Fire safety early warning system based on artificial intelligence video recognition

Abstract

The invention discloses a fire safety early warning system based on artificial intelligence video recognition, and relates to the technical field of intelligent fire protection and computer vision crossing. The smoldering stage recognition system comprises a smoldering stage recognition unit, a smoldering stage recognition unit, a smoldering stage recognition unit, a smoldering stage recognition unit and a smoldering stage recognition unit, the smoldering stage recognition unit comprises a smoldering stage recognition unit, a smoldering stage recognition unit, a smoldering stage recognition unit and a smoldering stage recognition unit, the smoldering stage recognition unit comprises a smoldering stage recognition unit and a smoldering stage recognition unit, the smoldering stage recognition unit comprises a smoldering stage recognition unit and a smoldering stage recognition unit, the smoldering stage recognition unit comprises a smoldering stage recognition unit, the smoldering stage recognition unit comprises a smoldering stage recognition unit and a smoldering stage recognition unit, the smoldering stage recognition unit comprises a smoldering stage recognition unit, the smoldering stage Compared with a traditional threshold value method (larger than or equal to 0.5 mg / m), early warning is carried out 7-12 minutes ahead of time, and the early warning speed is increased.

Description

Technical Field

[0001] The present invention relates to the technical field of intersection of smart firefighting and computer vision, and specifically to a fire safety early warning system based on artificial intelligence video recognition. Background Art

[0002] Existing safety warning systems have certain defects: Single sensor limitations: Traditional smoke / temperature sensing systems have a response delay of 20-30 seconds and a false alarm rate of up to 15% (such as accidental contact with kitchen steam), and video analysis technology bottlenecks: Public patent CN117152906A is an artificial intelligence-based video image fire alarm system, including a control host. The present invention can monitor the interior of a building by setting up floor monitoring on each floor of the building. The video information collected by the floor monitoring is saved and backed up by the storage end and at the same time, it is transmitted to the industrial computer. The fire identification module set up inside the industrial computer uses video processing, scene recognition, smoke recognition and flame recognition to identify the fire location on the floor, and then makes a fire judgment. Compared with fire detection equipment such as fire detectors, the fire identification module can make judgments based on the scene where the flame is located, the smoke generated and the burning condition of the flame. The fire detection has high accuracy and is not prone to misjudgment. By setting up a fire alarm platform, fire information sharing is established between multiple buildings, users and residents of the corresponding buildings, and fire departments, and static image recognition is used to reduce fire losses. The missed reporting rate is greater than 8%. The existing video system lacks dynamic behavior analysis and cannot identify early fire conditions (such as smoke spread in the unopened fire stage), and the linkage mechanism is insufficient. Summary of the Invention

[0003] In response to the shortcomings of the existing technology, the present invention provides a fire safety early warning system based on artificial intelligence video recognition, which solves the problems raised by the background technology.

[0004] To achieve the above objectives, the present invention is implemented through the following technical solutions: a fire safety early warning system based on artificial intelligence video recognition, comprising: Multispectral sensing module: a heterogeneous acquisition array integrating visible light / infrared / TOF sensors; Dynamic analysis engine: uses an improved spatiotemporal joint modeling method of YOLOv7-ASFF and ST-GCN; 3D positioning unit: A fusion positioning system based on an improved SfM algorithm and UWB (error < 0.5m).

[0005] Preferably, the dynamic analysis engine comprises: Flame feature extraction layer: constructs the stroboscopic feature vector (5-15Hz) and edge entropy matrix; Smoke diffusion model: The improved optical flow method (ICOF) is used to calculate the concentration gradient field.

[0006] The hierarchical warning strategy includes: Four-level response mechanism: automatically triggers different linkage strategies according to the fire development stage; Digital twin verification: Verify the feasibility of the emergency plan through FDS simulation (error <15%).

[0007] Preferably, the heterogeneous acquisition array includes: Multimodal sensor suite; including a visible light camera (supporting H.266 / VVC encoding), a dual-band infrared thermal imager (7.5-13μm wavelength range), and a TOF ranging module (accuracy ±2cm). Each sensor is independently equipped with an industrial-grade protective housing (IP67 protection rating). Dynamic protocol parsing module; built-in Modbus TCP, KNX, DALI three protocol stacks, supports automatic identification of RS-485 / RS-232 / CAN bus data frame structure, abnormal data packet retransmission times ≤ 3 times; Edge computing node; uses the NVIDIA Jetson AGX Orin hardware platform, integrated with a multi-threaded data stream processing unit (supporting 4-channel 8K video parallel decoding).

