A building fire intelligent detection and positioning method based on multi-source data fusion

Abstract

The application discloses a building fire intelligent detection and positioning method based on multi-source data fusion, which comprises the following steps: collecting two types of differentiated data of active and passive through a double perception layer architecture, and constructing a full-coverage information complementary perception network; establishing a dynamic game mechanism based on the two types of differentiated data, judging the data source reliability in real time, and generating a high-reliability fusion feature set; performing parallel research and judgment based on the fusion feature set through multi-head deep learning research and judgment, and outputting a fire information research and judgment result; and generating a labeled training sample by using the fire information research and judgment result to optimize the dynamic game and multi-head deep learning research and judgment parameters, and correcting historical misjudgment. The application aims to solve the technical problems of high false alarm rate, poor positioning accuracy, inability to identify fire source types, lack of self-adaptive ability, high deployment cost in complex environments, existence of monitoring blind area and the like of the existing fire detection system.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensing technology, and in particular to an intelligent detection and location method for building fires based on multi-source data fusion. Background Technology

[0002] Traditional building fire detection systems primarily rely on single point sensors such as smoke and heat detectors, triggering alarms based on preset thresholds. These systems have significant limitations in complex building environments: high false alarm rates, inability to distinguish between actual fires and interfering sources like cooking fumes and steam; lack of location capabilities, only providing area-based alarms and failing to guide precise rescue efforts; slow response to early-stage fires; and inability to identify the type of ignition source. Recent multi-sensor fusion technologies often employ fixed-weight data fusion strategies, lacking adaptability to sensor performance degradation and dynamic environmental changes, making long-term reliability difficult to guarantee. Furthermore, existing systems generally lack self-learning capabilities, unable to learn from historical data to evolve their performance. In addition, traditional detection networks have coverage blind spots, and deployment and maintenance costs are high in large, complex spaces. These shortcomings severely restrict the improvement of fire prevention and control effectiveness, necessitating an innovative solution capable of accurate perception, intelligent decision-making, and continuous evolution. Summary of the Invention

[0003] This invention provides a method for intelligent detection and location of building fires based on multi-source data fusion, comprising:

[0004] Passive and active sensing data are collected, and after clock synchronization, instantaneous fluctuation elimination, and noise elimination, they are encapsulated in a standardized format to construct a full-coverage complementary sensing network.

[0005] The system identifies the precise spatial coordinates of two types of differentiated data, establishes a causal relationship graph based on physical laws according to the sequential response relationship of different sensors, inserts virtual probe nodes in the virtual perception enhancement stage, constructs a spatiotemporal graph model, and calculates the fire confidence and core contribution by constructing an intelligent fusion mechanism based on hierarchical dynamic game, generating a fusion feature set.

[0006] Based on the fusion characteristics, multi-head deep learning is used to determine and locate the fire area and its corresponding fire protection zone, and to accurately locate the fire source, fire intensity, spread trend, and the status and location of monitoring personnel, and output the fire information analysis results.

[0007] Arbitration signal source data is collected, and a self-evolving knowledge base is constructed based on the confidence level of the signal source and the fire information analysis results, thereby optimizing the parameters of the dynamic game and multi-head deep learning model.

[0008] The above-described intelligent building fire detection and location method based on multi-source data fusion includes the following: collecting two types of differentiated data—active and passive—through a dual-sensor layer architecture to construct a fully covered, complementary information sensing network, comprising:

[0009] A passive sensing layer is constructed using a radio frequency gateway and a high-precision magnetometer to collect passive data.

[0010] An active sensing layer is constructed using a miniature electronic nose array, an acoustic-optical sensing unit, and radar to collect active data.

[0011] The above-described intelligent building fire detection and location method based on multi-source data fusion includes a dynamic game mechanism established based on two types of differentiated data to determine the credibility of the data source in real time and generate a high-credibility fusion feature set, including:

[0012] By associating two separate types of differentiated data into a unified spatiotemporal dataset, a spatiotemporal graph model is constructed.

[0013] Using independent information sources as participants, the fire confidence level of each participating group is calculated, and the core contribution of each participant to the participating group is calculated, outputting a high-confidence fusion feature set.

[0014] The above-described intelligent building fire detection and location method based on multi-source data fusion includes parallel analysis based on the fused feature set using multi-head deep learning, outputting fire information analysis results, including:

[0015] Based on the fusion characteristics, the fire area and its corresponding fire protection zone can be quickly located using a decision-making and positioning model.

[0016] After locating the area, deep feature extraction and analysis are performed through multi-head deep learning to output fire information analysis results.

