Communication and perception integrated model construction method and communication and perception integrated model

Abstract

The invention provides a construction method of a communication perception integrated model and the communication perception integrated model, and relates to the field of communication, and the method comprises the steps: employing a set modulation mode to transmit a plurality of OFDM frames between a transmitter and a receiver, and obtaining a frequency domain OFDM signal segment corresponding to each OFDM frame through the receiver, and each OFDM frame comprises a data subcarrier and a pilot frequency subcarrier; each sample instance comprises a frequency domain OFDM signal segment, and a real symbol and a real distance corresponding to the frequency domain OFDM signal segment; the communication and perception integrated model comprises a communication network architecture and a perception network architecture which share the same complex value residual time sequence convolutional network; the communication network architecture carries out equalization processing based on the frequency domain OFDM signal segment of the input communication and perception integrated model so as to output a prediction symbol, and the perception network architecture carries out distance estimation based on the frequency domain OFDM signal segment of the input communication and perception integrated model so as to output a prediction distance between the transmitter and the receiver. Therefore, joint nonlinear equalization and transmitting-receiving end distance estimation are realized.

Description

Technical Field

[0001] This disclosure relates to the field of communication technology, specifically to a method for constructing an integrated communication and sensing model and the integrated communication and sensing model itself. Background Technology

[0002] In terahertz systems, free-space path loss increases quadratically with frequency, resulting in attenuation far exceeding that of microwave systems. Furthermore, practical terahertz transceivers suffer from various hardware limitations. Key front-end components such as low-noise amplifiers, subharmonic mixers, and frequency multipliers often exhibit non-ideal frequency responses due to gain ripple, conversion loss variations, and bandwidth-dependent impedance mismatch. These effects lead to amplitude and phase distortion in the subcarrier at the frequency margins. More importantly, frequency up-conversion and down-conversion introduce significant phase noise (PN), which accumulates rapidly with increasing multiplication order. Designing robust terahertz ISAC (communication-sensing integrated) systems requires a comprehensive solution to these dual challenges at both the hardware and propagation levels.

[0003] Traditional ISAC frameworks, originally developed for microwave or millimeter-wave bands, struggle to address the unique challenges of the terahertz band. Particularly in signal processing, these frameworks often overlook severe nonlinear effects, hardware limitations, and the impact of dynamic industrial environments. While Orthogonal Frequency Division Multiplexing (OFDM) has become the preferred solution due to its spectral efficiency and resistance to multipath fading, the deployment of terahertz ISAC systems in smart industrial scenarios still faces numerous limitations. Dense obstacles and reflective surfaces lead to strong frequency-selective fading, and the use of directional antennas cannot fully compensate for the rapid attenuation of the signal-to-noise ratio (SNR) during transmission. Furthermore, industrial applications typically involve diverse communication and sensing tasks, requiring flexible modulation schemes ranging from robust low-order quadrature amplitude modulation (QAM) to high-throughput high-order QAM formats. These needs underscore the necessity of constructing a unified signal processing framework—one capable of achieving robust equilibrium and accurate sensing in heterogeneous environments.

[0004] In recent years, Temporal Convolutional Networks (TCNs) have emerged as an alternative to Recurrent Neural Network (RNN)-based equalizers. Originally developed for sequence modeling, this technique achieves parallel computation through dilated convolutions while still capturing long-range dependencies. Compared to traditional RNNs and their variants, TCNs avoid recursive gradient propagation, significantly improving training efficiency and stability. When applied in the frequency domain, TCN-based equalizers can utilize the structural characteristics of OFDM signals to model distortion patterns between subcarriers. Despite these advantages, most existing TCN equalizers are designed for real-valued inputs and lack mechanisms to handle the complex distortions unique to terahertz systems. Particularly noteworthy is the scarcity of research exploring TCNs with recurrent training or employing residual learning schemes specifically tailored to the nonlinear characteristics of terahertz systems, leaving ample room for research in this field.

[0005] Although recent studies have proposed ISAC neural waveform design schemes based on DFT-s-OFDM embedded sensing features, these schemes are often limited to specific waveform designs and cannot achieve joint reasoning for communication and sensing tasks. Summary of the Invention

[0006] The main purpose of this disclosure is to provide a method for constructing a communication-sensing integrated model and a communication-sensing integrated model, so as to achieve joint nonlinear equalization and transmitter-receiver distance estimation.

[0007] To achieve the above objectives, the first aspect of this disclosure provides a method for constructing a communication-aware integrated model based on a complex-valued residual temporal convolutional network, the method comprising: The spatial distance between the transmitter and receiver of the terahertz communication system is adjusted sequentially. After each adjustment, multiple OFDM frames are transmitted between the transmitter and receiver using a set modulation scheme. The receiver obtains the frequency domain OFDM signal segment corresponding to each OFDM frame. Each OFDM frame contains a data subcarrier and a pilot subcarrier. Construct a training sample set containing multiple sample instances, each sample instance containing a frequency domain OFDM signal segment, as well as the corresponding real symbol and real distance of the frequency domain OFDM signal segment; A communication-sensing integrated model to be trained is constructed. The communication-sensing integrated model includes a communication network architecture and a sensing network architecture that share the same complex-valued residual temporal convolutional network. The input of the communication-sensing integrated model is a frequency domain OFDM signal segment. The communication network architecture performs equalization processing on the frequency domain OFDM signal segment input to the communication-sensing integrated model to output predicted symbols. The sensing network architecture performs distance estimation on the frequency domain OFDM signal segment input to the communication-sensing integrated model to output the predicted distance between the transmitter and the receiver. The communication and sensing integrated model to be trained is trained using the training sample set to obtain the ideal communication and sensing integrated model.

[0008] In some embodiments of this disclosure, the communication network architecture includes: The communication feature acquisition module is used to acquire the communication features of the frequency domain OFDM signal segment of the input communication sensing integrated model; Among them, the complex-valued residual temporal convolutional network is used to process the communication features output by the communication feature acquisition module, so that the communication network architecture is based on the predicted symbol obtained after equalization processing of the processing result output of the complex-valued residual temporal convolutional network.

[0009] In some embodiments of this disclosure, the perception network architecture includes: The sensing feature acquisition module is used to acquire the sensing features of the frequency domain OFDM signal segment of the input communication sensing integrated model; Among them, the complex-valued residual temporal convolutional network is used to process the perceptual features output by the perceptual feature acquisition module, so that the perceptual network architecture outputs the predicted distance between the transmitter and the receiver based on the processing result of the complex-valued residual temporal convolutional network.

