A big data fire safety monitoring management method and system

Abstract

This invention discloses a big data-based fire safety monitoring and management method and system. The method involves sequentially cleaning, denoising, and normalizing the collected data to obtain standard multimodal data. An improved multi-scale convolutional neural network is used to extract environmental trend features, spatial distribution features, and temporal periodic features from the standard multimodal data. These features are then aligned and fused to obtain feature multimodal data. A risk state assessment model based on a CTHMM continuous-time hidden Markov model is established, defining the system's safety state as a hidden state. The feature multimodal data is used as an observation sequence to output a fire risk assessment index. Based on the fire risk assessment index, a hierarchical decision-making mechanism is used to classify fire risks and trigger corresponding early warning mechanisms according to the risk level. This transforms fire management from passive response to proactive early warning, significantly reducing the probability of fire occurrence.

Description

Technical Field

[0001] This invention relates to the field of forestry ecological restoration monitoring technology, and in particular to a big data-based fire safety monitoring and management method and system. Background Technology

[0002] Currently, fire safety monitoring mainly relies on single-sensor technologies, such as smoke sensors and infrared flame detectors. Single-modal detection is susceptible to environmental interference, leading to high false alarm rates. It cannot distinguish between real fires and environmental disturbances, lacks real-time monitoring of fire equipment status, and cannot predictively assess fire risks, making it difficult to cope with complex and changing environmental conditions. While existing technologies typically employ fixed-weight multimodal fusion, they cannot dynamically adjust to the environment and lack predictive maintenance capabilities, hindering the improvement of the efficiency and accuracy of intelligent fire safety monitoring. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing a big data-based fire safety monitoring and management method and system.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned big data-based fire safety monitoring and management method:

[0005] Collect environmental perception data, equipment status data, building and personnel dynamic data, and external and historical data. The collected data are then cleaned, denoised, and normalized to obtain standard multimodal data.

[0006] An improved multi-scale convolutional neural network is used to extract environmental trend features, spatial distribution features, and temporal periodic features from the standard multimodal data, and the features are aligned and fused to obtain feature multimodal data.

[0007] A risk status assessment model based on the CTHMM continuous-time hidden Markov model is established, the safety status of the system is defined as the hidden state, the feature multimodal data is used as the observation sequence, and the fire risk assessment index is output.

[0008] Based on the aforementioned fire risk assessment index, a hierarchical decision-making mechanism is used to classify fire risks, and a corresponding level of early warning mechanism is triggered according to the risk level.

[0009] Furthermore, in the aforementioned big data-based fire safety monitoring and management method, the collection of environmental perception data, equipment status data, building and personnel dynamic data, and external and historical data involves sequentially cleaning, denoising, and normalizing the collected data to obtain standard multimodal data, including:

[0010] Collect environmental perception data, equipment status data, building and personnel dynamic data, and external and historical data. Environmental perception data includes temperature, humidity, smoke concentration, harmful gas concentration, and light intensity. Equipment status data includes equipment operating parameters, fault codes, and maintenance records. Building and personnel dynamic data includes the distribution of fire protection facilities, personnel density, and movement trajectories. External and historical data includes fire cases, hazard records, and early warning and response results.

[0011] Linear interpolation is used to complete short-term missing data, abnormal data is identified based on the 3σ principle, wavelet transform is used to denoise and retain real abrupt change signals, Kalman filtering is used to smooth video trajectory data, occlusion-induced jump points are removed, and high-frequency fluctuation parameters are averaged to obtain preprocessed data.

[0012] Preprocessed data is mapped to the [0,1] interval based on Min-Max normalization to obtain standard multimodal data.

[0013] Furthermore, in the aforementioned big data-based fire safety monitoring and management method, the improved multi-scale convolutional neural network is used to extract the environmental trend features, spatial distribution features, and temporal periodic features of the standard multimodal data, and the features are aligned and fused to obtain feature multimodal data, including:

[0014] The improved multi-scale convolutional neural network sets up three types of convolutional kernels and adds an SE attention mechanism after the convolutional layer to enhance the weights of key features;

[0015] Temperature and humidity time series data are processed by 1×3 convolution kernels to extract rising and falling trends and obtain environmental trend features;

[0016] High-risk area clustering features are extracted by convolving the building area grid data with 3×3 convolution kernels to obtain spatial distribution features;

[0017] Based on the 5×5 convolution kernel combined with the sliding window, the daily and weekly regular fluctuations are extracted to obtain the time-series periodic features.