[0008] Preferably, the visible light camera array comprises: Main camera: 8K resolution (7680×4320 pixels), frame rate 25fps, support AI stabilization (compensation accuracy 0.05° / s); Auxiliary camera: fisheye lens (FOV ≥ 180°), built-in distortion correction algorithm (polynomial fitting order ≤ 5).

[0009] Preferably, the array data acquisition method comprises: Heterogeneous data preprocessing: Time series data (such as temperature and humidity) is normalized using a sliding window (window length = 3s) Three-level noise reduction is implemented for video data: spatial domain wavelet transform (Daubechies-4 basis function) + frequency domain adaptive filtering (cutoff frequency ≥ 15 Hz).

[0010] Preferably, the multi-source data fusion adopts an improved Kalman filter, and the state equation is: ; Where Rk is the covariance matrix of heterogeneous data, including: Video data: ; Environmental data: .

[0011] The present invention provides a fire safety early warning system based on artificial intelligence video recognition. Compared with the existing technology, it has the following advantages: 1. This fire safety early warning system based on artificial intelligence video recognition uses an improved YOLOv7-ASFF model (mAP@0.5=98.7%) to identify smoke in the smoldering stage (smoke concentration 0.1-0.5mg / m³). Compared with the traditional threshold method (requires ≥0.5mg / m³), it can provide early warning 7-12 minutes earlier, speeding up the warning process.

[0012] 2. This fire safety warning system, based on AI video recognition, features false alarm rate control: a multi-protocol adaptive parsing algorithm reduces the false alarm rate to 0.12 times per thousand hours. Edge computing efficiency: the Jetson AGX Orin platform enables parallel processing of four channels of 8K video (21.3fps), a 144% improvement over traditional solutions (8.7fps). Knowledge graph application: the emergency plan matching accuracy rate is 98% (BERT+GraphSAGE model), a 50% improvement over manual decision-making (65%). Automatic mode switching is supported for 16 typical scenarios. DETAILED DESCRIPTION

[0013] The following is a clear and complete description of the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0014] The present invention provides a technical solution: a fire safety early warning system based on artificial intelligence video recognition, comprising: Multispectral sensing module: a heterogeneous acquisition array integrating visible light / infrared / TOF sensors; Dynamic analysis engine: uses an improved spatiotemporal joint modeling method of YOLOv7-ASFF and ST-GCN; 3D positioning unit: A fusion positioning system based on an improved SfM algorithm and UWB (error < 0.5m).

[0015] The dynamic analysis engine includes: Flame feature extraction layer: constructs the stroboscopic feature vector (5-15Hz) and edge entropy matrix; Smoke diffusion model: The improved optical flow method (ICOF) is used to calculate the concentration gradient field.

[0016] The hierarchical warning strategy includes: Four-level response mechanism: automatically triggers different linkage strategies according to the fire development stage; Digital twin verification: Verify the feasibility of the emergency plan through FDS simulation (error <15%).

[0017] Heterogeneous acquisition arrays include: Multimodal sensor suite; including a visible light camera (supporting H.266 / VVC encoding), a dual-band infrared thermal imager (7.5-13μm wavelength range), and a TOF ranging module (accuracy ±2cm). Each sensor is independently equipped with an industrial-grade protective housing (IP67 protection rating). Dynamic protocol parsing module; built-in Modbus TCP, KNX, DALI three protocol stacks, supports automatic identification of RS-485 / RS-232 / CAN bus data frame structure, abnormal data packet retransmission times ≤ 3 times; Edge computing node; uses the NVIDIA Jetson AGX Orin hardware platform, integrated with a multi-threaded data stream processing unit (supporting 4-channel 8K video parallel decoding).

[0018] The visible light camera array includes: Main camera: 8K resolution (7680×4320 pixels), frame rate 25fps, support AI stabilization (compensation accuracy 0.05° / s); Auxiliary camera: fisheye lens (FOV ≥ 180°), built-in distortion correction algorithm (polynomial fitting order ≤ 5).

