Equalizer model construction method based on transfer learning and equalizer model

Abstract

The invention provides an equalizer model construction method based on transfer learning and an equalizer model, and relates to the field of communication, and the method comprises the steps: transmitting a plurality of source OFDM frames between a transmitter and a receiver of a terahertz communication system through employing a source modulation format, and constructing a source training sample set; based on the source training sample set, pre-training a to-be-trained equilibrium model constructed based on the complex value residual time sequence convolutional network to obtain a source equilibrium model; transmitting a plurality of target OFDM frames by adopting a target modulation format different from the source modulation format, obtaining target frequency domain OFDM signal segments corresponding to the target OFDM frames, and constructing a target training sample set; performing feature freezing on a part of functional layers in the source equalization model, training the source equalization model by adopting the target training sample set, and adjusting parameters of a part of functional layers which are not subjected to feature freezing in the source equalization model to obtain a target equalizer model adaptive to a source modulation format and a target modulation format, and the generalization ability of the constructed equalizer model is improved.

Description

Technical Field

[0001] This disclosure relates to the field of communication technology, specifically to a method for constructing an equalizer model based on transfer learning and an equalizer model. Background Technology

[0002] Neural network-based equalizers generally suffer from insufficient generalization ability across modulation formats, which limits their application in dynamic ISAC (Sensing-Communication Integrated Systems) systems. Existing equalizer models are trained and evaluated based on specific modulation formats, assuming that the training data is independently and identically distributed. However, in smart industrial environments, modulation formats often need to be dynamically adjusted according to the capabilities of terminal devices and task requirements. For example, low-latency control signals and high-throughput data streams may coexist within the same communication system. In this case, the assumption of consistent data distribution no longer holds, making it impractical to train a dedicated equalizer model for each modulation format, especially in terahertz communication systems where the cost of collecting labeled data is particularly high. Summary of the Invention

[0003] The main purpose of this disclosure is to provide a method for constructing an equalizer model based on transfer learning and an equalizer model, so as to improve the generalization ability of the constructed equalizer model.

[0004] To achieve the above objectives, a first aspect of this disclosure provides a method for constructing an equalizer model based on transfer learning, the method comprising: Multiple source OFDM frames are transmitted between the transmitter and receiver of a terahertz communication system using a source modulation format, and the source frequency domain OFDM signal segment corresponding to each source OFDM frame is obtained by the receiver to construct a source training sample set. Based on the source training sample set, the equilibrium model to be trained, constructed based on the complex residual temporal convolutional network, is pre-trained to obtain the source equilibrium model. Using a target modulation format different from the source modulation format, multiple target OFDM frames are transmitted between the transmitter and receiver, and the target frequency domain OFDM signal segment corresponding to each target OFDM frame is obtained through the receiver to construct a target training sample set; While freezing the features of some functional layers in the source equalization model, the source equalization model is trained using the target training sample set. The parameters of the functional layers in the source equalization model that have not had their features frozen are adjusted to obtain a target equalizer model that can adapt to both the source modulation format and the target modulation format.

[0005] In some embodiments of this disclosure, the source equilibrium model includes: The TCN network layer consists of multiple Res-TCN modules stacked from top to bottom in the order of input and output. The TCN network layer is used to handle: statistical decision boundaries, power amplification nonlinearity, phase noise, and IQ imbalance. The complex-valued convolutional post-processing layer is coupled to the output of the lowest-level Res-TCN module among multiple Res-TCN modules; the complex-valued convolutional post-processing layer is used for frequency domain feature projection and symbol mapping.

[0006] In some embodiments of this disclosure, the number of Res-TCN modules included in the TCN network layer is S; While freezing features in some functional layers of the source equilibrium model, the source equilibrium model is trained using the target training sample set. The parameters of the functional layers in the source equilibrium model that were not subject to feature freezing are adjusted, including: Features are frozen for the bottom N1 Res-TCN modules in the TCN network layer. The source equalization model is trained using the target training sample set to adjust the parameters of the top M1 Res-TCN modules and the complex-valued convolutional post-processing layer in the TCN network layer of the source equalization model; where N1+M1≤S.

