Fire-fighting supervision and management integrated platform based on artificial intelligence

Abstract

The invention relates to the field of fire safety, and discloses a fire supervision and management integrated platform based on artificial intelligence, the platform accurately models a fire state through a multi-modal data fusion technology in combination with a dynamic Bayesian network and a particle filter algorithm, and reduces false alarms and missing alarms through an adaptive alarm mechanism. The platform adopts a high-strength encryption and security authentication technology to ensure secure transmission of fire data, and in order to deal with the problem of network instability, the invention also designs a data caching function to ensure that important data is not lost when the network is interrupted, and the security of the fire data is ensured through real-time prediction, flexible adjustment and secure transmission. According to the invention, the fire-fighting monitoring and emergency response capability of a high-risk area can be obviously improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fire safety, and particularly to an integrated fire supervision and management platform based on artificial intelligence. Background Art

[0002] With the acceleration of the industrialization and urbanization processes, fires have become an important hidden danger in the safety management of various places. Traditional fire monitoring and alarm systems usually rely on temperature and humidity sensors, smoke detectors, and other single types of sensors to detect fire signals. Although these traditional systems have played a certain role in early fire warnings, due to the limitations of their technical architectures, they show certain deficiencies in complex environments.

[0003] Most of the existing fire monitoring systems rely on monitoring methods based on single sensors, and usually use thresholds of physical parameters such as smoke concentration, temperature, or humidity to trigger alarms. However, the application of single sensors is prone to false alarms or missed alarms in many actual situations. For example, temperature and humidity sensors may misjudge as a fire when the ambient temperature fluctuates or the air humidity changes, while smoke sensors may not be able to identify the initial stage of a fire, especially when there is no obvious smoke or high temperature in the fire. Therefore, this detection method based on a single sensor often cannot meet the changing and complex fire monitoring requirements.

[0004] At the same time, the fire status prediction models in the existing technology are relatively simple, and usually trigger alarm signals by setting fixed alarm thresholds. When certain physical parameters are detected to exceed the set thresholds, the system issues an alarm. However, due to the differences in different environments and fire types, fixed thresholds often cannot effectively adapt to all scenarios. For example, in different buildings and different areas, the occurrence mechanisms and propagation modes of fires may be completely different, which leads to the alarm sensitivity of the existing systems being too rigid to be dynamically adjusted, and prone to false alarms or missed alarms.

[0005] In terms of data transmission, most traditional fire alarm systems use relatively simple wireless or wired network transmission methods, usually lacking sufficient data encryption and transmission security guarantees. In some high-risk areas, especially in special environments such as factories and chemical warehouses, the leakage, tampering, or loss of data may lead to extremely serious consequences. The communication protocols in the existing technology usually lack sufficient security protection and cannot effectively cope with risks such as network attacks and information tampering, thus causing potential information leakage and incorrect transmission of alarm signals during the transmission process.

[0006] In addition, fire monitoring and alarm systems often need to transmit fire alarm signals and relevant data to the monitoring center in real time. However, the systems in the existing technologies usually cannot ensure the timeliness of data transmission when the network environment is unstable. In remote areas, mountainous areas or large industrial areas, due to the coverage and bandwidth problems of communication networks, fire alarm signals may miss the best response time due to network interruption or transmission delay. The system cannot perform effective real-time processing at critical moments, delaying the timeliness of fire emergency response.

[0007] Therefore, the present invention proposes an integrated platform for fire supervision and management based on artificial intelligence to solve the deficiencies of the existing technologies. Summary of the Invention

[0008] Aiming at the deficiencies of the existing technologies, the present invention provides an integrated platform for fire supervision and management based on artificial intelligence, which solves the problems of false alarms and missed alarms caused by single sensors in the existing technologies, the limitations of fire state prediction, and the deficiencies in the security and timeliness of data transmission.

[0009] To achieve the above objectives, the present invention is realized through the following technical solutions: An integrated platform for fire supervision and management based on artificial intelligence, comprising:

[0010] A data acquisition module, configured to acquire multi-modal sensor data, the data including temperature and humidity data, smoke concentration data, infrared video data, and visible light video data, and perform standardization processing on the acquired data;

[0011] A data modeling and processing module, connected to the data acquisition module, configured to establish a conditional probability relationship between the fire state and the observed data based on a dynamic Bayesian network, perform dynamic evolution modeling on time series data based on a stochastic differential equation, and recursively update the posterior probability of the fire state through a particle filtering algorithm;

[0012] A decision-making and feedback module, connected to the data modeling and processing module, configured to generate a fire alarm signal according to the posterior probability of the fire state, and adjust the model parameters of the data modeling and processing module through a feedback mechanism;

[0013] A communication module, connected to the decision-making and feedback module, configured to transmit fire alarm signals, fire state prediction data, and receive remote instructions.

[0014] Preferably, the data acquisition module includes:

[0015] Multi-type sensors, including temperature and humidity sensors, smoke sensors, infrared thermal imagers, and visible light cameras;

[0016] A data standardization unit, configured to perform unit unification, denoising processing, and time serialization processing on data from different sources;

[0017] A feature extraction unit for extracting flame features and smoke features from video data and converting them into numerical input data.

