Search

CN-122022523-A - Fire disaster simulation and decision method and system based on multi-mode large model

CN122022523ACN 122022523 ACN122022523 ACN 122022523ACN-122022523-A

Abstract

The invention relates to the field of artificial intelligence application, and discloses a fire disaster simulation and decision system and method based on a multi-mode large model, which can integrate multi-source heterogeneous data, have intelligent reasoning capacity special for the field and support dynamic decision optimization. The method comprises the steps of collecting heterogeneous data and preprocessing, providing multi-mode fused structured data for subsequent fire perception and analysis, inputting the multi-mode fused structured data into a large language model subjected to forest fire prevention field instruction fine adjustment, performing forest fire prevention special instruction adjustment, combining the current multi-mode fused data after the large language model receives the adjusted instruction, dynamically outputting a fire evolution prediction sequence, simulating fire extinguishing strategies based on fire prediction results, evaluating and sequencing each fire extinguishing strategy, generating executable auxiliary decision suggestions, and generating an optimal scheme based on the auxiliary decision suggestions.

Inventors

  • WANG QIANG

Assignees

  • 天翼数字生活科技有限公司

Dates

Publication Date
20260512
Application Date
20260202

Claims (11)

  1. 1. A fire disaster simulation and decision method based on a multi-mode large model is characterized by comprising the following steps: step 1, collecting heterogeneous data and preprocessing, and providing multi-mode fused structured data for subsequent fire sensing and analysis; Step 2, inputting the multi-mode fused structured data into a large language model subjected to forest fire prevention field instruction fine adjustment, and performing forest fire prevention special instruction fine adjustment; step 3, after receiving the optimized instruction, the large language model dynamically outputs a fire evolution prediction sequence by combining current multi-mode fusion data; Step 4, based on the fire prediction result generated in the step 3, commanding a tuned large language model to simulate fire extinguishing strategies, and evaluating and sequencing each fire extinguishing strategy to generate executable auxiliary decision-making suggestions; And 5, generating an optimal scheme based on the auxiliary decision suggestion in the step 4.
  2. 2. The fire disaster simulation and decision method based on the multi-mode big model according to claim 1 is characterized in that the heterogeneous data comprise satellite thermal infrared image data, meteorological data, topographic data, unmanned aerial vehicle video streams and social media fire information, the preprocessing comprises the steps of carrying out fusion processing on all collected heterogeneous data to be unified to the same space-time coordinate system, carrying out unified resolution processing on image data in the heterogeneous data, and carrying out word segmentation and vectorization processing on text data in the heterogeneous data.
  3. 3. The multi-modal large model based fire simulation and decision making method of claim 2 wherein the large language model is a CLIP model.
  4. 4. A multi-modal large model based fire simulation and decision method as claimed in claim 3 wherein step 3 further comprises: step 3-1, extracting and fusing characteristics through the CLIP large model; step 3-2, dynamically predicting fire spread simulation based on the CLIP large model; wherein step 3-1 further comprises: Step 3-1-1, extracting the characteristics of 5 types of data sources to generate multi-mode characteristic vectors, wherein the method comprises the steps of extracting the characteristics of satellite thermal infrared image data and unmanned aerial vehicle video streams by using an image encoder of a CLIP, and extracting the characteristics of meteorological data, topographic data and social media fire information by using a text encoder of the CLIP; Step 3-1-2, performing feature fusion on the extracted multi-modal feature vectors through a cross-modal fusion mechanism of the CLIP to generate fused multi-modal feature vectors, wherein the fused multi-modal feature vectors comprise information of fire point positions, fire intensity, spreading directions and environmental conditions; wherein step 3-2 further comprises: step 3-2-1, inputting the fused multi-mode feature vector generated in the step 3-1 and the optimized instruction as prompt words into a CLIP model; step 3-2, the CLIP model combines the information contained in the fused multi-mode feature vector, and reasoning the prediction of the fire spreading track is carried out according to the prompt words; step 3-2-3, the CLIP model outputs a reasoning simulation result, and a basis is provided for the simulation of the suppression scheme; And 3-2-4, verifying the simulation result of the reasoning by the CLIP model.
  5. 5. The fire simulation and decision method based on the multi-modal large model as claimed in claim 4, wherein the feature vector of each mode is further subjected to dynamic weight adjustment in the process of feature fusion of the multi-modal feature vector.
  6. 6. The method for simulating and deciding fire based on a large multi-modal model as set forth in claim 1 further comprising dynamic decision optimization, triggering re-execution of steps 1 to 5 upon detection of a change in fire status to dynamically update the fire prediction and optimal fire extinguishing scheme.
  7. 7. A fire simulation and decision system based on a multi-modal large model, the system comprising: The multi-mode data fusion module is configured to acquire heterogeneous data and perform preprocessing, and provide multi-mode fused structured data for subsequent fire sensing and analysis; the instruction tuning module is configured to input the multi-mode fused structured data into a large language model subjected to instruction fine tuning in the forest fire prevention field to perform special instruction tuning for forest fire prevention; The dynamic fire evolution prediction module is configured to dynamically output a fire evolution prediction sequence by combining current multi-mode fusion data after the large language model receives the optimized instruction; The multi-strategy fire extinguishing simulation and evaluation module is configured to instruct the optimized large language model to simulate fire extinguishing strategies based on the fire prediction result generated by the dynamic fire evolution prediction module, evaluate and sort each fire extinguishing strategy and generate executable auxiliary decision suggestions; And the optimal extinguishing scheme generating module is used for generating an optimal scheme based on the auxiliary decision proposal of the multi-strategy fire extinguishing simulation and evaluation module.
  8. 8. The multi-modal large model based fire simulation and decision making system of claim 7 wherein the system further comprises: And the dynamic decision optimization module is configured to trigger the multi-mode data fusion module to be re-executed to the optimal extinguishing scheme generation module after detecting the change of the fire scene state so as to dynamically update the fire prediction and the optimal extinguishing scheme.
  9. 9. The electronic equipment is characterized by comprising a controller, wherein the controller comprises a processor, a communication interface, a memory and a communication bus, and the processor, the communication interface and the memory are communicated with each other through the communication bus; a memory for storing a computer program; a processor for implementing the steps of the multi-modal large model-based fire simulation and decision method as claimed in any one of claims 1 to 6 when executing a program stored on a memory.
  10. 10. A computer readable storage medium, on which a computer program is stored, which computer program, when being executed by a processor, implements the steps of the fire simulation and decision method based on a multimodal big model as claimed in any of the claims 1-6.
  11. 11. A computer program product, characterized in that the program when executed by a processor implements the steps of the fire simulation and decision method based on a multimodal big model as claimed in any of the claims 1-6.