[0019] The array data acquisition method comprises: Heterogeneous data preprocessing: Time series data (such as temperature and humidity) is normalized using a sliding window (window length = 3s) Three-level noise reduction is implemented for video data: spatial domain wavelet transform (Daubechies-4 basis function) + frequency domain adaptive filtering (cutoff frequency ≥ 15 Hz).

[0020] Multi-source data fusion uses an improved Kalman filter, and the state equation is: ; Where Rk is the covariance matrix of heterogeneous data, including: Video data: ; Environmental data: .

[0021] It includes a multispectral perception module, a dynamic analysis engine and a three-dimensional positioning unit.

[0022] The spatiotemporal joint modeling method is to combine the improved YOLOv7-ASFF and ST-GCN, as follows: Flame feature extraction layer: Constructs the stroboscopic feature vector (5-15Hz) and edge entropy matrix to accurately identify flame features.

[0023] Smoke diffusion model: This model uses the Improved Optical Flow (ICOF) method to calculate the concentration gradient field, enabling analysis of smoke diffusion. This combined spatiotemporal modeling approach enables comprehensive analysis and processing of fire safety-related characteristics from both temporal and spatial dimensions, facilitating more accurate detection of fires and related hazards, and improving the performance of fire safety early warning systems.

[0024] Multispectral perception module: A heterogeneous acquisition array integrating visible light / infrared / TOF sensors, specifically including a multimodal sensor group, a dynamic protocol parsing module, and an edge computing node.

[0025] The multimodal sensor suite includes a visible light camera (supporting H.266 / VVC encoding), a dual-band infrared thermal imager (7.5-13μm wavelength range), and a time-of-flight (TOF) ranging module (accuracy ±2cm). Each sensor is housed in an independent industrial-grade protective housing (IP67 rating). The main visible light camera boasts 8K resolution (7680×4320 pixels), a frame rate of 25fps, and AI image stabilization (compensation accuracy of 0.05° / s). The auxiliary camera uses a fisheye lens (FOV ≥ 180°) with a built-in distortion correction algorithm (polynomial fitting order ≤ 5).

[0026] Dynamic protocol analysis module: Built-in Modbus TCP, KNX, and DALI protocol stacks, supports automatic identification of RS-485 / RS-232 / CAN bus data frame structure, and the number of abnormal data packet retransmissions ≤ 3 times.

[0027] Edge computing node: uses the NVIDIA Jetson AGX Orin hardware platform, integrated with a multi-threaded data stream processing unit (supporting 4-channel 8K video parallel decoding).

[0028] Dynamic analysis engine: It adopts the spatiotemporal joint modeling method of improved YOLOv7-ASFF and ST-GCN, which specifically includes the flame feature extraction layer and smoke diffusion model.

[0029] Flame feature extraction layer: Constructs the stroboscopic feature vector (5-15Hz) and edge entropy matrix.

[0030] Smoke diffusion model: The improved optical flow method (ICOF) is used to calculate the concentration gradient field.

[0031] Hierarchical early warning strategy: including a four-level response mechanism and digital twin verification.

[0032] Four-level response mechanism: automatically triggers different linkage strategies according to the stage of fire development.

[0033] Digital twin verification: Verify the feasibility of the emergency plan through FDS simulation (error <15%).

[0034] Array data acquisition method: heterogeneous data preprocessing, time series data (such as temperature and humidity) are normalized using a sliding window (window length = 3s).

[0035] Three-level noise reduction is implemented for video data, including spatial domain wavelet transform (Daubechies-4 basis functions) + frequency domain adaptive filtering (cutoff frequency ≥ 15 Hz).

[0036] Multi-source data fusion: Using an improved Kalman filter, the state equation is: [Specific equation content], where Rk is the covariance matrix of heterogeneous data, including video data and environmental data.

[0037] The dynamic analysis engine utilizes an improved spatiotemporal joint modeling method of YOLOv7-ASFF and ST-GCN. The flame feature extraction layer accurately identifies flame characteristics by constructing stroboscopic feature vectors and edge entropy matrix. The smoke diffusion model uses an improved optical flow method to calculate the concentration gradient field, thereby enabling analysis of smoke diffusion.

