110 alarm receiving and handling intelligent auxiliary method based on artificial intelligence technology

Abstract

The invention discloses a 110 alarm receiving and handling intelligent auxiliary method based on an artificial intelligence technology, and belongs to the technical field of public safety and artificial intelligence crossing. The method comprises the steps of multi-mode alarm information acquisition and anti-interference preprocessing; dynamically cognizing and analyzing the alarm condition based on multi-modal fusion; carrying out digital twinning enabling multi-agent dynamic police force scheduling; real-time intelligent assistance of digital twinborn monitoring and alarm processing is realized; carrying out federated learning driven full-process redisk and model iterative optimization; and system security and data compliance guarantee are realized. Through deep integration of multi-mode perception, dynamic cognition, digital twinning and federated learning technologies, accurate analysis of alarm conditions, intelligent scheduling of police strength, whole-course disposal assistance and model closed-loop evolution are realized, the alarm receiving and disposal efficiency, the standardization level and the resource utilization rate are remarkably improved, meanwhile, data privacy and system safety are guaranteed, and the system is suitable for popularization and application. The method is suitable for intelligent police service upgrading of all levels of public security units.

Description

A Smart Assistance Method for 110 Emergency Response Based on Artificial Intelligence Technology Technical Field

[0001] This invention relates to the interdisciplinary field of public safety and artificial intelligence, specifically to a smart assistance method for handling 110 emergency calls that deeply integrates multimodal perception, dynamic cognition, digital twin monitoring, and federated learning with privacy protection. It is applicable to police units at different levels, including command centers and local police stations, covering the entire process of receiving, dispatching, handling, and reviewing calls. It enables precise analysis of incidents, optimal deployment of police resources, real-time assistance in handling incidents, and closed-loop model evolution, comprehensively improving the efficiency, standardization, and utilization of police resources in handling emergency calls, providing full-chain technical support for the upgrade of smart policing. Background Technology

[0002] As a core hub for maintaining social order and protecting the lives and property of the people, the 110 emergency response system is responsible for accepting and handling various matters, including criminal cases, public security incidents, and emergency requests for assistance. With the increasing complexity of social governance and the continuous growth in the daily number of emergency calls, traditional emergency response models and existing intelligent solutions are gradually revealing numerous bottlenecks, making it difficult to meet the practical demands for efficiency, precision, standardization, and compliance.

[0003] The core problems with existing technologies are as follows:

[0004] 1. Insufficient multimodal data fusion limits parsing accuracy: Traditional solutions rely on single text or voice data, failing to fully integrate multi-source information, resulting in weak parsing of ambiguous alarms, dialect alarms, and noisy environments. Some solutions attempt multimodal fusion, but lack anti-interference mechanisms, with dialect recognition accuracy below 60% and key element extraction completeness below 86%, easily leading to misjudgments of alarms.

[0005] 2. Insufficient scientific nature of dynamic dispatching and unbalanced resource allocation: Police dispatch is mostly based on the "proximity principle" or static plans, without taking into account dynamic factors such as police load, expertise, and real-time traffic. This results in a high rate of idle police force, delayed response to emergency situations, and difficulty in dealing with sudden scenarios.

[0006] 3. Inadequate cross-departmental coordination mechanism and inefficient non-police triage: The linkage between 110 and departments such as 12345, 119, and 120 relies on manual transfer, lacks intelligent triage and collaborative plans, non-police incidents occupy police resources, and the response time for cross-departmental handling is as long as several minutes.

[0007] 4. The contradiction between privacy compliance and intelligence is prominent: centralized AI training violates data security regulations, and simple desensitization or edge computing leads to a loss of more than 8% in model accuracy, making it difficult to balance privacy protection and intelligence performance.

[0008] 5. Lack of end-to-end closed-loop optimization and poor system adaptability: The system lacks a "handling-assessment-feedback-iteration" closed loop, and model and contingency plan updates rely on human experience, with a cycle of up to several months, making it unable to adapt to new types of police incidents and crime patterns.

[0009] 6. Insufficient coordination between frontline support and command: Frontline support is limited to basic contingency plan dissemination, making it difficult for the command center to visualize and grasp the dynamics of the scene, and novice dispatchers rely on long-term training to improve their professional skills.

[0010] In summary, existing technologies have significant shortcomings in areas such as multimodal analysis accuracy, the scientific nature of dynamic scheduling, cross-departmental collaboration efficiency, privacy compliance adaptation, closed-loop optimization capabilities, and frontline operational support. There is an urgent need for an intelligent assistance method that integrates deep multimodal fusion, digital twin visualization, federated learning for privacy protection, and a closed-loop iteration throughout the entire process, enabling a paradigm shift in 110 emergency response from "human experience-driven" to "data intelligence-driven." Summary of the Invention

[0011] This invention aims to address the pain points of existing technologies and provide a more innovative and practical intelligent assistance method. Specific objectives include:

[0012] 1. Accuracy rate of police report classification ≥ 98.5%, completeness of key element extraction ≥ 97%, accuracy rate of dialect recognition ≥ 92%, and fuzzy address positioning error ≤ 50 meters;

[0013] 2. The average response time for emergency incidents has been reduced by more than 35%, the balance of police force workload has been improved to more than 95%, and the accuracy rate of diverting non-police incidents is ≥96%;

[0014] 3. Cross-departmental (110 / 119 / 120 / 12345) emergency response time ≤ 3 seconds, and the standardization rate of front-line police officers' handling procedures has increased by 45%;

[0015] 4. Under the premise of privacy compliance (zero external transmission of raw data), the model accuracy loss is ≤3%, and the model iteration cycle is shortened to the weekly level;

[0016] 5. Adaptable to public security units of different sizes, reducing deployment costs and technical barriers.

