Intelligent fire initial fire identification method and system based on AI vision

Abstract

This invention discloses a smart fire prevention and control method and system based on AI vision, belonging to the field of fire prevention and control technology. This invention achieves accurate fire identification and response through a four-element acquisition architecture, cross-scale preprocessing, three-in-one AI inference, and dynamic closed-loop decision-making. The method includes: acquiring multimodal data using visible light, mid-wave infrared, laser speckle imaging, and gas sensing; outputting a four-dimensional purification feature set after laser speckle denoising, image registration, and spatiotemporal alignment; constructing a cross-modal fusion network, a dynamic threshold inferencer, and an edge incremental learning module to adaptively allocate feature weights and generate scene-adaptive thresholds to complete the initial fire assessment; and achieving risk classification and personalized early warning through a three-level decision-making mechanism. This invention can effectively identify open flames, trace smoke, and concealed fire sources, improving the comprehensiveness of perception, the accuracy of identification, and the adaptability of the scene, enabling early detection, early warning, and early response to fires.

Description

Technical Field

[0001] This invention relates to the field of fire detection technology, specifically to a smart fire detection method and system based on AI vision for initial fire detection. Background Technology

[0002] With the deep integration of artificial intelligence, computer vision and fire protection technology, smart fire protection has become the core development direction for improving fire safety prevention and control capabilities and achieving accurate identification and efficient handling of initial fires. At present, the fire safety needs of various buildings are increasing, which puts forward higher requirements for the accuracy of initial fire identification, response speed and scenario adaptability. Therefore, there is a need for smart fire protection initial fire identification methods and systems based on AI vision.

[0003] Existing technology, such as the invention patent application with publication number CN118557930B, discloses an AI-based fire analysis-based fire equipment control method and system, belonging to the field of intelligent safety control of fire equipment. The method includes acquiring fire monitoring video streams and real-time fire monitoring data; pre-setting a standard set of fire monitoring images and a standard set of fire monitoring data, and constructing a fire monitoring grid model and a fire monitoring threshold model; passing each frame of the fire monitoring image in the video stream through the fire monitoring grid model to obtain a fire monitoring image status label; obtaining fire monitoring anomalies based on the real-time fire monitoring data through the fire monitoring threshold model; and outputting fire equipment operation commands based on the fire monitoring image status labels and the fire monitoring anomalies. By constructing a fire analysis model, the method quickly identifies fire occurrences and accurately outputs fire equipment operation commands.

[0004] Regarding the above-mentioned solutions, the applicant of this invention has discovered at least the following technical problems: 1. In the prior art, most initial fire identification solutions adopt a single-modality or simple combination acquisition method, failing to form a systematic multimodal collaborative acquisition architecture. The acquisition of visual data such as visible light, infrared, and laser speckle with gas sensing data lacks a spatiotemporal synchronization calibration mechanism, resulting in deviations in data of different types and dimensions, making it difficult to achieve effective data fusion. At the same time, in the data preprocessing stage, the denoising algorithm for laser speckle signals lacks specific optimization for fire scenarios, easily eliminating weak speckle features of hidden fire sources. The registration accuracy of infrared and visible light images is low, and the alignment of multimodal data lacks a unified coordinate system construction and verification standard, which greatly reduces the accuracy of subsequent feature extraction and fails to provide high-quality feature data support for fire identification.

[0005] 2. Existing fire detection inference structures mostly employ single network models, failing to construct a collaborative inference system integrating cross-modal feature fusion, dynamic threshold inference, and edge incremental learning. They lack adaptive adjustment capabilities for multimodal feature weight allocation, making it difficult to adapt to differences in fire characteristics across various scenarios. Dynamic threshold generation often relies on fixed scenario baseline thresholds, without dynamically adjusting based on real-time environmental parameters of the target monitoring area, resulting in poor threshold adaptability and a high risk of misjudgment and missed detection. Furthermore, most inference models are deployed in the cloud, lacking lightweight edge learning modules, making it impossible to update model parameters on-site. This hinders their ability to adapt to fire detection needs in unknown scenarios or after environmental changes, and the inference response speed cannot meet the practical requirements for rapid initial fire detection.

[0006] 3. In existing technologies, fire risk classification often relies on a single indicator or simple threshold judgment, failing to construct a quantitative classification system based on multiple core indicators. The consideration of key indicators such as fire type, rate of characteristic change, and environmental impact is insufficient, resulting in classification results that lack scientific rigor and accuracy, and cannot accurately reflect the degree of fire risk in the target monitoring area. Regarding early warning and response decisions, a uniform response strategy is often adopted, without developing personalized solutions based on the actual situation of the target monitoring area, such as scene attributes, equipment deployment, and personnel density. The response to fires of different risk levels lacks specificity, making it difficult to achieve accurate initial fire response and risk management, and failing to fully meet the refined and personalized prevention and control needs of smart fire protection in various scenarios. Summary of the Invention

[0007] In view of the above-mentioned technical deficiencies, the purpose of this invention is to provide a method and system for intelligent fire prevention and initial fire identification based on AI vision.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In the first aspect, the present invention provides a smart fire prevention and control method for initial fire identification based on AI vision, including the following steps: Step 1, multimodal perception and acquisition: using a four-element acquisition architecture composed of a visible light camera, a mid-wave infrared thermal imager, a laser speckle imaging unit and a gas sensor, multimodal data is acquired from the target monitoring area at the current moment, thereby obtaining heterogeneous data of the target monitoring area.

[0009] Step 2, Cross-scale feature preprocessing: Based on the heterogeneous data of the target monitoring area, a cross-scale feature purification algorithm is used to sequentially complete the laser speckle signal denoising of the target monitoring area, the infrared and visible light image registration of the target monitoring area, and the multimodal data alignment of the target monitoring area, outputting a four-dimensional purified feature set composed of color features, thermal radiation features, speckle features, and dynamic features of the target monitoring area.

[0010] Step 3: Innovative AI Inference Operation: Construct a three-in-one inference structure consisting of a cross-modal feature fusion network, a dynamic threshold inferencer, and an edge incremental learning module. Adaptively allocate multimodal feature weights for the target monitoring area, generate scene-adaptive judgment thresholds for the target monitoring area in real time, complete the initial fire assessment of the target monitoring area, and output the suspected fire assessment result and corresponding multimodal feature data.

[0011] Step 4: Dynamic closed-loop decision optimization: A three-level decision mechanism of feature verification, time-series tracking and multi-source verification is adopted to analyze the fire risk classification of the target monitoring area and generate personalized early warning and response decisions for the target monitoring area.

[0012] In a second aspect, the present invention provides an AI vision-based intelligent fire prevention and early fire identification system, comprising the following modules: a multimodal perception and acquisition module: which uses a four-element acquisition architecture consisting of a visible light camera, a mid-wave infrared thermal imager, a laser speckle imaging unit and a gas sensor to acquire multimodal data of the target monitoring area at the current moment, thereby obtaining heterogeneous data of the target monitoring area.

