Fire-fighting alarm condition social strength scheduling method based on alarm condition coordinates and alarm condition classification

Abstract

The invention relates to a social strength scheduling method for a fire-fighting alarm condition based on a text analysis technology, and the method comprises the steps: firstly, monitoring an alarm condition inputted by a 119 wiring platform through a background server; when an alarm occurs, the background server analyzes and obtains an alarm coordinate; meanwhile, the alarm conditions are classified; and finally, scheduling the social strength according to the alarm condition coordinates and / or the alarm condition classification. The method for analyzing and obtaining the alarm condition coordinates by the background server comprises the following steps of: calling a map by the background server through an alarm condition address described by an alarm person, and obtaining the coordinates of an alarm condition occurrence position; the background server classifies the warning conditions to obtain a warning condition classification method, and the warning conditions are classified by adopting a semantic analysis method based on natural language processing NLP (Network Language Protocol) to analyze the content of the warning conditions and classify the warning conditions. According to the method, the social power is reasonably and fully scheduled through accurate alarm condition coordinates and alarm condition classification, and the scheduling efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the application of computer technology in social resource management, and in particular to a social force dispatching method for fire alarm situations based on alarm situation coordinates and alarm situation classification. Background Art

[0002] The efficient handling of fire alarms is of great significance to fire control and fire fighting. The existing fire alarm handling process is mostly handled by the fire brigade or the community property, lacking effective integration of human, material, information and other resources. In today's extremely information-rich society, how to use as many social resources as possible to handle fire alarms is a difficult problem faced in the unified command process. The main problems existing in the existing technology are: 1. Most of the existing fire alarm handling processes are relatively isolated, lacking a complete city-district-street-community-grid-property complete social personnel on-site and handling process.

[0003] 2. The existing social personnel for police dispatch need to be manually dispatched by the command center, which makes it difficult to achieve rapid response and handling of fire alarms.

[0004] In order to improve the efficiency of handling fire alarms, there are also some explorations in the existing technology, such as: The "A Smart Fire City Alarm Dispatch System" with publication number CN116739245A has a danger warning module: used to collect danger factors in the city monitoring area in real time, and then identify and predict the danger in the monitoring area, and then determine the warning danger information and analyze the probability of danger; the alarm center module: used to receive alarm information and receive warning danger information, and then preliminarily classify the alarm information and determine the location information; the fire dispatch module: used to compare and match the classification of the alarm information with the danger level in the database, evaluate the priority of the danger, select the optimal path according to the location information and the path weight, and then combine the priority and the optimal path to perform fire dispatch. In general, this technical solution performs reasonable fire dispatch according to the danger priority and the optimal path and analyzes the probability of danger, so as to improve firefighting efficiency.

[0005] The defects of this plan mainly consider the dispatch method of internal forces of the fire department. There is a lack of dispatch of social forces, and it is not easy for effective personnel to arrive near the crime scene quickly.

[0006] The "Fire Dispatch Method, Device, Equipment and Storage Medium" with publication number CN117196246A, the method includes: obtaining the alarm information of the current fire, and determining the fire information of the current fire based on the alarm information; based on the fire information, screening and obtaining the first fire unit, the police force level of the first fire unit is not less than the fire level of the fire; determining the optimal police dispatch path from the second geographical location corresponding to each first fire unit to the first geographical location, and sending the police dispatch task instruction and the optimal police dispatch path to each first fire unit, so that each first fire unit dispatches fire equipment to reach the first geographical location according to the optimal police dispatch path after receiving the police dispatch task instruction, and the optimal police dispatch path is the path with the shortest time from the second geographical location to the first geographical location. The defect of this scheme is mainly that it only considers how to quickly find the shortest path to the incident address of the fire alarm, and the accuracy of coordinate positioning is not further optimized.

[0007] The "A fire emergency response method and system" with publication number CN116957302A includes the following steps: obtaining accident information and on-duty ERT personnel information; determining fire equipment information and accident handling personnel information according to the accident level; obtaining the fire equipment ID and storage location through the fire equipment information, matching the on-duty ERT personnel ID with the fire equipment ID, and generating an emergency plan based on the matching result; notifying the on-duty ERT personnel of the emergency plan and obtaining feedback from the on-duty ERT personnel; updating the emergency plan based on the feedback information and the accident handling personnel information; recording the on-duty ERT personnel's accident handling process and broadcasting the handling results. It generates a reasonable and efficiently executable emergency plan by comprehensively analyzing the emergency accident situation, which is used to improve the response capability and efficiency of the emergency response team ERT, thereby ensuring the safety of life and property. The main defect of this solution is the lack of analysis of the alarm content, and the inability to quickly notify the corresponding ministries, commissions, bureaus and other handling agencies to arrive at the scene to assist in handling.