[0038] Example 1: Petrochemical Tank Farm Deployment Hardware configuration: 8 8K cameras (50m apart) + 4 thermal imagers (30° top-down angle); Edge nodes: deploy one Jetson AGX Orin per 200 square meters; Algorithm training: Dataset: Self-built 100,000-frame video containing 15 types of fire scenes (including thick smoke, explosion, smoldering, etc.); Model optimization: Knowledge distillation technology is used to reduce the number of ResNet-3D parameters to 1 / 53; Performance indicators: Flame detection: F1-score = 0.987 (traditional method 0.723); Response delay: average 0.68s (83% improvement over traditional systems); False alarm rate: 0.12 times / 1,000 hours (false alarm rate in kitchen steam scene <0.5%).

[0039] Example 2: Super high-rise building application Special design: Vertical channel monitoring: using fisheye lens (FOV ≥ 180°) + UWB positioning; Evacuation guidance: dynamic path planning algorithm (Dijkstra + smoke diffusion model); Linkage strategy: Emergency response: Automatically close the elevator shaft and start the pressure differential ventilation of the refuge floor; Extreme response: Activate the building facade sprinkler system.

[0040] Hardware deployment specifications: Front-end device spacing: scene type, visible light camera spacing, thermal imager coverage radius; Open space ≤100m ≤80m; Confined space ≤50m ≤60m; Power management: PoE+battery dual power supply (battery life > 72h).

[0041] Software algorithm optimization: Model compression: channel pruning (pruning rate 40%) + quantization (INT8); Inference speed: 8.7 FPS on the edge → 21.3 FPS after optimization; Data enhancement: Introducing smoke occlusion simulation (occlusion rate 0-80%); Enhanced lighting changes (brightness range 0-255, contrast 1-3).

[0042] Example 3: Application of underground integrated pipeline corridor Special design: Long distance transmission: using optical fiber ring network (transmission delay <10ms); Humidity compensation: Establishing a temperature-humidity-flame characteristic mapping model (R²=0.96); Emergency process: Emergency response: Automatically start longitudinal smoke exhaust (wind speed ≥ 2.5m / s); Extreme response: Close three adjacent fire zones (response time < 1.5s); Example 4: Smart City Platform Integration Data connection: Connect with the API of the city's IoT platform (such as Huawei OceanConnect); Realize the linkage between fire protection data and transportation and medical systems (response delay < 200ms); Big data analysis: Building a fire risk index model: ; Where F is building density, H is population density, P is historical fire frequency, and E is economic value.

[0043] The multimodal sensor group in the multispectral perception module is responsible for collecting different types of data. The visible light camera captures scene images, the infrared thermal imager detects temperature distribution, and the time-of-flight ranging module measures distance information. The dynamic protocol parsing module performs protocol analysis on the collected data to ensure accurate data transmission. The edge computing node performs preliminary data processing to reduce the burden of subsequent processing.

[0044] The hierarchical early warning strategy automatically triggers corresponding linkage strategies based on a four-level response mechanism according to the fire development stage, such as activating firefighting equipment and notifying relevant personnel. Digital twin verification verifies the feasibility of emergency plans through FDS simulation, ensuring effective response in the event of an actual fire.

[0045] The array data acquisition method pre-processes heterogeneous data to make it more suitable for subsequent analysis. Multi-source data fusion uses an improved Kalman filter to fuse data from different sources to improve data accuracy and reliability.

[0046] This system achieves efficient and accurate fire safety early warnings through the collaborative work of a multispectral sensing module, a dynamic analysis engine, and a three-dimensional positioning unit. This significantly improves early warning speed compared to traditional methods, enabling early detection of smoldering fires. Furthermore, a multi-protocol adaptive parsing algorithm effectively reduces false alarms and minimizes unnecessary interference. The efficient processing capabilities of edge computing nodes enhance the overall operational efficiency of the system, while the application of knowledge graphs improves the matching accuracy of emergency response plans and enhances the system's intelligence.

[0047] Example

[0048]

[0049]

[0050]

[0051]

[0052]

[0053]

[0054]

[0055]

[0056]

[0057]

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065] .