[0017] The above-described intelligent building fire detection and location method based on multi-source data fusion includes, after locating the area, performing deep feature extraction and analysis through multi-head deep learning to output fire information analysis results, including:

[0018] Using the fusion features as input, the probability that the fire source is located at the center point of the grid is calculated through a fusion regression model based on spatial grid classification and signal source tracing.

[0019] Using fused features as input, the fire intensity and spread trend can be detected by combining gas fingerprints and visual-thermodynamic features.

[0020] Using fused features as input, a continuous spatial analysis method based on field theory and topology is used to detect personnel status and location heatmaps.

[0021] The above-described intelligent building fire detection and location method based on multi-source data fusion includes generating labeled training samples using fire information analysis results to optimize dynamic game theory and multi-head deep learning analysis parameters, and correcting historical misjudgments, including:

[0022] Collect arbitration signal source data and calculate the basic confidence level of each signal source. Obtain the overall confidence score through pre-set confidence fusion rules.

[0023] A self-evolving knowledge base is constructed based on the arbitration signal sources above the confidence threshold and the fire information analysis results, and the parameters of the dynamic game and multi-head deep learning model are optimized.

[0024] The beneficial effects achieved by this invention are as follows:

[0025] A dynamic game-theoretic fusion mechanism intelligently assesses the real-time reliability of data from each sensor, making the system highly robust to sensor performance degradation and environmental interference, thus reducing false alarm rates. A multi-head deep learning network, through parallel analysis, outputs multi-dimensional situational information in a single operation, including precise coordinates of the fire source, material type, fire intensity, and personnel location, providing unprecedented decision support accuracy for firefighting and rescue operations. An automatic arbitration closed-loop self-evolution mechanism enables lifelong learning, continuously optimizing model parameters to adapt to new fire conditions and environmental changes, maintaining optimal performance over the long term. Through a dual-sensor layer architecture and a dynamic game-theoretic fusion mechanism, the system achieves a technological leap from passive alarm to proactive early warning, from regional monitoring to precise positioning, and from fixed thresholds to adaptive decision-making. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0027] Figure 1 This is a flowchart of a building fire intelligent detection and positioning method based on multi-source data fusion provided in Embodiment 1 of this application. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Example 1

[0030] like Figure 1 As shown, Embodiment 1 of this application provides a method for intelligent detection and location of building fires based on multi-source data fusion, including:

[0031] S1: Collect passive and active sensing data, and after clock synchronization, instantaneous fluctuation elimination, and noise elimination, encapsulate the data in a standardized format to construct a full-coverage complementary sensing network;

[0032] The process involves collecting passive and active sensing data, synchronizing them with a clock, eliminating instantaneous fluctuations and noise, and then encapsulating them in a standardized format to construct a fully covered complementary sensing network. This includes the following sub-steps:

[0033] S11: A passive sensing layer is constructed through an RF gateway and a high-precision magnetometer to collect passive data;

[0034] Through a LoRa IoT communication gateway network deployed within the building, all gateways are controlled to simultaneously perform signal sensing tasks in a distributed architecture while maintaining normal data communication. Continuous handshake communication between gateways captures real-time received radio frequency field disturbance data across all communication links, and timestamps each data packet with nanosecond-level precision. Simultaneously, a miniature high-precision MEMS magnetometer array deployed at key nodes of the building's load-bearing structure is activated to continuously collect geomagnetic field anomaly data at a preset sampling rate.

[0035] The radio frequency field disturbance data mainly includes abnormal fluctuations in the Received Signal Strength Index (RSSI) and characteristic offsets in the carrier phase, reflecting the refraction, scattering, and absorption effects of flames and high-temperature smoke on the wireless signal propagation path. The geomagnetic field anomaly data mainly includes persistent distortions in the geomagnetic field vector (intensity, tilt, and deflection), reflecting the magnetic abrupt changes that occur after the temperature of ferromagnetic materials such as steel bars inside the building exceeds the Curie point due to a fire.

[0036] For the LoRa radio frequency field, the normal fluctuation range of RSSI and phase for each link is obtained by estimating the Gaussian kernel density of historical data for 30 consecutive days during fire-free periods. For the geomagnetic field, a baseline vector distribution map is established by collecting static magnetic field data of buildings under thermally stable conditions. Then, passive sensing data containing key information such as radio frequency field disturbance data, geomagnetic field anomaly data, and the baseline vector distribution map of the normal fluctuation area are output.

[0037] S12: An active sensing layer is constructed using a miniature electronic nose array, an acoustic-optical sensing unit, and radar to collect active data.

