A deep learning-based OTFS receiver signal detection method and system

Abstract

The application discloses a kind of based on deep learning's OTFS receiver signal detection method and system.In the present application, communication is carried out in the scene of mobile vehicle, and the channel experiences the time and frequency fading of double selectivity;The present application proposes a kind of based on deep learning's OTFS receiver signal detection method to realize signal detection and communication under high mobility scene, adopts deep learning neural network architecture instead of traditional OTFS signal equalization and detection technology, while guaranteeing reliability, try to increase flexibility and efficiency.At the same time, the present application proposes a kind of cyclic shift SP method, for pre-processing signal data, to a certain extent, the problem that convolution kernel has the limitation of two-dimensional signal processing is improved.The present application utilizes deep learning and SP method, so that the performance of the proposed receiver is significantly better than the classical signal detection method under low SNR, provides an effective solution for stable and reliable signal receiver under high mobility scene.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, and relates to vehicular network communication, OTFS modulation theory, and receiver design based on deep learning. Specifically, it relates to a signal detection method and system for an OTFS receiver based on deep learning. Background Technology

[0002] With the Internet of Things gradually becoming a reality and communication networks evolving rapidly, the development of sixth-generation (6G) networks is considered to have entered its infancy, with high-speed, reliable, and stable communication gaining increasing popularity. 6G networks need to meet the requirements of High Mobility Wireless Communication (HMWC), including high-speed mobile communication scenarios such as vehicles, high-speed trains, and satellites. Unlike quasi-static channels with only frequency selectivity, modern communication paths involve complex scatterers at the transmitter and receiver, and the Doppler effect caused by relative motion between the transmitter, receiver, and scatterers results in biselective fading in HMWC scenarios, exhibiting both time-selectivity and frequency-selectivity. Furthermore, due to high-speed movement, this fading is often large-scale. This large-scale biselective fading presents significant challenges to communication in HMWC scenarios. Obtaining accurate channel estimation information, optimizing existing network deployments, and developing advanced signal processing technologies to ensure reliable and stable communication have become urgent problems to be solved.

[0003] In 4G and 5G cellular networks, Orthogonal Frequency Division Multiplexing (OFDM) modulation has become one of the commonly used multi-carrier modulation techniques. However, under the aforementioned channel characteristics, traditional OFDM technology in the time-frequency (TF) domain suffers from severe inter-carrier interference (ICI) in dual-select channels. High-speed movement and high-frequency bands cause high Doppler shift, which severely disrupts the orthogonality between OFDM subcarriers. Meanwhile, Orthogonal Time-Frequency Space (OTFS) modulation, also a multi-carrier modulation technique, is an emerging two-dimensional modulation scheme that maps data symbols to the delayed Doppler (DD) domain, and has become a promising solution for High-Wave-Current Multi-Channel Network (HMWC). OTFS transforms the TF domain channel, which experiences frequency and time fading, into a DD domain channel. Since signal modulation is in the DD domain, the channel parameters remain approximately constant. The channel response in the DD domain is a two-dimensional convolution of the symbol and the channel response, enhancing the stability and correlation of symbol information. This transformation helps overcome the inherent limitations of the OFDM TF domain processing framework.

[0004] OTFS-based channel equalization and signal detection play a crucial role in communication systems. It eliminates the influence of channel parameters on the transmitted signal, thereby recovering the original transmitted signal at the receiver. Its core idea is to use channel state information obtained from channel estimation to inversely compensate for distortions caused by the wireless channel, ultimately identifying and recovering the desired signal from the interference source. This is essential for reliable and stable communication. However, due to the complexity of two-dimensional signal processing, performing physical layer algorithms such as signal detection in the DD domain presents significant challenges. For example, OTFS-based channel estimation algorithms require matched filtering and windowing operations to improve the accuracy of two-dimensional signal processing.

[0005] In recent years, following its success in computer vision, graphics representation, and natural language processing, deep learning (DL) technology has attracted increasing attention for its application in the physical layer of wireless communication. Deep learning is characterized by its ability to automatically learn and extract complex patterns and features from data, more accurately represent complex nonlinear relationships, and lead to a paradigm shift in data processing. Wireless channel conditions and interference patterns are often highly dynamic and complex, posing a substantial challenge to accurate modeling using traditional techniques. Deep learning, however, can provide a fine-grained understanding of the multifaceted relationships between transmitted signals and received observations, thereby enabling more complex and effective signal processing techniques.

