A multimodal large model driven icing wind tunnel agent system

Abstract

The application provides a multi-modal large model driven icing wind tunnel intelligent agent system, comprising: a data layer for acquiring multi-source target data in an icing wind tunnel test process, processing the multi-source target data, and constructing a processed target data set; a model layer adopting a multi-modal large model as a central control, a plurality of icing wind tunnel field models as execution units, and constituting an intelligent agent collaborative architecture, wherein the multi-modal large model uniformly completes task scheduling and decision output, and each icing wind tunnel field model is responsible for the calculation task of the corresponding field; a technical support layer for integrating and processing the target data set in the icing wind tunnel test process, realizing the interactive calling of the data layer and the model layer; and a business layer for interacting with the user based on the output of the model layer. Through implementation of the scheme, the intelligent degree of the wind tunnel test can be effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of wind tunnel icing test technology, specifically relating to a multimodal large model-driven intelligent system for icing wind tunnels. Background Technology

[0002] With the development of digital twins, the Internet of Things and large model technology, leveraging the powerful cognitive reasoning capabilities of large models to drive domain simulation tools and build "industrial intelligent agent systems" for key equipment to achieve system simulation, prediction and intelligent decision-making has become an important direction for improving test efficiency and operational safety.

[0003] Icing wind tunnels are crucial testing facilities for studying the operational safety of aircraft as a whole or key components such as engines under icing weather conditions. During icing wind tunnel testing, it is necessary to simultaneously acquire multimodal, multi-source, and heterogeneous sensor data, including images, temperature, pressure, liquid water content, droplet size, and icing morphology, and rely on multiple specialized models for joint calculations and state simulations. However, existing icing wind tunnel testing systems generally suffer from the following shortcomings: Firstly, models in different domains operate independently and are fragmented. Task scheduling, data fusion, and decision output during the test heavily rely on human experience, resulting in low testing efficiency, slow response, and difficulty in coping with dynamic adjustment needs under complex weather conditions. Secondly, there is a lack of a unified understanding and correlation mechanism between multi-source data. Traditional rule-based or single deep learning model-based methods struggle to effectively integrate multimodal information such as images and time-series signals, failing to achieve cross-modal joint reasoning and adaptive scheduling, thus limiting the level of intelligence in the testing process. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a multimodal large model-driven intelligent agent system for icing wind tunnels, in order to meet the need to improve the level of intelligence in icing wind tunnel testing.

[0005] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a multimodal large model-driven intelligent agent system for icing wind tunnels, comprising: a data layer for acquiring multi-source target data during icing wind tunnel experiments and processing the multi-source target data to construct a processed target dataset; a model layer that adopts an intelligent agent collaborative architecture with a multimodal large model as the central control and multiple icing wind tunnel domain models as execution units, wherein the multimodal large model uniformly completes task scheduling and decision output, and each of the multiple icing wind tunnel domain models is responsible for the computational tasks of its corresponding domain; a technical support layer for integrating and processing the target dataset during icing wind tunnel experiments and realizing interactive calls between the data layer and the model layer; and a business layer for interacting with users based on the output of the model layer.

[0006] Optionally, the multi-source target data includes wind tunnel test site monitoring data, wind tunnel test environmental parameters, and equipment status; the multi-type icing wind tunnel domain models include a safety monitoring visual model, an aerodynamic flow field proxy model, and an icing and anti-icing numerical model; the safety monitoring visual model is used to determine safety risks based on wind tunnel test site monitoring data; the aerodynamic flow field proxy model is used to predict the aerodynamic flow field characteristics inside the wind tunnel based on wind tunnel test environmental parameters and equipment status; and the icing and anti-icing numerical model is used to predict the icing process based on environmental parameters and the aerodynamic flow field characteristics output by the aerodynamic flow field proxy model.

[0007] Optionally, the technical support layer is configured to include: a multimodal fusion module, used to integrate the target dataset in the data layer, construct an icing wind tunnel industry knowledge base, and align multimodal data during wind tunnel testing using a multimodal fusion algorithm; a retrieval enhancement generation module, used to automatically retrieve standard specifications and historical cases from the icing wind tunnel industry knowledge base when the multimodal large model in the model layer generates decisions or answers, and input the retrieval results as context into the multimodal large model; and a model scheduling module, used to generate standardized calling instructions based on the decisions output by the multimodal large model, using a function call interface to drive various icing wind tunnel domain models in the model layer to complete rapid predictions.

[0008] Optionally, the technical support layer is configured to further include: a status monitoring module, used to monitor the status of the icing wind tunnel test through deep learning and pattern recognition technology, perform pattern analysis on sensor data to identify abnormal operating conditions, and provide real-time feedback to the corresponding domain model of the model layer; and a result post-processing module, used to process the output results of the model layer to generate decision information suitable for use by icing wind tunnel operators.

