therefore, this metric can be used to constrain the deep learning network to update the parameters of the deep learning model.

[0137] S5. Compare the output data symbols with the labels of their transmitted signals to evaluate the output of the iterative data detection network model; specifically,

[0138] After obtaining the predicted output of the data detection task, the detection results are compared with the labels of the transmitted signals, and the output of the iterative channel estimation network model is evaluated using a loss function.

[0139] In data inspection tasks, cross-entropy loss is used as the loss function, and its definition is as follows:

[0140]

[0141] Cross-entropy loss measures the error between the predicted output of a deep learning model and the expected label data in a classification task. The smaller the probability that the network predicts the correct data symbol, the larger its value. Therefore, this metric can be used to update the parameters of the iterative data detection network model.

[0142] S6. After repeating steps S3 to S5 a set number of times, the optimized joint channel estimation and data detection deep learning model is obtained; specifically,

[0143] Deep learning models use backpropagation to calculate the value of the loss function, continuously updating the parameters of the deep learning model. The update process is as follows:

[0144] ;

[0145] By observing the channel prediction results output by the deep learning network on the training dataset and comparing the prediction results with the real channel, we can see the training status of the deep learning network.

[0146] Repeatedly update the deep learning model parameters until all data in the training set has been input;

[0147] Repeat the above process until the maximum number of iterations is reached.

[0148] S7. After inputting the received signal acquired in real time by the receiving antenna into the trained model, the channel state information and data detection results are obtained. Furthermore, the received signal of the receiving antenna is input into the trained joint channel estimation and data detection network to obtain the channel state information and predicted data symbols, thus completing the channel estimation and data detection tasks.

[0149] To further verify the results of this method and model, simulation experiments were conducted and compared with existing technologies, as detailed below:

[0150] The data signal-to-noise ratio, pilot signal-to-noise ratio, and other variables are continuously adjusted and compared with the channel estimation results of other algorithms. From Figure 4 As can be seen, the pilot design scheme provided in this embodiment, compared with other pilot design schemes, still produces better results while using fewer pilot resources. Furthermore, the proposed pilot design scheme also shows significant performance improvements when applied to other channel estimation algorithms. This result demonstrates that the pilot design scheme involved in this invention can help the OTFS channel estimation algorithm improve performance while using fewer pilot resources.

[0151] Furthermore, from Figure 5 As can be seen, the joint channel estimation and data detection network provided in this embodiment maintains performance comparable to other channel estimation algorithms even when the pilot signal-to-noise ratio is 3dB lower. Under the same pilot signal-to-noise ratio, its performance is significantly better than other algorithms, and it requires fewer iterative estimations. This result demonstrates that the joint channel estimation and data detection network in this invention exhibits superior performance.

[0152] In data inspection tasks, from Figure 6 It can be seen that the algorithm provided in this embodiment has a significant performance improvement compared to other data detection algorithms under the same pilot signal-to-noise ratio. Furthermore, the algorithm provided in this embodiment has performance metrics that better reflect ideal conditions. This result demonstrates that the joint channel estimation and data detection network of the present invention exhibits better performance and robustness.

[0153] Furthermore, from Figure 7 As can be seen, different networks show significant performance improvements after using the proposed data slicing operation. This result indicates that the data slicing operation involved in this invention is also applicable to other algorithms, helping to improve their performance and demonstrating universality and generalization.

[0154] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0155] Therefore, a computer-readable storage medium is disclosed, wherein computer-readable instructions are stored on the computer-readable storage medium, and when the computer-readable instructions are executed by a processor, all or part of the steps of the method are invoked.

[0156] The preferred embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and the devices and structures not described in detail should be understood as being implemented in a conventional manner in the art. Any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention using the methods and techniques disclosed above, or modify them into equivalent embodiments with equivalent changes, without departing from the scope of the present invention. This 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 present invention's technical solutions still fall within the protection scope of the present invention.

