Fire monitoring controller with self-adaptive alarm threshold adjusting function

Abstract

The invention discloses a fire monitoring controller with a self-adaptive alarm threshold adjustment function, and belongs to the technical field of fire monitoring. Comprising a data acquisition module used for acquiring visible light and infrared double-spectrum video streams of a monitoring area and environment temperature and humidity; the data processing module is used for executing scene classification, fire feature analysis, dynamic alarm threshold calculation and alarm decision according to the video stream and the environment data; the feedback learning module is used for optimizing the fire analysis model according to the false report and / or missing report data; and the alarm execution module triggers sound-light alarm and fire-fighting equipment linkage according to the alarm decision. According to the invention, adaptive adjustment of the alarm threshold is realized, the problems of poor threshold adaptation, poor environmental adaptability, high false alarm / missing alarm rate, response lag and the like of the device in the prior art are solved, and a high-reliability solution is provided for multi-scene fire prevention and control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fire monitoring, and particularly relates to a fire monitoring controller with adaptive alarm threshold adjustment function. BACKGROUND

[0002] As the core of the automatic fire alarm system, the core function of the fire alarm controller is to receive the detection signal in real time and determine whether a fire occurs. The traditional fixed threshold alarm has significant defects. At present, some improved schemes have realized the adjustable function of the basic threshold, for example, setting three different thresholds of high / medium / low to realize the preliminary adjustment of the threshold. However, the setting of the threshold is still relatively fixed, and it is difficult to adapt to complex monitoring environments, thereby leading to high false alarm / missed alarm rates, which greatly affects the monitoring effect. Some improved schemes use composite monitoring methods and consider the influence of environmental factors to realize certain adaptive threshold adjustment capability. However, the adjusted threshold is still difficult to adapt to complex monitoring environments, and false alarm / missed alarm often occurs, which causes certain troubles to the users and even economic losses.

[0003] At present, some improved schemes supporting self-learning have slow response speed and lack real-time response according to the false alarm severity and missed alarm, leading to poor dynamic adaptability. If various influencing factors are comprehensively considered to improve the monitoring effect, multiple sensors need to be cooperated with each other, and it is not realistic to consider the cost and other factors in actual practice. Therefore, how to realize the adaptation of multiple scenes, the adaptive alarm threshold, the timely response of feedback learning, and the reduction of false alarm / missed alarm rate on the basis of practicality has become the direction and key of the research of the technical personnel in the field. SUMMARY

[0004] Therefore, the present application is proposed in view of the problems and needs in the prior art. The purpose of the present application is realized through the following technical scheme:

[0005] The present application provides a fire monitoring controller with adaptive alarm threshold adjustment function, comprising:

[0006] A data acquisition module is configured to acquire a visible light and infrared dual-spectrum video stream of a monitored area and environmental temperature and humidity.

[0007] A data processing module is configured to perform scene classification, fire feature analysis, dynamic alarm threshold calculation, and alarm decision making according to the video stream and the environmental data.

[0008] A feedback learning module is configured to optimize a fire analysis model according to false alarm and / or missed alarm data.

[0009] An alarm execution module is configured to trigger sound and light alarms and fire-fighting equipment linkage according to the alarm decision making.

[0010] In the fire monitoring controller with adaptive alarm threshold adjustment function, the data processing module performs scene classification, specifically:

[0011] Canny edge detection is performed on the visible light video frame, and when the vertical line density is > 40 lines / m2, it is determined as a warehouse scene;

[0012] The infrared thermal imaging image is analyzed, and when the curved surface heat source proportion is > 15% or the high-temperature equipment density is > 5 / 100m2, it is determined as an industrial workshop scene;

[0013] HSV color space segmentation is performed on the visible light image, and when the green pixel proportion is > 30% and the building density is < 10%, it is determined as an outdoor scene.

[0014] In the fire monitoring controller with adaptive alarm threshold adjustment function, the fire feature analysis of the data processing module is specifically:

[0015] For flame features, in the visible light channel: whether the area flicker frequency contains 3-18Hz range, whether the RGB pixel R> G> B proportion is > 60%, and whether the edge Fourier descriptor satisfies the circularity < 0.7; in the infrared channel: whether the thermal radiation gradient ΔT is ≥ 5℃ / s;

[0016] For smoke features, in the visible light channel: whether the motion target rising track slope is > 55°, and whether the continuous 5-frame pixel contrast attenuation rate is > 25%; in the infrared channel: whether the low-temperature cluster mass center motion rate is > 0.3m / s;

[0017] The flame features and smoke features are input into the fire analysis model, and the fire probability value P is output, which is 0-1. fire The fire analysis model is MobileNetV3.