[0013] Cross-scale feature preprocessing module: Based on the heterogeneous data of the target monitoring area, it uses a cross-scale feature purification algorithm to sequentially complete the laser speckle signal denoising of the target monitoring area, infrared and visible light image registration of the target monitoring area, and multimodal data alignment of the target monitoring area, and outputs a four-dimensional purified feature set composed of color features, thermal radiation features, speckle features and dynamic features of the target monitoring area.

[0014] The innovative AI inference module is used to construct a three-in-one inference structure consisting of a cross-modal feature fusion network, a dynamic threshold inferencer, and an edge incremental learning module. It adaptively allocates multimodal feature weights for the target monitoring area, generates scene-adaptive judgment thresholds for the target monitoring area in real time, completes the initial fire assessment of the target monitoring area, and outputs the suspected fire assessment result and the corresponding multimodal feature data.

[0015] Dynamic closed-loop decision optimization module: It is used to analyze the fire risk classification of the target monitoring area by adopting a three-level decision mechanism of feature verification, time series tracking and multi-source verification, and generate personalized early warning and disposal decisions for the target monitoring area.

[0016] The beneficial effects of this invention are as follows: 1. This embodiment employs a four-element acquisition architecture consisting of a visible light camera, a mid-wave infrared thermal imager, a laser speckle imaging unit, and a gas sensor. Through a scientific equipment deployment strategy, it achieves multi-dimensional, full-scene heterogeneous data acquisition of the target monitoring area. It can acquire conventional visible light, infrared, and gas parameter data, and also accurately capture the thermal radiation speckle signals of concealed fire sources in hidden areas and non-metallic obstructed areas through the laser speckle imaging unit, thus overcoming the limitation of single-mode acquisition in covering concealed fires. Simultaneously, the four components work collaboratively with an industrial bus and edge computing unit, achieving spatiotemporal synchronization of data acquisition based on timestamp synchronization and spatial coordinate calibration. This effectively avoids deviations between different types of data, providing comprehensive, accurate, and high-quality heterogeneous data support for subsequent fire identification, significantly improving the ability to perceive initial fires, especially concealed fire sources and trace amounts of smoke.

[0017] 2. This embodiment of the solution utilizes a cross-scale feature purification algorithm to sequentially complete three preprocessing steps: laser speckle signal denoising, infrared and visible light image registration, and multimodal data alignment. This addresses the core issues of multimodal data purification and fusion. Specifically, laser speckle signal denoising employs an improved wavelet thresholding algorithm and sets a specific threshold range for fire-related speckle, effectively preserving subtle speckle features of concealed fire sources while eliminating environmental scattering and equipment noise interference. Infrared and visible light image registration significantly reduces image bias through adaptive feature point matching and homography matrix transformation. Multimodal data alignment, based on a unified coordinate system, achieves precise spatiotemporal alignment of various data types. The preprocessed four-dimensional purified feature set integrates four core features: color, thermal radiation, speckle, and dynamics. This significantly improves the accuracy and reliability of feature extraction in subsequent AI inference operations, providing high-quality feature support for initial fire assessment.

[0018] 3. This embodiment constructs a three-in-one inference structure consisting of a cross-modal feature fusion network, a dynamic threshold inferencer, and an edge incremental learning module. This overcomes the limitations of a single inference model and achieves intelligent and scene-adaptive initial fire assessment. The cross-modal feature fusion network adaptively allocates multi-modal feature weights through a multi-head attention mechanism, accurately uncovering the correlation between different features and improving the accuracy of fire feature fusion inference. The dynamic threshold inferencer, based on an improved BP neural network, combines real-time environmental parameters and scene baseline thresholds to generate scene-adaptive judgment thresholds for each frame, effectively avoiding misjudgments and missed judgments caused by fixed thresholds. The edge incremental learning module enables lightweight on-site updates of model parameters, quickly adapting to fire identification needs in unknown scenarios or after environmental changes. Deployed on edge computing units, it significantly improves inference response speed, meeting the practical needs of rapid initial fire identification and comprehensively improving the accuracy, real-time performance, and scene adaptability of initial fire assessment.

[0019] 4. This solution, through a three-tiered decision-making mechanism of feature verification, time-series tracking, and multi-source verification, accurately verifies suspected fire assessment results, effectively eliminating interfering factors and ensuring the accuracy of fire assessment. Based on this, a fire risk grading system is constructed, comprehensively considering four core indicators: fire category, detection confidence level, feature change rate, and environmental impact coefficient. Through quantitative scoring and grading standards, it accurately reflects the fire risk level of the target monitoring area, providing a scientific basis for subsequent early warning and response. Simultaneously, based on the risk grading results and combined with the actual situation of the target monitoring area, such as scene attributes, equipment deployment, and emergency exit distribution, differentiated and personalized early warning and response decisions are generated. This achieves precise handling of fires at different risk levels, covering the entire process from continuous monitoring and local early warning to emergency linkage alarms and personnel evacuation. This effectively improves the efficiency and precision of initial fire control, helping to achieve the core goal of smart fire protection: "early detection, early warning, and early response," and reducing fire safety hazards. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the implementation steps of the method of the present invention.

[0022] Figure 2 This is a schematic diagram of the system structure connection of the present invention. Detailed Implementation

[0023] 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 embodiments of the present invention, and not all embodiments. 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.

[0024] Examples of embodiments of the present invention Figure 1 As shown, the AI ​​vision-based intelligent fire prevention and control method for initial fire identification includes the following steps: Step 1, multimodal perception and acquisition: A four-element acquisition architecture consisting of a visible light camera, a mid-wave infrared thermal imager, a laser speckle imaging unit, and a gas sensor is used to acquire multimodal data of the target monitoring area at the current moment, thereby obtaining heterogeneous data of the target monitoring area.

[0025] In a specific embodiment, the four-element acquisition architecture consisting of a visible light camera, a mid-wave infrared thermal imager, a laser speckle imaging unit, and a gas sensor is configured as follows: The visible light camera, as the core visual acquisition unit, is deployed at the core location of the target monitoring area to collect visible light image data of the target monitoring area; the mid-wave infrared thermal imager is coaxially deployed with the visible light camera to synchronously collect infrared thermal imaging data and temperature field distribution data of the target monitoring area; the laser speckle imaging unit is deployed around the concealed areas and non-metallic shielding areas of the target monitoring area, emitting a 5-10mW low-power laser to collect thermal radiation speckle signals from concealed fire sources within the target monitoring area; the gas sensor is deployed at ventilation openings and flammable material storage areas within the target monitoring area to collect CO concentration data and smoke particle concentration data within the target monitoring area; the four components are connected to the edge computing unit via an industrial bus, achieving coordinated operation based on timestamp synchronization and spatial coordinate calibration, thus forming a complete four-element acquisition architecture.

