Lightweight OTFS end-to-end signal detection method and device for high-speed moving scene

Abstract

The invention discloses a lightweight OTFS end-to-end signal detection method and device for a high-speed mobile scene, belongs to the technical field of wireless communication, and aims to overcome the defects that an existing OTFS signal detection technology depends on channel state information prior, detection performance is deteriorated under a high-speed time-varying channel, and detection precision and lightweight end-side deployment cannot be considered at the same time. The method comprises the following steps: converting a time domain received signal to obtain a delay-Doppler domain complex signal two-dimensional matrix, splitting the delay-Doppler domain complex signal two-dimensional matrix into a real part matrix and an imaginary part matrix to construct a dual-channel input tensor, inputting the dual-channel input tensor into a lightweight detection model based on improved GhostNet fusion CBAM, outputting the soft decision probability of each bit of a single-frame signal, obtaining an original bit stream through threshold decision, and finally obtaining a final detection result. And channel state information priori and independent channel estimation links are not needed in the whole process. According to the OTFS signal detection method, multi-stage processing error accumulation can be eliminated, high-precision and low-complexity OTFS signal detection in a high-speed scene is realized, and the OTFS signal detection method can be deployed in a resource-limited terminal and is suitable for scenes of Internet of Vehicles, low-altitude communication, satellite communication and the like.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, specifically relating to signal detection technology for Orthogonal Time-Frequency Space (OTFS) communication systems. It is particularly suitable for high-speed mobile scenarios such as vehicle networking, satellite communication, and low-altitude communication. It can realize lightweight and robust end-to-end detection of OTFS signals on resource-constrained terminal devices, providing technical support for reliable communication in high-speed time-varying channel environments. Background Technology

[0002] As mobile communication evolves towards the 6th Generation (6G) era, high-speed mobile scenarios such as vehicle-to-everything (V2X) communication, high-speed rail communication, and low-altitude drone communication are becoming increasingly common, placing stringent demands on the Doppler frequency offset resistance, multipath fading resistance, and real-time performance of communication systems. Traditional Orthogonal Frequency Division Multiplexing (OFDM) technology is susceptible to dual selective fading in high-speed mobile scenarios, resulting in severe inter-carrier interference and a significant decrease in communication reliability. Time-Delay-Doppler (OTFS) technology, through signal modulation and processing in the Delay-Doppler (DD) domain, transforms the rapidly time-varying channel in the time-frequency domain into a sparse and stable channel, becoming one of the core technologies for solving communication challenges in high-speed mobile scenarios.

[0003] Signal detection is a crucial component of the OTFS communication system receiver, and its performance directly determines the overall reliability and real-time performance of the system. Existing OTFS signal detection technologies are mainly divided into two categories: traditional algorithms and deep learning-assisted algorithms. Traditional algorithms rely on accurate prior Channel State Information (CSI) and require complex serial steps such as channel estimation, equalization, and symbol demodulation to achieve signal recovery. This not only results in high system complexity but also makes them prone to error accumulation under high-speed, time-varying channels, leading to performance degradation. While deep learning-assisted algorithms simplify the processing flow through a data-driven approach, most solutions fail to balance detection accuracy with lightweight deployment requirements or do not achieve true end-to-end bit-level recovery, making them unsuitable for resource-constrained applications such as vehicle-mounted and airborne terminals.

[0004] Among the publicly available related technologies, existing solutions still have limitations and cannot simultaneously meet the end-to-end low latency, high accuracy, and lightweight deployment requirements of OTFS signal detection in high-speed mobile scenarios. For example, the detection scheme based on residual channel attention network in patent CN 118432999 A does not achieve direct end-to-end mapping from the received signal to the original bit stream. The processing flow still requires additional encoding / decoding and modulation / demodulation stages, and the network structure is complex with a large number of parameters and computational overhead, making it unsuitable for lightweight deployment of resource-constrained terminals. The transceiver scheme based on DDT deep learning model and pilot-assisted detection in patent CN 118694647 A still does not get rid of the indirect dependence on channel state information. Under high-speed time-varying channels, the detection performance is prone to decline due to CSI time-out. At the same time, the model training complexity is high, the hardware resource requirements are large, and the adaptability to low-cost single-antenna terminals is insufficient. These defects limit the large-scale application of OTFS technology in high-speed mobile scenarios such as vehicle networking and low-altitude communication.

[0005] In summary, existing patented technologies still have room for improvement in terms of end-to-end processing capabilities for OTFS signal detection, lightweight deployment adaptability, and robustness to high-speed time-varying channels. Therefore, this invention proposes a lightweight OTFS end-to-end signal detection method and apparatus for high-speed mobile scenarios. It completely eliminates the traditional serial processing steps and reliance on CSI, achieving high-precision, low-complexity signal detection through a customized lightweight deep learning architecture, thus meeting the real-time communication needs of resource-constrained terminals in high-speed mobile scenarios. Summary of the Invention

[0006] To address the shortcomings of existing OTFS signal detection technologies, such as the heavy reliance on prior channel state information, high system complexity, severe performance degradation under high-speed time-varying channels, and the inability of existing deep learning detection schemes to balance detection accuracy with lightweight deployment, the lack of true end-to-end bit-level recovery, and insufficient adaptability to high-speed mobile scenarios, this invention aims to provide a lightweight OTFS end-to-end signal detection method and apparatus for high-speed mobile scenarios. This method completely eliminates the traditional serial and independent signal processing stage at the OTFS receiver, establishing a direct end-to-end mapping from the received signal to the original bitstream. It achieves highly robust and high-precision OTFS signal detection without requiring prior channel state information. Furthermore, through a customized lightweight network architecture, it adapts to the deployment needs of resource-constrained devices such as vehicle-mounted terminals and airborne terminals, meeting the real-time communication requirements of high-speed mobile scenarios such as vehicle networking, satellite communication, and low-altitude communication.

