CN-121981269-A - Dynamic plan generating system driven by multi-mode large model in emergency management scene

CN121981269ACN 121981269 ACN121981269 ACN 121981269ACN-121981269-A

Abstract

The invention relates to the technical field of emergency management, in particular to a dynamic plan generating system driven by a multi-mode large model in an emergency management scene. The method comprises a static knowledge layer, a history retrieval layer, a static plan generation layer, a field data processing layer, a dynamic plan generation layer and a real-time evaluation feedback layer, wherein the static knowledge layer is used for constructing structural history data based on a pre-training large model, the history retrieval layer is used for carrying out semantic vector retrieval on the structural history data to output similar emergency event data, the static plan generation layer is used for receiving the similar emergency event data to generate a structural plan, the field data processing layer is used for collecting and analyzing field multi-mode data in real time to generate field multi-mode processing information, the dynamic plan generation layer is used for outputting a preliminary dynamic plan according to the structural plan and the field multi-mode processing information to output a final dynamic plan after the structural processing, and the real-time evaluation feedback layer is used for receiving the final dynamic plan and outputting execution effect data to the static knowledge layer and the dynamic plan generation layer. Through a six-layer architecture, the full-flow plan generation from multiplexing historical data to bottom dynamic optimization and then to execution effect closed-loop optimization is realized.

Inventors

XIE QIWEI
DUAN NA
FENG HAONAN
ZHANG XILE
WANG DENGDUO
LI CHANGFENG

Assignees

公安部第三研究所

Dates

Publication Date: 20260505
Application Date: 20260122

Claims (10)