[0017] To achieve the above objectives, this invention provides an intelligent assistance method for handling 110 emergency calls based on artificial intelligence technology, comprising the following steps:

[0018] Step S1: Multimodal alarm information collection and anti-interference preprocessing

[0019] Establish a multi-channel, all-dimensional alarm data collection terminal, simultaneously accessing multi-source data such as telephone alarms, internet alarms (including APP, mini-program, and video alarms), and IoT device early warnings (including monitoring and sensors), and combine anti-interference preprocessing technology to construct a high-quality alarm data matrix.

[0020] 1. Multi-source data acquisition

[0021] Voice information: Recordings of conversations between the caller and the dispatcher, and audio clips from video alarms, converted to WAV format, covering Mandarin and 6 major dialect regions;

[0022] Text information: text manually entered by the person reporting the incident, supplementary information from the dispatcher, device event tags, and related text from historical incidents;

[0023] Spatial and related information: mobile phone base station positioning, GPS / BeiDou dual-mode positioning, monitoring location, related personnel and vehicle information, historical police reports and geographic correlation data;

[0024] Visual and sensor information: video alarm footage, monitoring frame images, and IoT sensor early warning data (smoke, vibration, etc.).

[0025] 2. Anti-interference preprocessing procedure

[0026] Speech preprocessing: WaveNet noise reduction model suppresses noise (improving signal-to-noise ratio by 15dB), Fourier transform extracts spectral features, and VAD technology extracts effective speech; an embedded dialect adaptation module achieves recognition accuracy of ≥92% for 6 major dialect regions;

[0027] Text preprocessing: Jieba word segmentation, SnowNLP sentiment analysis, BERT entity recognition model, extraction of key entities, standardization of colloquial and dialect expressions;

[0028] Visual preprocessing: frame extraction and sharpness optimization, YOLOv8 target detection model identifies key targets such as people involved in the incident, vehicles, and weapons;

[0029] Data fusion and alignment: Multi-source information is associated based on timestamps, and missing data is filled in by interpolation or similar cases, and an acoustic-text consistency verification mechanism is introduced.

[0030] Step S2: Multimodal Fusion for Dynamic Recognition and Analysis of Police Situations

[0031] A police incident cognition engine based on "multimodal fusion + knowledge graph enhancement + dynamic graph neural network" is constructed to achieve incident classification, element extraction, risk assessment, and non-police-related triage.

[0032] 1. Police Situation Classification and Key Element Extraction

[0033] An improved VLP framework multimodal fusion model was adopted to classify six main categories, including criminal and public security cases, and 25 subcategories (including new types of police incidents), with a classification accuracy of ≥98.5%.

[0034] Based on knowledge graph and NER technology, 15 core elements (location, time, people involved, etc.) are automatically extracted. Fuzzy addresses are combined with Baidu Maps API and a self-built knowledge graph (50,000+ entities). Precise coordinates are generated through GNN inference, with a resolution success rate of ≥96%.

[0035] 2. Dynamic risk level assessment

[0036] Construct an assessment system encompassing five dimensions: personnel safety, property loss, social impact, difficulty of handling, and spatiotemporal dynamics, with 3-5 indicators for each dimension;

[0037] The basic weights are determined by the analytic hierarchy process (AHP), and the DGNN is adjusted in real time. The acoustic emotion vector and text semantics are integrated to calculate the emotion compensation factor, which is divided into four risk levels. The assessment time is ≤2.5 seconds.

[0038] 3. Intelligent triage of non-police incidents

[0039] Based on the responsibility boundary specification, a triage decision tree and semantic matching model (C4.5 algorithm + cosine similarity) are constructed.

[0040] Non-emergency requests are automatically routed to the 12345 platform (response time ≤ 1 second), while ambiguous emergency calls initiate a three-way call, with a route accuracy rate of ≥ 96%.

[0041] Step S3: Digital Twin-enabled Multi-Agent Dynamic Police Dispatch

[0042] A dynamic scheduling system based on "Multi-Agent Reinforcement Learning (MARL) + Digital Twin Monitoring" is constructed to replace the traditional "nearby dispatch" model.

[0043] 1. Real-time monitoring of police force and resource status

[0044] Real-time collection of dynamic information on police forces (location, status, equipment, expertise, etc.), with a data update frequency of ≤8 seconds / time; synchronous access to cross-departmental resource status such as 119, 120, and 12345.

[0045] 2. Construction of a Multi-Agent Reinforcement Learning Scheduling Model

[0046] With the goal of "shortest processing time, optimal load, highest success rate, and lowest resource consumption", the MAVFA algorithm is used for training, combined with more than 500,000 historical data, and a rescheduling heuristic algorithm is introduced to deal with dynamic interference.

[0047] Reward function design: R = 0.4 × response speed + 0.3 × success rate of handling + 0.15 × police force load balancing + 0.1 × resource consumption + 0.05 × standard compliance, dispatch plan generation time ≤ 1.5 seconds.

[0048] 3. Police force allocation and optimal route planning

[0049] For high-risk incidents, priority will be given to assigning specialized police officers and assessing the need for reinforcements; for ordinary incidents, available police officers will be assigned in a balanced manner.

[0050] By integrating real-time traffic data from Gaode Maps and improving the Dijkstra algorithm to generate the optimal route, the ETA error is ≤50 seconds. The pedestrian police route is marked with monitoring coverage points and avoidance areas.

[0051] 4. Cross-departmental collaborative scheduling

[0052] Automatically triggers corresponding departmental coordination based on the type of incident, pushes information via RESTful API, and generates collaborative contingency plans;

[0053] Establish a real-time communication channel with a response time of ≤3 seconds, and dynamically adjust based on a digital twin model during the collaboration process.

