Electromagnetic signal sorting and identifying system based on general large model

Abstract

The invention discloses an electromagnetic signal sorting and recognition system based on a general large model, and the system is characterized in that a task controller module splits a user task into sub-demands, and calls a series of sorting and recognition models meeting conditions to form a task dynamic solution; the model arrangement derivation module performs evaluation, optimization and fusion derivation on the model by utilizing heuristic methods such as multi-target collaborative decomposition based on a preliminary evaluation result, the heuristic learning module guides task reasonable arrangement and model iterative optimization through human-computer interaction, and the model evolution module realizes model scale compression and performance optimization. The key sample library provides structural data support for model construction, fine tuning and evaluation, and finally a closed-loop optimization mechanism of problem decomposition, collaborative solution, evaluation feedback and evolution promotion is formed. A large model is used for extracting signal multi-dimensional feature knowledge from mass marked / unmarked electromagnetic data, and the generalization ability and the rapid adaptive ability of the model are remarkably improved through parameterized knowledge storage and task customization setting.

Description

Technical Field

[0001] The present application relates to the technical field of signal analysis and processing, and in particular to an electromagnetic signal sorting and identification system based on a general large model. Background Art

[0002] With the introduction of the Transformer architecture, the number of parameters in deep learning models has surpassed the billion-level and is heading towards the trillion-level, driving the technological evolution of general artificial intelligence. In natural language processing, OpenAI's GPT-3 demonstrated zero-shot transfer learning capabilities in 2020, and Google's Switch Transformer significantly improved computational efficiency through a mixture of expert models (MoE) architecture. In computer vision, Meta's SAM model, launched in 2023, achieved zero-shot transfer for image segmentation, and DINOv2 surpassed traditional supervised methods in multiple visual tasks through self-supervised learning. Multimodal fields integrate the processing capabilities of multiple sources of information, including text and images.

[0003] At the application level, there are significant gaps in existing technologies: ① Traditional electromagnetic signal analysis relies on manual experience and fixed rule systems, which makes it difficult to cope with dynamic electromagnetic spectrum changes in complex environments; ② Although the AIP military large model released by the US Palantir company in 2023 has multi-source integration capabilities, its technical architecture has not been disclosed and does not involve core algorithms in the field of electromagnetic signals; ③ Domestic existing information systems generally have problems such as single data processing dimensions, rigid processing methods, and insufficient real-time decision-making capabilities. When faced with massive multimodal electromagnetic signals, feature extraction methods are difficult to update, extraction efficiency is low, and the extracted features are not generalizable enough. It is impossible to dynamically and autonomously construct target signal recognition solutions, and it is difficult to meet the timeliness requirements for identifying abnormal and unknown signals.

[0004] The current technical bottlenecks are mainly reflected in the following aspects: First, electromagnetic signals have time-varying and nonlinear characteristics, and traditional deep learning models find it difficult to capture complex correlations; second, sample data is scarce or zero samples, and conventional training methods are prone to overfitting and insufficient robustness; third, there is a lack of effective modeling methods for multimodal electromagnetic data fusion. These technical defects have seriously restricted the development of intelligent processing capabilities of the electromagnetic spectrum. Summary of the Invention

[0005] The purpose of this application is to provide an electromagnetic signal sorting and recognition system based on a general large model in order to overcome the existing technical defects. The large model is used to mine multi-dimensional signal feature laws from massive labeled / unlabeled electromagnetic data. Through parameterized knowledge storage and task customization settings, the model's generalization ability and environmental adaptability are significantly improved, solving the problems of missed signal sorting and inaccurate type judgment in complex electromagnetic environments.

[0006] The purpose of this application is achieved through the following technical solutions:

[0007] In the first aspect, the present application proposes an electromagnetic signal sorting and identification system based on a general large model, the system comprising a task controller module, a model orchestration and derivation module, a heuristic learning module, a model evolution module, and a key sample library;

[0008] The task controller module is connected to the universal large model base to decompose user tasks into sub-requirements and call the corresponding sorting and identification sub-models to form a dynamic solution. It then conducts collaborative solution of sub-task groups based on the dynamic solution and optimizes the model group and task organization through comprehensive evaluation feedback.

