A demand structuring method and device for a low, small and slow perception system in a sea area

Abstract

The application provides a demand structuring method and device for a sea area low-small-slow perception system. The method provided by the application comprises the following steps: constructing a structured template of the sea area low-small-slow perception system; instantiating the structured template; obtaining first-type missing instances and second-type missing instances based on a plurality of structured instances; constructing first-type training samples by taking the first-type missing instances as samples and corresponding structured instances as labels, and constructing second-type training samples by taking the second-type missing instances as samples and corresponding structured instances as labels; training an initial model by using the first-type training samples and the second-type training samples, to obtain a basic model; freezing part of parameters of the basic model, fine-tuning the parameters by using the first-type training samples and the second-type training samples, to obtain a first special-purpose model and a second special-purpose model; and selecting a target model according to a current input text, so that the target model outputs a complete structured instance based on the current input text, to improve modeling accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sea area perception, and in particular to a demand structuring method and device for a sea area low-small-slow perception system. BACKGROUND

[0002] With the increasing demand for sea area safety prevention and control, accurate perception and efficient prevention of low-small-slow targets have become key technical challenges. As an important means to cope with this challenge, the sea area low-small-slow perception system has gradually attracted widespread attention from all walks of life. However, the existing technology faces significant bottlenecks in actual application. Specifically, the low-small-slow perception system generates a large amount of multi-modal data in the running process, which covers information in perception, monitoring, defense, and other aspects. Although these data have important reference value, the utilization efficiency is very low, and it is difficult to extract effective information to support decision-making in a short time, especially in the complex and rapidly changing environment of the sea area, the modeling accuracy of the perception system is seriously affected. SUMMARY

[0003] Therefore, the present application provides a demand structuring method and device for a sea area low-small-slow perception system to improve the modeling accuracy of the sea area low-small-slow perception system.

[0004] Specifically, the present application is realized by the following technical solutions:

[0005] The first aspect of the present application provides a demand structuring method for a sea area low-small-slow perception system, which comprises:

[0006] constructing a structured template of the sea area low-small-slow perception system; the structured template is used to organize and describe the internal relationship and interaction between perception ability, prevention activity, prevention task, resource and organizational relationship;

[0007] instantiating the structured template to obtain a plurality of structured instances;

[0008] For each structured instance, missing processing is performed on a single field or a plurality of fields in the structured instance to obtain a first type of missing instance missing a single field and a second type of missing instance missing a plurality of fields;

[0009] constructing a first type of training sample with the first type of missing instance as a sample and the structured instance corresponding to the first type of missing instance as a label, and constructing a second type of training sample with the second type of missing instance as a sample and the structured instance corresponding to the second type of missing instance as a label;

[0010] training an initial model using the first type of training sample and the second type of training sample to obtain a basic model;

[0011] freeze part parameters of the base model, respectively fine-tune parameters of the base model with the first type of training samples and the second type of training samples to obtain a first special model adapted to missing a single field and a second special model adapted to missing multiple fields;

[0012] select a target model matching the current input text from the base model, the first special model and the second special model according to the current input text, so that the target model outputs a complete structured instance based on the current input text.

[0013] The second aspect of the application provides a demand structuring device for a low-small-slow perception system of a sea area, the demand structuring device comprising a construction module, a processing module and a prediction module;

[0014] The construction module is configured to construct a structured template of the low-small-slow perception system of the sea area, and the structured template is configured to organize and describe the internal relations and interactions between perception capabilities, prevention activities, prevention tasks, resources and organizational relationships;

[0015] The processing module is configured to instantiate the structured template to obtain a plurality of structured instances;

[0016] The processing module is configured to, for each structured instance, perform missing processing on a single field or multiple fields in the structured instance to obtain a first type of missing instance missing a single field and a second type of missing instance missing multiple fields;

[0017] The processing module is configured to construct a first type of training sample with the first type of missing instance as a sample and a structured instance corresponding to the first type of missing instance as a label, and construct a second type of training sample with the second type of missing instance as a sample and a structured instance corresponding to the second type of missing instance as a label;

[0018] The processing module is configured to train an initial model using the first type of training sample and the second type of training sample to obtain a base model;

[0019] The processing module is configured to freeze part parameters of the base model, respectively fine-tune parameters of the base model with the first type of training samples and the second type of training samples to obtain a first special model adapted to missing a single field and a second special model adapted to missing multiple fields;

[0020] The prediction module is configured to select a target model matching the current input text from the base model, the first special model and the second special model according to the current input text, so that the target model outputs a complete structured instance based on the current input text.

[0021] The third aspect of the present application provides a demand structuring device for a sea area low, small and slow perception system, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor executes the program to realize the steps of the method provided in any one of the first aspect of the present application.

[0022] The demand structuring method and device for a sea area low, small and slow perception system provided by the present application significantly improve the accuracy and adaptability of demand processing of the sea area low, small and slow perception system by constructing a structured template and instantiating, combining missing field processing and hierarchical model training. On the one hand, the structured template establishes a standardized association framework for core elements such as perception ability, prevention activity, prevention task, resource and organizational relationship, ensuring the systematicness and logicality of demand description and avoiding information fragmentation; on the other hand, a basic model is first trained based on the structured instance, and then a special model is trained for single field missing scenarios and multi-field missing scenarios, and through the collaborative selection of the basic model and the special model, efficient completion of different missing degree texts is realized, so that the output complete structured instance is more accurate and more suitable for actual business scenarios, providing solid demand analysis support for accurate perception and dynamic prevention of sea area low, small and slow targets. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 A flowchart of the demand structuring method for a sea area low, small and slow perception system provided by the present application is shown;

[0024] Figure 2 A schematic diagram of the structured template shown in an exemplary embodiment of the present application is shown;

[0025] Figure 3 A structural schematic diagram of the demand structuring device for a sea area low, small and slow perception system provided by the present application is shown. DETAILED DESCRIPTION

[0026] The exemplary embodiments will be described in detail herein with reference to the drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments are not intended to represent all implementations consistent with the present application.

