Multi-agent collaborative urban event handling system based on large model

Abstract

The invention discloses a multi-agent cooperative urban event handling system based on a large model, relates to the technical field of artificial intelligence, and solves the problems of tedious event task handling process and repeated dispatching of event tasks. The urban event disposal system comprises an intelligent agent module and a collaboration module. The agent module comprises an event reporting agent, an event acceptance agent, an event distribution agent, an event handling agent, an event checking agent and an event analysis agent; performing pre-recommendation of event handling departments on the event semantic vectors through a distribution reasoning model to generate a candidate department list; determining a business property, a liability list and an event responsible department of the city event through a dynamic discrimination algorithm; checking a problem area of an event report image through an automatic review model, and calculating probability values of event disposal completion and case settlement completion so as to assist a checker in checking an event state; the event task disposal process is simplified, and the situation of repeated dispatch of the event tasks is reduced.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a multi-agent collaborative urban event handling system based on a large model. Background Technology

[0002] With the acceleration of urbanization and the improvement of urban management, channels for public participation and reporting of urban issues have become more accessible, resulting in a corresponding increase in the number of incidents requiring handling. On the one hand, the issues raised by the public cover all aspects of the city, and staff members often lack a clear understanding of the incident classifications and corresponding departments. On the other hand, duplicate reports are frequent, requiring staff to manually merge multiple complaints from the same person or multiple people filing the same complaint. Furthermore, all stages of incident assignment, review, and statistical analysis rely on manual processes, leading to low overall efficiency, poor accuracy, and high workload for staff.

[0003] AI agents, centered around large language models, possess the ability to autonomously understand, perceive, plan, remember, and use tools, enabling them to automatically execute systems that complete complex tasks. The success of AI agents has provided strong support for the construction of swarm intelligence; multiple AI agents can collaborate and complement each other to accomplish more complex tasks than a single agent, such as software development and social simulation.

[0004] Patent No. CN2023116793619 discloses a multi-agent collaborative knowledge reasoning framework and system based on a large language model, relating to the fields of computer technology and artificial intelligence technology. The framework includes the following steps: for a given knowledge reasoning problem, accurate knowledge reasoning is performed through multiple rounds of reasoning and verification stages; the reasoning stage includes: multiple reasoners based on the large language model provide answers through a procedural reasoning process in the causal direction; the verification stage includes: multiple evaluators based on the large language model check whether the answers provided by the reasoners can be inferred from the problem through a causal consistency verification process in the non-causal direction, and use counterfactual reasoning to verify the correctness of the original answers. This application, based on multi-agent collaborative technology, enhances the reliability and causal consistency of knowledge-based reasoning in LLMs.

[0005] Patent No. CN2023101595507 discloses a method for regional collaborative power flow control of a large power grid based on a multi-agent policy gradient model. The method involves dividing the region and designing its state representation vector, local observation representation vector, and action representation vector. Based on the multi-agent policy gradient model, the local observation representation vector of each agent is used as the input of the first layer, and the output is a specific continuous action space vector, called a communication action. All communication actions are mapped and concatenated into a global policy communication information, which, along with the local observation representation vector, is used as the input of the second layer. The output continuous action is the final action executed by the regional agent in the environment. A simulated power grid operating environment is constructed based on a discretized power grid operation dataset. The model interacts with the simulated power grid operating environment, and batch sample data is collected for model training until convergence. This invention can effectively reduce the variance and randomness of multi-agent policy learning and improve its application effect in large-scale complex power grids.

[0006] However, the above patents have significant shortcomings in data processing, such as the inability to effectively merge similar city events and excessive backlog of city events in the handling process; this leads to cumbersome event handling procedures and repeated dispatch of event tasks, resulting in excessive workload and pressure on grassroots staff. Summary of the Invention

[0007] The purpose of this invention is to provide a multi-agent collaborative urban incident handling system based on a large model, which can simplify the cumbersome and complex urban incident handling process through effective collaboration of multiple agents, thereby effectively shortening the handling time of urban incidents and significantly improving the handling efficiency of urban incidents.

[0008] This invention utilizes the following technical solution:

[0009] A multi-agent collaborative urban event handling system based on a large model, comprising an agent module and a collaboration module;

[0010] The intelligent agent module is used to complete the reporting, acceptance, dispatch, handling, verification and analysis of urban events;

[0011] The intelligent agent module includes an event reporting intelligent agent, an event receiving intelligent agent, an event dispatch intelligent agent, an event handling intelligent agent, an event verification intelligent agent, and an event analysis intelligent agent; among them,

[0012] The event reporting agent is used to convert image information reported by the public when reporting city events into event text, and to assign the event text to the corresponding event type library;

[0013] The event handling agent is used to extract key information elements from the event text and transform the event text into an event semantic vector. At the same time, the event semantic vector of the event text is deduplicated and normalized.

[0014] The event dispatching intelligent agent is used to pre-recommend the event handling departments based on the event semantic vector according to the historical event dispatch information, so as to generate a list of candidate departments. Combined with key information elements and the list of candidate departments, the business nature, list of powers and responsibilities and the responsible department of the event are determined.

[0015] The event handling intelligent agent is used to intelligently extract similar historical events based on the key information elements of the event to be handled, compare the event to be handled with the handling methods and handling criteria of similar historical events, and generate guidance and prompts based on the comparison results. At the same time, it assists the department in charge of the event in handling the event and saving images.

[0016] The event verification agent is used to enhance the images before and after event processing, verify the problem areas in the two images, and calculate the probability values ​​of the event to be processed and the case closed, respectively, to assist the verification personnel in reviewing whether the event to be processed has been closed.