[0006] In practical applications, OTFS signal detection involves reversing the distortion caused by the wireless channel to the transmitted signal and distinguishing signal or waveform patterns in the interference background to recover the original transmitted signal. To achieve this goal, deep learning technology can directly obtain the mapping from the received signal to the equalization output, eliminating the need for explicit feature engineering and simplifying the overall equalization process. Furthermore, it can address the slow convergence speed of traditional algorithms through two fundamental features: sparsely connected neurons and shared weights within the receptive domain of each layer, thus improving computational efficiency. Therefore, research on deep learning-based OTFS signal detection methods is highly beneficial for current OTFS receiver research. Summary of the Invention

[0007] Purpose of the invention: The purpose of this invention is to provide a signal detection method and system for OTFS receivers based on deep learning. By utilizing deep learning and OTFS technology, reliable and stable communication can be achieved in vehicular network communication scenarios with high mobility, thereby improving the robustness and flexibility of signal detection.

[0008] Technical solution: To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0009] A deep learning-based OTFS receiver signal detection method includes the following steps:

[0010] The complex information in the two-dimensional DD domain signal is split into real and imaginary parts, and then concatenated in the third dimension to form a data format of (M,N,2); where M is the number of subcarriers and N is the number of time slots.

[0011] The cyclic shift (SP) method is used to divide the two-dimensional DD domain signal of the input data into 9 segments according to the maximum time delay value and the maximum Doppler frequency shift value. Based on the characteristics of two-dimensional cyclic circular convolution, the outer 8 segments are copied and the perimeter of the original frame is filled according to the law of DD domain information migration.

[0012] After the two-dimensional DD domain signal is filled, it is used as the input to the neural network in the deep learning-based OTFS signal receiver;

[0013] After obtaining the output of the neural network, the expanded part is removed to restore the original size, and the equalized symbol of the OTFS modulation symbol is obtained; feature extraction and signal recovery are performed on the signal constellation diagram.

[0014] Furthermore, in the SP method, the extension length of SP is the extension length of the maximum delay taps at both ends of the delay axis and the extension length of the maximum Doppler frequency taps at both ends of the Doppler axis.

[0015] Furthermore, in the SP method, the cutting and filling of the 9 segments are as follows: the length of the segments at both ends of the delay axis is the maximum delay tap, and the length of the segments at both ends of the Doppler axis is the maximum Doppler frequency tap; the segments in the delay axis and Doppler axis directions are filled in a cyclic shift manner.

[0016] Furthermore, the neural network in the deep learning-based OTFS signal receiver adopts a CNN network, which uses the two-dimensional convolutional kernel of the CNN to extract and fit features of the signal in the two-dimensional DD domain; the data input to the two-dimensional convolutional kernel module is preprocessed using the SP method.

[0017] Furthermore, the neural network in the deep learning-based OTFS signal receiver adopts the lightweight neural network MobileNet or MobileViT. Before the two-dimensional convolutional kernel in the network performs data input and preprocessing, the SP method is used to preprocess the input two-dimensional DD domain signal data.

[0018] Furthermore, the training dataset for deep learning neural networks is a set of received symbols from the receiver and modulated symbols from the transmitter.

[0019] This invention also provides a deep learning-based OTFS receiver signal detection system, comprising:

[0020] The preprocessing module is used to split the two-dimensional DD domain signal into real and imaginary parts, and splice them in the third dimension to form a (M,N,2) data format; and to use the SP method to cut the two-dimensional DD domain signal of the input data into 9 segments according to the maximum time delay value and the maximum Doppler frequency shift value, and according to the characteristics of two-dimensional circular convolution, copy the outer 8 segments and fill the perimeter of the original frame according to the law of DD domain information migration.

[0021] The deep learning module is used to fill the two-dimensional DD domain signal and use it as the input to the neural network in the deep learning-based OTFS signal receiver; after obtaining the output of the neural network, the expanded part is removed to restore the original size and obtain the equalized symbol of the OTFS modulation symbol; feature extraction and signal recovery are performed on the signal constellation diagram.

[0022] The present invention also provides a deep learning-based OTFS communication system for in-vehicle networks. The system uses OTFS technology for modulation, transmission and reception of communication signals, and at the receiving end, it uses a deep learning-based OTFS receiver signal detection method to recover and detect the signal.

[0023] The present invention also provides a computer system, including a memory, a processor, and a computer program / instructions stored in the memory and executable on the processor, wherein the computer program / instructions, when executed by the processor, implement the steps of the deep learning-based OTFS receiver signal detection method.

[0024] The present invention also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the deep learning-based OTFS receiver signal detection method.

[0025] Beneficial Effects: Compared with existing technologies, this invention has the following advantages: First, the deep learning-based OTFS signal detection receiver in this invention does not require prior CSI information and can directly learn and recover the original signal from a large amount of data. Second, this invention proposes a cyclic shift SP method for data preprocessing of two-dimensional DD domain signals, which improves the performance of deep learning network models to a certain extent and solves the limitations of two-dimensional convolutional kernels in processing edge information. Third, this invention can use lightweight neural networks to optimize the network deployment of the receiver, reducing parameter calculation while ensuring reliability and increasing flexibility. Attached Figure Description

[0026] Figure 1 This is a system communication scenario diagram according to an embodiment of the present invention.