[0054] Step S4: Real-time Intelligent Assistance for Digital Twin Monitoring and Emergency Response

[0055] Build a digital twin-based situational awareness map for police incident response, providing accompanying assistance and visual monitoring.

[0056] 1. Construction and Real-time Monitoring of Digital Twins

[0057] Create a "digital twin" of police incidents, integrating multi-source information such as police officers' GPS trajectories, on-site data, and handling status;

[0058] It builds 2D / 3D situation maps based on a game engine, supports multi-screen splitting and detail magnification; it has a built-in performance evaluation model, and automatically triggers graded warnings for anomalies.

[0059] 2. Frontline police officers provide real-time assistance.

[0060] Intelligent contingency plan push: The dynamic contingency plan library covers 65 types of handling guidelines and 45 types of legal guidelines, with a response time of ≤0.5 seconds;

[0061] Real-time risk warning: The LSTM algorithm is used to build a situational awareness model, with a warning accuracy of ≥92%;

[0062] Voice interaction: Supports recognition of 100+ police commands, with a response time of ≤0.8 seconds, improving document entry efficiency by 65%;

[0063] Evidence collection guidance: Generate a standardized checklist, verify its completeness, and achieve an accuracy rate of ≥94%.

[0064] 3. Intelligent knowledge assistance for dispatchers

[0065] Automatically matches laws and regulations, questioning scripts, and typical cases, and pushes them to the sidebar to assist novice dispatchers in making accurate inquiries.

[0066] Step S5: Full-process review and iterative optimization of federated learning-driven model

[0067] Construct a closed-loop system covering the entire process of "data collection - quality assessment - intelligent follow-up - federated feedback - model iteration".

[0068] 1. Intelligent assessment of disposal quality

[0069] We collected 25 evaluation indicators, scored them based on the random forest algorithm, and automatically located the problem links in low-scoring cases with an accuracy of ≥93%.

[0070] 2. Intelligent follow-up visits and feedback collection

[0071] The intelligent follow-up robot dials the phone number of the person who reported the incident, supports dialect interaction, extracts key feedback using the BERT model, and achieves a follow-up participation rate of ≥65% and an extraction accuracy rate of ≥96%.

[0072] 3. Privacy Protection and Model Iteration in Federated Learning

[0073] Deploy a federated learning framework, keep the raw data locally, introduce a differential privacy mechanism (ε=0.5), and the model accuracy loss is ≤3%;

[0074] The central server aggregates gradients to generate a global model, iterates weekly, and incremental training takes ≤3 hours, while simultaneously updating the pre-plan library and knowledge graph.

[0075] Step S6: System Security and Data Compliance Assurance

[0076] Build a multi-layered, end-to-end security protection system to ensure data security, law enforcement compliance, and system stability.

[0077] 1. Data encryption and privacy protection

[0078] Data is stored and transmitted using the AES-256 encryption algorithm, and transmission security is ensured by the SSL / TLS protocol.

[0079] The RBAC model implements three-level permission allocation, operation logs are stored on the blockchain, and sensitive information is anonymized 100%.

[0080] 2. System stability and disaster recovery backup

[0081] "Cloud + edge computing" hybrid deployment, with the basic layer adapted to Huawei Atlas500 equipment (operating in an environment of -20℃~60℃).

[0082] The dual-active backup system has a data synchronization latency of ≤10 seconds and a system availability of ≥99.99%; relying on the public security intranet and 5G private network, the transmission latency is ≤50ms.

[0083] Beneficial effects

[0084] 1. Significantly improved accuracy of multimodal parsing: 98.5% accuracy in police incident classification, ≥92% dialect recognition, ≥97% key element extraction, and ≤50-meter fuzzy address positioning error, solving the problem of parsing complex scenarios;

[0085] 2. Dynamic scheduling and resource utilization optimization: Emergency response time is reduced by 35%+, police force load balance is 95%+, efficiency of handling multiple incidents concurrently is improved by 30%, and police force idle rate is reduced by 25%;

[0086] 3. High efficiency in cross-departmental collaboration and triage: accuracy rate of non-police triage ≥96%, cross-departmental response time ≤3 seconds, and efficiency in handling complex police situations improved by 40%;

[0087] 4. Balance between privacy compliance and intelligence: No raw data is transmitted outside the system, the model accuracy loss is only 3%, and it has passed privacy and security testing and certification;

[0088] 5. Enhanced visualization and standardized handling: The standardization rate of police response has increased by 45%, the integrity of evidence has increased by 30%, and the training cycle for new officers has been shortened by 60%;

[0089] 6. Closed-loop iteration and strong scenario adaptability: Weekly iteration adapts to new types of police situations, edge computing reduces deployment costs, and it is suitable for public security units of different sizes.

[0090] 7. Improved public satisfaction and police efficiency: The quality score of handling police calls increased by 18%, public satisfaction was ≥97%, the time to receive a call was ≤40 seconds, and the efficiency of document entry increased by 65%. Attached Figure Description

[0091] Figure 1: Overall flowchart of the intelligent auxiliary method for 110 emergency response based on artificial intelligence technology of the present invention;

[0092] Figure 2: Flowchart of the detailed process for multimodal alarm information collection and anti-interference preprocessing in step S1;

[0093] Figure 3: Flowchart of the multimodal fusion process for dynamic cognition and analysis of police situations in step S2;

[0094] Figure 4: Flowchart of multi-agent dynamic police force dispatching empowered by digital twin in step S3;

[0095] Figure 5: Flowchart of the full process review and model iteration optimization driven by federated learning in step S5.