[0018] In the fire monitoring controller with adaptive alarm threshold adjustment function, the dynamic alarm threshold calculation of the data processing module is specifically:

[0019] According to the scene classification result, the basic threshold T is matched b , specifically: warehouse scene 0.6, industrial workshop scene 0.7, and outdoor scene 0.65;

[0020] Environmental compensation: when the relative humidity is > 85%, K env = 1.2; when the environmental temperature is > 50℃ and it is an industrial workshop scene, K env = 1.15; when the environmental temperature is > 40℃ and it is an outdoor scene, K env = 0.9; the K env is the environmental compensation coefficient;

[0021] Dynamic alarm threshold T d : T d = T b × K env × (1-0.05× P fire ).

[0022] In the fire monitoring controller with adaptive alarm threshold adjustment function, the smoke feature analysis includes strong wind interference coping, specifically:

[0023] Calculate the horizontal wind speed by 3 consecutive smoke motion vectors, and determine that the horizontal wind speed is greater than 2 m / s when the displacement is greater than 5 pixels / frame;

[0024] When the horizontal wind speed is greater than 2 m / s, close the visible light channel smoke trajectory slope analysis, strengthen the infrared channel flame detection, and reduce the ΔT threshold to 3 ℃ / s.

[0025] In the fire monitoring controller with adaptive alarm threshold adjustment function, the dynamic alarm threshold is also adjusted according to the day and night mode, specifically:

[0026] Calculate the average brightness value of the visible light video, and determine that it is daytime when the average brightness value is greater than 180 lux, K day = 1.1; determine that it is night when the average brightness value is less than 30 lux, K day = 0.85; when 180 lux≥average brightness value≥30 lux, K day = 1.0; the K day is a day and night mode compensation factor;

[0027] Dynamic alarm threshold correction: T d = T d × K day .

[0028] In the fire monitoring controller with adaptive alarm threshold adjustment function, the feedback learning module is specifically used for:

[0029] After manually marking false alarms, extract flame and smoke feature values and temperature and humidity data;

[0030] When P fire ≥1.5 times T d and is not a fire, mark it as a serious false alarm, immediately generate an adversarial sample, and complete fire analysis model retraining within 72 hours;

[0031] Otherwise, mark it as a common false alarm, save the flame and smoke feature data and temperature and humidity data, and trigger fire analysis model optimization when the cumulative number is greater than or equal to 2 times;

[0032] When the false negative is confirmed, the scene-based threshold is immediately reduced by 0.1, and the lower limit of the basic threshold is 0.4; the corresponding feature weight in the fire analysis model is increased by 15%, and the upper limit is 90%.

[0033] In the fire monitoring controller with the adaptive alarm threshold adjustment function described above, the alarm decision is specifically:

[0034] Third-level response: P fire When the dynamic alarm threshold is greater than or equal to 90 dB, the audible and visual alarms are started;

[0035] Second-level response: P fire When the dynamic alarm threshold is greater than or equal to 1.25 times, the sprinkler system is linked to handle on the basis of the third-level response;

[0036] First-level response: P fire When the dynamic threshold is greater than or equal to 1.5 times, the smoke exhaust fan and emergency broadcast are started on the basis of the second-level response.

[0037] Compared with the prior art, the beneficial effects of the present application are at least as follows:

[0038] The fire monitoring controller with the adaptive alarm threshold adjustment function of the present application classifies scenes through video stream images, distinguishes between industrial workshops, warehouses, and different outdoor scenes; extracts and analyzes flame features and smoke features, outputs fire probability values from a pre-trained fire analysis model, adjusts feature weights in the model considering the influence of strong winds on smoke feature analysis, and is more adaptable; matches the basic alarm threshold according to different scenes, considers the influence of environmental temperature and humidity factors and the real-time output of the fire probability value of the model on the basis of the basic threshold, generates a dynamic alarm threshold, and also considers the influence of light changes on visible light analysis to compensate for the dynamic alarm threshold in day and night modes; adopts a hierarchical alarm mechanism, and the response speed is faster; the model adopts a false alarm / missed alarm feedback learning optimization mechanism, and the false alarm suppression rate is significantly greater than the predetermined target of 35%, and the missed alarm suppression rate is also greatly improved, being greater than 70%.