[0027] The terms used in the present application are merely for the purpose of describing specific embodiments and are not intended to limit the present application. The singular forms "a", "an" and "the" used in the present application are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein means and includes any or all possible combinations of one or more associated listed items.

[0028] It should be understood that, although the terms first, second, third, etc. can be employed in this application to describe various information, the information should not be limited to these terms. These terms are only used to differentiate one piece of information from another. For example, without departing from the scope of the application, the first information can also be referred to as the second information, and similarly, the second information can also be referred to as the first information. Depending on the context, the word "if" as used herein can be interpreted as "when" or "upon" or "in response to determining".

[0029] Model-based systems engineering is an advanced systems engineering method that emphasizes model-centered design, development, verification and maintenance of systems and systems. In the construction of low, slow and small target perception system in the sea area, this method can help to sort out system requirements, design system architecture, optimize system performance, and realize communication and cooperation between different professional fields. However, the low, slow and small perception system in the sea area has complex environment, increasingly rich detection and perception means, complex system modeling structure, and lack of general standards, resulting in huge workload and low efficiency, especially in the description of requirements before modeling. Therefore, how to accurately construct the structured requirement knowledge base of the perception system is an important prerequisite for reducing system modeling.

[0030] Therefore, in view of the structured construction of the low, slow and small perception system in the sea area, a general requirement structured template for this professional field is proposed. By using artificial intelligence technology, sample training and learning are realized in this professional field. Key features are automatically and intelligently extracted from relevant data, and structured expression is realized through fine-tuning model, so as to improve the modeling efficiency and decision support ability of the sea perception system under different requirements.

[0031] The specific embodiments are given below to introduce the technical solutions of the application in detail.

[0032] Figure 1 The flow chart of the requirement structuring method for the low, slow and small perception system in the sea area provided by the present application is shown in FIG. 1. Please refer to Figure 1 The method provided by the present embodiment can include:

[0033] S101, constructing a structured template of the low, slow and small perception system in the sea area.

[0034] It should be noted that Figure 2 The schematic diagram of the structured template shown in an exemplary embodiment of the present application is shown in FIG. 2. Please refer to Figure 2The structured template forms the whole from independent individuals through interactive operation under certain constraints, and is used to organize and describe the internal relations and interactions between the perception ability, prevention activities, prevention tasks, resources and organizational relationships; the perception ability is used to describe the ability required to achieve a specific prevention target; the prevention activities are used to describe the prevention tasks performed to achieve a specific prevention target; the prevention tasks are used to describe the specific operations required to achieve a specific prevention target; the resources are used to describe various resources supporting the prevention activities; and the organizational relationships are used to describe the organizational coordination relationships among the prevention activities, prevention tasks and resources.

[0035] Before introducing the demand structuring method for the sea low, small and slow perception system provided by the embodiment, the sea low, small and slow perception system is simply introduced.

[0036] Specifically, the sea low, small and slow perception system is a comprehensive system for detecting, identifying, tracking and preventing low-altitude, small and slow targets at sea. Through data collection, processing and analysis by multiple perception devices (including radar systems, photoelectric devices, sonar systems, radio detectors, laser radars, environmental sensing devices, etc.), the system can master the situation information of the sea area in real time and assist decision-making.

[0037] Specifically, the sea low, small and slow perception system involves task strategy, organizational relationship, system resource, capability demand and prevention activity. In this system, each component realizes perception and prevention of low, small and slow targets through close cooperation and information flow.

[0038] It can be understood that the task strategy is the core of the sea low, small and slow perception system, which determines how to support the task through the organizational relationship and system resource, thereby guiding the prevention activity. The organizational relationship is responsible for coordinating the relationship between each prevention unit and ensuring efficient allocation and coordination of system resources. The system resource ensures the effective scheduling and application of various resources to meet the needs of prevention activities.

[0039] Further, the system resource supports the operation of the entire sea low, small and slow perception system, including effective management and control of perception devices (such as radars, photoelectric devices, unmanned aerial vehicles, etc.). The capability demand refers to the capability requirement of the sea low, small and slow perception system in different prevention environments, including perception accuracy, response speed, data processing capability, etc. The prevention activity ensures that the system can efficiently cope with different levels of threats through the coordination and execution of command tasks.

[0040] Referring to the foregoing description, the structured template defines the mutual relationship between the prevention task, organizational relationship, resource, perception ability and prevention activity, thereby providing a detailed framework and implementation path for the actual application of the sea low, small and slow perception system, and ensuring that the system can flexibly cope with various threats in complex sea environments.

[0041] In particular implementation, the requirement structured template in this scenario can be constructed based on the dodaf2.0 technology. It should be noted that in the embodiment, the corresponding structured template is established based on the characteristics of the sea low small and slow perception system.

[0042] For example, in a possible implementation, the specific implementation process of the present step can include:

[0043] (1) Standardized description of the type of perception capability, dependency relationship, and mapping relationship between perception capability and prevention activities.

