Multi-task electromagnetic model based on hybrid expert network

Abstract

The invention discloses a multi-task electromagnetic model based on a hybrid expert network, and belongs to a wireless communication technology. The model comprises a preprocessing module, a feature extraction module, a task output module and a pre-training-fine tuning learning strategy. The preprocessing module carries out standardization processing on the multi-source electromagnetic signals; the feature extraction module is based on a Transform structure, introduces a hybrid expert network to replace part of a traditional feedforward neural network, and dynamically selects an expert sub-network through a task specific routing mechanism; the task output module configures a special structure according to different task targets; in the pre-training stage, a mask auto-encoder is used for pre-training large-scale label-free data, and a downstream task is subjected to full-amount fine adjustment through small-scale label data. According to the method, multi-task collaborative learning and differential expression are realized, and the recognition performance, robustness and processing efficiency of the model in a complex electromagnetic environment are improved.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a multi-task electromagnetic model based on a hybrid expert network. Background Technology

[0002] Electromagnetic signals (EM signals) are physical quantities that propagate in space or a medium in the form of electromagnetic waves and carry information content. They are usually artificially modulated and generated by changing parameters such as the amplitude, frequency, and phase of a carrier wave to encode and transmit information. As a fundamental resource of modern information society, electromagnetic signals are the core carrier supporting the operation of wireless communication, radar detection, and satellite navigation. With the continuous expansion of electromagnetic applications, they are also playing an increasingly important role in emerging scenarios such as unmanned systems, intelligent sensing, and electronic countermeasures. Therefore, in-depth research on the time-frequency structure, characteristic representation, propagation, and interference mechanisms of electromagnetic signals is of great significance for improving the system's sensing capabilities, anti-interference capabilities, and autonomous decision-making capabilities.

[0003] The processing and understanding of electromagnetic signals has given rise to various analysis tasks, such as modulation identification, blind source separation, and radio frequency fingerprinting. These tasks play a crucial role in signal decoding, interference suppression, and spectrum management. Current electromagnetic analysis methods primarily focus on dedicated models optimized for specific tasks; each model is designed for a single task, achieving excellent performance on those tasks. However, in recent years, with the increasing complexity of electromagnetic environments and the rapid development of intelligent electromagnetic networks, electromagnetic signals have exhibited characteristics of multi-source, diversity, and dynamism. Against this backdrop, traditional "single-task, single-model" signal analysis methods have gradually revealed their limitations. Multiple independently trained small models lead to model redundancy and wasted computational power; the inability to share knowledge between tasks results in insufficient system adaptability and robustness. Furthermore, in multi-task concurrent application scenarios, the parallel operation of multiple models introduces problems such as increased management complexity and increased response time latency, hindering the construction of a unified intelligent electromagnetic analysis framework.

[0004] Based on the aforementioned needs, in recent years, academia and industry have begun to explore multi-task electromagnetic fundamental models. These models aim to handle multiple electromagnetic domain tasks within a unified network architecture, enabling collaborative learning across tasks ranging from signal modulation recognition to radio frequency fingerprinting. Compared to traditional single-task models, multi-task fundamental models can capture potential correlations between tasks in joint optimization, promoting information fusion and knowledge transfer through a shared representation space, thus enhancing the model's robustness and generalization. In this process, the model not only learns task-specific features but also learns universal electromagnetic features, providing crucial support for subsequent few-sample tasks, out-of-distribution signal recognition, and cross-domain adaptation.

[0005] Current research has achieved good performance through large-scale unlabeled data pre-training and small-scale labeled data fine-tuning, followed by full-scale fine-tuning in the downstream process, demonstrating the feasibility of the pre-training-fine-tuning paradigm in the field of electromagnetic signal processing. However, research in this field is still in its early exploratory stage and faces many challenges: the acquisition and labeling of electromagnetic signals is costly and relies on personnel with specialized backgrounds; different types of electromagnetic signals have significant differences in frequency, modulation methods, and physical characteristics, requiring models to cope with heterogeneous and dynamically changing data during the learning process; the knowledge sharing mechanism between different tasks is still unclear, and models are prone to performance trade-offs and information conflicts when uniformly processing multiple tasks. Based on the above problems and practical engineering needs, this invention proposes a multi-task electromagnetic fundamental model based on a hybrid expert network. Summary of the Invention