[0039] In summary, the present application realizes adaptive adjustment of the alarm threshold through scene classification, flame and smoke feature analysis to output fire probability values, real-time adjustment of the dynamic alarm threshold, and closed-loop processing of feedback learning optimization, solves the problems of poor device threshold adaptability, poor environmental adaptability, high false alarm / missed alarm rate, and response lag in the prior art, and provides a high-reliability solution for multi-scene fire prevention and control. BRIEF DESCRIPTION OF DRAWINGS

[0040] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings in which:

[0041] Figure 1 is a structural schematic diagram of a fire monitoring controller with adaptive alarm threshold adjustment function provided by an embodiment of the present application;

[0042] Figure 2 is a working flow schematic diagram of a fire monitoring controller provided by an embodiment of the present application;

[0043] Figure 3 is a false alarm event processing flow schematic diagram of a fire monitoring controller provided by an embodiment of the present application. DETAILED DESCRIPTION

[0044] Hereinafter, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely a part of the embodiments of the present application, and the present application is not limited by the example embodiments described herein. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without making creative efforts fall within the scope of protection of the present application.

[0045] Embodiment One

[0046] The traditional fixed threshold alarm has significant defects. At present, some improved schemes have realized the basic threshold adjustable function, but the threshold setting is still relatively fixed, which is difficult to adapt to complex monitoring environment, and thus leads to high false alarm / misreporting rate, which greatly affects the monitoring effect. Some improved schemes use composite monitoring means, consider the influence of environmental factors, and realize certain adaptive threshold adjustment capability. Not only is it only suitable for single or fixed scene, but also has poor scene adaptation capability, and the adjusted threshold is still difficult to adapt to complex monitoring environment, false alarm / misreporting often occurs, which causes certain disturbance to the user, and even economic loss.

[0047] At present, some improved schemes supporting self-learning have slow response speed, lack real-time response according to false alarm severity and misreporting, and thus have poor dynamic adaptability. If various influencing factors are considered to improve the monitoring effect, multiple sensors need to be cooperated with each other, which is not realistic in actual practice considering the cost and other factors. Therefore, how to realize multi-scene adaptation, adaptive alarm threshold, and timely response of feedback learning on the basis of practicality to reduce the false alarm / misreporting rate has become the research direction and key of the technical personnel in the field.

[0048] To solve the above problems, the fire monitoring controller with adaptive alarm threshold adjustment function provided by the embodiments of the present application comprises, as shown in Figure 1

[0049] The data acquisition module acquires the visible light and infrared dual-spectrum video stream of the monitored area by using a dual-spectrum integrated camera, and detects the environmental temperature and relative humidity by using a temperature and humidity sensor.

[0050] The data processing module performs scene classification (different scenes are identified according to the video stream data), fire feature analysis (fire and smoke features are analyzed according to the video stream, the features are input into a fire analysis model, and a fire probability value is output), dynamic alarm threshold calculation (a more adaptive calculation is performed by combining a basic threshold matched according to the scene, environmental factors, and the fire probability value output by the model, which can effectively reduce false alarms and missed alarms), and alarm decision (determined according to the relationship between the fire probability value output by the fire analysis model and the dynamic alarm threshold).

[0051] The feedback learning module optimizes the fire analysis model (incremental training, reduces false alarms and missed alarms) according to false alarm and / or missed alarm data (confirmed by manual review).

[0052] The alarm execution module triggers sound and light alarms and fire equipment linkage according to the alarm decision.

[0053] Embodiment two

[0054] The working process of the fire monitoring controller of the embodiments of the present application is shown in Figure 2 First, the data acquisition module acquires the visible light and infrared dual-spectrum video stream and the environmental temperature and humidity data, and then the data processing module performs scene classification and fire feature analysis. The scene classification (the scene includes a warehouse, an industrial workshop, and an open field) is as follows:

[0055] For warehouse scene recognition, Canny edge detection is performed on the visible light video frame to extract all edge segments. Then, the vertical line segment density is calculated: the line segments with an angle <5° to the vertical direction are screened, and the number per unit area is counted. If the vertical line density is >40 lines / ㎡, the warehouse scene is determined.

[0056] For industrial workshop scene recognition, the infrared thermal imaging image is analyzed to identify the curved surface heat source (such as a reaction kettle) with a temperature >100°, calculate the area ratio of the curved surface heat source, count the number of high-temperature equipment (temperature >80°), and calculate the density per 100㎡. If the curved surface heat source ratio is >15% or the high-temperature equipment density is >5 pieces / 100㎡, the industrial workshop scene is determined.

