Fire-fighting detection alarm system and method for emergency management

Abstract

The invention belongs to the technical field of fire protection, and particularly discloses a fire protection detection alarm system and method for emergency management, and the system comprises a multi-source sensing module which aligns image data with sensor data based on a timestamp; the self-adaptive generation module is used for calling a corresponding generator variant based on a scene label, and extracting morphological characteristics of flame / smoke from the enhanced image by adopting an improved YOLOv5 model; the multi-source fusion decision-making module fuses the three types of features through an attention mechanism, calculates fire confidence based on the fused features, sets three-level threshold values, sends out sound-light alarms of different intensities according to fire levels, pushes information containing fire positions to a fire control center, and calculates fire three-dimensional coordinates by combining sensor deployment positions and image coordinates; predicting a fire spreading range; and the self-optimization module evaluates the alarm performance. According to the fire-fighting detection alarm system and method for emergency management, the early recognition rate of a fire disaster is increased, and accurate judgment and early warning of the fire disaster are achieved.

Description

Technical Field

[0001] This invention relates to the field of fire protection technology, and in particular to a fire detection and alarm system and method for emergency management. Background Technology

[0002] Fire detection and alarm systems play a crucial role in emergency management, with applications spanning multiple fields and locations. Their core function is to minimize losses from fires and other fire-related accidents through real-time monitoring, rapid early warning, and coordinated response. For industrial environments such as chemical warehouses, which often store large quantities of flammable, explosive, corrosive, or toxic chemicals, complex fire hazards are common, making the application of fire detection and alarm systems particularly important.

[0003] In existing technologies, chemical warehouse fires are characterized by rapid fire spread, high explosion risk, strong toxicity, and difficulty in firefighting. Fire detection and alarm systems often rely on simple sensor alarms, such as point-type smoke and heat detectors. These systems are insensitive to the early characteristics of specific chemical fires and smoldering fire characteristics, lack the ability to identify environmental interference factors such as dust, vapor, and electromagnetic interference, and lack intelligent analysis and predictive capabilities, making it difficult to accurately assess and warn of early fire risks. Summary of the Invention

[0004] The purpose of this invention is to provide a fire detection and alarm system and method for emergency management, which can improve the early identification rate of fires and achieve accurate judgment and early warning of fires.

[0005] To achieve the above objectives, the present invention provides a fire detection and alarm system for emergency management, comprising: The multi-source sensing module aligns image data and sensor data based on timestamps; The adaptive generation module calls the corresponding generator variant based on the scene label and uses an improved YOLOv5 model to extract the morphological features of flames / smoke from the enhanced image, including area, contour, and motion trajectory. The multi-source fusion decision module first removes abnormal sensor data through the data cleaning unit, smooths the data using a sliding window filter, then fuses three types of features through an attention mechanism, calculates the fire confidence level based on the fused features, sets three-level thresholds, and finally issues audible and visual alarms of different intensities according to the fire level, pushes information including the fire location to the fire control center, calculates the three-dimensional coordinates of the fire by combining the sensor deployment location and image coordinates, and predicts the spread range of the fire in the next 5-10 minutes based on historical data and current features. The self-optimization module evaluates alarm performance, automatically collects new types of samples and updates the generator, and uses federated learning to update the recognition model.

[0006] Preferably, image data is obtained through an image acquisition unit and sensor data is obtained through an environmental sensing unit; The image acquisition unit includes an infrared camera and a visible light camera, covering the monitoring area and acquiring real-time image streams; The environmental sensing unit includes a temperature sensor, a smoke sensor, and a gas sensor, and collects environmental data at a frequency of 10 seconds per scan.

[0007] Preferably, the generator variant is as follows: Generator Variant 1: For low-light enhancement generator, based on CycleGAN, infrared images and low-light visible light images are converted into enhanced images to highlight flame / smoke features; Generator Variant 2: For the smoke penetration generator, it removes dense smoke obstruction based on Pix2Pix and restores the core area of ​​the flame; Generator Variant 3: For rare fire generators, feature samples of different burning substances are generated based on CGAN pre-training to help the identification model learn special fires.

[0008] Preferably, the CGAN pre-training is as follows: Sample generation: Using the CGAN generator, inputting low-light fire and early smoke labels, generating samples for the corresponding scenarios, and filtering samples with ≥90% similarity to real fire features through a feature matching model; Hybrid Training: Real samples and generated samples are mixed proportionally to train a ResNet-based fire detection model. The adversarial loss function is set as follows: ; Where L is the total loss function, To generate the adversarial loss, L_cls is the classification loss. , These are the weighting coefficients; Recognition Application: For input surveillance images, if low light is observed, CycleGAN is automatically invoked to generate an enhanced image, which is then fused with the features of the original image to output the recognition result.