[0009] The model orchestration derivation module is used to further comprehensively evaluate the model through multi-objective collaborative decomposition based on the evaluation results of the task controller and obtain the evaluation value. Then, based on the evaluation value, the subtask groups are orchestrated and optimized, and collaboratively integrated;

[0010] A heuristic learning module, used to guide large-scale model base optimization iterations through human-computer interaction and optimize the orchestration logic of the sorting and recognition business model;

[0011] The model evolution module is used to reduce the size and optimize the performance of electromagnetic recognition models through a multi-objective structural evolution algorithm. It also supports fast pruning in real-time scenarios and global optimization in offline scenarios.

[0012] The key sample library is used to store and manage structured signal samples and semantic descriptions, providing fine-tuning learning data for the large model base.

[0013] In one possible implementation, the electromagnetic signal sorting and identification system has two working modes, namely model-driven and business-driven. In the model-driven mode, the large model knowledge data is regularly updated through the key sample library, and the model optimization task is initiated through the task controller to complete the iteration of model evolution and orchestration logic, thereby realizing the update and upgrade of existing business applications; in the business-driven mode, the signal recognition task is initiated through the application service interface, the general large model is combined with the learned signal sorting and identification knowledge, and the corresponding business application is autonomously called to complete the real-time identification of the signal and the result evaluation.

[0014] In one possible implementation, when the task controller module identifies an unknown signal, it performs:

[0015] Trigger task decomposition and filter identification models that meet process boundaries for re-iterative calculation;

[0016] Build applications that dynamically generate solutions and use heuristic learning to control model derivation and orchestration optimization direction in real time until a local optimal solution is achieved.

[0017] In a possible implementation, the model evolution module is further configured to:

[0018] For real-time electromagnetic signal recognition scenarios, unstructured and structured pruning and compression models are performed through model and task evaluation.

[0019] For non-real-time scenarios, global and local collaborative optimization is performed through multi-objective evolutionary calculation to form a high-performance electromagnetic recognition model structure;

[0020] To meet the needs of distributed deployment, a large model is used to guide the generation of lightweight business basic models.

[0021] In one possible implementation, signal samples in the key sample library are processed through feature extraction, environmental noise pollution reduction, and signal association to generate structured data, providing knowledge rules for the large model base and training and testing data for model orchestration derivation.

[0022] In one possible implementation, the application service interface supports users in invoking signal recognition tasks and publishing optimized model updates to distributed deployment nodes.

[0023] In one possible implementation, the universal large model base is trained with electromagnetic signal data, and has the ability to capture the intrinsic characteristics of electromagnetic signals, mine multi-dimensional and multi-modal signal knowledge, and expand generalization capabilities through parameterized storage.

[0024] In one possible implementation, the model-driven process includes:

[0025] Submit structured data in the key sample library to the universal large model base;

[0026] Set model optimization tasks through the task controller, and perform multi-factor screening of signal samples in the key sample library based on signal type, target type, acquisition time, and applicable scenarios;

[0027] The filtered signal samples are input into the model orchestration derivation module to perform task completion effect evaluation and adaptively evolve the model orchestration logic, parameter structure, and application boundary description based on the evaluation results.

[0028] The evolved model is updated to the application service interface for user calls.

[0029] In one possible implementation, the business-driven process includes:

[0030] The user initiates a signal recognition task through the application service interface. After preprocessing and structural description of the target signal, the universal large model performs preliminary identification and result evaluation.

[0031] If the recognition result is abnormal or an unknown signal, the task controller is triggered to disassemble the task, select the business model that meets the process boundary, and re-iterate the recognition result;

[0032] Through the heuristic learning module, human-computer interaction is used to control the model derivation and arrangement direction, and optimize the model arrangement logic;

[0033] After completing the evolution of model capabilities, the updated model will be published through the application service interface to form a closed loop.

[0034] The above-mentioned main scheme of this application and its further options can be freely combined to form multiple schemes, all of which are schemes that can be adopted and protected by this application; and in this application, (non-conflicting options) can also be freely combined with each other and with other options. After understanding the scheme of this application, those skilled in the art will understand that there are many combinations based on existing technology and common knowledge, all of which are technical solutions to be protected by this application, and they are not exhaustive here.