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

[0008] First, this invention provides a lightweight OTFS end-to-end signal detection method for high-speed mobile scenarios, applied to the receiving end of an OTFS communication system. The method includes the following steps:

[0009] S1: Transform the time-domain received signal acquired by the OTFS receiver to obtain a two-dimensional complex matrix of the DD domain signal;

[0010] S2: The two-dimensional complex signal matrix in the DD domain is split into a real part matrix and an imaginary part matrix with the same dimension. The real part matrix is ​​used as the first channel and the imaginary part matrix is ​​used as the second channel to construct a dual-channel input tensor.

[0011] S3: Input the pre-trained lightweight OTFS end-to-end detection model into the dual-channel input tensor. The lightweight OTFS end-to-end detection model is a deep learning model based on an improved GhostNet and fused with a Convolutional Block Attention Module (CBAM).

[0012] S4: The lightweight OTFS end-to-end detection model outputs the soft decision probability of each bit in a single frame of OTFS signal through forward propagation. After hard decision by a preset threshold, the original bit stream of the transmitter is directly obtained, completing the end-to-end detection of the OTFS signal. No prior channel state information or independent channel estimation is required throughout the process.

[0013] Further, in step S1, the transformation process specifically involves: first, mapping the time-domain received signal to the time-frequency domain using the Wigner transform, and then converting the time-frequency domain signal into a two-dimensional complex signal matrix in the DD domain using the Symptotic Finite Fourier Transform (SFFT). The two dimensions of the DD domain complex signal matrix correspond to the delay domain and the Doppler domain, respectively, and each element in the matrix contains the amplitude and phase information of the signal at the corresponding position.

[0014] Furthermore, in step S2, the dimension of the dual-channel input tensor is 2×N×M, where N is the number of delay domain sampling points of the two-dimensional matrix of the complex signal in the DD domain, and M is the number of Doppler domain sampling points. This dual-channel input tensor completely preserves the two-dimensional joint features of the amplitude, phase, delay dimension and Doppler dimension of the OTFSDD domain signal, avoiding the loss of complex signal phase information caused by single-channel input.

[0015] Further, in step S3, the lightweight OTFS end-to-end detection model, along the signal transmission direction, sequentially includes an input adaptation layer, a multi-stage Ghost bottleneck feature extraction layer, a CBAM attention enhancement module, a feature aggregation layer, and a bit mapping output layer. The input adaptation layer employs a 3×3 standard convolution to perform preliminary feature fusion and extraction on the dual-channel input tensor, mapping the dual-channel features to multi-channel basic features. The multi-stage Ghost bottleneck feature extraction layer is composed of multiple stacked Ghost bottlenecks. These Ghost bottlenecks retain the lightweight core structure of generating inherent features through the main convolution of the Ghost module and generating Ghost features through depthwise convolution, specifically targeting OTFS. The DD domain 2D signal feature optimization method optimizes the convolution kernel size, channel expansion ratio, and number of stacked layers to complete feature extraction from shallow to deep layers with extremely low computational overhead. The CBAM attention enhancement module is embedded in the output of the mid-to-high-level Ghost bottleneck to perform dual recalibration of the features in both channel and spatial dimensions. The feature aggregation layer uses global average pooling to reduce the dimensionality and aggregate deep features, thereby reducing the number of model parameters and computational cost. The bit mapping output layer maps the aggregated features to a dimension that matches the total number of bits in a single frame of OTFS signal and outputs the soft decision probability of each bit.

[0016] Furthermore, the CBAM attention enhancement module is only embedded after the mid-to-high-level Ghost bottleneck where the number of feature channels reaches a preset threshold. Its fusion logic is as follows: the features output by the Ghost bottleneck are processed by batch normalization and activation functions, and then sequentially input into the channel attention unit and spatial attention unit of CBAM to complete the feature recalibration of the channel dimension and the delay-Doppler spatial dimension, respectively, and finally output the enhanced features to the next layer of the network; the channel attention unit performs global average pooling and global max pooling on the input features, and obtains the channel attention weights after passing through a shared fully connected layer and Sigmoid activation, thus completing the channel dimension recalibration; the spatial attention unit performs average pooling and max pooling on the channel recalibrated features, and after concatenating the pooling results, obtains the spatial attention weights after passing through a convolutional layer and Sigmoid activation, thus completing the spatial dimension recalibration.

[0017] Further, in step S4, the bit mapping output layer includes a fully connected layer and a Sigmoid activation layer. The output dimension of the fully connected layer is consistent with the total number of bits in a single frame OTFS signal. The Sigmoid activation layer maps the output value to the [0,1] interval to obtain the soft decision probability of each bit being 1. The preset threshold for the hard decision is 0.5. A soft decision probability greater than 0.5 is determined as bit 1, and a soft decision probability less than or equal to 0.5 is determined as bit 0.