[0006] This invention aims to overcome the problems of high model redundancy, insufficient knowledge sharing between tasks, and limited generalization ability in complex electromagnetic environments in existing electromagnetic signal processing methods. Specifically, it addresses the shortcomings of traditional deep learning models in simultaneously adapting to electromagnetic tasks such as modulation recognition, signal classification, and RF fingerprinting, as well as insufficient utilization of cross-task information. To this end, this invention proposes a multi-task electromagnetic model based on a hybrid expert network. Through dynamic expert selection and task adaptive routing mechanisms, it achieves collaborative learning and differentiated representation among multiple tasks, thereby improving the system's recognition performance, robustness, and processing efficiency in multi-source and diverse electromagnetic environments.

[0007] To achieve the above objectives, the present invention provides the following technical solution: This invention protects a multi-task electromagnetic model based on a hybrid expert network, comprising a preprocessing module, a feature extraction module, a task output module, and a pre-training-fine-tuning learning strategy module, specifically: The preprocessing module is used to process multi-source electromagnetic signals into a standardized input format. The feature extraction module is used to perform deep feature modeling and temporal relationship modeling on sequence signals using a multi-layer Transformer network structure. The task output module consists of multiple sets of task-specific structures, and adapts classifiers, regression layers or decoders according to the type of electromagnetic task. The pre-training-fine-tuning learning strategy module is used to train large-scale unlabeled electromagnetic data using a masked autoencoder during the pre-training stage, and to perform full fine-tuning using small-scale labeled data during the downstream task stage.

[0008] Furthermore, the specific processing procedure of the preprocessing module is as follows: The original I / Q signal is subjected to amplitude normalization. The normalized signal is then packaged and the packaged sequence is segmented into patches. A learnable positional code is superimposed on each patch to obtain the input features of the feature extraction module.

[0009] Furthermore, each layer of the multi-layer Transformer network structure includes a self-attention layer, a feedforward network, a normalization layer, and a residual layer connection. In particular, even-numbered layers introduce hybrid expert networks to replace traditional feedforward neural networks. Each hybrid expert layer contains 8 expert sub-networks. Expert selection weights are calculated through task-specific routing networks. For each token, the two experts with the largest weights are selected to participate in the calculation and are weighted and summed.

[0010] Preferably, a maximum capacity for processing tokens is first set for each expert. When the number of tokens allocated to an expert exceeds its capacity limit, the excess tokens will no longer enter the expert's forward computation, but will be directly passed through the residual structure. When the number of allocated tokens is insufficient, zero-padding alignment is performed on the expert's input.

[0011] Furthermore, in the routing distribution phase, an importance-based token sorting mechanism is introduced. The importance score of each token is calculated based on its weight in the routing network, and expert resources are allocated to the tokens in descending order of importance score. During the model training phase, in order to constrain the expert load, an auxiliary loss term is introduced in addition to the main task loss to participate in joint optimization.

[0012] Furthermore, a constraint is placed on the actual number of tokens received by each expert. A load statistics method based on probability estimation is introduced in the routing phase. That is, random perturbation is added to the routing score to estimate the probability that a token will be routed to each expert. The probability is then accumulated within the batch range as an approximate measure of the expert load.

[0013] Furthermore, a masked autoencoder is employed as a pre-training strategy, utilizing large-scale unlabeled electromagnetic data to pre-train the shared feature extraction module. During pre-training, the input patch is randomly masked, and only a small number of visible patches are used to reconstruct the masked portion, enabling the model to learn the global structure of the electromagnetic signal. The final loss in the pre-training phase is the weighted sum of the reconstruction loss and the auxiliary loss. After pre-training is completed, the process moves to the full fine-tuning phase for the downstream task.

[0014] Furthermore, the pre-training-fine-tuning learning strategy module adopts a masked autoencoder as a pre-training strategy and uses large-scale unlabeled electromagnetic data to pre-train the shared feature extraction module. During the pre-training process, the input patch is randomly masked, and only a small number of visible patches are used to reconstruct the masked part, so that the model learns the global structure of the electromagnetic signal.