[0035] This application discloses an electromagnetic signal sorting and recognition system based on a universal large model. Through a task controller module, user tasks are broken down into sub-requirements and a series of qualified sorting and recognition models are called upon to form a dynamic solution for the task. A model orchestration and derivation module, based on the task controller's preliminary evaluation results of the dynamic solution, uses heuristic methods such as multi-objective collaborative decomposition to evaluate, optimize, and fuse models in the solution to further improve task performance. A heuristic learning module guides the rational orchestration of tasks and iterative optimization of models through human-computer interaction to further meet user needs. A model evolution module achieves model size compression and performance optimization, and a key sample library provides structured data support for model construction, fine-tuning, and evaluation. Ultimately, a closed-loop optimization mechanism is formed, integrating problem decomposition, collaborative solution, evaluation feedback, and evolutionary improvement. This system utilizes a large model to extract multidimensional signal feature knowledge from massive amounts of labeled / unlabeled electromagnetic data. Through parameterized knowledge storage and task customization, the model's generalization and rapid adaptability are significantly improved, addressing the problems of missed signal sorting and inaccurate identification and classification under electromagnetic environmental conditions such as complex signals, missing observations, and environmental changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0037] Figure 1 A schematic diagram of an electromagnetic signal sorting and identification system based on a general large model proposed in an embodiment of the present application is shown.

[0038] Figure 2 The figure shows the workflow diagram of the model-driven system proposed in the embodiment of the present application.

[0039] Figure 3 The following shows a business-driven system workflow diagram proposed in an embodiment of the present application.

[0040] Figure 4 A schematic diagram showing the logical relationship between system components proposed in an embodiment of the present application is shown.

[0041] Figure 5 The application diagram of the electromagnetic signal sorting and identification system proposed in the embodiment of the present application is shown. DETAILED DESCRIPTION

[0042] The following describes the embodiments of the present application through specific examples. Those skilled in the art can easily understand the other advantages and effects of the present application from the content disclosed in this specification. The present application can also be implemented or applied through other different specific embodiments. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other unless they conflict.

[0043] Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of this application.

[0044] In existing technologies, the current technical bottlenecks are mainly reflected in the following aspects: first, electromagnetic signals have time-varying and nonlinear characteristics, and traditional deep learning models are difficult to capture complex correlations; second, sample data is scarce, and conventional training methods are prone to overfitting; third, there is a lack of effective modeling methods for multi-source heterogeneous data fusion. These technical defects have seriously restricted the development of intelligent electromagnetic spectrum management and control capabilities.

[0045] Therefore, in response to the problem of insufficient adaptability of traditional recognition methods caused by the emergence of new-system equipment and the upgrading of signal antagonism in complex electromagnetic spectrum environments, this application proposes an electromagnetic signal sorting and recognition system based on a universal large model. By constructing a multi-module collaborative architecture driven by a large model, it breaks through the limitations of traditional template matching and local data training and achieves a comprehensive improvement in signal processing capabilities. It is described in detail below.

[0046] Please refer to Figure 1 , Figure 1 A schematic diagram of an electromagnetic signal sorting and identification system based on a general large model proposed in an embodiment of the present application is shown. The system includes a task controller module, a model orchestration and derivation module, a heuristic learning module, a model evolution module, and a key sample library;

[0047] The task controller module is connected to the general large model base to decompose user tasks into sub-requirements and call the corresponding sorting and identification sub-models to form a dynamic solution. It then collaboratively solves sub-task groups based on the dynamic solution and optimizes the model group and task organization through comprehensive evaluation feedback.

[0048] The specific functions of the task controller module are: First, based on signal recognition tasks input by users, such as unknown radiation source classification and interference signal tracing, combined with the large model's understanding of task scenarios, process constraints, and signal characteristics, complex tasks are broken down into independently executable sub-requirements; Second, according to the characteristics of the sub-requirements, adapted lightweight recognition models are called from the business model library to build a dynamic solution for multi-model collaboration to break through the capability boundaries of a single model; Third, based on the algorithm performance evaluation knowledge base, a reinforcement learning strategy is used to select the optimal module combination to drive the parallel / serial execution of subtask groups, realizing cross-model feature transfer and result fusion; Fourth, the performance of the model group is quantified through comprehensive task evaluation, and the evaluation results are fed back to the model evolution module and task orchestration subsystem in real time, forming a self-evolution closed loop of "task execution-effect quantification-model iteration".