[0018] For the lightweight OTFS end-to-end detection model, it is optimized during the offline training phase using a joint loss function combining the main loss and auxiliary loss. The expression of the joint loss function is as follows:

[0019]

[0020] in, The main loss is a weighted binary cross-entropy loss, which corresponds to the core optimization objective of bit-level detection. As an auxiliary loss, symbol-level mean squared error loss is used to help the model learn the constellation point mapping relationship of OTFS modulation symbols; These are auxiliary loss weighting coefficients used to balance the optimization priorities of the main loss and auxiliary loss.

[0021] Furthermore, during the offline training phase of the lightweight OTFS end-to-end detection model, an Adaptive Moment Estimation with Weight Decoupling (AdamW) optimizer with decoupled weight decay is used for parameter iterative updates. The training process employs a cosine annealing learning rate scheduling strategy with a preheating mechanism, and regularization is performed through a dropout layer and a weight decay mechanism to avoid model overfitting and improve the model's generalization ability to time-varying channels.

[0022] Correspondingly, the present invention also provides a lightweight OTFS end-to-end signal detection device for high-speed mobile scenarios. This device includes a DD domain signal conversion module, a dual-channel input construction module, an end-to-end signal detection module, and a bit decision output module, which are sequentially connected in communication. The DD domain signal conversion module is used to transform the time-domain received signal acquired by the OTFS receiver and output a two-dimensional complex DD domain signal matrix. The dual-channel input construction module is used to split the two-dimensional complex DD domain signal matrix into a real part matrix and an imaginary part matrix with consistent dimensions, using the real part matrix as the first channel and the imaginary part matrix as the second channel to construct a dual-channel input tensor. The end-to-end signal detection module has a pre-trained lightweight OTFS end-to-end detection model built-in. This model is a deep learning model based on an improved GhostNet and incorporating the CBAM attention mechanism, used to receive the dual-channel input tensor and output the soft decision probability of each bit in a single-frame OTFS signal. The bit decision output module is used to perform hard decision on the soft decision probabilities with a preset threshold, directly outputting the original bit stream from the transmitter to complete the end-to-end detection of the OTFS signal.

[0023] Furthermore, the device also includes an offline training module, which is communicatively connected to the end-to-end signal detection module and is used to complete the offline training and weight update of the lightweight OTFS end-to-end detection model. The offline training module has a built-in joint loss function optimization unit, an AdamW optimizer, and a learning rate scheduling unit, which are used for model loss calculation, parameter iterative update, and dynamic adjustment of the learning rate, respectively.

[0024] III. Beneficial Effects

[0025] Compared with the prior art, the present invention has the following beneficial technical effects:

[0026] This invention constructs a full-link end-to-end OTFS signal detection architecture, completely abandoning the traditional serial and independent processing steps of channel estimation → equalization → symbol demodulation → bit decoding at the OTFS receiver. It directly establishes an end-to-end mapping relationship between the received signal in the DD domain and the original bit stream at the transmitter, without requiring any prior channel state information. This eliminates the error accumulation caused by multi-level processing from the root, while greatly simplifying the system architecture of the OTFS receiver and significantly reducing the overall system computational overhead.

[0027] This invention balances detection accuracy and edge deployment requirements through a dual-channel input architecture and a customized lightweight network design. The dual-channel input architecture, designed for the physical characteristics of OTFS DD domain complex signals, fully preserves the amplitude, phase, and delay-Doppler two-dimensional joint features of the signal. Based on an improved GhostNet model incorporating the CBAM attention mechanism, a lightweight model achieves accurate extraction of effective signal features and effective suppression of noise interference with extremely low computational and storage overhead. It can be directly deployed on resource-constrained edge hardware devices such as vehicle terminals and airborne terminals.

[0028] This invention achieves strong robustness, high generalization, and excellent engineering feasibility in high-speed mobile scenarios through a training and optimization system specifically designed for communication scenarios. The joint loss function, designed with bit-level detection as its core objective, combined with the AdamW optimizer and a cosine annealing learning rate scheduling strategy with a warm-up mechanism, significantly improves the model's convergence accuracy and adaptability to high-speed time-varying channels, maintaining stable detection performance even under conditions of Doppler frequency offset and strong multipath fading. Simultaneously, the model adopts a hardware-friendly operating design, adaptable to various mainstream terminal hardware platforms, requiring no large-scale modification of existing OTFS communication systems, and can be widely applied to various high-speed mobile wireless communication scenarios such as vehicle-to-everything (V2X) communication, low-altitude communication, and satellite communication. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the transmission architecture and end-to-end signal detection principle of the OTFS communication system of the present invention;

[0030] Figure 2 This is a schematic diagram of the lightweight OTFS end-to-end signal detection method for high-speed mobile scenarios according to the present invention.

[0031] Figure 3 This is a schematic diagram of the construction of the dual-channel input tensor of the complex signal in the OTFS DD domain of the present invention. Detailed Implementation

[0032] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments are now described. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0033] This invention discloses a lightweight OTFS end-to-end signal detection method and apparatus for high-speed mobile scenarios, which can be widely applied to various high-speed mobile wireless communication scenarios such as vehicle-to-everything (V2V) communication, high-speed rail communication, low-altitude communication, satellite communication, and high-speed UAV communication. The following embodiments use a V2V direct communication scenario as the core example. The scenario is a highway environment where the relative speed between the two vehicles is set to 500 km / h, corresponding to a maximum Doppler frequency offset of 1840 Hz at a dedicated 5.9 GHz V2V carrier frequency. The communication system uses OTFS modulation. The specific implementation process of this invention is described in detail.