[0044] Specifically, the perception capability describes the required capabilities of the system and how these capabilities support specific tasks and goals. It shows the required capabilities from the method to the operation level and helps to identify system requirements and resource allocation.

[0045] Specifically, the perception capability model is used to describe the capabilities required to achieve the prevention goal. The perception capability model can include a variety of preset capabilities. Further, based on the capability classification model, the capabilities in the perception capability model are further decomposed, refined, and classified. In addition, the dependency relationship between capabilities is established to clarify the dependency relationship between capabilities. Further, the mapping between perception capability and prevention activities that enable these capabilities is established to ensure that prevention activities match the required capabilities and determine how to use various available capability elements to perform activities.

[0046] (2) Classification of resources of the sea low small and slow perception system, and clear characteristics and standardized description of each type of resource.

[0047] Specifically, the resources of the sea low small and slow perception system can be classified from multiple dimensions, mainly involving prevention element resources, environmental resources, meteorological resources, intelligence resources, electromagnetic situation information, command and control systems, communication systems, and other resource indicators. For example, Table 1 shows a resource classification table according to an example embodiment of the present application.

[0048] Table 1 Resource classification table corresponding to the sea low small and slow perception system

[0049] Resource category Description Example Prevention element resource Participate in the core prevention unit of perception and prevention Phased array radar, mechanical scanning radar, laser radar, photoelectric, radio spectrum detection, sonar, patrol unmanned aerial vehicle Environmental resources Geographical, hydrological, topographical, and other static or dynamic environmental information Sea conditions, channel distribution, island location Meteorological resources Weather changes, wind direction and speed, precipitation, and other meteorological parameters Fog area, rain belt, typhoon path, and other weather parameters Intelligence resources Prevention situation information from the intelligence system Opponent activity patterns, historical track database, and intelligence data interface Electromagnetic situation information Monitoring and analysis results of the electromagnetic spectrum in the airspace / sea area Radar signal characteristics, communication interference sources Command and control system Information hub for prevention command and control Command platform Communication system Transmission network supporting information exchange between prevention units Satellite communication, 4G base station or 5G base station, etc.

[0050] Specifically, the classification and standardized description of the resources of the sea low small and slow perception system provide a clear framework for the construction of the sea low small and slow perception system, ensuring that different resources can effectively coordinate and complement each other, thereby improving the overall perception capability of the system.

[0051] (3) Construction of the organization and coordination relationship between resources and resources, and between resources and prevention tasks in the sea low small and slow perception system.

[0052] Specifically, after classifying the resources of the sea low, small and slow perception system, it is necessary to build an organizational coordination relationship. In a possible implementation, the specific implementation process of this step can include:

[0053] Step 1, based on the issuing and feedback mechanism of the superior command node to the subordinate perception unit, the vertical relationship between resources and resources is constructed.

[0054] Specifically, the superior command node is a key level responsible for overall planning, issuing instructions and summarizing feedback. It will issue specific detection, tracking, identification and other task instructions to the subordinate perception unit according to the overall sea security situation and prevention task target, and clearly define the working range and focus of each subordinate perception unit. At the same time, the subordinate perception unit will feed back the real-time acquired device running state, environmental changes and other information to the superior command node in a timely manner. This two-way information flow and instruction transmission enables the vertically distributed resources to form an organic whole. The superior command node provides direction guidance and necessary support for the subordinate perception unit, and the subordinate perception unit provides accurate grassroots data for the superior command node, thereby ensuring the efficient coordination of the entire system in the vertical direction. In this step, based on the issuing and feedback mechanism of the superior command node to the subordinate perception unit, the vertical relationship between resources and resources is constructed.

[0055] Step 2, based on the information sharing and collaborative linkage mechanism between the same level prevention units, the horizontal relationship between resources and resources is constructed.

[0056] Specifically, the horizontal relationship focuses on realizing the complementary advantages and collaborative prevention of the same level prevention units. In the sea low, small and slow perception system, the same level prevention units include different types of detection equipment, each of which has different detection advantages and application scenarios.

[0057] Further, through the information sharing mechanism, the same level prevention units can exchange the target information obtained in real time. For example, after the radar discovers the approximate direction of the target, the photoelectric device can quickly and accurately identify the direction, and the radio detector can monitor the communication signals of the target, thereby realizing the all-around and multi-angle perception of the target. In this step, based on the information sharing and collaborative linkage mechanism between the same level prevention units, the horizontal relationship between resources and resources is constructed.

[0058] Step 3, based on the complementary and fusion mechanism between different types of resources, the cross relationship between resources and resources is constructed.

[0059] Specifically, the cross relationship is the key to fully exert the characteristics of various resources and improve the comprehensive perception ability of the system. Different types of resources differ in detection principles, performance parameters, and applicable environments. For example, radar has an advantage in long-distance detection, but its performance may decline in complex weather conditions, while photoelectric equipment is more accurate in close-range identification, and laser radar can provide three-dimensional contour information of the target.

[0060] Further, through the complementary and fusion mechanism, these different types of resources can complement each other and make up for their respective shortcomings. For example, in a foggy environment, the detection effect of radar is affected, at which time infrared photoelectric equipment can be combined for auxiliary detection, and sonar equipment can monitor underwater targets to form comprehensive coverage of air, surface, and underwater targets. This cross relationship enables various resources to work together, fully exert their respective strengths, and improve the adaptability of the entire low-altitude, small, and slow perception system to complex environments and diverse targets. In this step, based on the complementary and fusion mechanism between different types of resources, the cross relationship between resources and resources is established.