[0096] The module includes: 1. Multimodal information acquisition module; 2. Anti-interference preprocessing module; 3. Multimodal alarm recognition and analysis module; 4. Digital twin dispatch module; 5. Real-time alarm response assistance module; 6. Federated review and iteration module; and 7. Security and compliance assurance module. Detailed Implementation

[0097] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0098] I. System Hardware Configuration

[0099] The method described in this invention requires the following hardware system for implementation, balancing performance and cost, and is adaptable to public security units of different sizes:

[0100] 1. Data acquisition terminal:

[0101] Police command center alarm receiving terminal: includes high-definition voice acquisition equipment, 4K monitor, keyboard and mouse set, supports multi-channel alarm access;

[0102] Mobile police terminals: smartphones and tablets (supporting GPS / BeiDou dual-mode positioning, 4G / 5G communication, voice interaction, and high-definition photo and video recording), suitable for grassroots police officers working outdoors;

[0103] IoT monitoring equipment: high-definition cameras (supporting target detection and frame extraction), sound sensors, vibration sensors, and smoke sensors, covering key road sections and locations;

[0104] Internet alarm access server: Supports data access from multiple channels such as APP, mini-program, and video alarm, with a concurrent processing capacity of ≥1000 channels / second;

[0105] Edge computing device: Huawei Atlas500 edge computing box (deployed in grassroots police stations), supports operation in environments ranging from -20℃ to 60℃, and meets the needs of local data preprocessing and model inference.

[0106] 2. Computation Nodes:

[0107] Command centers in large and medium-sized cities: Deploy GPU server clusters (including 4-8 NVIDIA A100 GPU servers), with a single node computing power of ≥100 TFLOPS, for multimodal model training, reinforcement learning scheduling computation, and digital twin rendering;

[0108] Grassroots units: Relying on edge computing devices and lightweight cloud nodes, no independent GPU clusters are required, reducing deployment costs.

[0109] 3. Storage devices:

[0110] It adopts a distributed storage system with a capacity of ≥100TB, supports encrypted data storage and real-time backup, and has a read / write speed of ≥1GB / s;

[0111] Cloud backup nodes: Adopting a dual-active architecture, data synchronization latency is ≤10 seconds, ensuring that data is not lost in extreme cases.

[0112] 4. Communication network:

[0113] Relying on the public security intranet and 5G private network, data transmission latency is ≤50ms and network reliability is ≥99.99%;

[0114] It supports offline caching on mobile police terminals and automatically synchronizes data after the network is restored to avoid data loss during field operations.

[0115] 5. Display and command terminal:

[0116] Command center large screen: 4K resolution, supports multi-screen split display and 3D situation rendering, size ≥120 inches;

[0117] Commander's console: includes a high-performance computer and a touch screen display, supporting command issuance, dynamic early warning viewing, and progress monitoring.

[0118] II. Software Module Implementation

[0119] The software modules of the method described in this invention are developed using Python and C++ languages, and employ the TensorFlow and PyTorch deep learning frameworks. The implementation details of each core module are as follows:

[0120] 1. Multimodal information acquisition and anti-interference preprocessing module:

[0121] Speech interference resistance processing: The Librosa library is used to extract spectral features, the WaveNet noise reduction model is used to suppress interference, WebRTC-VAD technology is used to extract effective speech segments, and the dialect adaptation module is optimized based on the attention mechanism. The accuracy of effective speech extraction is ≥99%, and the accuracy of dialect recognition is ≥92%.

[0122] Text preprocessing: Jieba word segmentation tool for word segmentation, SnowNLP for sentiment analysis, BERT model for entity recognition training, entity extraction accuracy ≥97%;

[0123] Visual preprocessing: Key targets in the image are extracted based on the YOLOv8 object detection model, and the image is optimized using OpenCV. The target recognition accuracy is ≥95%.

[0124] Data fusion and alignment: Based on the timestamp synchronization algorithm, Redis caching is used to achieve real-time data association, with data processing latency ≤2 seconds.

[0125] 2. Multimodal alarm recognition and analysis module:

[0126] Multimodal fusion model: Based on the improved Vision-LanguagePre-training (VLP) framework, the input includes speech-to-text (implemented via iFlytek's speech-to-text API, with an accuracy of ≥98.5%), original text, location features, and visual features. After being encoded by a cross-modal attention mechanism, these features are fed into a fully connected layer classifier. The training set contains 150,000+ historical police cases. The model is trained using the Adam optimizer with a learning rate of 0.001 and 100 iterations, achieving a final classification accuracy of 98.7%.

[0127] Accurate Address Resolution: Based on Baidu Maps API and a self-built address knowledge graph (containing 50,000+ entities), topological relationship reasoning is achieved through graph neural networks (GNN), with a fuzzy address resolution success rate of ≥96%.

[0128] Dynamic risk assessment: The Analytic Hierarchy Process (AHP) and Dynamic Graph Neural Network (DGNN) are implemented using Python. The weights of the indicators are adjusted through expert scoring and data-driven methods, and the assessment time is controlled within 2.3 seconds.

[0129] Non-police triage model: Based on the C4.5 decision tree algorithm, the training data contains 30,000+ triage cases, and combined with cosine similarity calculation, the triage accuracy reaches 96.2%.

[0130] 3. Digital Twin Scheduling Module:

[0131] Multi-agent reinforcement learning model: Implemented in PyTorch, using the Multi-Agent Value Function Approximation (MAVFA) algorithm, with a state space dimension of 60 (including police location, incident characteristics, traffic status, cross-departmental resource status, etc.) and an action space dimension of 25. The reward function weights are set according to response speed (0.4), handling success rate (0.3), police workload (0.15), resource consumption (0.1), and compliance with regulations (0.05). The model is trained using 500,000+ historical data points, and the scheduling scheme generation time is ≤1.5 seconds.