[0009] Optionally, the business layer is configured to include: a suggestion module, which provides users with decision-making suggestions for icing wind tunnel tests based on user input and through interaction with the model layer; a safety supervision module, which provides early warning of safety hazards based on the real-time perception capabilities provided by the technical support layer; and a data archiving and management module, which uniformly archives and manages all test data, prediction results, operation records, and risk analysis information during the icing wind tunnel test process.

[0010] Optionally, the data layer pre-stores a knowledge graph built based on knowledge of the icing domain.

[0011] This invention provides a multimodal large-model driven intelligent agent system for icing wind tunnels. Using a multimodal large-model as the central control unit, it replaces traditional manual decision-making and distributed model scheduling. Leveraging its powerful cross-modal understanding, semantic reasoning, and task planning capabilities, it uniformly analyzes multi-source target data from the data layer, including images, numerical values, and text. It dynamically schedules various specialized models in the icing wind tunnel domain as execution units, forming an intelligent agent collaborative architecture of central decision-making and domain execution. As a result, the system can autonomously complete task decomposition, model invocation, data interaction, and result fusion in the experimental process, improving the automation and intelligence of icing wind tunnel experiments, reducing manual intervention, and increasing experimental efficiency and result reliability. Furthermore, the technical support layer enables standardized interaction between data and models, while the business layer provides an intuitive human-machine interface, further ensuring the system's operability and scalability in complex icing wind tunnel experimental scenarios.

[0012] Other advantages, objectives, and features of the invention will be set forth in the following description and will be apparent to those skilled in the art in some respects, or may be learned by practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0013] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration: Figure 1 This is a schematic diagram of a specific module structure of an icing wind tunnel intelligent agent system driven by a multimodal large model in this invention; Figure 2 This is an example flowchart of the knowledge graph construction process in this invention; Figure 3 This is a schematic diagram of the overall architecture of the three icing wind tunnel domain models and the multimodal large model in the model layer of this invention; Figure 4 This is a schematic diagram illustrating the linkage between the industry knowledge base and the model layer in this invention; Figure 5 This is a schematic diagram of the overall architecture of the multimodal large model-driven icing wind tunnel intelligent agent system in this invention. Detailed Implementation

[0014] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can also refer to the internal connection of two components; and they can refer to a wireless connection or a wired connection. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0016] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other. This invention provides a multimodal large model-driven intelligent agent system for icing wind tunnels, such as... Figure 1 As shown, it includes: Data layer 101 is used to acquire multi-source target data during the icing wind tunnel test, process the multi-source target data, and construct the processed target dataset; Model layer 102 adopts a collaborative architecture of intelligent agents, with a multimodal large model as the central control and multiple icing wind tunnel domain models as execution units. The multimodal large model uniformly completes task scheduling and decision output, while each type of icing wind tunnel domain model is responsible for the computational tasks of its corresponding domain. Technical support layer 103 is used to integrate and process the target dataset during the icing wind tunnel test, and to enable interactive calls between the data layer and the model layer. Business layer 104 is used to interact with the user based on the output of the model layer.

[0017] For example, the multi-source target data in data layer 101 includes standardized management discipline knowledge, relevant papers, patents and monographs, operating procedures and standard systems, etc., to provide the technical support layer with an industry knowledge base. In addition, it also includes wind tunnel test site monitoring data, wind tunnel test environmental parameters, equipment status, etc. These data are diverse, including various data such as temperature, humidity, wind speed, pressure, spray flow rate, and ice shape images. The multi-source target data can originate from public databases and retrieval platforms, wind tunnel sensors, equipment control systems, video monitoring equipment, and operator manual records. The data acquisition methods are common knowledge and will not be described in detail in this embodiment. Data layer 101 also needs to process the multi-source target data, specifically including preliminary cleaning and standardization, such as unifying time, converting units, removing redundant data and outliers, to give it a unified format and structure for easy subsequent use, thus obtaining the processed target dataset. Data layer 101 also needs to process and store massive amounts of data to provide sufficient support for subsequent model training and fine-tuning. Furthermore, through metadata management, it ensures data consistency and integrity, and improves the efficiency of accessing and retrieving data. In addition, data security and privacy protection are equally important. The data layer ensures that data is fully protected through methods such as encrypted storage and transmission, access control and permission management.