[0018] Preferably, the data modeling and processing module calculates the posterior probability of the fire state through a dynamic Bayesian network model, and the calculation includes:

[0019] Describing the dynamic relationship of the fire state changing over time according to the time transition probability of the fire state;

[0020] Establishing the correlation between multi-modal sensor data and the fire state according to the conditional probability of the observed data;

[0021] Updating the prior probability of the next moment according to the posterior probability of the fire state;

[0022] Wherein, the time transition probability is set according to historical fire data and the environmental risk level.

[0023] Preferably, the stochastic differential equation model is used to model the time dynamic changes of multi-modal sensor data, and the model includes:

[0024] A deterministic change term for describing the influence of the fire state on the observed value, and the change intensity is dynamically adjusted according to the sensor type and the fire state;

[0025] A noise term for modeling environmental interference and sensor errors, and the amplitude of the noise term is determined by experimental fitting or scenario optimization;

[0026] A prediction unit for calculating the possible values of the observed data at future moments through numerical solutions.

[0027] Preferably, the particle filter algorithm is used to recursively estimate the posterior probability of the fire state, and the algorithm includes the following steps:

[0028] Initializing a particle set, where each particle represents a possible value of the fire state and assigning uniform initial weights to the particles;

[0029] Sampling the particle states according to the time transition probability output by the dynamic Bayesian network;

[0030] Updating the weights of the particles according to the observed data and calculating the posterior probability of the fire state based on the weights of the particles;

[0031] Removing the particles with weights lower than the set threshold and generating a new particle set through resampling;

[0032] Outputting the posterior probability of the fire state according to the weighted average value of the particle set.

[0033] Preferably, the decision-making and feedback module generates an alarm signal according to the posterior probability of the fire status, and the alarm mechanism includes:

[0034] When the posterior probability of the fire status is higher than 0.8, a high-risk alarm signal is triggered;

[0035] When the posterior probability of the fire status is between 0.6 and 0.8, a medium-risk early warning signal is triggered;

[0036] When the posterior probability of the fire status is lower than 0.6, only record the fire status and maintain normal monitoring.

[0037] Preferably, the feedback optimization mechanism is used to adjust the parameters of the data modeling and processing module according to the accuracy of the fire alarm, and the adjustment rules include:

[0038] When a false alarm occurs, reduce the weight of the relevant sensors and adjust the conditional probability relationship of the dynamic Bayesian network;

[0039] When a missed alarm occurs, increase the weight of the relevant sensors and enhance the decisive change term in the stochastic differential equation model.

[0040] Preferably, the decision-making and feedback module includes a fire status visualization unit, and the visualization unit generates a fire heat map and a predicted fire spread path map and displays them through the user interface.

[0041] Preferably, the communication module sends the fire alarm signal and the fire status prediction result to the remote fire center through the wireless network and receives the dispatching instructions from the remote fire center.

[0042] Preferably, an integrated device for fire supervision and management based on artificial intelligence is applied to the platform according to any one of claims 1-9, and is characterized by including:

[0043] A data acquisition unit for acquiring multi-modal sensor data and standardizing the data;

[0044] A data modeling unit for dynamically modeling and recursively updating the fire status based on the dynamic Bayesian network, the stochastic differential equation, and the particle filter algorithm;

[0045] A decision-making unit for generating an alarm signal according to the posterior probability of the fire status and adjusting the modeling parameters;

[0046] A communication unit for sending the fire alarm signal, the prediction data, and receiving remote instructions.

[0047] The present invention provides an integrated platform for fire supervision and management based on artificial intelligence. It has the following beneficial effects:

[0048] 1. This technical solution uses a dynamic Bayesian network to recursively model the fire status and combines the particle filter algorithm to update the state probability in real time. Through these technical means, the present invention can accurately calculate the occurrence probability of a fire based on multi-modal sensor data (such as temperature, humidity, smoke concentration, video surveillance data). Compared with the fire monitoring systems using single sensors or rule-based judgments in the prior art, traditional solutions usually rely on only one data source, such as temperature and humidity sensors, resulting in their inability to comprehensively consider various influencing factors in complex environments and often causing false alarms or missed alarms. However, by integrating multi-source data, the present invention significantly reduces the occurrence of such situations, provides more accurate fire monitoring and early warning capabilities, and can more flexibly respond to changes in the occurrence of fires under different environments and conditions.

[0049] 2. The alarm mechanism of the present invention is designed with triple responses (high-risk alarm, medium-risk early warning, low-risk monitoring), and can accurately trigger the alarm system according to the probability of a fire occurring. In the initial stage of a fire, the system uses a lower-sensitivity alarm threshold to avoid false alarms, and when the fire risk increases significantly, the system will immediately adjust the alarm threshold to provide a more timely response. At the same time, the feedback optimization mechanism enables the system to adaptively adjust the modeling parameters according to the alarm results. For example, it dynamically adjusts the weights of sensors and changes the time transition probability of the fire status according to false alarms and missed alarms. This design is superior to the systems relying on fixed alarm thresholds in the prior art, can make real-time responses to the dynamic changes of different fire scenarios and environments, thus significantly reducing the frequency of false alarms and missed alarms, and improving the accuracy and response speed of alarms.