[0049] The model orchestration derivation module is used to further comprehensively evaluate the model through multi-objective collaborative decomposition and obtain the evaluation value based on the evaluation results of the task controller. It then arranges and optimizes the subtask groups based on the evaluation value and performs collaborative integration.

[0050] The specific functions of the model orchestration derivation module are as follows: First, based on the subtask requirements fed back by the task controller, a multi-dimensional evaluation index system including recognition accuracy, computational efficiency, memory usage, and scenario matching is constructed. A heuristic multi-objective collaborative decomposition algorithm is used to comprehensively score candidate models (such as signal sorting models, modulation recognition models, and target association models) to generate quantitative evaluation values ​​for each model. Second, based on evaluation value ranking and task priority strategies (such as prioritizing low-latency models for real-time tasks and ensemble models for high-precision tasks), the module selects the optimal single model or model combination from the business basic model library. At the same time, an adaptive weight allocation algorithm is used to dynamically adjust the contribution weight of the model in the collaborative task. Third, an ensemble learning method is used to perform decision-level fusion of the output results of multiple models. For conflicting recognition results, the large model base is called to perform feature recalibration and confidence arbitration, ultimately generating a unified recognition result after fusion and derivation. Fourth, the performance data of the fused derived model is fed back to the task controller to drive incremental learning of the business basic model library and optimize the weight configuration of multi-objective evaluation indicators, forming a continuous evolution of the model orchestration strategy.

[0051] The heuristic learning module is used to guide the optimization iteration of the large model base through human-computer interaction and optimize the orchestration logic of the sorting and recognition business model.

[0052] The heuristic learning module, through human-computer interaction within the large model base, guides the optimization and iteration of the large model base and the rational arrangement of the sorting and recognition business models. This heuristic interaction guides the direction and means of model arrangement and derivation, continuously testing and optimizing the currently required sorting and recognition functions, thereby forming a model arrangement and derivation combination that best meets user needs.

[0053] The model evolution module is used to compress the scale and optimize the performance of the electromagnetic recognition model through a multi-objective structural evolution algorithm, and supports fast pruning in real-time scenarios and global optimization in offline scenarios.

[0054] In real-time scenarios, the Model Evolution Module can rapidly prune, dynamically removing unimportant parts or connections from the model, reducing computational effort and latency. This allows the model to quickly process large amounts of data and make decisions while maintaining performance. In offline scenarios, the Model Evolution Module utilizes a global optimization algorithm to comprehensively optimize the model's overall structure and parameters to find the optimal model configuration.

[0055] The model evolution module uses scale compression and performance optimization, combined with fast pruning in real-time scenarios and global optimization in offline scenarios, to enable the electromagnetic recognition model to run efficiently in different scenarios while maintaining high performance.

[0056] The key sample library is used to store and manage structured signal samples and semantic descriptions, providing fine-tuning learning data for the large model base.

[0057] The key sample library is responsible for collecting, organizing, and storing electromagnetic signal samples. After preprocessing and structured description, these samples are stored in a standardized and easily accessible format. In addition to storing raw signal samples, the key sample library also provides detailed semantic descriptions for each sample. These semantic descriptions can include information such as signal type, source, acquisition time, and applicable scenarios. The structured signal samples and their semantic descriptions in the key sample library serve as an important data source for fine-tuning learning in the large model base. By using this data, the large model can continuously optimize its parameters and structure to better suit the electromagnetic signal sorting and recognition task.

[0058] Optionally, the electromagnetic signal sorting and identification system has two working modes, namely model-driven and business-driven. In the model-driven mode, the large model knowledge data is regularly updated through the key sample library, and the model optimization task is initiated through the task controller to complete the iteration of model evolution and orchestration logic, thereby realizing the update and upgrade of existing business applications; in the business-driven mode, the signal recognition task is initiated through the application service interface, the general large model is combined with the learned signal sorting and recognition knowledge, and the corresponding business application is autonomously called to complete the real-time identification of the signal and the result evaluation.