Description

Fire disaster simulation and decision method and system based on multi-mode large model Technical Field The invention relates to the technical field of artificial intelligence and fire prevention and control, in particular to a fire intelligent sensing, dynamic prediction and auxiliary decision making technology based on multi-mode data fusion and a large language model, which is applicable to application scenes of forest fire prevention monitoring, emergency response command and disaster risk intelligent management. Background At present, a forest fire monitoring and early warning system mainly relies on satellite remote sensing, a ground weather observation station and limited video monitoring equipment to perform fire sensing and risk assessment. The basic implementation principle of the system is that ground surface heat abnormal point data are acquired through meteorological satellites (such as MODIS, VIIRS and the like), and the possibility and potential spreading trend of fire occurrence are primarily judged by utilizing an empirical model (such as Canadian fire weather index system FWI) or a simplified physical spreading model in combination with meteorological parameters such as temperature, humidity, wind speed, precipitation and the like acquired by a ground meteorological station and with terrain data such as a Digital Elevation Model (DEM) and the like. On the basis, part of the system is introduced with a fixed camera or an infrared sensor to monitor the video of a local area, and an alarm is triggered when smoke or open fire is detected. In recent years, there have also been studies attempting to apply a machine learning method (such as a support vector machine, random forest, etc.) to analysis of historical fire data to improve accuracy of fire level prediction. However, these systems still have significant limitations in practical applications. Firstly, the data source on which the existing system depends is single, the data source is mainly concentrated on satellite thermal infrared images and static weather site data, and the fusion capability of dynamic and multidimensional information such as real-time aerial video of an unmanned aerial vehicle and information reported by social media users is lacked. The system has insufficient sensitivity for identifying the initial fire, and is easy to report when the cloud layer shields the satellite vision or the fire scale is small. Secondly, most of the current algorithm models for fire spread simulation and risk assessment are general models, and depth optimization is not performed on a specific application scene of forest fire prevention. For example, most systems employ a fixed spread rate formula, which is difficult to accurately reflect the effects of complex terrain, vegetation type changes, and sudden weather conditions (e.g., gusts) on fire behavior. At the same time, these models often do not have the ability to understand natural language instructions or generate structured decision advice, and cannot effectively support the tactical formulation of commanders. Third, the existing system functions are mainly concentrated on a monitoring-alarming link, and intelligent auxiliary decision-making capability for scheduling fire extinguishing resources, comparing and selecting a fire extinguishing scheme and dynamically adjusting the fire extinguishing scheme is lacked. Even if part of the system can output a fire risk map, only static risk levels are often provided, and executable coping strategies cannot be generated according to the real-time evolution condition of the fire scene. Finally, the response mechanism of the existing system is relatively slow, the update period is long (usually several hours), and the closed loop optimization capability is not provided. Once the fire scene environment is suddenly changed (such as wind direction suddenly changed and new fire points appear), the system is difficult to recalculate and push the updated prediction result and the corresponding advice in a short time, so that decision delay is caused, and the optimal putting out window is missed. In summary, the existing forest fire monitoring and early warning system has the key defects that (1) the data source is single, multi-mode real-time information is not fused, (2) the core algorithm lacks suitability for a forest fire prevention professional scene, prediction accuracy and practicality are insufficient, (3) the system only has a basic alarm function and cannot provide a dynamic and executable fire extinguishing auxiliary decision, and (4) the system response is slow, and the real-time updating and closed loop optimizing capability is lacking. These drawbacks directly restrict the implementation of early discovery, accurate prediction and efficient suppression of forest fires. Disclosure of Invention The following presents a simplified summary of one or more aspects in order to provide a basic understanding of such aspects. This summary is not an