[0061] Step 4, establish a mapping table between tasks and required resources.

[0062] Specifically, the mapping table is a corresponding relationship table used to associate the tasks and required resources in the low-altitude, small, and slow perception system. When building the mapping table, factors such as the importance and urgency of the task, the type, performance parameters, and quantity of the required resources are considered. For example, for important and urgent tasks, such as tracking suspected threat targets, resources with superior performance and fast response speed are preferentially matched.

[0063] Further, the mapping table supports adjustments based on real-time task status and resource constraints. For example, when the progress of a task changes, or some resources cannot be normally used due to failure, occupation, etc., the mapping table will be updated in time, and resources will be reallocated to ensure that the task can be continuously and effectively executed, thereby achieving optimal allocation of resources and efficient completion of tasks.

[0064] (4) For different types of defense targets, define the standardized expression of each type of defense task under the defense target; the standardized expression of each type of defense task includes task target, execution condition, required resources, and execution process.

[0065] Specifically, the defense target is a low-altitude, small, and slow target, including fast boats, unmanned aerial vehicles, unmanned boats, and underwater vehicles, which pose a threat to the safety of the sea area. Non-cooperative target behavior modeling is an important part of the defense system, which can help understand and predict the attack patterns of potential threats.

[0066] Further, for these non-cooperative targets, the standardized expression of various prevention tasks needs to cover task objectives, execution conditions, required resources, and execution processes. For our prevention tasks, there are passive prevention tasks and active prevention tasks. Passive prevention tasks include regional alert, concealment and camouflage, deployment of static countermeasures, etc. Active prevention tasks include dynamic search, target locking, active prevention, electronic interference, etc.

[0067] Specifically, the execution of each prevention task needs to have clear standards, including applicable conditions, execution processes, required resources, and expected effects. For example, passive prevention is suitable for situations where the prevention risk has not been determined or needs to be responded to in a low-key manner, while active prevention is suitable for situations where the risk of the other party needs to be prevented and needs to be responded to quickly. At the same time, task switching rules need to be developed to ensure that tasks can be flexibly adjusted in different prevention environments to meet changing prevention needs.

[0068] Through the standardized definition of prevention tasks, it can be ensured that the sea area low-small-slow perception system can efficiently identify, evaluate, and respond to various potential threats when facing various non-cooperative targets, thereby improving the perception ability and response speed of the prevention system.

[0069] (5) Constructing a logical relationship and a time sequence relationship of the prevention activities, and constructing a standardized execution process of the prevention activities according to the logical relationship and the time sequence relationship.

[0070] In specific implementation, a logical relationship of the prevention activities can be established. The logical relationship includes a sequential relationship, a parallel relationship, and a feedback relationship. A time sequence relationship of the prevention activities is established. The time sequence relationship includes response times of tasks, execution relationships of different tasks, and a task scheduling mechanism. According to task requirements, capability requirements, organizational relationships of the prevention activities, and the logical relationship and the time sequence relationship, a standardized execution process of the prevention activities is constructed.

[0071] Specifically, the sequential relationship reflects the progressive nature of task execution in the prevention activities, and the output result of a previous task is the input basis of a subsequent task.

[0072] Further, the parallel relationship emphasizes the coordinated advancement of multiple tasks within the same time period. For example, while tracking a target, target identification and environmental parameter collection can be carried out simultaneously. Tracking equipment continuously acquires target motion data, optical and electronic equipment synchronously identifies target characteristics, and meteorological sensors monitor real-time changes in sea weather. Parallel processing of the three can greatly shorten the overall task cycle and improve the rapid response capability to low-small-slow targets.

[0073] It can be understood that the feedback relationship is the key to realize the dynamic optimization of prevention activities. For example, when performing interception tasks, interception effect data is fed back to the command center in real time. If the interception does not meet the expectations, the command center can adjust the interception strategy according to the feedback information, reschedule resources or replace the interception method, forming a closed-loop mechanism of execution, feedback and adjustment, to ensure that the prevention activities can adapt to the complex and changing sea environment and target behavior.

[0074] Further, the logical relationship of the prevention activities can make the prevention activities efficiently promote in the regular scene and flexibly adjust in the complex and sudden scene, realize the unity of systematicness, efficiency and adaptability, upgrade the prevention system from the simple superposition of scattered tasks to the closed-loop system of organic cooperation, and improve the comprehensive ability to cope with various threats.

[0075] In addition, when establishing the time sequence relationship of the prevention activities, the response time of each task, the execution relationship of different tasks and the task scheduling mechanism can be determined to ensure efficient cooperation in the time dimension.

[0076] Further, the execution relationship of different tasks includes synchronous execution and asynchronous execution. Synchronous execution requires related tasks to be consistent at the time node. For example, when the command center issues an interception instruction, the communication system needs to transmit the instruction to all prevention equipment participating in the interception. Asynchronous execution allows prevention tasks to be flexibly connected in time. For example, while the target tracking task is ongoing, the threat assessment task can be started after accumulating certain data, without waiting for the tracking to be completed.

[0077] The task scheduling mechanism prioritizes the resource allocation of time-sensitive tasks. When multiple tasks are triggered at the same time, the device priority and computing power allocation are dynamically adjusted according to the task urgency and resource occupation to ensure that the prevention tasks are completed within the time constraint.