[0018] As an optional implementation, the data layer pre-stores a knowledge graph constructed based on knowledge of the icing domain. To ensure the icing wind tunnel intelligent agent system fully integrates knowledge of the icing domain, the data layer in this embodiment integrates various data sources beyond experimental data, including aerodynamic and icing knowledge, academic papers, patents, and industry standards. Specifically: First, relevant icing domain knowledge is obtained from public databases and retrieval platforms, which can be downloaded in batches, crawled, or API-called after keyword and topic searches. This knowledge, along with internal engineering documents such as standards, equipment manuals, reports, test procedures, and contingency plans, is then uniformly incorporated into the knowledge base. Next, the multi-source data undergoes denoising, deduplication, segmentation, entity extraction, and relation extraction to construct an icing knowledge graph. This knowledge is transformed into a structured form that the intelligent agent system can understand and process, and knowledge graph embedding technology is used to convert it into embedding vectors for use by expert tools during training and by the central processing unit during decision-making and reasoning.

[0019] The knowledge graph construction process, such as Figure 2As shown, the process begins with data preparation, which includes data collection, cleaning, and preprocessing to provide high-quality raw data for subsequent processing. Next comes the information extraction stage, where entity extraction and relation extraction are used to extract the core entities that constitute knowledge and the relationships between them from the preprocessed data. Finally, the knowledge representation stage completes the construction of triples and standardization, transforming the extracted knowledge into a structured and standardized representation, thus forming a knowledge graph that can be understood and applied by computers. It should be noted that knowledge graphs are also an important element in industry knowledge bases.

[0020] Model layer 102 adopts an intelligent agent collaborative architecture, which is constructed from a multimodal large model as the central control and decision-making core and an icing wind tunnel domain model. The icing wind tunnel domain model may include: a safety supervision visual model, an aerodynamic flow field proxy model, and an icing and anti-icing numerical model.

[0021] The multimodal large model is responsible for the system's task scheduling decisions and human-computer interaction. This model can not only understand natural language input, enabling intelligent question answering, experimental design, and experimental record generation, but also directly perceive multimodal data such as ice shape images and monitoring videos. Based on the prediction and recognition results of each model and combined with an industry knowledge base, it generates appropriate answers or decision suggestions, helping users more efficiently complete the scheduling of various system tools, experimental design, data analysis, and documentation, thereby reducing manual workload and improving document standardization.

[0022] In addition, the multimodal large model can perform intent recognition, task decomposition, and scheduling planning for complex instructions. That is, when receiving icing wind tunnel test tasks or real-time adjustment instructions, the multimodal large model, as the central control node, can first perform intent recognition and task decomposition on the input test task instructions or environmental monitoring data. The multimodal large model can determine the icing wind tunnel domain model to be called, the calling order, input and output requirements, and collaborative logic based on the constructed task graph, and generate standardized scheduling decision instructions. After receiving the feedback results from each domain model, it generates a decision based on all the feedback results.

[0023] It should be specifically noted that, in this embodiment, the method for generating decisions based on all feedback results can be as follows: The multimodal large model receives feedback results from multiple domain models. First, it completes semantic alignment and unified encoding through a preset multimodal fusion algorithm, transforming multiple feedback results into a unified feature vector, eliminating format differences, and establishing logical connections between the results. Then, the multimodal large model calls the icing wind tunnel industry knowledge base to verify the compliance and rationality of the fused results, checking whether they conform to industry standards and safety specifications, and judging the rationality of the results in conjunction with historical cases, providing a reference for decision-making. Next, based on the fusion results, knowledge base verification conclusions, and preset engineering logic rules, the multimodal large model comprehensively judges the experimental working condition, analyzes the causal relationship of each result, and weighs multiple factors such as safety, efficiency, and accuracy in conjunction with the experimental objectives to form a preliminary decision. Finally, the multimodal large model transforms the preliminary decision into executable instructions adapted to the experimental scenario, along with the decision basis.

[0024] In terms of implementation, considering the sensitivity and confidentiality requirements of wind tunnel test data, this multimodal large model adopts an open-source general-purpose large model base based on the Transformer architecture and is deployed privately on a local server. By fine-tuning instructions using professional corpora in the wind tunnel field within the intranet environment, the model achieves deep adaptation to domain-specific knowledge while ensuring that data does not leave the domain or leak.

[0025] The technical support layer 103 is used to engineer the integration of multi-source data and model capabilities, including multimodal fusion, visual-language alignment, data governance and knowledge base construction, data prediction agents, and perception and pattern recognition.

[0026] The business layer 104 is geared towards the final application output, providing functions such as dynamic early warning of safety anomalies, prediction of test results, intelligent Q&A, wind tunnel operation skills training, and expansion of fluid machinery platforms. This reduces the trial and error costs of test design and parameter tuning, improves test efficiency and accuracy, enhances on-site safety management and remote supervision capabilities, and supports migration and expansion on different test platforms.

