Fire early warning and intelligent fire extinguishing method based on image recognition

Abstract

The invention provides a fire early warning and intelligent fire extinguishing method based on image recognition. The fire early warning and intelligent fire extinguishing method comprises the steps that a camera network is used for covering a target monitoring area, flame and smoke characteristic parameters are input, noise is removed through image preprocessing, and image data are standardized. A heat source point is selected as a camera installation position in combination with a fire propagation mode, and layout is optimized. And calling image data, performing real-time analysis based on a dynamic flame sequence, extracting abnormal response, calculating a fire risk value by using a mode recognition algorithm, and generating peak fire probability data. And collecting real-time fire characteristic data, comparing the data with a risk value after filtering and noise reduction, and setting multi-level threshold values to generate an alarm result. And analyzing a false alarm reason, optimizing the position of the camera and the starting condition of the fire extinguishing device, and generating a fire risk and fire extinguishing efficiency report. According to the invention, the accuracy and timeliness of fire early warning can be improved, the false alarm rate is reduced, and the fire extinguishing efficiency is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of intelligent security and fire protection technology, and more specifically, to an early warning and intelligent fire extinguishing method based on image recognition. Background Art

[0002] In today's society, fire safety is a constant focus. With the acceleration of urbanization and the continuous expansion of building scale, the potential risk of fire is also increasing. Traditional fire warning and extinguishing methods rely primarily on smoke detectors and temperature sensors. While these devices can provide fire alarms to a certain extent, they have certain limitations. For example, smoke detectors may cause false alarms due to interference such as dust or steam, while temperature sensors may not be able to detect abnormal temperature increases in the early stages of a fire. Furthermore, these traditional methods typically only trigger alarms after a fire has already occurred and produced noticeable smoke or temperature changes, failing to provide early warning. This can result in fires not being detected in their early stages and spreading rapidly, causing greater losses.

[0003] In recent years, image recognition technology has been widely applied in various fields, showing great potential in fire early warning. Real-time image acquisition and analysis of the monitored area through a camera network enables early identification of flames and smoke. However, existing fire early warning systems based on image recognition still have several issues. For one thing, image data is easily affected by factors such as lighting and background interference, resulting in reduced recognition accuracy. Furthermore, the issue of optimizing camera layout has not been fully addressed. How to rationally position cameras based on the fire propagation pattern within the monitored area to effectively monitor potential fire areas remains an urgent challenge. Furthermore, existing systems also have shortcomings in fire risk assessment and alarm threshold setting, failing to dynamically adjust based on real-time data, resulting in a high rate of false alarms.

[0004] Existing fire warning systems lack accurate recognition of early fire characteristics and are susceptible to environmental interference. Camera layout lacks scientific optimization methods, preventing effective coverage of potential fire areas. Fire risk assessment and alarm threshold settings are not dynamic and flexible enough, resulting in high false alarm rates. These issues limit the effectiveness of existing fire warning and extinguishing technologies in practical applications. A more efficient, accurate, and intelligent approach to early fire warning and extinguishing is urgently needed. Summary of the Invention

[0005] The present invention provides a fire early warning and intelligent fire extinguishing method based on image recognition, comprising: Step 1: Cover the target monitoring area with a camera network, input flame and smoke characteristic parameters, set the monitoring range, and use image preprocessing tools to remove noise and standardize the image data of the target monitoring area; Step 2: Based on the fire propagation pattern of the target monitoring area, select the heat source point as the camera installation location, and perform layout optimization to obtain the camera installation layout; Step 3: Retrieve image data of the installation layout and target monitoring area, perform real-time image analysis based on the dynamic flame sequence, extract abnormal responses in potential fire areas, calculate fire risk values ​​using pattern recognition algorithms, and generate peak fire probability data for each potential fire area; Step 4: Collect real-time fire characteristic data from the installed cameras, perform filtering and noise reduction processing to obtain a real-time data set, compare and analyze the real-time fire characteristic data of each heat source point in the real-time data set with the corresponding fire risk value, and set multi-level thresholds to generate alarm analysis results; Step 5: Based on the alarm data of the entire monitoring process, analyze the causes of false alarms, optimize the camera installation location and fire extinguishing device activation conditions, and generate fire risk reports and fire extinguishing efficiency reports.

[0006] Furthermore, when executing step 1, the following steps are included: Step 2.1: Use a 3D point cloud scanner to collect spatial topological data of the target monitoring area and generate a point cloud density distribution map; Step 2.2: Build a 3D structural model based on the point cloud density distribution map, and import the thermal conductivity coefficient and air convection parameters of the building materials; Step 2.3: Retrieve the flame spectrum characteristics and smoke particle motion model from the feature database and map them to the thermodynamic simulation node of the 3D structure model; Step 2.4: Set the heat source monitoring conditions in the image processing tool, define the monitoring angle range and pitch angle tolerance threshold; Step 2.5: Simulate the image acquisition distortion rate under different light intensities and generate a background interference feature template library.

[0007] Furthermore, when executing step 2, it includes: Step 3.1: Use a fluid dynamics model to simulate the fire smoke diffusion path and calculate the propagation time matrix from the heat source point to the evacuation channel; Step 3.2: Identify high fire risk areas based on the propagation time matrix and mark the temperature gradient mutation points; Step 3.3: Calculate the camera field of view overlap optimization function based on the distribution of temperature gradient mutation points; Step 3.4: Use the ray tracing algorithm to verify the blind area coverage of the camera layout and mark the coordinates of the uncovered areas; Step 3.5: Add the auxiliary camera node and recalculate the coverage matrix to output the final installation layout coordinate set.

[0008] Furthermore, when executing step 3, the following steps are included: Step 4.1: Load the flame morphology data from the historical fire database and construct the flame height fluctuation function and smoke turbulence model; Step 4.2: Inject the synthetic flame sequence at the heat source coordinates and superimpose the environmental interference noise simultaneously; Step 4.3: Use a multi-scale convolutional network to extract the flame edge sharpness features and smoke contour diffusion entropy; Step 4.4: Input the features into the bidirectional long short-term memory network to predict the fire spread speed and thermal radiation intensity increment; Step 4.5: Dynamically update the safety threshold based on the prediction results and mark the coordinates of high-risk areas.

[0009] Furthermore, when executing step 4, the following steps are included: Step 5.1: Use millimeter-wave radar to synchronously collect temperature field distribution data and align it with the camera image data in time and space; Step 5.2: Apply wavelet packet decomposition filtering to the multi-source data to separate high-frequency noise and low-frequency characteristic components; Step 5.3: Fuse the image RGB features with the millimeter-wave temperature gradient features to generate a multimodal fire feature vector; Step 5.4: Dynamically divide the warning level according to the fire risk probability value, and generate the first-level alarm threshold and the second-level fire extinguishing activation threshold; Step 5.5: When the multimodal feature vector triggers two-level thresholds simultaneously, the positioning tracking mode is activated.

[0010] Furthermore, when executing step 5, it includes: Step 6.1: Create a spatiotemporal distribution map of false alarms and correlate it with humidity sudden change events in meteorological data. Step 6.2: Extract the texture feature similarity matrix of the false alarm area and match it with the steam source category in the interference source feature library; Step 6.3: Optimize the camera position fitness function through genetic algorithm to avoid high similarity interference areas; Step 6.4: Deploy the movable camera stand and adjust the physical position in real time according to the path planning instructions; Step 6.5: Verify the reduction in false positive rate under the new layout and output the topology optimization report.

[0011] Furthermore, step 6.3 includes: Step 7.1: Construct the Voronoi diagram of the heat source point distribution and calculate the area of ​​the polygon-covered holes; Step 7.2: Identify discrete heat source clusters using the DBSCAN clustering algorithm and mark the cluster boundary coordinates; Step 7.3: Deploy wide-angle camera nodes at the edge of the cluster and adjust the pitch angle to cover the hole area; Step 7.4: Use the robotic arm collaborative control system to execute camera displacement and provide real-time feedback on positioning accuracy data; Step 7.5: When the positioning accuracy continues to exceed the preset tolerance, a manual calibration protocol is triggered.