[0027] Figure 2This is a model diagram of the OTFS communication system according to an embodiment of the present invention.

[0028] Figure 3 This is a format diagram of the input signal data in an embodiment of the present invention.

[0029] Figure 4 This is a model diagram of the OTFS signal detection system based on deep learning in an embodiment of the present invention.

[0030] Figure 5 This is a schematic diagram of the cyclic shift SP method in an embodiment of the present invention.

[0031] Figure 6 These are signal constellation diagrams for different stages in embodiments of the present invention.

[0032] Figure 7 This is a graph showing the average bit error rate of the DD domain signal in an embodiment of the present invention.

[0033] Figure 8 This is a comparison chart of BER performance under different modulation orders in the embodiments of the present invention.

[0034] Figure 9 This is a comparison chart of BER performance under imperfect CSI in an embodiment of the present invention.

[0035] Figure 10 This is a comparison chart of BER performance without encoding in an embodiment of the present invention.

[0036] Figure 11 This is a comparison chart of BER performance under RS ​​encoding in an embodiment of the present invention.

[0037] Figure 12 This is a comparison chart of BER performance under BCH encoding in an embodiment of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the embodiments of this invention are described in detail below with reference to the accompanying drawings. These embodiments are implemented based on the technical solutions of this invention, providing detailed implementation methods and specific operating procedures. It should be understood that the specific examples described herein are merely illustrative of this invention, but the scope of protection of this invention is not limited to the following embodiments.

[0039] This invention considers a highly mobile vehicular network communication scenario where vehicles move at a certain speed, scatterers exist between communication paths, and the communication channel experiences biselective fading. Due to the presence of scatterers and the high-speed movement of vehicles, the communication path between the vehicle and road infrastructure changes in real time, and the channel parameters are not fixed at any given time. Traditional TF domain symbol modulation techniques can no longer meet the requirements for stable and reliable communication under mobile conditions. Therefore, OTFS modulation of the DD domain modulation signal can be applied to high-speed mobile vehicular network communication scenarios, and a deep learning-based signal receiver system replaces the traditional signal detection system for channel equalization and signal detection.

[0040] Therefore, the present invention discloses a deep learning-based OTFS receiver signal detection method. This method first decomposes the complex information of the two-dimensional DD domain signal into real and imaginary parts, and then concatenates them in the third dimension. Next, using the SP method, the two-dimensional DD domain signal of the input data is divided into nine segments according to the maximum time delay and the maximum Doppler frequency shift. Based on the characteristics of two-dimensional circular convolution, the outer eight segments are copied, and the original frame perimeter is filled according to the DD domain information migration pattern. The filled two-dimensional DD domain signal is then used as the input to the neural network in the deep learning-based OTFS signal receiver. After obtaining the output of the neural network, the expanded portion is removed, restoring the original size to obtain the equalized symbol of the OTFS modulation symbol. Feature extraction and signal recovery are then performed on the signal constellation diagram.

[0041] Furthermore, this method considers the highly mobile vehicular network communication scenario. The movement of vehicles and the presence of scatterers cause large-scale changes and fading in the communication path, making the channel dynamic and variable. Unlike traditional OFDM modulation techniques that modulate symbols in the TF domain, this method uses OTFS (Optical Transform-Time Filter) modulation in the DD domain. The channel response in the DD domain is a two-dimensional convolution of the symbol and the channel response, enhancing the stability and correlation of symbol information. This transformation helps overcome the inherent limitations of the OFDM TF domain processing framework.

[0042] This invention utilizes deep learning to extract relevant features from DD domain signals without requiring prior channel state information (CSI). The goal of the deep learning-based signal detection receiver is to recover the received symbols at the receiver end into equalized symbols that approximate the modulation symbols at the transmitter end, performing feature extraction and signal recovery on the signal constellation diagram. During training and fitting, the two-dimensional DD domain signal is split into real and imaginary parts, which are then concatenated as new input.

[0043] In practical implementation, the two-dimensional convolutional kernel of CNN can be used to extract and fit the signal in the two-dimensional DD domain. At the same time, the lightweight neural network MobileNet can be used to optimize the network deployment, and MobileViT can be used to combine the spatial sensing bias and correlation features between local pixels of CNN with the self-attention mechanism of Transformer, so as to extract both local and global information with a small number of parameters.

[0044] Because time delay and Doppler effects cause information at the edges of the DD domain signal to diffuse to the opposite side during convolution operations to capture feature information, 2D convolutional kernels with limited receptive domains cannot capture this variation. After the 2D DD domain signal is filled and used as new input data, the SP method allows the 2D convolutional kernel to effectively capture the diffusion relationship of edge information. After training, the data is restored to its original size by removing the extended segments. The SP method can improve the problem of the limited receptive domain of 2D convolutional kernels, thus achieving more reliable signal detection.