[0017] An event analysis agent is used to connect multiple event sources and dynamically generate event scenarios based on a sentence or a question, in order to discover the root cause of the event problem and provide real-time event handling suggestions. Multiple event sources include high-frequency subjects, group hotspots, repeated complaints, security emergencies, sensitive demands, hot trends, continuous hotspots, and topics of concern.

[0018] The collaboration module is used to coordinate urban events through multiple intelligent agents.

[0019] Preferably, the event reporting agent includes an image transcribing unit and a type matching unit; the image transcribing unit uses a preset image understanding and conversion model to convert the image information when the public reports urban events into event text; the type matching unit uses a matching orientation algorithm to assign the converted event text to the corresponding event type library; the event type library includes natural disasters, accidents, public health events, and social security events.

[0020] Preferably, the event handling intelligent agent includes an information extraction unit and an information processing unit; the information extraction unit uses a preset intelligent vector model to extract key information elements from the event text and converts the event text into an event semantic vector; the key information elements include the event occurrence time, the event occurrence address, and the items or behaviors of the people involved in the event; the information processing unit uses a vector similarity retrieval algorithm to deduplicate and normalize the event semantic vector of the event text.

[0021] Preferably, the event dispatching intelligent agent includes a text parsing unit and a dispatching perception unit; the text parsing unit uses text semantic algorithms and dispatching inference models to pre-recommend the event handling departments based on historical event dispatching information, and generates a candidate department list by using the event semantic vectors that have completed information processing; the dispatching perception unit uses dynamic discrimination algorithms to determine the business nature, list of powers and responsibilities, and responsible department of the city event based on key information elements and the candidate department list.

[0022] Preferably, the dispatch inference model first extracts key information elements, business nature, list of powers and responsibilities, and responsible department from historical event dispatch information:

[0023] Text = [A, B, C, D] (1)

[0024] A=[t,p,g,h] (2)

[0025] Where Text represents historical event assignment information, A represents key information elements, B represents the nature of the business, C represents the list of authority and responsibility, D represents the department responsible for the event, t represents the time of the event, p represents the location of the event, g represents the items involved in the event, and h represents the actions of the people involved in the event.

[0026] Then, the correlation between key information elements in the historical event assignment information is calculated:

[0027]

[0028]

[0029] Among them, DIS t () indicates the correlation between the occurrence time of the event, i represents the sequence number of the historical event, n represents the total number of historical events, s() represents the deviation coefficient, and DIS p () indicates the degree of relevance of the event's location. Let α represent the Hadamard inner product, and α() represent the street-level deviation coefficient. Indicates vector AND;

[0030] Finally, the obtained correlations and extracted key information elements, business nature, list of responsibilities, and responsible departments for events are input into the learning and training layer to obtain a weight matrix. The learning and training layer includes 3 CBL blocks, 2 residual blocks, 3 Inception blocks, 2 Transformer encoding blocks, 1 batch standard layer, and a Silu activation function. The CBL block includes 1 5x5 convolutional layer with a stride of 2, 4 batch standard layers, and a ReLU activation function. Subsequently, the weight matrix is ​​input into 3 normalization layers, 2 fully connected layers, and a Softmax activation function to perform inference verification on the data weight matrix, thereby generating a candidate department list.

[0031] Preferably, the dynamic discrimination algorithm first determines the business nature of the event to be processed based on its key information elements:

[0032]

[0033] Where NE represents the business nature function, DFT() represents the Fourier function, ⊙ represents the Hadamard outer product, and ∏ represents the multinomial vector product;

[0034] Then, the list of powers and responsibilities is adjusted according to the nature of the business;

[0035]

[0036] Among them, lrr ′ This represents the list of responsibilities and obligations to be calibrated, where m represents the number of calibrations, β represents the calibration threshold, RD() represents the calibration function, j represents the list number, γ represents the total number of responsibilities and obligations, and lrr j This represents the list of rights and responsibilities for item j;

[0037] Finally, the department responsible for the incident is determined from the list of candidate departments based on the list of responsibilities and authorities.

[0038]

[0039] Among them, DRI ′ The DRI represents the department responsible for the event, l represents the candidate department number, δ represents the total number of candidate departments, ε represents the correlation coefficient between departments, and DRI represents the total number of candidate departments. l This indicates the l-th candidate department.

[0040] Preferably, the event handling intelligent agent includes an event extraction unit, a strategy handling unit, and a process assistance unit; the event extraction unit uses a clustering algorithm to intelligently extract similar historical events based on the key information elements of the event to be handled; the strategy handling unit uses an expert prior model to compare the event to be handled with the handling methods and handling criteria of similar historical events, and forms guidance and prompts based on the comparison results; the process assistance unit displays the entire process of handling similar historical events and assists the department responsible for the event in executing the handling of the event and retaining images.

[0041] Preferably, the event verification intelligent agent includes an image enhancement unit, a scene matching unit, and a verification and judgment unit; the image enhancement unit uses an anti-sharpening mask algorithm to enhance the image before and after event processing; the scene matching unit uses an image matching algorithm to fuse local and global features of the two enhanced images; the verification and judgment unit uses an automatic review model to verify the problem areas of the two images and calculates the probability values ​​of the pending event being handled and closed, respectively, to assist the verification personnel in reviewing whether the pending event has been closed.

[0042] Preferably, the automated review model includes an image judge and a case closure checker; the image judge first vectorizes the images before and after event processing:

[0043]

[0044]

[0045] Wherein, IM1 represents the image before event processing, a represents the row coordinate of the image before event processing, b represents the column coordinate of the image before event processing, z represents the number of rows and columns, IM2 represents the image after event processing, c represents the row coordinate of the image after event processing, and d represents the column coordinate of the image after event processing.