[0008] The "A method and device for alarm reminder and processing" with publication number CN117540960A obtains the alarm information provided by the alarmer, extracts the alarmer's location information and accident type; verifies the location information based on mobile phone positioning technology to determine the location of the alarm; sets an alarm dispatch range corresponding to the area type and area characteristics based on the area type and area characteristics of the location of the alarm; sends alarm reminder information to the target alarm object that is within the alarm dispatch range and matches the accident type; and plans the alarm route based on the location of the alarm and the location of the target alarm object. The main defect of the scheme is that this method is suitable for accident alarms and relies on the address positioning of the alarmer's mobile phone device, but lacks analysis of the address and is easily interfered by false addresses. At the same time, fire alarms have more stringent requirements on address accuracy.

[0009] The "Multi-feature fusion named entity recognition method for police text" with publication number CN115146644A mainly includes the following steps: extracting police text information; constructing a named entity recognition dataset for police text and dividing it into training set, validation set and test set; model training; model testing; using the trained multi-feature fusion named entity recognition model to perform named entity recognition on the remaining unlabeled police texts in the police database. This scheme depends on the form of police text, and the BILSTM-CRF model used has strong generalization ability, but requires a large number of standard samples.

[0010] The above solutions also have the following problems: not fully utilizing all map resources, not analyzing the location coordinates, resulting in insufficient positioning accuracy. Summary of the invention

[0011] In order to solve the problems existing in the prior art, the present invention proposes a social force dispatching method for fire alarms based on alarm coordinates and alarm classification. The method fully dispatches various resources and efficiently handles alarms by using artificial intelligence technology, map coordinate positioning technology, etc.

[0012] The present invention provides a social force dispatching method for fire alarm situations based on alarm situation coordinates and alarm situation classification, comprising the following steps: First, the backend server monitors the fire alarm information input by the 119 switchboard; in the following text, the fire alarm information is referred to as the alarm information; Then, when an alarm occurs, the background server analyzes and obtains the coordinates of the alarm; at the same time, the background server classifies the alarm to obtain the alarm classification; Finally, social forces are dispatched according to the police situation coordinates and / or police situation classification; The method for the backend server to analyze and obtain the coordinates of the alarm is that the backend server uses the alarm address described by the alarm caller to call the map and obtain the coordinates of the location where the alarm occurred; then combined with the location of the alarm caller, the coordinates of the location where the alarm occurred are located, i.e., the coordinates of the alarm; The map includes Baidu Map, Amap and / or Tiandi Map; the steps of locating the police situation coordinates include: 101) Obtain the address of the crime scene, and obtain the city where the crime scene occurred from the address of the crime scene; 102) Get the coordinates of the map feedback: call the address resolution interface API of Baidu Map, Amap and Tiandi Map to get the coordinates of Baidu Map, Amap and Tiandi Map feedback; 103) Coordinate filtering of map API feedback: Determine whether the coordinates fed back by the map are within the city area of ​​the crime scene by checking whether the coordinates are within the electronic fence of the city where the crime scene occurred, and remove the coordinates outside the city area; 104) Calculate address resolution coordinates: If the coordinates returned by the three map APIs are not within the city limits of the city where the crime occurred, the coordinates will be empty; If the coordinates returned by a map API are within the city limits of the city where the crime occurred, the coordinates are those returned by the map API; If two coordinates returned by the map API are within the city limits of the crime scene, the coordinate will be the midpoint of the two coordinates; If three coordinates returned by the map API are within the city limits of the crime scene, the coordinates are the midpoint of these three coordinates; 105) Confirm the alarm coordinates: The fire alarm has a coordinate of the alarm person's location. If the coordinate of the alarm person's location and the address resolution coordinate are within 2 kilometers, the address resolution coordinate is used as the alarm coordinate. Otherwise, the coordinate of the alarm person's location is used as the alarm coordinate. The backend server classifies the police situation and obtains the police situation classification by using a semantic analysis method based on natural language processing (NLP) to analyze the police situation content and classify the police situation, and the steps include: 201) building an analysis model; 202) deploying the analysis model and 203) using the analysis model; 201) The steps to construct an analytical model include: 2011) Classification data collection: collect fire alarm text data for annotation; before annotation, clean the collected data and convert the cleaned data into a unified format; 2012) Constructing the analysis model architecture: The analysis model uses the RoBERTa model of the Transformer architecture; Use the LoRA algorithm to give the analysis model the ability to classify police information and add a classification layer: add a fully connected layer as a classification layer on top of the last layer of RoBERTa to output the probabilities of each police category; the number of neurons in this fully connected layer is equal to the number of categories of police classification; 2013) Training analysis models; Dataset division: divide the data set obtained in step 2011) into a training set, a validation set and a test set; Set training parameters: learning rate, batch size, number of iterations, and LoRA parameters; Loss function: Use cross entropy loss function; Optimizer: Select AdamW optimizer; During training: freeze all parameters of the RoBERTa model and only train the newly added classification layer parameters; During the training process of the analysis model, once the required performance is achieved on the validation set, the training is stopped and saved; at this time, the model is deployed to the actual application scenario for real-time processing and classification of fire alarm text data; 2014) Testing and Validation: Before deployment, analytical models are validated using test data to ensure that their performance in production is consistent with their performance on the training and validation sets; 2015) Continuous monitoring and improvement: After the analysis model is deployed, its performance is continuously monitored; if the performance of the analysis model is found to have degraded or new problems have emerged, the analysis model is retrained with new fire alarm text data or the parameters of the analysis model are adjusted; 203) Methods of using analytical models include: The input to the analysis model is the text content of the fire alarm; this text comes from a variety of sources including: transcriptions of emergency calls, posts on social media, or voice transcriptions from surveillance cameras; The output of the analysis model is the classification result of the fire alarm situation; the classification result is one or more category labels, which respectively represent the type or urgency of the alarm situation; When a new fire alarm text is generated, it is input into the deployed analysis model; the analysis model performs semantic analysis on the input text and classifies it according to the features and rules learned during the training process; finally, the classification results are used to guide the dispatch of social forces.