[0015] Furthermore, the full fine-tuning stage of the downstream task adopts an end-to-end joint training approach, synchronously updating routing parameters and expert parameters, enabling the hybrid expert network to adaptively adjust the expert selection strategy according to different tasks.

[0016] Furthermore, the task output module is configured with different structures according to the task type: For classification tasks such as modulation recognition, a fully connected classification layer is used to output the probability distribution of each category; For regression-related tasks such as parameter estimation, a regression output layer is used to generate continuous values. For signal reconstruction tasks, a decoder structure is used to output the sequence results.

[0017] In another aspect, the present invention protects a computer-readable storage medium storing a computer program that, when executed by a processor, implements any of the methods described above.

[0018] Technical effects of the present invention Compared with the prior art, the present invention has the following advantages: The present invention has the following advantages: This invention proposes a multi-task electromagnetic model based on a hybrid expert network, capable of simultaneously adapting to various electromagnetic tasks such as modulation recognition, passive localization, and signal denoising under a unified shared backbone structure. By introducing a shared hybrid expert network and employing a task-specific routing mechanism, the model achieves dynamic feature division and efficient knowledge sharing among different tasks. This avoids the problem of mutual interference between tasks in traditional multi-task models and significantly improves the model's parameter utilization and inference efficiency. Compared to the fixed-structure Transformer, this invention uses learnable expert selection routing in the feature extraction module, enabling the model to automatically select the optimal expert combination for different tasks, improving the adaptability and discriminative ability of feature representation. Large-scale unlabeled pre-training using a MAE masked autoencoder enables the model to acquire universal electromagnetic representation capabilities, maintaining stable performance even under data scarcity or complex channel conditions. This pre-training-fine-tuning paradigm effectively enhances the model's task transfer and generalization capabilities, achieving high-precision recognition, high robustness, and low deployment cost in multi-task electromagnetic scenarios, demonstrating engineering application value.

[0019] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of a multi-task electromagnetic model based on a hybrid expert network in an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of the attention mechanism in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] This invention provides a multi-task electromagnetic model based on a hybrid expert network. Its core concept lies in constructing a unified multi-task electromagnetic analysis framework, using a shared expert network as a general feature extraction module, and dynamically selecting experts through customized routes for different tasks, thereby achieving collaborative learning of various electromagnetic tasks within the same model. This invention first uses large-scale unlabeled electromagnetic data for pre-training to obtain universally applicable electromagnetic signal representations; then, it fine-tunes the model using small-scale labeled data in downstream tasks, selecting and combining shared experts through task-specific routes to achieve differentiated modeling for each task. Through the synergistic effect of shared experts and task routing, this invention enhances the deep knowledge sharing capability among multiple tasks while avoiding task conflicts and redundant computations, thus improving the model's generalization ability and overall processing efficiency in complex electromagnetic environments.

[0026] Example: The multi-task electromagnetic model based on a hybrid expert network proposed in this invention mainly consists of a preprocessing module, a feature extraction module, a task output module, and a pre-training-fine-tuning learning strategy. Through the collaborative design of these modules, unified modeling and differentiated representation of various electromagnetic signal tasks are achieved.

[0027] The preprocessing module unifies the multi-source electromagnetic signals into a standardized input format suitable for deep model learning. First, amplitude normalization is performed on the raw I / Q signals to ensure signals collected from different acquisition devices and under different signal-to-noise ratios maintain a consistent numerical scale. Then, several signal samples are bundled together, combining multiple signals into a single training sample to improve GPU utilization. Next, the preprocessed sequence is divided into multiple fixed-length patches to reduce the length of the input sequence. To explicitly incorporate temporal information into the sequence, this module further superimposes learnable positional codes onto each patch, enabling the model to capture the temporal order of the signals. This module obtains the normalized signal input, which serves as the input to the feature extraction module.