[0045] When using neural networks to extract features from two-dimensional (DD) domain signals, the limited receptive field of the two-dimensional convolutional kernel limits its ability to aggregate local information. This results in the neural network's inability to effectively process edge symbols, even though the information of these edge symbols may be located on the other side of the frame. Applying the SP method to preprocess the input data moves the information constituting the edge symbols to their adjacent positions, allowing the neural network model to aggregate this information effectively.

[0046] Furthermore, for lightweight neural networks MobileNet and MobileViT, which also contain a large number of two-dimensional convolutional kernel modules, the SP method is used to preprocess the input two-dimensional DD domain signal data before the two-dimensional convolutional kernels are used for data input and preprocessing.

[0047] Based on the same inventive concept, this invention also discloses a deep learning-based OTFS receiver signal detection system, comprising: a preprocessing module for splitting the two-dimensional DD domain signal into real and imaginary parts and splicing them in the third dimension; and using the SP method to cut the input data's two-dimensional DD domain signal into 9 segments according to the maximum time delay value and the maximum Doppler frequency shift value, copying the outer 8 segments according to the characteristics of two-dimensional circular convolution, and filling the original frame perimeter according to the DD domain information migration rule; a deep learning module for using the filled two-dimensional DD domain signal as the input to the neural network in the deep learning-based OTFS signal receiver; and after obtaining the output of the neural network, removing the expanded part to restore the original size, obtaining the equalized symbol of the OTFS modulation symbol; and performing feature extraction and signal recovery on the signal constellation diagram.

[0048] This invention also discloses a deep learning-based OTFS (On-The-Flight Forwarding System) communication system for vehicular networks, considering a vehicular network communication scenario where a vehicle is traveling at a certain speed. This scenario involves scatterers between communication paths and biselective fading in the communication channel. Due to the presence of scatterers and the high-speed movement of the vehicle, the communication path between the vehicle and road infrastructure changes in real time, and the channel parameters are not fixed at any given time. This system uses OTFS technology for modulation, transmission, and reception of communication signals. Simultaneously, at the receiving end, a deep learning-based OTFS signal detection receiver is used to recover and detect the signal. The following detailed description of this invention embodiment is provided in conjunction with a specific scenario.

[0049] like Figure 1 As shown in the system communication scenario diagram, this embodiment of the invention considers a highly mobile vehicular network communication scenario. The downlink communication path between the transmitter Tx1 and the receiver vehicle Rx can be equivalent to two paths: one is a direct path, and the other, after passing through scatterer 1, also serves as a propagation path, ultimately superimposed at Rx. Simultaneously, due to the continuous movement of the vehicle, these two communication paths are constantly changing and exhibit frequency-selective fading and time-selective fading. Similarly, the uplink communication path between the transmitter Tx2 and the road communication facilities also changes continuously due to the vehicle's movement. Furthermore, the communication path between Tx2 and Rx is also affected by the vehicle scatterer 2, causing the communication information ultimately superimposed at the receiver to interfere with each other, and each path undergoes large-scale biselective fading. Under these conditions, the OTFS technique, which can sparsify the delayed and Doppler-varied modulated signals in the DD domain, can be applied.

[0050] like Figure 2 As shown in the OTFS communication system model diagram, the communication process within the inner box is the general process of traditional OFDM technology, while the process within the outer box is implemented by adding preprocessing ISFFT and post-processing SFFT to OFDM. OFDM modulates symbols in the time-frequency (TF) domain, while OTFS modulates symbols in the delay-Doppler (DD) domain.

[0051] OTFS communication system model:

[0052] At the transmitting end, the transmitted data bit stream *u* first passes through an encoder for channel coding, increasing redundancy while improving anti-interference and error correction capabilities. Then, a mapper modulates the symbols into a two-dimensional signal, such as Quadrature Amplitude Modulation (QAM). At this point, the transmitted symbol in the DD domain is represented as... That is, transmitting signals in a two-dimensional DD domain with M subcarriers and N time slots. The signals are then obtained from the DD domain symbols using a two-dimensional ISFFT. to TF field symbol It can be represented as:

[0053]

[0054] After adding the send window function, the above formula can be expressed as:

[0055]

[0056] Among them W tx [m,n] represents the send window function.

[0057] Next, the waveform g will be emitted. tx The Heisenberg transform of (t) is applied to X TF This yields the discrete-time signal x(t). Specifically, the Heisenberg transform corresponds to the inverse fast Fourier transform (IFFT) implemented in OFDM modulation. Therefore, the time-domain signal x(t) can be expressed as:

[0058]

[0059] Where g tx (t-nT) represents the received pulse waveform, where T is the time slot duration and Δf is the subcarrier spacing.