[0034] Example 1

[0035] This embodiment provides a specific implementation of a lightweight OTFS end-to-end signal detection method for high-speed mobile scenarios, which is divided into two main stages: offline training of the lightweight OTFS end-to-end detection model and online end-to-end signal detection. First, the basic principles of the OTFS communication system on which this invention is based are explained, and then the specific implementation process is described in detail.

[0036] OTFS is a wireless communication technology that modulates information in the delay-Doppler (DD) domain. Its core principle is to modulate the transmitted symbols onto a delay-Doppler two-dimensional grid in the DD domain, and then use a family of two-dimensional Fourier transforms to map the time-frequency domain to the time domain. In high-speed mobile scenarios, compared to traditional OFDM technology, it exhibits stronger resistance to Doppler frequency offset and better channel stability. The complete transmission process of an OTFS communication system consists of two main stages: modulation at the transmitter and demodulation at the receiver. For its system architecture and signal detection methods, please refer to [link to documentation]. Figure 1 The core mathematical model of OTFS is as follows:

[0037] 1. Transmitter OTFS modulation model

[0038] Let N be the number of delay domain sampling points and M be the number of Doppler domain sampling points in the OTFS system. The bit stream to be transmitted at the transmitter, after constellation modulation, is mapped to an N×M dimensional complex symbol matrix in the DD domain. , where matrix elements Let be the transmitted modulation symbol at the k-th delay grid point and the l-th Doppler grid point in the DD domain, where k = 0, 1, ..., N−1 is the delay domain index and l = 0, 1, ..., M−1 is the Doppler domain index.

[0039] The transmitter maps the DD-domain transmission symbol matrix to the time-frequency (TF) domain using an inverse symplectic finite Fourier transform (ISFFT) to obtain the time-frequency domain transmission matrix. The mathematical expression for ISFFT is:

[0040]

[0041] Where n = 0, 1, ..., N−1 is the time-domain symbol index, and m = 0, 1, ..., M−1 is the frequency-domain subcarrier index. Time-frequency domain transmission matrix The Wigner transform, also known as the Heisenberg transform, maps the transmitted signal to the time domain. The expression is:

[0042]

[0043] in, For OFDM symbol duration, For subcarrier spacing, This is the pulse shaping function for the emitted pulse.

[0044] 2. Time-varying channel transmission model

[0045] The wireless channel adopts a linear time-varying channel model, and its time-varying channel impulse response is: ,in, For path delay, For time. Time-domain received signal. The expression is:

[0046]

[0047] in, It is Additive White Gaussian Noise (AWGN).

[0048] 3. OTFS demodulation model at the receiver end

[0049] The receiver receives the signal in the time domain. Demodulation processing, the inverse process of the transmitter, is performed by first mapping the time-domain received signal to the time-frequency domain using the inverse Wigner transform, thus obtaining the time-frequency domain received matrix. Then, the time-frequency domain signal is converted into a two-dimensional complex signal matrix in the DD domain using SFFT. The mathematical expression for SFFT is:

[0050]

[0051] in, The received complex signal at the k-th delay grid point and the l-th Doppler grid point in the DD domain is the output signal of step S1 of the present invention.

[0052] The input-output relationship of the DD domain can be simplified as follows:

[0053]

[0054] Where H is the DD domain channel matrix, ⊙ is the Hadamard product, and W is the DD domain Gaussian white noise matrix. Traditional OTFS detection algorithms require obtaining prior information about H through channel estimation, and then recovering the transmitted bit stream through equalization and demodulation. However, this invention does not require prior channel state information and directly establishes an end-to-end mapping from Y to the original bit stream at the transmitter through a deep learning model.

[0055] In this embodiment, the core basic parameters of the OTFS communication system are set as follows:

[0056] 1. Carrier frequency =5.9 GHz, subcarrier spacing =15 kHz, symbol duration ≈66.67μs, Cyclic Prefix (CP) length =4.69 μs, compatible with V2X communication standards for vehicle-to-everything (V2X) networks;

[0057] 2. OTFS frame structure: the number of sampling points in the delay domain is N=32, the number of sampling points in the Doppler domain is M=32, that is, the size of the two-dimensional matrix of the complex signal in the DD domain is 32×32;

[0058] 3. Modulation method: Quadrature amplitude modulation (QPSK) is used. A single frame of OTFS signal contains N×M=1024 modulation symbols, corresponding to a total number of bits L=2048 bits;

[0059] 4. Channel Model: The Extended Vehicular Amodel (EVA) defined by the 3GPP standard is adopted. This model is the standard channel model for high-speed mobile scenarios in vehicle-to-everything (V2X) communication and includes 9 distinguishable multipath channels. The standard parameters of the multipath channels are shown in the table below:

[0060] Multipath number Relative delay (ns) Normalized power (dB) 1 0 0.0 2 30 -1.5 3 150 -1.4 4 310 -3.6 5 370 -0.6 6 710 -9.1 7 1090 -7.0 8 1730 -12.0 9 2510 -16.9

[0061] The channel Doppler spectrum adopts the classic Jakes spectrum, with a maximum Doppler frequency offset range of 0~1840 Hz, corresponding to a relative moving speed of 0~500 km / h between two vehicles, and a signal-to-noise ratio (SNR) range of 0~20 dB, covering typical high-speed moving scenarios of highway vehicle-to-everything (V2X) networks.

[0062] 5. The sampling rate of the receiver's RF front-end is matched with the system bandwidth. The sampling rate is set to 1.92 MHz, and the length of the acquired time-domain received signal is consistent with the OTFS frame length.