[0046] Then, the coordinates of the image before event processing are mapped onto the image after event processing to obtain the total mapped coordinate pairs of the images:

[0047] IM0={|IM1→IM2|}=impart[V,W] (10)

[0048] Where IM0 represents the total image mapping coordinate pair, → represents the mapping symbol, impart[] represents the image segmentation function, V represents the image similar mapping coordinate pair, and W represents the image different mapping coordinate pair;

[0049] Then, based on the proportion of different mapped coordinate pairs in the total mapped coordinate pairs of the image, the probability value of the city event being handled is calculated, and the obtained probability value is compared with a set threshold: if the probability value is greater than the set threshold, the city event is judged to have been handled and set as a successful case; if the probability value is less than or equal to the set threshold, the city event is judged to have been unhandled or mishandled and set as a failed case.

[0050] Subsequently, the case closure checker inputs successful cases, failed cases, and case closure elements into the feature extraction layer to obtain feature values. Case closure elements include verification completion, extended approval, and suspended approval. The feature extraction layer consists of three 5x5 convolutional layers and two batch normalization layers. Then, the obtained feature values ​​are input into the learning and training layer for iterative training to obtain the feature weight matrix. The learning and training layer consists of two Inception blocks of different depths, three residual blocks, three batch normalization layers, and a MISH activation function. Finally, the feature weight matrix is ​​passed through four fully connected layers, a random deactivation layer, and a softmax activation function to complete the overall verification and case closure assistance effect of the case closure checker.

[0051] Preferably, the collaboration module includes a communication interaction unit, an information sharing unit, a cooperation strategy unit, and a comprehensive management unit; the communication interaction unit is responsible for communication operations between multiple agents using a communication protocol; the information sharing unit uses a position tracking algorithm to identify and record information between multiple agents and shares the information; the cooperation strategy unit uses a scheduling decision algorithm to design and implement cooperation strategies between agents and ensures smooth collaborative work between agents; the comprehensive management unit uses a state monitoring algorithm to manage the operating status and collaborative relationships of agents and ensures effective collaboration between agents.

[0052] This invention uses a dispatch inference model to pre-recommend event handling departments based on event semantic vectors to generate a candidate department list; it uses a dynamic discrimination algorithm to determine the business nature, list of responsibilities and authorities of urban events, and the responsible department for the event; it verifies the problem areas of the event-reported images by automatically reviewing the model, and calculates the probability values ​​of event completion and case closure to assist the verification personnel in reviewing the event status; it simplifies the event task handling process and reduces the situation of repeated dispatch of event tasks. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0054] Figure 1 Flowchart of a multi-agent collaborative urban incident handling system;

[0055] Figure 2 This is a schematic diagram illustrating the principle of an urban incident handling system. Detailed Implementation

[0056] The present invention will now be described in detail with reference to the accompanying drawings and embodiments:

[0057] like Figures 1 to 2 As shown, the present invention provides a multi-agent collaborative urban event handling system based on a large model, comprising an agent module and a collaboration module.

[0058] The intelligent agent module is used to complete the reporting, acceptance, dispatch, handling, verification and analysis of urban events;

[0059] The intelligent agent module includes an event reporting intelligent agent, an event receiving intelligent agent, an event dispatch intelligent agent, an event handling intelligent agent, an event verification intelligent agent, and an event analysis intelligent agent; among them,

[0060] The event reporting agent is used to convert image information reported by the public when reporting city events into event text, and to assign the event text to the corresponding event type library;

[0061] In this invention, the event reporting intelligent agent includes an image transcribing unit and a type matching unit; the image transcribing unit uses a preset image understanding and conversion model to convert the image information when the public reports urban events into event text; the type matching unit uses a matching orientation algorithm to assign the converted event text to the corresponding event type library; the event type library includes natural disasters, accidents, public health events, and social security events;

[0062] The event handling agent is used to extract key information elements from the event text and transform the event text into an event semantic vector. At the same time, the event semantic vector of the event text is deduplicated and normalized.

[0063] In this invention, the event handling intelligent agent includes an information extraction unit and an information processing unit. The information extraction unit uses a preset intelligent vector model to extract key information elements from the event text and converts the event text into an event semantic vector. The key information elements include the event occurrence time, the event occurrence address, and the behavior of the items or people involved in the event. The information processing unit uses a vector similarity retrieval algorithm to deduplicate and normalize the event semantic vector of the event text.

[0064] The event dispatching intelligent agent is used to pre-recommend the event handling departments based on the event semantic vector according to the historical event dispatch information, so as to generate a list of candidate departments. Combined with key information elements and the list of candidate departments, the business nature, list of powers and responsibilities and the responsible department of the event are determined.

[0065] In this invention, the event dispatching intelligent agent includes a text parsing unit and a dispatching perception unit. The text parsing unit uses text semantic algorithms and a dispatching inference model to pre-recommend the event handling departments based on historical event dispatching information by processing the event semantic vectors after information processing, so as to generate a candidate department list. The dispatching perception unit uses a dynamic discrimination algorithm to determine the business nature, list of powers and responsibilities, and responsible department of the urban event based on key information elements and the candidate department list.

[0066] In this embodiment, the business nature includes urban management, social affairs management services, agricultural and rural development and urbanization promotion, comprehensive social security management, emergency conflict, work safety and ecological environment protection;

[0067] Urban management includes water supply manhole covers, sewage manhole covers, rainwater manhole covers, rainwater grates, power manhole covers, damaged advertising signs, street light manhole covers, telecommunications manhole covers, electric jack manhole covers, network manhole covers, heating manhole covers, isolated trees, water supply oil (gas) transfer signs, gas pressure regulating stations (boxes), cross-river pipelines, natural gas pipelines, etc.