[0060] Furthermore, for a linear time-varying (LTV) channel containing p parseable paths, the channel response in the DD domain can be expressed as:

[0061]

[0062] Where h p , τ p and v p These represent the complex fading coefficients, time delay, and Doppler frequency associated with the p analyzable paths, respectively.

[0063] Therefore, the time-domain received signal y(t) after passing through the LTV channel can be expressed as:

[0064] y(t)=∫∫H DD (τ,ν)x(t-τ)e j2πv(t-τ) dτdv+ω(t)

[0065] Where ω(t) represents a mean of 0 and a variance of σ. 2 Additive white Gaussian noise (AWGN). The time-domain received signal y(t) is first converted to the TF domain by the Wegener transform. Similar to the Heisenberg transform, the Wegener transform corresponds to the Fast Fourier Transform (FFT) implemented in OFDM modulation. Then, a matched window g is added. rx (t), then g rx The cross-fuzzy function Y(t,f) between y(t) and y(t) can be expressed as:

[0066]

[0067] Then, the received window function W rx The [m,n] values ​​are summed, and samples are taken at intervals t = nT and f = mΔf to obtain the matched filter output in the TF domain:

[0068] Y TF [m,n]=W rx [m,n]Y(t,f)| t＝nT,f＝mΔf

[0069] Therefore, the input-output relationship of OTFS can be derived as follows:

[0070]

[0071] Where H m,n [m′,n′] is g tx (t), H DD (τ,ν) and g rx The combined effect of (t). Let G represents tx (t) and g rx The cross-mixing function of (t) is written as:

[0072]

[0073] Therefore, H m,n [m′,n′] can be expressed by the formula:

[0074]

[0075] (mm′)Δf-ν)e j2π(v+m′Δf)((n-n′)T-τ) e j2πvn′T dτdv

[0076] Applying the Wigner transform—that is, the FFT—to the receiver, the received symbol in the TF domain can be represented as:

[0077]

[0078] After that, Y TF The expression for converting from a two-dimensional SFFT to the DD domain is:

[0079]

[0080] Finally, the data symbol Y on the DD field of the receiving end. DD After passing through the demapper, the signal is restored to a one-dimensional sequence, which is then decoded by the decoder to obtain the final output bitstream.

[0081] like Figure 3 As shown in the format diagram of the input signal data, in traditional deep learning neural network training, two-dimensional convolutional kernels typically process three-dimensional signals, where the size of the third dimension is usually 2 or 3, representing black-and-white and color images in image processing, respectively. However, unlike traditional image data, signals in the DD domain... The data in Y consists entirely of complex numbers, that is, it is expressed in the form of a + bj. DD The number of channels is 1, and the format is usually only in the form of complex numbers (M,N). However, two-dimensional convolutional kernels generally cannot process two-dimensional data with complex elements and 1 channel. In order to meet the input data format requirements of two-dimensional convolutional kernels and avoid the situation where complex information cannot be processed, the complex information of the received signal can be split into real and imaginary parts, and then concatenated in the same dimension as channel information.

[0082] Deep learning signal detection methods:

[0083] Y DD After splitting into real and imaginary parts, the components are then concatenated in the third dimension to form an input format of (M,N,2). This input is then fed into a deep learning-based OTFS signal detection receiver, allowing the two-dimensional convolutional kernel to process DD domain signals in a manner similar to processing black and white images.

[0084] Therefore, let r be the two-dimensional received symbol, Re(r) be the real part of the received symbol, and Im(r) be the imaginary part of the received symbol, then the input data x of the neural network... d It can be represented as:

[0085] x d =[Re(r),Im(r)]

[0086] Among them are

[0087] Subsequently, the training dataset for the deep learning neural network consists of the set of received symbols from the receiver and the modulation symbols from the transmitter. In other words, the received symbol x... d As the input signal to a deep learning neural network, the modulation symbol y d As labels used by deep learning neural networks to determine whether the learning and fitting are correct or not. Therefore, the dataset Ω can be represented as:

[0088]

[0089] Where x d With y d The superscripts of N represent corresponding data indices. B This indicates the total number of samples in the dataset.

[0090] After inputting the training dataset Ω into the deep learning-based OTFS signal detection receiver, the two-dimensional convolutional kernels in the convolutional layers extract the corresponding data features from the input data. The purpose of training is to optimize the network parameters based on the training dataset generated by the traditional OTFS transmitter, enabling the trained model to achieve excellent bit error rate (BER) performance and improve the network's generalization ability. Specifically, the goal of the deep learning-based OTFS signal detection receiver is to learn the functional relationship using the training weights θ. Therefore, the output data of the neural network It can be represented as:

[0091]

[0092] The reliability of wireless communication systems is primarily reflected in the bit rate (BER). Therefore, the goal of an intelligent receiver for OTFS communication is to minimize the BER between the input and output bit streams. Consequently, the objective of a deep learning-based OTFS signal detection receiver is to minimize y. d and The error between them is considered to minimize the final BER as much as possible. This can be expressed as:

[0093]

[0094] Where L represents the mean squared error (MSE) loss function, and the method for calculating the MSE loss is also given, N D L represents the length of the data, and the goal of a neural network is to minimize L.