[0009] Preferably, the three types of features include: Image features: presence, area, and diffusion rate of flames / smoke; Sensor characteristics: temperature change rate, smoke concentration, gas concentration; Environmental characteristics: spatial layout, ventilation conditions, and historical fire data.

[0010] Preferably, the multi-source fusion decision module also includes an emergency linkage unit, which automatically triggers linkage measures, including one or more of the following: starting the sprinkler system, closing the ventilation duct, turning on the emergency lighting, and unlocking the evacuation route.

[0011] Preferably, the three-level threshold is as follows: Warning level: Confidence level 30%-50%; Alarm level: Confidence level 50%-80%; Emergency level: Confidence level ≥ 80%.

[0012] Preferably, the self-optimization module includes a performance monitoring unit, a sample update unit, and a model iteration unit. The performance monitoring unit specifically comprises: Statistical analysis of alarm accuracy, response time, and false alarm rate; The sample update unit is specifically as follows: When a new type of fire occurs, samples are automatically collected and the generator is updated; The model iteration unit is specifically: Federated learning is used to update the recognition model, ensuring that the edge device and the cloud model are optimized in sync.

[0013] This invention also provides a fire alarm detection method for emergency management, comprising the following steps: S1. Multi-source data acquisition: S11. Synchronously collect visible light / infrared images, temperature, smoke concentration, and gas concentration data of the monitoring area; S12. Add timestamps and location tags to the raw data and transmit it to the processing terminal; S2. Image Enhancement and Feature Extraction: S21. Select the appropriate generator variant based on the environmental parameters; S22. Input the original image into the generator, and output the enhanced image; S23. Extract morphological and motion features of flames / smoke from the enhanced image; S3. Multi-source data fusion decision-making: S31. Filter the sensor data and calculate the temperature change rate and concentration gradient; S32. Integrate image features, sensor features, and environmental features to calculate the fire confidence level; S33. Determine the fire severity level based on confidence level; S4. Intelligent Alarm and Linkage: S41. Activate the corresponding alarm mode according to the fire level and push location information; S42. Predict the fire's spread and generate emergency response recommendations; S43. Trigger the corresponding level of linkage measures; S5. System self-optimization: S51. Regularly evaluate alarm performance, and initiate model optimization when the false alarm rate exceeds 5%; S52. Collect new fire samples and update the generator and recognition model; S53. Achieve co-evolution between edge nodes and cloud models through federated learning.

[0014] Therefore, the fire detection and alarm system and method for emergency management described above have the following beneficial effects: This invention has detection and recognition capabilities: based on multi-scene image recognition technology using generative adversarial networks, it improves the accuracy of image recognition through adaptive sample generation, cross-scene feature enhancement, and end-to-end joint training, achieving an accuracy rate of over 95% in recognizing flames, smoke, and gases, and can accurately detect fire characteristic parameters in the early stages of a fire.

[0015] This invention has alarm and linkage capabilities: it can issue an alarm signal within 10 seconds and accurately display the alarm location; the linkage response time is short and the linkage accuracy reaches 100%.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a fire detection and alarm system for emergency management according to the present invention. Detailed Implementation

[0018] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0020] Example 1 like Figure 1 As shown, an emergency management fire detection and alarm system includes: The multi-source sensing module aligns image data and sensor data based on timestamps to ensure spatiotemporal consistency. It obtains raw image data through the image acquisition unit and sensor data through the environmental sensing unit. The image acquisition unit consists of an infrared camera and a visible light camera, covering the monitoring area and acquiring a real-time image stream. Temperature sensors, smoke sensors, and gas sensors (detecting CO and combustible gas concentrations) are deployed as environmental sensing units, acquiring environmental data at a frequency of 10 seconds per acquisition.

[0021] The adaptive generation module calls the corresponding generator variant based on scene labels and uses an improved YOLOv5 model to extract the morphological features (area, contour, motion trajectory) of flames / smoke from the enhanced image. Generator variant 1: For the low-light enhancement generator, it uses CycleGAN to convert infrared and low-light visible light images into enhanced images to highlight flame / smoke features; Generator variant 2: For the smoke penetration generator, it uses Pix2Pix to remove dense smoke obstructions and restore the core area of ​​the flame; Generator variant 3: For the rare fire generator, it uses CGAN pre-training to generate feature samples of burning different materials (wood, chemicals, electrical appliances) to help the recognition model learn special fire conditions.