[0010] In some embodiments of this disclosure, the communication feature acquisition module includes a communication encoder and a communication attention module; wherein, the communication encoder is used to extract communication features based on the frequency domain OFDM signal segment of the input communication sensing integrated model to obtain a first communication feature, and the communication attention module is used to calculate the global average pooling activation value of the real part of the first communication feature using complex convolution and batch normalization algorithms to obtain the communication attention weight of the first communication feature; The perceptual feature acquisition module includes a perceptual encoder and a perceptual attention module. The perceptual encoder is used to extract perceptual features based on the frequency domain OFDM signal segment of the input communication and perception integrated model to obtain the first perceptual feature. The perceptual attention module is used to calculate the global average pooling activation value of the real part of the first perceptual feature using complex value convolution and batch normalization algorithms to obtain the perceptual attention weight of the first perceptual feature. The integrated communication and perception model also includes a cross-attention module, which is used to: process the first communication feature based on communication adaptive adjustment of attention weight and perception attention weight to obtain the second communication feature, and use the second communication feature as the communication feature output by the communication feature acquisition module; and process the first perception feature based on perception adaptive adjustment of attention weight and communication attention weight to obtain the second perception feature, and use the second perception feature as the perception feature output by the perception feature acquisition module.

[0011] In some embodiments of this disclosure, the magnitudes of the communication adaptive adjustment attention weight and the perception adaptive adjustment attention weight are dynamically and adaptively adjusted based on the predicted distance between the transmitter and receiver, the channel signal-to-noise ratio status information between the transmitter and receiver, the predicted phase structure information and frequency information of the frequency domain OFDM signal segment predicted by the integrated communication and perception model during operation.

[0012] In some embodiments of this disclosure, the integrated communication-aware model includes a supervision module, which includes: The communication loss calculation module is used to calculate the communication loss value between the predicted symbols output by the communication network architecture and the actual symbols. ; The perception loss calculation module is used to calculate the perception loss value between the predicted distance output by the perception network architecture and the true distance. ; The overall loss calculation module is used to calculate losses based on set coefficients. The overall loss of the integrated communication and sensing model is calculated using the following formula. :

[0013] Furthermore, during the training process of the integrated communication and sensing model using the training sample set, if the overall loss... If the loss is less than the set loss threshold, then the training of the communication-sensing integrated model to be trained is completed, and the trained communication-sensing integrated model is taken as the ideal communication-sensing integrated model.

[0014] In some embodiments of this disclosure, the communication loss calculation module uses the mean square error loss function to calculate the communication loss value between the predicted symbol and the true symbol output by the communication network architecture. ; The perception loss calculation module uses the Huber loss function to calculate the perception loss value between the predicted distance output by the perception network architecture and the true distance. .

[0015] In some embodiments of this disclosure, a complex-valued residual temporal convolutional network is used to process the communication features output by the communication feature acquisition module and the perception features output by the perception feature acquisition module, and output fused features. The communication network architecture also includes: a communication prediction output module, used to process based on fused features and output predicted symbols; The perception network architecture also includes a distance prediction output module, which processes the fused features and outputs the predicted distance between the transmitter and the receiver.

[0016] In some embodiments of this disclosure, the modulation method is set to include QPSK modulation and 16QAM modulation; The training sample set includes: a first training sample set obtained using QPSK modulation after each adjustment of the spatial spacing between the transmitter and receiver based on the same terahertz communication system, and a second training sample set obtained using 16QAM modulation.

[0017] The second aspect of this disclosure provides a communication sensing integrated model based on a complex-valued residual temporal convolutional network, which is constructed using any of the construction methods for a communication sensing integrated model based on a complex-valued residual temporal convolutional network shown in the first aspect of this disclosure.

[0018] The terahertz wireless communication system provided in this embodiment adjusts the spatial distance between the transmitter and receiver of the terahertz communication system sequentially. Each adjustment involves acquiring the frequency-domain OFDM signal segment corresponding to each OFDM frame via the receiver. Each OFDM frame contains data subcarriers and pilot subcarriers, ensuring that each sample instance in the constructed training sample set contains the frequency-domain OFDM signal segment, as well as the corresponding real symbol and real distance. The integrated communication-sensing model includes a communication network architecture and a sensing network architecture sharing the same complex-valued residual temporal convolutional network. The input to the integrated communication-sensing model is the frequency-domain OFDM signal segment. The communication network architecture performs equalization processing based on the input frequency-domain OFDM signal segment of the integrated communication-sensing model to output predicted symbols. The sensing network architecture performs distance estimation based on the input frequency-domain OFDM signal segment of the integrated communication-sensing model to output the predicted distance between the transmitter and receiver. Then, the training sample set is used to train the integrated communication-sensing model to obtain an ideal integrated communication-sensing model, enabling joint nonlinear equalization and transceiver distance estimation. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the specific embodiments or related technologies of this disclosure, the accompanying drawings used in the description of the specific embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating a method for constructing a communication-sensing integrated model based on a complex-valued residual temporal convolutional network, as provided in an embodiment of this disclosure; Figure 2 A schematic block diagram of a terahertz ISAC overall system provided in an embodiment of this disclosure; Figure 3 A digital signal processing flowchart for generating OFDM signals is provided as an embodiment of this disclosure; Figure 4 This is a transmitter radio frequency flowchart provided in one embodiment of the present disclosure; Figure 5 This is a receiver radio frequency flowchart provided in one embodiment of the present disclosure; Figure 6 A digital signal processing flowchart for a receiver provided in an embodiment of this disclosure; Figure 7 A schematic diagram of the architecture of a communication-aware integrated model provided in an embodiment of this disclosure; Figure 8 This is a block diagram of a communication network architecture provided in an embodiment of the present disclosure; Figure 9 This is a block diagram of a perception network architecture provided in an embodiment of the present disclosure; Figure 10 A schematic diagram illustrating how the BER performance of 8-Gbaud OFDM-QPSK transmission varies with distance when using different methods, according to an embodiment of this disclosure; Figure 11 A schematic diagram illustrating the variation of BER performance of 16-Gbaud OFDM-QPSK transmission with distance when using different methods, as provided in an embodiment of this disclosure; Figure 12 A schematic diagram illustrating the variation of BER performance of 4G-baud OFDM-16QAM transmission with distance when using different methods, as provided in an embodiment of this disclosure; Figure 13 This is a comparative schematic diagram showing the range estimation accuracy of different methods provided for an embodiment of the present disclosure.