[0007] In some embodiments of this disclosure, the number of Res-TCN modules included in the TCN network layer is S; While freezing features in some functional layers of the source equilibrium model, the source equilibrium model is trained using the target training sample set. The parameters of the functional layers in the source equilibrium model that were not subject to feature freezing are adjusted, including: Feature freezing is performed on the high-level M2 Res-TCN modules and complex-valued convolutional post-processing layer of the TCN network layer. The source equalization model is trained using the target training sample set to adjust the parameters of the low-level N2 Res-TCN modules of the TCN network layer in the source equalization model; where N2+M2≤S.

[0008] In some embodiments of this disclosure, the total number of Res-TCN modules included in the TCN network layer is S equal to 6 and M2 equal to 3.

[0009] In some embodiments of this disclosure, M1 equals M2, N1 equals N2, and the sum of M1 and N1 equals S; The bottom N1 Res-TCN modules in the TCN network layer are used to handle: power amplification nonlinearity, phase noise, and IQ imbalance; The M1 high-level Res-TCN modules in the TCN network layer are used to handle: statistical decision boundaries.

[0010] In some embodiments of this disclosure, the complex-valued convolution post-processing layer outputs in the IQ domain; The source equalization model also includes a decision loss function, which determines whether to adjust the parameters of the functional layers in the source equalization model that have not undergone feature freezing by calculating the distance between the predicted value and the true value of the output of the complex-valued convolution post-processing layer in the IQ domain, so as to obtain a target equalizer model that can be adapted to both the source modulation format and the target modulation format.

[0011] In some embodiments of this disclosure, the source modulation format includes an OFDM-QPSK modulation format, and the baud rate of the OFDM-QPSK modulation format includes at least one baud rate; The target modulation format includes: OFDM-16QAM modulation format, and the baud rate of OFDM-16QAM modulation format includes at least one baud rate.

[0012] In some embodiments of this disclosure, the source OFDM frame includes source data subcarriers and source pilot subcarriers; the source training sample set includes multiple source sample instances, and the source sample instances include: source frequency domain OFDM signal segments, the real symbols corresponding to the source frequency domain OFDM signal segments, and the real distance between the transmitter and receiver corresponding to the source frequency domain OFDM signal segments; The target OFDM frame contains target data subcarriers and target pilot subcarriers; the target training sample set includes multiple target sample instances, which include: target frequency domain OFDM signal segments, the real symbols corresponding to the target frequency domain OFDM signal segments, and the real distance between the transmitter and receiver corresponding to the target frequency domain OFDM signal segments; The equalization model includes a communication network architecture and a sensing network architecture that share the same complex-valued residual temporal convolutional network. The communication network architecture is used for communication equalization processing of the source frequency domain OFDM signal segment and the target frequency domain OFDM signal segment based on the input equalization model. The sensing network architecture is used to sense the distance between the transmitter and the receiver based on the source frequency domain OFDM signal segment and the target frequency domain OFDM signal segment based on the input equalization model.

[0013] The second aspect of this disclosure provides an equalizer model based on transfer learning, which is constructed using any of the transfer learning-based equalizer model construction methods provided in the first aspect of this disclosure.

[0014] The equalizer model construction method and equalizer model based on transfer learning provided in this disclosure first pre-trains an equalizer model to be trained based on a complex-valued residual temporal convolutional network using a source training sample set to obtain a source equalizer model. Then, using a target modulation format different from the source modulation format, multiple target OFDM frames are transmitted between the transmitter and receiver, and the target frequency domain OFDM signal segment corresponding to each target OFDM frame is obtained by the receiver to construct a target training sample set. Subsequently, while freezing the features of some functional layers in the source equalizer model, the source equalizer model is trained using the target training sample set to adjust the parameters of the functional layers in the source equalizer model that have not undergone feature freezing, so as to obtain a target equalizer model that can adapt to both the source modulation format and the target modulation format, thereby improving the generalization ability of the constructed equalizer model and enabling it to adapt to the communication equalization process of different modulation formats. Attached Figure Description

[0015] 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.