[0018] Furthermore, in the aforementioned big data-based fire safety monitoring and management method, the improved multi-scale convolutional neural network is used to extract the environmental trend features, spatial distribution features, and temporal periodic features of the standard multimodal data, and the features are aligned and fused to obtain feature multimodal data, including:

[0019] By unifying the step size of environmental trend characteristics, spatial distribution characteristics and temporal periodic characteristics at different sampling frequencies through linear interpolation, unified feature data is obtained.

[0020] The unified feature data is concatenated along the channel dimension, and the correlation between features is learned through the Cross-Attention mechanism to output multimodal feature data.

[0021] Furthermore, in the aforementioned big data-based fire safety monitoring and management method, the establishment of a risk state assessment model based on a CTHMM continuous-time hidden Markov model, defining the system's safety state as a hidden state, using the characteristic multimodal data as an observation sequence, and outputting a fire risk assessment index includes:

[0022] The risk state assessment model based on the CTHMM continuous-time hidden Markov model divides the safe state into four hidden states: safe, low risk, medium risk and high risk.

[0023] The feature multimodal data is mapped to a probability distribution using a Gaussian mixture model (GMM), a transition probability matrix between hidden states is defined, and historical fire data is used for initialization.

[0024] Furthermore, in the aforementioned big data-based fire safety monitoring and management method, the establishment of a risk state assessment model based on a CTHMM continuous-time hidden Markov model, defining the system's safety state as a hidden state, using the characteristic multimodal data as an observation sequence, and outputting a fire risk assessment index includes:

[0025] The transition probability and observation probability of CTHMM are iteratively optimized using the EM expectation-maximization algorithm until the log-likelihood function of the validation set converges.

[0026] The Viterbi algorithm is used to solve for the most likely sequence of hidden states, calculate the posterior probability of each state, map the hidden states to a risk index of 0-100, and output the fire risk assessment index.

[0027] Furthermore, in the aforementioned big data-based fire safety monitoring and management method, the step of classifying fire risks based on the fire risk assessment index using a hierarchical decision-making mechanism and triggering corresponding early warning mechanisms according to the risk level includes:

[0028] Based on the aforementioned fire risk assessment index, a hierarchical decision-making mechanism is used to classify fire risks into four levels: Level 1 risk, Level 2 risk, Level 3 risk, and Level 4 risk.

[0029] Furthermore, in a big data-based fire safety monitoring and management system, the fire safety monitoring and management system includes the following modules:

[0030] The multimodal data acquisition module is used to collect environmental perception data, equipment status data, building and personnel dynamic data, and external and historical data. The collected data is cleaned, denoised, and normalized in sequence to obtain standard multimodal data.

[0031] The feature alignment and fusion module is used to extract the environmental trend features, spatial distribution features and temporal periodic features of the standard multimodal data using an improved multi-scale convolutional neural network, and to align and fuse the features to obtain feature multimodal data.

[0032] The risk index assessment module is used to establish a risk state assessment model based on the CTHMM continuous-time hidden Markov model, define the safety state of the system as the hidden state, use the feature multimodal data as the observation sequence, and output the fire risk assessment index.

[0033] The fire safety classification module is used to classify fire risks based on the fire risk assessment index using a hierarchical decision-making mechanism, and to trigger the corresponding level of early warning mechanism according to the risk level.

[0034] Furthermore, in a big data-driven fire safety monitoring and management system, the risk index assessment module includes the following sub-modules:

[0035] The iterative submodule is used to iteratively optimize the transition probability and observation probability of CTHMM using the EM expectation-maximization algorithm until the log-likelihood function of the validation set converges.

[0036] The output submodule is used to solve for the most likely hidden state sequence using the Viterbi algorithm, calculate the posterior probability of each state, map the hidden states to a risk index of 0-100, and output the fire risk assessment index.

[0037] Furthermore, in a big data-driven fire safety monitoring and management system, the fire safety classification module includes the following sub-modules:

[0038] The classification submodule is used to classify fire risks based on the fire risk assessment index using a hierarchical decision-making mechanism, including Level 1 risk, Level 2 risk, Level 3 risk, and Level 4 risk.

[0039] Its beneficial effects lie in the fact that, through multi-source acquisition and standardized preprocessing, it solves the problems of fragmented and low-quality traditional data, ensuring the accuracy and reliability of the analytical foundation; in the feature extraction stage, it improves the multi-scale CNN combined with an attention mechanism to accurately capture multi-dimensional features of time series, space, and trends, and strengthens data correlation after fusion, breaking through the limitations of single feature analysis and improving the comprehensiveness of risk identification; in the risk assessment stage, the CTHMM model transforms the safety status into a quantitative index, achieving accurate risk classification, effectively reducing false alarms and missed alarms, and avoiding resource waste or hazard omissions; the hierarchical decision-making and early warning mechanism matches response measures according to risk level, significantly shortening response time, while closed-loop management feeds back into model iteration, continuously optimizing monitoring accuracy. Overall, it realizes the transformation of fire management from passive response to proactive early warning, significantly reducing the probability of fire occurrence. Attached Figure Description