[0012] Furthermore, setting the fire extinguishing start threshold in step 5.4 includes: Step 8.1: Analyze the topological correlation between flame morphology and smoke diffusion characteristics through graph convolutional networks; Step 8.2: Match the combustion reaction equation in the flame type database and calculate the oxygen demand index; Step 8.3: If it is an open flame, start the high-pressure water mist injection system and adjust the water mist particle size distribution according to the oxygen demand index; Step 8.4: If it is an electrical fire, inject perfluorohexanone gas and adjust the injection flow rate based on the flame height dynamic calculation function; Step 8.5: Monitor the temperature decay rate of the area covered by the fire extinguishing agent in real time and use feedback to adjust the injection parameters.

[0013] Furthermore, the pattern recognition algorithm in step 3 includes: Step 9.1: Use the Inception-v3 network to extract multi-scale texture features and generate a 256-dimensional feature vector; Step 9.2: Capture global context information through dilated spatial pyramid pooling and output a 512-dimensional feature map; Step 9.3: Concatenate the feature vector and the feature map and input them into the spatiotemporal attention module to calculate the inter-frame motion weight; Step 9.4: Generate a spatial weight mask based on the Euclidean distance between the pixel and the flame core; Step 9.5: Perform Softmax normalization on the weighted fusion features, output the fire risk probability value and mark the high-risk areas.

[0014] Furthermore, the generation of the dynamic flame sequence includes: Step 10.1: Construct a flame morphology variational autoencoder to learn the dynamic distribution of flames in the latent space. Step 10.2: Inject Poisson noise and temperature gradient labels into the input of the generator of the GAN. Step 10.3: Constrain the temporal coherence of the synthesized sequence through a cycle consistency loss function; Step 10.4: Add dust occlusion model and rain refraction effect to simulate environmental interference; Step 10.5: Calculate the Wasserstein distance between the synthetic data and the real data, and reconstruct the generator network when it exceeds the allowable error threshold.

[0015] The above embodiments of the present invention have at least the following beneficial effects: 1. By collecting spatial topological data and constructing a 3D structural model through a 3D point cloud scanner, it is possible to accurately simulate the smoke diffusion path of a fire, identify high-risk fire areas, and optimize camera layout, thereby effectively improving the coverage and accuracy of fire monitoring and solving the problem of blind spots caused by unreasonable camera layout in existing technologies.

[0016] 2. The dynamic flame sequence is analyzed using a multi-scale convolutional network and a bidirectional long short-term memory network, which can accurately extract flame and smoke characteristics, predict the fire spread speed and the increase in thermal radiation intensity, and dynamically update the safety threshold accordingly, thereby significantly improving the accuracy and real-time performance of fire risk assessment, and solving the problems of fire risk assessment in existing technologies that are not dynamic and flexible enough and have a high false alarm rate.

[0017] 3. Combining the image RGB features with the millimeter-wave temperature gradient features, a multimodal fire feature vector is generated, and the warning level is dynamically divided according to the fire risk probability value, realizing the fusion analysis and accurate alarm of multi-source data. At the same time, the camera position is optimized through genetic algorithms to avoid high-similarity interference areas, effectively reducing the false alarm rate, and solving the problem of image data being easily affected by environmental interference and frequent false alarms in existing technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which: Figure 1 A flowchart of a fire early warning and intelligent fire extinguishing method based on image recognition provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0019] The technical solutions of this application will be described clearly and completely below, in conjunction with the accompanying drawings. It should be understood that the described embodiments represent only a portion of the embodiments of this application, and not all of them. The components of this application, generally described and illustrated in the drawings herein, may be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of this application provided in the drawings is not intended to limit the scope of the claimed application, but rather merely represents selected embodiments of this application. All other embodiments derived by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application. It should be noted that similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings. Furthermore, in the description of this application, the terms "first," "second," etc., are used solely to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0020] like Figure 1 As shown, this application proposes a fire early warning and intelligent fire extinguishing method based on image recognition, which includes the following steps: Step 1: Cover the target monitoring area with a camera network, input flame and smoke characteristic parameters, set the monitoring range, and use image preprocessing tools to remove noise and standardize the image data of the target monitoring area; Step 2: Based on the fire propagation pattern of the target monitoring area, select the heat source point as the camera installation location, and perform layout optimization to obtain the camera installation layout; Step 3: Retrieve image data of the installation layout and target monitoring area, perform real-time image analysis based on the dynamic flame sequence, extract abnormal responses in potential fire areas, calculate fire risk values ​​using pattern recognition algorithms, and generate peak fire probability data for each potential fire area; Step 4: Collect real-time fire characteristic data from the installed cameras, perform filtering and noise reduction processing to obtain a real-time data set, compare and analyze the real-time fire characteristic data of each heat source point in the real-time data set with the corresponding fire risk value, and set multi-level thresholds to generate alarm analysis results; Step 5: Based on the alarm data of the entire monitoring process, analyze the causes of false alarms, optimize the camera installation location and fire extinguishing device activation conditions, and generate fire risk reports and fire extinguishing efficiency reports.

[0021] It should be noted that camera network coverage of the target surveillance area refers to the coordinated operation of multiple cameras to form a spatial monitoring matrix. This can be achieved by deploying a combination of PTZ cameras and fixed thermal imagers to ensure comprehensive coverage of the surveillance area and adapt to varying ambient lighting conditions. Image preprocessing tools remove noise and standardize image data to eliminate image quality degradation caused by environmental interference. This can be achieved using a Gaussian filter combined with a histogram equalization algorithm, improving the accuracy of subsequent feature extraction.

[0022] Fire propagation modeling refers to modeling the physical laws of fire spread. This can be achieved through computational fluid dynamics simulation combined with a database of material combustion characteristics, providing a thermodynamic basis for camera placement. Real-time image analysis of dynamic flame sequences captures the temporal variations in flame morphology. This can be achieved through the optical flow method, which tracks the motion vectors of flame edges, enhancing the timeliness of early fire identification. Pattern recognition algorithms calculate fire risk values ​​by quantifying visual features into fire probabilities. This can be achieved through the use of deep convolutional neural networks combined with support vector machine classifiers, enabling quantitative assessment of fire severity. Multi-level threshold generation of alarm analysis results establishes a phased response mechanism. This can be achieved through the use of sliding window statistics combined with Bayesian decision theory, balancing false alarm rate and response speed. Optimizing camera installation locations and fire extinguishing device activation conditions creates a closed-loop feedback system. This can be achieved through the use of reinforcement learning algorithms combined with training on historical false alarm data, continuously improving the system's adaptability.

[0023] The core innovation of this application lies in the construction of a full-link intelligent fire prevention and control system from environmental perception to decision-making optimization. The fire propagation model is used to guide the camera layout to achieve spatial coverage optimization, and the dynamic image analysis and multi-level threshold decision-making mechanism are combined to improve the accuracy of early recognition. The closed-loop feedback is used to continuously iterate the system parameters, and finally a fire warning and fire extinguishing collaborative control system with self-learning capabilities is formed.

[0024] Specifically, during the data acquisition phase, the target monitoring area is covered by a camera network. Flame and smoke characteristic parameters are input and the monitoring range is set to establish an accurate physical model foundation for subsequent analysis. Image preprocessing tools are then used to remove noise and standardize the image data of the target monitoring area, eliminating environmental noise interference and improving the quality of the original data.

[0025] Furthermore, based on the fire propagation patterns in the target monitoring area, heat source points were selected as camera installation locations, and the camera installation layout was optimized. This step ensured that the spatial distribution of monitoring equipment was consistent with fire spread patterns, ensuring priority coverage of high-risk areas.