[0068] Social affairs management services include cracking down on false advertising, cracking down on sales promotions with prizes, price fraud, price gouging, management of the list of businesses with abnormal operations, online license display, investigation of product labels, wildlife trade, trademark infringement, supervision and management of food production licenses, food production using non-food raw materials, food with added drugs, purchase inspection system, improper business practices, trademark legality, environmental inspection of catering establishments, and food sampling inspection of catering establishments, etc.

[0069] The promotion of agriculture, rural areas and urbanization includes forest rights disputes, property supervision, property dispute resolution, village affairs disclosure, unlicensed mining, cross-border mining, rural domestic waste collection and transportation, illegal construction in rural areas, water and soil protection, rectification of illegal construction in rural areas, soil pollution prevention and control, water pollution prevention and control, village toilet renovation, drainage of township and village roads, special rectification of straw burning ban, missing river chief system publicity boards, forest patrol, etc.

[0070] Comprehensive social security management includes maintaining community order, patrolling the security environment around schools, resolving conflicts and disputes, processing temporary residence permits, addressing risks and hidden dangers of illegal fundraising, preventing drowning of minors, security patrols, joint security escort for schools, etc.

[0071] Safety in production includes school gas safety management, restaurant gas safety management, urban residential gas safety management, supervision of rural temple fairs and gatherings, administrative supervision of pharmacies, investigation of safety hazards in community and rural housing, cooperation in the supervision of oil and gas pipelines, investigation of safety hazards in construction site facilities, renovation of old urban gas pipelines, unauthorized operation of liquefied gas cylinder filling, safety inspection of medical devices, enforcement against illegal operations, and cooperation in road traffic safety management, etc.

[0072] Emergency response includes handling food safety incidents, emergency response to public health emergencies, investigation and handling of special equipment accidents, emergency response to sudden incidents, strengthening the construction of rural public fire protection facilities, post-disaster relief, geological disaster prevention and control, flood control and drought relief, forest fire hazard investigation and fire prevention, etc.

[0073] Ecological and environmental protection includes restaurants not installing fume purification devices, unorganized emissions of waste gas, alarms from environmental monitoring equipment, excessive volume of audio equipment, and road dust, etc.

[0074] A list of powers and responsibilities refers to a system that, based on laws and regulations, clarifies the powers and responsibilities of government departments, the methods of fulfilling those responsibilities, and the circumstances for accountability, and is supported by mechanisms such as public disclosure, dynamic adjustment, and performance evaluation to promote the comprehensive, accurate, and effective performance of government departments in accordance with the law. The list of powers and responsibilities includes administrative licensing, administrative penalties, administrative enforcement, administrative levies, administrative payments, administrative adjudication, administrative confirmation, administrative rewards, administrative inspections, and administrative filings.

[0075] The event handling intelligent agent is used to intelligently extract similar historical events based on the key information elements of the event to be handled, compare the event to be handled with the handling methods and handling criteria of similar historical events, and generate guidance and prompts based on the comparison results. At the same time, it assists the department in charge of the event in handling the event and saving images.

[0076] In this invention, the event handling intelligent agent includes an event extraction unit, a strategy handling unit, and a process assistance unit. The event extraction unit uses a clustering algorithm to intelligently extract similar historical events based on the key information elements of the event to be handled. The strategy handling unit uses an expert prior model to compare the event to be handled with the handling methods and criteria of similar historical events, and generates guidance and prompts based on the comparison results. The process assistance unit displays the entire process of handling similar historical events and assists the department responsible for the event in executing the handling and image retention of the event.

[0077] The event verification agent is used to enhance the images before and after event processing, verify the problem areas in the two images, and calculate the probability values ​​of the event to be processed and the case closed, respectively, to assist the verification personnel in reviewing whether the event to be processed has been closed.

[0078] In this invention, the case is closed after being handled by the department responsible for the incident and then determined by the investigators.

[0079] In this invention, the event verification intelligent agent includes an image enhancement unit, a scene matching unit, and a verification and judgment unit. The image enhancement unit uses an anti-sharpening mask algorithm to enhance the image before and after event processing. The scene matching unit uses an image matching algorithm to fuse local and global features of the two enhanced images. The verification and judgment unit uses an automatic review model to verify the problem areas of the two images and calculates the probability values ​​of the event to be processed being handled and closed, respectively, to assist the verification personnel in reviewing whether the event to be processed has been closed.

[0080] In this embodiment, the workflow of the unsharpening mask algorithm is as follows:

[0081] First, the image is divided into two parts. The first part is the low-frequency unsharp mask obtained by low-pass filtering (smoothing and blurring techniques) of the image. The second part is the high-frequency part obtained by subtracting the unsharp mask from the original image. Then, the high-frequency part is amplified using a magnification factor (i.e., contrast gain CG) and added to the unsharp mask. Finally, the enhanced image is obtained.