[0066] In different application scenarios, deployment in petrochemical storage tank areas achieves high-precision flame detection and rapid response through reasonable hardware configuration and algorithm optimization. Applications in super-high-rise buildings are specially designed for vertical channel monitoring and evacuation guidance, and system performance is improved through software algorithm optimization. Applications in underground integrated pipeline corridors solve problems such as long-distance transmission and humidity compensation, ensuring the rapid and accurate execution of emergency procedures. Smart city platform integration enables the linkage of fire data with other systems, enhancing the city's overall fire safety management capabilities. With the continuous advancement of technology, this system can be further optimized and expanded. In terms of sensors, more advanced and sensitive sensors can be developed to improve the accuracy and comprehensiveness of data collection. In terms of algorithms, deep learning models are continuously improved to enhance the ability to recognize complex fire scenarios and the timeliness of early warnings.

[0067] In terms of system integration, we will strengthen integration with more fields, such as integrating with smart grids to achieve fire protection for power facilities; deeply integrating with intelligent security systems to build a more comprehensive security protection network. At the same time, we will use big data and cloud computing technologies to conduct in-depth analysis of large amounts of fire data, identify potential fire risks, and provide more scientific decision-making basis for fire safety management.

[0068] Furthermore, with the widespread adoption of 5G technology, high-speed data transmission and real-time processing of system data will be achieved, further improving the system's responsiveness and performance. Through continuous technological innovation and improvement, this fire safety early warning system based on AI video recognition will play a greater role in the field of fire safety, providing stronger support for protecting people's lives and property.

[0069] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0070] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. Fire safety early warning system based on artificial intelligence video recognition, characterized by: Include: Hierarchical warning strategy, visible light camera array, multi-source data fusion and multi-spectral perception module, among which the multi-spectral perception module: a heterogeneous acquisition array integrating visible light / infrared / TOF sensors, Dynamic analysis engine: adopts the spatiotemporal joint modeling method of improved YOLOv7-ASFF and ST-GCN, 3D positioning unit: A fusion positioning system based on an improved SfM algorithm and UWB (error < 0.5m).

2. The fire safety early warning system based on artificial intelligence video recognition according to claim 1 is characterized by: The dynamic analysis engine includes: Flame feature extraction layer: constructs the stroboscopic feature vector (5-15Hz) and edge entropy matrix; Smoke diffusion model: The improved optical flow method (ICOF) is used to calculate the concentration gradient field.

3. The fire safety early warning system based on artificial intelligence video recognition according to claim 1 is characterized by: The hierarchical warning strategy includes: Four-level response mechanism: automatically triggers different linkage strategies according to the fire development stage; Digital twin verification: Verify the feasibility of the emergency plan through FDS simulation (error <15%).

4. The fire safety early warning system based on artificial intelligence video recognition according to claim 3 is characterized by: The heterogeneous acquisition array includes: Multimodal sensor suite; including a visible light camera (supporting H.266 / VVC encoding), a dual-band infrared thermal imager (7.5-13μm wavelength range), and a Time of Flight (TOF) ranging module (accuracy ±2cm). Each sensor is housed in an independent industrial-grade protective housing (IP67 protection rating). Dynamic protocol parsing module; built-in Modbus TCP, KNX, DALI three protocol stacks, supports automatic identification of RS-485 / RS-232 / CAN bus data frame structure, abnormal data packet retransmission times ≤ 3 times; Edge computing node; uses the NVIDIA Jetson AGX Orin hardware platform, integrated with a multi-threaded data stream processing unit (supporting 4-channel 8K video parallel decoding).

5. The fire safety early warning system based on artificial intelligence video recognition according to claim 1 is characterized in that: The visible light camera array comprises: Main camera: 8K resolution (7680×4320 pixels), frame rate 25fps, support AI stabilization (compensation accuracy 0.05° / s); Auxiliary camera: fisheye lens (FOV ≥ 180°), built-in distortion correction algorithm (polynomial fitting order ≤ 5).

6. The fire safety early warning system based on artificial intelligence video recognition according to claim 1 is characterized in that: The data acquisition method of the array comprises: Heterogeneous data preprocessing: Time series data (such as temperature and humidity) is normalized using a sliding window (window length = 3s); Three-level noise reduction is implemented for video data: spatial domain wavelet transform (Daubechies-4 basis function) + frequency domain adaptive filtering (cutoff frequency ≥ 15 Hz).

7. The fire safety early warning system based on artificial intelligence video recognition according to claim 1 is characterized by: The multi-source data fusion adopts an improved Kalman filter, and the state equation is: ； Where Rk is the covariance matrix of heterogeneous data, including: Video data: ; Environmental data: .

Images