[0078] Finally, according to the task requirements, capability requirements, organizational relationships of the prevention activities, as well as the logical relationship and time sequence relationship, a standardized execution process of the prevention activities is constructed, and the comprehensive operation specifications of various elements are integrated. The process is task requirement-oriented, relies on the vertical, horizontal and cross organizational relationships among resources, and integrates logical relationships and time sequence relationships into specific operation links.

[0079] Further, the entire process defines standardized steps, time nodes and cooperation rules to ensure that different posts and different equipment form a unified action criterion in the prevention activities, meet the requirements of precision and timeliness of prevention tasks, and realize efficient allocation of resources through organizational relationships, and finally improve the overall efficiency of the sea low, slow target prevention activities.

[0080] (6) The standardized elements corresponding to the perception ability, resources, organizational relationships, prevention tasks and prevention activities are integrated to form the structured template.

[0081] Specifically, the structured template realizes the unified association and mapping of each element, ensures the accurate matching of the perception ability and resource configuration, the organizational relationship and task allocation, the activity process and task requirement, and avoids disconnection of elements. The structured template provides a standardized framework for rapid construction and dynamic adjustment of the system, reduces repeated design, improves overall collaborative efficiency, facilitates standardized management and quantitative evaluation of the prevention process, and enhances the stability and scalability of the system.

[0082] As can be understood from the foregoing description, the structured template covers five dimensions of ability, resource, organization, task, and activity, has clear structure, and defines various resource indicators, ability descriptions, task processes, and strategy rules, with uniform standards. In addition, the structured template supports flexible configuration and rapid deployment of multiple scenarios, multiple tasks, and multiple resources, and will provide a reference for demand modeling and system design for subsequent perception system construction in different sea areas, and provide support for the structured needs of subsequent perception scenarios.

[0083] S102, instantiating the structured template to obtain a plurality of structured instances.

[0084] In specific implementation, structured data matching the structured template can be extracted from technical materials corresponding to the sea low-small-slow perception system, attribute parameters of low-small-slow targets, attribute parameters of detection equipment, and environmental parameters corresponding to the sea low-small-slow perception system, and then the structured data is combined according to the structured template to obtain a structured instance.

[0085] Specifically, the above technical materials relate to system capability requirements, prevention schemes, and other technical materials related to prevention activities.

[0086] Further, in extracting structured data, different types of data can be uniformly formatted and denoised to obtain processed data, and then for image data in the processed data, structured data matching the structured template is extracted through an image processing model; for text data in the processed data, structured data matching the structured template is identified through a text analysis model.

[0087] It should be noted that the technical materials corresponding to the sea low-small-slow perception system, the attribute parameters of the low-small-slow targets, the attribute parameters of the detection equipment, and the environmental parameters corresponding to the sea low-small-slow perception system are multi-modal data, mainly involving image data and text data. In specific implementation, the perception ability, prevention activities, prevention tasks, resources, and organizational relationships and other related contents can be extracted from the multi-modal data.

[0088] Specifically, multimodal data has different formats and units, so all data needs to be converted into a unified format for subsequent processing and analysis. For image data, text can be extracted to construct image-text pairs. For text data, it needs to be converted into a standard text format, ensuring that all character encodings are consistent to avoid garbled text issues.

[0089] Specifically, denoising processing includes image data denoising and text data denoising. Image data denoising reduces noise interference with subsequent image analysis, ensuring the validity of the image data. Text data denoising refers to the process of cleaning and standardizing meaningless or interfering content in text information.

[0090] Furthermore, for the image data in the processed data, relevant information is extracted using an image processing model. For example, text information in the image can be extracted using an OCR tool, such as the model identifier of a small, slow-moving target or the parameter text of a detection device. If the image originates from a PDF document, the document will be parsed simultaneously to obtain the corresponding text description of the image, ultimately constructing an image-text pair.

[0091] For the text data in the processed data, a text analysis model is used to identify structured data that matches the structured template. For example, the text data can be segmented, stop words can be removed, and then a text analysis model (such as BERT or LSTM) can be used to perform deep analysis on the text to extract key information. Then, the extracted data is organized according to the structured template and converted into a structured format.

[0092] Furthermore, after extracting the relevant structured data, the structured data is combined according to the requirements of the structured template to form structured instances, providing a foundation for subsequent model training.

[0093] S103. For each structured instance, perform missing processing on a single field or multiple fields in the structured instance to obtain a first type of missing instance with a missing single field and a second type of missing instance with multiple missing fields.

[0094] It can be understood that each structured implementation instance includes fields such as perception ability, prevention activity, prevention task, resource and organization relationship, and each field has a definite value. For each structured instance, a single field or multiple fields therein can be processed for missing. For example, for each structured instance, any one of the fields of perception ability, prevention activity, prevention task, resource and organization relationship can be missing (that is, other fields remain unchanged, and only the value of a single field is removed or marked as missing), to obtain a first type of missing instance missing a single field. For another example, for each structured instance, two or more of the fields of perception ability, prevention activity, prevention task, resource and organization relationship can be missing, to obtain a second type of missing instance missing multiple fields.

[0095] S104, constructing a first type of training sample with the first type of missing instance as a sample and the structured instance corresponding to the first type of missing instance as a label, and constructing a second type of training sample with the second type of missing instance as a sample and the structured instance corresponding to the second type of missing instance as a label.