[0040] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0041] Figure 1 This is a schematic diagram of the first embodiment of a big data-based fire safety monitoring and management method according to the present invention;

[0042] Figure 2 This is a schematic diagram of a second embodiment of a big data-based fire safety monitoring and management method according to the present invention;

[0043] Figure 3 This is a schematic diagram of the first embodiment of a big data fire safety monitoring and management system according to the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0045] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," and "this" used herein may also include the plural forms. It should be further understood that the terminology used in this specification includes the presence of features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0046] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, a big data-based fire safety monitoring and management method includes the following steps:

[0047] Step 101: Collect environmental perception data, equipment status data, building and personnel dynamic data, and external and historical data. Clean, denoise, and normalize the collected data in sequence to obtain standard multimodal data.

[0048] Specifically, in this embodiment, environmental perception data, equipment status data, building and personnel dynamic data, and external and historical data are collected. Environmental perception data includes temperature, humidity, smoke concentration, harmful gas concentration, and light intensity. Equipment status data includes equipment operating parameters, fault codes, and maintenance records. Building and personnel dynamic data includes the distribution of fire protection facilities, personnel density, and movement trajectories. External and historical data includes fire cases, hazard records, and early warning and response results. Short-term missing data is completed using linear interpolation, abnormal data is identified based on the 3σ principle, real abrupt changes are preserved by wavelet transform denoising, video trajectory data is smoothed by Kalman filtering, jump points caused by occlusion are removed, and high-frequency fluctuation parameters are averaged to obtain preprocessed data. Based on Min-Max normalization, the preprocessed data is mapped to the [0,1] interval to obtain standard multimodal data.

[0049] Comprehensive and targeted data collection is achieved through categorized acquisition: Environmental sensing data is acquired through a sensor network deployed throughout the building, covering key indicators such as temperature (accuracy ±0.5℃), humidity (accuracy ±3%RH), smoke concentration (mg / m³), harmful gas concentration (e.g., CO, ppm), and light intensity. The acquisition frequency is dynamically adjusted based on the risk level of each area. In high-risk areas, such as electrical rooms and flammable material warehouses, data is collected every 1-5 seconds to capture real-time changes; in ordinary areas, such as office areas and corridors, data is collected every 10-30 seconds, balancing monitoring accuracy and resource consumption. Equipment status data is collected in real-time through IoT interfaces of fire-fighting equipment, sprinkler systems, smoke detectors, fire extinguishers, fire pumps, etc., including operating parameters such as sprinkler water pressure and pump current, as well as equipment fault codes and maintenance records. Any changes in equipment status, such as abnormal pressure or fault triggering, are immediately uploaded to ensure real-time tracking of equipment status. This data is relevant to both building and human activities. Personnel dynamic data is divided into two categories: static and dynamic. Building static data is imported through BIM models, covering information such as floor structure, room layout, evacuation route location, fire extinguisher location, and fire hydrant distribution. It is updated synchronously within 24 hours after building renovation to ensure the accuracy of spatial data. Personnel dynamic data is collected through video surveillance, facial recognition technology to identify personnel, heat maps to show density distribution, and access control systems, including personnel density, flow trajectory, and dwell time. The collection frequency is linked with environmental perception data in the same area to achieve synchronous matching of spatiotemporal data. External and historical data serve as supplementary support. External data is connected to meteorological departments to obtain environmental factors affecting fire risk such as rainfall and wind speed, as well as the urban fire pipeline pressure monitoring system, which is updated daily to reflect changes in the external environment. Historical data is imported from the past 5 years, including fire cases, hazard records, and early warning and response results of this building and similar buildings, and is supplemented and updated quarterly to provide historical reference for risk assessment. Regarding the deployment and interface specifications of data acquisition equipment, the sensors adopt the LoRa / NB-IoT low-power protocol, which reduces energy consumption and ensures data transmission stability. They are deployed according to a standard of at least one comprehensive environmental sensor per 50㎡ to ensure no blind spots in the building's monitoring. All sensors are uniformly connected to an edge gateway for data aggregation. The device interfaces strictly comply with the "Technical Standard for Internet of Things Systems for Fire Protection Facilities," supporting Modbus / OPCUA protocols to ensure compatible data transmission for different brands and types of fire protection equipment, enabling real-time data upload to the cloud platform. Data storage employs a dual mechanism of edge caching and distributed storage. Collected data is first cached at edge nodes to avoid data loss due to network interruptions, and then synchronized to distributed databases such as HBase. Original data backups are retained for at least 6 months to provide a basis for subsequent data traceability and model iteration.