[0063] Offline training of a lightweight OTFS end-to-end detection model. The core objective of this stage is to complete the training and optimization of the GhostNet-CBAM model to obtain a pre-trained weight model that can be used for online inference. The specific steps are as follows:

[0064] Step 1: Training Dataset Construction

[0065] The OTFS communication dataset covering the above-mentioned high-speed mobile scenarios in vehicle-to-everything (V2X) networks was generated using the MATLAB communication simulation platform. The specific implementation is as follows:

[0066] 1. Sample composition: Each sample contains three core data types: [DD domain dual-channel input tensor, transmitter raw bitstream label, and transmitter modulation symbol label].

[0067] 2. Scene coverage: The dataset is based on the above EVA extended vehicle channel model, covering the full combination of moving speed 50~500km / h, signal-to-noise ratio 0~20dB, 9-path time-varying multipath channel, and QPSK modulation. 10,000 independent samples are generated for each scene, with a total sample size of 1.2 million.

[0068] 3. Data Preprocessing: For the two-dimensional complex signal matrix in the DD domain of each sample, it is split into a 32×32 real part matrix and a 32×32 imaginary part matrix to construct a 2×32×32 dual-channel input tensor; the original bitstream label at the transmitting end is binarized and encoded, with a value of 0 or 1; the modulation symbol label at the transmitting end is normalized to obtain the normalized coordinates of the QPSK constellation points;

[0069] 4. Dataset partitioning: The dataset is randomly divided into a training set (200,000 samples), a validation set (20,000 samples), and a test set (20,000 samples) in an 8:1:1 ratio. The training set is used for iterative updates of model parameters, the validation set is used to monitor model convergence and generalization ability, and the test set is used for final model performance evaluation.

[0070] Step 2: Construction and initialization of a lightweight OTFS end-to-end detection model

[0071] In this embodiment, a lightweight OTFS end-to-end detection model based on an improved GhostNet and fused with the CBAM attention mechanism is used. The overall input is a 2×32×32 dual-channel input tensor, and the output is a 2048-dimensional bit soft-decision probability vector. The specific structure is as follows:

[0072] 1. Input adaptation layer: 3×3 standard convolution is used, with 16 kernels, stride 1, padding 1, 2 input channels and 16 output channels. Preliminary feature fusion is performed on the dual-channel input tensor, and a 16×32×32 basic feature map is output.

[0073] 2. Multi-stage Ghost bottleneck feature extraction layer: A total of 4 Ghost bottlenecks are stacked. Each Ghost bottleneck retains the lightweight core structure of generating inherent features by main convolution + generating Ghost features by depth convolution. The main convolution adopts 1×1 standard convolution, and the depth convolution adopts 3×3 depth separable convolution, which greatly reduces the computational cost and parameter count of the model.

[0074] 3. CBAM Attention Enhancement Module: The preset channel number threshold is 64. The CBAM attention enhancement module is only embedded at the output end of the high-level Ghost bottleneck in groups 3 and 4. The specific fusion logic is as follows:

[0075] The features output by the Ghost bottleneck are first processed by a batch normalization layer and a ReLU activation function;

[0076] Input channel attention unit: The input features are subjected to global average pooling and global max pooling respectively to obtain two 1×C×1×1 feature vectors, where C is the number of input channels. After passing through two shared fully connected layers and a sigmoid activation function, the channel attention weights are obtained and multiplied with the input features to complete the channel dimension recalibration.

[0077] Input Spatial Attention Unit: After channel recalibration, the features are subjected to average pooling and max pooling in the channel dimension to obtain two 1×H×W feature maps. After concatenation along the channel dimension, they are passed through a 7×7 convolutional layer and a Sigmoid activation function to obtain spatial attention weights. These weights are multiplied with the input features to complete the recalibration of the delayed-Doppler spatial dimension. Finally, the enhanced features are output to the next layer of the network.

[0078] Feature aggregation layer: Global average pooling is used to perform global average pooling on the 128×4×4 deep feature map to obtain a 128-dimensional aggregated feature vector, thus completing feature dimensionality reduction and aggregation.

[0079] Bit-mapped output layer: consists of two fully connected layers and a Sigmoid activation layer. The first fully connected layer has an input dimension of 128 and an output dimension of 512. The second fully connected layer has an input dimension of 512 and an output dimension of 2048. The Sigmoid activation layer maps the output values ​​of the fully connected layers to the range of 0 to 1 and outputs the soft decision probability of each bit being 1.

[0080] During the model initialization phase, the weights of each convolutional layer and fully connected layer are initialized using the He normal distribution, the bias term is initialized to 0, the weights of the batch normalized layer are initialized to 1, and the bias term is initialized to 0.

[0081] Step 3: Setting the loss function, optimizer, and training strategy for model training

[0082] 1. Joint Loss Function Setting: A joint loss function combining the main loss and auxiliary loss is adopted, and its expression is:

[0083]

[0084] In this embodiment, the auxiliary loss weight coefficient λ is set to 0.1 to balance the optimization priorities of the main loss and the auxiliary loss, ensuring that the core optimization objective of the model is bit-level detection accuracy.

[0085] 2. Main Loss Function: A weighted binary cross-entropy loss is used, directly corresponding to the binary classification task of bit-level detection. Optimizing this loss is equivalent to directly minimizing the system bit error rate, expressed as:

[0086]

[0087] Among them, the number of training batch samples B=64, and the total number of OTFS signal bits in a single frame L=2048; This is the true label of the j-th bit of the i-th sample, with a value of 0 or 1; The soft decision probability of the i-th sample and j-th bit in the model output; bit weight coefficients. In this embodiment, the default value is 1. For the error-prone bits corresponding to high Doppler frequency offset in high-speed mobile scenarios, it can be set to 1.2~1.5 to enhance the model's learning ability for error-prone bits.