[0057] ​For the recognition of outdoor scenes, the green pixels with hue H∈[60°, 150°] are extracted by HSV color space segmentation of the visible light image; the building outline is recognized by edge detection, and the building area ratio is calculated; if the green pixel ratio is > 30% and the building density is < 10%, it is determined as an outdoor scene.

[0058] Among them, the fire feature analysis, the specific process is as follows:

[0059] Based on the image of the video stream, the flame feature and the smoke feature can be extracted, and the occurrence of fire can be more accurately identified by comprehensive analysis of the two;

[0060] For the flame feature, in the visible light channel:

[0061] The pixel flicker frequency in the ROI region is analyzed, and whether it contains 3-18Hz component is verified by FFT;

[0062] The pixel ratio that meets the condition of RGB pixel R>G>B is counted, and whether the pixel ratio is >60% is judged;

[0063] The flame edge point set is extracted, and the circularity of the Fourier descriptor is calculated, if the circularity is <0.7, it is judged as non-spherical flame;

[0064] In the infrared channel: whether the thermal radiation gradient ΔT (temperature difference at the same position / time interval) is ≥5℃ / s is calculated;

[0065] For the smoke feature, in the visible light channel:

[0066] The motion target centroid is tracked, the rising trajectory straight line is fitted, and the slope angle is calculated, whether the slope angle is >55°;

[0067] The attenuation rate of the contrast of 5 consecutive frames of pixels is calculated, and whether the attenuation rate is >25% is judged;

[0068] In the infrared channel: track the low-temperature mass (<50°), calculate the motion rate by the centroid displacement, and judge whether the motion rate is >0.3m / s;

[0069] Finally, the 7 feature vectors of the flame feature and the smoke feature are input into the fire analysis model, and the fire probability value P is output fire The fire analysis model is MobileNetV3, the model is a pre-trained model, and the weight distribution of the 7 features in the model is dynamically adjusted according to the scene type (such as when the environmental humidity is >85%, the smoke feature weight is reduced by 20%-30% to avoid high-temperature water vapor interference; when in night mode, the flame thermal radiation gradient weight is increased by 25%-35%), preferably, one distribution method is shown in Table 1.

[0070] Table 1 Weight distribution of 7 features in fire analysis model

[0071] Warehouse Industrial plant Wilderness Visible light smoke rising trajectory slope 25％ 10％ 5％ Visible light pixel contrast decay rate 20％ 5％ 15％ Infrared low temperature clump movement rate 15％ 10％ 10％ Visible light flame flicker frequency 10％ 25％ 25％ Infrared flame thermal radiation gradient 10％ 30％ 25％ Pixel proportion of flame color R>G>B condition 10％ 15％ 15％ Flame edge irregularity 10％ 5％ 5％

[0072] Preferably, considering the influence of strong wind environment on smoke features, the application includes strong wind interference coping when analyzing smoke features, specifically:

[0073] Calculate the mean value of the centroid horizontal displacement through the continuous 3 frames of smoke motion vectors, and when the displacement is > 5 pixels / frame, determine that the horizontal wind speed is > 2 m / s;

[0074] When the horizontal wind speed is > 2 m / s, close the visible light channel smoke trajectory slope analysis (strong wind interference causes the trajectory to be flat), strengthen the infrared channel flame detection, and reduce the thermal radiation gradient ΔT threshold from 5℃ / s to 3℃ / s to accelerate the response.

[0075] Next, dynamic alarm threshold calculation is performed, specifically as follows:

[0076] According to the above scene classification result, match the basic threshold T b , specifically: warehouse scene 0.6 (high flammability of goods and large accumulation, high sensitivity is required), industrial workshop scene 0.7 (there are high temperature interference sources, etc., the sensitivity needs to be appropriately reduced), outdoor scene 0.65 (environmental interference factors are complex and have a greater impact, the sensitivity needs to be set to moderate);

[0077] Environmental compensation (considering the influence of environmental factors temperature and humidity):

[0078] When the relative humidity is > 85%, K env = 1.2 (suppress the influence of water vapor on smoke identification in high humidity environment);

[0079] When the ambient temperature is > 50℃ and it is an industrial workshop scene, K env = 1.15 (suppress high temperature interference);

[0080] When the ambient temperature is > 40℃ and it is an outdoor scene, K env = 0.9 (fire risk increases at high temperature); K env is the environmental compensation coefficient, the basic value is 1.0;

[0081] Dynamic alarm threshold T d : T d = T b × K env × (1-0.05× P fire ) (considering the influence of model output fire probability value, the higher the probability value, the greater the possibility of fire, so the threshold needs to be dynamically and appropriately reduced).