[0026] It should be noted that the visible light camera is a high-definition industrial-grade camera with a resolution of no less than 1080P and a frame rate of 25fps, ensuring clear capture of conventional fire conditions in the target monitoring area. The mid-wave infrared thermal imager uses a model with a wavelength of 3-5μm, and its coaxial deployment with the visible light camera enables precise alignment of the acquisition angle. The synchronously acquired temperature field distribution data has an accuracy of ±0.5℃, which can accurately capture the weak temperature anomalies in the early stages of a fire. The 5-10mW low-power laser emitted by the laser speckle imaging unit meets safety standards and will not interfere with on-site personnel and equipment. Its deployment position precisely covers concealed areas and non-metallic obstruction areas, effectively capturing hidden areas that conventional visual equipment cannot identify. The system uses thermal radiation speckle signals from concealed fire sources to fill gaps in the collection of hidden fire information. A high-precision electrochemical gas sensor with a sampling frequency of 1Hz is used to capture subtle changes in CO and smoke particle concentrations in real time, providing early warnings of gas anomalies in the early stages of a fire. All four components communicate at high speed with an industrial bus and edge computing unit, with a communication latency of ≤50ms. Combined with timestamp synchronization and spatial coordinate calibration, the system enables coordinated operation of the four types of data acquisition units, ensuring accurate matching of heterogeneous data in time and space. This provides synchronous, complete, and high-precision multimodal data support for subsequent cross-scale feature preprocessing and AI inference, further guaranteeing the comprehensiveness and accuracy of initial fire identification.

[0027] In a specific embodiment, the acquisition of heterogeneous data of the target monitoring area is carried out as follows: RGB color images of the target monitoring area are continuously acquired using a visible light camera as the first type of heterogeneous data; infrared grayscale images and pixel temperature values ​​of the target monitoring area are acquired using a mid-wave infrared thermal imager as the second type of heterogeneous data; laser speckle imaging unit is used to acquire laser speckle time-domain signals and spatial images of the target monitoring area as the third type of heterogeneous data; CO concentration and smoke particle concentration values ​​of the target monitoring area are acquired using a gas sensor at a preset sampling frequency as the fourth type of heterogeneous data; the four types of heterogeneous data are integrated to form a heterogeneous dataset of the target monitoring area, thus completing the acquisition of heterogeneous data.

[0028] It should be noted that the visible light camera continuously captures RGB color images with a resolution of at least 1080P and a frame rate of 25fps, clearly capturing the color, shape, and distribution characteristics of open flames and smoke within the target monitoring area. Its RGB channel values ​​can be directly used for subsequent color feature extraction. The mid-wave infrared thermal imager captures infrared grayscale images with 16-bit grayscale levels, a pixel temperature detection range of -20℃ to 150℃, and an accuracy of ±0.5℃. This not only presents the temperature distribution differences in the target area but also accurately captures the initial fire situation through pixel temperature values. The subtle high-temperature anomaly provides core data for extracting thermal radiation features. The laser speckle imaging unit collects laser speckle time-domain signals with a sampling frequency set to 100Hz, and the spatial image resolution is consistent with that of the visible light camera. This allows for the accurate capture of speckle intensity and density changes caused by thermal radiation from concealed fire sources. The time-domain signal and spatial image work together to provide complete data support for speckle feature extraction. The gas sensor collects data at a preset sampling frequency of 1Hz, with CO concentration detection range of 0-500ppm and smoke particle concentration detection range of 0.01-10mg / m³. 3 The data collection accuracy is ≤±5%, which can capture subtle changes in gas composition caused by initial fires in real time and detect potential fire hazards in advance. During the collection of four types of heterogeneous data, UTC timestamps are added simultaneously to ensure data time synchronization. During integration, the data is associated and matched according to the time dimension and spatial coordinates to form a heterogeneous dataset with standardized structure, complete data, and spatiotemporal synchronization. The attributes and uses of each type of data are clearly distinguished to avoid data confusion. This provides accurate, comprehensive, and directly callable data support for subsequent cross-scale feature purification, multimodal feature fusion, and AI inference operations, ensuring the accuracy and efficiency of initial fire identification.

[0029] Step 2, Cross-scale feature preprocessing: Based on the heterogeneous data of the target monitoring area, a cross-scale feature purification algorithm is used to sequentially complete the laser speckle signal denoising of the target monitoring area, the infrared and visible light image registration of the target monitoring area, and the multimodal data alignment of the target monitoring area, outputting a four-dimensional purified feature set composed of color features, thermal radiation features, speckle features, and dynamic features of the target monitoring area.

[0030] In a specific embodiment, the process of sequentially completing laser speckle signal denoising in the target monitoring area, infrared and visible light image registration in the target monitoring area, and multimodal data alignment in the target monitoring area is as follows: First, laser speckle signal denoising is performed. A wavelet thresholding improvement algorithm is used to perform wavelet decomposition on the time-domain signal of laser speckle in the target monitoring area, set a specific threshold range for fire speckle, eliminate environmental scattering and equipment noise interference, and retain speckle feature details of concealed fire sources.

[0031] Secondly, infrared and visible light image registration is performed. Corner and edge features of the infrared and visible light images of the target monitoring area are extracted. Feature point adaptive matching algorithm is used to complete feature point matching. Image space coordinate transformation is achieved through homography matrix to complete registration and reduce image deviation.

[0032] Finally, multimodal data alignment is performed. Based on the timestamp, the denoised laser speckle signal, the registered infrared and visible light images are synchronized with the gas parameter data collected at the same time. Based on the spatial coordinates, various types of data are mapped to the unified coordinate system of the target monitoring area to complete the spatiotemporal alignment of multimodal data.

[0033] It should be noted that, firstly, laser speckle signal denoising is performed by using a wavelet thresholding improvement algorithm to decompose the time-domain signal of laser speckle in the target monitoring area into wavelet coefficients. The specific process is as follows: Wavelet basis function selection: The db4 wavelet is selected as the basis function to decompose the time-domain signal of laser speckle in the target monitoring area into 5 layers of wavelet coefficients. The first and second layers are high-frequency coefficients, corresponding to equipment noise / environmental scattering interference, and the third to fifth layers are low-frequency coefficients, corresponding to the effective speckle characteristics of concealed fire sources.

[0034] Threshold range setting: Based on the fire speckle feature library, the specific threshold range for fire speckle is set to [0.02, 0.15] (after normalization). Coefficients below 0.02 are judged as equipment electronic noise, and coefficients above 0.15 are judged as strong environmental scattering interference.