[0050] The collected raw data undergoes systematic optimization to eliminate data defects and standardize data format, ultimately generating a standard multimodal dataset that meets the model input requirements. During the data cleaning phase, common data problems are categorized and handled: For missing values, short-term and long-term missing values ​​are distinguished. When the missing time is ≤5 minutes, linear interpolation is used to fill in the missing data based on valid data before and after the missing value, ensuring data continuity. When the missing time is >5 minutes, the data segment is marked as invalid, and an equipment fault warning is triggered, reminding staff to check for sensor or transmission link problems. Outlier removal is based on the 3σ principle, such as sensor false alarms of extreme high temperatures or sudden humidity drops. Further verification is performed in conjunction with equipment status. If the equipment is offline, the outlier data is determined to be caused by transmission or equipment failure and is directly removed. If the equipment is operating normally, the data is retained and marked for verification, with staff confirming whether it is a genuine anomaly based on the on-site situation. Standardizing the format unifies three key dimensions of the data: timestamps are standardized to UTC+8 timezone to avoid data confusion across time zones; data units are standardized, such as converting all temperatures to °C and concentrations to standard units to eliminate potential issues related to dimensional differences; and the data structure adopts JSON format, including core fields such as acquisition time, device ID, region ID, and data value, ensuring data traceability and ease of parsing. In the noise reduction stage, targeted algorithms are used based on the noise characteristics of different data types: environmental perception data contains noise generated by high-frequency sensor jitter, such as minor temperature fluctuations, which is denoised using wavelet transform to filter out high-frequency noise while fully preserving real risk signals such as sudden increases in smoke concentration and temperature; personnel dynamic data is prone to trajectory jumps due to video surveillance obstruction, so the Kalman filter algorithm is used to smooth the trajectory data, correcting anomalies caused by obstruction and restoring the true movement path of personnel; high-frequency fluctuation parameters in equipment status data are processed using the moving average method to extract the steady-state characteristics of equipment operation, avoiding interference from instantaneous fluctuations in the analysis. During the normalization stage, adaptation methods are adopted for different types of data: numerical data uses Min-Max normalization, mapping the data to the [0,1] interval, for example, temperature is transformed according to the rule of 0℃→0, 80℃→1, eliminating the impact of differences in the scale of different indicators on model training; categorical data uses one-hot encoding, converting discrete categories into vector forms that the model can recognize. Finally, the preprocessing stage outputs a standard multimodal dataset containing three major categories of features: time series features, spatial features, and static attributes, providing high-quality, standardized input data for subsequent multimodal feature extraction and risk assessment models.

[0051] Step 102: Use an improved multi-scale convolutional neural network to extract environmental trend features, spatial distribution features, and temporal periodic features from standard multimodal data, and align and fuse the features to obtain feature multimodal data;

[0052] Specifically, in this embodiment, the improved multi-scale convolutional neural network uses three types of convolutional kernels and adds an SE attention mechanism after the convolutional layers to enhance the weights of key features. Temperature and humidity time-series data are processed using 1×3 convolutional kernels to extract rising and falling trends, yielding environmental trend features. High-risk area clustering features are extracted using 3×3 convolutional kernels on building area grid data, yielding spatial distribution features. Daily and weekly fluctuations are extracted using 5×5 convolutional kernels combined with a sliding window, yielding time-series periodic features. Linear interpolation is used to unify the step size of environmental trend features, spatial distribution features, and time-series periodic features at different sampling frequencies, resulting in unified feature data. This unified feature data is then concatenated along the channel dimension, and the cross-attention mechanism is used to learn the correlations between features, outputting multimodal feature data.

[0053] The improved multi-scale CNN model construction process revolves around the core architecture designed for accurate feature extraction and a clearly defined execution flow. In the model structure design section 2.1.1, three types of differentiated convolutional kernels are used to achieve multi-dimensional feature capture: 1×3 convolutional kernels focus on time-series data such as temperature and humidity, effectively extracting short-term upward or downward trends in environmental parameters, such as the risk signal of continuously rising temperatures; 3×3 convolutional kernels are used for building area grid data, such as the smoke concentration distribution matrix of each room, accurately identifying the spatial clustering characteristics of high-risk areas, such as a concentrated increase in smoke concentration on a certain floor; 5×5 convolutional kernels, combined with a 1-hour sliding window, can capture long-term periodic fluctuations, such as the periodic feature of a sudden increase in population density during weekday morning rush hours. To improve feature effectiveness, the model adds a key improvement: embedding an SE attention mechanism after the convolutional layers, which strengthens the weights of key risk features such as smoke concentration and high temperature, while suppressing irrelevant noise features such as weak light changes, further optimizing feature extraction accuracy. In the feature extraction process in 2.1.2, the input data needs to be reorganized into standard multimodal data according to the time-space dimension to form a data matrix of environment, equipment and personnel of each building area at a certain time. Then, through the forward propagation operation of the model, three types of feature maps corresponding to trend, spatial and periodic features are output, and the dimensions of the three types of feature maps are unified as [batch_size,64,64,128], which lays the foundation for subsequent feature processing.