[0021] It should be noted that the elements in the attached diagram are schematic and not drawn to scale. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present disclosure, the technical solutions of the present disclosure will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present disclosure, and not all embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present disclosure.

[0023] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] In this disclosure, the terms “upper,” “lower,” “left,” “right,” “front,” “rear,” “top,” “bottom,” “inner,” “outer,” and “middle,” etc., indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings. These terms are primarily for the purpose of better describing this disclosure and its embodiments, and are not intended to limit the indicated devices, elements, or components to having a specific orientation, or to be constructed and operated in a specific orientation.

[0025] Furthermore, in addition to indicating location or positional relationship, some of the aforementioned terms may also have other meanings. For example, the term "above" may also be used in certain circumstances to indicate a dependency or connection. Those skilled in the art can understand the specific meaning of these terms in this disclosure according to the specific circumstances.

[0026] Furthermore, the terms "set up," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral structure; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection via an intermediate medium, or an internal connection between two devices, components, or parts. Those skilled in the art can understand the specific meaning of these terms in this disclosure according to the specific circumstances.

[0027] It should be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other. This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0028] As described in the background section of this disclosure, the disadvantages of the related technologies are as follows: (1) Hardware deployment is difficult and terminal devices are diverse; (2) Most existing TCN equalizers are designed for real-valued inputs and lack a mechanism to handle the composite distortion unique to terahertz systems; (3) There is no residual learning scheme specifically for the nonlinear characteristics of terahertz; (4) The network model has high complexity, large parameter overhead, and poor stability, making it less applicable to resource-constrained terahertz systems; (5) The related methods lack key features such as passive sensing capability, cross-modulation robustness, and compact single-input single-output (SISO) architecture, which are the core elements for industrial-grade ISAC deployment to achieve large-scale application.

[0029] To address the challenges of hardware impairment, diverse terminal device requirements, and stringent sensing constraints faced by terahertz ISAC systems in smart industrial deployments, the inventors of this publication have developed a unified neural receiver architecture based on complex-valued residual temporal convolutional networks (CV Res-TCN) as an integrated communication and sensing model. This integrated model can simultaneously achieve robust equalization, adaptive modulation, and passive sensing functions. Unlike traditional models that suffer from limited processing order and unstable training, the TCN architecture possesses advantages such as parallel computing capabilities, stability, and effective long-distance modeling. By extending TCN to the complex domain and introducing residual connection technology, this architecture achieves high fidelity in symbol recovery, exhibits excellent performance in cross-modulation generalization, and can accurately infer distance from OFDM pilot structures—all these features are integrated into a lightweight shared design, perfectly adaptable to heterogeneous industrial terminal devices.

[0030] refer to Figure 1 This disclosure provides a method for constructing a communication-sensing integrated model based on a complex-valued residual temporal convolutional network, aiming to achieve joint nonlinear equalization and transmitter-receiver distance estimation. The construction method mainly includes the following steps: In S110, the spatial distance between the transmitter and receiver of the terahertz communication system is adjusted sequentially. After each adjustment, multiple OFDM frames are transmitted between the transmitter and receiver using a set modulation method. The receiver obtains the frequency domain OFDM signal segment corresponding to each OFDM frame. Each OFDM frame contains a data subcarrier and a pilot subcarrier. In S120, a training sample set containing multiple sample instances is constructed. Each sample instance contains a frequency domain OFDM signal segment, as well as the real symbol and real distance corresponding to the frequency domain OFDM signal segment. In S130, a communication-sensing integrated model to be trained is constructed. The communication-sensing integrated model includes a communication network architecture and a sensing network architecture that share the same complex-valued residual temporal convolutional network. The input of the communication-sensing integrated model is a frequency domain OFDM signal segment. The communication network architecture performs equalization processing based on the frequency domain OFDM signal segment input to the communication-sensing integrated model to output predicted symbols. The sensing network architecture performs distance estimation based on the frequency domain OFDM signal segment input to the communication-sensing integrated model to output the predicted distance between the transmitter and the receiver. In S140, the training sample set is used to train the communication and sensing integrated model to be trained, so as to obtain the ideal communication and sensing integrated model.

[0031] In the above scheme, the spatial distance between the transmitter and receiver of the terahertz communication system is adjusted sequentially. Each adjustment involves acquiring the frequency-domain OFDM signal segment corresponding to each OFDM frame via the receiver. Each OFDM frame contains data subcarriers and pilot subcarriers, ensuring that each sample instance in the constructed training sample set contains the frequency-domain OFDM signal segment, as well as the corresponding real symbol and real distance. The integrated communication-sensing model includes a communication network architecture and a sensing network architecture sharing the same complex-valued residual temporal convolutional network. The input to the integrated communication-sensing model is the frequency-domain OFDM signal segment. The communication network architecture performs equalization processing based on the input frequency-domain OFDM signal segment of the integrated communication-sensing model to output predicted symbols. The sensing network architecture performs distance estimation based on the input frequency-domain OFDM signal segment of the integrated communication-sensing model to output the predicted distance between the transmitter and receiver. Then, the training sample set is used to train the integrated communication-sensing model to obtain an ideal integrated communication-sensing model, enabling joint nonlinear equalization and transceiver distance estimation.

[0032] The following section provides a detailed explanation of each of the above steps in conjunction with the accompanying drawings.

[0033] There are several ways to configure a terahertz communication system. For example, refer to... Figure 2 The terahertz ISAC hardware and software hybrid test platform shown can be used as a terahertz communication system in the process of building an integrated communication and sensing model according to the embodiments of this disclosure. It mainly includes two modules: a transmitter and a receiver. This terahertz communication system simulates a typical short-range, high-speed terahertz link between edge terminals in a smart factory, simultaneously supporting communication and sensing functions under actual hardware limitations.