[0096] Specifically, a complete structured instance and a missing instance of missing data corresponding to the complete structured instance are paired to take the missing instance as an input of the model and the complete structured instance as a target output, so as to construct a training sample.

[0097] It can be understood that in the embodiment, two types of training samples are constructed through processing of the first type of missing instance and the second type of missing instance. Specifically, the first type of training sample is: the missing instance missing a single field as an input and the complete structured instance as a label; and the second type of training sample is: the instance missing multiple fields as an input and the complete structured instance as a label.

[0098] S105, training an initial model by using the first type of training sample and the second type of training sample, to obtain a basic model.

[0099] It should be noted that in a possible implementation, the number of the first type of training sample and the number of the second type of training sample are not much different. The basic model learns how to predict and restore the missing field according to the known field by training the input data of the missing field.

[0100] Optionally, in a possible implementation, the specific implementation process of the step can include:

[0101] Step 1, constructing an initial model and configuring LoRA technology to fine-tune the initial model, to obtain a fine-tuned model.

[0102] Specifically, the specific structure of the initial model is set according to actual needs, and the specific structure of the initial model is not limited in the embodiment.

[0103] LoRA technology is an efficient fine-tuning method that adapts the pre-trained initial model by introducing a low-rank matrix, allowing quick adjustments when processing new tasks without changing a large number of model parameters, thereby saving computing resources and improving the efficiency of fine-tuning.

[0104] The core idea of LoRA technology is to decompose the weight matrix of the original model into a low-rank matrix, so that only the low-rank part of the matrix needs to be updated during fine-tuning, while the remaining part of the initial model remains unchanged.

[0105] Step 3, training the fine-tuning model using the first type of training samples and the second type of training samples to obtain the base model.

[0106] It should be noted that during the training of the fine-tuning model, batch training strategy and gradient accumulation strategy can be used. Among them, the batch training strategy refers to dividing the entire training sample into several batches, and inputting each batch of data into the model for training at a time. Through batch training, hardware resources can be used more efficiently and model training can be accelerated. For example, in one possible implementation, the batch size batch_size=16, that is, the model processes 16 training samples at a time during training. It should be noted that the number of first type of training samples and the number of second type of training samples are not significantly different during one training.

[0107] Further, gradient accumulation is a technique commonly used in large model training, which can simulate large batch training under memory constraints. Gradient accumulation strategy allows the use of a smaller batch size (e.g. 16), and accumulates the gradient through multiple forward propagations, and finally performs a backward propagation. This can effectively reduce the memory usage.

[0108] Optionally, in one possible implementation, real-time training visualization can also be performed during the training process to display key indicators during the training process; according to the key indicators, dynamically adjusting the specified hyperparameters in the training process; periodically saving the intermediate state of the model during the training process.

[0109] In specific implementation, during the training process, the training progress of the model can be displayed through visualization tools or charts to display key indicators such as loss, accuracy, and gradient changes. Further, the specified hyperparameters can be dynamically adjusted according to the key indicators, and the specified hyperparameters can include learning rate, Batch Size, etc.

[0110] In a specific implementation, the learning rate can be dynamically adjusted based on the loss. When the loss decreases slowly, it can be that the learning rate is too large, causing the model to oscillate around the optimal solution. In this case, the learning rate can be reduced. When the loss oscillates, it can be that the learning rate is too small, causing the model to train too slowly or stagnate. In this case, the learning rate can be appropriately increased.

[0111] It should be noted that the specified hyperparameters in the training process are dynamically adjusted based on the key indicators, which aims to optimize the hyperparameters according to the training of the model, thereby improving the training efficiency and avoiding overfitting or underfitting.

[0112] In addition, during the training process, the intermediate state of the model can be saved regularly, for example, in one possible implementation, the intermediate state of the model is saved once every epoch or every several steps. By regularly saving the intermediate state of the model, it can help to resume training from the recently saved state when problems occur during the training process, avoiding the need to start training from scratch.

[0113] S106, freeze part of the parameters of the base model, respectively fine-tune the parameters of the base model with the first type of training sample and the second type of training sample, to obtain a first special model adapted to missing a single field and a second special model adapted to missing multiple fields.

[0114] It should be noted that freezing part of the parameters means that part of the parameters remain unchanged during fine-tuning, and only the remaining parameters are updated.

[0115] In a specific implementation, the part that needs to be frozen is set according to actual needs, which is not limited in this embodiment. For example, part of the layers in the base model or a specific module in the base model can be frozen.

[0116] Referring to the foregoing description, the first type of training sample is a sample missing a single field, and the second type of training sample is a sample missing multiple fields. In this step, a data set can be constructed based on the first type of training sample for fine-tuning the special model adapted to missing a single field; similarly, a data set can be constructed based on the second type of training sample for fine-tuning the special model adapted to missing multiple fields.

[0117] S107, according to the current input text, selecting a target model matching the current input text from the base model, the first special model and the second special model, so that the target model outputs a complete structured instance based on the current input text.

[0118] Optionally, in a possible implementation, the specific implementation process of the step can include: judging whether the current input text is the first type of missing text missing a single field or the second type of missing text missing multiple fields; selecting the first special model as the target model when the current input text is the first type of missing text, and selecting the second special model as the target model when the current input text is the second type of missing text; and selecting the basic model as the target model when the current input text is neither the first type of missing text nor the second type of missing text.