[0054] A two-step process is employed to achieve the collaborative integration of multi-source features, forming a comprehensive risk feature profile. In the feature alignment stage, optimizations are made to address the heterogeneity of multimodal data: for temporal alignment, linear interpolation is used to unify all features to a step size of 10 seconds per second, ensuring data synchronization across time dimensions; for spatial alignment, using the building BIM model coordinates as a reference, features extracted from each region are mapped to a unified spatial grid with a precision of 1m×1m, ensuring accurate spatial correspondence of features from different regions and avoiding feature misalignment due to differences in spatial references. In the feature fusion stage, feature integration is completed in two levels: primary fusion is achieved through channel-level stitching, merging the three types of feature maps—trend, spatial, and periodic—at the channel level to achieve initial feature aggregation; advanced fusion introduces a cross-attention mechanism, learning the relationships between different features to uncover potential risk correlations, ultimately outputting a fused feature vector with dimensions [batch_size, 1024]. This stage ultimately generates multimodal feature data containing spatiotemporal correlation information, providing comprehensive and collaborative feature input for subsequent risk assessment models.

[0055] Step 103: Establish a risk status assessment model based on the CTHMM continuous-time hidden Markov model, define the safety status of the system as the hidden state, use the feature multimodal data as the observation sequence, and output the fire risk assessment index.

[0056] Specifically, this embodiment uses a risk state assessment model based on a CTHMM (Continuous-Time Hidden Markov Model) to classify safe states into four hidden states: safe, low-risk, medium-risk, and high-risk. A Gaussian Mixture Model (GMM) is used to map the feature multimodal data into a probability distribution, defining the transition probability matrix between hidden states and initializing it with historical fire data. The EM (Expectation-Maximization) algorithm is used to iteratively optimize the transition probabilities and observation probabilities of the CTHMM until the log-likelihood function of the validation set converges. The Viterbi algorithm is used to solve for the most probable hidden state sequence, calculating the posterior probability of each state, mapping the hidden states to a risk index of 0-100, and outputting the fire risk assessment index.

[0057] The model parameter definition is the foundation for the CTHMM risk status assessment model to function, focusing on clarifying the core settings of hidden states, observation sequences, and transition probabilities. Regarding the definition of hidden and safe states, the model divides the system's safe state into four clearly defined levels to ensure accurate and controllable risk grading: S0 is a safe state, representing that all monitored indicators, such as temperature, smoke concentration, and equipment operating parameters, are within normal ranges, with no potential risks; S1 is a low-risk state, with only a slight anomaly in a single indicator, such as a local area temperature slightly exceeding the normal threshold, but without other abnormal signals; S2 is a medium-risk state, characterized by multiple indicators being abnormal simultaneously, such as a continuous rise in temperature and an increasing trend in smoke concentration, indicating enhanced risk correlation; S3 is a high-risk state, approaching the conditions for a fire, typically characterized by sustained high temperatures, excessive smoke concentration, and possible abnormal responses from some fire-fighting equipment. In terms of setting the observation sequence and transition probability, the observation sequence directly adopts the feature multimodal data vector output in step 2. Through the Gaussian mixture model (GMM), it is mapped to a probability distribution that meets the model requirements, realizing the transformation of feature data into observation signals. The state transition probability matrix is ​​initialized with historical fire data and follows the actual law of risk evolution. For example, the probability of transitioning from low risk to medium risk in S1 is significantly higher than that in S2, and the probability of transitioning from safe to high risk in S3 is higher, ensuring that the state switching conforms to the risk progression logic in the real scenario.

[0058] Through scientific data partitioning and algorithm optimization, the accuracy and stability of model evaluation are ensured. In the dataset construction phase, the model uses a historical dataset containing both normal building operation data and past fire case data, balancing risk-free and risky scenarios to ensure the comprehensiveness of training samples. The dataset is divided into a 7:3 ratio: training set (70%) and validation set (30%). Each data set is labeled with a corresponding hidden state label; for example, based on fire development patterns, monitoring data from 10 minutes before a fire is uniformly labeled as S3 (high risk), providing a clear supervised learning objective for the model. In the parameter optimization phase, the model uses the Expectation-Maximization (EM) algorithm for iterative optimization, focusing on adjusting two key parameters of the CTHMM: state transition probability (optimizing the probability of switching between different hidden states) and observation probability (adjusting the probability distribution of observation sequences corresponding to each hidden state in the multimodal data). During iteration, the log-likelihood loss of the validation set is used as the convergence criterion. Iteration stops when the loss value no longer decreases significantly and tends to stabilize, indicating that the model parameters have reached their optimal state and can accurately capture the correlation between data and risk states.