[0028] The feature extraction module is built on a Transformer architecture, consisting of multiple stacked self-attention encoders, used to extract deep temporal features of electromagnetic signals. In network design, this invention employs an even number of layers to introduce a hybrid expert network instead of the traditional feedforward neural network (FFN), enhancing the model's ability to express multi-task-differentiated features. Each hybrid expert layer contains multiple shared expert sub-networks, with consistent expert structures but independent parameters; simultaneously, a task-specific routing mechanism is set up to dynamically select and combine experts based on the feature requirements of the current task. The Transformer's self-attention layer is responsible for extracting global correlations, while the hybrid expert structure provides differentiated deep representations, enabling the model to achieve knowledge sharing and flexible division of labor among multiple tasks within a unified architecture. This module obtains a general representation of the electromagnetic signal, which serves as the input to the task output module.

[0029] The task output module is used for specialized feature decoding and result generation based on different task objectives. This module consists of multiple task-specific structures, which can employ classifiers, regression layers, decoders, or other adapted structures depending on the task type. For example, a fully connected classifier head can be used for modulation recognition tasks; a lightweight decoder structure can be used for signal recovery or reconstruction tasks. This module outputs the final task result.

[0030] The pre-training-fine-tuning learning strategy consists of two parts: a pre-training phase and a downstream fine-tuning phase. This invention employs a masked autoencoder as the pre-training strategy, utilizing large-scale unlabeled electromagnetic data to pre-train the shared feature extraction module. During pre-training, the input patches are randomly masked, and only a small number of visible patches are used to reconstruct the masked parts, enabling the model to learn the global structure of electromagnetic signals. The pre-training process does not require task labels, which helps the model obtain a general electromagnetic representation. Subsequently, a full fine-tuning approach is used in the downstream task, training the model end-to-end with task-specific data. This allows the task routing mechanism to adaptively select expert combinations and optimize the task output module.

[0031] Taking a ground-based electromagnetic monitoring center as an example, this multi-task electromagnetic model, in actual deployment, consists of a broadband receiving antenna, a radio frequency front-end, a data acquisition card, and a GPU server. The radio signals received by the antenna are down-converted by the radio frequency front-end and digitized by the acquisition card before being transmitted to the server in real-time in baseband I / Q format. The preprocessing program on the server performs amplitude normalization, segment packaging, and patching of the data before inputting it into the multi-task electromagnetic model of this invention. The model first uses a feature extraction network to extract a general electromagnetic representation, and then inputs the general electromagnetic representation into the corresponding task output module according to the task type to complete inference. The inference results are written to the central database for use by upper-level business systems such as electromagnetic situation monitoring, interference alarms, and spectrum management.

[0032] Specific implementation steps: First, the acquired multi-source electromagnetic signals are preprocessed. Each acquired signal is assumed to be a complex sequence. ,in Due to differences in acquisition equipment, sampling environment, and signal-to-noise ratio conditions, amplitude normalization is performed on the original signal to eliminate the impact of inconsistencies in dimensions on model training. The calculation method is as follows: After normalization, the signals are packaged, combining multiple signals into a single 6000-bit sample to improve GPU utilization. Subsequently, the sequence is divided into fixed-length patches, and a learnable positional encoding is overlaid on each patch, allowing the model to preserve temporal order information when processing sequence data. Finally, the segmented and encoded patches are input into the model as input features for the feature extraction module. .

[0033] Figure 2This is a schematic diagram of the attention mechanism in an embodiment of the present invention. The feature extraction module is constructed using a multi-layer Transformer network structure. Each layer includes a self-attention module, a feedforward network, a normalization layer, and residual connections, used for deep feature modeling and temporal relationship modeling of sequence signals. In the network design, even-numbered layers introduce a Mixture of Experts (MoE) module to replace the traditional feedforward network structure, thereby enhancing the model's ability to express multi-task differences. Each MoE layer contains 8 expert sub-networks. Each expert has a consistent structure and independent parameters. The model uses a task-specific routing network. For input features Calculate the expert selection weights: in, w r It's the routing weight. b r It's routing bias; The routing network selects the two experts with the highest weights for each token to participate in the calculation based on the expert weight results, and then performs a weighted sum of their outputs to obtain the final output of the MoE layer: Expert is an expert network, which consists of two fully connected layers, consistent with standard feedforward neural networks.

[0034] After feature extraction by a 12-layer Transformer network, a general representation of the signal can be obtained. This representation, as a shared feature, is further input into the task output module to complete predictions for different downstream tasks. During the pre-training phase, the task output module is configured as a decoder network consisting of 8 Transformer layers to reconstruct the occluded input signal, thereby achieving unsupervised learning of the electromagnetic signal features.