[0035] Thresholding rules: Use a soft thresholding function for high-frequency coefficients. Processing, among which, The adaptive threshold is set to 0.03. The original speckle coefficient to be purified. The low-frequency coefficients of the speckle coefficient after purification are retained without further processing.

[0036] Signal reconstruction: The processed high-frequency coefficients and low-frequency coefficients are subjected to inverse wavelet transform to reconstruct the denoised laser speckle time-domain signal. The denoising effect is verified by the signal-to-noise ratio (SNR), which requires SNR ≥ 25dB. Otherwise, the threshold range is readjusted. Finally, environmental scattering and equipment noise interference are eliminated, and the speckle feature details of the concealed fire source are preserved.

[0037] Secondly, infrared and visible light image registration is performed to extract corner and edge features from the infrared and visible light images of the target monitoring area. Feature point adaptive matching algorithm is used to complete feature point matching, and image space coordinate transformation is achieved through homography matrix. The specific process is as follows: Feature extraction stage: Corner feature extraction: An improved Shi-Tomasi corner detection algorithm is used, with a corner response value threshold of 0.05. Corner detection is performed on the infrared and visible light images of the target monitoring area, and corners with response values ​​≥0.05 are selected as candidate feature points.

[0038] Edge feature extraction: The Canny edge detection algorithm is used, with dual thresholds set to [50, 150], to extract edge contour features of the two types of images and assist in corner feature matching.

[0039] Feature point matching: Adaptive matching is performed using FLANN (Fast Nearest Neighbor Search Library) combined with RANSAC (Random Sample Consensus) algorithm. The matching distance threshold is set to 20, and incorrect matching pairs are eliminated. The correct matching rate is required to be ≥85%.

[0040] The matched feature point pairs are filtered, and ≥40 pairs of valid feature points with uniform spatial distribution are retained.

[0041] Coordinate transformation step: Based on effective feature point pairs, the homography matrix H (3×3 matrix) is solved using the least squares method, as shown in the formula. ,in, These are the pixel coordinates of the infrared image. These are the pixel coordinates of the visible light image.

[0042] The infrared image is mapped to the pixel coordinate system of the visible light image through a homography matrix. The image is resampled using bilinear interpolation. The final output is a registered infrared-visible fusion image. The registration deviation is required to be ≤2 pixels. Registration is completed to reduce image deviation.

[0043] Finally, multimodal data alignment is performed. Based on the timestamp, the denoised laser speckle signal, the registered infrared and visible light images are synchronized with the gas parameter data collected at the same time. Based on the spatial coordinates, various types of data are mapped to the unified coordinate system of the target monitoring area. The specific process is as follows: Time synchronization: All acquisition devices are equipped with a high-precision clock module to stamp each frame / group of acquired data with a UTC timestamp.

[0044] Based on the acquisition timestamp of the visible light image, the laser speckle signal and gas parameter data are interpolated and aligned to the time dimension of the image frame, with a time synchronization error of ≤100ms.

[0045] The time-synchronized data is grouped into time windows (1s), with each group containing 10 frames of laser speckle signals, 1 frame of registration image, and 1 set of gas parameters.

[0046] Spatial alignment step: Construct a unified coordinate system for the target monitoring area: with the geometric center point of the target monitoring area as the origin, the horizontal axis to the right is the X-axis, the vertical axis upward is the Y-axis, and the vertical axis outward is the Z-axis, with the unit being meters.

[0047] Based on the device calibration parameters (intrinsic parameters + extrinsic parameters), the pixel coordinates of visible light / infrared images are converted into physical coordinates in a unified coordinate system. The formula is as follows: ,in, For the camera intrinsic parameter matrix, It is an extrinsic parameter matrix. For depth value, These are pixel coordinates.

[0048] The acquisition location of the laser speckle signal (coordinates of the laser emission point + depth of the scattering point) and the installation location of the gas sensor are mapped to the same unified coordinate system.

[0049] Alignment verification step: Verify that the spatial overlap of the multimodal data after spatiotemporal alignment is ≥90% and the time synchronization accuracy is ≤100ms. Otherwise, recalibrate the device clock or calibrate the parameters to finally complete the spatiotemporal alignment of the multimodal data.

[0050] In a specific embodiment, the output of the four-dimensional purification feature set, composed of color features, thermal radiation features, speckle features, and dynamic features of the target monitoring area, is specifically processed as follows: From the registered visible light image, the RGB channel values ​​and HSV color space parameters of the target monitoring area are extracted to construct a color feature vector; from the infrared thermal imaging data, the average temperature, temperature gradient, and area ratio of high-temperature anomalies in the target monitoring area are extracted to construct a thermal radiation feature vector; from the denoised laser speckle data, the speckle density, intensity variance, and flicker frequency of the target monitoring area are extracted to construct a speckle feature vector; through the pixel motion trajectory of consecutive frames of the target monitoring area, the motion speed, motion direction, and area change rate are extracted to construct a dynamic feature vector; the color feature vector, thermal radiation feature vector, speckle feature vector, and dynamic feature vector are concatenated along the channel dimension and standardized into a feature matrix of a unified dimension to form the four-dimensional purification feature set of the target monitoring area, which is then output.

[0051] It should be noted that the RGB channel values ​​extracted from the registered visible light image range from 0 to 255, and the hue values ​​in the HSV color space range from 0 to 360°, while the saturation and brightness values ​​range from 0 to 1. Extraction focuses on the suspected fire area within the target monitoring region, while also considering background contrast. The constructed color feature vector is 6-dimensional (3D RGB + 3D HSV), accurately capturing the color differences between open flames and trace amounts of smoke. When extracting the average temperature, temperature gradient, and area ratio of high-temperature anomalies from infrared thermal imaging data, the threshold for high-temperature anomalies is set to 15°C above the ambient temperature. The temperature gradient is calculated using the temperature difference between adjacent pixels, and the area ratio is calculated as the ratio of the number of pixels in the high-temperature anomaly area to the total number of pixels in the target monitoring region. The constructed thermal radiation feature vector is 3-dimensional, accurately representing the thermal radiation intensity and distribution characteristics of the initial fire. From the denoised laser... When extracting speckle density, intensity variance, and flicker frequency from speckle data, speckle density is calculated based on the number of specks per unit area, intensity variance reflects the dispersion of speckle grayscale values, and flicker frequency is calculated based on the period of speckle intensity change. The constructed speckle feature vector is 3-dimensional, focusing on capturing the subtle speckle change characteristics of concealed fire sources. The pixel motion trajectory of continuous frame images of the target monitoring area is analyzed by optical flow and frame difference methods. The extracted motion speed is expressed in pixels / frame, and the motion direction is represented by angles from 0 to 360°. The area change rate is calculated as the ratio of the area difference of the suspected fire area in adjacent frames to the area of ​​the previous frame. The constructed dynamic feature vector is 3-dimensional, which can accurately capture the dynamic change trend of the fire. After the four types of feature vectors are constructed, they are all normalized to the [0,1] interval using the Min-Max normalization method to eliminate the differences in the dimensions of different features. Then, they are spliced ​​according to the channel dimension to form a 15-dimensional (6+3+3+3) feature matrix.