[0022] CGAN pre-training, as detailed below: 1. Sample generation: Using the CGAN generator, input the labels "low light fire" and "early smoke" to generate samples for the corresponding scenarios. The feature matching model is used to select samples with a similarity of ≥90% to the features of real fires.

[0023] 2. Hybrid Training: A ResNet-based fire detection model is trained by mixing real samples (5000 images) and generated samples (3000 images) in a 6:4 ratio, using an adversarial loss function, i.e., the total loss function. Set to: ; in, To generate the adversarial loss, L_cls is the classification loss. =0.3, =0.7.

[0024] 3. Recognition Application: If the input monitoring image shows "low light", CycleGAN is automatically called to generate an enhanced image, and the recognition result is output after fusing the features of the original image.

[0025] The multi-source fusion decision-making module first removes abnormal sensor data (such as sudden changes in values ​​from a single sensor) through a data cleaning unit, smooths the data using a sliding window filter, then fuses three types of features through an attention mechanism, calculates the fire confidence level based on the fused features, sets three-level thresholds, and finally issues audible and visual alarms of different intensities according to the fire severity level, pushing information including the fire location to the fire control center. Combining sensor deployment locations and image coordinates, it calculates the three-dimensional coordinates of the fire with an error ≤1 meter, and predicts the spread range of the fire in the next 5-10 minutes based on historical data and current characteristics. In addition, an emergency linkage unit is set up to automatically trigger linkage measures (such as activating the sprinkler system, closing ventilation ducts, turning on emergency lighting, and unlocking evacuation routes).

[0026] The three types of characteristics include: Image features: presence, area, and diffusion rate of flames / smoke; Sensor features: temperature change rate, smoke concentration, and gas concentration; Environmental features: spatial layout (e.g., presence of flammable materials), ventilation conditions, and historical fire data.

[0027] Level 3 thresholds include: Warning level (confidence level 30%-50%): Suspected fire, initiate secondary verification; Alarm level (confidence 50%-80%): Confirms fire and triggers audible and visual alarms; Emergency Level (Confidence ≥ 80%): The fire is large, and emergency response should be activated.

[0028] The specific method for calculating confidence level is as follows: 1. Quantize the three types of features and use them as input for confidence calculation: Image feature weighting S_img: ; Where α1+α2+α3=1, in this embodiment α1=0.5, α2=0.3, α3=0.2, which can be dynamically adjusted; P_img is the probability of flame / smoke presence; M_img is the feature matching degree, that is, the feature similarity between the generated enhanced image and the real fire sample database; D_img is the dynamic feature score, that is, the spread speed and area change rate of flame / smoke, which is normalized to the range [0,1]. The faster the spread and the more intense the area growth, the higher the score.

[0029] Sensor feature weights are divided into S_sensor: ; Wherein, β1+β2+β3=1, in this embodiment β1=0.4, β2=0.3, β3=0.3, and β1 can be increased in high temperature scenarios; T_sensor is the temperature anomaly score, the deviation rate between the current temperature and the baseline temperature, which is 1 when it exceeds the threshold, otherwise it is normalized to [0,1] proportionally; Q_sensor is the smoke concentration score, the proportion of smoke concentration exceeding the alarm threshold, in the range [0,1]; G_sensor is the gas concentration score, the normalized value of CO or combustible gas concentration, in the range [0,1].

[0030] Environmental feature weighting S_env: ;

[0031] in, In this embodiment , E_env is the environmental risk coefficient, based on the warehouse space layout, flammable material distribution, and ventilation conditions, ranging from [0,1] (e.g., 0.8 for chemical warehouses by default); H_env is the historical matching degree, the similarity between the current feature and historical fire data, ranging from [0,1].

[0032] 2. The fire confidence score is calculated by dynamically weighting and fusing the three feature scores obtained in step 1 above through an attention mechanism.

[0033] The weights are automatically adjusted based on the reliability of the features: When the image clarity is high (no dense smoke / low light), increase the image feature weight (S_img); When sensor data changes continuously and stably (without sudden changes), increase the sensor feature weight (S_sensor). When the risk level of a scenario is high (such as a flammable fuel depot), the weight of environmental features (S_env) should be increased.

[0034] The condition is satisfied that S_img+S_sensor+S_env=1.

[0035] 3. Introduce a time-sliding window to correct the confidence level. Take the confidence scores [C1, C2, C3, C4, C5] for the 5 most recent moments (1 second apart); Calculate the weighted average C_final: (Recent data has higher weight); If C_final exceeds the same threshold three times consecutively, the fire level is confirmed.

