A fire department alert and response system, method, and storage medium

Abstract

The application provides a fire-fighting alarm response system and method, comprising: in response to an alarm communication, extracting an alarm element based on a preset inquiry rule, and generating a pre-alarm work order including the alarm element; and processing the alarm element by using a preset determination model; specifically: constructing a fire alarm dialogue model with Transform as a basic architecture, regarding the alarm question and answer as a sequence generation question, reasoning the fire situation that should be inquired next in a self-recurrence manner according to the obtained fire situation, until the model considers that enough information has been obtained to judge the fire grade; the above steps are executed for multiple times, the understanding of the fire scene situation is gradually deepened, and the real-time fire grade evaluation is generated on this basis. The system deviation existing in the strategy learning method of the Markov decision process and reinforcement learning in the existing alarm automatic alarm response or type system is solved, and the defect that the response result deviates from the actual operator's behavior to a certain extent is solved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a fire alarm response system, method and storage medium. Background Technology

[0002] In the field of disaster emergency management, especially in fire emergency response scenarios, real-time and accurate information acquisition and assessment are crucial for improving emergency response efficiency. Traditional fire emergency responses typically rely on manual operation and judgment, but with the development of artificial intelligence technology, automated emergency response and fire situation assessment have become possible. In this project, the core concept of the model is to acquire real-time information about the fire scene through automated interactive dialogue, including but not limited to key information such as the area of ​​the fire source, the number of people trapped, smoke conditions, and available fire-fighting facilities. Through interactive question-and-answer sessions, the model can gradually deepen its understanding of the fire scene and generate a real-time fire severity assessment based on this understanding. Summary of the Invention

[0003] This invention proposes a fire alarm response system, method, storage medium, and management system, which solves the problems of existing automatic alarm response systems and methods that use Markov decision processes and reinforcement learning for policy learning, resulting in processing results that deviate to some extent from the actual operator behavior, as well as low model efficiency and lack of interpretability.

[0004] The technical solution of the present invention is implemented as follows: a fire alarm response method, comprising the following steps:

[0005] In response to alarm communication, the alarm elements are extracted based on preset query rules, and a pre-received alarm work order including the alarm elements is generated;

[0006] The alarm elements are processed using a preset judgment model. Specifically, a fire alarm dialogue model with Transformer as its basic architecture is constructed. The alarm response is treated as a sequence generation question. The next fire information to be inquired about is inferred from the fire information already obtained in an autoregressive manner until the model believes that enough information has been obtained to make a fire level judgment. The above steps are executed multiple times to gradually deepen the understanding of the fire scene situation and generate a real-time fire level assessment on this basis.

[0007] As a preferred technical solution, the model adopts Transformer as the basic architecture, treats the emergency response question as a sequence generation problem, and uses an autoregressive approach to infer the next fire situation to be inquired about based on the fire situation already obtained, until the model believes that it has obtained enough information to determine the fire level.

[0008] As a preferred technical solution, the fire alarm dialogue model uses the explicit state of the fire as the model input to generate implicit states during training. The "END" symbol represents the end signal, indicating the end of state generation and transition to fire classification reasoning. During model training, the prediction symbol [S] sequence is used to realize the autoregressive generation of the implicit state sequence, which can be regarded as learning state reasoning of multiple stages at once. The level prediction symbol [D] is used to learn and generate the target level based on all fire state information. The explicit state refers to the necessary inquiries and status responses during the alarm reception stage, while the implicit state refers to the optional inquiries and status responses during the alarm reception stage.

[0009] As a preferred technical solution, the model first obtains the explicit state of the fire. By concatenating state prediction symbols [S], it calculates the probability distribution of the implicit states and queries the state with the highest probability among the implicit states. At this time, the user will answer True, False, or uncertain. For uncertain information, the above steps are repeated to find the next state with the highest probability and query the user. Based on the user's True or False answer, the model concatenates the newly obtained implicit states and then uses the state prediction symbols to predict the next state. When the termination signal is inferred from the state symbols, or when the confidence level of the predicted state probability is lower than the threshold, the model stops querying and switches to level determination, predicting the target fire level by concatenating level prediction symbols [D].