[0013] The step of dispatching social forces according to the police situation coordinates and / or the police situation classification includes: 1) Locate the area to which the alarm belongs through the alarm coordinates; The coordinate range of the area is collected and entered in advance; the area includes the coordinate range of each district, street, community, grid, community, school, business district and / or factory area in the city; 2) When an emergency occurs within the area to which the emergency belongs, the emergency contact person corresponding to the emergency category will be notified by outbound calling; at the same time, the emergency handling link will be sent to the emergency contact person corresponding to the emergency category via SMS; The emergency contact information for the area to which the alarm belongs is entered in advance; 3) According to the alarm coordinates and classification, notify the corresponding grid members, urban management team members, fire safety volunteers and / or property fire safety specialists near the alarm coordinates. These personnel carry GPS mobile terminal devices for positioning; 4) The personnel notified in step 2) and step 3) rush to the scene as soon as possible after receiving the notification; The first person to arrive at the crime scene corrects the coordinates of the police situation and re-locates the crime scene. The accurate crime scene that has been re-located is sent back to the backend server, which then notifies other people who need to rush to the scene to arrive at the accurate crime scene. 5) After arriving at the scene, the personnel invite other personnel to join the operation through the background server. The invited personnel receive outbound calls and / or SMS notifications; 6) The personnel arriving at the scene describe the situation on the scene and inform the backend server; the backend server updates the description of the situation on the scene in real time and promptly informs the fire protection system personnel and other social personnel.

[0014] The beneficial effect of this method is that it can accurately locate the location of the alarm by integrating the triangular coordinates of the alarm person and the alarm address described by the alarm person, and at the same time, classify the alarm. Through accurate alarm coordinates and alarm classification, social forces can be reasonably and fully dispatched, which greatly saves resources and improves dispatch efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of the social force dispatching method of this embodiment. DETAILED DESCRIPTION

[0016] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0017] refer to Figure 1 In this example, a social force dispatching method for fire alarm based on alarm coordinates and alarm classification includes the following steps: 1) The backend server monitors the alarm information entered by the 119 switchboard; 2) When an alarm occurs: The backend server calls the map through the alarm address described by the caller to obtain the coordinates of the location where the alarm occurred; then combined with the location of the caller, the coordinates of the location where the alarm occurred are analyzed, i.e., the coordinates of the alarm; 3) Collect and enter the coordinate range of the area in advance; the area includes the coordinate range of each district, street, community, grid, as well as the residential area, school, business district and / or factory area in the city; Locate the area to which the police incident belongs through the coordinates of the incident location; 4) Enter the emergency contact person of the area in advance; When an emergency occurs within the area, the emergency contact person of the corresponding emergency classification will be notified through an outbound call; at the same time, the emergency handling link will be sent to the emergency contact person through SMS; 5) According to the alarm coordinates and classification, notify the corresponding grid members, urban management team members, fire safety volunteers and property fire safety specialists who are carrying GPS mobile terminal devices near the alarm coordinates; 6) Step 4) and step 5) notify the personnel to rush to the scene quickly after receiving the notification; The first person to arrive at the crime scene corrects the address of the police coordinates and accurately locates the crime scene. The accurately located crime scene is sent back to the backend server, which then notifies other people who need to rush to the scene. 7) After the personnel arrive at the scene, they will invite other personnel to join the operation through the background server. The invited personnel will receive outbound calls and SMS notifications; 8) Personnel arriving at the scene describe the situation on site and inform the backend server; the backend server updates the description of the situation on site in real time and promptly informs fire protection system personnel and other social personnel.