[0027] This invention provides a multimodal large-model driven intelligent agent system for icing wind tunnels. Using a multimodal large-model as the central control unit, it replaces traditional manual decision-making and distributed model scheduling. Leveraging its powerful cross-modal understanding, semantic reasoning, and task planning capabilities, it uniformly analyzes multi-source target data from the data layer, including images, numerical values, and text. It dynamically schedules various specialized models in the icing wind tunnel domain as execution units, forming an intelligent agent collaborative architecture of central decision-making and domain execution. As a result, the system can autonomously complete task decomposition, model invocation, data interaction, and result fusion in the experimental process, improving the automation and intelligence of icing wind tunnel experiments, reducing manual intervention, and increasing experimental efficiency and result reliability. Furthermore, the technical support layer enables standardized interaction between data and models, while the business layer provides an intuitive human-machine interface, further ensuring the system's operability and scalability in complex icing wind tunnel experimental scenarios. As an optional implementation method, multi-source target data includes wind tunnel test site monitoring data, wind tunnel test environmental parameters, and equipment status; various icing wind tunnel domain models include safety supervision visual models, aerodynamic flow field proxy models, and icing and anti-icing numerical models. A visual model for safety supervision is used to identify safety risks based on on-site monitoring data from wind tunnel tests. An aerodynamic flow field proxy model is used to predict the aerodynamic flow field characteristics inside a wind tunnel based on wind tunnel test environment parameters and equipment status. A numerical model for icing and de-icing is used to predict the icing process based on environmental parameters and aerodynamic flow field characteristics output by an aerodynamic flow field proxy model.

[0028] For example, the safety monitoring visual model focuses on the safety of personnel, equipment, and materials at the wind tunnel testing site. By performing target detection and pattern recognition on test data, equipment status, and personnel behavior, it monitors potential safety risks in real time and issues early warning signals. This model incorporates computer vision algorithms, feeding risk categories and images back to a central large-scale model, allowing the intelligent agent to automatically adjust monitoring strategies based on changes in operating conditions, ensuring safe and stable operation during wind tunnel testing. Specifically, for high-frequency risks at wind tunnel testing sites such as "personnel violations, messy electrical wiring, slippery ground due to water accumulation, and smoke and fire hazards," the safety monitoring visual model uses YOLO-based target detection to analyze the video stream in real time. First, it identifies personnel and their key protective elements (such as safety helmets and work clothes), temporary power supply elements such as cables or power strips, abnormal smoke targets, and slippery ground areas, outputting target categories, locations, and confidence levels. Based on this, the system combines target tracking with preset area constraint rules to perform cross-frame pattern recognition. For example, when personnel enter a restricted area or have inadequate protection for more than a set duration, it is considered a violation; when cables are piled up in a passageway and remain there, it is considered a tripping hazard; when a water accumulation area is located in a passageway and its area exceeds a threshold, a slippery ground warning is triggered; and when smoke persists and shows a spreading trend, a fire warning is triggered. These risk events are then fed back to the central big data model in a structured form (including risk category, warning level, video keyframe evidence, and risk triggering rules). The central big data model then generates handling suggestions based on the test situation and operating procedures.

[0029] The aerodynamic flow field proxy model is responsible for rapidly predicting the aerodynamic flow field inside the wind tunnel based on inputs such as environmental parameters and equipment status, replacing expensive traditional CFD calculations and achieving real-time digital twins. These prediction results are crucial for the design of test schemes, equipment debugging and optimization, and the analysis of icing and de-icing effects. This model can not only predict the airflow distribution in the wind tunnel but also further predict the aerodynamic characteristics of the test specimen, such as lift and drag, pressure distribution, and flow field uniformity. In icing tests, the aerodynamic flow field model can compare the changes in aerodynamic characteristics before and after icing, helping researchers analyze the impact of icing on the aerodynamic performance of the test specimen and providing a scientific basis for the design of de-icing strategies. Specifically, the model takes as input the inflow velocity or inlet total pressure, total temperature, fan speed, valve opening, test section boundary conditions, and the operating status of subsystems such as spraying or cooling, and can combine the geometric and constraint information of the wind tunnel test section for conditional encoding. The model uses convolutional neural networks or neural operators to learn the mapping relationship between operating conditions and flow field distribution, outputting velocity field, pressure field, and necessary turbulence statistics within seconds.