[0059] In a model-driven approach, the system regularly retrieves the latest structured data from key sample libraries, continuously updating the general large model base and associated small business models to ensure data diversity and dynamic improvement of knowledge rules. During this process, the task controller proactively initiates model optimization tasks, screening key samples based on multiple dimensions such as signal type, target type, acquisition time, and applicable scenarios. This drives the model orchestration derivation module to adaptively evolve the model orchestration logic, parameter structure, and application boundaries. Ultimately, the optimized model is updated to the application service interface, achieving periodic iterative improvements in model performance.

[0060] In a business-driven approach, when a user initiates a real-time signal recognition task through an application service interface, the system dynamically generates a solution by combining the prior knowledge of the general large model with the business model. If an abnormal or unknown signal is encountered, the task controller triggers a task decomposition and model re-iteration process, screening business models that meet the boundaries and recalculating the recognition results. Simultaneously, the heuristic learning module's human-computer interaction mechanism optimizes the model orchestration direction. Once optimization is complete, the updated model capabilities are published through the service interface, forming a closed loop from task triggering to model evolution, ensuring rapid system response and continuously improving recognition accuracy in complex electromagnetic environments.

[0061] Optionally, when the task controller module identifies an unknown signal, it performs:

[0062] Trigger task decomposition and filter identification models that meet process boundaries for re-iterative calculation;

[0063] Build applications that dynamically generate solutions and use heuristic learning to control model derivation and orchestration optimization direction in real time until a local optimal solution is achieved.

[0064] The task controller module is used when processing unknown signals. When the system identifies unknown signals, if the result is abnormal or the signal is unknown, the task controller module will trigger a series of operations, including task decomposition and screening of business models that meet the process boundaries and re-iterative calculation until the optimized result is obtained, thereby improving the system's recognition ability and adaptability to complex signals.

[0065] Optionally, the model evolution module is also used to:

[0066] For real-time electromagnetic signal recognition scenarios, unstructured and structured pruning and compression models are performed through model and task evaluation.

[0067] For non-real-time scenarios, global and local collaborative optimization is performed through multi-objective evolutionary calculation to form a high-performance electromagnetic recognition model structure;

[0068] To meet the needs of distributed deployment, a large model is used to guide the generation of lightweight business basic models.

[0069] By comprehensively evaluating the algorithms and tasks of redundant models, we employ both unstructured and structured pruning techniques to rapidly compress the models. Unstructured pruning precisely removes neurons or connections that contribute less to signal recognition, while structured pruning can trim certain aspects of the model's structure.

[0070] Computational bionics based on multi-objective evolutionary computing achieves high-performance global electromagnetic recognition model structure optimization through global and local collaborative optimization. Multi-objective evolutionary computing can simultaneously consider multiple model performance indicators, such as recognition accuracy and computational efficiency, and through continuous iterative optimization, find the optimal model structure that meets these indicators.

[0071] By using the large model to guide the learning of the small model of the sorting and identification business, and through technologies such as knowledge distillation, the knowledge and feature extraction capabilities of the large model are transferred to the small model, a lightweight business basic model that meets various signal recognition businesses and application scenarios can be generated.

[0072] Optionally, signal samples in the key sample library are processed through feature extraction, environmental noise pollution reduction, and signal association to generate structured data, providing knowledge rules for the large model base and training and testing data for model orchestration derivation.

[0073] Signal samples in the key sample library undergo a series of traditional sorting and identification processes, including feature extraction and optimization, environmental contamination reduction, and signal correlation. First, feature extraction and optimization techniques identify and extract the most representative and discriminative key features from the raw electromagnetic signal samples. Next, environmental contamination reduction techniques are used to remove noise and interference from the signals. These environmental factors can negatively impact accurate signal recognition, so this process improves signal quality and identifiability. Finally, signal correlation is used to correlate and integrate signals collected from different sources or at different times.

[0074] Optionally, the application service interface supports users to call signal recognition tasks and publish optimized model updates to distributed deployment nodes.

[0075] The application service interface provides users with a convenient way to call signal recognition tasks. Users can use this interface to initiate signal recognition requests and submit electromagnetic signal samples to be identified. During the call, users can configure task parameters based on their specific needs, such as specifying the signal type, setting the recognition priority, and determining the level of detail for the output results. After receiving the user request, the application service interface sends the task to the system's task controller module for processing. Based on the task requirements, the task controller invokes the appropriate signal recognition model for analysis and recognition, and then provides feedback to the user through the application service interface, satisfying their signal sorting and recognition needs.