[0059] The model transforms abstract state judgments into intuitive and timely quantitative risk indicators. During the real-time inference phase, the model continuously receives the feature multimodal data output from step 2 and uses the Viterbi algorithm to solve for the most probable hidden state sequence corresponding to the current data. This algorithm can filter out the sequence with the highest probability from numerous potential state sequences and simultaneously calculate the posterior probability of each hidden state, further clarifying the confidence level of the current safety state and avoiding bias from single-state judgments. In calculating the risk assessment index, the model transforms the abstract hidden states into a quantitative risk index of 0-100, establishing a clear mapping relationship: S0, safe (0-20); S1, low risk (21-40); S2, medium risk (41-70); S3, high risk (71-100), allowing managers to intuitively perceive the level of risk. To ensure real-time response to risk changes, this risk assessment index is updated every 10 seconds, dynamically tracking subtle changes in risk within the building and promptly capturing risk escalation signals. This provides accurate and timely quantitative basis for subsequent risk-level-based tiered decision-making and early warning mechanisms.

[0060] Step 104: Based on the fire risk assessment index, fire risks are classified using a hierarchical decision-making mechanism, and corresponding early warning mechanisms are triggered according to the risk level.

[0061] Specifically, in this embodiment, fire risks are classified into four levels based on the fire risk assessment index and a hierarchical decision-making mechanism, including Level 1 risk, Level 2 risk, Level 3 risk, and Level 4 risk.

[0062] Using a risk assessment index as the core, the system first clarifies the four-level risk classification standards, then designs targeted decision-making logic and execution plans, and finally achieves continuous optimization through closed-loop management, forming a full-process risk response system. The risk classification standards are based on a risk index of 0-100: Level 1 risk corresponds to an index of 0-20, representing no significant risk, and the system only needs to maintain normal monitoring; Level 2 risk corresponds to 21-40, representing minor anomalies, mostly caused by occasional events such as equipment false alarms, requiring localized investigation; Level 3 risk corresponds to 41-70, representing potential hazards, indicating risks requiring on-site intervention, such as abnormal local indicators combined with equipment status fluctuations; Level 4 risk corresponds to 71-100, representing high fire risk, approaching the conditions for a fire, requiring emergency response. Based on this classification standard, a targeted hierarchical decision-making mechanism was constructed: For Level 1 risks, the system automatically records monitoring data without triggering additional warnings, generating only a routine monitoring report every hour to balance management efficiency and resource consumption; for Level 2 risks, the decision-making logic first compares historical data to determine whether it is an occasional anomaly such as temporary fluctuations in sensor data, and then pushes a notification to the property management personnel's APP, requiring on-site verification and feedback within one hour to promptly eliminate false alarms or handle minor issues; for Level 3 risks, the decision-making process combines spatial characteristics to accurately locate high-risk areas, such as the office on the east side of the 3rd floor, and associates the operating status of electrical equipment in that area. During execution, an audible and visual alarm is triggered in the area, and a notification is pushed... Work orders are sent to fire inspectors, requiring them to arrive on-site within 30 minutes and simultaneously shut off non-essential power to the area to reduce risk. For Level 4 risks, the decision-making logic focuses on activating the emergency plan, calculating the optimal evacuation route based on real-time personnel distribution and evacuation route status. The execution process involves both internal and external coordination: internally, the building's audible and visual alarms are activated, evacuation instructions are broadcast repeatedly via emergency broadcasts, elevators are automatically forced to the ground floor, and access control systems are automatically unlocked to ensure smooth evacuation; externally, the system automatically dials 119 (the emergency number in China), simultaneously uploading key information such as the building's precise location, current risk index, and real-time personnel density, and also promptly notifies the property manager and the local fire command center to ensure coordinated multi-party response. Furthermore, to continuously optimize the mechanism, a closed-loop management process for early warning has been established: the entire process of recording the early warning trigger time, specific handling procedures, and final results (e.g., Level 3 risk → troubleshooting within 20 minutes). This data is fed back to the historical database for monthly updates to the transition probability of the CTHMM model, continuously improving the accuracy of risk assessment and decision-making, forming a complete closed loop of monitoring-early warning-response-optimization.