[0082] In this embodiment, the workflow of the image matching algorithm is as follows:

[0083] 1) The input consists of N uniformly sampled local patches;

[0084] 2) Extract features from local regions and then input them into mini-PointNet for further feature extraction; in the fully connected layers of mini-PointNet, the value of the next layer neuron is calculated by matrix multiplication of the input value matrix and parameter matrix of the previous layer, and applied to each input neuron;

[0085] 3) Use a max pooling layer to aggregate the local features of each local region into global features;

[0086] 4) Concatenate the global features to the various local features;

[0087] 5) Use a set of MLPs to fuse local and global features to obtain the feature representation of the image, and integrate and output the feature representation into the feature space to complete image matching;

[0088] An event analysis agent is used to connect multiple event sources and dynamically generate event scenarios based on a sentence or a question, in order to discover the root cause of the event problem and provide real-time event handling suggestions. Multiple event sources include high-frequency subjects, group hotspots, repeated complaints, security emergencies, sensitive demands, hot trends, continuous hotspots, and topics of concern.

[0089] The collaboration module is used to coordinate the processing of urban events through multiple intelligent agents;

[0090] In this invention, the collaboration module includes a communication interaction unit, an information sharing unit, a cooperation strategy unit, and a comprehensive management unit. The communication interaction unit is responsible for communication operations between multiple agents using a communication protocol. The information sharing unit uses a position tracking algorithm to identify and record information between multiple agents and shares the information. The cooperation strategy unit uses a scheduling decision algorithm to design and implement cooperation strategies between agents and ensures smooth collaborative work between agents. The comprehensive management unit uses a status monitoring algorithm to manage the operating status and collaborative relationships of agents and ensures effective cooperation between agents.

[0091] Furthermore, the dispatch inference model first extracts key information elements, business nature, list of powers and responsibilities, and responsible department from historical event dispatch information:

[0092] Text = [A, B, C, D] (1)

[0093] A=[t,p,g,h] (2)

[0094] Where Text represents historical event assignment information, A represents key information elements, B represents the nature of the business, C represents the list of authority and responsibility, D represents the department responsible for the event, t represents the time of the event, p represents the location of the event, g represents the items involved in the event, and h represents the actions of the people involved in the event.

[0095] Then, the correlation between key information elements in the historical event assignment information is calculated:

[0096]

[0097]

[0098] Among them, DIS t () indicates the correlation between the occurrence time of the event, i represents the sequence number of the historical event, n represents the total number of historical events, s() represents the deviation coefficient, and DIS p () indicates the degree of relevance of the event's location. Let α represent the Hadamard inner product, and α() represent the street-level deviation coefficient. Indicates vector AND;

[0099] Finally, the obtained correlations and extracted key information elements, business nature, list of responsibilities, and responsible departments for events are input into the learning and training layer to obtain a weight matrix. The learning and training layer includes 3 CBL blocks, 2 residual blocks, 3 Inception blocks, 2 Transformer encoding blocks, 1 batch standard layer, and a Silu activation function. The CBL block includes 1 5x5 convolutional layer with a stride of 2, 4 batch standard layers, and a ReLU activation function. Subsequently, the weight matrix is ​​input into 3 normalization layers, 2 fully connected layers, and a Softmax activation function to perform inference verification on the data weight matrix, thereby generating a candidate department list.

[0100] In this invention, the working principle of the dispatch reasoning model is as follows: First, key information elements are extracted from historical event dispatch information, including the nature of the event, business type, list of powers and responsibilities, and responsible department, etc., and processed and parsed for subsequent calculation and reasoning; using the calculation model, the key information elements in the historical event dispatch information are calculated and analyzed to obtain the correlation between different information elements for subsequent reasoning and verification.

[0101] Then, the extracted key information elements and correlations are input into the learning and training layer, and various deep learning algorithms and techniques are used for training and learning to obtain the weight matrix; the calculated weight matrix is ​​input into the normalization layer, the fully connected layer and the Softmax activation function for inference verification to generate a candidate department list.

[0102] Finally, based on the reasoning and verification results, the generated list of departments is returned as the processing and assignment result;

[0103] Furthermore, the dynamic discrimination algorithm first determines the business nature of the event to be processed based on its key information elements:

[0104]

[0105] Where NE represents the business nature function, DFT() represents the Fourier function, ⊙ represents the Hadamard outer product, and ∏ represents the multinomial vector product;

[0106] Then, the list of powers and responsibilities is adjusted according to the nature of the business;

[0107]

[0108] Among them, lrr ′ This represents the list of responsibilities for completing the calibration, where m represents the number of calibrations, α represents the calibration threshold, RD() represents the calibration function, j represents the list number, γ represents the total number of responsibilities, and lrr j This represents the list of rights and responsibilities for item j;

[0109] Finally, the department responsible for the incident is determined from the list of candidate departments based on the list of responsibilities and authorities.

[0110]

[0111] Among them, DRI ′ The DRI represents the department responsible for the event, l represents the candidate department number, δ represents the total number of candidate departments, ε represents the correlation coefficient between departments, and DRI represents the total number of candidate departments. l Indicates the l-th candidate department;

[0112] In this invention, the departments responsible for handling events include street (township) level urban operation centers, district / county (city) level urban operation centers, city urban operation centers, etc.

[0113] In this invention, the dynamic discrimination algorithm works as follows: First, the business nature of the event to be processed needs to be determined based on its key information elements. This process can be described by constructing a certain classifier model. For example, the specific business nature of the event to be processed can be determined by the probability distribution in historical events. After determining the business nature of the event to be processed, the list of responsibilities needs to be corrected. The list of responsibilities includes the responsibilities and business scope of each department. By correcting the list of responsibilities, it can be ensured that the scope of responsibility of each department matches the business nature of the event to be processed.

[0114] After determining the list of responsibilities, suitable departments need to be dynamically selected from the candidate department list. This can be achieved through different strategies and algorithms. For example, selection can be done through priority allocation, capability matching, workload balancing, etc.