[0119] For example, in a possible implementation, the input text can be description information for describing requirements, and the basic model is directly selected as the target model; for another example, the input text can be a missing structured instance and a prompt word, the prompt word can indicate a currently missing field, based on the prompt word, it can be determined whether a single field or multiple fields are currently missing, and then a matched model is selected based on this.

[0120] According to the foregoing description, the method provided by the embodiment has at least the following advantages:

[0121] (1) Improving the modeling accuracy of structured instances

[0122] By constructing a structured template, instantiating the structured template, constructing training samples by using the structured instance, training the basic model, the first special model and the second special model based on the training samples, and finally outputting a complete structured instance based on the input text by using the basic model, the first special model and the second special model, the modeling accuracy can be improved by combining the basic model and multiple special models.

[0123] (2) Improving the flexibility and adaptability of the model, and ensuring the completeness and accuracy of the structured instance

[0124] By training the basic model, the first special model and the second special model, the scheme can dynamically select the most suitable model for reasoning according to the specific situation of the input text. In this way, multiple models can adapt to various different scenarios, improve the flexibility and adaptability of the system, and ensure that the generated structured instance is more complete and accurate.

[0125] (3) Better transfer learning and knowledge transfer

[0126] In the scheme, when constructing multiple special models, the strategy of transfer learning is adopted by fine-tuning the basic model, so that each special model can utilize the knowledge learned in the basic model. Through the strategy of transfer learning, each model can further optimize on the basis of the existing powerful feature representation when fine-tuning for different missing scenarios, thereby reducing the training time and improving the performance of the model.

[0127] The demand structuring method for the sea low-small-slow perception system provided by the embodiment significantly improves the accuracy and adaptability of demand processing of the sea low-small-slow perception system. On the one hand, the structured template establishes a standardized association framework for core elements such as perception ability, prevention activity, prevention task, resource, and organizational relationship, ensuring the systematicness and logicality of demand description and avoiding information fragmentation. On the other hand, a basic model is first trained based on the structured instance, and then special models are trained for single-field missing scenarios and multi-field missing scenarios. Through the collaborative selection of the basic model and the special model, efficient completion of different missing degree texts is realized, and the output complete structured instance is more accurate and more suitable for actual business scenarios, providing solid demand analysis support for accurate perception and dynamic prevention of sea low-small-slow targets.

[0128] In addition to providing a demand structuring method for the sea low-small-slow perception system, the application also provides a demand structuring device for the sea low-small-slow perception system. Figure 3 The demand structuring device for the sea low-small-slow perception system provided by the application is shown in the figure. Please refer to Figure 3 The demand structuring device for the sea low-small-slow perception system provided by the embodiment includes a construction module 310, a processing module 320, and a prediction module 330.

[0129] The construction module 310 is configured to construct a structured template of the sea low-small-slow perception system. The structured template is used to organize and describe the internal relationship and interaction between perception ability, prevention activity, prevention task, resource, and organizational relationship.

[0130] The processing module 320 is configured to instantiate the structured template to obtain a plurality of structured instances.

[0131] The processing module 320 is configured to perform missing processing on a single field or multiple fields in each structured instance to obtain a first type of missing instance missing a single field and a second type of missing instance missing multiple fields.

[0132] The processing module 320 is configured to construct a first type of training sample by taking the first type of missing instance as a sample and taking the structured instance corresponding to the first type of missing instance as a label, and construct a second type of training sample by taking the second type of missing instance as a sample and taking the structured instance corresponding to the second type of missing instance as a label.

[0133] The processing module 320 is configured to train an initial model by using the first type of training sample and the second type of training sample, to obtain a basic model.

[0134] The processing module 320 is configured to freeze part of parameters of the basic model, and fine-tune parameters of the basic model by using the first type of training sample and the second type of training sample respectively, to obtain a first special-purpose model adapted to missing a single field and a second special-purpose model adapted to missing multiple fields.

[0135] The prediction module 330 is configured to select a target model matched with the current input text from the basic model, the first special-purpose model and the second special-purpose model according to the current input text, so that the target model outputs a complete structured instance based on the current input text.

[0136] The apparatus of the embodiment can be used to execute the method. Figure 1 The steps of the method embodiment, the specific implementation principles and implementation processes are similar, and will not be described here.

[0137] The implementation processes of the functions and roles of the units in the apparatus are specifically described in the implementation processes of the corresponding steps in the above method, and will not be described here.

[0138] Further, the application also provides a structured device for a sea area low-small-slow perception system, including a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor executes the program to realize the steps of any one of the methods provided in the first aspect of the application.

[0139] For the device embodiment, since it basically corresponds to the method embodiment, the related parts are described in the part of the method embodiment. The device embodiments described above are only schematic, and the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, that is, they can be located in one place, or distributed on multiple network units. According to actual needs, part or all of the modules can be selected to achieve the purpose of the scheme of the application. Those skilled in the art can understand and implement it without creative labor.

[0140] The above only is the preferred embodiment of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A demand structuring method for a low, small and slow perception system oriented to a sea area, characterized in that, The method comprises: constructing a structured template of a sea area low, small and slow awareness system; the structured template is used to organize and describe the internal relations and interactions between awareness capabilities, prevention activities, prevention tasks, resources and organizational relationships; instantiating the structured template to obtain a plurality of structured instances; for each structured instance, performing missing field processing on a single field or a plurality of fields in the structured instance to obtain a first type of missing instance missing a single field and a second type of missing instance missing a plurality of fields; constructing a first type of training sample by taking the first type of missing instance as a sample and the structured instance corresponding to the first type of missing instance as a label, and constructing a second type of training sample by taking the second type of missing instance as a sample and the structured instance corresponding to the second type of missing instance as a label; training an initial model using the first type of training sample and the second type of training sample to obtain a base model; freezing part of the parameters of the base model, and respectively fine-tuning the parameters of the base model using the first type of training sample and the second type of training sample to obtain a first special-purpose model adapted to missing a single field and a second special-purpose model adapted to missing a plurality of fields; selecting a target model matching the current input text from the base model, the first special-purpose model and the second special-purpose model according to the current input text, so that the target model outputs a complete structured instance based on the current input text.