[0063] Its beneficial effects lie in the fact that, through multi-source acquisition and standardized preprocessing, it solves the problems of fragmented and low-quality traditional data, ensuring the accuracy and reliability of the analytical foundation; in the feature extraction stage, it improves the multi-scale CNN combined with an attention mechanism to accurately capture multi-dimensional features of time series, space, and trends, and strengthens data correlation after fusion, breaking through the limitations of single feature analysis and improving the comprehensiveness of risk identification; in the risk assessment stage, the CTHMM model transforms the safety status into a quantitative index, achieving accurate risk classification, effectively reducing false alarms and missed alarms, and avoiding resource waste or hazard omissions; the hierarchical decision-making and early warning mechanism matches response measures according to risk level, significantly shortening response time, while closed-loop management feeds back into model iteration, continuously optimizing monitoring accuracy. Overall, it realizes the transformation of fire management from passive response to proactive early warning, significantly reducing the probability of fire occurrence.

[0064] Please see Figure 2 In a big data-driven fire safety monitoring and management method, an improved multi-scale convolutional neural network is used to extract environmental trend features, spatial distribution features, and temporal periodic features from standard multimodal data. These features are then aligned and fused to obtain feature-rich multimodal data, including the following steps:

[0065] Step 201: The improved multi-scale convolutional neural network sets up three types of convolutional kernels and adds an SE attention mechanism after the convolutional layer to enhance the weights of key features;

[0066] Step 202: Process the temperature and humidity time series data with a 1×3 convolution kernel to extract the rising and falling trends and obtain environmental trend features;

[0067] Step 203: Use a 3×3 convolution kernel to convolve the grid data of the building area to extract the clustering features of high-risk areas and obtain the spatial distribution features;

[0068] Step 204: Extract daily and weekly regular fluctuations based on 5×5 convolution kernels combined with sliding windows to obtain time-series periodic features.

[0069] The above describes an embodiment of the fire safety monitoring and management method based on big data according to the present invention. Please refer to [link / reference]. Figure 3 In a big data-driven fire safety monitoring and management system, the fire safety monitoring and management system includes the following modules:

[0070] The multimodal data acquisition module is used to collect environmental perception data, equipment status data, building and personnel dynamic data, and external and historical data. The collected data is cleaned, denoised, and normalized in sequence to obtain standard multimodal data.

[0071] The feature alignment and fusion module is used to extract environmental trend features, spatial distribution features and temporal periodic features from standard multimodal data using an improved multiscale convolutional neural network, and to align and fuse the features to obtain feature multimodal data.

[0072] The risk index assessment module is used to establish a risk status assessment model based on the CTHMM continuous-time hidden Markov model. It defines the safety status of the system as the hidden state, uses feature multimodal data as the observation sequence, and outputs the fire risk assessment index.

[0073] The fire safety classification module is used to classify fire risks based on the fire risk assessment index and a hierarchical decision-making mechanism, and to trigger the corresponding level of early warning mechanism according to the risk level.

[0074] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A big data-based fire safety monitoring and management method, characterized in that, The fire safety monitoring and management method includes the following steps: Collect environmental perception data, equipment status data, building and personnel dynamic data, and external and historical data. The collected data are then cleaned, denoised, and normalized to obtain standard multimodal data. An improved multi-scale convolutional neural network is used to extract environmental trend features, spatial distribution features, and temporal periodic features from the standard multimodal data, and the features are aligned and fused to obtain feature multimodal data. A risk status assessment model based on the CTHMM continuous-time hidden Markov model is established, the safety status of the system is defined as the hidden state, the feature multimodal data is used as the observation sequence, and the fire risk assessment index is output. Based on the aforementioned fire risk assessment index, a hierarchical decision-making mechanism is used to classify fire risks, and a corresponding level of early warning mechanism is triggered according to the risk level. The risk state assessment model based on the CTHMM continuous-time hidden Markov model is established, defining the system's safety state as a hidden state, using the characteristic multimodal data as an observation sequence, and outputting a fire risk assessment index, including: The risk state assessment model based on the CTHMM continuous-time hidden Markov model divides the safe state into four hidden states: safe, low risk, medium risk and high risk. The feature multimodal data is mapped to a probability distribution using a Gaussian mixture model (GMM), a transition probability matrix between hidden states is defined, and it is initialized using historical fire data. The risk state assessment model based on the CTHMM continuous-time hidden Markov model is established, defining the system's safety state as a hidden state, using the characteristic multimodal data as an observation sequence, and outputting a fire risk assessment index, including: The transition probability and observation probability of CTHMM are iteratively optimized using the EM expectation-maximization algorithm until the log-likelihood function of the validation set converges. The Viterbi algorithm is used to solve for the most likely sequence of hidden states, calculate the posterior probability of each state, map the hidden states to a risk index of 0-100, and output the fire risk assessment index.