[0035] After completing the model backbone structure, the model is pre-trained using a masked autoencoder (MAE) method. Specifically, the input patches are proportionally... In random occlusion, the model reconstructs the occluded portion using only the visible, unoccluded portion. Its reconstruction loss is defined as: in, This represents the actual value of the masked patch. This indicates the model's prediction results. ZmIt is a masked patch. Through this training method, the model can learn the global structure of electromagnetic signals without relying on manually labeled data, providing a feature representation foundation with good generalization ability for subsequent downstream tasks.

[0036] During the training of the hybrid expert network, to avoid excessive concentration of routing processes on a few experts, constraints need to be placed on expert participation. First, a maximum capacity for processing tokens is set for each expert. When the number of tokens allocated to a particular expert exceeds its capacity limit, the excess tokens are no longer included in that expert's forward computation but are directly passed through the residual structure. When the allocated number is insufficient, zero-padding is applied to the expert's input to ensure consistency in parallel computation. Furthermore, in the routing distribution phase, an importance-based token ranking mechanism is introduced. This involves calculating the importance score of each token based on its weight in the routing network, and then allocating expert resources to tokens in descending order of importance score, prioritizing the allocation of the top-1 experts, followed by the top-2 experts, thus completing one round of the routing process.

[0037] To constrain expert workload during model training, an auxiliary loss term is introduced in addition to the main task loss for joint optimization. First, an expert importance constraint term is constructed. Within a batch (a single batch of data), the soft weights of each sample routed to each expert are accumulated, serving as the importance statistic for that expert in that batch. Based on this, the Importance Loss is constructed, defined as follows: in, This represents the statistical set of importance of all experts within the current batch. std(·) and mean(·) represent the standard deviation and mean, respectively, and var(.) represents the variance. This constraint is used to regulate the distribution of different experts in the routing weights.

[0038] Furthermore, to constrain the actual number of tokens received by each expert, a load statistics method based on probability estimation is introduced during the routing phase. This involves adding random perturbations to the routing score to estimate the probability that a token will be routed to each expert, and then accumulating these probabilities within a batch as an approximate measure of expert load. Based on the estimated load distribution, a load constraint loss (Load Loss) is constructed, which is also expressed using the squared coefficient of variation. Here, load represents the distribution of load estimates for each expert in the current batch.

[0039] The final loss during the pre-training phase is a weighted sum of the reconstruction loss and the auxiliary loss: After pre-training, the system enters the full fine-tuning phase for downstream tasks. During fine-tuning, the feature extraction module and task output module are jointly trained end-to-end using electromagnetic data with task annotations, enabling the hybrid expert network to adaptively adjust the expert selection strategy according to different tasks. The task output module is configured with different structures based on the task type: for classification tasks such as modulation recognition, a fully connected classification layer is used to output the probability distribution of each category; for regression tasks such as parameter estimation, a regression output layer is used to generate continuous values; and for signal reconstruction tasks, a decoder structure is used to output the sequence results. During training, routing parameters and expert parameters are updated synchronously through a backpropagation mechanism, allowing shared experts to complete parameter adjustments under multi-task constraints, thereby obtaining a multi-task electromagnetic model adapted to downstream tasks.

[0040] Once fine-tuning is complete and convergence conditions are met, the trained model parameters are saved and deployed to the actual electromagnetic signal processing system. During system operation, the real-time acquired electromagnetic signals are input into the preprocessing module for standardization, and then sent to the feature extraction module and task output module. According to the task configuration, the corresponding signal analysis function is automatically executed, and information such as modulation type, feature parameters, or reconstruction results are output, completing the multi-task electromagnetic modeling process.

[0041] In another aspect, the present invention protects a computer-readable storage medium storing a computer program that, when executed by a processor, implements any of the methods described above.