[0050] 3. By adopting the AES encryption algorithm, TLS protocol, and two-way authentication mechanism, the present invention ensures the confidentiality and integrity of fire data during transmission. All fire alarm signals and prediction data are encrypted to prevent being stolen or tampered with in an open network environment. Compared with the relatively simple encryption methods in traditional systems, the high-strength encryption technology and multi-layer security mechanisms of the present invention effectively avoid possible network attacks, data leakage, and tampering problems, ensuring the reliable transmission of information. Especially when it comes to fire early warning information, any leakage or tampering of information may lead to catastrophic consequences, while the present invention can ensure the security and stability of the fire alarm system through end-to-end encryption and authentication.

[0051] 4. In some special environments, the communication network may be unstable. For example, in remote areas or disaster sites, the network connection is prone to interruption, which may cause delays or losses in real-time data transmission. To address this issue, the present invention adopts a data caching mechanism that automatically caches fire alarm signals, status data, and remote instructions when the network is interrupted, and ensures that the data can be seamlessly synchronized to the target terminal after the network is restored. Through this solution, the system will not lose important data due to network interruption, ensuring the continuity and stability of fire response. Compared with the existing solutions lacking effective caching and fault tolerance mechanisms, the caching mechanism of the present invention greatly improves the reliability of the system in an unstable network environment, avoids potential risks caused by data loss, and can flexibly adjust the system behavior through remote instructions to ensure real-time fire monitoring and response capabilities even in harsh environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is an architecture diagram of an integrated platform for fire supervision and management based on artificial intelligence. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] Next, the technical solutions of the present invention will be clearly and completely described in conjunction with the drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0054] Please refer to the attached Figure 1 , an embodiment of the present invention provides an integrated platform for fire supervision and management based on artificial intelligence. The following will detail each module of the platform of the present invention.

[0055] In the integrated platform for fire supervision and management of the present invention, the data acquisition module is the starting point of the entire system. Its main function is to collect multi-modal fire monitoring data and perform standardization and feature extraction processing on it to provide high-quality input data for the subsequent data modeling and processing module. The performance of the data acquisition module directly determines the overall accuracy and real-time nature of fire monitoring. This module ensures the stable operation of the platform in complex environments through the collaborative work of multiple sensors and the combination of multi-modal data processing technologies.

[0056] In this embodiment, the data acquisition module specifically includes the following components and functional implementations:

[0057] Data acquisition unit

[0058] Generally, the data acquisition unit consists of multiple types of sensors, which are distributed at key positions in the monitoring area according to different data requirements. The types of sensors include but are not limited to the following:

[0059] Temperature and humidity sensor

[0060] It is used to measure the temperature and humidity changes in the environment in real time. The temperature range is usually [-40°C, 120°C], and the humidity range is [0%, 100%]. The sensor should be installed away from high-temperature equipment or humid areas to reduce the impact of environmental interference. For example, when a fire occurs, the temperature rise rate can reach 10°C / s or higher, and the humidity may decrease rapidly, and these dynamic characteristics provide an important basis for subsequent modeling.

[0061] Smoke sensor

[0062] It is used to detect the concentration of smoke particles in the air, with the unit of ppm (parts per million concentration). As an option, the sensitivity range of this sensor can be dynamically adjusted between 1 ppm - 10 ppm. For example, when monitoring industrial areas, a high-sensitivity mode can be selected; while in ordinary indoor environments, the sensitivity can be reduced to reduce false alarms.

[0063] Infrared thermal imager

[0064] Specifically, this sensor is used to capture infrared image data of heat sources in the monitoring area, and its resolution is usually 640×480 pixels or higher. The infrared data can provide two-dimensional distribution information of the surface temperature of the target area, such as local abnormal temperature rise signals generated in the early stage of a fire. The dynamic range of the infrared thermal imager is [-20°C, 1500°C], and the detection sensitivity can reach 0.1°C.

[0065] Visible light camera

[0066] It is used to obtain visible light video data of the monitoring area, with a resolution of usually 1920×1080 pixels and a frame rate of 30 fps or higher. This device is mainly used to capture the dynamic characteristics of flames and smoke. The installation position of the visible light camera should avoid direct sunlight areas to reduce optical distortion of the video signal.

[0067] In some embodiments, the data acquisition frequency of the sensor can be adjusted according to the monitoring scenario. For example, in high-risk areas, the sampling frequencies of the temperature and humidity sensor and the smoke sensor can be set to 10 Hz or higher; while in ordinary areas, the sampling frequency can be reduced to 1 Hz to save computing resources.

[0068] Data normalization unit

[0069] As an option, the data normalization unit in the data acquisition module is used to process data from different types of sensors to ensure its consistency and high quality. The main functions of the data normalization unit include:

[0070] Unit conversion

[0071] The output data of different sensors usually have different units. For example, temperature is expressed in degrees Celsius (°C), smoke concentration is expressed in ppm, and video data is expressed in pixel matrices. The purpose of unit conversion is to unify all data into a standard format that can be processed by the system.

[0072] For example, the smoke concentration data can be standardized through the formula

[0073]

[0074] where C std : the standardized concentration value; C raw : the original concentration value; C min : the lowest concentration detection value of the sensor; C max : the highest concentration detection value of the sensor.