[0034] At the transmitter end, the transmitter is divided into two main parts: transmitter digital signal processing and transmitter radio frequency (RF). The generation of OFDM signals is as follows: Figure 3The digital signal processing flow shown is illustrated, and the structure of the transmitter RF module is as follows: Figure 4 As shown. Exemplarily, the transmitter can first use a pseudo-random bit sequence as the original data source, and then modulate the bit stream used as the original data source to the target modulation scheme. Synchronization and passive sensing functions are achieved by embedding a zero-correlation sequence (ZC) into a predefined pilot subcarrier. After serial-to-parallel conversion, inverse fast Fourier transform, and cyclic prefix (CP) insertion, the signal is digitally up-converted to generate a 10 GHz intermediate frequency signal, which is then transmitted to a 64 GSa / s arbitrary waveform generator. This arbitrary waveform generator performs digital-to-analog conversion, outputting a 10 GHz real-valued analog intermediate frequency signal. This analog intermediate frequency signal is then up-converted to the terahertz band by a subharmonic mixer operating at 170-260 GHz. A 17.5 GHz local oscillator drives a ×6 active frequency multiplier to generate a 105 GHz local oscillator signal, which is mixed with the intermediate frequency signal to form a broadband terahertz signal. The generated terahertz signal has a maximum bandwidth of 16 Gbaud and a center frequency of 220 GHz (i.e., 220 ± 8 GHz). The terahertz signal is amplified by a low-noise amplifier with a working frequency of 195-230 GHz and a gain of 25 dB, and then transmitted over short distances in the air using a horn antenna with a working frequency of 170-260 GHz and a gain of 25 dBi.

[0035] The receiver employs a symmetrical structural design, and the receiver RF module structure is as follows: Figure 5 As shown. Exemplarily, the incident terahertz signal is acquired by the exact same antenna, amplified by a low-noise amplifier with a gain of 25dB, and down-converted to a 10 GHz intermediate frequency signal using the same subharmonic mixer and local oscillator configuration as the transmitter. The signal is then amplified a second time using an intermediate frequency amplifier with a gain of 35dB, and finally digitized using an 80 GSa / s real-time oscilloscope with sufficient analog bandwidth, thus fully capturing the broadband waveform signal.

[0036] The digitized interference samples are input into the receiver-side DSP link implemented by mathematical software, such as... Figure 6 As shown, the processing flow includes down-conversion, resampling, synchronization calibration, CP cancellation, frequency offset estimation, fast Fourier transform, phase compensation, and channel estimation. It is worth noting that in practical terahertz transceivers, hardware devices introduce significant frequency response non-uniformity, thus affecting communication and sensing performance. These device-level non-ideal characteristics generate amplitude and phase distortion on OFDM subcarriers, thereby reducing demodulation fidelity and affecting pilot-based sensing reliability. To address this issue, the CV Res-TCN technology proposed in this disclosure is applied to all received frequency domain subcarrier sequences, achieving nonlinear equalization and transceiver distance estimation.

[0037] For example, the communication-sensing integrated model constructed in this embodiment is based on a unified CV Res-TCN network, which operates on the receiver's frequency domain subcarrier sequence. The CV Res-TCN network adopts a shared encoder structure and is equipped with a task-specific output branch. The overall network includes two main functions: the communication network functions inherent in the communication network architecture and the sensing network functions inherent in the sensing network architecture, simultaneously supporting communication and sensing tasks. The communication network architecture and the sensing network architecture are inherently related because they both operate on the same received terahertz OFDM signal (frequency domain OFDM signal segment) and share underlying signal impairment features. To fully utilize these inherent connections while minimizing redundant computation, this embodiment proposes a unified multi-task CV Res-TCN framework as the communication-sensing integrated model. This model extracts common feature representations through a shared backbone network and integrates task-specific head-layer networks for both communication and sensing objectives. Unlike traditional task-isolated architectures, this communication-sensing integrated model achieves its advantages through parameter reuse, joint optimization, and task-aware information interaction.

[0038] There are several ways to set up a communication network architecture, some of which are illustrated below.

[0039] For example, the communication network architecture may include: a communication feature acquisition module, which is used to acquire the communication features of the frequency domain OFDM signal segment of the input communication sensing integrated model; wherein, a complex-valued residual temporal convolutional network is used to process the communication features output by the communication feature acquisition module, so that the communication network architecture obtains the predicted symbol after equalization processing based on the processing result output of the complex-valued residual temporal convolutional network.

[0040] For example, the perception network architecture may include: a perception feature acquisition module, which is used to acquire the perception features of the frequency domain OFDM signal segment of the input communication perception integrated model; wherein, a complex-valued residual temporal convolutional network is used to process the perception features output by the perception feature acquisition module, so that the perception network architecture outputs the predicted distance between the transmitter and the receiver based on the processing result of the complex-valued residual temporal convolutional network.

[0041] There are several ways to configure the communication feature acquisition module, and some of these methods are illustrated below.

[0042] For example, refer to Figure 7The communication feature acquisition module includes a communication encoder and a communication attention module. The communication encoder is used to extract the first communication feature based on the frequency domain OFDM signal segment of the input communication sensing integrated model. The communication attention module is used to calculate the global average pooling activation value of the real part of the first communication feature using complex convolution and complex batch normalization algorithms to obtain the communication attention weight of the first communication feature.

[0043] There are several ways to configure the perceptual feature acquisition module. Some of these methods are illustrated below.

[0044] For example, refer to Figure 7 The perceptual feature acquisition module includes a perceptual encoder and a perceptual attention module. The perceptual encoder is used to extract perceptual features based on the frequency domain OFDM signal segment of the input communication and perception integrated model to obtain the first perceptual feature. The perceptual attention module is used to calculate the global average pooling activation value of the real part of the first perceptual feature using complex value convolution and complex value batch normalization algorithms to obtain the perceptual attention weight of the first perceptual feature.

[0045] like Figure 7 As shown, this embodiment of the disclosure designs a dual-branch encoder to handle heterogeneous inputs across tasks: the communication encoder receives flattened frequency-domain OFDM symbols processed by fast Fourier transform, phase correction, and pilot equalization; the sensing encoder extracts channel state information (CSI) from the pilot positions. Each branch employs complex-valued convolution and complex-valued batch normalization modules to generate task-specific feature maps. (1) (2) in, Indicates communication input characteristics The extraction results (representing the first communication feature or the first communication feature map). Representing perceived input features The extraction results (representing the first perceptual feature or the first perceptual feature map).