[0016] Figure 1 A flowchart illustrating a method for constructing an equalizer model based on transfer learning, 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 block diagram of an equalizer model including a multi-task CV Res-TCN network architecture provided in an embodiment of this disclosure; Figure 8 This is a schematic block diagram of a communication network structure provided in an embodiment of the present disclosure; Figure 9 A schematic block diagram of a perception network structure provided in an embodiment of this disclosure; Figure 10A flowchart illustrating the training process of an equalizer model based on transfer learning, provided in one embodiment of this disclosure. Figure 11 This is a schematic diagram comparing experimental data of cross-modulation transfer learning effects of QPSK and 16QAM modulation formats under different frozen TCN layer configurations according to an embodiment of this disclosure; Figure 12 This diagram illustrates a performance comparison of small-sample transfer learning (SSTL) using 5% and 10% OFDM-16QAM data for fine-tuning according to an embodiment of this disclosure.

[0017] It should be noted that the elements in the attached diagram are schematic and not drawn to scale. Detailed Implementation

[0018] 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.

[0019] 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.

[0020] 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.

[0021] 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.

[0022] 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.

[0023] 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.

[0024] In smart industrial environments, heterogeneous wireless terminals exhibit significant differences in communication requirements, ranging from low-order modulation schemes used to control critical links to high-order formats supporting high-throughput tasks such as real-time monitoring or AR / VR streaming. This diverse need necessitates adaptive ISAC receivers capable of handling multiple modulation schemes with low latency and minimal training overhead. However, traditional neural equalizers are typically trained for fixed modulation formats (such as QPSK or 16QAM). The learned feature representations and decision boundaries are essentially correlated with the symbolic geometry and statistical properties of the training constellation diagram. Therefore, directly applying such models to unknown modulation formats often leads to significant performance degradation, requiring costly retraining for each new modulation configuration.

[0025] To address this issue, various strategies have been proposed to promote modulation generalization learning. Transfer learning (TL), as a mainstream method, fine-tunes a trained modulation scheme using a small amount of labeled data. While this method reduces the need for full retraining, its performance is limited by the similarity between the source and target domains. Meta-learning, as an emerging solution, shows potential in few-sample pilot demodulation tasks. The Bayesian active meta-learning framework proposed by Cohen et al. achieves rapid model adaptation through uncertainty estimation, but these methods have mainly been validated in simulated OFDM environments and have not yet been fully applied to hardware-constrained terahertz scenarios. Recent research has also explored transfer learning in the field of channel estimation, whose adaptation goals are similar to modulation format-specific generalization, but existing research on this area remains relatively scarce.

[0026] The disadvantages of the existing technology are as follows: (1) For the same algorithm data, different modulation formats are required to retrain, and the cost of collecting labeled data is high; (2) The generalization ability of cross-modulation format is insufficient and is limited by the similarity between the source domain and the target domain; (3) It has poor applicability in the terahertz field.

[0027] To overcome the shortcomings of existing technologies, this invention proposes a transfer learning-based equalizer model construction method employing adaptive modulation training. This method enables the network to achieve robust performance across multiple modulation formats while maintaining low hardware complexity and training cost. By utilizing residual convolutional blocks, temporal local modeling, and a constellation-aware loss function, this model requires only 10% of the labeled data to achieve efficient transfer between OFDM Quadrature Phase Shift Keying (QPSK) and OFDM-16QAM signals. This design not only achieves lightweight cross-format adaptation but is also particularly suitable for terahertz ISAC system network structures in heterogeneous, low-overhead environments such as smart industrial scenarios.

[0028] refer to Figure 1 This disclosure provides a method for constructing an equalizer model based on transfer learning, which mainly includes the following steps: In S110, multiple source OFDM frames are transmitted between the transmitter and receiver of the terahertz communication system using source modulation format, and the source frequency domain OFDM signal segment corresponding to each source OFDM frame is obtained by the receiver to construct a source training sample set. In S120, the source equilibrium model is obtained by pre-training the equilibrium model to be trained based on the source training sample set and constructing the temporal convolutional network based on complex residuals. In S130, a target modulation format different from the source modulation format is used to transmit multiple target OFDM frames between the transmitter and the receiver, and the target frequency domain OFDM signal segment corresponding to each target OFDM frame is obtained by the receiver to construct a target training sample set. In S140, while freezing the features of some functional layers in the source equalization model, the source equalization model is trained using the target training sample set. The parameters of the functional layers in the source equalization model that have not been frozen are adjusted to obtain a target equalizer model that can adapt to both the source modulation format and the target modulation format.