[0132] Digital twin monitoring: Based on the Unity game engine, a visual situation map is built, which supports 2D / 3D switching, and receives police officers' GPS trajectory and monitoring data in real time, with a dynamic update frequency of ≤8 seconds / time;

[0133] Route planning: Integrates with Gaode Map's real-time traffic API and is based on an improved Dijkstra algorithm, achieving an optimal route planning accuracy of ≥97% and an ETA prediction error of ≤50 seconds;

[0134] Cross-departmental collaboration: Data interaction with the 119, 120, and 12345 platforms is achieved through RESTful APIs, with a response time of ≤3 seconds.

[0135] 4. Real-time Emergency Response Assistance Module:

[0136] Contingency plan database: Stored using a MySQL database, it covers 65 types of emergency response guidelines and 45 types of legal guidelines. It uses a keyword matching algorithm to accurately push contingency plans with a response time of ≤0.5 seconds.

[0137] Risk warning: Based on the LSTM model, the system predicts the development of events and integrates police feedback and monitoring data, achieving a warning accuracy rate of ≥92%.

[0138] Voice interaction: Integrates iFlytek's voice interaction API, supports the recognition of 100+ police commands, and has a response time of ≤0.8 seconds;

[0139] Evidence collection guidance: Implemented through a mobile APP, it supports real-time uploading and verification of photos, videos, and audio evidence, with an evidence integrity verification accuracy rate of ≥94%.

[0140] 5. Federated Review and Iteration Module:

[0141] Quality assessment: Based on the random forest algorithm, 25 evaluation indicators were input, the model accuracy was ≥93%, and Python was used to automate scoring and problem localization;

[0142] Intelligent follow-up robot: Based on iFlytek's speech synthesis and recognition technology, it supports dialect interaction, the follow-up script can be customized, and the feedback information extraction adopts the BERT model with an accuracy of ≥96%;

[0143] Federated learning: Based on the FedAvg algorithm improvement, a differential privacy mechanism (ε=0.5) is introduced. Each branch office uses a lightweight BERT model for local training, and the central server securely aggregates gradients. Incremental training takes ≤3 hours.

[0144] Contingency plans and knowledge graph updates: By analyzing feedback data and best practices using natural language processing technology, new emergency response procedures and entity relationships are automatically added, and the updates are performed weekly.

[0145] 6. Security and Compliance Assurance Module:

[0146] Data encryption: AES-256 encryption algorithm is used for storage and transmission, SSL / TLS protocol ensures transmission security, and transparent encryption technology is used for storage encryption;

[0147] Access control: A three-level permission allocation is implemented based on the RBAC model, and operation logs are stored using blockchain technology, making them tamper-proof;

[0148] Sensitive information masking: Based on regular expressions and entity recognition, it automatically hides sensitive content such as ID card numbers and minors' names, with a 100% desensitization rate;

[0149] Disaster recovery backup: The dual-active architecture ensures system availability ≥99.99% and data synchronization latency ≤10 seconds.

[0150] III. Demonstration of Practical Application Scenarios

[0151] To verify the practical effectiveness of the method of this invention, the application process and effects of the method are explained in detail below in conjunction with four typical police incident scenarios:

[0152] Scenario 1: Level I Emergency (Shopping mall fire, cross-departmental collaboration)

[0153] 1. Multimodal Information Acquisition and Preprocessing: The alarm caller reports via video, "Fire on the 5th floor of XX shopping mall, thick smoke, many people trapped." The system simultaneously acquires video footage, audio clips, and the alarm caller's GPS location (accurate to the mall entrance). Smoke and temperature sensors within the mall push alerts, and surrounding surveillance automatically captures and uploads images of the thick smoke. The preprocessing module performs noise reduction and transcription of the audio, extracting the text "Fire on the 5th floor of XX shopping mall, many people trapped"; it extracts frames from the video footage, identifying the range of thick smoke and areas where people are gathered; it correlates and aligns the audio, text, location, and sensor data; acoustic detection indicates the alarm caller is experiencing "panic" (90% confidence level), triggering a consistency check with no anomalies, and generating an alarm data matrix.

[0154] 2. Emergency Situation Dynamic Recognition and Analysis: The multimodal fusion model determined the emergency situation type to be "Disaster / Accident - Fire," extracting key elements: location of the incident (5th floor of XX shopping mall, precise coordinates XXX), involved personnel (multiple trapped individuals), risk of casualties, involvement of hazardous materials (flammable items), and the spread of dense smoke at the scene; the knowledge graph was used to associate the mall's structural information (the 5th floor is a food court with gas pipelines) and the distribution of nearby fire stations (the nearest fire station is 1.8 kilometers away); dynamic risk assessment combined with traffic congestion index (0.2, smooth traffic) and pedestrian flow (weekend peak, 1000+ people), with an urgency score of 98, classifying it as a Level I risk; it was confirmed that multi-departmental collaboration was required, and no diversion was necessary.

[0155] 3. Dynamic Police Force Dispatch and Digital Twin Monitoring: The police force status perception unit collects data on the nearest patrol team (idle, 0.8 km away) and two officers (fully equipped) from the nearest police station; cross-departmental resource status shows one fire truck idle at the nearest fire station and three ambulances on standby at the 120 emergency center. The multi-agent dispatch model generates the following scheme: Patrol officers are assigned to arrive first to evacuate surrounding crowds (estimated arrival time 1 minute 30 seconds), police officers from the police station provide reinforcements (estimated arrival time 2 minutes 10 seconds), and coordination is established with the 119 fire truck (estimated arrival time 3 minutes 50 seconds) and the 120 ambulance (estimated arrival time 4 minutes 30 seconds); the route planning module plans the optimal route for the fire trucks, avoiding construction sections. The command center monitors the location and progress of all parties in real time through the digital twin situation map, generating a "response situation map."