[0046] For example, refer to Figure 7To achieve cross-task collaboration, one embodiment of this disclosure proposes a lightweight cross-task cross-attention module. This module enhances features by selectively injecting importance weights from one task to another. The cross-attention module is used to: process a first communication feature based on communication adaptive adjustment of attention weights and perception attention weights to obtain a second communication feature, and use the second communication feature as the communication feature output by the communication feature acquisition module; and to process a first perception feature based on perception adaptive adjustment of attention weights and communication attention weights to obtain a second perception feature, and use the second perception feature as the perception feature output by the perception feature acquisition module.

[0047] It should be noted that the values ​​of the communication adaptive adjustment attention weight and the perception adaptive adjustment attention weight can be dynamically preset and adjusted. For example, in some embodiments, the magnitudes of the communication adaptive adjustment attention weight and the perception adaptive adjustment attention weight can be dynamically adaptively adjusted based on the predicted distance between the transmitter and receiver, the channel signal-to-noise ratio status information between the transmitter and receiver, or the predicted phase structure information and frequency information of the frequency domain OFDM signal segment predicted by the integrated communication and perception model during operation.

[0048] For example, one embodiment of this disclosure first generates task-specific communication attention weights by calculating the global average pooling activation value of the real part of each feature map. and perceived attention weight : (3) (4) in This represents the Sigmoid function. This represents global average pooling over the time dimension.

[0049] Subsequently, reference Figure 7 By introducing a cross-attention module to execute a cross-attention mechanism to adjust the feature response, the following second communication feature is obtained. Second sensory features : (5) (6) in, This indicates that communication adaptively adjusts attention weights. This indicates that attention weights are adjusted adaptively based on perception. and These represent learnable scalars for different tasks, allowing for adaptive adjustment of attention weights. This design avoids direct feature fusion, maintaining task independence while allowing each task to learn useful structural information from another. The communication task can learn long-range CSI structures at low SNR from the perception task; the perception task can utilize the frequency domain structure and phase patterns learned from the communication task to make distance regression more accurate.

[0050] For example, refer to Figure 7 Optimized second communication feature Second sensory features The input is then used to train a shared CV Res-TCN backbone network (the same complex-valued residual temporal convolutional network) and a multi-task decoder to optimize the parameters of the CV Res-TCN backbone network, and based on the second communication features. Second sensory features Output fused features.

[0051] For example, the integrated communication and sensing model may further include a supervision module, which may include a communication loss calculation module, a sensing loss calculation module, and an overall loss calculation module. The communication loss calculation module is used to calculate the communication loss value between the predicted symbols output by the communication network architecture and the true symbols. The perception loss calculation module is used to calculate the perception loss value between the predicted distance output by the perception network architecture and the true distance. The overall loss calculation module is used to calculate losses based on set coefficients. The overall loss of the integrated communication and sensing model is calculated using the following formula. :

[0052] For example, the overall multi-task loss can be expressed as: (7) in This represents a hyperparameter used to balance communication reliability and sensing accuracy. This represents the mean square error between the predicted symbol output by the equalization process and the true constellation symbol, while This represents the mean square error between the predicted distance and the actual distance.

[0053] During the training process of the integrated communication and sensing model to be trained using the training sample set, if the overall loss... If the loss is less than the set loss threshold, the training of the communication-sensing integrated model to be trained is considered complete, and the trained communication-sensing integrated model is taken as the ideal communication-sensing integrated model. This ensures that the parameter settings of the ideal communication-sensing integrated model obtained after training are more suitable and can meet the application requirements.

[0054] The communication loss calculation module calculates the communication loss value between the predicted symbols and the true symbols output by the communication network architecture. There are various methods that can be used. For example, in some embodiments, the communication loss calculation module uses the mean squared error loss function to calculate the communication loss value between the predicted symbol and the true symbol output by the communication network architecture. This is to enhance robustness to outliers.

[0055] The perceptual loss calculation module calculates the perceptual loss value between the predicted distance output by the perceptual network architecture and the true distance. There are various methods that can be used. For example, in some embodiments, the perception loss calculation module uses the Huber loss function to calculate the perception loss value between the predicted distance output by the perception network architecture and the true distance. This is to enhance robustness to outliers.

[0056] For example, the complex-valued residual temporal convolutional network can also be used to process the communication features output by the communication feature acquisition module and the perceptual features output by the perceptual feature acquisition module, and output fused features. In this case, the communication network architecture can further include a communication prediction output module, which processes the fused features and outputs predicted symbols. The perceptual network architecture can further include a distance prediction output module, which processes the fused features and outputs the predicted distance between the transmitter and receiver. Through this approach, the advantages of the communication network architecture and the perceptual network architecture sharing the same complex-valued residual temporal convolutional network—facilitating parameter sharing, mutual compensation, and correction—can be leveraged to obtain more accurate fused features, making the final output predicted symbols and predicted distances closer to the true values.

[0057] For example, refer to Figure 8 The communication network architecture can include four components: an input sliding window buffer, an initial preprocessing layer, stacked Res-TCN modules, and a final communication branch (communication prediction output module). The overall steps of the communication network architecture are as follows: The input is a complex sequence of length 2H+1, centered on the subcarrier, represented as a symmetric sliding window: (8) in Indicates subcarrier The received frequency domain symbols. This window can capture inter-carrier interference, frequency domain correlation, and residual phase drift.

[0058] The input sequence first passes through an initial preprocessing layer, which consists of a complex-valued convolutional layer and a complex-valued batch normalization (CBN) layer. The complex-valued convolutional layer projects the original subcarriers into a latent modulation-aware feature space: (9) This layer extracts spectral features while preserving inherent phase information. A CBN layer is then applied to stabilize the training process under different link conditions.

[0059] The output signal is then passed to stacked Res-TCN modules. Each Res-TCN module contains dilated convolutional layers, complex-valued normalization layers, CReLU layers, and non-normalized complex-valued convolutional layers, thereby enabling hierarchical feature learning among subcarriers. Each Res-TCN module can be represented as: (10) in This indicates complex-valued activation. Each complex-valued convolution mixes the real and imaginary parts in the following way: (11) Meanwhile, residual connections enhance gradient flow and ensure the stability of deep training: (12) Finally, a post-convolutional layer aggregates the local context around the central subcarrier and predicts the equalized complex-valued symbol. This structure enables the network to learn conjugate symmetric features caused by IQ imbalance, phase distortion, and spectral regeneration. Dilated convolutions help model long-term phase drift and power amplifier distortion with memory effects. This communication-aware integrated model network is trained to minimize the mean square error between predicted and true symbols. (13) Therefore, the communication-aware integrated model network implicitly learns nonlinear inverse mappings: (14) in , , These represent phase noise, power amplifier nonlinearity, and terahertz channel fading, respectively.