Claims

1. A method for OTFS joint channel estimation and data detection based on deep learning in a satellite scenario, characterized in that: Comprise the following steps: S1, according to the pilot design scheme under the satellite scene, construct the data set including training set and test set; The pilot design scheme is to increase the guard interval at the pilot symbol position, which is specifically expressed by the following formula: ， wherein, represents a transmitted signal in the delay-Doppler domain, respectively represent Doppler and delay indices, respectively represent a pilot symbol and a data symbol located at Doppler index , and delay index , respectively represent Doppler and delay indices of a pilot symbol. S2, construct a joint channel estimation and data detection deep learning model, including coarse channel estimation network model, iterative data detection network model and data aided channel estimation network model; The specific process is as follows: Through pilot channel extraction module and noise reduction sub network, the coarse channel estimation network is built; Through interference elimination module, channel feature extraction sub network, slice processing module and data detection sub network, the iterative data detection network is built; Through data preprocessing module and iterative channel estimation sub network, the data aided channel estimation network is built; The coarse channel estimation network, iterative data detection network and data aided channel estimation network are connected in series to obtain the complete joint channel estimation and data detection network; S3, the training data of different networks in the training set are input into the corresponding network model, the feature extraction function of the deep learning model is trained, and the channel state information and data symbol output are predicted from the input data; The specific method of training deep learning model is as follows: First, the received signal is preprocessed, and the real part and the imaginary part of the received signal are split and spliced in the data channel dimension to form tensor data; For coarse channel estimation network, the preprocessed received signal is directly selected as the input training set data of coarse channel estimation network; The least mean square error is taken as the loss function, and the channel estimation network model is trained, and the specific formula is as follows: ， wherein, denotes a loss value of a minimum mean square error loss function, denotes an equivalent channel matrix prediction output of a fully connected deep learning network, denotes a label equivalent channel matrix; For iterative data detection network, the preprocessed received signal and channel state information label data are selected as the input training set data of iterative data detection network; Cross entropy loss is taken as the loss function, and the data detection model is trained, and the specific formula is as follows: ， wherein represents a loss value of a cross-entropy loss function, represents a total number of data symbols, is a class of a data symbol, is a sign function, represents an observation sample belongs to a class a prediction probability; For data aided channel estimation network, the preprocessed received signal and data symbol label data are selected as the input training set data of data aided channel estimation network; At the same time, the training set data of different networks are input into the corresponding network model, and the three different networks are trained in parallel; S4, the equivalent channel matrix is recovered from the output channel state information, and the equivalent channel matrix and its corresponding channel state information label are compared, and the output of coarse channel estimation network model and data aided channel estimation network model is evaluated; S5, the output data symbol and its label are compared, and the output of iterative data detection network model is evaluated; S6, repeat steps S3-S5 for a set number of times to obtain the optimized joint channel estimation and data detection deep learning model; The calculation value of the loss function is back propagated, and the deep learning model parameters are updated, and the updating process is specifically expressed by the following formula: ， wherein, is the parameter of the th deep learning model, is the training learning rate, is the first order derivative of the deep learning model parameter, is the output of the loss function, including and ; S7, the received signal obtained by the receiving antenna in real time is input into the trained model to obtain the channel state information and data detection result.

2. The method of claim 1, wherein the method is based on deep learning in a satellite scenario. The construction method of data set is as follows: Firstly, the OTFS data frame in the delay-Doppler domain is simulated by a two-dimensional matrix with size M×N, which represents the data symbols sent by the transmitting antenna to the receiving antenna in the actual communication system; Next, the data symbols and pilot symbols are inserted according to the pilot design scheme; the matrix after the insertion is sequentially subjected to ISFFT transformation and Heisenberg transformation, so as to convert the delay-Doppler domain signal into time-frequency domain signal and time domain signal; The data at the transmitting antenna end is constructed, the channel parameters are set according to the NTN-TDL scene in the 3GPP standard, and the channel model is initialized and constructed; The transmitting signal of the transmitting antenna is input into the channel model, and the receiving signal at the receiving antenna affected by the channel and environmental noise is further obtained; After obtaining the output result of the channel model, the receiving antenna sequentially converts the time domain signal into time-frequency domain signal and delay-Doppler domain signal through Wigner transformation and SFFT transformation; On the basis of the obtained delay-Doppler domain signal, the receiving signal data is stored, and the training set and test set of the deep learning model are constructed.

3. The device for joint channel estimation and data detection based on deep learning of OTFS in satellite scenario, used for executing the method for joint channel estimation and data detection based on deep learning of OTFS in satellite scenario according to claim 1, characterized in that: The method comprises a coarse channel estimation network, an iterative data detection network and a data-aided channel estimation network connected in series, wherein, The coarse channel estimation network comprises a pilot signal extraction module and a noise reduction subnetwork; The iterative data detection network comprises an interference cancellation module, a channel feature extraction subnetwork, a feature fusion module, a data slicing module and a data detection subnetwork; The data-aided channel estimation network comprises a data preprocessing module and an iterative channel estimation subnetwork.

4. The deep learning based OTFS joint channel estimation and data detection apparatus according to claim 3, characterized in that: The noise reduction subnetwork, the data detection subnetwork and the iterative channel estimation subnetwork are all composed of U-Net networks; the feature extraction subnetwork is composed of multiple convolution units.

5. A computer-readable storage medium, characterized in that: The computer readable storage medium stores computer readable instructions, and the computer readable instructions are executed by the processor to perform the steps of the method of claim 1 or 2.

Images