[0075] Denoising processing

[0076] Generally, the data output by sensors will be interfered by environmental noise. For example, a smoke sensor may give false alarms due to water vapor. The denoising processing uses the Kalman filtering algorithm, and its state update formula is:

[0077] X k = X k-1 + K k (Z k - HX k-1 )

[0078] where X k : the current state estimate value; X k-1 : the previous state estimate value; Z k : the current observation value; H: the observation matrix; K k : the Kalman gain, which is calculated from the error covariance.

[0079] Time serialization

[0080] The data normalization unit aligns the data of different sensors through timestamps to ensure that the multi-modal data has a consistent time resolution. As an implementation method, the resolution of time serialization can be set to 100 ms or 1 s according to the monitoring scenario.

[0081] Feature extraction unit

[0082] The feature extraction unit mainly processes the flame and smoke features in video data. Since video data contains a large amount of invalid information, it is necessary to extract the key features related to fire through a deep learning model

[0083] Flame Feature Extraction: Flames usually have unique color and dynamic shape change characteristics. Flame features are extracted through a Convolutional Neural Network (CNN), and its loss function uses cross-entropy loss:

[0084]

[0085] where L is the loss value; y i : the true label of the i-th sample; the flame feature value predicted by the model; N: the number of samples.

[0086] Smoke Feature Extraction: The diffusion pattern of smoke shows low-level evolution and gradually increasing dynamic characteristics. When extracting smoke features, the inter-frame difference image can be calculated:

[0087] ΔI = |I t - I t-1 |

[0088] where ΔI is the inter-frame difference image; I t : the video frame at time t; I t-1 : the video frame at time t-1.

[0089] Multi-modal Data Collaborative Acquisition

[0090] In some embodiments, the collaborative acquisition of sensors can be dynamically triggered by logical rules. For example, when the smoke concentration exceeds the threshold, the infrared thermal imager will start in high frame rate mode to capture the possible fire source location. Significant changes in temperature and humidity data can also trigger the high-frequency sampling mode of video data for analyzing the specific source of the fire signal.

[0091] The standardized and feature-extracted data generated by the data acquisition module will be transmitted to the data modeling and processing module in a unified format. Among them, numerical data (such as temperature, humidity, smoke concentration) is directly used as the input of the dynamic Bayesian network and stochastic differential equation; the feature extraction results of video data are used as auxiliary information for updating the particle filter weights. This connection method ensures the efficient utilization of multi-modal data in subsequent modeling.

[0092] Through the above detailed description, the data acquisition module realizes the complete functions from data acquisition to standardized processing and then to feature extraction, and can provide high-quality multi-modal data input for the fire supervision and management integration platform of the present invention.

[0093] Data Modeling and Processing Module

[0094] The data modeling and processing module is an important core part of the present invention to achieve intelligent fire supervision and management. This module is responsible for comprehensively analyzing and dynamically modeling the multi-modal data collected from the data acquisition module, and generating the posterior probability value P(H t |E1:t ), providing reliable input for subsequent decision-making and feedback modules. This module combines the Dynamic Bayesian Network (DBN), Stochastic Differential Equation (SDE), and Particle Filter (PF) to achieve recursive estimation, dynamic evolution, and real-time update of the fire state. Generally, the standardized data received by this module from the data acquisition module includes temperature and humidity, smoke concentration, as well as flame and smoke characteristics extracted from video images. After modeling, these data generate a high-precision fire state probability.

[0095] Dynamic Bayesian Network Modeling

[0096] In this embodiment, the Dynamic Bayesian Network is used to establish the conditional probability relationship between the fire state H t and the observed data E t . Through recursive calculation, the Dynamic Bayesian Network model obtains the posterior probability P(H t |E 1:t ) of the fire state.

[0097] Its recursive formula is:

[0098]

[0099] where P(H t |E 1:t ): the posterior probability of the fire state at time t; P(E t |H t ): the likelihood probability of the current observed data E t under the fire state H t ; P(H t |H t-1 ): the time transition probability of the fire state, indicating the relationship between the fire state H t-1 and H t ; P(E t ): the marginal probability of the current observed data, used for normalization.

[0100] The time transition probability P(H t |H t-1 ) is modeled using an exponential distribution, and its formula is:

[0101]

[0102] where |H t - H t-1 |: represents the change in the fire state over time; τ: the time smoothing coefficient, used to adjust the smoothness of the change in the fire state.

[0103] The observed likelihood probability P(E t |H t )

[0104] Observation likelihood \(P(E t |H t ) represents the probability of the observation data \(E t occurring under the fire state \(H t . Assuming the conditional independence of the observation data, the likelihood probability can be expressed as:

[0105]

[0106] where \(n\): the number of sensors; \(E t,i : the observation value of the \(i\)-th sensor at time \(t\).

[0107] For a single observation value \(E t,i , it is modeled using a Gaussian distribution, and its formula is:

[0108]

[0109] where \(\mu i (H t ) is the expected value of the observation of the \(i\)-th sensor under the fire state \(H t ; the variance of the observation value, representing the sensor noise level.

[0110] State prior update

[0111] The prior probability of the fire state is updated from the posterior probability of the previous moment:

[0112] P(H t |H t-1 ) \(\propto P(H t-1 |E 1:t-1 )P(H t |H t-1 )

[0113] In some embodiments, the parameters of the time transition probability \(P(H t |H t-1 ) can be dynamically adjusted according to the monitoring scenario. For example, in a high-risk area, a lower time smoothing coefficient \(\tau\) can be set to improve the response ability to rapid changes in the fire state.