[0036] The self-optimization module includes a performance monitoring unit, a sample update unit, and a model iteration unit.

[0037] Performance monitoring unit: Statistical analysis of alarm accuracy, response time, false alarm rate, and other indicators.

[0038] Sample update unit: When a new type of fire occurs (such as the combustion of new chemicals), it automatically collects samples and updates the generator.

[0039] Model Iteration Unit: Federated learning is used to update the recognition model, ensuring that the edge device and the cloud model are optimized synchronously.

[0040] Example 2 A fire alarm detection method for emergency management includes the following steps: S1. Multi-source data acquisition: S11. Synchronously collect visible light / infrared images, temperature, smoke concentration, and gas concentration data of the monitoring area; S12. Add timestamps and location tags to the raw data and transmit it to the processing terminal.

[0041] S2. Image Enhancement and Feature Extraction: S21. Select the corresponding generator variant based on environmental parameters (such as light intensity and smoke concentration); S22. Input the original image into the generator, and output the enhanced image; S23. Extract morphological and motion features of flames / smoke from the enhanced image.

[0042] S3. Multi-source data fusion decision-making: S31. Filter the sensor data and calculate dynamic characteristics such as temperature change rate and concentration gradient; S32. Integrate image features, sensor features, and environmental features to calculate the fire confidence level; S33. Determine the fire level (early warning / alarm / emergency) based on the confidence level.

[0043] S4. Intelligent Alarm and Linkage: S41. Activate the corresponding alarm mode according to the fire level and push location information; S42. Predict the fire's spread and generate emergency response recommendations; S43. Trigger the corresponding level of linkage measures (such as sprinkler or ventilation control).

[0044] S5. System self-optimization: S51. Regularly evaluate alarm performance, and initiate model optimization when the false alarm rate exceeds 5%; S52. Collect new fire samples and update the generator and recognition model; S53. Achieve co-evolution between edge nodes and cloud models through federated learning.

[0045] Example 3 For fire detection scenarios in chemical warehouses, an emergency management fire detection and alarm system is proposed. 32 sensor nodes (including temperature, smoke, and CO sensors) are deployed at 5m×5m intervals inside the warehouse. Eight high-definition visible light cameras and four infrared cameras are installed to achieve blind-spot-free coverage. An edge computing terminal is deployed in the warehouse control room and linked with the fire control system.

[0046] Experimental testing: When an electrical short circuit in a corner of the warehouse triggered an early fire (temperature 38℃→52℃, slight smoke), sensor nodes and cameras simultaneously collected data. Due to the low smoke concentration and uneven lighting, the system invoked a low-light enhancement generator to highlight the faint smoke features in the image, extracting smoke down to 0.3m. 2The suspected smoke area was identified. Fusion data showed: temperature change rate +2℃ / min, CO concentration 0.002%, image feature matching degree 82%, and calculated fire confidence level 65% (alarm level). The system triggered an audible and visual alarm, pushed the fire location (shelf number), predicted that the fire would spread to adjacent shelves within 5 minutes, and automatically activated the sprinkler system in that area. Upon receiving the alarm, the fire control center dispatched personnel with fire extinguishers to the scene, extinguishing the initial fire within 3 minutes.

[0047] In this test, the system's alarm response time from the occurrence of a fire was 2.7 seconds, the positioning error was 0.8 meters, there were no missed alarms or false alarms, and the linkage measures were accurately triggered.

[0048] Therefore, the present invention employs the above-mentioned emergency management fire detection and alarm system and method to improve the early fire identification rate and achieve accurate fire judgment and early warning.

[0049] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A fire detection and alarm system for emergency management, characterized in that, include: The multi-source sensing module aligns image data and sensor data based on timestamps; The adaptive generation module calls the corresponding generator variant based on the scene label and uses an improved YOLOv5 model to extract the morphological features of flames / smoke from the enhanced image, including area, contour, and motion trajectory. The multi-source fusion decision module first removes abnormal sensor data through the data cleaning unit, smooths the data using a sliding window filter, then fuses three types of features through an attention mechanism, calculates the fire confidence level based on the fused features, sets three-level thresholds, and finally issues audible and visual alarms of different intensities according to the fire level, pushes information including the fire location to the fire control center, calculates the three-dimensional coordinates of the fire by combining the sensor deployment location and image coordinates, and predicts the spread range of the fire in the next 5-10 minutes based on historical data and current features. The self-optimization module evaluates alarm performance, automatically collects new types of samples and updates the generator, and uses federated learning to update the recognition model.