[0018] In step 2), the classification of the police situation is to analyze the content of the police situation using a semantic analysis method based on natural language processing (NLP) to classify the police situation, and the steps include: 2.2.1) Constructing the analysis model; (1) Classified data collection: A large amount of fire alarm text data is collected and annotated. This data should cover different types of alarms, such as fire, emergency rescue, chemical leakage, etc. Data sources may include historical alarm records, transcriptions of alarm call recordings, relevant information on social media, etc. The collected data is cleaned to remove noise, irrelevant information, and duplicates.

[0019] (2) Training and building the model architecture: Select a large model suitable for the police classification task and load its pre-trained weights. This can be a large model that performs well in the field of natural language processing (NLP) or text classification. However, the pre-trained large model is still lacking in professional police classification, and the Lora fine-tuning algorithm is needed to give the large model the ability to classify police content. Lora fine-tuning refers to adding a small number of trainable parameters to the large model. Due to the small number of these newly added parameters, the training cost is significantly reduced, and at the same time, similar effects can be achieved as fine-tuning of all model parameters.

[0020] The police classification dataset is divided into training set, validation set and test set. Usually, the training set is used to train the model, the validation set is used to adjust hyperparameters and monitor model performance, and the test set is used to evaluate the generalization ability of the model.

[0021] Set fine-tuning parameters: learning rate, batch size, number of iterations (epoch), and lora parameters for training. During the fine-tuning process, if the model achieves satisfactory performance on the validation set, the training can be stopped and the weights can be saved. At this point, the model can be deployed to actual application scenarios for real-time processing and classification of fire alarm text data.

[0022] (3) Continuous monitoring and improvement: After the model is deployed, its performance needs to be continuously monitored. If the performance of the model is found to have degraded or new problems have arisen, the model can be retrained using new data or the model parameters can be adjusted. In addition, new features or algorithms can be introduced to improve the performance of the model.

[0023] 2.2.2) Deployment of analytical models; After building and training the analysis model, the next step is to deploy the model to the actual production environment so that it can classify fire alarm text in real time. The deployment process usually includes the following steps: (1) Choose appropriate hardware and software: Choose an appropriate server deployment model based on the complexity of the large model and the expected load.

[0024] (2) Loading model configuration interface: Write code in the production environment to load the previously serialized model. This usually involves reading the model file and deserializing it into a usable model object using the corresponding framework or library. Configure the input and output interfaces of the model according to actual needs. For the fire alarm classification task, an API interface that can receive text input and return classification results may be required.

[0025] (3) Testing and validation: Before deployment, use test data to validate the model to ensure that its performance in the production environment is consistent with its performance on the training set and validation set. After deployment, set up a monitoring mechanism to track the performance and stability of the model. If problems with the model or performance degradation are found, timely maintenance and updates are carried out.

[0026] 2.2.3) Use analytical models; The input to the analysis model is the text content of fire alarms. These texts come from various sources, such as transcriptions of emergency calls, posts on social media, and voice transcriptions from surveillance cameras.

[0027] The output of the analytical model is the classification result of the fire alarm. Depending on the training data and design of the model, the output may be one or more class labels, indicating the type or urgency of the alarm. For example, the output may be one or more of the labels such as "fire", "emergency rescue", "chemical leak", etc.

[0028] When new fire alarm text is generated, it is input into the deployed model. The model will perform semantic analysis on the input text and classify it according to the features and rules learned during the training process. Finally, the model will output the classification results to the corresponding interface or system, and guide the dispatch of social forces based on the classification results. Application examples in specific scenarios are as follows: 1) Notify the Emergency Management Bureau If the analysis model identifies the word "factory" or its synonyms (such as "factory", "industrial zone", etc.) in the alarm text, it will recommend notifying the emergency bureau to handle the situation. This is because flammable and explosive items may be stored in the factory. Once a fire or other accident occurs, the emergency bureau's professional forces are needed to respond and deal with it quickly.