[0042] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0043] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-task electromagnetic model based on a hybrid expert network, characterized in that, It includes a preprocessing module, a feature extraction module, a task output module, and a pre-training-fine-tuning learning strategy module. The preprocessing module is used to process multi-source electromagnetic signals into a standardized input format. The feature extraction module is used to perform deep feature modeling and temporal relationship modeling on sequence signals using a multi-layer Transformer network structure. The task output module consists of multiple sets of task-specific structures, and adapts classifiers, regression layers or decoders according to the type of electromagnetic task. The pre-training-fine-tuning learning strategy module is used to train large-scale unlabeled electromagnetic data using a masked autoencoder during the pre-training stage, and to perform full fine-tuning using small-scale labeled data during the downstream task stage.

2. The multi-task electromagnetic model based on a hybrid expert network according to claim 1, characterized in that, The specific processing procedure of the preprocessing module is as follows: The original I / Q signal is subjected to amplitude normalization. The normalized signal is then packaged and the packaged sequence is segmented into patches. A learnable positional code is superimposed on each patch to obtain the input features of the feature extraction module.

3. The multi-task electromagnetic model based on a hybrid expert network according to claim 1, characterized in that, Each layer of the multi-layer Transformer network structure includes a self-attention layer, a feedforward network, a normalization layer, and a residual layer connection. In particular, even-numbered layers introduce hybrid expert networks to replace traditional feedforward neural networks. Each hybrid expert layer contains 8 expert sub-networks. Expert selection weights are calculated through task-specific routing networks. For each token, the two experts with the largest weights are selected to participate in the calculation and are weighted and summed.

4. The multi-task electromagnetic model based on a hybrid expert network according to claim 3, characterized in that, First, a maximum capacity of tokens that can be processed is set for each expert. When the number of tokens allocated to an expert exceeds its capacity limit, the excess tokens will no longer enter the expert's forward computation, but will be directly passed through the residual structure. When the number of tokens allocated is insufficient, zero-padding alignment is performed on the expert's input.

5. The multi-task electromagnetic model based on a hybrid expert network according to claim 3, characterized in that, During the routing distribution phase, an importance-based token sorting mechanism is introduced. The importance score of each token is calculated based on its weight in the routing network, and expert resources are allocated to the tokens in descending order of importance score. During the model training phase, in order to constrain the expert load, an auxiliary loss term is introduced in addition to the main task loss to participate in joint optimization.

6. The multi-task electromagnetic model based on a hybrid expert network according to claim 4, characterized in that, The number of tokens actually received by each expert is constrained. A load statistics method based on probability estimation is introduced in the routing phase. That is, random perturbation is added to the routing score to estimate the probability that a token is routed to each expert. The probability is accumulated within the batch range as an approximate measure of expert load.

7. The multi-task electromagnetic model based on a hybrid expert network according to claim 1, characterized in that, A masked autoencoder is used as a pre-training strategy, and the shared feature extraction module is pre-trained using large-scale unlabeled electromagnetic data. During the pre-training process, the input patch is randomly masked, and only a small number of visible patches are used to reconstruct the masked part, so that the model learns the global structure of the electromagnetic signal. The final loss in the pre-training phase is the weighted sum of the reconstruction loss and the auxiliary loss. After pre-training is completed, the process moves to the full fine-tuning phase for the downstream task.

8. The multi-task electromagnetic model based on a hybrid expert network according to claim 7, characterized in that, The pre-training-fine-tuning learning strategy module uses a masked autoencoder as the pre-training strategy. It uses large-scale unlabeled electromagnetic data to pre-train the shared feature extraction module. During the pre-training process, the input patch is randomly masked, and only a small number of visible patches are used to reconstruct the masked part, so that the model learns the global structure of the electromagnetic signal.

9. The multi-task electromagnetic model based on a hybrid expert network according to claim 7, characterized in that, The full fine-tuning phase of the downstream task adopts an end-to-end joint training approach, synchronously updating routing parameters and expert parameters, enabling the hybrid expert network to adaptively adjust the expert selection strategy according to different tasks.

10. The multi-task electromagnetic model based on a hybrid expert network according to claim 1, characterized in that, The task output module is configured with different structures according to the task type: For classification tasks such as modulation recognition, a fully connected classification layer is used to output the probability distribution of each category; For regression-related tasks such as parameter estimation, a regression output layer is used to generate continuous values. For signal reconstruction tasks, a decoder structure is used to output the sequence results.

Images