[0029] In the above scheme, the equalization model to be trained is first pre-trained based on the source training sample set to construct the equalization model to be trained using a complex-valued residual temporal convolutional network, thus obtaining the source equalization model. Then, using a target modulation format different from the source modulation format, multiple target OFDM frames are transmitted between the transmitter and receiver, and the target frequency domain OFDM signal segment corresponding to each target OFDM frame is obtained by the receiver to construct the target training sample set. Afterwards, while freezing the features of some functional layers in the source equalization model, the source equalization model is trained using the target training sample set, and the parameters of the functional layers in the source equalization model that have not been feature-frozen are adjusted to obtain a target equalizer model that can adapt to both the source and target modulation formats, thereby improving the generalization ability of the constructed equalizer model and enabling it to adapt to the communication equalization process of different modulation formats.

[0030] The following section provides a detailed description of the equalizer model construction method based on transfer learning disclosed herein, with reference to the accompanying drawings.

[0031] 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 the equalizer model in 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, and supports both communication and sensing functions under actual hardware constraints.

[0032] 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 3 The digital signal processing flow shown is illustrated, and the structure of the transmitter RF module is as follows: Figure 4As 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 into a predefined pilot subcarrier. After serial-to-parallel conversion, inverse fast Fourier transform, and cyclic prefix insertion, the signal is digitally up-converted to generate a 10GHz intermediate frequency signal, which is then transmitted to a 64GSa / s arbitrary waveform generator. This arbitrary waveform generator performs digital-to-analog conversion, outputting a 10GHz 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 105GHz 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 16Gbaud and a center frequency of 220GHz (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.

[0033] 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.

[0034] 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, cyclic prefix elimination, 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 nonuniformity, which affects 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.

[0035] For example, the equalizer 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 employs a shared encoder structure and is equipped with a task-specific output branch. The equalizer model includes two main functions: communication network functions inherent in the communication network architecture and sensing network functions inherent in the sensing network architecture, simultaneously supporting both 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 band) 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 equalizer model. This equalizer 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 equalizer model achieves advantages through parameter reuse, joint optimization, and task-aware information interaction.

[0036] 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 from the pilot positions. Each branch employs complex-valued convolution and complex-valued batch normalization modules to generate task-specific feature maps.

[0037] 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).

[0038] 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.

[0039] 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 the sensing branch, which contains a complex-valued convolutional layer, a global average pooling layer, a flattening layer, and a linear layer.

[0040] The source equalization model construction method based on the adaptive modulation TL strategy proposed in this disclosure is based on the principle of fine-tuning the parameters of the pre-trained source equalization model to obtain the target equalization model, thereby achieving fast equalizer adaptation between different modulation formats. This TL method is based on the CV Res-TCN equalizer model, and its core principle is that although the constellation structures of different modulation formats differ, the distortion phenomena induced by terahertz (including nonlinear distortion, spectral leakage, and phase noise) are essentially consistent across different modulation formats. These distortion features are effectively encoded in the intermediate residual feature map of the TCN network layer.

[0041] There are various ways to configure the equalization model. For example, in some embodiments, the source equalization model may include: a TCN network layer and a complex-valued convolutional post-processing layer; wherein, the TCN network layer includes multiple Res-TCN modules stacked from top to bottom in the order of input and output; the TCN network layer is used to process: statistical decision boundaries, power amplification nonlinearity, phase noise, and IQ imbalance; the complex-valued convolutional post-processing layer is coupled to the output of the bottommost Res-TCN module among the multiple Res-TCN modules; the complex-valued convolutional post-processing layer is used for frequency domain feature projection and symbol mapping.

[0042] set up and These represent the source training sample set obtained based on the source modulation format and the target training sample set obtained based on the target modulation format, respectively. Although the true constellation points... and While differences exist in amplitude and angular resolution, terahertz damage shares common structural characteristics across different modulation schemes. Assuming residual feature maps of the intermediate TCN network layer... It encodes modulation-independent distortion patterns, requiring only adaptive adjustments to the specific modulation format in the readout layer (TCN network layer and complex-valued convolutional post-processing layer, also known as the output layer). Its training process includes the following steps: pre-training, feature freezing, and readout layer fine-tuning.