[0038] Simultaneously, through a multimodal data collaborative acquisition process using the active sensing layer, the distributed micro-electronic nose array is controlled to enter a high-frequency sampling mode. Multiple integrated cross-sensitive MEMS gas sensors continuously monitor the concentrations of characteristic gases such as CO, CO2, and HCN, and the built-in processor calculates key gas ratio sequences in real time, forming a dynamically changing gas fingerprint spectrum. Simultaneously, the infrared thermal imaging camera and microphone array in the acoustic-optical sensing unit are activated to capture temperature field distribution data, collect ambient audio, and generate a Mel-spectrum map in real time. A millimeter-wave radar module transmits a frequency-modulated continuous wave and receives the echo signal. After processing with a fast Fourier transform, point cloud data containing distance, velocity, and angle is generated, and micro-motion signals in the 0.1-0.5Hz frequency band are specifically extracted for respiration detection.

[0039] All sensing units are equipped with a precision clock synchronization mechanism to assign a uniform, high-precision timestamp to each data sample. A sliding window mean filter is used to eliminate instantaneous fluctuations in gas sensor readings; non-uniformity correction is used to eliminate fixed pattern noise in thermal imaging data; finally, the processed multimodal data is encapsulated in a standardized format to generate an active sensing dataset containing gas concentration vectors, temperature field matrices, audio feature vectors, and radar point cloud clusters.

[0040] S2: Connect the precise spatial coordinates of two types of differentiated data, establish a causal relationship graph based on physical laws according to the sequential response relationship of different sensors, insert virtual probe nodes in the virtual perception enhancement stage, construct a spatiotemporal graph model, and calculate the fire confidence and core contribution by constructing an intelligent fusion mechanism based on hierarchical dynamic game to generate a fusion feature set;

[0041] The process involves: identifying the precise spatial coordinates of two types of differentiated data; establishing a causal relationship graph based on physical laws according to the sequential response relationships of different sensors; inserting virtual probe nodes during the virtual perception enhancement stage to construct a spatiotemporal graph model; and calculating the fire confidence and core contribution by constructing an intelligent fusion mechanism based on hierarchical dynamic game theory to generate a fusion feature set. This includes the following sub-steps:

[0042] S21: Link the two types of dispersed, differentiated data into a unified spatiotemporal dataset and construct a spatiotemporal graph model;

[0043] Based on two types of differentiated data (passive sensing data and active sensing data), a precise clock synchronization mechanism is used to align all data streams to a unified time axis and assign precise spatial coordinates to each data point.

[0044] By analyzing the sequential response relationships of different sensor data over time, a causal relationship graph based on physical laws is established. Specifically, a bidirectional inference network is constructed: in the backward inference path, a three-layer 128-dimensional temporal LSTM network is used to backtrack and calculate the most probable signal propagation path, starting from the currently detected anomalous signal and combining architectural space topology and fluid dynamics constraints; in the forward verification path, based on the estimated signal source location, the sensor nodes to which the signal should reach and its expected response intensity are predicted using a physical field propagation model.

[0045] The virtual sensing enhancement stage then begins. Virtual probe nodes are automatically inserted into the frontal region of the physical signal propagation. These nodes deduce the expected physical field state of the region based on physical laws. Through multiple rounds of message passing in a 256-dimensional feature space using a graph convolutional network, the state estimates of these virtual nodes are continuously optimized, forming a predictive coverage of unmonitored areas.

[0046] The output is an enhanced spatiotemporal graph containing both actual observation data and inferred data. The node features include not only sensor readings but also predicted states based on physical laws; the edge weights reflect the dynamic correlation strength between sensors based on causal relationships.

[0047] S22: Using independent information sources as participants, calculate the fire confidence level of each participating group and the core contribution of each participant to the participating group, and output a high-confidence fusion feature set.

[0048] The core of establishing an intelligent fusion mechanism based on hierarchical dynamic game theory lies in transforming traditional static data fusion into a dynamic game process with cognitive reasoning capabilities. Each independent information source (spatiotemporal graph node) in the spatiotemporal graph model is defined as a game participant, including radio frequency disturbance nodes, geomagnetic monitoring points, gas sensors, and acoustic-optical sensing units, etc., and a multi-dimensional feature description is established for each participant, including real-time data quality, historical reliability, and physical characteristics.