[0095] like Figure 4The diagram shows a deep learning-based OTFS signal detection system model. The transmitter generates an input data 0 / 1 bitstream, which is then channel-coded by an encoder and constellation-mapped by a mapper. Taking 4-QAM modulation as an example, after passing through the encoder and mapper, the input bitstream is distributed across four regions on the coordinate axis, as shown in the diagram—the constellation diagram. At this point, the signal has become a two-dimensional DD domain signal. This signal is then OTFS modulated (ISFFT+IFFT), and then passes through an LTV channel with added noise and interference. The signal and the biselective fading channel coefficients are convolved in two dimensions. The signal after passing through the channel is demodulated by OTFS at the receiver and enters the designed deep learning-based OTFS signal detection receiver. It can be seen that the demodulated two-dimensional DD domain signal, after being subjected to biselective fading interference, exhibits a chaotic and disordered form on the constellation diagram. Using this received symbol as input data for a neural network for deep learning, the output equalized symbol essentially restores the four-constellation distribution pattern similar to the transmitter's modulation symbol. The equalization symbols are then processed by a decoder and demapper to obtain the output bitstream. Finally, the BER (Bit Rate) of the output bitstream and the input bitstream is calculated to evaluate the signal detection performance. In the entire OTFS communication process, the portion after the channel until the output bitstream is replaced by the deep learning-based OTFS signal detection receiver system designed in this invention.

[0096] like Figure 5 As shown in the schematic diagram of the circular shift SP method, since the neural network extracts data features through filters, its receptive field is limited, and it can only aggregate local information. When the neural network is applied to process OTFS frames at the receiving end, this characteristic causes the neural network model to be unable to effectively process edge symbols, the information of which may be located on the other side of the entire frame. For example, as shown in the figure, due to Doppler shift, the information of region 6 will spread to region 4. This change migrates from one edge to another opposite edge, and the size of the convolution kernel is often much smaller than the size of the DD domain signal, making it insufficient to capture this change relationship. To address this, the present invention designs the SP method to resolve the conflict between the limited receptive field of CNN and the two-dimensional DD domain signal characteristics of OTFS. First, the received OTFS frame is cut into 9 segments. Then, according to the characteristics of two-dimensional circular convolution, the outer 8 segments are copied and filled into the perimeter of the original frame according to the law of DD domain information migration. After filling, the segments are used as new inputs to train the neural network. At this time, when the original convolution kernel performs feature extraction at the previous position, it can capture the information of regions 4 and 6 that were originally separated by two ends.

[0097] Circular shift SP method:

[0098] In the SP extension, Q represents the maximum delay tap, and P represents the maximum Doppler frequency tap. The extension length of SP is the extension length of the maximum delay taps at both ends of the delay axis and the extension length of the maximum Doppler frequency taps at both ends of the Doppler axis.

[0099] Let E(·,(Q,P)) be the SP spread function with maximum time delay tap Q and maximum Doppler frequency tap P, and simultaneously... This is the inverse process of SP expansion; that is, after obtaining the output of the neural network, the SP-expanded portion is removed, restoring the data to the same format as the label data for loss calculation. Thus, the input data of the neural network can be written as:

[0100] x sp =E(x) d (Q,P))

[0101] Where x sp This represents the input data after SP expansion.

[0102] After the data undergoes SP expansion before being input into the neural network and SP removal after passing through the neural network, the final output data can be rewritten as:

[0103]

[0104] By employing the SP method, information constituting edge symbols is moved to their adjacent positions, allowing the neural network to effectively aggregate this information. This enables the capture of information lost at the edges when the neural network model uses convolutional windows for feature extraction, significantly improving channel equalization and signal detection performance. Furthermore, since MobileNet and MobileViT both contain a certain number of two-dimensional convolutional kernels, this invention also applies the SP method to lightweight neural networks, achieving further performance improvements.