[0088] 3. Auxiliary Loss Function: A symbol-level mean squared error loss is used. Symbol-level features are extracted from the output of the Ghost bottleneck in the fourth group of the model. After mapping, the coordinates of the predicted symbol constellation points are obtained. This is used to assist the model in learning the constellation point mapping relationship of OTFS modulation symbols, accelerating model convergence. The expression is:

[0089]

[0090] The total number of modulation symbols in a single frame OTFS signal is S=1024; The coordinates of the constellation points of the original modulation symbol of the k-th transmitter of the i-th sample; The coordinates of the predicted symbol constellation points output by the intermediate layer of the model.

[0091] 4. Optimizer and training strategy settings:

[0092] The core optimizer uses the AdamW optimizer, with weight decay coefficients set to 1e−4, β1=0.9, β2=0.999, and initial learning rate set to 1e−3. The learning rate scheduling adopts a cosine annealing learning rate scheduling strategy with a warm-up mechanism. The warm-up rounds are set to 5 rounds, and the total number of training rounds is set to 100 rounds. During the warm-up phase, the learning rate linearly increases from 1e−6 to 1e−3, and during the subsequent training phases, the learning rate smoothly decays to 1e−6 with the help of the cosine function.

[0093] Regularization strategy: Add a dropout layer after the first fully connected layer of the bit-mapped output layer, with a dropout rate of 0.2, in conjunction with the weight decay mechanism of the AdamW optimizer, to suppress model overfitting;

[0094] Early stopping mechanism: After each training round, the model's bit error rate performance is tested on the validation set. When the bit error rate on the validation set does not decrease for 10 consecutive rounds, the model is considered to have converged, training is stopped early, and the weights of the model with the best performance on the validation set are saved for subsequent online inference.

[0095] Step 4: Model Training and Performance Validation. Input the training set samples into the completed model, and perform forward propagation, loss calculation, backpropagation, and parameter iterative updates according to the settings described above. After training, perform performance validation on the test set.

[0096] IV. Implementation of Online End-to-End Signal Detection

[0097] In this embodiment, the receiver of the vehicle-mounted OTFS communication terminal in the Internet of Vehicles uses the lightweight OTFS end-to-end detection model trained above to complete online end-to-end signal detection. The specific steps are as follows, see below. Figure 2 :

[0098] S1: Transform the time-domain received signal acquired by the OTFS receiver to obtain a two-dimensional complex matrix of the DD domain signal;

[0099] Specifically, the vehicle-mounted OTFS receiver acquires the time-domain received signal through the radio frequency front end. First, it maps the time-domain received signal to the time-frequency domain through the inverse Wigner transform, and then converts the time-frequency domain signal into a 32×32 DD domain complex signal two-dimensional matrix through SFFT. The rows of the matrix correspond to the delay domain, the columns correspond to the Doppler domain, and each element is a complex value containing the amplitude and phase information of the corresponding position signal.

[0100] S2: Decompose the two-dimensional complex signal matrix in the DD domain into a real part matrix and an imaginary part matrix of the same dimension. Using the real part matrix as the first channel and the imaginary part matrix as the second channel, construct a dual-channel input tensor. See [link to documentation]. Figure 3 ;

[0101] Specifically, the 32×32 DD domain complex signal two-dimensional matrix obtained in step S1 is split into a real part matrix and an imaginary part matrix, both with dimensions of 32×32. The real part matrix is ​​used as the first channel and the imaginary part matrix is ​​used as the second channel to construct a dual-channel input tensor with dimensions of 2×32×32. The tensor is normalized in the same way as in the training stage, and the values ​​are mapped to the −11 interval. This tensor completely preserves the two-dimensional joint features of the amplitude, phase, delay dimensions and Doppler dimension of the OTFS DD domain signal.

[0102] S3: Input the dual-channel input tensor into the pre-trained lightweight OTFS end-to-end detection model. The lightweight OTFS end-to-end detection model is a deep learning model based on an improved GhostNet and incorporating the CBAM attention mechanism.

[0103] Specifically, the preprocessed dual-channel input tensor is input into the pre-trained lightweight OTFS end-to-end detection model. The model propagates forward sequentially through the input adaptation layer, the multi-level Ghost bottleneck feature extraction layer, the CBAM attention enhancement module, the feature aggregation layer, and the bit mapping output layer, outputting a 2048-dimensional soft decision probability vector. Each element in the vector corresponds to the soft decision probability of each bit being 1 in a single frame of the OTFS signal.

[0104] S4: The lightweight OTFS end-to-end detection model outputs the soft decision probability of each bit in a single frame of OTFS signal through forward propagation. After hard decision by a preset threshold, the original bit stream of the transmitter is directly obtained, and the end-to-end detection of the OTFS signal is completed. No prior channel state information or independent channel estimation is required throughout the process.

[0105] Specifically, for each bit output by the model, a preset threshold of 0.5 is used for hard decision: a soft decision probability greater than 0.5 is determined as bit 1; a soft decision probability less than or equal to 0.5 is determined as bit 0. The original bit stream from the transmitter is directly obtained through hard decision, completing end-to-end detection of the OTFS signal. No prior channel state information is required throughout the process, and independent channel estimation, equalization, and symbol demodulation stages are unnecessary, significantly simplifying the receiver processing flow.