[0115] Finally, after the screening process is complete, the department responsible for the incident can be determined. This department will be in charge of handling the entire incident and will specifically assign tasks. After determining the responsible department, the information needs to be relayed to the relevant department for follow-up work.

[0116] Furthermore, the automated review model includes an image judge and a case closure checker; the image judge first vectorizes the images before and after event processing:

[0117]

[0118]

[0119] Wherein, IM1 represents the image before event processing, a represents the row coordinate of the image before event processing, b represents the column coordinate of the image before event processing, z represents the number of rows and columns, IM2 represents the image after event processing, c represents the row coordinate of the image after event processing, and d represents the column coordinate of the image after event processing.

[0120] Then, the coordinates of the image before event processing are mapped onto the image after event processing to obtain the total mapped coordinate pairs of the images:

[0121] IM0={|IM1→IM2|}=impart[V,W] (10)

[0122] Where IM0 represents the total image mapping coordinate pair, → represents the mapping symbol, impart[] represents the image segmentation function, V represents the image similar mapping coordinate pair, and W represents the image different mapping coordinate pair;

[0123] Then, based on the proportion of different mapped coordinate pairs in the total mapped coordinate pairs of the image, the probability value of the city event being handled is calculated, and the obtained probability value is compared with a set threshold: if the probability value is greater than the set threshold, the city event is judged to have been handled and set as a successful case; if the probability value is less than or equal to the set threshold, the city event is judged to have been unhandled or mishandled and set as a failed case.

[0124] Subsequently, the case closure checker inputs successful cases, failed cases, and case closure elements into the feature extraction layer to obtain feature values. Case closure elements include verification completion, extension approval, and suspended approval. The feature extraction layer consists of three 5x5 convolutional layers and two batch normalization layers. Then, the obtained feature values ​​are input into the learning and training layer for iterative training to obtain the feature weight matrix. The learning and training layer consists of two Inception blocks of different depths, three residual blocks, three batch normalization layers, and a MISH activation function. Finally, the feature weight matrix is ​​passed through four fully connected layers, a random deactivation layer, and a softmax activation function to complete the overall verification and case closure assistance effect of the case closure checker.

[0125] Verification and completion refers to the process by which the city operation center at the same level as the responsible entity for handling the incident verifies the handling status through methods such as comparing photos before and after the handling, dispatching grid personnel, telephone inquiries, video conferencing, on-site inspections, and intelligent verification. Only those that meet the completion standards can be completed; those that do not meet the completion standards after verification are returned to the original responsible entity.

[0126] Extension of approval means that the department responsible for the incident should submit an extension application within the processing time limit; applications submitted after the deadline will not be accepted. Extension applications should follow the principle of "one extension per incident," and the same incident can only be extended once. The reason for the extension needs to be filled in in the city's incident handling system. The department responsible for the incident should select the number of days for the extension based on the actual situation and upload supporting materials. The maximum extension period is 1 time the original processing time limit. The department responsible for the incident at the same level is responsible for reviewing the extension application.

[0127] Suspended approval refers to the process where, for events that cannot be resolved in a timely manner due to objective limitations and subjective efforts, the responsible department can apply for suspension of the case, stating the reasons and uploading supporting materials, which will then be reviewed and approved by the department responsible for the event at the same level. Once the reason for suspension disappears, the department responsible for the event should continue to handle the event.

[0128] In this invention, the automatic review model works as follows: First, the images of the event to be processed before and after processing are vectorized. Then, the coordinates of the image before processing are mapped onto the image after processing to obtain the total mapped coordinate pairs. Based on the proportion of different mapped coordinate pairs in the total mapped coordinate pairs, the probability value of the city event being handled is calculated. Subsequently, the obtained probability value is compared with a set threshold to determine whether the city event has been handled.

[0129] For successful cases, failed cases, and case closure elements, feature extraction is performed on the case closure checker to obtain feature values. Case closure elements include verification completion, extended approval, and suspended approval. The feature extraction layer uses three 5×5 convolutional layers and two batch normalization layers. Subsequently, the obtained feature values ​​are input into the learning and training layer for iterative training to obtain the feature weight matrix. The learning and training layer uses two Inception blocks of different depths, three residual blocks, three batch normalization layers, and the MISH activation function. Finally, the feature weight matrix is ​​passed sequentially through four fully connected layers, a random deactivation layer, and a softmax activation function to complete the overall verification of the case closure checker and its case closure assistance effect.

[0130] In this invention, the image understanding conversion model, matching orientation algorithm, intelligent vector model, vector similarity retrieval algorithm, text semantic algorithm, clustering algorithm, expert prior model, anti-sharpening mask algorithm, image matching algorithm, communication protocol, position tracking algorithm, scheduling decision algorithm and status monitoring algorithm are all commonly used technical means in this field, and will not be described in detail here.

[0131] Example:

[0132] When using the urban incident handling system to process urban incidents, firstly, the incident reporting agent uses the image transcribing unit to convert the image information reported by the public into incident text using a preset image understanding and conversion model. Then, the type matching unit uses a matching orientation algorithm to assign the converted incident text to the corresponding incident type library. Next, the incident receiving agent uses the information extraction unit to extract key information elements from the incident text using a preset intelligent vector model and converts the incident text into an incident semantic vector. Finally, the information processing unit uses a vector similarity retrieval algorithm to deduplicate and normalize the incident semantic vector of the incident text.