[0026] Subsequently, image data from the installation layout and target monitoring area is retrieved and real-time image analysis is performed based on dynamic flame sequences. This real-time analysis method, based on dynamic flame sequences, captures the temporal variations in flame morphology. By extracting abnormal responses from potential fire areas, a pattern recognition algorithm is used to calculate fire risk values, generating peak fire probability data for each potential fire area. This step quantifies abstract fire data into probability values, enabling quantitative assessment of early-stage fire activity.

[0027] Next, real-time fire signature data from the installed cameras is collected and filtered for noise reduction to create a real-time dataset. The real-time fire signature data for each heat source in the dataset is compared with the corresponding fire risk value, and multi-level thresholds are set to generate alarm analysis results. This multi-level threshold setting mechanism considers the correlation between real-time data and historical risk values ​​while establishing a phased response strategy to effectively balance false alarms and missed alarms.

[0028] Finally, based on the alarm data collected throughout the monitoring process, the system analyzes the causes of false alarms, optimizes camera installation locations and fire extinguishing device activation conditions, and generates fire risk and fire extinguishing efficiency reports. This closed-loop feedback mechanism leverages false alarm data analysis to continuously optimize system parameters, forming a dynamic, iterative intelligent decision-making model.

[0029] As a result, through the two-way interaction of data flow and decision flow between each step, a complete technical chain from environmental perception to intelligent decision-making is built, realizing the whole process management of early fire warning and intelligent fire extinguishing.

[0030] As a preferred embodiment, the solution of this application is specifically implemented as follows: First, deploy an array of high-definition network cameras covering the target monitoring area. The cameras use a 1080p resolution and a frame rate of 30 fps. Flame characteristic parameters are set to the infrared spectral range of 450-1100 nm, and smoke characteristic parameters are set to a grayscale value change rate threshold of 0.1 / s. The monitoring range is set to 120° horizontally and 90° vertically within the camera's field of view. Gaussian filters are used to reduce noise in the captured images, and histogram equalization is used for image standardization.

[0031] Furthermore, computational fluid dynamics software was used to simulate the fire propagation pattern in the target area, identifying the locations of heat sources with the fastest temperature gradients. Cameras were then installed at these locations, and their layout was optimized using a genetic algorithm to achieve a field of view overlap of over 95%.

[0032] The optimized camera layout and real-time image data stream are then used. Deep learning models such as YOLOv5 are used to detect and track targets in dynamic flame sequences, extracting flame edge, color, and motion features. A convolutional neural network is used to classify these features, calculate the fire risk value for each potential fire area, and generate a risk heat map.

[0033] Next, real-time data from the installed cameras is collected and denoised using a wavelet transform. This processed real-time data is compared and analyzed with the previously generated risk value. Three alarm thresholds are set: a level one warning is issued when the risk value exceeds 0.3, a level two alarm is issued when it exceeds 0.6, and an automatic fire extinguishing system is triggered when it exceeds 0.8.

[0034] Finally, alarm data from the entire monitoring process is recorded, including information such as alarm time, location, and type. Data mining techniques are used to analyze the causes of false alarms, such as changes in ambient light and reflective objects. Based on the analysis, the camera's installation angle and position are adjusted to optimize the triggering conditions of the fire extinguishing device. A risk assessment report is generated, including indicators such as the distribution of high-risk areas, false alarm rate, and average response time, as well as a fire extinguishing efficiency report with data such as fire extinguishing agent usage efficiency and fire extinguishing success rate.

[0035] Furthermore, when executing step 1, it includes: collecting spatial topological data of the target monitoring area through a three-dimensional point cloud scanner to generate a point cloud density distribution map; constructing a three-dimensional structural model based on the point cloud density distribution map, and importing the thermal conductivity coefficient and air convection parameters of the building materials; calling the flame spectrum characteristics and smoke particle motion model from the feature database, and mapping them to the thermodynamic simulation nodes of the three-dimensional structural model; setting the heat source point monitoring conditions in the image processing tool, defining the monitoring angle range and pitch angle tolerance threshold; simulating the image acquisition distortion rate under different light intensities, and generating a background interference feature template library.

[0036] The acquisition accuracy of the 3D point cloud scanner can be set to contain at least 1,000 point cloud data points per cubic meter. A point cloud density distribution map is generated through lidar scanning. Furthermore, the grid resolution of the point cloud density distribution map can be configured to 5 mm. When constructing the 3D structural model, the thermal conductivity of building materials is obtained from the material database. For example, the thermal conductivity of concrete is taken as 1.7 W / (m·K). Air convection parameters are dynamically adjusted based on the number and location of ventilation openings in the monitored area. Flame spectral features are retrieved, including the flame's radiant intensity curve in the infrared band. The smoke particle motion model uses a diffusion equation based on Brownian motion. When setting the heat source monitoring conditions, the monitoring angle range can be defined as 120 degrees horizontally, with a pitch angle tolerance threshold of ±5 degrees. The background interference feature template library is generated by simulating three typical scenarios: direct sunlight at noon, diffuse light at dusk, and artificial lighting at night. The color temperature offset and contrast attenuation coefficient of the image acquisition device are calculated for each scenario.

[0037] Specifically, a 3D point cloud scanner is used to scan the target area from multiple angles, acquiring spatial geometric data and generating a density distribution map, thereby establishing a 3D model containing the building's structural details. During the model construction phase, the thermal conductivity of the building materials is imported into the corresponding structural units. For example, the thermal conductivity of steel is assigned a value of 50 W / (m·K) and that of wood is assigned a value of 0.12 W / (m·K). Air convection parameters are set to a flow rate range of 0.1-2.5 m / s based on the regional ventilation conditions. The flame spectrum characteristics extract radiation intensity distribution data within a specific wavelength range from a preset database. The smoke particle motion model calculates the diffusion velocity and concentration gradient changes of particles in three-dimensional space and maps them to the model's thermodynamic nodes for coupled simulation. In the image processing tool, the monitoring angle range is set to cover a 120-degree sector-shaped area horizontally. The pitch angle tolerance threshold is determined through geometric optics calculations to ensure that the camera installation position meets the vertical adjustment margin of ±5 degrees. For different lighting conditions, the ray tracing algorithm is used to simulate interference features including reflected glare, shadow occlusion and color temperature changes, and a template library containing 100 typical scenes is generated for subsequent image denoising.

[0038] As a preferred embodiment, the solution of this application is specifically implemented as follows: A 3D point cloud scanner collects spatial topological data of the target monitoring area and generates a point cloud density distribution map. The point cloud scanner uses LiDAR technology to scan the entire monitoring area at a sampling rate of millions of points per second, generating high-precision 3D point cloud data.

[0039] A 3D structural model was constructed based on the point cloud density distribution map, and the thermal conductivity and air convection parameters of the building materials were imported. Using computer-aided design software, the point cloud data was converted into an editable 3D mesh model. The thermal conductivity of various building materials was retrieved from the material database. For example, the thermal conductivity of concrete is 1.6 W / (m·K) and that of wood is 0.15 W / (m·K). Air convection parameters were set based on the indoor temperature and airflow velocity.

[0040] Flame spectral features and smoke particle motion models are retrieved from the feature database and mapped to the thermodynamic simulation nodes of the 3D structural model. Flame spectral features include the characteristic wavelengths of different burning substances, while the smoke particle motion model is constructed based on fluid mechanics principles. These features and models are mapped to pre-set thermodynamic simulation nodes in the 3D model for subsequent fire simulation analysis.

[0041] In the image processing tool, set the hotspot monitoring conditions and define the monitoring angle range and pitch angle tolerance threshold. You can set the camera's horizontal monitoring angle range to 120 degrees, the vertical pitch angle range to -30 degrees to +60 degrees, and the tolerance threshold to ±5 degrees to ensure complete monitoring coverage.