[0076] The application service interface is also responsible for publishing the updated signal recognition model after system optimization to application units at all levels. When the system's signal recognition model has been iteratively optimized, the application service interface will send update notifications to application units at all levels to ensure that users can keep up to date with the latest version of the model. The application service interface supports automatic or manual deployment of updated models to application units at all levels. Automatic deployment ensures that all users can use the latest model in a timely manner, while manual deployment provides users with more control and is suitable for scenarios that require updates at specific times. In addition, the application service interface also provides model version management capabilities, allowing users to choose to use a specific version of the model for signal recognition to maintain business continuity and stability. When releasing updates, the application service interface will ensure the compatibility of the new model with the existing system to avoid system failures or recognition errors caused by model updates.

[0077] Optionally, the general large model base is trained through electromagnetic signal data, and has the ability to capture the intrinsic characteristics of electromagnetic signals, mine multi-dimensional and multi-modal signal knowledge, and expand generalization capabilities through parameterized storage.

[0078] The universal large-scale model base is trained using extensive electromagnetic signal data, capturing the inherent characteristics of multimodal signals. These characteristics include frequency, amplitude, phase, and other features. By storing parameters, the model solidifies these features and their associated patterns. When encountering new signals, the model can use these parameters for inference, enabling rapid recognition and classification, thereby expanding its generalization capabilities.

[0079] Optional, model-driven processes include:

[0080] Submit structured data in the key sample library to the universal large model base;

[0081] Set model optimization tasks through the task controller, and perform multi-factor screening of signal samples in the key sample library based on signal type, target type, acquisition time, and applicable scenarios;

[0082] The filtered signal samples are input into the model orchestration derivation module to perform task completion effect evaluation and adaptively evolve the model orchestration logic, parameter structure, and application boundary description based on the evaluation results.

[0083] The evolved model is updated to the application service interface for user calls.

[0084] Figure 2 The figure shows the model-driven system workflow proposed in the embodiment of the present application. Model-driven means that the system will regularly evolve and iterate the general large model base and multiple small business models. First, the latest structured data in the key sample library is submitted to the general large model base to improve the data diversity and knowledge rules of the large model; secondly, the large model scheduling task controller is used to set the model optimization task, and the signal samples in the key sample library are screened by multiple factors, such as signal type, target type, acquisition time, applicable scenarios, etc. The screened signal samples are sent to the model orchestration derivation module, and the corresponding task completion effect evaluation is performed. According to the evaluation results, the model orchestration logic, model parameter structure and application boundary description are adaptively evolved, and updated to the application service interface for user call.

[0085] Optional business-driven processes include:

[0086] The user initiates a signal recognition task through the application service interface. After preprocessing and structural description of the target signal, the universal large model performs preliminary identification and result evaluation.

[0087] If the recognition result is abnormal or an unknown signal, the task controller is triggered to disassemble the task, select the business model that meets the process boundary, and re-iterate the recognition result;

[0088] Through the heuristic learning module, human-computer interaction is used to control the model derivation and arrangement direction, and optimize the model arrangement logic;

[0089] After completing the evolution of model capabilities, the updated model will be published through the application service interface to form a closed loop.

[0090] Figure 3The business-driven system workflow diagram proposed in the embodiment of the present application is shown. Business-driven means that when the user has a need to identify a specific target or an abnormal unknown signal, a signal recognition task is initiated to the system. First, the business application unit calls the application service interface to identify multiple target signals. After preprocessing and structured description, the general large model first identifies the signal based on the currently learned signal sorting and recognition knowledge, and evaluates the result. When the identification result is abnormal or faces an unknown target signal, the large model initiates a task for further processing, calls the task controller to disassemble the task and screen the business model that meets the process and boundary to re-iterate and calculate the recognition result, and continuously seeks better processing procedures, models, structures and parameters. In this process, in addition to evaluating the effect of the recognition results, the heuristic learning module can also be called through human-computer interaction to control the direction of model orchestration and derivation. Finally, the model orchestration optimization and the capability evolution of the basic model library are completed and updated to the application service interface; the business application unit realizes the closed loop of specific signal, target recognition and model update by accessing the updated service interface and recognition model.