[0052] Step 3: Innovative AI Inference Operation: Construct a three-in-one inference structure consisting of a cross-modal feature fusion network, a dynamic threshold inferencer, and an edge incremental learning module. Adaptively allocate multimodal feature weights for the target monitoring area, generate scene-adaptive judgment thresholds for the target monitoring area in real time, complete the initial fire assessment of the target monitoring area, and output the suspected fire assessment result and corresponding multimodal feature data.

[0053] In a specific embodiment, the construction of the three-in-one inference structure—comprising a cross-modal feature fusion network, a dynamic threshold inferencer, and an edge incremental learning module—is carried out as follows: A cross-modal feature fusion network is built, with an attention-weighted fusion Transformer at its core. Four dedicated feature input branches are set: color feature vector, thermal radiation feature vector, speckle feature vector, and dynamic feature vector. Multi-head attention adaptively allocates the multi-modal feature weights of the target monitoring area, and the fused features are output through the feature fusion layer. A dynamic threshold inferencer is built, with an improved BP neural network at its core. The input layer receives environmental parameters of the target monitoring area, the hidden layer performs feature mapping, and the output layer outputs the relevant threshold for fire assessment. An edge incremental learning module is built, based on a pruned AWF-Transformer network. The core parameters of the feature weight layer and the threshold generator are retained, and a lightweight fine-tuning algorithm is designed to achieve on-site updates of model parameters. All three are deployed in the same edge computing unit, establishing the correlation between the output of the fusion network and the input of the dynamic threshold inferencer, as well as the correlation between the decision feedback and the input of the edge incremental learning module, thus forming the three-in-one inference structure.

[0054] In a specific embodiment, the output of the suspected fire determination result and the corresponding multimodal feature data value is specifically output as follows: An environmental parameter feature library for each common scenario is pre-constructed, storing the mapping relationship between light intensity, humidity, temperature, and interference source type and corresponding benchmark thresholds for each common scenario; light intensity, ambient temperature, humidity, and smoke concentration of the target monitoring area are collected in real time using a four-element acquisition architecture to identify and match the scenario type of the target monitoring area; the matched scenario benchmark threshold and real-time environmental parameters are input into a dynamic threshold inference engine, and calculations are performed using an improved BP neural network, outputting the fire confidence threshold, speckle feature threshold, and temperature feature threshold of the target monitoring area for each frame.

[0055] It should be noted that the RGB channel values ​​extracted from the registered visible light image range from 0 to 255, and the hue values ​​in the HSV color space range from 0 to 360°, while the saturation and brightness values ​​range from 0 to 1. Extraction focuses on the suspected fire area within the target monitoring region, while also considering background contrast. The constructed color feature vector is 6-dimensional (3D RGB + 3D HSV), accurately capturing the color differences between open flames and trace amounts of smoke. When extracting the average temperature, temperature gradient, and area ratio of high-temperature anomalies from the infrared thermal imaging data, the threshold for high-temperature anomalies is set to 15°C above the ambient temperature. The temperature gradient is calculated using the temperature difference between adjacent pixels, and the area ratio is calculated as the ratio of the number of pixels in the high-temperature anomaly area to the total number of pixels in the target monitoring region. The constructed thermal radiation feature vector is 3-dimensional, accurately characterizing the thermal radiation intensity and distribution characteristics of the initial fire. From the denoised laser speckle data... When extracting speckle density, intensity variance, and scintillation frequency, speckle density is calculated based on the number of specks per unit area, intensity variance reflects the dispersion of speckle grayscale values, and scintillation frequency is calculated based on the period of speckle intensity change. The constructed speckle feature vector is 3-dimensional, focusing on capturing the subtle speckle change characteristics of concealed fire sources. The pixel motion trajectory of continuous frame images of the target monitoring area is analyzed by optical flow and frame difference methods. The extracted motion speed is expressed in pixels / frame, and the motion direction is represented by angles from 0 to 360°. The area change rate is calculated as the ratio of the area difference of the suspected fire area in adjacent frames to the area of ​​the previous frame. The constructed dynamic feature vector is 3-dimensional, which can accurately capture the dynamic change trend of fire (open flame expansion, smoke rise). After the four types of feature vectors are constructed, they are all normalized to the [0,1] interval using the Min-Max normalization method to eliminate the differences in different feature dimensions, and then spliced ​​according to the channel dimension to form a 15-dimensional (6+3+3+3) feature matrix.

[0056] The improved BP neural network is used for computation. Each frame outputs the fire confidence threshold, speckle feature threshold, and temperature feature threshold for the target monitoring area. The specific computation process is as follows: Improved BP neural network architecture design: Network layers: A 4-layer fully connected architecture of "input layer - hidden layer 1 - hidden layer 2 - output layer" is adopted, where: Input layer: 8-dimensional, the input parameters are the real-time environmental feature vector of the target monitoring area, including light intensity (normalized to [0,1]), ambient temperature (°C, normalized to [0,1]), relative humidity (%, normalized to [0,1]), wind speed (m / s, normalized to [0,1]), and smoke background concentration (mg / m³). 3 (Normalized to [0,1]), device runtime (h, normalized to [0,1]), scene type code (1-5 correspond to warehouse / shopping mall / substation / factory / other respectively), and historical misjudgment rate (%, normalized to [0,1]).

[0057] Hidden layer 1: 64 neurons, ReLU activation function, and a Dropout layer (dropout rate 0.2) is added to prevent overfitting.

[0058] Hidden layer 2: There are 32 neurons, the activation function is ReLU, and L2 regularization (coefficient 0.001) is added as a constraint parameter.

[0059] Output layer: 3-dimensional, corresponding to fire confidence threshold, speckle feature threshold, and temperature feature threshold respectively. The activation function is Sigmoid, and the output value is normalized to the range [0.5, 0.95].

[0060] The core improvement is to introduce an adaptive learning rate and momentum term into the backpropagation stage of the traditional BP neural network. The initial learning rate is set to 0.01 and dynamically adjusted according to the convergence speed of the loss function (multiplied by 1.05 when convergence is slow and multiplied by 0.9 when oscillation occurs). The momentum term is set to 0.9 to accelerate weight updates and avoid local optima.

[0061] Network training process: Training dataset construction: Collect 100,000 sets of environmental feature-threshold label data for different scenarios (sunny / low light / night / smoke interference / high humidity) and different fire stages (no fire / trace smoke / open flame / hidden fire source) in the target monitoring area, and divide them into training set, validation set and test set in a 7:2:1 ratio.