[0029] 2) Notify the Public Security Bureau If the analysis model identifies the presence of "jumping from a building" or its synonyms (such as "suicide" or "jumping off a cliff") in the police report, it will recommend notifying the public security bureau to handle the case. This is because such police reports often involve personal safety and social stability, and require the assistance of the public security bureau to maintain and control order on the scene.

[0030] 3) Not notifying social forces to arrive at the scene In some cases, such as when the text mentions "save cats and dogs" or words with the same meaning, the analysis model may determine that the alarm is not an emergency fire incident, and therefore does not recommend notifying social forces to arrive at the scene. This helps avoid unnecessary waste of resources and confusion, while ensuring that real emergencies can be handled promptly and effectively.

[0031] In step 2), the map includes Baidu map API, Amap API and / or Tiandi map API; the police situation coordinate positioning step includes: 2.1.1) Obtain the crime scene address of the police report; 2.1.2) Get the coordinates of the map feedback: Call the address resolution interface API of Baidu Map, Amap and Tiandi Map to obtain the coordinates fed back by Baidu Map, Amap and Tiandi Map; 2.1.3) Coordinate filtering of map API feedback: The three map feedback coordinates are judged to be within the city area by whether the coordinates are within the electronic fence of the city, and the coordinates outside the city area are eliminated; 2.1.4) Calculate the address resolution coordinates: If the coordinates returned by the three map APIs are not within the city area, the coordinates are empty; If the coordinates returned by a map API are within the city area, the coordinates are those returned by the map API; If two coordinates returned by the map API are within the city area, the coordinate is the midpoint of the two coordinates; If three coordinates returned by the map API are within the city area, the coordinate is the midpoint of these three coordinates; 2.1.5) Confirm the alarm coordinates: The fire alarm has the coordinates of the location of the caller. If the coordinates of the caller's location and the address resolution coordinates are within 2 kilometers, the address resolution coordinates are used. Otherwise, the coordinates of the caller's location are used as the alarm coordinates.

[0032] Example of link SMS in step 4): [×× District Urban Transportation Center] 20××-××-×× 00:29:39 received a fire alarm: smoke coming out of the hood, located at "×× District ×× Road near ×× Gas Station". The automatic location alarm occurred in "×× District ×× Street ×× Community × Grid (Exclusive Grid)", for reference only. Please click the link to view details and report information: https: / / ×××××× In step 7), a record table of the city-level address book is built. Inviting other personnel to join the operation means inviting personnel in the address book; here, screening is performed by searching the organizational structure name, or fuzzy query by name or mobile phone number, or manually entering the mobile phone number for invitation. The invited personnel receive an invitation SMS, which includes relevant information about the police situation and a link to the police situation handling.

[0033] In step 8), the on-site personnel describe the on-site situation as follows: On-site: The backend server guides the on-site personnel to select whether electricity, gas, company or factory is involved; Information verification: The backend server guides on-site personnel to select whether there are people trapped or whether there is an open fire; Disposal completed: the backend server guides the on-site personnel to choose whether to use the micro fire station; Archiving: The backend server guides on-site personnel to choose whether to set it as a classic case.

[0034] The analysis model is described in further detail below.

[0035] 1. For fire alarm text data, data cleaning mainly adopts the following methods: 1. Remove irrelevant information: For example, information such as time, location, and caller information that has no impact on the classification of the police situation.

[0036] Example: The original alarm data is "20xx xx month xx day xx hour xx minute xx second Mr. X (12345678910) reported: At the north gate of Jia Road, a community, the circus loudspeaker is disturbing the residents, please help." The cleaned data is "At the north gate of Jia Road, a community, the circus loudspeaker is disturbing the residents, please help." 2. Remove noise: such as duplicates, spelling errors, redundant words, etc.

[0037] 3. Synonym replacement: Replace synonyms in the text to increase the diversity of the data.

[0038] 4. Adjust sentence order: shuffle the word order in the sentence to improve the robustness of the model.

[0039] 5. Data format conversion: Convert the cleaned data into the "text\tlabel" format to facilitate model reading and training.

[0040] 2. Build the analysis model architecture: When choosing an analysis model, consider using the Transformer-based RoBERTa (Robustly optimizedBERT approach) model. RoBERTa is an improvement on BERT with stronger text understanding capabilities. RoBERTa's optimization enables it to perform better in processing long texts and diverse data, and is suitable for the complexity of police data. RoBERTa cancels the Next Sentence Prediction (NSP) task in BERT and focuses on Masked Language Modeling (MLM), which improves the training efficiency and effectiveness of the model. RoBERTa uses larger batch processing in training, which further improves the stability and accuracy of the model.