[0010] As a preferred technical solution, the visible characteristics of a fire include the number of people trapped and the area of ​​the fire. One or more of the following visible characteristics can be added as needed: the name of the person who reported the fire, the phone number of the person who reported the fire, the time of the fire, the address of the fire, the location of the person who reported the fire, the floor on fire, the number of people trapped, the number of people injured and the extent of their injuries, and the smoke conditions.

[0011] As a preferred technical solution, the hidden state includes open flame, smoke, white smoke, black smoke, explosive hazard, surrounding combustibles, spread, high-rise buildings, collapse, whether fire-fighting facilities are intact and whether water sources are sufficient. The hidden state is adjusted according to the climate zone, season, time period, building density, and security alert level of the area.

[0012] As a preferred technical solution, the alarm classification results include duplicate alarms, non-duplicate alarms, ordinary alarms, and important alarms; when the alarm classification result is determined to be a duplicate alarm, a second confirmation of the duplicate alarm is performed, and the current alarm communication ends after obtaining a positive response to the duplicate alarm.

[0013] As a preferred technical solution, the system uses an interactive voice response component to transfer questions and answers and obtain the results. When the question and answer results meet preset conditions, the person receiving the fire alarm voice communication extracts alarm elements based on the preset query rules; when the question and answer results do not meet the preset conditions, a manual response is initiated.

[0014] A fire alarm response system includes the following modules:

[0015] Alarm Receiving Center Module: Used to receive alarm information and early warning information, extract alarm elements based on preset query rules, and generate a pre-received alarm work order including the alarm elements;

[0016] Fire alarm dialogue model module: It is pre-built with a fire alarm dialogue model based on Transformer as the basic architecture; it is used to process the alarm elements received by the alarm receiving center module, treat the alarm question and answer as a sequence generation question, and use an autoregressive method to infer the next fire situation to be inquired about based on the fire situation already obtained, until the model believes that it has obtained enough information to make a fire level judgment; it executes the above steps multiple times to gradually deepen the understanding of the fire scene situation, and generates a real-time fire level assessment on this basis;

[0017] The hazard information display module is used to acquire and display hazard scene information in real time based on information from the hazard warning module of the alarm receiving center module and the fire alarm dialogue model module using GIS map management technology.

[0018] Hazard warning module: Used to collect hazard factors in the urban monitoring area in real time, and then identify and predict the hazards in the monitoring area, and then judge the warning hazard information and analyze the probability of the hazard occurring.

[0019] A non-temporary storage medium is provided, which, when the instructions in the storage medium are executed by a fire alarm response system, enables the device to perform the aforementioned fire alarm response method.

[0020] Compared with existing technologies, this solution has the following advantages:

[0021] This paper introduces explicit and implicit traits into the alarm elements, thus treating fire alarm response as a multi-step reasoning problem. Considering that the call-handling process can be naturally viewed as a sequence generation, this work reformulates automatic call-handling as a sequence generation task. Unlike previous reinforcement learning methods, the multi-step query process is explicitly modeled as a task of generating a sequence containing fire alarm information. This improves the efficiency and interpretability of the multi-step reasoning process. It addresses the problem of capturing uncertainty in the reasoning process and determining the accuracy level with limited labeled data and a minimal number of dialogue turns. Furthermore, it resolves the systemic bias inherent in existing automatic alarm response or typology systems that use Markov decision processes and reinforcement learning for policy learning, which leads to responses that deviate to some extent from actual operator behavior.

[0022] Unlike traditional multi-round question-and-answer systems that only require accuracy, the goal of fire emergency response question-and-answer systems is to determine the main characteristics and severity of a fire in as few rounds of questioning as possible, thereby enabling precise dispatch of firefighting forces and resources. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a model architecture diagram of the fire alarm dialogue model based on the Transformer architecture of the present invention.