[0114] Stochastic differential equation modeling

[0115] In this embodiment, the stochastic differential equation is used to describe the dynamic evolution process of multi-modal sensor data over time, especially to model the changes in physical quantities such as temperature and smoke concentration. The mathematical form of the stochastic differential equation is:

[0116] dE t,i = f(E t,i ,t)dt + g(E t,i ,t)dWt

[0117] Among them, E t,i : the i-th observed variable, such as temperature, smoke concentration; f(E t,i ,t): the deterministic change term, representing the direct impact of the fire state on the observed variable; g(E t,i ,t): the noise term, representing random interference; W t : Brownian motion, describing randomness.

[0118] The deterministic change term f(E t,i ,t)

[0119] The deterministic change term f(E t,i ,t) is determined by the fire state H t and the response coefficient α i of the observed variable:

[0120] f(E t,i ,t) = α i H t

[0121] Among them, α i : the response coefficient of the i-th sensor, representing the degree of influence of the fire state on its observed value.

[0122] The noise term g(E t ,i,t)

[0123] The noise term g(E t,i ,t) depends on the accuracy of the sensor, and its form is usually 3

[0124] g(E t,i ,t) = β i

[0125] Among them, β i : the noise amplitude, reflecting the inherent error of the sensor.

[0126] Numerical solution

[0127] The stochastic differential equation is discretized by the Euler-Maruyama method to obtain the following form:

[0128]

[0129] Among them, Δt: time step; ξ: a random variable sampled from the standard normal distribution.

[0130] Particle filter recursive update

[0131] The particle filter is an important algorithm for recursively estimating the posterior probability of the fire state. Its main steps include particle initialization, state prediction, weight update, and resampling.

[0132] Particle initialization

[0133] At the initial moment, N particles are generated, and each particle represents a possible value of the fire state The weights are initialized to a uniform distribution:

[0134]

[0135] State prediction

[0136] According to the time transition probability P(H t |H t-1 ), the state of the particles is predicted to obtain new particles

[0137] Weight update

[0138] According to the likelihood probability of the observed data P(E t |H t ), the weight of each particle is updated:

[0139]

[0140] The weights are normalized to:

[0141]

[0142] Resampling

[0143] Particles with weights below the threshold are removed, and a new particle set is regenerated according to the normalized weights

[0144] Posterior probability calculation

[0145] The posterior probability of the fire state is obtained by weighted averaging of the particles:

[0146]

[0147] The output of the data modeling and processing module is the posterior probability P[H t |E 1:t ). These results are passed to the decision-making and feedback module to trigger the alarm mechanism or for further analysis. At the same time, the weight information in the particle filter update process can assist in optimizing the dynamic adjustment of the alarm threshold

[0148] Through the above detailed disclosure, the data modeling and processing module of the present invention realizes accurate modeling and recursive prediction of the fire state

[0149] Decision-making and feedback module

[0150] The decision-making and feedback module is the key control part of the present invention, and its main function is to calculate the posterior probability P(Ht |E 1:t ) is analyzed to generate corresponding alarm signals. Meanwhile, the modeling parameters are adjusted through false alarm and missed alarm feedback to improve the accuracy and robustness of the overall system. This module is closely connected to the data modeling and processing module, uses the modeling results to drive the alarm response, and realizes parameter adaptive optimization through the analysis of historical data to form a dynamic closed-loop control system.

[0151] In this embodiment, the decision-making and feedback module includes a multi-level alarm mechanism, a feedback optimization mechanism, dynamic threshold adjustment, and an interaction part with other modules. Its technical implementation method is as follows.

[0152] Multi-level alarm mechanism

[0153] In this embodiment, the decision-making and feedback module uses the posterior probability P(H t |E 1:t ) to implement multi-level alarm determination. The alarm mechanism includes three levels: high-risk alarm, medium-risk early warning, and low-risk monitoring.

[0154] Generally, the high-risk alarm is triggered when the posterior probability P(H t |E 1:t ) exceeds the threshold P high . The threshold P high is usually set to 0.8 to ensure that the system responds immediately after the fire risk reaches a high confidence level. The medium-risk early warning is triggered when the posterior probability P(H t |E 1:t ) is within the threshold interval [P mid1 , P mid2 , and the common setting is [0.6, 0.8]. The low-risk monitoring is triggered when the posterior probability P(H t |E 1:t ) is lower than P low (such as 0.6), and only records the fire status without generating external alarm signals.

[0155] As a specific implementation method, the judgment logic of the alarm level can be expressed as:

[0156] Among them, P high : The posterior probability threshold of the high-risk alarm; P mid1 , P mid2 : The threshold interval of the medium-risk early warning; P low : The upper limit threshold of the low-risk monitoring.

[0157] Generation of alarm signals. The alarm signals include the following information

[0158] The posterior probability of the fire status P(H t |E1:t );

[0159] Alarm level (high risk, medium risk or low risk);

[0160] Possible location of the fire occurrence;

[0161] Dynamic prediction data, such as the direction and scope of fire spread.

[0162] As an option, the alarm signal is sent to the remote fire center or the user device in an encrypted form through the communication module to ensure information security and transmission stability. Specifically, the encryption of the alarm signal can adopt the Advanced Encryption Standard (AES) with a key length of 256 bits.