[0043] In the pre-training process, the source equalization model is obtained by pre-training the equalization model to be trained based on the complex-valued residual temporal convolutional network using the source training sample set. Feature freezing and readout layer fine-tuning are performed during the training of the source equalization model using the target training sample set. At the same time, the parameters of the functional layers in the source equalization model that have not undergone feature freezing are fine-tuned to obtain a target equalizer model that can adapt to both the source modulation format and the target modulation format.

[0044] For example, the source equalizer model obtained for the target modulation format By freezing features and training the source equalizer model using the target sample set, the output layer of the final source equalizer model is replaced with a new layer, and only the target sample set is used. A small subset of the output layer is fine-tuned, specifically as follows:

[0045] This ensures rapid adaptability while preserving core distortion compensation knowledge.

[0046]

[0047] To achieve cross-format TL, the equalizer model network output is directly represented in the IQ domain:

[0048] This design avoids reliance on hard-coded symbol indexes and enhances compatibility with non-standard constellation diagrams.

[0049] For example, the complex-valued convolutional post-processing layer outputs in the IQ domain. The source equalization model also includes a decision loss function, which determines whether to adjust the parameters of the functional layers in the source equalization model that have not undergone feature freezing by calculating the distance between the predicted and true values ​​of the complex-valued convolutional post-processing layer output in the IQ domain, thereby obtaining a target equalizer model that can adapt to both the source and target modulation formats. Embodiments of this disclosure improve generalization under noisy and distorted conditions by introducing a soft decision loss function that measures the Euclidean distance of each candidate symbol in the IQ plane.

[0050] For example, the number of Res-TCN modules contained in the TCN network layer can be S, where S can be any positive integer such as 5, 6, 7, 8, or 9.

[0051] In some embodiments, a Top-adaptation strategy can be employed. Specifically, while freezing features of some functional layers in the source equalization model, the source equalization model is trained using a target training sample set. When adjusting the parameters of the functional layers in the source equalization model that have not undergone feature freezing, features can be frozen for the bottom N1 Res-TCN modules in the TCN network layer, and the source equalization model can be trained using the target training sample set to adjust the parameters of the higher M1 Res-TCN modules and the complex-valued convolutional post-processing layer in the TCN network layer of the source equalization model; where N1 + M1 ≤ S. M1 and N1 are both positive integers greater than or equal to 1, and their sum is less than or equal to the total number of Res-TCN modules, S.

[0052] In other embodiments, a bottom-adaptation strategy can be employed. Specifically, while freezing features of some functional layers in the source equalization model, the source equalization model is trained using a target training sample set. When adjusting the parameters of the functional layers in the source equalization model that have not undergone feature freezing, features can be frozen for the high-level M² Res-TCN modules and the complex-valued convolutional post-processing layer of the TCN network layer. The source equalization model is then trained using the target training sample set to adjust the parameters of the low-level N² Res-TCN modules in the TCN network layer of the source equalization model; where N² + M² ≤ S. M² and N² are both positive integers greater than or equal to 1, and their sum is less than or equal to the total number of Res-TCN modules, S.

[0053] In some embodiments, the total number of Res-TCN modules included in the TCN network layer, S, can be equal to 6, and M2 can be equal to 3.

[0054] For example, M1 equals M2, N1 equals N2, and the sum of M1 and N1 equals S. The bottom N1 Res-TCN modules in the TCN network layer are used to handle: power amplification nonlinearity, phase noise, and IQ imbalance. The top M1 Res-TCN modules in the TCN network layer are used to handle: statistical decision boundaries.

[0055] There are various ways to specify the type of source modulation format. For example, the source modulation format may include OFDM-QPSK modulation format, and the baud rate of OFDM-QPSK modulation format includes at least one baud rate.

[0056] The type of target modulation format can be varied. For example, the target modulation format may include OFDM-16QAM modulation format, and the baud rate of OFDM-16QAM modulation format includes at least one baud rate.