[0025] Figure 2 This is a schematic diagram of the reasoning method of the fire alarm dialogue model based on the Transformer architecture of the present invention;

[0026] Figure 3 This is a schematic diagram of a fire alarm response system according to the present invention. Detailed Implementation

[0027] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0028] This application discloses a fire alarm response method. In response to alarm communication, it extracts alarm elements based on preset query rules and generates a pre-alarm work order including the alarm elements. The alarm elements are then processed using a preset judgment model. Specifically, a fire alarm dialogue model with Transformer as its basic architecture is constructed. The alarm response is treated as a sequence generation question, and the next fire information to be inquired about is inferred in an autoregressive manner based on the fire information already obtained, until the model believes that enough information has been obtained to determine the fire level. The above steps are executed multiple times to gradually deepen the understanding of the fire scene situation and generate a real-time fire level assessment based on this.

[0029] The model uses the Transformer architecture as its basic framework, treating emergency response questions as sequence generation problems. It employs an autoregressive approach to infer the next fire situation to inquire about based on the obtained fire information, continuing until the model deems it has sufficient information to determine the fire severity level. Referring to fire emergency response guidelines, fire severity classification primarily depends on the burned area and the number of people trapped. However, in reality, other factors such as smoke conditions and the type of building on fire can also influence the actual severity level and the number of vehicles dispatched, potentially leading to escalation or additional dispatches. Therefore, this model treats the burned area and the number of people trapped as explicit states, which need to be obtained in the first two rounds of the dialogue. Other fire situations are considered implicit states, requiring prediction round by round based on previous dialogue information. Explicit states refer to the necessary inquiries and responses during the emergency response phase, while implicit states refer to the optional inquiries and responses during the emergency response phase.

[0030] The architecture of the model is as follows Figure 1 As shown, the explicit state of the fire is used as the model input to train and generate implicit states, where the "END" symbol represents the end signal, indicating the end of state generation and transition to fire classification reasoning. During training, the prediction symbol [S] sequence is used to achieve autoregressive generation of the implicit state sequence, which can be regarded as learning state reasoning of multiple stages at once; the level prediction symbol [D] is used to learn and generate the target level based on all fire state information.

[0031] The working principle of the model in the inference phase is as follows: Figure 2As shown, the model first obtains the visible state of the fire (number of trapped people, burned area), calculates the probability distribution of the hidden states by concatenating state prediction symbols [S], and queries the state with the highest probability among the hidden states. At this time, the user will answer True, False, or uncertain. For uncertain information, the above steps are repeated to find the next state with the highest probability and query the user again. Based on the user's True or False answer, the model concatenates the newly obtained hidden states and then uses the state prediction symbols to predict the next state. When the end signal is inferred from the state symbols, or when the confidence of the predicted state probability is lower than the threshold, the model will stop querying and switch to level determination, predicting the target fire level by concatenating the level prediction symbols [D].

[0032] The visible characteristics of a fire include the number of people trapped and the area burned. Depending on the needs, one or more of the following visible characteristics may be added: the name of the person who reported the fire, their phone number, the time of the report, the address of the incident, the location of the person who reported the fire, the floor on fire, the number of people trapped, the number of injured and the extent of their injuries, and the smoke conditions. The hidden characteristics include open flames, smoke, white smoke, black smoke, explosive hazards, surrounding combustibles, fire spread, high-rise buildings, collapse, the integrity of fire-fighting facilities, and the adequacy of water supply. The hidden characteristics are adjusted based on the climate zone, season, time of day, building density, and the security alert level of the area.