[0041] Use the LoRA (Low-Rank Adaptation) fine-tuning algorithm to give the analysis model the ability to classify police information; Low-rank matrix factorization: Two small low-rank matrices are inserted into some layers of the model. The product of these matrices approximates the weight change of the original layer, thereby updating the model.

[0042] Freeze the original model parameters: During fine-tuning, the parameters of the original model remain unchanged, and only the inserted low-rank matrix is ​​updated. This greatly reduces the number of parameters that need to be adjusted, making the fine-tuning process more efficient.

[0043] Fast convergence: Since only a small number of parameters are adjusted, LoRA fine-tuning converges faster and requires less computing resources.

[0044] Strong adaptability: LoRA can enable the model to handle specific tasks (such as police classification) without significantly increasing the complexity of the model.

[0045] 3. Model training: 1. Choose the right hardware and software: According to the complexity of the large model and the expected load, choose the right server to deploy the model. The server should have at least x64 architecture, 32G memory, and 32G GPU memory. Choose an environment framework compatible with the model: Python-based PyTorch and transformers frameworks are suitable for large model fine-tuning and reasoning, and Python-based flask framework provides interface services.

[0046] 2. Dataset division: Divide the fire alarm text dataset into training set, validation set and test set. The common division ratio is 80% training set, 10% validation set and 10% test set.

[0047] 3. Preprocessing and data enhancement: Clean the training data, remove noise, replace synonyms, and adjust sentences to enhance data diversity and the generalization ability of the model.

[0048] 4. Model selection and initialization: Select RoBERTa as the base model and load its pre-trained weights. RoBERTa is particularly suitable for processing complex fire alarm text data due to its optimized pre-training strategy and stronger text understanding ability.

[0049] For the RoBERTa model, in order to better adapt to the fire alarm classification task, the following adaptive modifications are made: Add a classification layer: Add a fully connected layer on top of the last layer of RoBERTa to output the probability of each alarm category. The number of neurons in this fully connected layer is equal to the number of alarm classification categories.

[0050] Parameter freezing and thawing: Based on LoRa technology, during the fine-tuning process, all parameters of the RoBERTa model are frozen and only the newly added classification layer parameters are trained. This can maintain the pre-trained knowledge of the RoBERTa model while reducing the computational cost of training.

[0051] Specifically in this example, the overall fine-tuning and training strategy is analyzed as follows: 1. Training parameter settings: Learning rate: Select an appropriate learning rate, such as 1e-5 or 2e-5, to ensure the model convergence speed and stability. Batch size: Select an appropriate batch size, such as 32 or 64, based on hardware resources. Epochs: Set the number of training cycles, which can be 3 to 5 epochs, adjusted according to the performance of the validation set.

[0052] 2. Early stopping mechanism: Performance monitoring: At the end of each epoch, use the validation set to evaluate model performance (such as accuracy, F1 score). Early stopping condition: If the validation set performance does not improve within a certain number of batches (for example, 1000 batches), stop training. This can prevent the model from overfitting and save computing resources.

[0053] 3. Loss function and optimizer: Loss function: Use the cross-entropy loss function, which is suitable for multi-classification tasks. Optimizer: Select the AdamW optimizer, which combines the advantages of the Adam optimizer and improves on weight decay, making it suitable for fine-tuning pre-trained models.

[0054] Once the desired performance indicator (such as accuracy or F1 score) is achieved on the validation set, stop training and save the model weights. Save the fine-tuned analysis model weights as a file for easy loading and deployment. In the actual application scenario, prepare the operating environment, including the necessary hardware and software support, such as servers, API frameworks (such as Flask or FastAPI), etc. Load the model into the actual system, set up the API interface, receive real-time fire alarm text data, perform classification processing, and return the results to the relevant system or personnel.

[0055] This method comprehensively considers the triangular coordinates of the caller and the police address described by the caller to calculate the location of the police incident. Compared with the existing methods, the police incident location obtained is more accurate, and the police incident classification is used to accurately screen the resources involved in police incident processing.

[0056] By entering the coordinate range of the responsibility area and the emergency contacts within the responsibility area in advance, the relevant emergency contacts can be quickly located through the alarm coordinates.

[0057] By entering the duty information of various ministries and commissions in advance, when the police situation contains relevant keywords, the duty personnel of the relevant ministries and commissions will be notified, so that the personnel of the public security, hospitals, gas, electricity and other relevant units involved in the police situation can arrive at the scene faster to deal with the problem.