2. The fire safety monitoring and management method based on big data as described in claim 1, characterized in that, The collected environmental perception data, equipment status data, building and personnel dynamic data, and external and historical data are sequentially cleaned, denoised, and normalized to obtain standard multimodal data, including: Collect environmental perception data, equipment status data, building and personnel dynamic data, and external and historical data. Environmental perception data includes temperature, humidity, smoke concentration, harmful gas concentration, and light intensity. Equipment status data includes equipment operating parameters, fault codes, and maintenance records. Building and personnel dynamic data includes the distribution of fire protection facilities, personnel density, and movement trajectories. External and historical data includes fire cases, hazard records, and early warning and response results. Linear interpolation is used to complete short-term missing data, abnormal data is identified based on the 3σ principle, wavelet transform is used to denoise and retain real abrupt change signals, Kalman filtering is used to smooth video trajectory data, occlusion-induced jump points are removed, and high-frequency fluctuation parameters are averaged to obtain preprocessed data. Preprocessed data is mapped to the [0,1] interval based on Min-Max normalization to obtain standard multimodal data.

3. The fire safety monitoring and management method based on big data as described in claim 1, characterized in that, The improved multi-scale convolutional neural network is used to extract environmental trend features, spatial distribution features, and temporal periodic features from the standard multimodal data, and the features are aligned and fused to obtain feature multimodal data, including: The improved multi-scale convolutional neural network sets up three types of convolutional kernels and adds an SE attention mechanism after the convolutional layer to enhance the weights of key features; Temperature and humidity time series data are processed by 1×3 convolution kernels to extract rising and falling trends and obtain environmental trend features; High-risk area clustering features are extracted by convolving the building area grid data with 3×3 convolution kernels to obtain spatial distribution features; Based on the 5×5 convolution kernel combined with the sliding window, the daily and weekly regular fluctuations are extracted to obtain the time-series periodic features.

4. The fire safety monitoring and management method based on big data as described in claim 1, characterized in that, The improved multi-scale convolutional neural network is used to extract environmental trend features, spatial distribution features, and temporal periodic features from the standard multimodal data, and the features are aligned and fused to obtain feature multimodal data, including: By unifying the step size of environmental trend characteristics, spatial distribution characteristics and temporal periodic characteristics at different sampling frequencies through linear interpolation, unified feature data is obtained. The unified feature data is concatenated along the channel dimension, and the correlation between features is learned through the Cross-Attention mechanism to output multimodal feature data.

5. The fire safety monitoring and management method based on big data as described in claim 1, characterized in that, The method of classifying fire risks based on the fire risk assessment index using a hierarchical decision-making mechanism and triggering corresponding early warning mechanisms according to the risk level includes: Based on the aforementioned fire risk assessment index, a hierarchical decision-making mechanism is used to classify fire risks into four levels: Level 1 risk, Level 2 risk, Level 3 risk, and Level 4 risk.

6. A big data-based fire safety monitoring and management system, used to execute the big data-based fire safety monitoring and management method as described in claim 1, characterized in that, The fire safety monitoring and management system includes the following modules: The multimodal data acquisition module is used to collect environmental perception data, equipment status data, building and personnel dynamic data, and external and historical data. The collected data is cleaned, denoised, and normalized in sequence to obtain standard multimodal data. The feature alignment and fusion module is used to extract the environmental trend features, spatial distribution features and temporal periodic features of the standard multimodal data using an improved multi-scale convolutional neural network, and to align and fuse the features to obtain feature multimodal data. The risk index assessment module is used to establish a risk state assessment model based on the CTHMM continuous-time hidden Markov model, define the safety state of the system as the hidden state, use the feature multimodal data as the observation sequence, and output the fire risk assessment index. The fire safety classification module is used to classify fire risks based on the fire risk assessment index using a hierarchical decision-making mechanism, and to trigger the corresponding level of early warning mechanism according to the risk level. The risk index assessment module includes the following sub-modules: The iterative submodule is used to iteratively optimize the transition probability and observation probability of CTHMM using the EM expectation-maximization algorithm until the log-likelihood function of the validation set converges. The output submodule is used to solve for the most likely hidden state sequence using the Viterbi algorithm, calculate the posterior probability of each state, map the hidden states to a risk index of 0-100, and output the fire risk assessment index.

7. The big data-based fire safety monitoring and management system as described in claim 6, characterized in that, The fire safety classification module includes the following sub-modules: The classification submodule is used to classify fire risks based on the fire risk assessment index using a hierarchical decision-making mechanism, including Level 1 risk, Level 2 risk, Level 3 risk, and Level 4 risk.

Images