[0156] 4. Real-time Emergency Assistance: Frontline police officers receive the "Fire Response Plan" on their terminals, prompting them to "wear protective equipment, prioritize evacuation of the public, stay away from gas pipelines, and provide real-time fire updates"; the risk warning module analyzes the smoke spread speed based on monitoring footage, indicating "Smoke will spread to the 4th floor within 5 minutes, accelerate evacuation"; officers report via voice, "Arrived at the mall entrance, beginning evacuation of surrounding crowds," which is automatically entered into the system; evidence collection guidance prompts the collection of photos of the fire scene, videos of the smoke area, and statistics on the number of trapped personnel. After the arrival of firefighters and ambulances (119 and 120), information is synchronized through real-time communication channels, and firefighting and rescue operations are carried out according to the collaborative plan.

[0157] 5. Review and Iteration: The emergency response was completed (fire extinguished, all trapped personnel rescued, 2 people with minor injuries sent to the hospital). Data collected by the review module: alarm reception time 38 seconds, complete element extraction, dispatch accuracy 100%, arrival time 1 minute 25 seconds, response time 45 minutes, public satisfaction 97 points. Intelligent follow-up collected feedback on "timely evacuation and efficient rescue." The federated learning module added this case to the training set, optimizing cross-departmental collaboration weights; the contingency plan library was supplemented with "shopping mall peak-hour fire evacuation strategies," and the knowledge graph was updated with suggestions for optimizing shopping mall fire exits.

[0158] Scenario 2: Level III emergency (burglary, assisted by a novice dispatcher)

[0159] 1. Multimodal Information Acquisition and Preprocessing: The caller reports the crime by phone, describing in Cantonese, "My house was broken into and things were stolen; my computer and cash are gone." The system simultaneously collects the call audio and mobile phone base station location (near XX residential area). The dispatcher adds "The caller lives in apartment 201, building 3, XX residential area." The preprocessing module recognizes the speech through a dialect-adaptive module and transcribes it into standard text. Text preprocessing extracts key entities: "apartment 201, building 3, XX residential area," "break-in," "computer," and "cash." After data fusion and alignment, a structured summary is generated.

[0160] 2. Dynamic Recognition and Analysis of Police Incidents: The multimodal fusion model determined the incident type to be "Criminal - Burglary," extracting key elements: location of the incident (Building 3, Apartment 201, XX Community, precise coordinates XXX), time of the incident (within 1 hour before the caller), persons involved (unknown, 1-2 people), stolen items (computer, cash), and scene condition (door pried open). Dynamic risk assessment, combined with the community's historical crime records (no theft records in the past 3 months) and low pedestrian traffic, resulted in an urgency score of 72, classifying it as a Level III risk; confirming it falls within the 110 (police) jurisdiction. The dispatcher's sidebar displays "Burglary Interrogation Script," "Key Points for Scene Preservation," and relevant articles of the Criminal Law on theft, assisting the dispatcher in supplementing inquiries with "amount of stolen cash, computer model, and whether the scene was disturbed."

[0161] 3. Dynamic Police Deployment: The police force status perception unit identifies the nearest patrol officer to the community (idle, 1.2 km away, skilled in criminal investigation), with a traffic congestion index of 0.3 (smooth traffic). The deployment model generates a solution: assign this officer to the scene (estimated arrival time 3 minutes 20 seconds), without requiring reinforcements. The route planning module generates the optimal walking route and marks the community's surveillance coverage points.

[0162] 4. Real-time Assistance for Police Response and Digital Twin Monitoring: Police officers receive a "burglary response plan" on their terminals, prompting them to "protect the scene, collect pry marks on the door lock, inquire with neighbors about witnesses, and retrieve community surveillance footage." Voice interaction allows officers to query the "burglary evidence collection list" and report via voice, "Arrived at the scene, door pried open, scene undamaged." Evidence collection guidance prompts officers to collect photos of the door lock, the location of stolen items, and footprints at the scene. Officers collect evidence according to regulations and upload it to the cloud. The command center monitors the officers' movements through a digital twin situation map, confirming their arrival along the planned route, and the response status is updated in real-time to "in progress."

[0163] 5. Review and Iteration: The case was completed (case filed and investigation completed, pry marks and surveillance footage collected). Review data showed: call response time 42 seconds, dispatch accuracy 100%, response time 25 minutes, and public satisfaction 95%. Follow-up feedback indicated "professional police officers and meticulous questioning." The model was iterated to optimize the element extraction weights for theft reports, and the contingency plan database was supplemented with "questioning scripts for burglaries in dialect areas."

[0164] Scenario 3: Level IV emergency (lost item, request for help, reporting in dialect)

[0165] 1. Multimodal Information Collection and Preprocessing: The person reporting the incident uses a WeChat mini-program, manually inputting dialect text (in Minnan dialect, "I lost a red backpack on the XX bus, containing my ID card and bank cards"), and simultaneously uploading their bus payment record. The system collects the text information, the reporting person's location (XX subway station), and the bus information associated with the payment record (XX bus route, currently at XX stop). The preprocessing module converts the dialect text into standard semantics, "A red backpack containing my ID card and bank cards was lost on the XX bus," extracts key entities, and associates it with the bus's real-time location to generate an incident data matrix.

[0166] 2. Police Situation Dynamic Recognition and Analysis: The model determines the police situation type as "Emergency Assistance - Lost Items", extracting key elements: lost items (red backpack, ID card, bank card), location (XX bus route), time (1 hour ago), and current bus location (XX stop); the risk assessment score is 60 points, classifying it as Level IV risk; no diversion is required.

[0167] 3. Dynamic Police Dispatch: The police force status perception unit detects the patrol officer closest to the XX station (idle, 0.6 km away). The dispatch model, combined with the police force load (current load 30%), generates a plan: assign the officer to the XX station to meet the driver (estimated arrival time 7 minutes).