[0049] A hierarchical game theory architecture is used to construct a dynamic game model. Initial gameplay is conducted within local spatial clusters. Each cluster consists of participants with strong spatial correlation. The core contribution of each participant within the cluster is calculated using a local game theory algorithm based on improved Shapely values. The calculation formula is as follows:

[0050] Indicates participants The ultimate core contribution of (i.e., a data source); This represents the environmental adaptation coefficient, a global scaling factor used to adjust the baseline level of all participants' contributions based on the overall complexity and disturbance level of the current environment; for example, it may be reduced when environmental disturbances are severe to lower the initial confidence assumptions of all features. This represents the total number of participants within the local spatial cluster where the game is currently taking place; Represents the set of participants The participants currently being evaluated were excluded. Any subsequent subset; Representing conditional mutual information, quantified in the known participating group All data Under these conditions, participants Data Able to make the final decision (e.g., whether it's a fire) How much additional, unique information does it provide; the higher the value, the more irreplaceable the information provided. Let be the physical prior probability, representing the participant's... Observational data Given physical model parameters The probability of this occurring under conditions such as thermal diffusion equations and gas propagation models; if the data severely violates physical laws, this value will be very small, thus reducing its contribution.

[0051] It is a weighting parameter that controls The magnitude of the divergence influences the behavior; the larger the divergence, the more severe the penalty for behaviors that violate the predictions of the physics model. yes Divergence measures the difference between two probability distributions. and The differences between them; Based on participants The physical field state deduced from the data; Based on Participation Group The physical field state is deduced from the data. This is used to penalize data that causes the deduction results to deviate from the mainstream (participating group) Participants whose projection results are significantly inconsistent with the actual results. It is the temporal correlation gain coefficient, used to measure the temporal correlation function. The weights; It is a time-series correlation function used to evaluate participants. Data and Participation Group The causality of data over time series. For example, if The temperature rise always occurs slightly earlier than If the temperature rise of other sensors is greater, then this function value will be larger, indicating that... It may be located upstream of where the fire is spreading; This represents the topological influence factor, used to adjust the weights of the topological correlation function; The topology correlation function is based on the spatial layout and graph structure of the sensor network to calculate the participants. With Alliance The overall connectivity strength of all members; the closer the sensors are in space or the more closely they are associated along the signal propagation path, the larger this value is.

[0052] This algorithm introduces a multiphysics coupling consistency constraint, where a participant's core contribution receives an additional gain when their observation data is highly consistent with high-confidence arbitration evidence. Specifically, a two-layer neural network with an attention mechanism is constructed. The first layer of 256 neurons is responsible for extracting the feature representations of each participant, while the second layer of 128 neurons calculates the initial weight distribution using a softmax function. During training, a gradient descent algorithm with multi-task loss is used, with a learning rate of 0.001 and a batch size of 64.

[0053] During the global fusion phase, the system treats each local cluster as a supernode and engages in cross-regional collaborative game. The joint confidence of different participant combinations is calculated in real time, and a Monte Carlo tree search algorithm is used to optimize the Shapely value calculation process, reducing computational complexity. Specifically, a time series analysis mechanism is introduced to evaluate the participants' contributions using a sliding window, ensuring the temporal stability of the game outcome.

[0054] The final output is a high-confidence fusion feature set optimized by dynamic game theory. This feature set not only contains the weighted fusion of the perceived data, but also retains the contribution distribution and confidence assessment of each participant.

[0055] S3: Based on the fusion features, multi-head deep learning is used to determine and locate the fire area and its fire protection zone, and to accurately locate the fire source, fire intensity, spread trend, and monitoring personnel status and location, and output the fire information analysis results.

[0056] The process involves using multi-head deep learning to determine the location of the fire area and its corresponding fire protection zone based on fusion characteristics. This allows for precise location of the fire source, fire intensity, spread trend, and the status and location of monitoring personnel. The resulting fire information analysis results are then output, including the following sub-steps:

[0057] S31: Based on the fusion characteristics, the fire area and its corresponding fire compartment are quickly located through the decision-making and localization model;

[0058] A high-confidence fused feature set is input into a generative-discriminative dual-path inference network. The generative path constructs a physical prior based on parameterized fire dynamics equations. By solving the simplified Navier-Stokes equations, which include buoyancy and turbulence terms, it simulates the development of a thermal plume on a 0.1-meter resolution spatial grid using the finite volume method. Key parameters in the equations, such as the heat release rate, are adaptively adjusted through a learnable module, outputting a hypothetical heatmap driven by physical laws. The discriminative path employs a deep convolutional network with residual connections, containing 5 convolutional layers and 3 fully connected layers, using the ReLU activation function. By analyzing the spatial distribution characteristics of multi-source sensor data, it outputs a data-driven detection heatmap.