[0058] This method allows relevant social forces with handling capabilities to arrive at the scene earlier than the fire department, so that they can analyze and handle the situation in the first time, guide the fire department's forces, and avoid the situation from escalating.

Claims

1. A social force dispatching method for fire alarm based on alarm coordinates and alarm classification, characterized by the following steps: include: First, the backend server monitors the fire alarm information input by the 119 switchboard; in the following text, the fire alarm information is referred to as the alarm information; Then, when an alarm occurs, the background server analyzes and obtains the coordinates of the alarm; at the same time, the background server classifies the alarm to obtain the alarm classification; Finally, social forces are dispatched according to the police situation coordinates and / or police situation classification; The method for the backend server to analyze and obtain the coordinates of the alarm is that the backend server uses the alarm address described by the alarm caller to call the map and obtain the coordinates of the location where the alarm occurred; then combined with the location of the alarm caller, the coordinates of the location where the alarm occurred are located, i.e., the coordinates of the alarm; The map includes Baidu Map, Amap and / or Tiandi Map; the steps of locating the police situation coordinates include: 101) Obtain the address of the crime scene, and obtain the city where the crime scene occurred from the address of the crime scene; 102) Get the coordinates of the map feedback: call the address resolution interface API of Baidu Map, Amap and Tiandi Map to get the coordinates of Baidu Map, Amap and Tiandi Map feedback; 103) Coordinate filtering of map API feedback: Determine whether the coordinates fed back by the map are within the city area of ​​the crime scene by checking whether the coordinates are within the electronic fence of the city where the crime scene occurred, and remove the coordinates outside the city area; 104) Calculate address resolution coordinates: If the coordinates returned by the three map APIs are not within the city limits of the city where the crime occurred, the coordinates will be empty; If the coordinates returned by a map API are within the city limits of the city where the crime occurred, the coordinates are those returned by the map API; If two coordinates returned by the map API are within the city limits of the crime scene, the coordinate will be the midpoint of the two coordinates; If three coordinates returned by the map API are within the city limits of the crime scene, the coordinates are the midpoint of these three coordinates; 105) Confirm the alarm coordinates: The fire alarm has a coordinate of the alarm person's location. If the coordinate of the alarm person's location and the address resolution coordinate are within 2 kilometers, the address resolution coordinate is used as the alarm coordinate. Otherwise, the coordinate of the alarm person's location is used as the alarm coordinate. The backend server classifies the police situation and obtains the police situation classification by using a semantic analysis method based on natural language processing (NLP) to analyze the police situation content and classify the police situation, and the steps include: 201) building an analysis model; 202) deploying the analysis model; and 203) using the analysis model; 201) The steps to construct an analytical model include: 2011) Classification data collection: collect fire alarm text data for annotation; before annotation, clean the collected data and convert the cleaned data into a unified format; 2012) Constructing the analysis model architecture: The analysis model uses the RoBERTa model of the Transformer architecture; Use the LoRA algorithm to give the analysis model the ability to classify police information and add a classification layer: add a fully connected layer as a classification layer on top of the last layer of RoBERTa to output the probabilities of each police category; the number of neurons in this fully connected layer is equal to the number of categories of police classification; 2013) Training analysis models; Dataset division: divide the data set obtained in step 2011) into a training set, a validation set and a test set; Set training parameters: learning rate, batch size, number of iterations, and LoRA parameters; Loss function: Use cross entropy loss function; Optimizer: Select AdamW optimizer; During training: freeze all parameters of the RoBERTa model and only train the newly added classification layer parameters; During the training process of the analysis model, once the required performance is achieved on the validation set, the training is stopped and saved; at this time, the model is deployed to the actual application scenario for real-time processing and classification of fire alarm text data; 2014) Testing and Validation: Before deployment, analytical models are validated using test data to ensure that their performance in production is consistent with their performance on the training and validation sets; 2015) Continuous monitoring and improvement: After the analysis model is deployed, its performance is continuously monitored; if the performance of the analysis model is found to have degraded or new problems have emerged, the analysis model is retrained with new fire alarm text data or the parameters of the analysis model are adjusted; 203) Methods of using analytical models include: The input to the analysis model is the text content of the fire alarm; this text comes from a variety of sources including: transcriptions of emergency calls, posts on social media, or voice transcriptions from surveillance cameras; The output of the analysis model is the classification result of the fire alarm situation; the classification result is one or more category labels, which respectively represent the type or urgency of the alarm situation; When a new fire alarm text is generated, it is input into the deployed analysis model; the analysis model performs semantic analysis on the input text and classifies it according to the features and rules learned during the training process; finally, the classification results are used to guide the dispatch of social forces.