[0163] Feedback optimization mechanism

[0164] The feedback optimization mechanism is used to analyze the accuracy of the alarm, adjust the modeling parameters to improve the overall performance of the system. The feedback mechanism of this module includes false alarm handling and missed alarm handling.

[0165] False alarm handling

[0166] When the system generates a false alarm, it is necessary to reduce the weight of the relevant sensors to reduce interference. In the conditional probability formula of the dynamic Bayesian network model

[0167] Introduce a noise adjustment factor The adjustment rule is:

[0168]

[0169] where μ i (H t ) is the expected value of the i-th sensor in the fire state H t ; and the variance of the observed value, which is used to represent the noise level. The variance of the observed value, which is used to represent the noise level.

[0170] In the case of a false alarm, by increasing the value, the weight of the sensor in the conditional probability calculation can be reduced, thereby reducing its impact on the fire state.

[0171] The adjustment rule is:

[0172]

[0173] where Δσ > 0, representing the adjustment amplitude, which is dynamically determined by the false alarm frequency.

[0174] Missed alarm handling

[0175] In the case of a missed alarm, it is necessary to improve the sensitivity and weight of the sensor to enhance the discrimination ability of the fire state. The adjustment rule is:

[0176] μ i (H t ) ← μ i (H t ) - Δμ

[0177] Among them, Δμ > 0, indicating the adjustment amount of the expected value.

[0178] This adjustment makes the influence of the observed value deviating from the expected value on the fire state more significant, thereby reducing the false alarm probability.

[0179] Dynamic threshold adjustment

[0180] Generally, the alarm threshold P high , P mid1 , P mid2 , P low is fixed. However, in some embodiments, the threshold can be dynamically adjusted according to the environmental risk. For example, in high-risk areas such as chemical plants, the high-risk alarm threshold P high is reduced to 0.7; while in low-risk areas (such as ordinary residential areas), the low-risk monitoring threshold P low is increased to 0.65.

[0181] The mathematical formula for dynamic adjustment is:

[0182] P' high = P high × (1 - R)

[0183] P low = P low × (1 + R)

[0184] Among them, P' high , P' low : the adjusted threshold; R: the risk factor, with a value range of [0, 0.2], determined by the environmental risk level.

[0185] Specifically, in forest fire monitoring, the value of R can be dynamically adjusted according to seasonal factors. For example, in summer, the forest is dry and flammable, and the risk factor R can be taken as 0.15; while in the wet winter, R can be taken as 0.05.

[0186] The influence of the threshold on the feedback mechanism

[0187] During the feedback optimization process, the dynamic adjustment of the alarm threshold can also be linked with the modeling parameters. Generally, reducing the alarm threshold will increase the frequency of high-risk alarms. Therefore, it is necessary to correspondingly increase the time transfer smoothing coefficient τ of the dynamic Bayesian network.

[0188] The adjustment rule is:

[0189] τ ← τ × (1 + γ)

[0190] Among them, γ > 0 represents the adjustment ratio; τ is the time smoothing coefficient, which is used to smooth the dynamic changes of the fire state.

[0191] Through the above-mentioned linkage adjustment, the system achieves a balance between response sensitivity and robustness.

[0192] This module receives the posterior probability P(H t |E 1:t ) of the fire state from the data modeling and processing module. The generation of the alarm signal and the parameter optimization of the feedback result directly affect the data modeling and processing module. For example, the false alarm feedback result will dynamically adjust the noise term g(E t,i , t) in the stochastic differential equation model, and the missed alarm feedback result will optimize the particle weight distribution in the particle filter algorithm. In addition, the alarm signal of this module is transmitted to the remote fire center through the communication module to realize the visualization of the fire state and the alarm linkage.

[0193] Through the above detailed description, the decision-making and feedback module realizes the hierarchical response of the fire alarm, the closed-loop control of parameter optimization, and the adaptive adjustment of the system performance.

[0194] Communication module

[0195] In the present invention, the communication module is responsible for the transmission of the fire alarm signal, the sharing of the fire state data, and the reception and feedback of the remote instructions. As the information interaction bridge of the entire system, the communication module not only supports multiple network protocols to adapt to different application scenarios, but also ensures the transmission security and reliability through data encryption and integrity verification technologies. The collaborative work of this module with other modules ensures the smooth flow of information in the complex environment. In this embodiment, the technical implementation of the communication module includes parts such as data transmission function, data encryption and security authentication, remote instruction processing, network protocol adaptability, and data cache management.

[0196] Data transmission function

[0197] In this embodiment, the core function of the communication module is to transmit the fire alarm signal, the posterior probability P(H t |E 1:t ) of the fire state, and the fire trend prediction data generated from the decision-making and feedback module to the target terminal, and receive external instructions and feedback them to the internal modules of the system.

[0198] Generally, the format of the alarm signal includes:

[0199] Posterior probability of the fire state: For example, P(H t |E 1:t ) = 0.92, indicating a high-risk alarm;

[0200] Alarm level: including high risk, medium risk or low risk;

[0201] Fire location: Mark the specific location through GPS or Internet of Things nodes;

[0202] Fire spread prediction data: including the propagation speed v fire , the propagation direction θ fire and the possibly affected area A impact .