[0106] Example 2

[0107] This embodiment provides a lightweight OTFS end-to-end signal detection device for high-speed mobile scenarios. This device is a hardware implementation of the method described in Embodiment 1, applied to the receiving end of a vehicle-mounted OTFS communication terminal in a vehicle-to-everything (V2X) network, and adapted to high-speed mobile scenarios under the EVA extended vehicle channel model. The device includes a DD domain signal conversion module, a dual-channel input construction module, an end-to-end signal detection module, and a bit decision output module, which are connected in sequence. It also includes an offline training module that is connected in communication with the end-to-end signal detection module. The specific implementation of each module is as follows:

[0108] 1. DD Domain Signal Conversion Module: In this embodiment, the module has a built-in Wigner inverse transform and SFFT operation core, which is used to transform the time-domain received signal acquired by the OTFS receiver RF front end and output a 32×32 DD domain complex signal two-dimensional matrix; the module's input interface is connected to the output of the analog-to-digital converter of the receiver RF front end, and the output interface is connected to the dual-channel input construction module.

[0109] 2. Dual-channel input construction module: In this embodiment, the module has a built-in complex number decomposition and tensor splicing operation unit, which is used to decompose the two-dimensional matrix of the DD domain complex signal into a real part matrix and an imaginary part matrix with the same dimensions. The real part matrix is ​​used as the first channel and the imaginary part matrix is ​​used as the second channel to construct a 2×32×32 dual-channel input tensor and complete the normalization preprocessing of the tensor. The input interface of the module is connected to the DD domain signal conversion module, and the output interface is connected to the end-to-end signal detection module.

[0110] 3. End-to-end signal detection module: In this embodiment, the module has a pre-trained lightweight OTFS end-to-end detection model built in. The model structure is completely consistent with the model built in Embodiment 1. It is adapted to the feature extraction requirements in the EVA channel scenario and is used to receive dual-channel input tensors. It outputs the soft decision probability of each bit in a single frame of OTFS signal through model forward propagation. The module's input interface is connected to the dual-channel input construction module, and its output interface is connected to the bit decision output module. It is also connected to the offline training module through the data interaction interface.

[0111] 4. Bit Decision Output Module: In this embodiment, the module has a built-in hard decision operation unit and a bit stream output interface, which is used to make a hard decision on the soft decision probability with a preset threshold of 0.5, and directly output the original bit stream of the transmitter to the baseband processing unit of the vehicle terminal to complete the end-to-end detection of the OTFS signal.

[0112] 5. Offline Training Module: In this embodiment, this module is implemented using a host computer server. It has a built-in joint loss function optimization unit, AdamW optimizer, and learning rate scheduling unit, which are used for model loss calculation, parameter iterative update, and dynamic adjustment of the learning rate, respectively. Based on the dataset of the EVA extended vehicle channel model, it completes the offline training and weight update of the lightweight OTFS end-to-end detection model. The optimal weight model after training is burned into the storage unit of the end-to-end signal detection module through the bus.

[0113] The device in this embodiment does not require prior channel state information or independent channel estimation and equalization processing units. Its overall hardware architecture is simplified, with low computation and storage overhead. It can be directly integrated into the vehicle-mounted OTFS communication terminal, adapting to high-speed mobile scenarios of vehicle networking under the EVA channel model and meeting real-time communication requirements.

[0114] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A lightweight end-to-end signal detection method for high-speed mobile scenarios, applied to the receiving end of an OTFS communication system, characterized in that, Includes the following steps: S1: Transform the time-domain received signal acquired by the OTFS receiver to obtain a two-dimensional complex matrix of the DD domain signal; S2: The two-dimensional complex signal matrix in the DD domain is split into a real part matrix and an imaginary part matrix with the same dimension. The real part matrix is ​​used as the first channel and the imaginary part matrix is ​​used as the second channel to construct a dual-channel input tensor. S3: Input the dual-channel input tensor into the pre-trained lightweight OTFS end-to-end detection model. The lightweight OTFS end-to-end detection model is a deep learning model based on improved GhostNet and fused with CBAM. S4: The lightweight OTFS end-to-end detection model outputs the soft decision probability of each bit in a single frame of OTFS signal through forward propagation. After hard decision by a preset threshold, the original bit stream of the transmitter is directly obtained, completing the end-to-end detection of the OTFS signal. No prior channel state information or independent channel estimation is required throughout the process.

2. The lightweight OTFS end-to-end signal detection method for high-speed mobile scenarios according to claim 1, characterized in that, The transformation process described in step S1 is as follows: first, the time-domain received signal is mapped to the time-frequency domain through Wigner transformation, and then the time-frequency domain signal is converted into a two-dimensional complex signal matrix in the DD domain through SFFT. The two dimensions of the two-dimensional complex signal matrix in the DD domain correspond to the delay domain and the Doppler domain, respectively.

3. The lightweight OTFS end-to-end signal detection method for high-speed mobile scenarios according to claim 1, characterized in that, The dimension of the dual-channel input tensor in step S2 is 2×N×M, where N is the number of delay domain sampling points of the two-dimensional matrix of the complex signal in the DD domain, and M is the number of Doppler domain sampling points. The dual-channel input tensor fully preserves the two-dimensional joint features of the amplitude, phase, delay dimension and Doppler dimension of the OTFS DD domain signal.