[0062] Loss function: The mean squared error (MSE) loss function is used, and the formula is as follows: ,in, The total number of training samples used in the loss calculation. For each training sample index, iterate through the samples from 1 to... , For the true threshold label, This is the network prediction value.

[0063] Training termination conditions: The validation set loss does not decrease for 10 consecutive rounds or the number of training rounds reaches 200 rounds. After training is completed, the optimal model parameters are saved. The threshold prediction error of the model on the test set is required to be ≤0.02.

[0064] Real-time computation and threshold output: Input data preprocessing: Normalize the 8-dimensional environmental feature vectors collected in each frame of the target monitoring area to eliminate dimensional differences.

[0065] Forward propagation operation: The preprocessed feature vector is input into the improved BP neural network, and then passes through the feature mapping of hidden layer 1 and hidden layer 2 in sequence. The output layer outputs a 3D threshold prediction value.

[0066] Post-threshold processing: Fire confidence threshold: The output value range is constrained to [0.7, 0.95], with a step size of 0.01 (e.g., 0.85, 0.86), which is used to determine whether the confidence of the fire identification result meets the standard.

[0067] Speckle feature threshold: The output value range is constrained to [0.5, 0.8], with a step size of 0.01, which corresponds to the effective judgment threshold of laser speckle features.

[0068] Temperature feature threshold: The output value range is constrained to [0.6, 0.9], with a step size of 0.01, which corresponds to the effective judgment threshold of infrared thermal radiation features.

[0069] Output rules: The computation time per frame (25fps) is ≤40ms. The output threshold is synchronized to the judgment module of the inference structure in real time. If a single frame computation times out, the threshold of the previous frame is used. If three consecutive frames time out, the model reset mechanism is triggered.

[0070] Threshold dynamic calibration: Calculate the threshold deviation rate hourly based on the fire situation assessment results (number of false alarms / missed alarms) of the target monitoring area, using the following formula: The deviation rate is the actual judgment result and the ideal judgment result.

[0071] If the deviation rate is ≥5%, an incremental learning approach is adopted, using the latest 200 sets of valid samples to fine-tune the network parameters and update the threshold output model to ensure that the threshold always adapts to the environmental changes in the target monitoring area.

[0072] Based on the three-in-one reasoning structure, the fire confidence level, speckle characteristics, and temperature characteristics of the target monitoring area are obtained. These characteristics are then compared with the corresponding thresholds for fire confidence level, speckle characteristics, and temperature characteristics. If all three are greater than or equal to the threshold, a suspected fire is identified in the target monitoring area, and a preliminary distinction is made between open flames, trace smoke, and concealed fire sources. Simultaneously, multimodal feature data corresponding one-to-one with the location and category of the suspected fire is extracted. If any one of these three characteristics is less than the threshold, no suspected fire is identified in the target monitoring area. Finally, based on the determination result, a suspected fire determination result including the presence or absence of a suspected fire, the preliminary category, and the detection confidence level, as well as the multimodal feature data of the target monitoring area precisely corresponding to the determination result, is output.

[0073] Step 4: Dynamic closed-loop decision optimization: A three-level decision mechanism of feature verification, time-series tracking and multi-source verification is adopted to analyze the fire risk classification of the target monitoring area and generate personalized early warning and response decisions for the target monitoring area.

[0074] In a specific embodiment, the analysis of the fire risk classification of the target monitoring area is carried out as follows: If it is determined that there is a suspected fire in the target monitoring area, the core fire indicators of the target monitoring area are extracted, including fire type, detection confidence level, characteristic change rate and environmental impact coefficient. The four core indicators are quantified into scores of 0-100 according to a unified standard. Among them, the fire type is assigned 100 points for open flame, 80 points for concealed fire source and 50 points for trace smoke. The detection confidence level is converted into a score by confidence level × 100. The characteristic change rate is quantified into a score by actual rate / preset danger threshold × 100. The environmental impact coefficient is obtained by weighted summation of three sub-indicators: quantity of flammable materials, personnel density and ventilation conditions. The weights of the three sub-indicators are 0.4, 0.3 and 0.3, respectively.

[0075] A linear weighted risk score is constructed, with the basic weights of the scenario set as fire category 0.4, detection confidence 0.1, feature change rate 0.3, and environmental impact coefficient 0.2. The total fire risk score of the target monitoring area is obtained and rounded to the nearest integer. The risk levels are divided according to the total risk score: 0-20 is Level I low risk, 21-40 is Level II low to medium risk, 41-60 is Level III medium to high risk, 61-80 is Level IV high risk, and 81-100 is Level V extremely high risk.

[0076] In a specific embodiment, the process of generating personalized early warning and response decisions for the target monitoring area is as follows: Based on the final result of the fire risk classification of the target monitoring area, combined with the scene attributes, equipment deployment, and emergency access distribution of the target monitoring area, differentiated response strategies are formulated.

[0077] For Level I low risk, decisions are generated for continuous monitoring and without the need for early warning, and the quaternary acquisition architecture is controlled to increase the sampling frequency.

[0078] For Level II low-to-medium risk, a decision is made to generate a local audible and visual warning and push notification information to notify on-site security personnel for verification.

[0079] For Level III medium-to-high risk, decisions are made to generate regional early warnings, activate smoke extraction systems, and close fire doors, and to coordinate with on-site fire-fighting equipment.

[0080] For Level IV high risk, decisions are made to generate a comprehensive early warning, activate the sprinkler system, and organize personnel evacuation, while simultaneously pushing alarm information to the fire control center.

[0081] For Level V extremely high risks, decisions are generated to initiate emergency response, automatically dial 119 to report the emergency, and activate emergency broadcasts. The entire handling process is recorded and synchronized to the cloud, completing the generation of personalized early warning and handling decisions for the target monitoring area.

[0082] Examples of embodiments of the present invention Figure 2As shown, the AI ​​vision-based intelligent fire prevention and early fire identification system includes the following modules: Multimodal perception and acquisition module: It uses a four-element acquisition architecture consisting of a visible light camera, a mid-wave infrared thermal imager, a laser speckle imaging unit, and a gas sensor to collect multimodal data of the target monitoring area at the current moment, thereby obtaining heterogeneous data of the target monitoring area.

[0083] Cross-scale feature preprocessing module: Based on the heterogeneous data of the target monitoring area, it uses a cross-scale feature purification algorithm to sequentially complete the laser speckle signal denoising of the target monitoring area, infrared and visible light image registration of the target monitoring area, and multimodal data alignment of the target monitoring area, and outputs a four-dimensional purified feature set composed of color features, thermal radiation features, speckle features and dynamic features of the target monitoring area.