[0133] Then, the event dispatching agent, through the text parsing unit, uses text semantic algorithms and a dispatching inference model based on historical event dispatching information to pre-recommend the event handling departments using the semantic vectors of the processed events, generating a candidate department list. The dispatching perception unit, through a dynamic discrimination algorithm, determines the business nature, responsibility list, and responsible department of the city event based on key information elements and the candidate department list. Next, the event handling agent, through the event extraction unit, uses a clustering algorithm to intelligently extract similar historical events based on the key information elements of the event to be handled. The strategy handling unit, through an expert prior model combined with the handling methods and basis of similar historical events, compares the events to be handled and generates guidance and prompts based on the comparison results. Finally, the process assistance unit displays the entire process of handling similar historical events and assists the responsible department in executing the handling and image retention of the events to be handled.

[0134] Subsequently, the event verification agent uses an image enhancement unit to enhance the images before and after event processing using an anti-sharpening mask algorithm. Then, a scene matching unit uses an image matching algorithm to fuse local and global features of the two enhanced images. Finally, an verification and judgment unit uses an automatic review model to verify the problem areas in the two images and calculates the probability values ​​for handling and closing the pending event, assisting investigators in reviewing whether the event should be closed. Next, the event analysis agent links to event sources such as high-frequency subjects, group hotspots, repeated complaints, security emergencies, sensitive demands, hot trends, ongoing hotspots, and topics of concern. It dynamically generates event scenarios using a single sentence or question to discover the root causes of event problems and provide real-time event handling suggestions.

[0135] Finally, the collaboration module is responsible for communication operations between multiple agents through the communication interaction unit using the communication protocol; it identifies and records information between multiple agents through the information sharing unit and shares the information; it designs and implements cooperation strategies between agents in conjunction with the cooperation strategy unit and ensures smooth collaboration between agents; and it manages the operating status and collaboration relationships of agents through the comprehensive management unit and ensures effective cooperation between agents.

Claims

1. A multi-agent collaborative urban incident handling system based on a large model, characterized in that: Includes an agent module and a collaboration module; The intelligent agent module is used to complete the reporting, acceptance, dispatch, handling, verification and analysis of urban events; The intelligent agent module includes an event reporting intelligent agent, an event receiving intelligent agent, an event dispatch intelligent agent, an event handling intelligent agent, an event verification intelligent agent, and an event analysis intelligent agent; among them, The event reporting agent is used to convert image information reported by the public when reporting city events into event text, and to assign the event text to the corresponding event type library; The event handling agent is used to extract key information elements from the event text and transform the event text into an event semantic vector. At the same time, the event semantic vector of the event text is deduplicated and normalized. The event dispatching intelligent agent is used to pre-recommend the event handling departments based on the event semantic vector according to the historical event dispatch information, so as to generate a list of candidate departments. Combined with key information elements and the list of candidate departments, the business nature, list of powers and responsibilities and the responsible department of the event are determined. The event handling intelligent agent is used to intelligently extract similar historical events based on the key information elements of the event to be handled, compare the event to be handled with the handling methods and handling criteria of similar historical events, and generate guidance and prompts based on the comparison results. At the same time, it assists the department in charge of the event in handling the event and saving images. The event verification agent is used to enhance the images before and after event processing, verify the problem areas in the two images, and calculate the probability values ​​of the event to be processed and the case to be closed, so as to assist the verification personnel in reviewing whether the event to be processed has been closed. An event analysis agent is used to connect multiple event sources and dynamically generate event scenarios based on a sentence or a question, in order to discover the root cause of the event problem and provide real-time event handling suggestions. Multiple event sources include high-frequency subjects, group hotspots, repeated complaints, security emergencies, sensitive demands, hot trends, continuous hotspots, and topics of concern. The collaboration module is used to coordinate urban events through multiple intelligent agents.

2. The urban event handling system based on a large model and multi-agent collaboration as described in claim 1, characterized in that: The event reporting agent includes an image transcribing unit and a type matching unit. The image transcribing unit uses a preset image understanding and conversion model to convert the image information reported by the public when reporting urban events into event text. The type matching unit uses a matching orientation algorithm to assign the converted event text to the corresponding event type library. The event type library includes natural disasters, accidents, public health events, and social security events.

3. The urban event handling system based on a large model and multi-agent collaboration as described in claim 1, characterized in that: The event handling agent includes an information extraction unit and an information processing unit. The information extraction unit uses a preset intelligent vector model to extract key information elements from the event text and converts the event text into an event semantic vector. The key information elements include the event occurrence time, the event occurrence address, and the behavior of the items or people involved in the event. The information processing unit uses a vector similarity retrieval algorithm to deduplicate and normalize the event semantic vector of the event text.

4. The urban event handling system based on a large model and multi-agent collaboration as described in claim 1, characterized in that: The event dispatching intelligent agent includes a text parsing unit and a dispatching perception unit. The text parsing unit uses text semantic algorithms and a dispatching inference model to pre-recommend the event handling departments based on historical event dispatching information and the event semantic vectors that have completed information processing, so as to generate a list of candidate departments. The dispatching perception unit uses a dynamic discrimination algorithm to determine the business nature, list of powers and responsibilities, and responsible department of the city event based on key information elements and the list of candidate departments.