[0030] The main task of the numerical model for icing and de-icing is to predict the icing process based on environmental parameters and guide the design of de-icing systems. Based on icing mechanisms and flow field information, the model calculates various variables such as temperature, airflow velocity, and liquid water content to predict the shape and growth rate of ice under different experimental conditions. Simultaneously, it incorporates existing de-icing technologies, such as heating and electrical pulses, to simulate their effectiveness in removing ice. These predictions can effectively guide the selection of de-icing strategies in actual experiments and improve the efficiency and safety of icing tests. Specifically, a Transformer-based neural network structure is adopted, with environmental parameters as input conditions, including temperature, velocity, liquid water content, and average volume diameter. It can further integrate wall-related features (such as local velocity, pressure gradient, and droplet collection coefficient distribution) output by the aerodynamic flow field proxy model. The state of the test specimen surface along the chord or spanwise direction is represented as a sequence or mesh embedded unit. Through a self-attention mechanism, the non-local correlation of icing in space and the evolution law in time are captured. Thus, without the need for complex multiphysics iterative calculations, the geometric parameters and growth trend of ice shape, such as ice thickness distribution, leading edge ice corner position, and surface roughness changes of the test specimen, can be quickly output.

[0031] The overall architecture of the three models mentioned above and the multimodal large model is as follows: Figure 3 As shown, the multimodal large model, the safety supervision visual model, the aerodynamic flow field proxy model, and the icing and anti-icing numerical model are called through functions.

[0032] This invention provides a multimodal, large-model-driven intelligent system for icing wind tunnels, achieving full-process coverage of on-site monitoring, flow field prediction, and icing process simulation. By defining the correlation between flow field feature inputs for icing and anti-icing / de-icing numerical models through a defined aerodynamic flow field proxy model, a professional data linkage link between domain models is constructed. This avoids deviations in results from independent model calculations, improves the accuracy of icing process prediction, and makes the acquisition and retrieval of multi-source data more aligned with the engineering realities of icing wind tunnel experiments, reducing the cost of processing invalid data. As an optional implementation, the technical support layer is configured to include: The multimodal fusion module is used to integrate the target dataset in the data layer, build an industry knowledge base for icing wind tunnels, and align the multimodal data in the wind tunnel test process through a multimodal fusion algorithm. The retrieval enhancement generation module is used to automatically retrieve standard specifications and historical cases from the icing wind tunnel industry knowledge base when generating decisions or answers in the multimodal large model at the model layer, and input the retrieval results as context into the multimodal large model; The model scheduling module is used for decision-making based on the output of a multimodal large model. It uses a function call interface to generate standardized calling instructions to drive various icing wind tunnel domain models in the model layer to complete rapid predictions.

[0033] For example, the technical support layer tightly integrates the basic data of the data layer and the intelligent reasoning capabilities of the model layer. Through multimodal fusion, intelligent perception and pattern recognition and other technical means, it ensures that the icing wind tunnel intelligent agent system can operate efficiently and accurately in actual experiments, and provides continuous data feedback and self-learning capabilities for model optimization and upgrading.

[0034] At the data layer, raw data from wind tunnel tests, including sensor data, video images, and test logs, have been standardized and cleaned into a unified format. The multimodal fusion module in the technical support layer further integrates heterogeneous data from the data layer, such as images, text, and knowledge graphs, to establish an industry knowledge base. This allows the data to provide rich input for subsequent model training, such as... Figure 4 As shown, the industry knowledge base contains an ice-free knowledge graph and experimental data, which form a closed loop with the data flow of the model layer. The industry knowledge base serves as one of the inputs to the model layer, and the output of the model layer enters the industry database, which further enables the system to have the ability to self-evolve and realize two-way verification between virtual simulation and physical experiments.

[0035] In one specific embodiment, the multimodal fusion module can integrate image and video data from wind tunnel tests with text data using multimodal fusion techniques such as visual-language alignment. This provides accurate information for aerodynamic flow field prediction and icing state analysis, improving the model's generalization ability to unknown data. For example, in icing tests, the technical support layer can extract icing surface features from monitoring video images and combine them with the anti-icing and de-icing model to provide the optimal de-icing method based on icing characteristics.

[0036] Meanwhile, the technical support layer introduces a retrieval-enhanced generation module. When the central large model generates decisions or answers, the system automatically retrieves standard specifications and historical cases from the knowledge base, using the search results as context input to the model. This minimizes the "illusion" phenomenon and ensures the professionalism and compliance of the generated content. For example, the technical support layer can utilize industry knowledge bases to provide icing and de-icing models with background knowledge such as icing mechanisms and flow field characteristics, making the ice shape prediction results more accurate and consistent with practical engineering applications.

[0037] At the model layer, the multimodal large model of the icing wind tunnel intelligent agent system completes inference analysis, task decomposition, and scheduling decisions based on the input real-time data, clarifying the professional models to be called and the execution requirements. The model scheduling module in the technical support layer receives the scheduling decision instructions output by the multimodal large model, transforms them into standardized function call instructions, completes interface adaptation, data transmission, and model triggering, drives professional models in various fields to complete rapid prediction, and enables the system to respond to changes in measurement data in real time.