4. The lightweight OTFS end-to-end signal detection method for high-speed mobile scenarios according to claim 1, characterized in that, The lightweight OTFS end-to-end detection model, along the signal transmission direction, includes, in sequence, an input adaptation layer, a multi-level Ghost bottleneck feature extraction layer, a CBAM attention enhancement module, a feature aggregation layer, and a bit mapping output layer. The input adaptation layer uses a 3×3 standard convolution to perform preliminary feature fusion and extraction on the dual-channel input tensor, mapping the dual-channel features to multi-channel basic features. The multi-stage Ghost bottleneck feature extraction layer is composed of multiple Ghost bottlenecks stacked together. The Ghost bottleneck retains the lightweight core structure of Ghost module main convolution generating inherent features + depth convolution generating Ghost features. The convolution kernel size, channel expansion ratio and number of stacked layers are optimized for OTFS DD domain two-dimensional signal features. The CBAM attention enhancement module is embedded in the output of the mid-to-high-level Ghost bottleneck and is used to perform dual recalibration of features in both channel and spatial dimensions. The feature aggregation layer employs global average pooling to reduce the dimensionality and aggregate deep features. The bit-mapping output layer is used to map the aggregated features to a dimension that matches the total number of bits in a single-frame OTFS signal, and outputs the soft decision probability for each bit.

5. The lightweight OTFS end-to-end signal detection method for high-speed mobile scenarios according to claim 4, characterized in that, The CBAM attention enhancement module is embedded only after the mid-to-high-level Ghost bottleneck where the number of feature channels reaches a preset threshold. Its fusion logic is as follows: the features output by the Ghost bottleneck are processed by batch normalization and activation functions, and then sequentially input into the channel attention unit and spatial attention unit of CBAM to complete the feature recalibration of the channel dimension and the delay-Doppler spatial dimension, respectively, and finally output the enhanced features to the next layer of the network; the channel attention unit performs global average pooling and global max pooling on the input features, and after passing through a shared fully connected layer and Sigmoid activation, it obtains the channel attention weights and completes the channel dimension recalibration; The spatial attention unit performs channel-dimensional average pooling and max pooling on the channel-recalibrated features. The pooling results are concatenated and then passed through a convolutional layer and Sigmoid activation to obtain spatial attention weights, thus completing the spatial dimension recalibration.

6. The lightweight OTFS end-to-end signal detection method for high-speed mobile scenarios according to claim 4, characterized in that, The bit-mapped output layer includes a fully connected layer and a Sigmoid activation layer. The output dimension of the fully connected layer is consistent with the total number of bits in a single frame of the OTFS signal. The Sigmoid activation layer maps the output value to the [0,1] interval to obtain the soft decision probability of each bit being 1. The preset threshold for the hard decision is 0.

5. A soft decision probability greater than 0.5 is determined as bit 1, and a soft decision probability less than or equal to 0.5 is determined as bit 0.

7. The lightweight OTFS end-to-end signal detection method for high-speed mobile scenarios according to claim 1, characterized in that, The lightweight OTFS end-to-end detection model is optimized during the offline training phase using a joint loss function that combines the main loss and auxiliary loss. The expression for the joint loss function is as follows: in, The main loss is a weighted binary cross-entropy loss, which corresponds to the core optimization objective of bit-level detection. As an auxiliary loss, symbol-level mean squared error loss is used to help the model learn the constellation point mapping relationship of OTFS modulation symbols; These are auxiliary loss weighting coefficients used to balance the optimization priorities of the main loss and auxiliary loss.

8. The lightweight OTFS end-to-end signal detection method for high-speed mobile scenarios according to claim 7, characterized in that, In the offline training phase of the lightweight OTFS end-to-end detection model, the AdamW optimizer is used for parameter iterative updates. The training process adopts a cosine annealing learning rate scheduling strategy with a preheating mechanism, and regularization is performed through dropout layer and weight decay mechanism to avoid model overfitting.

9. A lightweight OTFS end-to-end signal detection device for high-speed mobile scenarios, characterized in that, It includes a DD domain signal conversion module, a dual-channel input construction module, an end-to-end signal detection module, and a bit decision output module that are connected in sequence. The DD domain signal conversion module is used to transform the time-domain received signal acquired by the OTFS receiver and output a two-dimensional matrix of DD domain complex signals. The dual-channel input construction module is used to split the two-dimensional matrix of the complex signal in the DD domain into a real part matrix and an imaginary part matrix with the same dimensions, and construct a dual-channel input tensor with the real part matrix as the first channel and the imaginary part matrix as the second channel. The end-to-end signal detection module has a built-in pre-trained lightweight OTFS end-to-end detection model. The model is a deep learning model based on an improved GhostNet and fused with the CBAM attention mechanism. It is used to receive dual-channel input tensors and output the soft decision probability of each bit in a single frame of OTFS signal. The bit decision output module is used to make a hard decision based on a preset threshold of the soft decision probability, and directly output the original bit stream from the transmitter to complete the end-to-end detection of the OTFS signal.

10. The lightweight OTFS end-to-end signal detection device for high-speed mobile scenarios according to claim 9, characterized in that, It also includes an offline training module, which is communicatively connected to the end-to-end signal detection module and is used to complete the offline training and weight update of the lightweight OTFS end-to-end detection model. The offline training module has a built-in joint loss function optimization unit, AdamW optimizer and learning rate scheduling unit, which are used for model loss calculation, parameter iterative update and dynamic adjustment of learning rate, respectively.

Images