[0057] To verify the generalization ability of the proposed adaptive modulation TL strategy across different modulation schemes, this disclosure presents complete experiments on both QPSK and 16QAM modulation schemes. Under the same hardware and channel conditions, each modulation format was tested independently, and equal amounts of sample data were collected, serving as the source training sample set and the target training sample set, respectively. The resulting balanced dataset contains multiple complex-valued OFDM signal segments, each corresponding to a known transmission label (real symbol) and distance parameters.

[0058] Table 1 below shows the experimental results after testing. Based on the actual frame structure used in the test platform of the terahertz communication system, a rigorous comparative analysis was conducted on three typical OFDM configurations. All configurations used the same IFFT size, pilot mode, and cyclic prefix length to ensure the fairness and consistency of the evaluation benchmark. Although Case A and Case B differ in modulation format and baud rate, both achieved the same network throughput of 12.04 Gbps, highlighting the trade-off between spectral efficiency and baud rate spread, and demonstrating the superiority of the adaptive modulation TL strategy.

[0059] Table 1 Effective parameters of OFDM signals

[0060] To evaluate the modulation generalization capability of the equalizer after adopting a transfer learning strategy, a pre-trained low-order modulation model (source equalizer model) is adapted to a high-order modulation format using limited data. Specifically, a CV Res-TCN model (an equalizer model based on CV Res-TCN) is first trained on an 8-Gbaud OFDM-QPSK signal, and then the source equalizer model is fine-tuned to 4-Gbaud OFDM-16QAM using a small target training set (target domain data (5% or 10%)). This method simulates a real-world application scenario in smart industrial networks—a strategy with significant advantages due to the difficulty in collecting and labeling large datasets covering all modulation schemes.

[0061] For example, the original CV Res-TCN architecture (equalizer model) consists of six stacked Res-TCN modules followed by a complex-valued convolutional post-processing layer responsible for frequency domain feature projection and symbol mapping. To determine the optimal transfer granularity, this invention proposes two transfer learning strategies: the first is Top-Adaptation, which freezes the lower-level TCN modules near the input and fine-tunes only the higher-level TCN modules and the post-processing layer; the second is Bottom-Adaptation, which freezes the higher-level TCN modules and the post-processing layer and fine-tunes only the lower-level modules near the input. While maintaining the trainability of the remaining layers, the number of frozen TCN blocks is systematically adjusted from 1 to 5.

[0062] Figure 11 The study compared the cross-modulation transfer learning effects of QPSK and 16QAM modulation methods under different frozen TCN layer configurations. Experimental data showed that Bottom-Adaptation consistently outperformed Top-Adaptation. Figure 11As shown, Bottom-Adaptation consistently outperforms Top-Adaptation across all configurations, performing best when the first three Res-TCN modules are frozen. This phenomenon reveals a key characteristic of the representation hierarchy in CV Res-TCN. The lower-level Res-TCN modules are primarily responsible for encoding channel-specific physical distortions, including power amplification nonlinearity, phase noise, and IQ imbalance. These characteristics are closely related to the statistical properties of the input signal, and the statistical properties of QPSK and 16QAM modulation formats differ significantly. In the Top-Adaptation strategy, freezing these lower-level TCN modules confines the model to a suboptimal channel representation that is no longer applicable to higher-order modulations, leading to performance degradation. In contrast, the top-level TCN blocks and the complex-valued convolutional post-processing layer primarily handle symbol constellation mapping and statistical decision boundaries. In the Bottom-Adaptation strategy, freezing these higher-level modules and fine-tuning the lower-level modules closer to the input allows the model to adapt to new signal distributions while maintaining symbol structure stability. Furthermore, due to the presence of residual connections, the frozen higher-layer modules can still retain useful representations, while the unfrozen lower-layer modules can compensate for the mismatch in lower-layer channel characteristics.

[0063] To quantify data efficiency, this invention evaluates the performance of few-shot transfer learning (SSTL) under different proportions of adaptive data. Using the optimal bottom-adaptation strategy and freezing three Res-TCN modules, the proportion of OFDM-16QAM training data (target training set) used for fine-tuning was adjusted from 5% to 10% of the complete dataset. Figure 12 As shown, even using only 10% of the OFDM-16QAM training data (the target training set), SSTL still achieves a bit error rate (BER) comparable to CV Res-TCN trained from scratch with full supervision. Even using only 5% of the data, SSTL still achieves competitive performance. These results demonstrate that the pre-trained QPSK backbone network (source equalizer model) can retain channel representations with good generalization capabilities, enabling accurate symbol recovery even with scarce supervised samples. Compared to the baseline model without CV Res-TCN, the BER reduction of SSTL-5% and SSTL-10% is significant, confirming the effectiveness and practicality of cross-modulation TL in terahertz ISAC systems.