[0038] This invention provides a multimodal large-model driven intelligent agent system for icing wind tunnels. The multimodal fusion module integrates and aligns heterogeneous data from multiple sources, and the constructed industry knowledge base provides unified domain knowledge support for the entire system, solving the problems of scattered data and lack of professional knowledge support in traditional experimental systems. The retrieval enhancement generation module provides context for the multimodal large model through real-time retrieval from the knowledge base, fundamentally reducing the knowledge illusion of the large model and ensuring the professionalism and compliance of decision-making and reasoning. The model scheduling module relies on function call interfaces to achieve standardized driving of domain models, replacing manual task decomposition and model scheduling, improving the response speed of experimental needs and the accuracy of model calls, and realizing efficient collaborative linkage of multiple models. As an optional implementation, the technical support layer is configured to further include: The status monitoring module is used to monitor the status of the icing wind tunnel test through deep learning and pattern recognition technology, perform pattern analysis on sensor data to identify abnormal operating conditions, and provide real-time feedback to the corresponding domain model in the model layer. The post-processing module is used to process the output results of the model layer and generate decision information suitable for use by operators of the icing wind tunnel.

[0039] For example, the status monitoring module is used to monitor the status of the entire wind tunnel test process in real time by matching the multi-source heterogeneous sensor data of the icing wind tunnel test with the working condition feature library in the icing knowledge graph through a pre-trained deep learning anomaly detection model. The deep learning anomaly detection model can be a Transformer-based temporal anomaly detection model or a CNN-LSTM fusion model.

[0040] Specifically, the status monitoring module of the technical support layer, addressing the temporal and multi-dimensional coupling characteristics of multi-source sensor data in icing wind tunnel tests, employs a CNN-LSTM fusion model to extract features from time-series data such as pressure, temperature, and wind speed. Simultaneously, a pre-trained visual anomaly detection model identifies ice shape anomalies and personnel violations in the wind tunnel monitoring video. The two types of identification results are matched in real time with the features of historical normal and historical fault conditions stored in the icing knowledge graph to achieve accurate identification and early warning of abnormal conditions. When anomalies such as flow field distortion and excessive ice shape are identified, the module pushes the abnormal data and risk level to the safety supervision visual model in real time for secondary verification. At the same time, it sends an alarm to the central multimodal large model, which calls the aerodynamic flow field proxy model and the icing numerical model to perform emergency simulation and generate a temporary test adjustment plan to ensure the safety and stability of the entire test process.

[0041] The post-processing module is used to normalize, fuse, and visualize the output results of models in various fields of the model layer. Combined with the test specifications and operating standards in the industry knowledge base, it generates structured decision information suitable for icing wind tunnel operators, including: real-time adjustment suggestions for test parameters, safety risk classification and early warning, automatic recording reports of the test process, and correlation analysis reports of ice shape and aerodynamic performance, and pushes the decision information to the human-computer interaction terminal.

[0042] Specifically, the post-processing module of the technical support layer performs data alignment, confidence level fusion, and format standardization on the multi-source heterogeneous results output by the model layer, such as ice shape detection results from the visual model, ice thickness simulation data from the numerical model, and aerodynamic performance calculation results from the surrogate model. Combined with the icing wind tunnel test specifications in the industry knowledge base, it automatically generates decision information that conforms to the user habits of operators. For example, it generates actionable adjustment suggestions for ice shape anomalies during the test; and it automatically generates a complete test report containing test parameters, ice shape evolution, aerodynamic performance changes, and safety warning records throughout the entire test, which greatly reduces the workload of data analysis and report writing for operators and improves test efficiency and standardization.

[0043] Meanwhile, the abnormal operating condition data collected by the status monitoring module and the decision information generated by the result post-processing module will be synchronously fed back to the industry knowledge base to update the operating condition feature library in the icing knowledge graph and optimize the training data of the anomaly detection model, thereby further improving the accuracy of status monitoring and the precision of decision information.

[0044] This invention provides a multimodal large model-driven intelligent agent system for icing wind tunnels. The status monitoring module uses deep learning and pattern recognition technologies to achieve real-time monitoring of the test status and accurate identification of abnormal conditions. Combined with real-time feedback to the domain model, it enables rapid response and collaborative handling of abnormal conditions, improving the safety and stability of the icing wind tunnel test process and solving the problems of lagging identification and untimely handling in traditional manual monitoring. The result post-processing module standardizes and visualizes the multi-source heterogeneous results output by the model layer, generating decision information suitable for operators. It transforms the professional calculation results of the model into operational suggestions and early warning information that can be directly executed in engineering, reducing the data analysis threshold for operators. As an optional implementation, the service layer is configured to include: The suggestion module is used to provide users with decision-making suggestions for icing wind tunnel tests based on user input and through interaction with the model layer; The safety monitoring module is used to provide early warnings of potential safety hazards based on the real-time perception capabilities provided by the technical support layer. The data archiving and management module is used to uniformly archive and manage all test data, prediction results, operation records, and risk analysis information during the icing wind tunnel test.