2. The fire detection and alarm system for emergency management according to claim 1, characterized in that: Image data is obtained through the image acquisition unit, and sensor data is obtained through the environmental sensing unit. The image acquisition unit includes an infrared camera and a visible light camera, covering the monitoring area and acquiring real-time image streams; The environmental sensing unit includes a temperature sensor, a smoke sensor, and a gas sensor, and collects environmental data at a frequency of 10 seconds per scan.

3. The fire detection and alarm system for emergency management according to claim 1, characterized in that, The generator variants are specifically: Generator Variant 1: For low-light enhancement generator, based on CycleGAN, infrared images and low-light visible light images are converted into enhanced images to highlight flame / smoke features; Generator Variant 2: For the smoke penetration generator, it removes dense smoke obstruction based on Pix2Pix and restores the core area of ​​the flame; Generator Variant 3: For rare fire generators, feature samples of different burning substances are generated based on CGAN pre-training to help the identification model learn special fires.

4. The fire detection and alarm system for emergency management according to claim 3, characterized in that, The CGAN pre-training process is as follows: Sample generation: Using the CGAN generator, inputting low-light fire and early smoke labels, generating samples for the corresponding scenarios, and filtering samples with ≥90% similarity to real fire features through a feature matching model; Hybrid Training: Real samples and generated samples are mixed proportionally to train a ResNet-based fire detection model. The adversarial loss function is set as follows: ； Where L is the total loss function, To generate the adversarial loss, L_cls is the classification loss. , These are the weighting coefficients; Recognition Application: For input surveillance images, if low light is observed, CycleGAN is automatically invoked to generate an enhanced image, which is then fused with the features of the original image to output the recognition result.

5. The fire detection and alarm system for emergency management according to claim 1, characterized in that, The three types of characteristics include: Image features: presence, area, and diffusion rate of flames / smoke; Sensor characteristics: temperature change rate, smoke concentration, gas concentration; Environmental characteristics: spatial layout, ventilation conditions, and historical fire data.

6. The fire detection and alarm system for emergency management according to claim 1, characterized in that: The multi-source fusion decision-making module also includes an emergency linkage unit, which automatically triggers linkage measures, including starting the sprinkler system, closing the ventilation ducts, turning on the emergency lighting, and unlocking the evacuation passages, among one or more of these.

7. The fire detection and alarm system for emergency management according to claim 1, characterized in that, The specific threshold for level three is as follows: Warning level: Confidence level 30%-50%; Alarm level: Confidence level 50%-80%; Emergency level: Confidence level ≥ 80%.

8. The fire detection and alarm system for emergency management according to claim 1, characterized in that, The self-optimization module includes a performance monitoring unit, a sample update unit, and a model iteration unit. The performance monitoring unit specifically includes: Statistical analysis of alarm accuracy, response time, and false alarm rate; The sample update unit is specifically as follows: When a new type of fire occurs, samples are automatically collected and the generator is updated; The model iteration unit is specifically: Federated learning is used to update the recognition model, ensuring that the edge device and the cloud model are optimized in sync.

9. A fire detection and alarm method for emergency management, employing the fire detection and alarm system for emergency management as described in any one of claims 1-8, characterized in that, Includes the following steps: S1. Multi-source data acquisition: S11. Synchronously collect visible light / infrared images, temperature, smoke concentration, and gas concentration data of the monitoring area; S12. Add timestamps and location tags to the raw data and transmit it to the processing terminal; S2. Image Enhancement and Feature Extraction: S21. Select the appropriate generator variant based on the environmental parameters; S22. Input the original image into the generator, and output the enhanced image; S23. Extract morphological and motion features of flames / smoke from the enhanced image; S3. Multi-source data fusion decision-making: S31. Filter the sensor data and calculate the temperature change rate and concentration gradient; S32. Integrate image features, sensor features, and environmental features to calculate the fire confidence level; S33. Determine the fire severity level based on confidence level; S4. Intelligent Alarm and Linkage: S41. Activate the corresponding alarm mode according to the fire level and push location information; S42. Predict the fire's spread and generate emergency response recommendations; S43. Trigger the corresponding level of linkage measures; S5. System self-optimization: S51. Regularly evaluate alarm performance, and initiate model optimization when the false alarm rate exceeds 5%; S52. Collect new fire samples and update the generator and recognition model; S53. Achieve co-evolution between edge nodes and cloud models through federated learning.

Images