[0060] For example, a schematic block diagram of a perception network architecture is shown below. Figure 9 As shown, it includes pilot extraction, complex-valued convolutional layer modules, Res-TCN modules, and a sensing branch (sensing prediction output module). After pilot extraction, the channel estimation matrix is ​​sent to the sensing network as input, and its expression is: (15) in This represents the number of pilot subcarriers in each OFDM symbol, where T is the number of time-coordinated pilot symbols. Each input... It is reshaped into a two-dimensional complex tensor and processed through a series of operations, which is the same as the architecture of a communication network.

[0061] Complex-valued convolutional layers capture the joint amplitude and phase patterns between pilot subcarriers and implicitly learn cross-subcarrier phase gradients through adaptive nonlinear filters, similar in principle to delay estimation methods based on Fast Fourier Transform. The Res-TCN module employs dilated convolutional sequences with skip connections to capture phase drift patterns and simulate the structured evolution of frequency-domain pilots in OFDM symbols. The final output is sent to a perceptual branch containing a complex-valued convolutional layer, a global average pooling layer, a flattening layer, and a linear layer. This branch maps the extracted high-dimensional features to the final distance estimate. .

[0062] Regarding the type of modulation scheme mentioned above, various types can be used. For example, the modulation scheme can include QPSK modulation and 16QAM modulation. In this case, the training sample set can include: a first training sample set obtained using QPSK modulation after each adjustment of the spatial spacing between the transmitter and receiver based on the same terahertz communication system, and a second training sample set obtained using 16QAM modulation. Through this method, the final communication-sensing integrated model based on complex-valued residual temporal convolutional networks can be adapted to application scenarios using both QPSK and 16QAM modulation.

[0063] During the training process of the Communication-Sensing Integrated Model (CV Res-TCN Network Model), to simulate the actual deployment communication conditions in a smart industrial environment, some embodiments of this disclosure employ unique and complementary data acquisition strategies to adapt to the different characteristics of communication and sensing tasks. For the communication task, the spatial distance between the transmitter and receiver horn antennas is systematically adjusted from 30cm to 110cm, with each adjustment interval being 20cm. At each distance point, 100 OFDM frames are transmitted using a fixed modulation scheme, and the generated intermediate frequency waveform is recorded using an 80GSa / s real-time oscilloscope. Each OFDM frame simultaneously contains data subcarriers and pilot subcarriers, thereby supporting simultaneous evaluation of equalization capabilities and passive sensing capabilities.

[0064] In contrast, passive sensing tasks require sophisticated distance regression analysis, necessitating dense and continuous coverage of distance values. To this end, some embodiments of this disclosure collect a large amount of OFDM pilot response data as frequency-domain OFDM signal segments by randomly sampling the distance between the transmitter and receiver (ranging from 30 cm to 110 cm). This strategy ensures that the sensing model fully observes distance diversity during training and generalizes well to unseen test distances. The actual distance between the transmitter and receiver is measured using a laser rangefinder with centimeter-level accuracy. To ensure data consistency and reduce multipath interference, all measurements are performed in a controlled line-of-sight environment with fixed antenna alignment and consistent hardware parameters.

[0065] To evaluate cross-modulation generalization capability, some embodiments of this disclosure have been used to conduct complete experiments on both QPSK and 16QAM modulation schemes. Under the same hardware and channel conditions, each modulation scheme was tested independently, and an equal amount of sample data was collected. The resulting balanced dataset contains multiple complex-valued OFDM signal segments, each corresponding to a known transmission label (real symbol) and distance parameter (real distance).

[0066] Each sample instance contains a complex-valued OFDM signal segment and its corresponding symbol label (for equalization processing) and a true distance value (for the sensing task). After processing by the receiver's DSP, the extracted frequency-domain subcarrier sequence is formatted into a fixed-length sliding window. These windows serve as input to the neural network, and their aligned true symbols and propagation distances serve as dual-task supervision labels. The dataset is randomly divided into training, validation, and test sets in an 80:10:10 ratio. By setting multiple sets of distance points, signal-to-noise ratio levels, and modulation schemes, the dataset is designed to fully reflect the signal diversity of THz ISAC systems in real-world scenarios, thereby supporting comprehensive training and evaluation of the proposed CV Res-TCN framework (communication-sensing integrated model) in both tasks.

[0067] During inference, the communication-aware integrated model extracts a sliding window centered on complex-valued subcarriers around the target symbol or pilot location as input. Each subcarrier is represented by real and imaginary parts, and the input window is flattened into a two-dimensional tensor of shape [W,2], where W=31 represents the number of subcarriers in the window. The network architecture of the communication-aware integrated model includes: an initial 128-channel one-dimensional complex-valued convolutional layer (kernel size 3), 6 Res-TCN modules, a final complex-valued convolutional layer for dimensionality reduction, and two task-specific heads—a linear layer containing complex-valued symbol equalization and a regression head for distance estimation.

[0068] For example, this integrated communication-aware model is implemented based on the PyTorch framework and trained using the Adam optimizer. The mean squared error loss function is used for the communication equalization task, while the Huber loss function is employed for the distance regression task to enhance robustness against outliers. During training, random deactivation is not enabled because data diversity provides sufficient regularization, and batch normalization is applied after each convolutional layer. The entire network is trained simultaneously on both tasks using a multi-task learning framework.

[0069] The final integrated communication and sensing model contains fewer than 50,000 trainable parameters, making it lightweight enough to perform real-time inference on edge AI processors or embedded DSP cores. During inference, the communication equalization and distance sensing tasks are executed in parallel, enabling efficient integration into industrial terahertz communication-sensing terminals.

[0070] The following experimental results are used to evaluate the effectiveness of the above construction method.

[0071] To comprehensively evaluate the effectiveness of the integrated communication and sensing model of the terahertz ISAC system in this disclosure, experiments were conducted from two dimensions: equalization performance under typical modulation schemes and passive sensing accuracy tests under different operating conditions. The experimental data, sourced from a hybrid hardware and software testing platform, fully demonstrates that the system, under actual deployment constraints, can both ensure communication stability and achieve centimeter-level precision sensing capabilities.