[0203] Specifically, the fire spread speed can be estimated according to the time evolution of the fire state, and the formula is:

[0204]

[0205] where v fire : the fire spread speed; ΔH t : the change in the fire state; Δt: the time interval.

[0206] The fire propagation direction θ fire is calculated according to the spatial gradient of the change in the fire state, and the formula is:

[0207]

[0208] where are the gradients of the fire state in the x and y directions.

[0209] The possibly affected area A of the fire impact can be calculated according to the fire spread speed and the spread range:

[0210]

[0211] where R fire = v fire × t predict : the fire spread radius; t predict : the prediction time.

[0212] These data are sent to the remote fire center or user-end devices through the MQTT protocol. As an option, the HTTP / HTTPS protocol can be used to support real-time data query.

[0213] Data Encryption and Security Authentication

[0214] To ensure the security of the transmitted data, the communication module adopts a multi-layer encryption mechanism. Generally, AES (Advanced Encryption Standard) is selected for data encryption, and the encryption formula is:

[0215] C = E k (M)

[0216] where C: the encrypted ciphertext; ·M: the original plaintext data; ·E k: Encryption function, based on the key k.

[0217] At the receiving end, the decryption formula is:

[0218] M = D k (C)

[0219] where D k : Decryption function, corresponding to the encryption key k.

[0220] As an option, the communication module also supports the TLS (Transport Layer Security) authentication mechanism to verify the identities of both communication parties through digital certificates. Generally, the authentication process of the TLS protocol includes the following steps:

[0221] Certificate exchange: Both communication parties exchange digital certificates;

[0222] Key negotiation: Generate a shared key through the Diffie-Hellman algorithm;

[0223] Data encryption: Encrypt data transmission based on the shared key.

[0224] The formula for Diffie-Hellman key exchange is:

[0225] K = g ab mod p

[0226] where K: Shared key; g: Generator; ·a, b: Private keys of both parties; p: Prime modulus.

[0227] In this way, the communication module achieves the confidentiality, integrity, and anti-eavesdropping capabilities of data transmission.

[0228] Remote instruction processing

[0229] The communication module is not only responsible for data transmission but also supports receiving remote instructions for dynamically adjusting system parameters or triggering specific operations.

[0230] The remote instruction format includes the following fields:

[0231] Instruction type: Such as adjusting the alarm threshold, modifying the sensor sampling frequency, starting the emergency mode, etc.; Target module: Specify the module to be adjusted, such as the data modeling and processing module; Parameter value: Provide the new threshold or adjustment range.

[0232] As a specific implementation method, after the instruction is received, it needs to pass an integrity check to ensure that no tampering has occurred during the transmission process. Generally, the integrity check uses the CRC (Cyclic Redundancy Check) algorithm, and its formula is:

[0233] R(x) = P(x) mod G(x)

[0234] Among them, R(x): check remainder; P(x): polynomial representation of the data to be checked; G(x): generating polynomial.

[0235] When R(x) = 0, the data is determined to be complete; otherwise, it is determined to be incorrect.

[0236] Network protocol adaptability

[0237] The communication module supports multiple communication protocols to adapt to different application scenarios and network conditions. Generally, the following protocols are supported:

[0238] MQTT: Suitable for lightweight data transmission scenarios, such as Internet of Things monitoring nodes

[0239] HTTP / HTTPS: Used for real-time data query of large-scale terminals:

[0240] LoRa: Suitable for long-distance and low-power data transmission, with a transmission rate of 0.3 kbps to 50 kbps and a coverage range of up to 15 kilometers;

[0241] 5G: Suitable for high-bandwidth and low-latency application scenarios, supporting real-time video stream transmission and high-frequency instruction interaction.

[0242] In a possible implementation, the communication module dynamically switches protocols according to network conditions. For example, in forest fire monitoring, long-distance nodes prefer to select the LoRa protocol; while in the urban fire protection system, 5G or Wi-Fi networks are preferred to ensure low latency.

[0243] Data cache management

[0244] To cope with network interruptions or transmission delays, in this embodiment, the communication module designs a data caching function. Generally, the cached content includes alarm signals, posterior probability P(H t |E 1:t ), fire spread data, and received remote instructions. The data cache is managed using a circular queue, and the cache depth is dynamically adjusted according to the storage capacity, usually ranging from 100 to 1000 records. As a specific implementation, the cached data is sorted by timestamp, and the latest records are preferentially transmitted. The data cache policy formula is as follows:

[0245] Priority = t current -t timestamp

[0246] Among them, t current : current time; t timestamp : timestamp of the data record.

[0247] When the network resumes, the communication module sends the cached data to the target terminal in priority order.

[0248] The communication module receives the alarm signal generated by the decision-making and feedback module, and combines the prediction data provided by the data modeling and processing module to optimize the alarm content.

[0249] The present invention also provides an integrated fire supervision and management device based on artificial intelligence. The following will describe the specific implementation manners of each device unit in combination with the working process of the device of the present invention.

[0250] The data acquisition unit is used to collect multi-modal sensor data and standardize the data;

[0251] The data modeling unit is used to dynamically model and recursively update the fire state based on the dynamic Bayesian network, stochastic differential equation and particle filter algorithm;

[0252] The decision-making unit is used to generate an alarm signal according to the posterior probability of the fire state and adjust the modeling parameters;

[0253] The communication unit is used to send fire alarm signals, prediction data and receive remote instructions.