[0091] Figure 4 A schematic diagram of the logical relationship between the system components proposed in the embodiment of the present application is shown, which shows its main modules and workflow. The figure includes traditional sorting and identification business processes, key sample libraries, large model-based sorting and identification business processes, task controllers, model autonomous orchestration, model evolution modules, heuristic learning modules, model derivation modules, and signal sorting and identification application service interfaces. The traditional sorting and identification business process, which contains multiple modules, such as semantic annotation, feature extraction and optimization modules, environmental pollution reduction modules, signal association identification modules and signal sorting modules. These modules are responsible for preprocessing, feature extraction and preliminary identification of signals, and storing structured signal samples in the key sample library. As the core data storage module of the system, the key sample library is used to store and manage structured signal samples and their semantic descriptions, providing data support for the system.

[0092] The task controller plays a central role in the large-model-based sorting and recognition workflow. It receives task input and, based on task requirements, invokes the corresponding business model to perform signal sorting and recognition. The task controller interacts with the autonomous model orchestration module, the model evolution module, and the heuristic learning module to dynamically orchestrate, optimize, and evolve the model. The autonomous model orchestration module is responsible for selecting the appropriate model from multiple feature extraction models based on task requirements and executing the task. The model evolution module is responsible for iteratively optimizing the model to improve its performance and adaptability. The heuristic learning module optimizes model orchestration and derivation through user interaction and guided optimization.

[0093] The model derivation module generates new business models based on model orchestration and derivation strategies for specific signal sorting and recognition tasks. These business models can be dynamically updated and optimized to adapt to changing signal environments and task requirements. The signal sorting and recognition application service interface connects the system with external application units. Users can use this interface to initiate signal recognition tasks, receive task results, and provide feedback. The application service interface also supports human-computer interaction, allowing users to interact with the system. A universal large model base serves as the underlying support for the entire system, providing common model capabilities and knowledge support for each module.

[0094] Figure 5 An application diagram of the electromagnetic signal sorting and identification system proposed in an embodiment of the present application is shown. It can be used for tasks such as signal sorting, target model status identification, and modulation identification after signal acquisition by a broadband digital receiver. It can also process and analyze offline electromagnetic signal data to form a target type. The type judgment results can be displayed and viewed by the display and control seat. At the same time, the iterative update of the general large model, the sorting and identification basic model, and the evolution model is completed. The system model results can also be applied to lightweight platforms to identify radiation sources in an outward and approaching manner.

[0095] By applying both system model-driven and business-driven working modes, intelligent sorting, identification and processing of radar, communication and navigation signals can be achieved, solving the problems of missed sorting and low recognition rate in complex electromagnetic environments with numerous signals and spectrum overlapping, providing more accurate signal and target identification results for electromagnetic spectrum management and control.

[0096] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0097] First, the team deeply couples traditional signal sorting and recognition processes with general large-scale model technology, leveraging the large-scale model's powerful data processing capabilities to capture valuable information from large amounts of labeled and unlabeled electromagnetic data. By storing electromagnetic signal knowledge in a large number of parameters and customizing them for specific tasks, the model's generalization capabilities are greatly expanded, enabling it to quickly adapt to environmental changes and cope with unknown electromagnetic environments.

[0098] Second, it avoids repeated training and integration of algorithm models in multi-task scenarios, greatly reducing the complexity of system integration. By unifying multi-type signal processing processes, optimizing signal processing algorithms, and iterating signal recognition models, the system architecture and operational procedures are simplified.

[0099] Third, when processing signals in complex electromagnetic environments, this system offers significant advantages over traditional methods and small models. It can better adapt to dynamically changing environments, improving the accuracy of signal recognition and type determination, thus resolving the issues of low signal recognition accuracy and inaccurate type determination in complex electromagnetic environments often encountered by traditional methods.

[0100] Fourth, it effectively enhances the comprehensive sorting and identification capabilities of radar, communication, and navigation signals. By mining inherent characteristic patterns from massive amounts of signal data and correlating multi-dimensional and multi-modal signal patterns, the system can more accurately sort and identify various electromagnetic signals, providing more precise signal and target identification results for intelligent electromagnetic spectrum processing.