[0042] Simulate image acquisition distortion rates under varying light intensities to generate a template library of background interference features. Using illumination simulation software, simulate varying lighting conditions from 0 lux to 100,000 lux, analyze image distortion, and build a template library containing background interference features under various lighting conditions, providing a reference benchmark for subsequent image processing.

[0043] Furthermore, a fluid dynamics model is used to simulate the fire smoke diffusion path and calculate the propagation time matrix from the heat source point to the evacuation channel; based on the propagation time matrix, high fire risk areas are identified and temperature gradient mutation points are marked; based on the distribution of temperature gradient mutation points, the camera field of view angle overlap rate optimization function is calculated; the blind spot coverage of the camera layout is verified through a ray tracing algorithm, and the coordinates of the uncovered area are marked; auxiliary camera nodes are added and the coverage matrix is ​​recalculated to output the final installation layout coordinate set.

[0044] The fluid dynamics model uses the finite volume method to discretize the Navier-Stokes equations. The simulation time step for the fire smoke diffusion path is set between 0.1 and 1 seconds, for example, a 0.5-second step to capture detailed smoke flow. The propagation time matrix uses the Dijkstra algorithm to calculate the shortest time path from the heat source to the evacuation route, with the time weighting determined by the smoke flow velocity and temperature distribution. Temperature gradient abrupt changes are labeled by setting a threshold. For example, a temperature change rate exceeding 5°C / second between adjacent grid cells is marked as a abrupt change. The camera field of view overlap optimization function uses a geometric coverage model to match the camera pitch angle tolerance threshold with the spatial coordinates of the temperature gradient abrupt change points. For example, a convex optimization algorithm is used to determine the minimum number of cameras required to cover all abrupt changes. The ray tracing algorithm uses the Monte Carlo path tracing method, with the number of ray reflections set between 3 and 5. For example, 4 reflections balances computational accuracy and efficiency. The addition of auxiliary camera nodes is achieved through a greedy algorithm, which prioritizes deploying nodes at locations with the largest coverage gain in the coordinates of uncovered areas. For example, each iteration selects locations where the coverage blind area reduction rate exceeds 30%.

[0045] Specifically, the fluid dynamics model establishes a three-dimensional unsteady turbulence model, linking the combustion rate of the heat source with the thermal conductivity of the building material to generate a dynamic smoke diffusion path. During the calculation of the propagation time matrix, the exit coordinates of the evacuation corridor are set as the target node, and the coordinates of the heat source are used as the starting node. A propagation time heat map is generated using a time accumulation function. Temperature gradient mutation points are identified based on the spatiotemporal distribution characteristics of the propagation time matrix. A sliding window method is used to detect areas where the temperature change rate exceeds a preset threshold, for example, within a 3×3 grid. The camera field of view overlap optimization function calculates the intersection area between each camera's coverage area and the temperature gradient mutation point, and iteratively optimizes the coverage rate to achieve a coverage rate of at least 95%. During the ray tracing algorithm verification, the camera's installation height and viewing angle parameters are input into the ray emission model, and the coordinates of the areas not covered by the light are calculated using a collision detection algorithm. Auxiliary camera nodes are deployed using a dynamic programming approach, linking coverage matrix updates with the elimination of blind spot coordinates. For example, the coverage matrix rank is recalculated after each node is added to evaluate the effectiveness of the layout optimization. Therefore, through the precise quantification of the propagation time matrix, the spatial annotation of the temperature gradient mutation points, the dynamic optimization of the field of view angle overlap rate, and the reverse verification of the blind spot coverage rate, the final output installation layout coordinate set can achieve full coverage of high fire risk areas while avoiding the deployment of redundant cameras.

[0046] As a preferred embodiment, the solution of this application is specifically implemented as follows: a fire smoke diffusion model is established using the computational fluid dynamics software ANSYS Fluent, with a grid resolution of 0.1m. A CAD model of the building structure and material thermal properties are imported. Heat source points are set in the model to simulate the smoke diffusion process under different fire intensities. The propagation time matrix from the heat source point to each evacuation route is calculated. Based on the propagation time matrix, a K-means clustering algorithm is used to identify high-fire risk areas, and temperature gradient mutation points are annotated using thermal imaging image processing techniques. Based on the spatial distribution of temperature gradient mutation points, a camera field of view angle overlap optimization function is constructed, with the objective of maximizing overlap coverage. A ray tracing algorithm is used to verify the camera layout plan, simulating the propagation path of light in three-dimensional space, calculating blind spot coverage, and marking the coordinates of uncovered areas. For uncovered areas, auxiliary camera nodes are added, and the node positions are optimized using a genetic algorithm. The coverage matrix is ​​recalculated until the coverage requirement is met. Finally, a set of camera installation layout coordinates is output, including the spatial coordinates and orientation angle of each camera.

[0047] Furthermore, when executing step 3, it includes: loading flame morphology data from the historical fire database, constructing a flame height fluctuation function and a smoke turbulence model; injecting a synthetic flame sequence at the coordinates of the heat source point, and synchronously superimposing environmental interference noise; using a multi-scale convolutional network to extract the flame edge sharpness characteristics and the smoke contour diffusion entropy value; inputting the characteristics into a bidirectional long short-term memory network to predict the fire spread speed and the thermal radiation intensity increment; dynamically updating the safety threshold according to the prediction results, and marking the coordinates of high-risk areas.

[0048] The flame height fluctuation function is constructed based on the statistics of flame height variations recorded in a historical fire database. For example, a Gaussian mixture model is used to fit the flame height distribution at different combustion stages, with the flame height variance parameter set to a range of 0.5-1.2 meters. The smoke turbulence model is established by calculating the vortex intensity coefficient of the historical smoke diffusion path, for example, using the Reynolds stress tensor to model turbulent kinetic energy. Synthetic flame sequences are generated by injecting Poisson noise and temperature gradient labels at the coordinates of the heat source. For example, dynamic flame sequences are generated at a rate of 30 frames per second, with the signal-to-noise ratio of ambient interference noise kept within the range of 15-25 dB. A multi-scale convolutional network uses a parallel convolutional structure, with a large-scale convolution kernel of 7×7 pixels to capture macroscopic smoke diffusion trends and a small-scale convolution kernel of 3×3 pixels to extract flame edge sharpness features. The bidirectional long short-term memory network has a time window of 10 seconds, with 128 hidden units in both the forward and backward propagation layers. The dynamic update of the safety threshold is achieved by calculating the exponentially weighted moving average of the fire spread rate. For example, the threshold update mechanism is triggered when the predicted speed exceeds the baseline value by 20%.

[0049] Specifically, when loading the historical fire database, flame morphology data is normalized and then fed into the wave function generation module. The time series of flame height data is decomposed into trend and periodic terms, for example, using empirical mode decomposition to isolate fluctuation components at different time scales. Synthetic flame sequences are injected using a generative adversarial network. The generator network receives the coordinates of the heat source and environmental parameters as input, and the discriminator network performs adversarial training on the synthetic sequences against real flame sequences. During the feature fusion stage of the multi-scale convolutional network, feature maps at different scales are weighted and concatenated using a channel-wise attention mechanism, for example, using a SE module to calculate channel weight coefficients. The output layer of the bidirectional long short-term memory network is connected to a time-distributed fully connected layer, mapping the hidden state at each time step into an incremental prediction of the fire spread rate and thermal radiation intensity. The safety threshold update process combines real-time prediction results with the historical safety threshold database, for example, using a Kalman filter algorithm to dynamically adjust the threshold. High-risk areas are marked by calculating the difference between the incremental thermal radiation intensity and the building structure's thermal load capacity. When this difference falls below a 15% safety margin, a coordinate marking instruction is triggered.