2. The social force dispatching method for fire alarm situations based on alarm situation coordinates and alarm situation classification according to claim 1 is characterized by: The step of dispatching social forces according to the police situation coordinates and / or the police situation classification includes: 1) Locate the area to which the alarm belongs through the alarm coordinates; The coordinate range of the area is collected and entered in advance; the area includes the coordinate range of each district, street, community, grid, community, school, business district and / or factory area in the city; 2) When an emergency occurs within the area to which the emergency belongs, the emergency contact person corresponding to the emergency category will be notified by outbound calling; at the same time, the emergency handling link will be sent to the emergency contact person corresponding to the emergency category via SMS; The emergency contact information for the area to which the alarm belongs is entered in advance; 3) According to the alarm coordinates and classification, notify the corresponding grid members, urban management team members, fire safety volunteers and / or property fire safety specialists near the alarm coordinates. These personnel carry GPS mobile terminal devices for positioning; 4) The personnel notified in step 2) and step 3) rush to the scene as soon as possible after receiving the notification; The first person to arrive at the crime scene corrects the coordinates of the police situation and re-locates the crime scene. The accurate crime scene that has been re-located is sent back to the backend server, which then notifies other people who need to rush to the scene to arrive at the accurate crime scene. 5) After arriving at the scene, the personnel invite other personnel to join the operation through the background server. The invited personnel receive outbound calls and / or SMS notifications; 6) The personnel arriving at the scene describe the situation on the scene and inform the backend server; the backend server updates the description of the situation on the scene in real time and promptly informs the fire protection system personnel and other social personnel.

3. The social force dispatching method for fire alarm based on alarm coordinates and alarm classification according to claim 2 is characterized in that The link SMS format in step 2) is: "On ×× year-×× month-×× day, ×× hour: ×× minute: ×× second, a fire alarm was received at "×× district ×× address" at ×× hour: ×× minute: ×× second: "Alarm description", located at "rough address". The alarm was automatically located at "×× district ×× street ×× community × grid (exclusive grid)", for reference only. Please click the link to view details and report information: https: / / ××××××".

4. The social force dispatching method for fire alarm situations based on alarm situation coordinates and alarm situation classification according to claim 2 is characterized by: In step 5), a record table of the city-level address book is constructed. Inviting other personnel to join the operation means inviting the personnel in the address book. Here, filtering is performed by searching the organizational structure name, or fuzzy query is performed by name or mobile phone number, or the mobile phone number is manually entered for invitation. The invited personnel receive an invitation SMS, which includes relevant information about the police incident and a link to the police incident handling.

5. The social force dispatching method for fire alarms based on alarm coordinates and alarm classification according to claim 2 is characterized by: In step 6), the on-site personnel described the on-site situation as: On-site: The backend server guides the on-site personnel to select whether electricity, gas, company or factory is involved; Information verification: The backend server guides on-site personnel to select whether there are people trapped or whether there is an open fire; Disposal completed: the backend server guides the on-site personnel to choose whether to use the micro fire station; Archiving: The backend server guides on-site personnel to choose whether to set it as a classic case.

6. The social force dispatching method for fire alarm situations based on alarm situation coordinates and alarm situation classification according to claim 1 is characterized by: The following are some application examples in specific scenarios: a. Notify the Emergency Management Bureau If the analysis model identifies the word "factory" or its synonyms in the alert text, the emergency department will be notified to participate in the handling; b. Notify the Public Security Bureau If the analysis model identifies the presence of "jumping from a building" or words with synonyms in the police report, the public security bureau will be notified to handle the case. c. Not notifying social forces to arrive at the scene If the analysis model identifies that the alarm text mentions "save the cat", "save the dog" or synonyms thereof, it is judged that the alarm does not belong to an emergency fire incident, and it is not recommended to notify social forces to arrive at the scene.

7. The social force dispatching method for fire alarms based on alarm coordinates and alarm classification according to claim 1 is characterized by: In the step 211), data cleaning of the fire alarm text data includes: Remove irrelevant information: remove information that has no impact on the classification of police situations; Remove noise: including duplicates, misspellings, and / or redundant words; Synonym replacement: replace synonyms in text data to increase data diversity; Adjust sentence order: shuffle the word order in the sentence to improve the robustness of the model; Finally, perform data format conversion: convert the cleaned data into the "text\tlabel" format to facilitate model reading and training.

Images