[0168] 4. Real-time Assistance in Police Response: The system pushes out a "Lost and Found Response Plan," prompting officers to "contact the bus driver, check the in-vehicle surveillance footage, register the lost item information, and verify the identity of the person reporting the incident." Police officers can use voice commands to find the bus driver's contact information and communicate with the driver in real time to confirm that the backpack has been found. By reporting "Backpack found, all items intact" via voice, the system guides the collection of photos of the backpack and the driver's statement, while simultaneously contacting the person reporting the incident for verification.

[0169] 5. Review and Iteration: The incident was resolved (the person who reported the incident retrieved their backpack). Review data showed: call response time 32 seconds, dispatch accuracy 100%, response time 18 minutes, and public satisfaction 96%. Follow-up feedback indicated a desire to contact the driver more quickly. The model was optimized to improve the police force matching strategy for lost item reports, prioritizing officers familiar with the public transportation system; the contingency plan database was supplemented with a "rapid response process for lost items on public transportation."

[0170] Scenario 4: Level II Incident (Knife Fight, Dialect-Based Reporting + Escalating Risk)

[0171] 1. Multimodal Information Acquisition and Preprocessing: The caller reports the incident by phone, describing in Wu dialect that "there's a fight with knives at the XX night market, many people are watching, and someone is bleeding." The system simultaneously collects the call audio and mobile phone base station location (XX night market area); nearby surveillance cameras capture footage of the fight (2 people with knives, many people watching), and uploads it simultaneously. The preprocessing module uses a dialect adaptive module to transcribe the speech, performs noise reduction, and extracts the text; target detection in the surveillance footage identifies the knife-wielding individuals and the injured; after data fusion and alignment, an alarm data matrix is ​​generated, and acoustic detection reveals the caller's "anger + panic" emotions (88% confidence level).

[0172] 2. Police Situation Dynamic Recognition and Analysis: The model determines the police situation type as "Public Security - Knife Fighting," extracting key elements: location of the incident (city center square at XX night market, precise coordinates XXX), involved persons (2 people with knives, 1 person injured, 20+ onlookers), casualties, and the use of weapons (knives); dynamic risk assessment, combined with night market pedestrian flow (500+ people) and traffic congestion index (0.7, congested), results in an urgency score of 92, classifying it as a Level II risk; no diversion is required.

[0173] 3. Dynamic Police Deployment: The police force status perception unit collects data on the two patrol teams closest to the night market (one team is idle, 1.1 km away; the other is responding to an incident, 1.5 km away), and three officers from the nearest police station (equipped with riot control expertise, riot shields, and restraint equipment). The deployment model generates the following plan: the idle patrol team is assigned to arrive at the control site first (estimated arrival time 3 minutes 10 seconds), officers from the police station provide reinforcements (estimated arrival time 4 minutes 20 seconds), and an ambulance is coordinated (estimated arrival time 5 minutes 10 seconds); the route planning module plans routes for the patrol teams to avoid congested areas.

[0174] 4. Real-time Assistance and Dynamic Early Warning in Police Response: Police officers receive a "Contingency Plan for Handling Knife-wielding Fights" on their terminals, prompting them to "wear riot gear, prioritize controlling the knife-wielding individuals, prevent injuries to bystanders, and treat the injured." The risk warning module, based on surveillance footage analysis, alerts officers that "one of the knife-wielding individuals is emotionally agitated and may injure bystanders; take precautions." Upon arrival, officers report via voice, "Arrived at the scene; knife-wielding individuals are still in a standoff; one injured person is on the ground," which is automatically entered into the system. Evidence collection guidance prompts officers to collect photos of the fight scene, weapons, and witness contact information. During the response, if one knife-wielding individual resists, and officers request backup, the command center, after confirming the situation via a digital twin situation map, dispatches another nearby team of officers to complete their mission early and provide backup, ensuring control of the scene.

[0175] 5. Review and Iteration: The incident was successfully handled (the knife-wielding assailant was apprehended, and the injured were taken to the hospital). Review data showed: call response time 40 seconds, dispatch accuracy 100%, arrival time 3 minutes and 5 seconds, handling time 20 minutes, and public satisfaction 95%. Follow-up feedback indicated that "the police acted decisively, effectively preventing the situation from escalating." The federated learning module added this case to the training set to optimize the reinforcement strategy for Level II incidents; the contingency plan database was supplemented with "tactics for handling knife fights in night markets."

[0176] IV. Performance Testing and Comparative Analysis

[0177] To verify the superiority of the method of this invention, a pilot test was conducted for six months at the public security command center of a sub-provincial city and three grassroots police stations. Traditional emergency response systems and existing single-mode intelligent auxiliary solutions (a certain brand's dispatch system based on single-modal parsing) were selected as control groups, while the method of this invention was used as the experimental group. The test results are shown in the table below:

[0178] Test Indicators | Control Group 1 (Traditional System) | Control Group 2 (Single Intelligent Solution) | Experimental Group (Invention Method) | Improvement (vs. Traditional System) | Incident Classification Accuracy | 89.2% | 93.5% | 98.7% | 9.5% | Key Element Extraction Completeness | 85.7% | 90.2% | 97.3% | 11.6% | Dialect Recognition Accuracy | 58.3% | 75.1% | 92.4% | 34.1% | Average Emergency Response Time | 6 minutes 45 seconds | 5 minutes 30 seconds | 4 minutes 08 seconds | 38.6% | Non-Police Incident Triage Accuracy | 82.3% | 88.7% | 96.2% | 13.9% | Cross-Departmental Response Time | 45 seconds | 20 seconds | 3 seconds | 93 0.3% Emergency Response Document Entry Efficiency: 3 minutes / case, 1.8 minutes / case, 1.05 minutes / case; 65.0% Emergency Response Process Standardization Rate: 55.2%, 72.1%, 99.8%, 44.6%; Police Load Balancing: 68.5%, 82.3%, 95.7%, 27.2%; Public Satisfaction: 92.1%, 94.3%, 97.5%, 5.4%; Model Iteration Cycle: 3 months, 1 month, 1 week; 91.7% Privacy Compliance: Non-compliant (centralized data storage), Partially compliant (simple anonymization), Fully compliant (zero external transmission of raw data); Low-cost Grassroots Deployment (primarily manual), Medium (requires GPU nodes), Medium-low (supports edge computing). surface