[0064] In some embodiments, the source OFDM frame includes source data subcarriers and source pilot subcarriers; the source training sample set includes multiple source sample instances, each including: a source frequency domain OFDM signal segment, the corresponding real symbol of the source frequency domain OFDM signal segment, and the real distance between the transmitter and receiver corresponding to the source frequency domain OFDM signal segment. The target OFDM frame may include target data subcarriers and target pilot subcarriers; the target training sample set includes multiple target sample instances, each including: a target frequency domain OFDM signal segment, the corresponding real symbol of the target frequency domain OFDM signal segment, and the real distance between the transmitter and receiver corresponding to the target frequency domain OFDM signal segment. As described above, the equalization model at this time may include: a communication network architecture and a perception network architecture sharing the same complex-valued residual temporal convolutional network. The communication network architecture is used for communication equalization processing based on the source frequency domain OFDM signal segment and the target frequency domain OFDM signal segment of the input equalization model, and the perception network architecture is used for perceiving the distance between the transmitter and receiver based on the source frequency domain OFDM signal segment and the target frequency domain OFDM signal segment of the input equalization model.

[0065] By sequentially adjusting the spatial distance between the transmitter and receiver of the terahertz communication system, and acquiring the frequency-domain OFDM signal segment corresponding to each OFDM frame through the receiver after each adjustment (each OFDM frame contains data subcarriers and pilot subcarriers), 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 equalizer model includes a communication network architecture and a sensing network architecture sharing the same complex-valued residual temporal convolutional network. The input to the equalizer model is the frequency-domain OFDM signal segment. The communication network architecture performs equalization processing based on the frequency-domain OFDM signal segment input to the equalizer model to output predicted symbols, while the sensing network architecture performs distance estimation based on the frequency-domain OFDM signal segment input to the equalizer model to output the predicted distance between the transmitter and receiver. Then, the training sample set is used to train the equalizer model to obtain an ideal equalizer model, enabling joint nonlinear equalization and transmitter-receiver distance estimation.

[0066] This concludes the description of the equalizer model construction method based on transfer learning. It should be understood that the equalizer model construction method based on transfer learning may include other steps besides those shown above, all of which fall within the protection scope of the equalizer model construction method based on transfer learning provided in this disclosure.

[0067] From the above description, it can be seen that the present disclosure achieves the following technical effects: (1) It is compatible with mixed formats, has strong generalization ability, and can be used in different terahertz ISAC system algorithms after fine-tuning; (2) It has high data efficiency; (3) It has fast convergence speed; (4) It has high modulation flexibility.

[0068] Example 2 This disclosure also provides an equalizer model based on transfer learning, which is constructed using any of the transfer learning-based equalizer model construction methods provided in Embodiment 1 of this disclosure.

[0069] 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 an equalizer model based on transfer learning, characterized in that, include: Multiple source OFDM frames are transmitted between the transmitter and receiver of a terahertz communication system using a source modulation format, and the source frequency domain OFDM signal segment corresponding to each source OFDM frame is obtained through the receiver to construct a source training sample set. Based on the source training sample set, the equilibrium model to be trained, constructed based on the complex residual temporal convolutional network, is pre-trained to obtain the source equilibrium model. Using a target modulation format different from the source modulation format, multiple target OFDM frames are transmitted between the transmitter and the receiver, and the target frequency domain OFDM signal segment corresponding to each target OFDM frame is obtained through the receiver to construct a target training sample set; While freezing the features of some functional layers in the source equalization model, the source equalization model is trained using the target training sample set. The parameters of the functional layers in the source equalization model that have not been feature-frozen are adjusted to obtain a target equalizer model that can adapt to the source modulation format and the target modulation format.