5. The urban event handling system based on a large model and multi-agent collaboration as described in claim 4, characterized in that: The dispatch reasoning model first extracts key information elements, business nature, list of powers and responsibilities, and responsible department from historical event dispatch information: Text = [A, B, C, D] (1) A=[t,p,g,h] (2) Where Text represents historical event assignment information, A represents key information elements, B represents the nature of the business, C represents the list of authority and responsibility, D represents the department responsible for the event, t represents the time of the event, p represents the location of the event, g represents the items involved in the event, and h represents the actions of the people involved in the event. Then, the correlation between key information elements in the historical event assignment information is calculated: Among them, DIS t () indicates the correlation between the occurrence time of the event, i represents the sequence number of the historical event, n represents the total number of historical events, s() represents the deviation coefficient, and DIS p () indicates the degree of relevance of the event's location. Let α represent the Hadamard inner product, and α() represent the street-level deviation coefficient. Indicates vector AND; Finally, the obtained correlations and extracted key information elements, business nature, list of responsibilities, and responsible departments for events are input into the learning and training layer to obtain a weight matrix. The learning and training layer includes 3 CBL blocks, 2 residual blocks, 3 Inception blocks, 2 Transformer encoding blocks, 1 batch standard layer, and a Silu activation function. The CBL block includes 1 5x5 convolutional layer with a stride of 2, 4 batch standard layers, and a ReLU activation function. Subsequently, the weight matrix is ​​input into 3 normalization layers, 2 fully connected layers, and a Softmax activation function to perform inference verification on the data weight matrix, thereby generating a candidate department list.

6. The urban event handling system based on a large model and multi-agent collaboration as described in claim 4, characterized in that: The dynamic discrimination algorithm first determines the business nature of the event to be processed based on its key information elements: Where NE represents the business nature function, DFT() represents the Fourier function, ⊙ represents the Hadamard outer product, and ∏ represents the multinomial vector product; Then, the list of powers and responsibilities is adjusted according to the nature of the business; Where lrr′ represents the list of responsibilities and obligations completed for calibration, m represents the number of calibrations, β represents the calibration threshold, RD() represents the calibration function, j represents the list number of responsibilities and obligations, γ represents the total number of responsibilities and obligations, and lrr j This represents the list of rights and responsibilities for item j; Finally, the department responsible for the incident is determined from the list of candidate departments based on the list of responsibilities and authorities. Where DRI′ represents the department responsible for the event, l represents the sequence number of the candidate department list, δ represents the total number of candidate departments, and ε represents the correlation coefficient of each department. l This indicates the l-th candidate department.

7. The urban event handling system based on a large model and multi-agent collaboration as described in claim 1, characterized in that: The event handling intelligent agent includes an event extraction unit, a strategy handling unit, and a process assistance unit; The event extraction unit uses a clustering algorithm to intelligently extract similar historical events based on the key information elements of the event to be processed; the strategy handling unit uses an expert prior model combined with the handling methods and handling criteria of similar historical events to compare the events to be processed, and forms guidance and prompts based on the comparison results. The process support unit displays the entire process of handling similar historical events and assists the responsible department in handling and recording the events.

8. The urban event handling system based on a large model and multi-agent collaboration as described in claim 1, characterized in that: The event verification agent includes an image enhancement unit, a scene matching unit, and a verification and judgment unit; The image enhancement unit uses an anti-sharpening mask algorithm to enhance the image before and after event processing; the scene matching unit uses an image matching algorithm to fuse local and global features of the two enhanced images. The verification and judgment unit uses an automatic review model to verify the problem areas of the two images and calculates the probability values ​​of the pending events being handled and closed, respectively, to assist the verification personnel in reviewing whether the pending events have been closed.

9. The urban event handling system based on a large model and multi-agent collaboration as described in claim 8, characterized in that: The automated review model includes an image analyzer and a case closure checker; the image analyzer first vectorizes the images before and after event processing: Wherein, IM1 represents the image before event processing, a represents the row coordinate of the image before event processing, b represents the column coordinate of the image before event processing, z represents the number of rows and columns, IM2 represents the image after event processing, c represents the row coordinate of the image after event processing, and d represents the column coordinate of the image after event processing. Then, the coordinates of the image before event processing are mapped onto the image after event processing to obtain the total mapped coordinate pairs of the images: IM0={|IM1→IM2|}=impart[V,W] (10) Where IM0 represents the total image mapping coordinate pair, → represents the mapping symbol, impart[] represents the image segmentation function, V represents the image similar mapping coordinate pair, and W represents the image different mapping coordinate pair; Then, based on the proportion of different mapped coordinate pairs in the total mapped coordinate pairs of the image, the probability value of the city event being handled is calculated, and the obtained probability value is compared with a set threshold: if the probability value is greater than the set threshold, the city event is judged to have been handled and set as a successful case; if the probability value is less than or equal to the set threshold, the city event is judged to have been unhandled or mishandled and set as a failed case. Subsequently, the case closure checker inputs successful cases, failed cases, and case closure elements into the feature extraction layer to obtain feature values. Case closure elements include verification completion, extended approval, and suspended approval. The feature extraction layer consists of three 5x5 convolutional layers and two batch normalization layers. Then, the obtained feature values ​​are input into the learning and training layer for iterative training to obtain the feature weight matrix. The learning and training layer consists of two Inception blocks of different depths, three residual blocks, three batch normalization layers, and a MISH activation function. Finally, the feature weight matrix is ​​passed through four fully connected layers, a random deactivation layer, and a softmax activation function to complete the overall verification and case closure assistance effect of the case closure checker.

10. The urban event handling system based on a large model and multi-agent collaboration as described in claim 1, characterized in that: The collaborative module includes a communication interaction unit, an information sharing unit, a cooperation strategy unit, and a comprehensive management unit. The communication interaction unit is responsible for communication operations between multiple agents using a communication protocol. The information sharing unit uses a location tracking algorithm to identify and record information between multiple agents and shares the information. The cooperation strategy unit uses a scheduling decision algorithm to design and implement cooperation strategies between agents and ensures smooth collaborative work between agents. The comprehensive management unit uses a state monitoring algorithm to manage the operating status and collaborative relationships of agents and ensures effective cooperation between agents.

Images