[0059] The outputs of the two pathways are fused using an adaptive attention gating mechanism based on sensor data quality. The calculation of the gating weights comprehensively considers the sensor's signal-to-noise ratio, historical reliability, and spatiotemporal consistency index. After obtaining the fused heatmap, a multi-round hypothesis testing loop is initiated, with each loop containing three meticulously designed stages: In the hypothesis generation stage, multiple high-probability regions are extracted from the fused heatmap as candidate fire source hypotheses; in the evidence prediction stage, for each candidate hypothesis, the expected data patterns that each sensor should observe under the condition that the hypothesis is true are deduced from the sensor responses, including temperature change curves, gas diffusion patterns, and signal perturbation characteristics; in the hypothesis testing stage, the posterior probability of each hypothesis is calculated, and the degree of matching between the predicted pattern and the actual observation is characterized by a likelihood function.

[0060] The entire hypothesis testing process is implemented through a gated recurrent unit with 128 hidden states. This network is trained using the time backpropagation algorithm with a cross-entropy loss function, a learning rate of 0.001, a batch size of 32, and 100 training epochs. The network dynamically adjusts the weights of each hypothesis by learning the temporal dependencies of the arbitration evidence. After obtaining a stable hypothesis distribution, dynamic perception guidance based on information entropy is implemented. By calculating the information gain of each monitoring area, the key regions that best distinguish competing hypotheses are identified, and the sampling frequency and data fusion weights of the sensors within these regions are proactively adjusted. This proactive perception strategy enables the system to quickly focus on the most discriminative evidence. The final output includes localization results containing multiple competing hypotheses and their confidence levels, with each hypothesis accompanied by a detailed spatial range probability distribution and evidence support analysis.

[0061] S32: After locating the area, deep feature extraction and analysis are performed through multi-head deep learning to output fire information analysis results.

[0062] The process of locating the area involves multi-head deep learning analysis to extract and evaluate deep features, and then outputting fire information analysis results, including the following sub-steps:

[0063] S321: Using the fusion features as input, the probability that the fire source is located at the center point of the grid is calculated through a fusion regression model based on spatial grid classification and signal source tracing.

[0064] Based on a high-reliability fused feature set, precise fire source localization is performed within the initial area. First, the target area is divided into a 3D grid with 0.1-meter precision, with each grid point corresponding to a potential fire source location. For each grid point, its spatial relationship with each sensor is calculated, including distance, obstacle distribution along the propagation path, and signal attenuation characteristics. A forward propagation algorithm is then established from the grid point to the observations of each sensor.

[0065] Based on the aforementioned forward propagation algorithm, a multi-dimensional constrained optimization problem is constructed, with the objective function being to minimize the combined difference between theoretical sensor readings and actual observations. Specifically, the loss function comprises three terms: a data fitting term measures the mean square error between theoretical and actual observations; a spatial smoothing term ensures the continuity of probabilities between adjacent grid points through Laplace regularization; and a physical constraint term utilizes heat conduction and fluid dynamics equations to guarantee that the solution conforms to physical laws. The optimization process employs a two-stage strategy: first, a global search method is used to determine potential regions on a coarse-grained grid, and then a quasi-Newton method is used for fine-tuning on a fine-grained grid.

[0066] During the optimization process, an adaptive weighting mechanism is introduced to dynamically adjust the weight of each sensor data in the loss function based on its reliability and timeliness. Sensor data with high reliability is assigned a larger weight, while contradictory sensor data is weighted less. Simultaneously, the uncertainty of the localization result is assessed by analyzing the Hessian matrix of the loss function. When the principal eigenvalue exceeds a threshold, a multi-initial-value optimization strategy is automatically initiated to avoid getting trapped in local optima.

[0067] The final output is the probability of fire source presence at each grid point, and the most likely location and confidence interval of the fire source are calculated based on the probability distribution. This process also generates an analysis of the contribution of each sensor's data to the location results, providing interpretability.

[0068] S322: Using fused features as input, it detects fire intensity and predicts spread trends through gas fingerprinting and visual-thermodynamic features;

[0069] After identifying the fire source, a correlation matrix of fire elements is established, using parameters such as gas concentration changes, temperature gradients, and optical characteristics as nodes. A causal network is constructed by analyzing the time-delay correlations between parameters. A sliding time window is used to calculate the transfer entropy between each parameter, determining the causal direction and intensity, and forming a dynamically weighted directed graph structure.