[0050] As a preferred embodiment, the solution of this application is specifically implemented as follows: First, flame morphology data from a historical fire database was loaded to construct a flame height fluctuation function and a smoke turbulence model. Specifically, multiple time-varying data series of flame height were extracted from the database. Their spectral characteristics were analyzed using Fourier transforms to obtain the primary frequency components. Based on these frequency components, a flame height fluctuation function was constructed: H(t) = Asin(ωt + φ) + Bsin(2ωt + φ) + C, where A, B, and C are amplitude coefficients, ω is the fundamental frequency, and φ is the phase. For the smoke turbulence model, the Navier-Stokes equations were used to describe fluid motion, combined with the k-ε turbulence model to calculate the turbulent kinetic energy and dissipation rate.

[0051] Next, a synthetic flame sequence is injected at the heat source coordinates, while simultaneously superimposing environmental interference noise. A generative adversarial network (GAN) generates synthetic sequences that resemble real flames and inserts them at the heat source coordinates. Environmental interference noise, such as Gaussian white noise and salt and pepper noise, is superimposed on the synthetic sequence at a preset ratio.

[0052] Furthermore, a multi-scale convolutional network is used to extract flame edge sharpness features and smoke contour diffusion entropy. The network architecture consists of multiple parallel convolution branches, each using kernels of different sizes, such as 3x3, 5x5, and 7x7. This approach simultaneously captures subtle variations in flame edges and large-scale features of smoke diffusion. For flame edge sharpness, the variance of the image gradient amplitude is calculated; for smoke contour diffusion entropy, the Shannon entropy formula is used to calculate the uncertainty of the contour pixel distribution.

[0053] The extracted features are then fed into a bidirectional long short-term memory (Bi-LSTM) network to predict the fire spread rate and thermal radiation intensity increment. The Bi-LSTM network consists of two LSTM layers, one forward and one backward, each containing multiple LSTM units. The network input is a time series feature vector, and the output is the predicted fire spread rate and thermal radiation intensity increment.

[0054] Finally, based on the prediction results, the safety threshold is dynamically updated, and the coordinates of high-risk areas are marked. The predicted fire spread rate and thermal radiation intensity increment are compared with preset baseline values. If they exceed a certain percentage, the safety threshold is raised accordingly. High-risk areas are marked using a clustering algorithm, clustering coordinate points with higher predicted risk values ​​to identify continuous high-risk areas.

[0055] Furthermore, when executing step 4, it includes: synchronously collecting temperature field distribution data through millimeter wave radar, and aligning it with the camera image data in time and space; performing wavelet packet decomposition filtering on multi-source data to separate high-frequency noise and low-frequency feature components; fusing image RGB features with millimeter wave temperature gradient features to generate a multimodal fire feature vector; dynamically dividing the warning level according to the fire risk probability value, generating a first-level alarm threshold and a second-level fire extinguishing start threshold; when the multimodal feature vector triggers the two-level threshold at the same time, activating the positioning and tracking mode.

[0056] Synchronous data acquisition between the millimeter-wave radar and camera is achieved through timestamp matching, for example, using GPS synchronization signals or network time protocols, ensuring that data collection intervals are less than 50 milliseconds apart. Wavelet packet decomposition and filtering utilizes Daubechies wavelet basis functions to perform a four-layer decomposition, filtering out components with frequencies above 100 Hz as high-frequency noise while retaining low-frequency features between 0.1 and 10 Hz. Image RGB features and millimeter-wave temperature gradient features are integrated through feature concatenation. For example, the 128-dimensional feature vector of the three RGB channels and the 64-dimensional feature vector of the temperature gradient are concatenated into a 192-dimensional multimodal vector. Dynamic warning levels are assigned based on threshold updates using a Bayesian probability model. For example, a fire risk probability exceeding 0.7 triggers a level 1 alarm, while a fire risk exceeding 0.9 activates level 2 fire suppression. When positioning and tracking mode is activated, a particle filter algorithm is used to calculate the fire source coordinates in real time. For example, the millimeter-wave radar's beamforming technology achieves positioning accuracy with a spatial resolution of 0.5 meters.

[0057] Specifically, during the spatiotemporal alignment of the millimeter-wave radar and camera, the radar's polar coordinate data is converted to the camera's pixel coordinate system using a coordinate transformation matrix. For example, joint calibration using a calibration plate eliminates matching errors caused by viewing angle differences. The low-frequency feature components after wavelet packet decomposition and filtering are input into the feature fusion module. The RGB features are extracted using a convolutional neural network to extract statistical features of the flame color histogram, and the temperature gradient features are calculated using a sliding window to calculate the rate of change of temperature differences between adjacent sampling points. After the multimodal fire feature vector is generated, it is input into a support vector machine classifier for joint classification. For example, a valid fire signal is identified when the confidence difference between the RGB and temperature gradient features is less than 0.15. During the dynamic thresholding process, the first-level alarm threshold is dynamically adjusted based on the ambient temperature baseline. For example, for every 5°C increase in the baseline temperature, the alarm threshold increases by 10%. When both thresholds are triggered simultaneously, the positioning and tracking mode calculates the fire source trajectory through multi-sensor data fusion. For example, by combining the camera's optical flow features with the radar's Doppler velocity information, a three-dimensional prediction of the fire's spread direction is generated.

[0058] The fire risk probability value is dynamically divided into warning levels, generating a first-level alarm threshold and a second-level fire extinguishing activation threshold. A risk probability value of 0.6 can be set as the first-level threshold and 0.8 as the second-level threshold, and these values ​​can be dynamically adjusted based on historical data.

[0059] When the multimodal feature vector triggers two thresholds simultaneously, the positioning and tracking mode is activated. The Kalman filter algorithm is used to track the fire source in real time and accurately locate the fire area.

[0060] Furthermore, the present application also includes constructing a Voronoi diagram of the distribution of heat source points and calculating the area of ​​the holes covered by polygons; identifying discrete heat source clusters through the DBSCAN clustering algorithm and marking the cluster boundary coordinates; deploying wide-angle camera nodes at the cluster boundaries and adjusting the pitch angle to cover the hole area; using a robotic arm collaborative control system to execute camera displacement and provide real-time feedback on positioning accuracy data; when the positioning accuracy continues to exceed the preset tolerance, a manual calibration protocol is triggered.

[0061] The Voronoi diagram for heat source distribution can be constructed using the Fortune algorithm to generate a spatial subdivision structure, where each heat source corresponds to a Voronoi cell. The cell area can be calculated using Green's formula integration, for example, by calculating the area of ​​a polygon using a sequence of edge vertex coordinates. The density threshold for the DBSCAN clustering algorithm can be set to include at least three heat sources per square meter, and the neighborhood radius parameter can be dynamically adjusted based on the size of the monitored area, for example, setting an EPS value of 0.5 meters for a 100-square-meter area. Wide-angle camera nodes can be deployed at the geometric center of the cluster boundary, with a pitch adjustment range of ±30 degrees and an expandable field of view of up to 120 degrees. Positioning accuracy feedback data for the robotic arm collaborative control system can include six-axis attitude parameters, with a data update frequency of at least 10 Hz and preset tolerances of ±2 mm for position error and ±0.5 degrees for angle error. The manual calibration protocol trigger can be set to trigger five consecutive instances of feedback data outside the tolerance range, at which point a calibration command is pushed to the operator terminal via the HMI interface.

[0062] More specifically, in analyzing the spatial distribution of heat sources in fire risk areas, the monitoring area is first divided into multiple polygonal cells using a Voronoi diagram spatial partitioning algorithm, with each cell corresponding to the sphere of influence of a heat source. By calculating the area differences between each cell, coverage holes can be accurately identified. For example, coverage holes are identified when the area difference between adjacent cells exceeds 50% of the average. Subsequently, the density-based DBSCAN algorithm is used to cluster discretely distributed heat sources. By setting the minimum sample number and neighborhood radius parameters, dense clusters can be effectively distinguished from noise points. For example, setting min_samples=5 and eps=0.7 meters accurately marks cluster boundaries. When deploying wide-angle cameras at cluster boundaries, a geometric center point positioning algorithm is used to determine the optimal installation position. The camera's field of view is then adjusted to cover the gap area using a pitch adjustment mechanism. For example, increasing the pitch angle by 15 degrees extends the monitoring range by 3 meters. When the robotic arm performs displacement operations, encoders collect the rotation angles of each joint in real time and calculate the end-arm pose using a forward kinematic model. If the deviation between the feedback data and the target position consistently exceeds a preset threshold—for example, if the positioning error exceeds 1.5mm three times in a row—the system automatically generates a calibration request and interrupts the automatic adjustment process. This collaborative mechanism, based on spatial topology analysis and dynamic feedback, increases blind spot coverage to over 98% while reducing the false alarm rate to below 0.3%.