[0072] First, regarding the evaluation of equalization performance under OFDM-QPSK and 16QAM modulation.

[0073] To verify the robustness and universality of the proposed CV Res-TCN equalizer (communication-sensing integrated model), comparative experiments were conducted at different wireless link distances. Figures 10 to 12 The bit error rate (BER) performance in three typical scenarios is demonstrated: 8-Gbaud OFDM-QPSK, 16-Gbaud OFDM-QPSK, and 4-Gbaud OFDM-16QAM. All tests were conducted on a line-of-sight terahertz link using the same hardware configuration. Given that the proposed CV Res-TCN operates in the frequency domain, frequency-domain neural equalizers were specifically selected as a benchmark for fair comparison, including frequency-domain deep convolutional neural networks (FDDCNN) and FD-ANN models, with BER performance as the evaluation metric.

[0074] like Figure 10As shown, the equalizer based on CV Res-TCN (communication-sensing integrated model) significantly reduces the BER of 8-Gbaud OFDM-QPSK signals across all distances (0.3-1.1m), fully demonstrating its ability to suppress nonlinearity and frequency-selective distortion. Notably, it performs best at 0.5m, rather than 0.3m. This is primarily attributed to the saturation effect of the low-noise amplifier at short-range receivers—exceeding the linear dynamic range of the front-end device when the signal power is too high. Specifically, CV Res-TCN reduces the average BER by 66.8%, significantly outperforming FD-ANN's 40.5% and FD-CNN's 55.7% under the same conditions. The largest improvement occurs at 0.3m, where CV Res-TCN reduces the BER by 81.2% compared to the unequalized baseline. These results strongly suggest that CV Res-TCN is most effective for compensating for joint nonlinearity and frequency-domain distortion in THz links, especially for short-range links prone to receiver saturation.

[0075] like Figure 11 As shown, the CV Res-TCN equalizer (communication-aware integrated model) proposed in this disclosure exhibits strong generalization ability and robustness when processing 16-Gbaud OFDM-QPSK signals. These signals are prone to RF impairments due to their higher symbol rates and wider bandwidths. Despite high distortion and path loss, the communication-aware integrated model network constructed in this disclosure effectively suppresses nonlinearity and spectral impairments at all test distances. Particularly noteworthy is the achievement of the lowest BER again at 0.5m, indicating that the mid-range design avoids receiver front-end overload while maintaining sufficient link budget. Quantitatively, CV Res-TCN reduces the average BER by 75.8%, significantly outperforming FDANN's 40.3% and FD-CNN's 49.5% under the same conditions. The maximum gain exceeds 82.6% at 0.9m, demonstrating the network's anti-interference capability in high-loss propagation environments. These results show that CV Res-TCN not only adapts to high data rate scenarios but also outperforms other frequency-domain-based neural equalization methods in complex terahertz environments.

[0076] Figure 12 The results of ablation experiments using 4-Gbaud OFDM-16QAM signals are presented, aiming to separate the contributions of complex-valued modeling and residual learning in the proposed equalizer. This low-code-rate, high-order modulation scheme is particularly susceptible to spectral non-uniformity and symbol decision ambiguity, making equalization processing exceptionally difficult in terahertz channel impairment environments. Figure 12As shown, CV Res-TCN consistently outperforms Real-valued Res-TCN and ComplexTCN across all test distances, achieving the lowest BER at 0.6m: 8.85 × 10⁻⁶. -4 Specifically, CV Res-TCN reduces the average BER by 39.9% compared to the baseline (without equalization), outperforming Real-valued Res-TCN (29.6%) and Complex TCN (32.6%). The maximum relative gain exceeds 50.3% at 0.6m. These results demonstrate that the synergistic effect of complex-valued feature extraction and residual gradient flow is crucial for achieving robust equalization in scenarios involving mixed nonlinearity and frequency-selective distortion.

[0077] Next, we will evaluate the passive sensing performance under OFDM-QPSK and 16QAM modulation.

[0078] In the multi-task CV Res-TCN network (communication-sensing integrated model) constructed according to the embodiments of this disclosure, passive sensing is implemented as a lightweight regression head. The algorithm is based on shared backbone features extracted from pilot subcarriers. To evaluate its performance, experiments were conducted using 16-Gbaud OFDM-QPSK waveforms under different transmission distances and received signal-to-noise ratios (SNRs). To accurately reflect the actual deployment environment, the measured transmission distance was first converted into the corresponding received SNR value. Specifically, the noise power was calculated using a cyclic cross-correlation method based on ZC pilots, and the signal power was obtained from the difference between the total received symbol power and the noise power. Subsequently, the root mean square error (RMSE) of the predicted distance was calculated at each SNR level.

[0079] like Figure 13 As shown, the CV Res-TCN sensing network architecture proposed in this disclosure significantly outperforms the baseline range-Doppler method based on two-dimensional fast Fourier transform under different signal-to-noise ratio (SNR) conditions. Notably, even in high SNR scenarios, the integrated communication-sensing model constructed in this disclosure maintains sub-centimeter accuracy, fully demonstrating its excellent resistance to noise and nonlinear interference. The observed decreasing trend in RMSE indicates that millimeter-level accuracy can be achieved by further improving the SNR. Experimental results confirm that even under challenging terahertz propagation conditions, the multi-task CV Res-TCN method proposed in this invention can effectively achieve accurate passive distance estimation, demonstrating superior performance advantages.

[0080] This concludes the description of the integrated communication sensing model based on complex-valued residual temporal convolutional networks. It should be understood that, in addition to the steps shown above, the integrated communication sensing model based on complex-valued residual temporal convolutional networks may include other steps, all of which fall within the protection scope of the integrated communication sensing model based on complex-valued residual temporal convolutional networks provided in this disclosure.

[0081] From the above description, it can be seen that the present invention achieves the following technical effects: (1) The ideal communication sensing integrated model (multi-task CV Res-TCN architecture) is constructed, which can simultaneously realize cross-modulation signal equalization and passive distance estimation under a unified framework; (2) A lightweight cross-task channel attention module is proposed, which can selectively inject importance weights from one task to another, and can realize cross-task collaboration; (3) The network model is lightweight, has good stability, and the parameters are perfectly adapted to terahertz systems and heterogeneous industrial terminal equipment; (4) The proposed multi-task CV Res-TCN has significant gains after being applied to terahertz systems.