[0045] For example, the business layer is the top layer of the icing wind tunnel intelligent agent system. It is responsible for transforming the capabilities of the technical support layer and the model layer into practical applications, directly providing operational intelligent decision support for wind tunnel test operators. The main function of this layer is to ensure that the model's output can be effectively applied in wind tunnel tests and seamlessly integrated with the actual workflow, thereby improving the efficiency, safety, and intelligence level of the test.

[0046] The business layer integrates multiple functions such as intelligent question answering, test prediction, and safety monitoring to help operators quickly obtain key data and information during icing wind tunnel tests. The business layer includes a suggestion module; during the test design phase, it can interact with the model layer to provide researchers with intelligent suggestions and decision support, optimizing the development of test plans and test outlines.

[0047] In the actual operation of wind tunnel testing, the safety supervision function of the business layer is particularly important. Therefore, a safety supervision module is also configured at the business layer. Combined with the real-time perception capabilities provided by the technical support layer, the business layer can monitor the status of personnel, equipment, and materials at the test site in real time. By integrating safety protection measures such as access control, tool abandonment detection, fire detection, smoke detection, and personnel entry and exit monitoring at the wind tunnel test site, the system can promptly detect safety hazards and issue early warnings. Whether it is an abnormality in test equipment, a violation of personnel operation, or an abnormal state of materials, the business layer can provide real-time feedback through intelligent means to ensure the safety and stability of the test process.

[0048] Furthermore, the business layer can combine the model layer's predictions with real-time data to assist operators in dynamically adjusting operating conditions and equipment parameters during testing, ensuring the test proceeds smoothly according to the predetermined objectives. In case of abnormal situations, the business layer will issue timely warnings and adjust the test plan based on model recommendations to avoid potential risks.

[0049] The business layer also includes a data archiving and management module. During the icing wind tunnel test, all test data, prediction results, operation records, risk analyses, and other information will be archived and managed by the business layer. This data will not only be used for the analysis and decision support of the current test, but will also serve as a reference for future test design and improvement.

[0050] This invention provides a multimodal, large-model-driven intelligent agent system for icing wind tunnels. The suggestion module provides customized experimental decision suggestions based on user input and interaction with the model layer, offering intelligent support for experimental scheme design and operating condition adjustments, optimizing the experimental design process, and reducing trial-and-error costs. The safety monitoring module relies on the real-time perception capabilities of the technical support layer to achieve accurate early warning of safety hazards, transforming the system's technical monitoring capabilities into actual experimental safety management capabilities, comprehensively ensuring the safety of personnel, equipment, and the experimental process at the test site. The data archiving and management module achieves unified archiving and management of data and information throughout the entire experimental process, ensuring not only the integrity and traceability of experimental data but also providing rich historical data support for subsequent experimental design, model optimization, and knowledge graph updates. Finally, as Figure 5 The diagram shows the overall architecture of the multimodal large-scale model-driven icing wind tunnel intelligent agent system. The data layer acquires various types of data, including general knowledge of disciplines, industry papers, patents, operating procedures, standards, and experimental operation data. Within the model layer, the multimodal large-scale model serves as the central hub, while icing wind tunnel domain models such as the safety monitoring visual model, the initiation flow field proxy model, and the icing and anti-icing numerical model serve as expert tools, i.e., execution units. The technical support layer mainly completes multimodal fusion and retrieval enhancement generation, using function call interfaces to generate standardized calling instructions, driving various icing wind tunnel domain models in the model layer to complete rapid predictions. The business layer is the top layer of the system, responsible for transforming the capabilities of the technical support layer and the model layer into practical applications, directly providing operable intelligent decision support for wind tunnel test operators. The main function of this layer is to ensure that the model's output can be effectively applied in wind tunnel tests and seamlessly integrated with the actual workflow, improving the efficiency, safety, and intelligence level of the test.

[0051] The icing wind tunnel intelligent agent system can provide important functional support such as intelligent question answering, operation guidance, terminal control, safety supervision, and analysis and decision-making before, during and after the test.

[0052] Before the test, the central big data model can automatically generate test plans and outlines based on professional knowledge, operating standards, and historical test data, providing operators with suggestions and guidance for test design. This includes selecting appropriate icing parameters, controlling equipment parameters, optimizing test logs, and determining suitable anti-icing and de-icing strategies. Furthermore, the central big data model possesses powerful text understanding and generation capabilities. Based on professional knowledge in the field of icing and wind tunnel testing technology, it can provide operators with real-time, accurate intelligent question-and-answer services, offering immediate information support for icing wind tunnel tests.