[0063] As a preferred embodiment, a Voronoi diagram of the heat source distribution is first constructed to calculate the area of ​​the polygonal coverage holes. Specifically, the coordinates of the heat source points are used as seed points. The distance from any point on the plane to each seed point is calculated to divide the plane into several polygonal regions. Furthermore, the area within each polygonal region that is not covered by the camera's field of view is calculated as the hole area.

[0064] Next, we used the DBSCAN clustering algorithm to identify discrete heat source clusters and annotate the cluster boundary coordinates. The DBSCAN algorithm, a density-based spatial clustering method, can effectively identify point clusters of arbitrary shapes. Thus, we divided the heat source points into several clusters based on their density distribution characteristics, and annotated the boundary coordinates of each cluster.

[0065] Next, wide-angle camera nodes are deployed at the cluster boundaries, and their pitch angles are adjusted to cover the hole areas. Specifically, wide-angle cameras with large fields of view are selected and installed at key locations along the cluster boundary. Furthermore, the camera's pitch angle is adjusted to ensure that its field of view covers the previously identified hole areas.

[0066] Then, a robotic arm collaborative control system executes camera movement, providing real-time feedback on positioning accuracy. The robotic arm system precisely controls the camera's spatial position and posture. Based on the output of the optimization algorithm, the robotic arm is driven to move the camera to the target location. Simultaneously, a position sensor collects the camera's actual coordinates in real time and compares them with the target coordinates to calculate positioning accuracy.

[0067] Finally, when positioning accuracy consistently exceeds a preset tolerance, a manual calibration protocol is triggered. For example, a threshold for acceptable positioning error can be set. Furthermore, if positioning errors exceed this threshold for multiple consecutive samplings, a calibration request is automatically generated, notifying relevant personnel for manual intervention and precise calibration.

[0068] This application further proposes to construct a Voronoi diagram of the distribution of heat source points and calculate the area of ​​the voids covered by polygons; identify discrete heat source clusters through the DBSCAN clustering algorithm and mark the cluster boundary coordinates; deploy wide-angle camera nodes at the cluster boundary and adjust the pitch angle to cover the void area; use a robotic arm collaborative control system to execute camera displacement and provide real-time feedback on positioning accuracy data; and trigger a manual calibration protocol when the positioning accuracy continues to exceed the preset tolerance.

[0069] When constructing a Voronoi diagram of the heat source distribution, the Fortune algorithm can be used to generate Thiessen polygons. The overlap ratio between the area of ​​each polygon unit and its adjacent units is calculated. For example, areas with a heat source density below 0.5 per square meter are identified as coverage holes. When processing using the DBSCAN clustering algorithm, the neighborhood radius eps is set to 1.2 meters and the minimum number of samples min_samples is set to 5. Isolated heat sources with a density below the threshold are classified as discrete heat source clusters, and the coordinates of the cluster convex hull vertices are extracted as boundary annotation points. The wide-angle camera node's field of view can be set to 170 degrees, with a pitch adjustment range of ±30 degrees. A spherical projection model is used to calculate the three-dimensional spatial mapping relationship between the coverage area and the holes. The robotic arm collaborative control system uses a six-axis robotic arm equipped with a camera gimbal. A laser tracker acquires displacement coordinates in real time, with a positioning accuracy error within ±2 mm. A calibration protocol is triggered when the displacement error exceeds 3 mm for five consecutive times.

[0070] After generating a Voronoi diagram of the heat source distribution, coverage holes can be identified by calculating the area differences of polygonal cells. For example, areas where the area of ​​a hole exceeds 150% of the average area of ​​adjacent cells are marked as monitoring blind spots. The DBSCAN clustering algorithm performs density analysis on discretely distributed heat sources. After extracting the cluster boundary coordinates, a wide-angle camera is deployed 0.5 meters outside the boundary line. The pitch angle is adjusted to ensure that the monitoring range covers the three-dimensional space of the hole area. The robotic arm collaborative control system drives the camera movement based on the optimized coordinate data. The real-time positioning accuracy data is compared with the three-dimensional structural model. If the cumulative error exceeds the safety threshold, it automatically switches to manual calibration mode to recalibrate the reference points. This process combines spatial segmentation with dynamic adjustment to ensure that the camera layout covers both high-density heat source areas and discrete cluster boundaries, eliminating monitoring blind spots caused by uneven heat source distribution and improving the spatial adaptability of the fire warning system.

[0071] As a preferred embodiment, a Voronoi diagram of the heat source distribution is first constructed, and the area of ​​the polygon coverage holes is calculated. Specifically, a Delaunay triangulation algorithm can be used to generate a Voronoi diagram of the heat source points, and then the coverage holes are quantified by calculating the area of ​​the Voronoi polygons. For example, an area threshold can be set, and when the area of ​​a Voronoi polygon exceeds the threshold, it is identified as a coverage hole.

[0072] The DBSCAN clustering algorithm is used to identify discrete heat source clusters and annotate their boundary coordinates. The DBSCAN algorithm can classify heat source clusters based on their density, with parameters such as the neighborhood radius and the minimum number of points set as clustering parameters. For each identified cluster, the boundary coordinates are determined using either the convex hull algorithm or the α-shape algorithm.

[0073] Camera movement is performed using a collaborative control system with a robotic arm. The robotic arm utilizes a 6-degree-of-freedom structure, and an inverse kinematics algorithm calculates joint angles to achieve precise camera positioning. The system provides real-time feedback on positioning accuracy data, including position and attitude errors.

[0074] When positioning accuracy consistently exceeds a preset tolerance, a manual calibration protocol is triggered. A time window and error threshold can be set. If positioning error consistently exceeds the threshold within this time window, the system automatically issues a calibration request. Manual calibration can be performed through the remote operation interface, adjusting the robot's zero position and calibration parameters.

[0075] Furthermore, when setting the fire extinguishing initiation threshold, the present application includes: analyzing the topological correlation between flame morphological characteristics and smoke diffusion characteristics through a graph convolutional network; matching the combustion reaction equation in the flame type database and calculating the oxygen demand index; if it is an open flame, starting the high-pressure water mist injection system and adjusting the water mist particle size distribution according to the oxygen demand index; if it is an electrical fire, injecting perfluorohexanone gas and adjusting the injection flow rate based on the flame height dynamic calculation function; real-time monitoring of the temperature decay rate of the area covered by the fire extinguishing agent, and feedback adjustment of the injection parameters.

[0076] The graph convolutional network can be configured as a topology with five layers of convolutional kernels, each with a size of 3×3. The spatial distribution correlation between flame edges and smoke particles is captured through an edge weight matrix. The combustion reaction equation is matched by least squares fitting the flame spectral characteristics to a standard combustion chain reaction model. For example, the oxygen consumption coefficient of the methane combustion equation (CH4+2O2→CO2+2H2O) is extracted as the basis for calculating the oxygen demand index. The high-pressure water mist injection system regulates the water mist particle size distribution by controlling the driving voltage of the piezoelectric atomizer. When the oxygen demand index exceeds 0.85, the median particle size of the water mist is adjusted to below 50 microns to enhance oxygen isolation. The injection flow rate calculation function for perfluorohexanone gas is defined as the product of the square of the flame height and the gas diffusion coefficient. For example, when the flame height is 1.2 meters, the injection flow rate is set to 0.8 cubic meters per second. The temperature decay rate is monitored using an infrared thermal imaging array, which collects temperature gradient data of the area covered by the fire extinguishing agent every 200 milliseconds. When the temperature decay rate is lower than 15°C per minute, the injection pressure increase command is triggered.