[0082] Example 2 This disclosure also provides a communication sensing integrated model based on a complex-valued residual temporal convolutional network. This communication sensing integrated model is constructed using any of the construction methods for a communication sensing integrated model based on a complex-valued residual temporal convolutional network shown in Embodiment 1 of this disclosure.

[0083] Although embodiments of the present disclosure have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present disclosure, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for constructing a communication-sensing integrated model based on complex-valued residual temporal convolutional networks, characterized in that, include: The spatial distance between the transmitter and receiver of the terahertz communication system is adjusted sequentially. After each adjustment, multiple OFDM frames are transmitted between the transmitter and receiver using a set modulation scheme. The receiver obtains the frequency domain OFDM signal segment corresponding to each OFDM frame. Each OFDM frame contains a data subcarrier and a pilot subcarrier. Construct a training sample set containing multiple sample instances, each sample instance containing the frequency domain OFDM signal segment, as well as the real symbol and real distance corresponding to the frequency domain OFDM signal segment; A communication-sensing integrated model to be trained is constructed. The communication-sensing integrated model includes a communication network architecture and a sensing network architecture that share the same complex-valued residual temporal convolutional network. The input of the communication-sensing integrated model is the frequency domain OFDM signal segment. The communication network architecture performs equalization processing on the frequency domain OFDM signal segment input to the communication-sensing integrated model to output predicted symbols. The sensing network architecture performs distance estimation on the frequency domain OFDM signal segment input to the communication-sensing integrated model to output the predicted distance between the transmitter and the receiver. The communication and sensing integrated model to be trained is trained using the training sample set to obtain an ideal communication and sensing integrated model.

2. The method as described in claim 1, characterized in that, The communication network architecture includes: A communication feature acquisition module is used to acquire the communication features of the frequency domain OFDM signal segment input to the integrated communication sensing model; The complex-valued residual temporal convolutional network is used to process the communication features output by the communication feature acquisition module, so that the communication network architecture obtains the predicted symbol after equalization processing based on the processing result output of the complex-valued residual temporal convolutional network.

3. The method as described in claim 2, characterized in that, The perception network architecture includes: The sensing feature acquisition module is used to acquire the sensing features of the frequency domain OFDM signal segment input to the integrated communication sensing model; The complex-valued residual temporal convolutional network is used to process the perceptual features output by the perceptual feature acquisition module, so that the perceptual network architecture outputs the predicted distance between the transmitter and the receiver based on the processing result of the complex-valued residual temporal convolutional network.

4. The method as described in claim 3, characterized in that, The communication feature acquisition module includes a communication encoder and a communication attention module; wherein, the communication encoder is used to extract communication features based on the frequency domain OFDM signal segment input to the integrated communication sensing model to obtain a first communication feature, and the communication attention module is used to calculate the global average pooling activation value of the real part of the first communication feature using complex convolution and batch normalization algorithms to obtain the communication attention weight of the first communication feature. The perception feature acquisition module includes a perception encoder and a perception attention module; wherein, the perception encoder is used to extract perception features based on the frequency domain OFDM signal segment input to the integrated communication perception model to obtain a first perception feature, and the perception attention module is used to calculate the global average pooling activation value of the real part of the first perception feature using complex value convolution and batch normalization algorithms to obtain the perception attention weight of the first perception feature. The integrated communication and perception model further includes a cross-attention module, which is used to: process the first communication feature based on communication adaptive adjustment of attention weights and the perception attention weights to obtain a second communication feature, and use the second communication feature as the communication feature output by the communication feature acquisition module; and to process the first perception feature based on perception adaptive adjustment of attention weights and the communication attention weights to obtain a second perception feature, and use the second perception feature as the perception feature output by the perception feature acquisition module.

5. The method as described in claim 4, characterized in that, The magnitudes of the communication adaptive adjustment attention weight and the perception adaptive adjustment attention weight are dynamically and adaptively adjusted based on the predicted distance between the transmitter and the receiver, the channel signal-to-noise ratio status information between the transmitter and the receiver, and the predicted phase structure information and frequency information of the frequency domain OFDM signal segment predicted by the integrated communication and perception model during operation.

6. The method as described in claim 4, characterized in that, The integrated communication and sensing model includes a supervision module, which includes: The communication loss calculation module is used to calculate the communication loss value between the predicted symbol and the real symbol output by the communication network architecture. ; The perception loss calculation module is used to calculate the perception loss value between the predicted distance output by the perception network architecture and the true distance. ; The overall loss calculation module is used to calculate losses based on set coefficients. The overall loss of the integrated communication and sensing model is calculated using the following formula. : Furthermore, during the training process of the integrated communication and sensing model to be trained using the aforementioned training sample set, if the overall loss... If the loss is less than the set loss threshold, then the training of the communication-sensing integrated model to be trained is completed, and the trained communication-sensing integrated model is taken as the ideal communication-sensing integrated model.

7. The method as described in claim 6, characterized in that, The communication loss calculation module uses the mean squared error loss function to calculate the communication loss value between the predicted symbol output by the communication network architecture and the true symbol. ; The perception loss calculation module uses the Huber loss function to calculate the perception loss value between the predicted distance output by the perception network architecture and the true distance. .

8. The method as described in claim 4, characterized in that, The complex-valued residual temporal convolutional network is used to process the communication features output by the communication feature acquisition module and the perception features output by the perception feature acquisition module, and output fused features. The communication network architecture further includes: a communication prediction output module, used to process the fused features and output the prediction symbols; The perception network architecture further includes a distance prediction output module, used to process the fused features and output the predicted distance between the transmitter and the receiver.

9. The method as described in claim 1, characterized in that, The modulation schemes include QPSK modulation and 16QAM modulation; The training sample set includes: a first training sample set obtained using QPSK modulation after each adjustment of the spatial spacing between the transmitter and the receiver based on the same terahertz communication system, and a second training sample set obtained using 16QAM modulation.

10. A communication-sensing integrated model based on complex-valued residual temporal convolutional networks, characterized in that, The communication-aware integrated model is constructed using the construction method of the communication-aware integrated model based on complex-valued residual temporal convolutional network as described in any one of claims 1-9.

Images