[0070] Based on the constructed causal network, a fire intensity assessment algorithm based on information flow accumulation is employed. Specifically, the energy transfer path during fire development is quantified by calculating the information flux between nodes. Focusing on the in-degree and out-degree weights of key parameter nodes, a quantitative index of fire intensity is obtained by calculating the closed integral of the information flow in the network. Particular attention is paid to the bidirectional coupling relationship between gas concentration nodes and temperature nodes, and the abrupt changes in fire intensity are captured by analyzing the time derivative of their mutual information.

[0071] Subsequently, in predicting the spread trend, a topological potential field-based prediction method was used. The built-up space was discretized into a potential field grid, and the topological potential energy of each grid point was calculated based on the correlation strength of parameters determined in the causal network. By solving the evolution equation of the potential energy gradient, the propagation path of the fire in the built-up space was simulated. The evolution equation formula is as follows:

[0072] Fire situation function Regarding time The partial derivatives of . It is a scalar field that describes a position in space. and time The intensity of the fire; Indicates a spatial region boundary Closed surface integrals on the surface are used to calculate the information flow into or out of the boundary region, reflecting the boundary effect of fire spread; Representing the propagating entropy flow, it is a vector field, representing the position... and time The causal information flow density; the transfer entropy is used to measure the causal influence from other locations to the current location, specifically reflecting the intensity of information transfer between fire parameters (such as temperature and gas concentration); Represents the conditional potential function The gradient; It is a scalar function, representing the state at a given position. and time Under the condition of location, Influence potential; gradient calculation Extract the direction and rate of change of the potential function in space to guide the propagation path of information flow;

[0073] The weighting coefficients of the entropy-constrained topological diffusion term are usually dynamically adjusted according to the complexity of the fire environment to balance the influence of information flow and physical diffusion. It is a graph structure or spatial region that represents the topological network of the built environment (such as rooms, corridors, etc.). This represents the coupling parameter, which controls the scaling ratio of mutual information and entropy difference, and is related to the time scale of fire development; Indicates temperature and gas concentration Mutual information between the fire spreaders indicates that the larger the value, the more obvious the abrupt changes in fire intensity. When the mutual information is higher than the maximum entropy, the spread is enhanced; conversely, the spread is weakened. This represents the maximum entropy value, used to constrain the entropy production rate of fire spread and ensure that the prediction conforms to the second law of thermodynamics. Represents the diffusion coefficient tensor, which is a location-dependent function that describes the hindering or promoting effect of building structures (such as walls, doors, and windows) on the spread of fire. It represents the tensor product, which is used to combine the diffusion coefficient and the fire situation to reflect anisotropic diffusion (i.e., the diffusion direction depends on the building topology). This represents the fire situation function.

[0074] The parameters in the equation are coupled with each other, reflecting the deep integration of information theory and physics. It has high computational complexity, involving operations such as integration, gradient, Laplace operator and exponential function, and is suitable for fire spread prediction in complex building environments.

[0075] The process involves dynamically correcting predictions by updating the causal network structure and parameters in real time. Each time new observational data is acquired, the information flow distribution within the network is recalculated, and the potential field parameters are adjusted to maintain the spatiotemporal continuity of the predictions. The final output includes fire intensity index, confidence level of spread direction, and analysis data of key influencing parameters.

[0076] S323: Using fused features as input, a continuous spatial analysis method based on field theory and topology is used to detect personnel status and location heatmaps.

[0077] Personnel location and status identification are performed using a continuous spatial analysis method based on field theory and topology. Specifically, the building space is modeled as a continuous multiphysics field function space, where the state of each point is jointly described by the temperature field, gas concentration field, and radio frequency perturbation field. Anomalies in the field distribution are identified by solving for the divergence and curl of the physical fields; these anomalies represent local disturbances caused by personnel to the physical environment. Specifically, the gradient vector streamlines of the temperature field are calculated; the presence of personnel alters the streamline distribution, forming thermodynamic singularities. Simultaneously, the isoconcentration curvature changes of the gas concentration field are analyzed; personnel respiration produces localized periodic curvature fluctuations. At the same time, the topological invariants of the differential geometric features are calculated to accurately locate the personnel's position.

[0078] To analyze personnel movement, a partial differential equation for the evolution of the physical field over time is established. The propagation characteristics of field distribution changes are analyzed, where rapidly propagating field changes correspond to personnel movement, and localized periodic changes correspond to respiratory activity. Furthermore, by analyzing the phase relationships between different physical field changes, the coherence function between field variables is calculated to distinguish between autonomous movement and field disturbances caused by environmental factors.