[0033] The alarm classification results include duplicate alarms, non-duplicate alarms, ordinary alarms, and important alarms. When the alarm classification result determines it to be a duplicate alarm, a secondary confirmation of the duplicate alarm is performed. Once a positive response is received regarding the duplicate alarm, the current alarm communication ends. Based on an interactive voice response component, a question-and-answer session is conducted to obtain the question-and-answer results. When the question-and-answer results meet preset conditions, the person conducting the fire alarm pre-receiver voice communication extracts alarm elements based on the preset query rules; when the question-and-answer results do not meet the preset conditions, a manual response is initiated.

[0034] At the same time, a fire alarm response system for implementing this fire alarm response method is provided, including the following modules:

[0035] Alarm Receiving Center Module: Used to receive alarm information and early warning information, extract alarm elements based on preset query rules, and generate a pre-received alarm work order including the alarm elements;

[0036] Fire alarm dialogue model module: It is pre-built with a fire alarm dialogue model based on Transformer as the basic architecture; it is used to process the alarm elements received by the alarm receiving center module, treat the alarm question and answer as a sequence generation question, and use an autoregressive method to infer the next fire situation to be inquired about based on the fire situation already obtained, until the model believes that it has obtained enough information to make a fire level judgment; it executes the above steps multiple times to gradually deepen the understanding of the fire scene situation, and generates a real-time fire level assessment on this basis;

[0037] The hazard information display module is used to acquire and display hazard scene information in real time based on information from the hazard warning module of the alarm receiving center module and the fire alarm dialogue model module using GIS map management technology.

[0038] Hazard warning module: Used to collect hazard factors in the urban monitoring area in real time, and then identify and predict the hazards in the monitoring area, and then judge the warning hazard information and analyze the probability of the hazard occurring.

[0039] How it works: Due to the presence of latent states, this task can be viewed as a multi-step reasoning problem. The challenge lies in capturing the uncertainty in the reasoning process and determining the accuracy rating with very little labeled data and as few dialogue turns as possible. Most previous methods typically treat this problem as a continuous decision-making process (Markov Decision Process) and employ reinforcement learning for policy learning. However, reinforcement learning is biased in how it queries fire information and only makes this rating based on the final accuracy reward, deviating to some extent from actual operator behavior. Considering that the switching process can be naturally viewed as a sequence generation, this work reformulates automated switching as a sequence generation task. Unlike previous reinforcement learning methods, the multi-step query process is explicitly modeled as a task of generating sequences containing fire information. This improves the efficiency and interpretability of the multi-step reasoning process.

[0040] Training and testing data for the above model:

[0041] Due to the unique nature of fire emergency response, obtaining authentic dialogue records is challenging. Therefore, all training and testing data in this project are derived from manually constructed dialogues. The dialogue data comprises 10 fire scenarios, including markets, vehicles, high-rise buildings, and underground structures. The construction process referenced the rules for fire emergency response Q&A in the fire emergency rescue manual, such as key questions, severity levels and decision descriptions, and dispatch prompts.

[0042] Model Input Sample

[0043] Specific data examples are as follows:

[0044] {

[0045] "type":"Market Fire",

[0046] "explicit_inform_slots":{

[0047] "Number of people trapped = 0":true,

[0048] "burning area = 0-200":true

[0049] },

[0050] "implicit_inform_slots":{

[0051] "open flame":true,

[0052] "smoke":true,

[0053] "white smoke":false,

[0054] "Black smoke":true,

[0055] "Explosive hazard":true,

[0056] "Surrounding combustibles":false,

[0057] "spread":false,

[0058] "High-rise buildings":true,

[0059] "collapse":false,

[0060] "Fire protection facilities are complete": false

[0061] "Water supply is sufficient":true

[0062] },

[0063] "fire_tag":1

[0064] }

[0065] The above status descriptions are the elements of questions asked by the fire dispatcher. True / False represents the feedback from the person who called the fire department. For example, in the above example, "Number of people trapped: 0": True means that the dispatcher asked "Are there any people trapped?" and the person who called the fire department answered "No". The dispatcher continued to ask "Does the burning area exceed 200 square meters?" and the person who called the fire department answered "No". The dispatcher continued to ask "Can you see open flames?" and the person who called the fire department answered "I can see open flames"... Finally, the dispatcher ended the questioning and determined the fire level to be 1 ("fire_tag":1).