2. The method as described in claim 1, characterized in that, The source equilibrium model includes: The TCN network layer comprises multiple Res-TCN modules stacked from top to bottom in the order of input and output; the TCN network layer is used to process: statistical decision boundaries, power amplification nonlinearity, phase noise, and IQ imbalance. A complex-valued convolutional post-processing layer is coupled to the output of the lowest-level Res-TCN module among the multiple Res-TCN modules; the complex-valued convolutional post-processing layer is used for frequency domain feature projection and symbol mapping.

3. The method as described in claim 2, characterized in that, The number of Res-TCN modules contained in the TCN network layer is S; While freezing features in some functional layers of the source equalization model, the source equalization model is trained using the target training sample set, and the parameters of the functional layers in the source equalization model that have not undergone feature freezing are adjusted, including: Features are frozen for the bottom N1 Res-TCN modules in the TCN network layer, and the source equalization model is trained using the target training sample set to adjust the parameters of the top M1 Res-TCN modules in the TCN network layer and the complex-valued convolutional post-processing layer in the source equalization model; wherein, N1+M1≤S.

4. The method as described in claim 2, characterized in that, The number of Res-TCN modules contained in the TCN network layer is S; While freezing features in some functional layers of the source equalization model, the source equalization model is trained using the target training sample set, and the parameters of the functional layers in the source equalization model that have not undergone feature freezing are adjusted, including: Feature freezing is performed on the M2 high-level Res-TCN modules of the TCN network layer and the complex-valued convolution post-processing layer. The source equalization model is trained using the target training sample set to adjust the parameters of the N2 low-level Res-TCN modules of the TCN network layer in the source equalization model; where N2+M2≤S.

5. The method as described in claim 4, characterized in that, The total number of Res-TCN modules contained in the TCN network layer is S equal to 6, and M2 equals 3.

6. The method according to any one of claims 2-5, characterized in that, M1 is equal to M2, N1 is equal to N2, and the sum of M1 and N1 is equal to S; The bottom N1 Res-TCN modules in the TCN network layer are used to handle: power amplification nonlinearity, phase noise, and IQ imbalance; The M1 high-level Res-TCN modules in the TCN network layer are used to process: statistical decision boundaries.

7. The method as described in claim 2, characterized in that, The complex-valued convolution post-processing layer outputs in the IQ domain; The source equalization model also includes a decision loss function, which determines whether to adjust the parameters of the functional layers in the source equalization model that have not undergone feature freezing by calculating the distance between the predicted value and the true value of the output of the complex-valued convolution post-processing layer in the IQ domain, so as to obtain the target equalizer model that can adapt to the source modulation format and the target modulation format.

8. The method as described in claim 1, characterized in that, The source modulation format includes OFDM-QPSK modulation format, and the baud rate of the OFDM-QPSK modulation format includes at least one baud rate; The target modulation format includes: OFDM-16QAM modulation format, wherein the baud rate of the OFDM-16QAM modulation format includes at least one baud rate.

9. The method as described in claim 1, characterized in that, The source OFDM frame includes source data subcarriers and source pilot subcarriers; the source training sample set includes multiple source sample instances, and the source sample instances include: the source frequency domain OFDM signal segment, the real symbol corresponding to the source frequency domain OFDM signal segment, and the real distance between the transmitter and the receiver corresponding to the source frequency domain OFDM signal segment; The target OFDM frame includes a target data subcarrier and a target pilot subcarrier; the target training sample set includes multiple target sample instances, and the target sample instances include: the target frequency domain OFDM signal segment, the real symbol corresponding to the target frequency domain OFDM signal segment, and the real distance between the transmitter and the receiver corresponding to the target frequency domain OFDM signal segment; The equalization model includes a communication network architecture and a sensing network architecture that share the same complex-valued residual temporal convolutional network. The communication network architecture is used to perform communication equalization processing based on the source frequency domain OFDM signal segment and the target frequency domain OFDM signal segment input to the equalization model. The sensing network architecture is used to sense the distance between the transmitter and the receiver based on the source frequency domain OFDM signal segment and the target frequency domain OFDM signal segment input to the equalization model.

10. An equalizer model based on transfer learning, characterized in that, The equalizer model is constructed using the equalizer model construction method based on transfer learning as described in any one of claims 1-9.

Images