[0079] For personnel status identification, attractor analysis is used to project the trajectories of multiphysics fields in phase space onto a low-dimensional manifold. The Lyapunov exponent and fractal dimension of the trajectories are analyzed to quantify personnel activity states. Periodic attractors correspond to regular breathing, mixed attractors to irregular movement, and stable points to a static state. The final output is a continuous probability density function of personnel distribution within the building space, along with a personnel status classification.

[0080] S4: Collect arbitration signal source data, and construct a self-evolving knowledge base based on the confidence level of the signal source and the fire information analysis results, so as to optimize the parameters of dynamic game and multi-head deep learning model.

[0081] The process includes collecting arbitration signal source data, constructing a self-evolving knowledge base based on signal source confidence and fire information analysis results, and optimizing the parameters of the dynamic game and multi-head deep learning model. This includes the following sub-steps:

[0082] S41: Collect arbitration signal source data and calculate the basic confidence level of each signal source, and obtain the overall confidence score through pre-set confidence fusion rules;

[0083] Listen to alarm signals from completely independent traditional fire alarm systems within the building (such as manual alarm buttons and ordinary smoke detectors). If these independent systems also alarm after this system alarms, it constitutes highly confident arbitration evidence. Listen to the activation signals of fire protection facilities such as sprinkler system water flow switches and smoke exhaust fan starts. The activation of these devices is strong physical evidence of a real fire, with the highest confidence level. Monitor persistent distortions of the geomagnetic field and persistent high temperatures displayed by distributed fiber optic thermography at load-bearing structures. These are physical effects that only occur in the later stages of a fire and cannot be simulated, constituting high-confidence evidence. Monitor the development of events over a period of time after an alarm. If parameters such as temperature, gas concentration, and signal disturbances are observed to conform to the fire dynamics model (monotonically increasing, continuously spreading), it forms medium-to-high confidence evidence of internal processes. Conversely, if the signal disappears rapidly, it may be transient interference. Then, assign a basic confidence score to each type of arbitration signal. Once an arbitration time window (e.g., 15 minutes after an alarm) has ended, a set of arbitration signals is collected, and a final arbitration decision ("real fire" or "false alarm") is generated through a pre-defined confidence fusion rule. Simultaneously, the overall confidence score for this arbitration is calculated.

[0084] S42: Construct a self-evolving knowledge base based on the arbitration signal sources and fire information analysis results that are above the confidence threshold, and optimize the parameters of the dynamic game and multi-head deep learning model.

[0085] Decision data packets with an overall confidence score exceeding a preset threshold are stored in the core training set of the knowledge base. The data packet content includes:

[0086] The original multimodal sensor data (after alignment); the Shapley value sequence of each information source during the dynamic game fusion process; the final decision output of the system (alarm, location, type judgment, etc.); the final judgment and confidence score of the automatic arbitration.

[0087] The knowledge base is linked to the model version to avoid conflicts between old and new data. The ratio of real fires to false alarms is actively monitored. If a certain type of sample is too few, its weight can be appropriately increased during incremental learning to avoid model bias. Low-quality or outdated samples are also cleaned up regularly.

[0088] Then, the incremental learning process is triggered periodically, or when a certain number of new high-confidence samples are added to the knowledge base; specifically,

[0089] Using new samples, the parameters in hierarchical dynamic games can be optimized. For example, it can be learned that a new type of interference source (such as disinfectant spray) will always cause false alarms in gas sensors, thereby automatically reducing its initial contribution in similar situations in new games.

[0090] For multi-head deep learning models, the existing model weights are fine-tuned with new samples using a small learning rate.

[0091] Once optimization is complete, real-time data is processed to make a new round of decisions that are theoretically more accurate. These new decisions are then evaluated by an automatic arbitration mechanism and may generate new learning samples, thus initiating the next evolutionary cycle.

[0092] Corresponding to the above embodiments, the present invention provides a computer storage medium, including: at least one memory and at least one processor;

[0093] The memory is used to store one or more program instructions;

[0094] The processor is used to run one or more program instructions to execute a building fire intelligent detection and location method based on multi-source data fusion.

[0095] Corresponding to the above embodiments, this embodiment of the invention provides a computer-readable storage medium containing one or more program instructions, which are executed by a processor to provide a method for intelligent detection and location of building fires based on multi-source data fusion.

[0096] The embodiments disclosed in this invention provide a computer-readable storage medium storing computer program instructions. When the computer program instructions are executed on a computer, the computer executes the above-described intelligent building fire detection and location method based on multi-source data fusion.