[0066] In specific application scenarios, the model will only receive the fire scenario type (type) and explicit state (explicit_inform_slots), and will complete the inference of the implicit state step by step based on the contents of the above two fields.

[0067] The performance metrics charts are as follows:

[0068] Dacc (accuracy of rating) FRec (Front-end Fire Recall Rate) ATurn (Average Number of Turns in a Conversation) 0.785 0.811 10.2

[0069] Expected application scenarios

[0070] In fire emergency response scenarios, timely, accurate, and efficient information collection and processing are crucial for ensuring personnel safety and mitigating property damage. This model aims to provide precise decision support to fire emergency responders through an intelligent question-and-answer system, thereby optimizing resource allocation and emergency response procedures. Specific application scenarios include the following:

[0071] First, the model can be applied to rapid assessment in the early stages of a fire. By automatically collecting key information from the scene, the model can generate detailed reports on the fire situation in real time and provide initial fire severity predictions to dispatchers, helping to quickly understand the severity of the fire and providing data support for subsequent rescue operations.

[0072] Furthermore, the application of models can optimize the allocation of fire rescue resources. By analyzing the collected data, models can provide emergency response centers with recommendations on how to optimize rescue resources such as personnel, vehicles, and equipment to ensure that rescue resources are used most effectively, shorten response time, and improve rescue efficiency.

[0073] Furthermore, the long-term application of the model is expected to accumulate a large amount of valuable fire emergency response data, providing important references for future fire early warning and risk assessment. Through deep learning of historical cases, the model can continuously optimize its prediction accuracy and response efficiency, providing strong technical support for fire emergency management.

[0074] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method of responding to a fire alarm, characterized by, The method comprises the following steps: In response to the alarm communication, the alarm elements are extracted based on the preset inquiry rules, and a pre-alarm work order including the alarm elements is generated; The alarm elements are processed by using a preset determination model; specifically, a fire alarm dialogue model is constructed with a Transformer as a basic architecture, the alarm question and answer are regarded as a sequence generation question, and the next fire situation that should be inquired is inferred in a self-recurrent manner according to the obtained fire situation until the model considers that enough information has been obtained to make a fire grade judgment; the above steps are executed multiple times to gradually deepen the understanding of the fire scene situation, and a real-time fire grade evaluation is generated on this basis; The fire alarm dialogue model takes the explicit state of the fire as the model input to train and generate the implicit state, wherein the "END" symbol represents an end signal indicating the generation of the end state and the transition to fire classification inference; in the model training process, the implicit state sequence is generated in a self-recurrent manner by using the prediction symbol [S] sequence, which can be regarded as learning multiple stage state inferences at one time; the target grade is learned and generated based on all fire state information by using the grade prediction symbol [D], wherein the explicit state refers to necessary inquiries and state replies in the alarm stage, and the implicit state refers to optional inquiries and state replies in the alarm stage; the model first obtains the explicit state of the fire, calculates the probability distribution of the implicit state by splicing the state prediction symbol [S], and inquires the state with the highest probability in the implicit state; at this time, the user will reply True, False or uncertain to the question; for uncertain information, the above steps are repeated to find the next state with the highest probability to inquire the user; according to the True or False answer of the user, the model splices the newly obtained implicit state and uses the state prediction symbol to predict the next state; when the end signal is inferred from the state symbol or when the confidence of the predicted state probability is lower than a threshold, the model stops inquiring and transitions to grade determination, and the target fire grade is predicted by splicing the grade prediction symbol [D].

2. A method of responding to a fire alarm according to claim 1, wherein The model uses a Transformer as a basic architecture, regards the alarm question and answer as a sequence generation question, and infers the next fire situation that should be inquired in a self-recurrent manner according to the obtained fire situation until the model considers that enough information has been obtained to make a fire grade judgment.