1. The dynamic plan generating system driven by the multi-mode large model in the emergency management scene is characterized by comprising the following components: A static knowledge layer (1) for constructing structured historical data based on the pre-trained large model; The history retrieval layer (2) is connected with the static knowledge layer (1) and is used for carrying out semantic vector retrieval on the structured history data to output similar emergency event data (G); a static plan generation layer (3) connected to the history retrieval layer (2) for receiving the similar emergency event data to generate a structured plan (J); the field data processing layer (4) is used for collecting and analyzing the field multi-mode data in real time to generate field multi-mode processing information; A dynamic plan generating layer (5) which is connected with the static plan generating layer (3) and the on-site data processing layer (4) and outputs a preliminary dynamic plan according to the structuring plan (J) and the on-site multi-mode processing information, and outputs a final dynamic plan after structuring processing; and the real-time evaluation feedback layer (6) is connected with the dynamic plan generation layer (5) and the static knowledge layer (1), receives the final dynamic plan and outputs execution effect data to the static knowledge layer (1) and the dynamic plan generation layer (5).
2. The multi-modal large model driven dynamic plan generation system in an emergency management scenario according to claim 1, wherein the static knowledge layer (1) comprises, A data cleansing unit (11) cleansing the historical emergency event data (A) to output cleansed structured historical data (B); The large model knowledge extraction unit (12) is connected with the output end of the data cleaning unit (11) and is used for directionally extracting core knowledge from the cleaned historical data (B) based on a pre-training large model to generate a structured knowledge item (C); And the static knowledge storage library (13) is connected with the large model knowledge extraction unit (12) through a structured data protocol, the structured knowledge item (C) is stored in association with the corresponding history plan, an update event stamp is set, and the static knowledge storage library (13) is connected with the history retrieval layer (2).
3. The multi-modal large model driven dynamic plan generation system in an emergency management scenario according to claim 1, wherein the history retrieval layer (2) comprises, The semantic vector extraction unit (21) receives new emergency event text data (D) which are input from the outside and marked with classification labels, loads a pre-trained BERT model or a localization large model, and converts the new emergency event text data (D) into semantic vectors (E); The emergency event classification label screening unit (22) is connected with the semantic vector extraction unit (21) through a vector transmission protocol, and screens all candidate emergency event libraries (F) with the same label from the static knowledge storage library (13) of the static knowledge layer (1) according to the classification labels; The vector retrieval and sorting unit (23) is connected with the database query interface of the emergency event classification label screening unit (22), the cosine similarity of the semantic vector (E) and each event vector in the candidate emergency event library (F) is calculated, the similar emergency event data (G) is output according to the cosine similarity, the similar emergency event data (G) comprises a similar emergency event text and corresponding similar plan data and is accompanied by a similarity score, and the output end of the vector retrieval and sorting unit (23) is connected with the static plan generation layer (3) through a structured data interface.
4. The multi-modal large model driven dynamic plan generation system in an emergency management scenario according to claim 1, wherein the static plan generation layer (3) comprises, A plan input unit (31) for performing format unification processing on similar plan data in the similar emergency event data (G) and outputting a standardized similar plan data set (H); The large model integration unit (32) is connected with the plan input unit (31), loads the pre-training large model, and generates a preliminary static plan (I) according to the standardized similar plan data set (H), the customized integration prompt word and the similar emergency event text in the similar emergency event data (G); the static plan output unit (33) is connected with the text processing interface of the large model integration unit (32) and is used for carrying out structural optimization on the preliminary static plan (I) and marking execution priority to obtain the structural plan (J), and the static plan output unit (33) is connected with the dynamic plan generation layer (5) and the static knowledge storage library (13) of the static knowledge layer (1).
5. The multi-modal large model driven dynamic plan generation system in an emergency management scenario according to claim 1, wherein the on-site data processing layer (4) comprises, The multi-source data docking unit (41) is connected with an external data source (S) to acquire geographic information, field video or pictures and convert the geographic information, the field video or the pictures into standardized data; The multi-mode large model information extraction unit (42) is connected with the multi-source data docking unit (41), loads a pre-training model, inputs the standardized data and extracts prompt words, and generates structural feature data; the key database association unit (43) is connected with the multi-mode large model information extraction unit (42) and used for comparing the structural feature data, outputting a comparison result and integrating the geographic information, the structural feature data and the comparison result; And a report generation and output unit (44) connected with the key database association unit (43) for integrating multi-source information, generating on-site multi-mode processing information and outputting the on-site multi-mode processing information to the dynamic plan generation layer (5).
6. The multi-modal large model driven dynamic plan generation system in an emergency management scenario according to claim 1, wherein the dynamic plan generation layer (5) comprises: a dynamic input unit (51) connected to the static plan generation layer (3) and the on-site data processing layer (4) for performing format alignment after receiving the structured plan (J) and the on-site multi-mode processing information; the large model generating unit (52) is connected with the dynamic input unit (51) and is used for triggering large model fusion to generate a flow path preliminary dynamic plan through customizing generation prompt words based on the pre-training large model; And the dynamic plan output unit (53) is connected with the large model generation unit (52), performs structuring treatment on the preliminary dynamic plan, classifies and presents core contents, synchronizes additional data to generate a time stamp and a risk level, outputs a final dynamic plan to the on-site disposal terminal, and transmits the final dynamic plan to the real-time evaluation feedback layer (6).
7. The multi-modal large model driven dynamic plan generation system in an emergency management scenario according to claim 1, wherein the real-time assessment feedback layer (6) comprises: an execution data acquisition unit (61) connected to a site treatment terminal (7), the execution data being acquired by the site treatment terminal (7); the effect evaluation unit (62) is connected with the execution data acquisition unit (61) and evaluates the execution effect according to a preset index to generate execution effect data; a feedback updating unit (63) updates the effective treatment cases of the execution effect data to a static knowledge repository (13) of the static knowledge layer (1).
8. The multi-modal large model driven dynamic plan generation system in an emergency management scenario of claim 2, wherein the historical emergency event data (a) includes a disposition feedback sheet and historical plan text; The data cleaning unit (11) filters redundant data in the historical emergency event data (A), corrects a data format and standardizes the data format, and outputs the cleaned structured historical data (B) after deficiency value complementation, wherein the deficiency value complementation comprises same-type event mean value filling and expert manual verification; The large model knowledge extraction unit (12) loads the pre-training large model, inputs customized prompt words, extracts core knowledge and generates a structured knowledge item (C), wherein the core knowledge comprises emergency strength configuration, equipment material list and multi-department collaboration logic; The static knowledge storage library (13) adopts a distributed database architecture, establishes a secondary index, stores the structured knowledge item (C) in association with a corresponding history plan, establishes an association mapping relation of a label-knowledge-history plan, and sets a data updating time stamp.
9. A multi-modal large model driven dynamic plan generation system in an emergency management scenario according to claim 3, characterized in that the vector dimension of the semantic vector (E) covers event impact scope, number of personnel involved, key risk source type; The semantic vector extraction unit (21) converts the new emergency event text data (D) into the semantic vector (E) and reserves event detail features through vector coding; the vector search ranking unit (23) calculates cosine similarity by the following formula: Cosine similarity = V_new.V_i- (||V/u) New| New (New) I Wherein V_new is the semantic vector (E), and V_new is the L2 norm of V_new; V _ i is the event vector and, i v_i is the L2 norm of V i; the vector search ranking unit (23) performs descending ranking according to the cosine similarity.
10. The multi-modal large-model driven dynamic plan generation system in an emergency management scenario according to claim 1, wherein the external data source (S) is provided by a data acquisition device comprising one or more of an image acquisition device, a smoke alarm and temperature sensor, a positioning terminal, a portable weather station, power switch status data, hydrant emergency status data.