[0084] The innovative AI inference module is used to construct a three-in-one inference structure consisting of a cross-modal feature fusion network, a dynamic threshold inferencer, and an edge incremental learning module. It adaptively allocates multimodal feature weights for the target monitoring area, generates scene-adaptive judgment thresholds for the target monitoring area in real time, completes the initial fire assessment of the target monitoring area, and outputs the suspected fire assessment result and the corresponding multimodal feature data.

[0085] Dynamic closed-loop decision optimization module: It is used to analyze the fire risk classification of the target monitoring area by adopting a three-level decision mechanism of feature verification, time series tracking and multi-source verification, and generate personalized early warning and disposal decisions for the target monitoring area.

[0086] The examples described in this invention are not limited to the specific embodiments listed above. The examples are merely illustrative to facilitate understanding of the invention and do not constitute a limitation on the scope of protection of this invention. Any modifications, equivalent substitutions, etc., made within the spirit and principles of this invention should be included within the scope of protection.

[0087] The above description is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in this specification, they should all fall within the protection scope of the present invention.

Claims

1. A smart fire prevention method for initial fire identification based on AI vision, characterized in that, Includes the following steps: Step 1, Multimodal Sensing and Acquisition: A four-element acquisition architecture consisting of a visible light camera, a mid-wave infrared thermal imager, a laser speckle imaging unit, and a gas sensor is used to acquire multimodal data of the target monitoring area at the current moment, thereby obtaining heterogeneous data of the target monitoring area; Step 2, Cross-scale feature preprocessing: Based on the heterogeneous data of the target monitoring area, a cross-scale feature purification algorithm is used to sequentially complete the laser speckle signal denoising of the target monitoring area, the infrared and visible light image registration of the target monitoring area, and the multimodal data alignment of the target monitoring area, and output a four-dimensional purified feature set composed of color features, thermal radiation features, speckle features and dynamic features of the target monitoring area; Step 3: Innovative AI Inference Operation: Construct a three-in-one inference structure consisting of a cross-modal feature fusion network, a dynamic threshold inferencer, and an edge incremental learning module. Adaptively allocate multimodal feature weights for the target monitoring area, generate scene-adaptive judgment thresholds for the target monitoring area in real time, complete the initial fire assessment of the target monitoring area, and output the suspected fire assessment result and corresponding multimodal feature data. Step 4: Dynamic closed-loop decision optimization: A three-level decision mechanism of feature verification, time-series tracking and multi-source verification is adopted to analyze the fire risk classification of the target monitoring area and generate personalized early warning and response decisions for the target monitoring area.

2. The AI ​​vision-based intelligent fire prevention method for initial fire identification according to claim 1, characterized in that, The four-element acquisition architecture, consisting of a visible light camera, a mid-wave infrared thermal imager, a laser speckle imaging unit, and a gas sensor, is composed of the following components: A visible light camera is deployed as the core visual acquisition unit in the core location of the target monitoring area to collect visible light image data of the target monitoring area; a mid-wave infrared thermal imager is coaxially deployed with the visible light camera to synchronously collect infrared thermal imaging data and temperature field distribution data of the target monitoring area; a laser speckle imaging unit is deployed around the concealed areas and non-metallic shielding areas of the target monitoring area, emitting a 5-10mW low-power laser to collect thermal radiation speckle signals of concealed fire sources in the target monitoring area; gas sensors are deployed at ventilation openings and flammable material storage areas in the target monitoring area to collect CO concentration data and smoke particle concentration data in the target monitoring area; the four units are connected to the edge computing unit through an industrial bus, and achieve coordinated linkage based on timestamp synchronization and spatial coordinate calibration, forming a complete four-element acquisition architecture.

3. The AI ​​vision-based intelligent fire prevention method for initial fire identification according to claim 2, characterized in that, The specific process for acquiring heterogeneous data from the target monitoring area is as follows: The target monitoring area is continuously acquired by a visible light camera as the first type of heterogeneous data; the infrared grayscale image and pixel temperature value of the target monitoring area are acquired by a mid-wave infrared thermal imager as the second type of heterogeneous data. The laser speckle imaging unit acquires the time-domain signal and spatial image of the target monitoring area as the third type of heterogeneous data; the gas sensor acquires the CO concentration and smoke particle concentration values ​​of the target monitoring area at a preset sampling frequency as the fourth type of heterogeneous data. The four types of heterogeneous data are integrated to form a heterogeneous dataset of the target monitoring area, thus completing the acquisition of heterogeneous data.

4. The AI ​​vision-based intelligent fire prevention method for initial fire identification according to claim 3, characterized in that, The process of sequentially performing laser speckle signal denoising in the target monitoring area, infrared and visible light image registration in the target monitoring area, and multimodal data alignment in the target monitoring area is as follows: First, laser speckle signal denoising is performed. A wavelet threshold improvement algorithm is used to decompose the time-domain signal of laser speckle in the target monitoring area into wavelets, set a special threshold range for fire speckle, eliminate environmental scattering and equipment noise interference, and retain speckle feature details of hidden fire sources. Secondly, infrared and visible light image registration is performed. Corner and edge features of infrared and visible light images of the target monitoring area are extracted. Feature point adaptive matching algorithm is used to complete feature point matching. Image space coordinate transformation is achieved through homography matrix to complete registration and reduce image deviation. Finally, multimodal data alignment is performed. Based on the timestamp, the denoised laser speckle signal, the registered infrared and visible light images are synchronized with the gas parameter data collected at the same time. Based on the spatial coordinates, various types of data are mapped to the unified coordinate system of the target monitoring area to complete the spatiotemporal alignment of multimodal data.

5. The AI ​​vision-based intelligent fire prevention method for initial fire identification according to claim 4, characterized in that, The output target monitoring area comprises a four-dimensional purification feature set, consisting of color features, thermal radiation features, speckle features, and dynamic features. The specific output process is as follows: From the registered visible light image, the RGB channel values ​​and HSV color space parameters of the target monitoring area are extracted to construct a color feature vector. From the infrared thermal imaging data, the mean temperature, temperature gradient, and area ratio of high-temperature anomalies in the target monitoring area are extracted to construct a thermal radiation feature vector. From the denoised laser speckle data, the speckle density, intensity variance, and flicker frequency of the target monitoring area are extracted to construct a speckle feature vector. Through the pixel motion trajectory of continuous frame images of the target monitoring area, the motion speed, motion direction, and area change rate are extracted to construct a dynamic feature vector. The color feature vector, thermal radiation feature vector, speckle feature vector, and dynamic feature vector are concatenated according to the channel dimension and standardized into a feature matrix of a unified dimension to form a four-dimensional purification feature set of the target monitoring area and output it.