[0105] like Figure 6 The signal constellation diagrams at different stages are shown, illustrating the signal constellation diagrams at different stages under 4-QAM modulation and uncoded OTFS signal conditions with an SNR of 20dB. Three of these diagrams are related to... Figure 6 The three constellations in the picture Figure 1A one-to-one correspondence. It can be seen that, in (a), the modulated signal at the transmitting end, due to 4-QAM modulation, is distributed equidistantly in the four quadrants of the constellation diagram. Meanwhile, the received signal at the receiving end, due to interference from channel biselective fading and the influence of additive white Gaussian noise, has a chaotic constellation distribution in the constellation diagram of (b). After channel equalization by the deep learning-based OTFS signal detection receiver, it can be seen that the constellation diagram of the equalized signal obtained in (c) is very close to the original modulated signal at the transmitting end. Therefore, it can be seen that the deep learning-based OTFS signal detection method of this invention is effective.

[0106] like Figure 7 The average bit error rate (BER) of the DD domain signal is shown in the figure. To better illustrate the limitations of the SP method, the concept of average BER is introduced. Specifically, the two-dimensional DD domain signal is represented as a two-dimensional matrix. Furthermore, the BER is not calculated based on the total number of erroneous bits in a single frame, but rather on the BER of each element in the two-dimensional matrix within a total test set of frames. For example, if the DD domain signal is represented as a matrix [M,N], and element [1,1] is taken as the object of study, then the average BER of this element is the number of erroneous bits at position [1,1] across all test set frames, divided by the total length of the test set frames. Figure 7 As shown, the average BER of an uncoded 4-QAM OTFS signal after CNN-based signal detection at a signal-to-noise ratio of 20dB is demonstrated in the DD domain with and without SP padding. In (a), without SP padding, due to the influence of Doppler frequency shift and the limited receptive field of the 2D convolution kernel, it can be clearly seen that the average BER is higher on both sides of the signal in the DD domain, and the average BER is even higher closer to the edge. Conversely, in (b), after SP padding, the feature of higher average BER at the edge disappears, replaced by a lower BER across the entire plane, and the number of scatter points of the average BER at each level is also much less than that without SP padding. This shows that the data preprocessing method of SP is effective in solving the problem of limited information capture by the convolution window and loss of global information.

[0107] like Figure 8The graph comparing BER performance under different modulation orders shows the BER performance of OTFS signal detection using traditional signal detection algorithms and deep learning-based OTFS signal detection methods under high-order QAM modulation without coding. It can be seen that as the modulation order increases, the BER of traditional signal detection algorithms MRC and LMMSE increases significantly and remains at a high BER value. Conversely, the CNN-based signal detection method maintains relatively stable BER performance even in high-order modulation, demonstrating the stability of the deep learning-based OTFS receiver signal detection method of this invention in high-order modulation. Furthermore, for each modulation order, the BER performance using the SP method is consistently better than that without it, further demonstrating its advantage.

[0108] like Figure 9 The comparison chart of BER performance under imperfect CSI shows the BER performance of traditional signal detection methods and CNN-based signal detection methods for uncoded OTFS signals under 4-QAM modulation when imperfect channel state information (CSI) is introduced. Since traditional signal detection algorithms rely on accurate CSI, the BER performance of MRC and LMMSE deteriorates sharply when the channel estimation error parameter ε appears, making them highly sensitive to the accuracy of CSI. Conversely, because the deep learning-based OTFS signal detection method of this invention does not require prior CSI information, reliable signal recovery can be achieved even with channel estimation errors.

[0109] like Figure 10 The comparison chart of BER performance under uncoded conditions shows the comparison of BER performance of OTFS signals under uncoded and 4-QAM modulation methods with traditional signal detection algorithms and signal detection methods of different network models. It can be seen that under low signal-to-noise ratio (SNR≤12dB) conditions, the deep learning-based OTFS receiver signal detection method exhibits better BER performance than the traditional signal detection methods MRC and LMMSE when perfect CSI is achieved. Meanwhile, the signal detection method with SP extension also shows better BER performance compared to the method without SP extension. In the horizontal comparison of network models, MobileViT shows the best BER performance, followed by CNN, and then MobileNetV1 and MobileNetV2, but the overall performance is relatively close, demonstrating the advantages of lightweight neural networks and helping to optimize network deployment and ensure flexibility.

[0110] like Figure 11The comparison chart of BER performance under RS ​​coding shows the BER performance comparison of OTFS signal detection under 4-QAM modulation with RS coding. When the EVA channel coding of the vehicular network exhibits a certain stability, the deep learning-based signal detection receiver of this invention achieves better BER performance than the traditional OTFS receiver while demonstrating the effectiveness of the neural network. In particular, at an SNR of 12dB, MobileViT, MobileNetV1 with SP extension, and CNN all achieved a BER of 0 for signal detection. At the same SNR, the traditional signal detection algorithms LMMSE and MRC only achieved 10. -3 The BER performance is orders of magnitude higher. Furthermore, in other signal-to-noise ratio aspects, the deep learning-based signal detection algorithm of this invention still demonstrates better BER performance than traditional signal detection algorithms, reflecting its effectiveness.