[0179] The test results show that the method of this invention significantly outperforms traditional systems and existing single intelligent solutions in core indicators such as accuracy of police incident identification, response speed, cross-departmental collaboration efficiency, standardization level, and privacy compliance, fully verifying the practicality, innovation, and superiority of the method. It not only meets the handling needs of high-volume and complex scenarios in large and medium-sized cities but also adapts to the low-cost deployment needs of grassroots police stations, effectively supporting the comprehensive intelligent upgrade of 110 emergency response work.

[0180] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A smart assistance method for handling 110 emergency calls based on artificial intelligence technology, characterized in that: Includes the following steps: S1: Multimodal alarm information collection and anti-interference preprocessing, simultaneously accessing multi-source data from voice, text, spatial correlation, vision, and sensors, and performing anti-interference preprocessing to construct a high-quality alarm data matrix; S2: Multimodal fusion-based dynamic alarm cognition and analysis, based on multimodal fusion models, knowledge graphs, and dynamic graph neural networks to achieve alarm classification, key element extraction, risk level assessment, and intelligent triage of non-police alarms; S3: Digital twin-enabled multi-agent dynamic police force dispatch, based on multi-agent reinforcement learning models and digital twin monitoring, to achieve real-time perception of police force status, dispatch plan generation, optimal path planning, and cross-departmental collaborative dispatch; S4 Digital twin monitoring and real-time intelligent assistance for police response: Construct and display a digital twin of the police situation, providing frontline police officers with intelligent contingency plan push, real-time risk warning, voice interaction and evidence collection guidance; S5: Federated learning-driven full-process review and model iteration optimization, based on the federated learning framework to achieve intelligent assessment of disposal quality, intelligent follow-up feedback collection and closed-loop iteration under model privacy protection; S6: System security and data compliance assurance, building a full-process security protection system through data encryption, access control, and disaster recovery backup.

2. The intelligent assistance method for 110 emergency response based on artificial intelligence technology according to claim 1, characterized in that: In step S1, the anti-interference preprocessing includes: speech preprocessing using the WaveNet noise reduction model and dialect adaptive module to achieve a dialect recognition accuracy of ≥92%; text preprocessing using jieba word segmentation, SnowNLP sentiment analysis and BERT entity recognition; visual preprocessing using the YOLOv8 object detection model; multi-source data is aligned based on timestamp association and an acoustic-text consistency verification mechanism is introduced.

3. The intelligent assistance method for 110 emergency response based on artificial intelligence technology according to claim 1, characterized in that: In step S2, the alarm classification adopts an improved Vision-LanguagePre-training multimodal fusion model with a classification accuracy of ≥98.5%; the key element extraction is based on knowledge graph and named entity recognition technology with a parsing success rate of ≥96%; the dynamic risk level assessment adopts the analytic hierarchy process and dynamic graph neural network with an assessment time of ≤2.5 seconds.

4. The intelligent assistance method for 110 emergency response based on artificial intelligence technology according to claim 1, characterized in that: In step S3, the multi-agent reinforcement learning scheduling model adopts a multi-agent value function approximation algorithm, constructing a reward function based on response speed, handling success rate, police load balance, resource consumption and compliance with regulations, with a scheduling scheme generation time ≤ 1.5 seconds; the path planning integrates real-time traffic data and improves the Dijkstra algorithm, with an ETA error ≤ 50 seconds.

5. The intelligent assistance method for 110 emergency response based on artificial intelligence technology according to claim 1, characterized in that: Step S3 also includes cross-departmental collaborative scheduling, which uses RESTful API to push information and generate collaborative plans with the 119, 120, and 12345 platforms, with a linkage response time of ≤3 seconds.

6. The intelligent assistance method for 110 emergency response based on artificial intelligence technology according to claim 1, characterized in that: In step S4, the digital twin constructs a two-dimensional / three-dimensional visualization situation map based on a game engine, with a dynamic update frequency of ≤8 seconds / time; the intelligent contingency plan library covers at least 65 types of handling guidelines and 45 types of legal guidelines, with a push response time of ≤0.5 seconds.

7. The intelligent assistance method for 110 emergency response based on artificial intelligence technology according to claim 1, characterized in that: In step S5, the federated learning framework introduces a differential privacy mechanism, where the privacy budget ε=0.5 and the model accuracy loss is ≤3%; the model adopts weekly iteration, and the incremental training time is ≤3 hours.

8. The intelligent assistance method for 110 emergency response based on artificial intelligence technology according to claim 1, characterized in that: Step S5 also includes an intelligent follow-up robot that supports dialect interaction, extracts feedback information based on the BERT model, has a follow-up participation rate of ≥65%, and an extraction accuracy of ≥96%.

9. The intelligent assistance method for 110 emergency response based on artificial intelligence technology according to claim 1, characterized in that: In step S6, data encryption uses the AES-256 algorithm, access control is based on the RBAC model to implement three-level permission management, operation logs are stored using blockchain, and system availability is ≥99.99%.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1-9.

Images