6. The AI ​​vision-based intelligent fire prevention method for initial fire identification according to claim 5, characterized in that, The construction process of the three-in-one inference structure, which consists of a cross-modal feature fusion network, a dynamic threshold inferencer, and an edge incremental learning module, is as follows: A cross-modal feature fusion network was constructed, with an attention-weighted fusion Transformer as the core. Four dedicated feature input branches were set up, including color feature vector, thermal radiation feature vector, speckle feature vector, and dynamic feature vector. The multi-head attention mechanism was used to adaptively allocate the multi-modal feature weights of the target monitoring area, and the fused features were output through the feature fusion layer. A dynamic threshold inference engine is constructed, with an improved BP neural network as its core. The input layer receives environmental parameters of the target monitoring area, the hidden layer performs feature mapping, and the output layer outputs the threshold related to fire determination. An edge incremental learning module is also constructed, based on a pruned AWF-Transformer network. The core parameters of the feature weight layer and threshold generator are retained, and a lightweight fine-tuning algorithm is designed to achieve on-site updates of model parameters. All three components are deployed in the same edge computing unit, establishing a fusion relationship between the network output and the input of the dynamic threshold inference engine, as well as a relationship between the decision feedback and the input of the edge incremental learning module, forming a three-in-one inference structure.

7. The AI ​​vision-based intelligent fire prevention method for initial fire identification according to claim 6, characterized in that, The output process for determining the suspected fire situation and the corresponding multimodal feature data values ​​is as follows: A pre-built environmental parameter feature library for each common scenario stores the mapping relationship between light intensity, humidity, temperature, and interference source type and corresponding benchmark thresholds for each common scenario. The light intensity, ambient temperature, humidity, and smoke concentration of the target monitoring area are collected in real time through a four-element acquisition architecture to identify and match the scenario type of the target monitoring area. The matched scenario benchmark threshold and real-time environmental parameters are input into a dynamic threshold inference engine, which is then processed by an improved BP neural network. Each frame outputs the fire confidence threshold, speckle feature threshold, and temperature feature threshold of the target monitoring area. Based on the three-in-one reasoning structure, the fire confidence level, speckle characteristics, and temperature characteristics of the target monitoring area are obtained. These characteristics are then compared with the corresponding thresholds for fire confidence level, speckle characteristics, and temperature characteristics. If all three are greater than or equal to the threshold, a suspected fire is identified in the target monitoring area, and a preliminary distinction is made between open flames, trace smoke, and concealed fire sources. Simultaneously, multimodal feature data corresponding one-to-one with the location and category of the suspected fire is extracted. If any one of these three characteristics is less than the threshold, no suspected fire is identified in the target monitoring area. Finally, based on the determination result, a suspected fire determination result including the presence or absence of a suspected fire, the preliminary category, and the detection confidence level, as well as the multimodal feature data of the target monitoring area precisely corresponding to the determination result, is output.

8. The AI ​​vision-based intelligent fire prevention method for initial fire identification according to claim 7, characterized in that, The analysis of the fire risk classification of the target monitoring area is carried out in the following specific process: If a suspected fire is detected in the target monitoring area, the core fire indicators for the target monitoring area are extracted, including fire type, detection confidence level, characteristic change rate, and environmental impact coefficient. The four core indicators are quantified into scores of 0-100 according to a unified standard. Specifically, the fire type is assigned 100 points for open flame, 80 points for concealed fire source, and 50 points for trace smoke. The detection confidence level is converted into a score by multiplying the confidence level by 100. The characteristic change rate is quantified into a score by multiplying the actual rate by the preset danger threshold by 100. The environmental impact coefficient is obtained by weighted summation of three sub-indicators: quantity of flammable materials, personnel density, and ventilation conditions. The weights of the three sub-indicators are 0.4, 0.3, and 0.3, respectively. A linear weighted risk score is constructed, with the basic weights of the scenario set as fire category 0.4, detection confidence 0.1, feature change rate 0.3, and environmental impact coefficient 0.

2. The total fire risk score of the target monitoring area is obtained and rounded to the nearest integer. The risk levels are divided according to the total risk score: 0-20 is Level I low risk, 21-40 is Level II low to medium risk, 41-60 is Level III medium to high risk, 61-80 is Level IV high risk, and 81-100 is Level V extremely high risk.

9. The AI ​​vision-based intelligent fire prevention method for initial fire identification according to claim 8, characterized in that, The specific process for generating personalized early warning and response decisions for the target monitoring area is as follows: Based on the final results of the fire risk classification of the target monitoring area, and in combination with the scene attributes, equipment deployment, and emergency access distribution of the target monitoring area, differentiated response strategies are formulated. For Level I low risk, decisions are generated for continuous monitoring and without the need for early warning, and the quaternary acquisition architecture is controlled to increase the sampling frequency; For Level II low to medium risk, a decision is made to generate a local audible and visual warning and push notification information to notify on-site security personnel for verification; For Level III medium-to-high risk, decisions are made to generate regional early warnings, activate smoke extraction systems, and close fire doors, and to coordinate with on-site fire-fighting equipment. For Level IV high risk, generate a comprehensive early warning, activate the sprinkler system and organize personnel evacuation, and simultaneously push alarm information to the fire control center; For Level V extremely high risks, decisions are generated to initiate emergency response, automatically dial 119 to report the emergency, and activate emergency broadcasts. The entire handling process is recorded and synchronized to the cloud, completing the generation of personalized early warning and handling decisions for the target monitoring area.

10. A system utilizing the AI ​​vision-based intelligent fire initial fire identification method according to any one of claims 1-9, characterized in that, Includes the following modules: Multimodal sensing and acquisition module: It is used to acquire multimodal data of the target monitoring area at the current moment by using a four-element acquisition architecture consisting of a visible light camera, a mid-wave infrared thermal imager, a laser speckle imaging unit and a gas sensor, thereby obtaining heterogeneous data of the target monitoring area; Cross-scale feature preprocessing module: Based on the heterogeneous data of the target monitoring area, it uses a cross-scale feature purification algorithm to sequentially complete the laser speckle signal denoising of the target monitoring area, infrared and visible light image registration of the target monitoring area, and multimodal data alignment of the target monitoring area, and outputs a four-dimensional purified feature set composed of color features, thermal radiation features, speckle features and dynamic features of the target monitoring area; Innovative AI inference module: used to build a three-in-one inference structure of cross-modal feature fusion network, dynamic threshold inferencer and edge incremental learning module, adaptively allocate multimodal feature weights of target monitoring area, generate scene-adaptive judgment thresholds of target monitoring area in real time, complete the initial judgment of fire in target monitoring area, and output the suspected fire judgment result and corresponding multimodal feature data. Dynamic closed-loop decision optimization module: It is used to analyze the fire risk classification of the target monitoring area by adopting a three-level decision mechanism of feature verification, time series tracking and multi-source verification, and generate personalized early warning and disposal decisions for the target monitoring area.

Images