Description

Dynamic plan generating system driven by multi-mode large model in emergency management scene Technical Field The invention relates to the technical field of emergency management, in particular to a dynamic plan generating system driven by a multi-mode large model. Background In emergency management work, timely, scientific and targeted emergency plans are critical to effectively cope with various emergencies. The common method for generating the plan in the current emergency management field mainly comprises the following two steps: Training an auxiliary decision model through the historical cases. The prior art discloses an emergency treatment decision information generation method, an emergency treatment decision information generation device, emergency treatment decision information generation equipment and a storage medium, which are characterized in that historical emergency treatment cases are collected and preprocessed to generate an emergency treatment case corpus, a model training set is built by using the corpus, an auxiliary decision model is obtained through training, entity relation extraction is carried out on pre-configured emergency plan information to build an emergency treatment knowledge graph, and finally, the emergency treatment decision information is determined by combining the auxiliary decision model and the emergency treatment knowledge graph according to description information of an event to be processed. The core is ‌ knowledge extraction and reasoning ‌, namely training a model through historical cases, and combining a knowledge graph to generate ‌ decision information. And performing similarity calculation by using the historical case library, and regenerating an emergency plan based on user input. The prior art discloses a large-model-based emergency plan generation method for sudden biological safety events, which comprises the steps of collecting related event information and an emergency plan generation training set, training by using a Transformer model as a basic structure to obtain an emergency plan generation model, then constructing a prompting word template for user input, constructing an emergency plan data vector library by adopting a text vectorization technology, finally calculating the similarity between user input and the content of the vector library, extracting data with highest correlation, combining real-time space information and user input, and generating a complete emergency plan by using the large model. The scheme focuses on ‌ searching and generating, finds the most relevant historical plan fragments by calculating the similarity between user input and a historical plan library, and then directly generates a complete targeted emergency plan ‌ ‌ by utilizing the generating capacity of a large model and combining real-time information ‌. The main core ideas of the methods are that text data and structured data are mainly analyzed to construct an emergency plan library, and the following technical problems exist: first, there is a problem of static knowledge dependence and real-time response lag. In the technical aspect, the method for generating the plan in the existing emergency management scene generally highly depends on a historical case library, a knowledge graph or a pre-training model, and the knowledge library and the model have static properties in nature and cannot capture and reflect dynamic changes of an emergency scene in real time, so that the adjustment of the plan is delayed from actual demands. In addition, the existing method lacks effective fusion capability for multi-mode real-time data, and dynamic information acquired on an emergency site cannot be incorporated into a plan decision flow, so that decision basis is limited to static historical data only. Taking the dangerous chemical factory leakage event as an example, the diffusion range of leakage substances cannot be identified through on-site real-time video data, and the personnel evacuation area and the warning range cannot be dynamically adjusted by combining the wind direction and the topographic data acquired in real time, so that the generated plan adjustment lags behind the on-site actual disposal requirement, and the emergency response efficiency is affected. From the data layer analysis, the update period of the historical case library is longer and is limited by the data acquisition range and the scene coverage, so that all possible event variants in the emergency management scene are difficult to cover, the generalization capability of the existing plan generation model is limited, and the existing plan generation model cannot be adapted to diversified emergency scene. Secondly, the problem of insufficient multi-mode data fusion exists. The traditional plan generating method takes text data processing as a core, can only analyze text data such as event description text, history plan text and the like, has obvious limitation on the processing capacity of non-text data