3. A method for responding to a fire alarm call as claimed in claim 1, wherein The explicit state of the fire includes the number of trapped persons and the burning area, and one or more of the following explicit properties can be added as needed: the name of the alarm person, the alarm person's phone number, the alarm time, the alarm address, the alarm person's location, the burning floor, the number of trapped persons, the number of injured persons and the degree of injury, and the smoke condition.

4. A method for responding to a fire alarm call as claimed in claim 1, wherein The implicit state includes open fire, smoke, white smoke, black smoke, explosive danger, surrounding combustible material, spread, high-rise building, collapse, completeness of fire fighting facilities, and sufficiency of water source, and the implicit state is adjusted according to the climate zone, season, time period, building density of the location, and safety alert level of the region.

5. A method for responding to a fire alarm call as claimed in claim 1, wherein, The alarm classification result includes a repeated alarm, a non-repeated alarm, a common alarm, and an important alarm; when the alarm classification result is determined as a repeated alarm, secondary confirmation of the repeated alarm is performed, and when a positive response to the repeated alarm is obtained, the current alarm communication is ended.

6. A method for responding to a fire alarm call as claimed in claim 1, wherein The interactive voice response component is used to perform a transfer question and answer, obtain a question and answer result, and when the question and answer result meets a preset condition, extract an alarm element based on the preset inquiry rule; when the question and answer result does not meet the preset condition, an artificial response is started.

7. A fire department alert and response system, characterized by The following modules are included: An alarm center module: used to receive alarm information and receive pre-alarm information, extract an alarm element based on a preset inquiry rule, and generate a pre-alarm work order including the alarm element; A fire alarm dialogue model module: a fire alarm dialogue model with a Transformer as a basic architecture is preset; the alarm element received by the alarm center module is processed, the alarm question and answer is regarded as a sequence generation question, the next fire situation that should be inquired about is inferred according to the obtained fire situation in a self-recurrence manner, and the process is continued until the model considers that enough information has been obtained to make a fire grade judgment; the above steps are executed multiple times to gradually deepen the understanding of the fire scene situation, and a real-time fire grade evaluation is generated on this basis; the fire alarm dialogue model takes the explicit state of the fire as the model input to train and generate the implicit state, where the "END" symbol represents an end signal indicating the end state generation and transition to fire classification inference; in the model training process, the prediction symbol [S] sequence is used to realize the self-recurrence generation of the implicit state sequence, which can be regarded as learning multiple stage state inferences at one time; the grade prediction symbol [D] is used to learn and generate the target grade based on all fire state information, where the explicit state refers to necessary inquiries and state replies in the alarm stage, and the implicit state refers to optional inquiries and state replies in the alarm stage; the model first obtains the explicit state of the fire, calculates the probability distribution of the implicit state by splicing the state prediction symbol [S], and inquires about the state with the highest probability in the implicit state; at this time, the user will reply True, False, or uncertain to the question; for uncertain information, the above steps are repeated to find the next highest probability state to inquire the user; according to the True or False answer of the user, the model splices the newly obtained implicit state and uses the state prediction symbol to predict the next state; when the end signal is inferred from the state symbol, or when the confidence of the predicted state probability is lower than a threshold, the model stops inquiring and transitions to grade determination, and the target fire grade is predicted by splicing the grade prediction symbol [D]; An information display module: used to obtain and display real-time information of the dangerous situation site using GIS map management technology based on information of the alarm center module, the pre-alarm module, and the fire alarm dialogue model module. The risk early warning module is used for collecting risk factors of a city monitoring area in real time, identifying and predicting risks of the monitoring area, judging early warning risk information and analyzing the probability of the risk.

8. A non-transitory storage medium, characterized in that, When the instructions in the storage medium are executed by the fire alarm response system of claim 7, the system can perform any one of the fire alarm response methods of claims 1 to 6.

Images