[0111] like Figure 12 The comparison chart of BER performance under BCH encoding shows the BER performance of the OTFS signal detection method with BCH encoding under 4-QAM modulation. At an SNR of 12dB, MobileViT and CNN also achieved BER performance on the order of magnitude. Furthermore, under these conditions, the deep learning-based OTFS receiver signal detection method of this invention still achieved better signal detection performance than traditional signal detection algorithms. Similarly, the use of the SP method also improved the performance of the deep learning-based OTFS signal detection method.

[0112] Those skilled in the art will understand that the technical solution of this invention, or the part that contributes to the prior art, can be embodied in the form of a software program product. This computer software program product is stored in a storage medium and includes several instructions to cause a computer system (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in the embodiments of this invention. The storage medium includes various media capable of storing computer programs, such as a USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0113] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based OTFS receiver signal detection method, characterized in that, Includes the following steps: The complex information in the two-dimensional time-delay Doppler (DD) domain signal is split into real and imaginary parts, and then concatenated in the third dimension to form a (M,N,2) data format; where M is the number of subcarriers and N is the number of time slots. The cyclic shift SP method is used to divide the two-dimensional DD domain signal of the input data into 9 segments according to the maximum time delay value and the maximum Doppler frequency shift value. Based on the characteristics of two-dimensional cyclic circular convolution, the outer 8 segments are copied and the perimeter of the original frame is filled according to the law of DD domain information migration. After the two-dimensional DD domain signal is filled, it is used as the input to the neural network in the deep learning-based OTFS signal receiver; After obtaining the output of the neural network, the expanded part is removed to restore the original size, and the equalized symbol of the OTFS modulation symbol is obtained; feature extraction and signal recovery are performed on the signal constellation diagram.

2. The OTFS receiver signal detection method based on deep learning according to claim 1, characterized in that, In the SP method, the extension length of SP is the extension length of the maximum delay taps at both ends of the delay axis and the extension length of the maximum Doppler frequency taps at both ends of the Doppler axis.

3. The OTFS receiver signal detection method based on deep learning according to claim 1, characterized in that, In the SP method, the cutting and filling of the 9 segments are as follows: the length of the segments at both ends of the delay axis is the maximum delay tap, and the length of the segments at both ends of the Doppler axis is the maximum Doppler frequency tap; the segments in the delay axis and Doppler axis directions are filled in a cyclic shift manner.

4. The OTFS receiver signal detection method based on deep learning according to claim 1, characterized in that, The neural network in the deep learning-based OTFS signal receiver adopts a CNN network, which uses the two-dimensional convolutional kernel of the CNN to extract and fit features of the signal in the two-dimensional DD domain; the data input to the two-dimensional convolutional kernel module is preprocessed using the SP method.

5. The OTFS receiver signal detection method based on deep learning according to claim 1, characterized in that, The neural network in the deep learning-based OTFS signal receiver adopts the lightweight neural network MobileNet or MobileViT. Before the two-dimensional convolutional kernel in the network performs data input and preprocessing, the SP method is used to preprocess the input two-dimensional DD domain signal data.

6. The OTFS receiver signal detection method based on deep learning according to claim 1, characterized in that, The training dataset for deep learning neural networks is a set of received symbols from the receiver and modulated symbols from the transmitter.

7. A deep learning-based OTFS receiver signal detection system, characterized in that, include: The preprocessing module is used to split the two-dimensional time-delay Doppler (DD) domain signal into real and imaginary parts, and then splice them together in the third dimension to form a (M,N,2) data format; where M is the number of subcarriers and N is the number of time slots. Furthermore, the cyclic shift SP method is used to divide the two-dimensional DD domain signal of the input data into 9 segments according to the maximum time delay value and the maximum Doppler frequency shift value. Based on the characteristics of two-dimensional cyclic circular convolution, the outer 8 segments are copied and the perimeter of the original frame is filled according to the law of DD domain information migration. The deep learning module is used to fill the two-dimensional DD domain signal and use it as the input to the neural network in the deep learning-based OTFS signal receiver; after obtaining the output of the neural network, the expanded part is removed to restore the original size and obtain the equalized symbol of the OTFS modulation symbol; feature extraction and signal recovery are performed on the signal constellation diagram.

8. A deep learning-based OTFS communication system for in-vehicle networks, characterized in that, The system uses OTFS technology for modulation, transmission, and reception of communication signals, and at the receiving end, it uses a deep learning-based OTFS receiver signal detection method according to any one of claims 1-5 to recover and detect the signal.

9. A computer system comprising a memory, a processor, and computer programs / instructions stored in the memory and executable on the processor, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the deep learning-based OTFS receiver signal detection method according to any one of claims 1-6.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the deep learning-based OTFS receiver signal detection method according to any one of claims 1-6.

Images