[0053] During the experiment, the central large-scale model generates equipment control command suggestions through real-time data input and inference. After safety verification or operator confirmation, it dynamically adjusts the test conditions and equipment parameters to ensure the experiment proceeds smoothly according to the predetermined objectives and achieves automation. The system can monitor changes in the test environment in real time and adjust the test settings according to changes in the aerodynamic flow field in the wind tunnel, the icing status of the test specimen, and the equipment status to ensure the accuracy and stability of the test process. If an anomaly occurs, the central large-scale model can issue timely warnings, provide safety feedback, and assist operators in making adjustments.

[0054] After the experiment, the intelligent agent system can not only automatically generate an experimental report, but also perform data analysis and summary of the experimental process. The central large-scale model will conduct further analysis on icing morphology and flow field changes based on the experimental data and prediction results, providing researchers with detailed analysis reports on icing mechanisms and anti-icing / de-icing effects, helping to optimize experimental plans and icing / de-icing technologies. Furthermore, the intelligent agent system can identify potential problems in the experiment and provide improvement suggestions for future experiments.

[0055] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.

Claims

1. A multimodal large-model driven intelligent agent system for icing wind tunnels, characterized in that, include: The data layer is used to acquire multi-source target data during the icing wind tunnel test, process the multi-source target data, and construct the processed target dataset. The model layer adopts a collaborative architecture of intelligent agents, with a multimodal large model as the central control and multiple icing wind tunnel domain models as execution units. The multimodal large model uniformly completes task scheduling and decision output, while each type of icing wind tunnel domain model is responsible for the computational tasks in its corresponding domain. The technical support layer is used to integrate and process the target dataset during the icing wind tunnel test, enabling interactive calls between the data layer and the model layer; The business layer is used to interact with the user based on the output of the model layer.

2. The multimodal large-model driven icing wind tunnel intelligent agent system according to claim 1, characterized in that, Multi-source target data includes wind tunnel test site monitoring data, wind tunnel test environmental parameters, and equipment status; various icing wind tunnel domain models include safety supervision visual models, aerodynamic flow field proxy models, and icing and anti-icing / de-icing numerical models. A visual model for safety supervision is used to identify safety risks based on on-site monitoring data from wind tunnel tests. An aerodynamic flow field proxy model is used to predict the aerodynamic flow field characteristics inside a wind tunnel based on wind tunnel test environment parameters and equipment status. A numerical model for icing and de-icing is used to predict the icing process based on environmental parameters and aerodynamic flow field characteristics output by an aerodynamic flow field proxy model.

3. The multimodal large model-driven icing wind tunnel intelligent agent system according to claim 1, characterized in that, The technical support layer is configured to include: The multimodal fusion module is used to integrate the target dataset in the data layer, build an industry knowledge base for icing wind tunnels, and align the multimodal data in the wind tunnel test process through a multimodal fusion algorithm. The retrieval enhancement generation module is used to automatically retrieve standard specifications and historical cases from the icing wind tunnel industry knowledge base when generating decisions or answers in the multimodal large model at the model layer, and input the retrieval results as context into the multimodal large model; The model scheduling module is used for decision-making based on the output of a multimodal large model. It uses a function call interface to generate standardized calling instructions to drive various icing wind tunnel domain models in the model layer to complete rapid predictions.

4. The multimodal large model-driven icing wind tunnel intelligent agent system according to claim 3, characterized in that, The technical support layer is configured to also include: The status monitoring module is used to monitor the status of the icing wind tunnel test through deep learning and pattern recognition technology, perform pattern analysis on sensor data to identify abnormal operating conditions, and provide real-time feedback to the corresponding domain model in the model layer. The post-processing module is used to process the output results of the model layer and generate decision information suitable for use by operators of the icing wind tunnel.

5. The multimodal large model-driven icing wind tunnel intelligent agent system according to claim 1, characterized in that, The business layer is configured to include: The suggestion module is used to provide users with decision-making suggestions for icing wind tunnel tests based on user input and through interaction with the model layer; The safety monitoring module is used to provide early warnings of potential safety hazards based on the real-time perception capabilities provided by the technical support layer. The data archiving and management module is used to uniformly archive and manage all test data, prediction results, operation records, and risk analysis information during the icing wind tunnel test.

6. A multimodal large-model driven icing wind tunnel intelligent agent system according to any one of claims 1-5, characterized in that, The data layer pre-stores a knowledge graph built based on knowledge of the icing domain.

Images