[0077] Flame morphological and smoke diffusion characteristics are fused through the node embedding layer of the graph convolutional network, generating a topological graph structure that incorporates spatial correlations, enabling accurate flame type identification. During the combustion reaction equation matching process, the oxygen demand index is calculated by multiplying the molar number of oxygen atoms by the combustion rate, providing a thermodynamic basis for selecting the extinguishing agent type. In open fire scenarios, the water mist particle size distribution is dynamically adjusted based on the oxygen demand index. For example, when the oxygen demand index is 0.9, the atomizer drive voltage is increased to 12V to produce water mist particles with a median size of 30 microns. For electrical fire scenarios, the flame height dynamic calculation function is derived by converting the pixel height of the flame outline to a known reference. For example, for every 0.5-meter increase in flame height, the injection flow rate increases by 0.3 cubic meters per second. The temperature decay rate of the extinguishing agent coverage area is determined through time series analysis of thermal imaging data. When the temperature decay rate in a local area falls below a preset threshold, the injection angle is automatically adjusted by 5 degrees to expand coverage. Through the above-mentioned parameter linkage mechanism, a precise match between the type of fire extinguishing agent and the injection parameters is achieved, and the spatial distribution efficiency of the fire extinguishing agent is optimized through closed-loop feedback control.

[0078] This application further proposes to perform the following steps in the pattern recognition algorithm: use the Inception-v3 network to extract multi-scale texture features to generate a 256-dimensional feature vector; capture global context information through void spatial pyramid pooling to output a 512-dimensional feature map; splice the feature vector and the feature map and input them into the spatiotemporal attention module to calculate the inter-frame motion weight; generate a spatial weight mask based on the Euclidean distance between the pixel point and the flame core; perform Softmax normalization on the weighted fusion feature to output the fire risk probability value and mark the high-risk area.

[0079] The Inception-v3 network uses a parallel multi-branch convolutional architecture to extract flame texture features at different scales. For example, it uses three convolutional kernels, 1×1, 3×3, and 5×5, to simultaneously process the input image. The output features of each branch are concatenated in the channel dimension to form a 256-dimensional feature vector. Dilated spatial pyramid pooling processes feature maps in four dilated convolutional layers with different dilation rates (1, 6, 12, and 18). The feature maps of multi-scale receptive fields are fused to output a 512-dimensional global feature map. The spatiotemporal attention module uses a three-dimensional convolutional kernel to analyze motion changes between consecutive frames. The temporal dimension weight coefficient is calculated using a gated recurrent unit to enhance the temporal correlation of flame flicker features. To generate the spatial weight mask, the coordinates of the flame core are determined by detecting the peak thermal radiation intensity. The weight of each pixel is negatively correlated with its Euclidean distance to the core point. For example, the weight decays to 30% of the initial value when the distance exceeds 50 pixels.

[0080] During real-time image analysis, the input dynamic flame sequence is first segmented into consecutive frames. The Inception-v3 network extracts multi-scale features from each frame, capturing subtle texture variations at the flame edge and the irregular patterns of smoke diffusion. The dilated spatial pyramid pooling module fuses multi-scale contextual information from the feature map to enhance modeling of the correlation between the flame and the diffusion path of the building structure. The resulting 768-dimensional fused features are then fed into the spatiotemporal attention module, which calculates the motion vector differences between adjacent frames to generate temporal weight coefficients to enhance the temporal correlation of the flame flicker frequency features. A spatial weight mask dynamically adjusts the feature contribution of each region based on the core location of the heat source, suppressing interference signals far from the fire source. Finally, the weighted fused features are mapped to a probability value between 0 and 1 using the Softmax function. A probability value exceeding 0.85 triggers a high-risk area flag.

[0081] More specifically, during real-time image analysis, the Inception-v3 network performs multi-branch convolutions on the input video frames. Three convolution kernels, 1×1, 3×3, and 5×5, are used in parallel to process the flame region. Multi-scale features, including flame edge sharpness and smoke particle density, are extracted, outputting a 256-dimensional feature vector. Furthermore, a spatial pyramid structure is constructed using dilated convolution layers with dilation rates of 6, 12, and 18, respectively. Multi-receptive field fusion captures the correlation between the flame and background elements, such as building vents and electrical equipment, generating a 512-dimensional global context feature map. These feature vectors and maps are concatenated along the channel dimension and fed into a 3D convolutional module with a temporal self-attention mechanism. This module dynamically assigns feature weights to each frame by calculating the similarity of the motion trajectories of flame pixels across five consecutive frames. A spatial weight mask is constructed based on the coordinates of the flame core region using a Gaussian function. The weight coefficient for pixels within 2 meters of the flame core is increased to 0.85, while the weight for pixels beyond 5 meters is reduced to below 0.3. Finally, the weighted fusion features are input into the Softmax classifier to output the fire risk probability value of each pixel. When the probability value exceeds 0.7, a red marking box is automatically generated to cover the high-risk area.

[0082] More specifically, the generation of dynamic flame sequences includes: constructing a variational autoencoder of flame morphology to learn the dynamic distribution of flames in the latent space; injecting Poisson noise and temperature gradient labels at the input of the generator of the generative adversarial network; constraining the temporal coherence of the synthesized sequence through the cycle consistency loss function; adding a dust occlusion model and rain refraction effect to simulate environmental interference; calculating the Wasserstein distance between the synthetic data and the real data, and reconstructing the generator network when the error threshold is exceeded.

[0083] The variational autoencoder can be constructed based on a convolutional neural network encoder and a deconvolutional decoder structure. The encoder outputs the mean and variance parameters in a latent space, for example, set to 128 dimensions. Poisson noise injected into the input of the generator of the generative adversarial network simulates the stochastic nature of flame flickering. The temperature gradient label can range from 100°C / m to 500°C / m, constraining the physical laws of flame morphology as it changes with temperature. The cycle consistency loss function uses a backpropagation through time algorithm to calculate the optical flow consistency error between adjacent frames, for example, with a maximum allowable inter-frame displacement difference of 5 pixels. The dust occlusion model controls particle density through a transparency parameter, and the rain refraction effect uses a ray deflection model to simulate the distortion of the flame outline by droplets. The Wasserstein distance is calculated using a gradient penalty method, with an error threshold set to 0.05. If the error exceeds this threshold, the weight parameters of the fully connected layers of the generator network are reinitialized.

[0084] In the process of synthesizing dynamic flame sequences, the encoder of a variational autoencoder first extracts the latent distribution features of flame morphology. The decoder then reconstructs the flame image by incorporating temperature gradient labels to control the correspondence between flame height and diffusion velocity. After receiving Poisson noise, the generator generates basic flame morphology through a residual network layer. The temperature gradient labels are used to adjust the feature map channel weights through a conditional batch normalization layer. A cycle consistency loss function calculates the difference in optical flow fields between three consecutive frames at the generator output, enforcing that the displacement of flame edges in adjacent frames conforms to fluid dynamics. A dust occlusion model randomly adds a semi-transparent occlusion layer during the image synthesis stage, and the refraction effect of rain is simulated through convolutional kernels that cause pixel shifts caused by droplets. The distribution difference between the synthesized data and real data is quantified using the Wasserstein distance. When the distance value exceeds a threshold, the generator's convolution kernel size is adjusted to 3×3 to enhance feature extraction. By simulating physical laws and environmental interference, this technical solution ensures temporal coherence in the generated dynamic flame sequences. It also covers a variety of complex scenarios, providing high-fidelity training data for fire recognition models, thereby improving the accuracy of fire risk probability calculations.