2. The method of claim 1, wherein, The method comprises: determining whether the current input text is a first type of missing text missing a single field or a second type of missing text missing a plurality of fields; when the current input text is the first type of missing text, selecting the first special-purpose model as the target model, and when the current input text is the second type of missing text, selecting the second special-purpose model as the target model; when the current input text is neither the first type of missing text nor the second type of missing text, selecting the base model as the target model.

3. The method of claim 1, wherein, The method comprises: extracting structured data matching the structured template from technical materials corresponding to the sea area low, small and slow awareness system, attribute parameters of low, small and slow targets, attribute parameters of detection equipment, and environmental parameters corresponding to the sea area low, small and slow awareness system; combining the structured data according to the structured template to obtain a structured instance.

4. The method of claim 1, wherein, The method comprises: standardizing the description of the types, dependency relationships and mapping relationships between awareness capabilities and prevention activities; classifying the resources of the sea area low, small and slow awareness system and explicitly describing the characteristics and standardized description of each type of resource; constructing the organizational coordination relationships between resources and resources, and between resources and prevention tasks in the sea area low, small and slow awareness system; For different types of prevention targets, define the standardized expression of each type of prevention task under the prevention target; the standardized expression of each type of prevention task includes task target, execution condition, required resources and execution process; Construct the logical relationship and time sequence relationship of the prevention activities, and construct the standardized execution process of the prevention activities according to the logical relationship and the time sequence relationship; Integrate the standardized elements corresponding to the perception ability, resources, organizational relationship, prevention task and prevention activity to form the structured template.

5. The method of claim 3, wherein, The structured data matched with the structured template from the technical data corresponding to the sea area low, small and slow perception system, the attribute parameters of the low, small and slow target, the attribute parameters of the detection equipment, and the environment parameters corresponding to the sea area low, small and slow perception system, includes: Uniformly format and denoise different types of data to obtain processed data; For image data in the processed data, extract structured data matched with the structured template through an image processing model; For text data in the processed data, identify structured data matched with the structured template through a text analysis model.

6. The method of claim 1, wherein, The training of the initial model using the first type of training sample and the second type of training sample includes: Build an initial model and configure LoRA technology to fine-tune the initial model to obtain a fine-tuned model; Train the fine-tuned model using the first type of training sample and the second type of training sample to obtain the base model.

7. The method of claim 6, wherein, Training the fine-tuned model using the first type of training sample and the second type of training sample to obtain the base model includes: Real-time training visualization of the training process to display key indicators in the training process; Dynamically adjust specified hyperparameters in the training process according to the key indicators; Periodically save the intermediate state of the model during the training process.

8. The method of claim 4, wherein, The construction of the organizational coordination relationship between resources and resources, and between resources and prevention tasks in the sea area low, small and slow perception system includes: Based on the downward and feedback mechanism of the superior command node to the subordinate perception unit, the vertical relationship between resources and resources is constructed; Based on the information sharing and collaborative linkage mechanism between prevention units at the same level, the horizontal relationship between resources and resources is constructed; Based on the complementary and fusion mechanism between different types of resources, the cross relationship between resources and resources is constructed; A mapping table between tasks and required resources is established; wherein, when establishing the mapping table between tasks and required resources, according to the importance, urgency and required resource type of the task, the matching resource is allocated to the task; the mapping table supports adjustment according to real-time task state and resource constraint condition.

9. A demand structuring device for a maritime low small slow awareness system, characterized in that The demand structuring device includes a construction module, a processing module and a prediction module; The construction module is used to construct a structured template of the sea area low, small and slow perception system; the structured template is used to organize and describe the internal relationship and interaction between perception ability, prevention activity, prevention task, resource and organizational relationship; The processing module is used to instantiate the structured template to obtain a plurality of structured instances; The processing module is configured to, for each structured instance, perform missing field processing on a single field or multiple fields in the structured instance to obtain a first type of missing instance missing a single field and a second type of missing instance missing multiple fields; The processing module is configured to construct a first type of training sample by taking the first type of missing instance as a sample and the structured instance corresponding to the first type of missing instance as a label, and construct a second type of training sample by taking the second type of missing instance as a sample and the structured instance corresponding to the second type of missing instance as a label; The processing module is configured to train an initial model by using the first type of training sample and the second type of training sample to obtain a base model; The processing module is configured to freeze part of parameters of the base model, fine-tune parameters of the base model by using the first type of training sample and the second type of training sample respectively, and obtain a first special-purpose model adapted to missing a single field and a second special-purpose model adapted to missing multiple fields; The prediction module is configured to select a target model matching the current input text from the base model, the first special-purpose model and the second special-purpose model according to the current input text, so that the target model outputs a complete structured instance based on the current input text. 10.A demand structuring device for a low, slow, and small maritime awareness system, comprising: A computer program product comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the steps of the method of any one of claims 1-8 when executing the program.

Images