[0254] The device of this embodiment can be used to execute the above platform embodiment, and its principle and technical effects are similar, so they will not be elaborated here.

[0255] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An integrated fire supervision and management platform based on artificial intelligence, characterized in that: include: A data acquisition module is used to collect multimodal sensor data, including temperature and humidity data, smoke concentration data, infrared video data and visible light video data, and perform standardization on the collected data; A data modeling and processing module, connected to the data acquisition module, is used to establish a conditional probability relationship between the fire state and the observed data based on a dynamic Bayesian network, to dynamically evolve the time series data based on a stochastic differential equation, and to recursively update the posterior probability of the fire state through a particle filter algorithm; A decision and feedback module, connected to the data modeling and processing module, for generating a fire alarm signal according to the posterior probability of the fire state, and adjusting the model parameters of the data modeling and processing module through a feedback mechanism; The communication module is connected to the decision and feedback module and is used to transmit fire alarm signals, fire status prediction data and receive remote commands.

2. According to claim 1, the fire supervision management integrated platform based on artificial intelligence is characterized in that: The data acquisition module comprises: Multiple types of sensors, including temperature and humidity sensors, smoke sensors, infrared thermal imagers, and visible light cameras; Data standardization unit, used to unify units, remove noise and process time series of data from different sources; The feature extraction unit is used to extract flame features and smoke features from the video data and convert them into numerical input data.

3. According to the artificial intelligence-based fire supervision and management integrated platform of claim 1, it is characterized in that: The data modeling and processing module calculates the posterior probability of the fire state through a dynamic Bayesian network model, and the calculation includes: According to the time transfer probability of fire state, the dynamic relationship of fire state changing with time is described; According to the conditional probability of the observed data, the association between multimodal sensor data and fire status is established; According to the posterior probability of the fire state, update the prior probability at the next moment; Wherein, the time transfer probability is set according to historical fire data and environmental risk level.

4. According to the artificial intelligence-based fire supervision and management integrated platform of claim 1, it is characterized in that: The stochastic differential equation model is used to model the temporal dynamic changes of multimodal sensor data, and the model includes: The decisive change term is used to describe the impact of the fire state on the observed value, and the change intensity is dynamically adjusted according to the sensor type and fire state; A noise term, used to model environmental disturbances and sensor errors, wherein the magnitude of the noise term is determined by experimental fitting or scenario optimization; The prediction unit is used to calculate the possible value of the observed data at a future time through a numerical solution.

5. According to the artificial intelligence-based fire supervision and management integrated platform of claim 1, it is characterized in that: The particle filter algorithm is used to recursively estimate the posterior probability of the fire state, and the algorithm includes the following steps: Initialize a collection of particles, each particle represents a possible value of the fire state, and assign uniform initial weights to the particles; The particle states are sampled according to the time transition probability output by the dynamic Bayesian network; Update the particle weights according to the observed data, and calculate the posterior probability of the fire state based on the particle weights; Particles with weights lower than the set threshold are eliminated, and a new set of particles is generated by resampling; The posterior probability of the fire state is output based on the weighted average of the particle ensemble.

6. According to the artificial intelligence-based fire supervision and management integrated platform of claim 1, it is characterized in that: The decision and feedback module generates an alarm signal according to the posterior probability of the fire state. The alarm mechanism includes: When the posterior probability of the fire state is higher than 0.8, a high-risk alarm signal is triggered; When the posterior probability of the fire state is between 0.6 and 0.8, a medium-risk warning signal is triggered; When the posterior probability of the fire state is lower than 0.6, only the fire state is recorded and normal monitoring is maintained.

7. The fire supervision and management integrated platform based on artificial intelligence according to claim 1 is characterized in that: The feedback optimization mechanism is used to adjust the parameters of the data modeling and processing module according to the accuracy of the fire alarm, and the adjustment rules include: When a false alarm occurs, the weight of the relevant sensor is reduced and the conditional probability relationship of the dynamic Bayesian network is adjusted; When false negatives occur, the weights of the relevant sensors are increased and the deterministic change terms in the stochastic differential equation model are enhanced.

8. The fire supervision management integrated platform based on artificial intelligence according to claim 1 is characterized in that: The decision and feedback module includes a fire status visualization unit, which generates a fire thermal map and a fire propagation path prediction map, and displays them through a user interface.

9. The fire supervision and management integrated platform based on artificial intelligence according to claim 1 is characterized in that: The communication module sends the fire alarm signal and the fire status prediction result to the remote fire fighting center through the wireless network, and receives the dispatching instruction of the remote fire fighting center.

10. An integrated fire supervision and management device based on artificial intelligence, applied to the platform according to any one of claims 1 to 9, characterized in that: include: A data acquisition unit, used to collect multimodal sensor data and standardize the data; A data modeling unit for dynamically modeling and recursively updating fire status based on dynamic Bayesian networks, stochastic differential equations, and particle filter algorithms; A decision-making unit, used to generate an alarm signal and adjust modeling parameters according to the posterior probability of the fire state; Communication unit, used to send fire alarm signals, prediction data and receive remote commands.

Images