[0085] Through the above technical solutions, this application effectively improves the physical rationality and temporal coherence of dynamic flame sequence generation. It captures the essential characteristics of flame dynamics through a variational autoencoder, and enhances the diversity of generated data and its compliance with thermodynamic laws by combining Poisson noise and temperature gradient labels. The cycle consistency loss function solves the problem of flame morphology mutation in traditional methods, and the introduction of dust and rain interference models improves the environmental adaptability of the synthesized data. The dynamic reconstruction mechanism based on the Wasserstein distance ensures continuous optimization of the generator, providing high-fidelity training data for the fire recognition model, significantly improving the calculation accuracy of fire risk probability values ​​and the reliability of early warning systems.

[0086] The above description is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. For those skilled in the art, various modifications and variations of the present application are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A fire early warning and intelligent fire extinguishing method based on image recognition, characterized in that: The steps include: Step 1: Cover the target monitoring area with a camera network, input flame and smoke characteristic parameters, set the monitoring range, and use image preprocessing tools to remove noise and standardize the image data of the target monitoring area; Step 2: Based on the fire propagation pattern of the target monitoring area, select the heat source point as the camera installation location, and perform layout optimization to obtain the camera installation layout; Step 3: Retrieve image data of the installation layout and target monitoring area, perform real-time image analysis based on the dynamic flame sequence, extract abnormal responses in potential fire areas, calculate fire risk values ​​using pattern recognition algorithms, and generate peak fire probability data for each potential fire area; Step 4: Collect real-time fire characteristic data from the installed cameras, perform filtering and noise reduction processing to obtain a real-time data set, compare and analyze the real-time fire characteristic data of each heat source point in the real-time data set with the corresponding fire risk value, and set multi-level thresholds to generate alarm analysis results; Step 5: Based on the alarm data of the entire monitoring process, analyze the causes of false alarms, optimize the camera installation location and fire extinguishing device activation conditions, and generate fire risk reports and fire extinguishing efficiency reports.

2. The method according to claim 1, characterized in that When executing step 1, include: Step 2.1: Use a 3D point cloud scanner to collect spatial topological data of the target monitoring area and generate a point cloud density distribution map; Step 2.2: Build a 3D structural model based on the point cloud density distribution map, and import the thermal conductivity coefficient and air convection parameters of the building materials; Step 2.3: Retrieve the flame spectrum characteristics and smoke particle motion model from the feature database and map them to the thermodynamic simulation node of the 3D structure model; Step 2.4: Set the heat source monitoring conditions in the image processing tool, define the monitoring angle range and pitch angle tolerance threshold; Step 2.5: Simulate the image acquisition distortion rate under different light intensities and generate a background interference feature template library.

3. The method according to claim 2, characterized in that When executing step 2, include: Step 3.1: Use a fluid dynamics model to simulate the fire smoke diffusion path and calculate the propagation time matrix from the heat source point to the evacuation channel; Step 3.2: Identify high fire risk areas based on the propagation time matrix and mark the temperature gradient mutation points; Step 3.3: Calculate the camera field of view overlap optimization function based on the distribution of temperature gradient mutation points; Step 3.4: Use the ray tracing algorithm to verify the blind area coverage of the camera layout and mark the coordinates of the uncovered areas; Step 3.5: Add the auxiliary camera node and recalculate the coverage matrix to output the final installation layout coordinate set.

4. The method according to claim 1, wherein When executing step 3, include: Step 4.1: Load the flame morphology data from the historical fire database and construct the flame height fluctuation function and smoke turbulence model; Step 4.2: Inject the synthetic flame sequence at the heat source coordinates and superimpose the environmental interference noise simultaneously; Step 4.3: Use a multi-scale convolutional network to extract the flame edge sharpness features and smoke contour diffusion entropy; Step 4.4: Input the features into the bidirectional long short-term memory network to predict the fire spread speed and thermal radiation intensity increment; Step 4.5: Dynamically update the safety threshold based on the prediction results and mark the coordinates of high-risk areas.

5. The method according to claim 1, wherein When executing step 4, include: Step 5.1: Use millimeter-wave radar to synchronously collect temperature field distribution data and align it with the camera image data in time and space; Step 5.2: Apply wavelet packet decomposition filtering to the multi-source data to separate high-frequency noise and low-frequency characteristic components; Step 5.3: Fuse the image RGB features with the millimeter-wave temperature gradient features to generate a multimodal fire feature vector; Step 5.4: Dynamically divide the warning level according to the fire risk probability value, and generate the first-level alarm threshold and the second-level fire extinguishing activation threshold; Step 5.5: When the multimodal feature vector triggers two-level thresholds simultaneously, the positioning tracking mode is activated.

6. The method according to claim 4, characterized in that When executing step 5, include: Step 6.1: Create a spatiotemporal distribution map of false alarms and correlate it with humidity sudden change events in meteorological data. Step 6.2: Extract the texture feature similarity matrix of the false alarm area and match it with the steam source category in the interference source feature library; Step 6.3: Optimize the camera position fitness function through genetic algorithm to avoid high similarity interference areas; Step 6.4: Deploy the movable camera stand and adjust the physical position in real time according to the path planning instructions; Step 6.5: Verify the reduction in false positive rate under the new layout and output the topology optimization report.

7. The method according to claim 6, characterized in that In step 6.3 include: Step 7.1: Construct the Voronoi diagram of the heat source point distribution and calculate the area of ​​the polygon-covered holes; Step 7.2: Identify discrete heat source clusters using the DBSCAN clustering algorithm and mark the cluster boundary coordinates; Step 7.3: Deploy wide-angle camera nodes at the edge of the cluster and adjust the pitch angle to cover the hole area; Step 7.4: Use the robotic arm collaborative control system to execute camera displacement and provide real-time feedback on positioning accuracy data; Step 7.5: When the positioning accuracy continues to exceed the preset tolerance, a manual calibration protocol is triggered.

8. The method according to claim 5, characterized in that When setting the fire extinguishing activation threshold in step 5.4, include: Step 8.1: Analyze the topological correlation between flame morphology and smoke diffusion characteristics through graph convolutional networks; Step 8.2: Match the combustion reaction equation in the flame type database and calculate the oxygen demand index; Step 8.3: If it is an open flame, start the high-pressure water mist injection system and adjust the water mist particle size distribution according to the oxygen demand index; Step 8.4: If it is an electrical fire, inject perfluorohexanone gas and adjust the injection flow rate based on the flame height dynamic calculation function; Step 8.5: Monitor the temperature decay rate of the area covered by the fire extinguishing agent in real time and use feedback to adjust the injection parameters.

9. The method according to claim 1, characterized in that The pattern recognition algorithm in step 3 includes: Step 9.1: Use the Inception-v3 network to extract multi-scale texture features and generate a 256-dimensional feature vector; Step 9.2: Capture global context information through dilated spatial pyramid pooling and output a 512-dimensional feature map; Step 9.3: Concatenate the feature vector and the feature map and input them into the spatiotemporal attention module to calculate the inter-frame motion weight; Step 9.4: Generate a spatial weight mask based on the Euclidean distance between the pixel and the flame core; Step 9.5: Perform Softmax normalization on the weighted fusion features, output the fire risk probability value and mark the high-risk areas.

10. The method according to claim 4, characterized in that The generation of dynamic flame sequences includes: Step 10.1: Construct a flame morphology variational autoencoder to learn the dynamic distribution of flames in the latent space. Step 10.2: Inject Poisson noise and temperature gradient labels into the input of the generator of the GAN. Step 10.3: Constrain the temporal coherence of the synthesized sequence through a cycle consistency loss function; Step 10.4: Add dust occlusion model and rain refraction effect to simulate environmental interference; Step 10.5: Calculate the Wasserstein distance between the synthetic data and the real data, and reconstruct the generator network when it exceeds the allowable error threshold.

Images