[0097] In this embodiment of the invention, the processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0098] The various methods, steps, and logic diagrams disclosed in the embodiments of this invention can be implemented or executed. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The processor reads information from the storage medium and, in conjunction with its hardware, completes the steps of the above methods.

[0099] The storage medium can be memory, such as volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.

[0100] Among them, non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory.

[0101] Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (Synchlink DRAM, SLDRAM), and direct memory bus RAM (DRRAM).

[0102] The storage media described in the embodiments of the present invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0103] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using a combination of hardware and software. When applied as software, the corresponding functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of computer programs from one place to another. Storage media can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0104] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent detection and location of building fires based on multi-source data fusion, characterized in that, include: Passive and active sensing data are collected, and after clock synchronization, instantaneous fluctuation elimination, and noise elimination, they are encapsulated in a standardized format to construct a full-coverage complementary sensing network. The system identifies the precise spatial coordinates of two types of differentiated data, establishes a causal relationship graph based on physical laws according to the sequential response relationship of different sensors, inserts virtual probe nodes in the virtual perception enhancement stage, constructs a spatiotemporal graph model, and calculates the fire confidence and core contribution by constructing an intelligent fusion mechanism based on hierarchical dynamic game, generating a fusion feature set. Based on the fusion characteristics, multi-head deep learning is used to determine and locate the fire area and its corresponding fire protection zone, and to accurately locate the fire source, fire intensity, spread trend, and the status and location of monitoring personnel, and output the fire information analysis results. Arbitration signal source data is collected, specifically including alarm signals from traditional fire alarm systems, action signals from fire-fighting facilities, and physical effects generated in the middle and late stages of a fire that cannot be simulated. A self-evolving knowledge base is constructed based on the confidence level of the signal source and the results of fire information analysis, thereby optimizing the parameters of dynamic game theory and multi-head deep learning models.

2. The intelligent detection and location method for building fires based on multi-source data fusion according to claim 1, characterized in that, By collecting two types of differentiated data—active and passive—through a dual-sensing layer architecture, a fully covered information-complementary sensing network is constructed, including: A passive sensing layer is constructed using a radio frequency gateway and a high-precision magnetometer to collect passive data. An active sensing layer is constructed using a miniature electronic nose array, an acoustic-optical sensing unit, and radar to collect active data.

3. The intelligent detection and location method for building fires based on multi-source data fusion according to claim 1, characterized in that, A dynamic game mechanism is established based on two types of differentiated data to determine the credibility of the data source in real time and generate a high-credibility fusion feature set, including: By associating two separate types of differentiated data into a unified spatiotemporal dataset, a spatiotemporal graph model is constructed. Using independent information sources as participants, the fire confidence level of each participating group is calculated, and the core contribution of each participant to the participating group is calculated, outputting a high-confidence fusion feature set.

4. The intelligent detection and location method for building fires based on multi-source data fusion according to claim 1, characterized in that, Based on the fused feature set, parallel analysis is performed using multi-head deep learning to output fire information analysis results, including: Based on the fusion characteristics, the fire area and its corresponding fire protection zone can be quickly located using a decision-making and positioning model. After locating the area, deep feature extraction and analysis are performed through multi-head deep learning to output fire information analysis results.

5. The intelligent detection and location method for building fires based on multi-source data fusion according to claim 4, characterized in that, After locating the area, deep feature extraction and analysis are performed using multi-head deep learning to output fire information analysis results, including: Using the fusion features as input, the probability that the fire source is located at the center point of the grid is calculated through a fusion regression model based on spatial grid classification and signal source tracing. Using fused features as input, the fire intensity and spread trend can be detected by combining gas fingerprints and visual-thermodynamic features. Using fused features as input, a continuous spatial analysis method based on field theory and topology is used to detect personnel status and location heatmaps.

6. The intelligent detection and location method for building fires based on multi-source data fusion according to claim 1, characterized in that, Using fire information analysis results to generate labeled training samples, the dynamic game and multi-head deep learning analysis parameters are optimized to correct historical misjudgments, including: Collect arbitration signal source data and calculate the basic confidence level of each signal source. Obtain the overall confidence score through pre-set confidence fusion rules. A self-evolving knowledge base is constructed based on the arbitration signal sources above the confidence threshold and the fire information analysis results, and the parameters of the dynamic game and multi-head deep learning model are optimized.

7. A computer-readable storage medium, characterized in that, It includes one or more program instructions, which are executed by a processor as described in any one of claims 1-6, for a building fire intelligent detection and location method based on multi-source data fusion.

Images