[0101] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. An electromagnetic signal sorting and identification system based on a general large model, characterized in that: The system includes a task controller module, a model orchestration derivation module, a heuristic learning module, a model evolution module and a key sample library; The task controller module is connected to the universal large model base to decompose user tasks into sub-requirements and call the corresponding sorting and identification sub-models to form a dynamic solution. It then conducts collaborative solution of sub-task groups based on the dynamic solution and optimizes the model group and task organization through comprehensive evaluation feedback. The model orchestration derivation module is used to further comprehensively evaluate the model through multi-objective collaborative decomposition based on the evaluation results of the task controller and obtain the evaluation value. Then, based on the evaluation value, the subtask groups are orchestrated and collaboratively integrated; A heuristic learning module, used to guide large-scale model base optimization iterations through human-computer interaction and optimize the orchestration logic of the sorting and recognition business model; The model evolution module is used to reduce the size and optimize the performance of electromagnetic recognition models through a multi-objective structural evolution algorithm. It also supports fast pruning in real-time scenarios and global optimization in offline scenarios. The key sample library is used to store and manage structured signal samples and semantic descriptions, providing fine-tuning learning data for the large model base.

2. The electromagnetic signal sorting and identification system according to claim 1, characterized in that: The electromagnetic signal sorting and identification system has two working modes: model-driven and business-driven. In the model-driven mode, the large model knowledge data is regularly updated through the key sample library, and the model optimization task is initiated through the task controller to complete the model evolution and the iteration of the orchestration logic, thereby realizing the update and upgrade of existing business applications. In a business-driven manner, the signal recognition task is initiated through the application service interface, the general large model is combined with the learned signal sorting and recognition knowledge, and the corresponding business application is autonomously called to complete the real-time identification of the signal and the result evaluation.

3. The electromagnetic signal sorting and identification system according to claim 1, characterized in that: When the task controller module identifies an unknown signal, it performs the following operations: Trigger task decomposition and filter identification models that meet process boundaries for re-iterative calculation; Build applications that dynamically generate solutions and use heuristic learning to control model derivation and orchestration optimization direction in real time until a local optimal solution is achieved.

4. The electromagnetic signal sorting and identification system according to claim 1, wherein: The model evolution module is also used to: For real-time electromagnetic signal recognition scenarios, unstructured and structured pruning and compression models are performed through model and task evaluation. For non-real-time scenarios, global and local collaborative optimization is performed through multi-objective evolutionary calculation to form a high-performance electromagnetic recognition model structure; To meet the needs of distributed deployment, a large model is used to guide the generation of lightweight business basic models.

5. The electromagnetic signal sorting and identification system according to claim 1, wherein: The signal samples in the key sample library generate structured data through feature extraction, environmental noise pollution reduction and signal association processing, providing knowledge rules for the large model base and training and testing data for model orchestration derivation.

6. The electromagnetic signal sorting and identification system according to claim 2, characterized in that: The application service interface supports users to call signal recognition tasks and publish optimized model updates to distributed deployment nodes.

7. The electromagnetic signal sorting and identification system according to claim 1, wherein: The general large model base is trained through electromagnetic signal data, and has the ability to capture the intrinsic characteristics of electromagnetic signals, mine the multi-dimensional knowledge of multimodal signals, and expand generalization capabilities through parameterized storage.

8. The electromagnetic signal sorting and identification system according to claim 2, wherein: The model-driven process includes: Submit structured data in the key sample library to the universal large model base; Set model optimization tasks through the task controller, and perform multi-factor screening of signal samples in the key sample library based on signal type, target type, acquisition time, and applicable scenarios; The filtered signal samples are input into the model orchestration derivation module to perform task completion effect evaluation and adaptively evolve the model orchestration logic, parameter structure, and application boundary description based on the evaluation results. The evolved model is updated to the application service interface for user calls.

9. The electromagnetic signal sorting and identification system according to claim 2, characterized in that: The business-driven approach includes: The user initiates a signal recognition task through the application service interface. After preprocessing and structural description of the target signal, the universal large model performs preliminary identification and result evaluation. If the recognition result is abnormal or an unknown signal, the task controller is triggered to disassemble the task, select the business model that meets the process boundary, and re-iterate the recognition result; Through the heuristic learning module, human-computer interaction is used to control the model derivation and arrangement direction, and optimize the model arrangement logic; After completing the evolution of model capabilities, the updated model will be published through the application service interface to form a closed loop.

Images