[0082] Considering the influence of light and other factors on feature analysis, the dynamic alarm threshold of the application is also adjusted according to the day and night mode, specifically:

[0083] The average brightness value of the visible light video frame is calculated, when the average brightness value > 180 lux, it is determined as daytime, K day = 1.1 (sufficient light, high false alarm rate, need to increase the alarm threshold);

[0084] When the average brightness value < 30 lux, it is determined as night, K day = 0.85 (need to improve sensitivity, reduce alarm threshold);

[0085] When 180 lux ≥ average brightness value ≥ 30 lux, K day = 1.0; K day is the day and night mode compensation factor;

[0086] The dynamic alarm threshold correction formula is: T d = T d × K day .

[0087] Then, according to the relationship between the fire probability value output by the fire analysis model and the dynamic alarm threshold, alarm decision and execution are made, specifically:

[0088] Three-level response: P fire ≥ dynamic alarm threshold, start sound and light alarm and decibel > 90 dB;

[0089] Secondary response: P fire ≥ 1.25 times the dynamic alarm threshold, on the basis of the three-level response, the sprinkler system is linked to handle;

[0090] Primary response: P fire ≥ 1.5 times the dynamic threshold, on the basis of the secondary response, start the smoke exhaust fan and emergency broadcast reminder.

[0091] In particular, for scenes identified as unclassified scenes other than the three scenes, the general mode is enabled, the basic alarm threshold is set to 0.6, the confidence of 7 features of flame and smoke is calculated, the feature weight is allocated according to the confidence using the entropy weight method, and the fire probability value is calculated according to the feature confidence and feature weight. At this time, the environmental compensation coefficient K env only considers the influence of high temperature water vapor, and needs to meet the conditions of relative humidity > 85%, thermal radiation gradient ΔT < 2 ℃ / s, color standard deviation > 0.25 to judge as high temperature water vapor interference, K env = 1.2, otherwise take the basic value 1.0; For wind interference and day and night mode compensation, feedback learning and other still use the above scene classification scheme.

[0092] Wherein, the flame flicker frequency in the fire is usually in the range of 3-18 Hz, but the flame main frequency is concentrated near 10 Hz, and the confidence calculation formula of the flame flicker frequency is: freis the flame flicker frequency;

[0093] The flame appears as R>G>B in the RGB color space, and the red channel dominates in the ideal case; according to IEEE-FS 2023, the proportion of pixels with R>G>B under visible light in an open fire is usually more than 50%, and about 75% is a typical value, so the confidence of the proportion of pixels with R>G>B is: ris the proportion of pixels with R>G>B;

[0094] The flame edge of the fire is irregular, and the circularity is low <0.7, and the circularity of the flame edge of the ideal fire is between 0.2-0.7, the closer to 0.7 (circular) the more it is not like a fire, and the lower the more irregular, the higher the possibility of fire; therefore, the confidence calculation formula is: ciris the circularity of the flame edge;

[0095] The temperature rises rapidly in the fire, and ΔT≥5℃ / s is a typical fire characteristic, so the confidence formula of the thermal radiation gradient ΔT is as follows: confΔT=1-e -0 . 5*ΔT ;

[0096] The smoke rising trajectory slope is greater than 55°, and the closer to 90° (vertical) the higher the confidence, and the confidence formula is: spis the smoke trajectory slope;

[0097] Real smoke causes pixel contrast attenuation >25%, and the greater the attenuation, the more likely it is smoke; according to NIST test data, the attenuation rate of real smoke is usually >30%; the confidence formula is: deis the pixel contrast attenuation rate;

[0098] The smoke movement rate is >0.3m / s, and the faster the rate the higher the confidence (but too fast may be air flow interference); according to the research, the real smoke rising rate is between 0.3-2m / s, and more than 2m / s may be strong wind interference; therefore, the confidence formula is:

[0099] v is the smoke movement rate (infrared low-temperature mass movement rate).

[0100] Example Three

[0101] The present application also optimizes the fire analysis model according to false alarm and / or missed alarm data, as follows:

[0102] False alarm processing flow, such as Figure 3The alarm event is manually reviewed, false alarms are marked, and flame and smoke feature values and temperature and humidity data are extracted.

[0103] wherein, when P fire ≥1.5 times T d and is not a fire (false alarm), it is marked as a serious false alarm, 10-20 times of counter samples are immediately generated, and fire analysis model retraining is completed within 72 hours;

[0104] Otherwise, it is marked as a common false alarm, flame and smoke feature data and temperature and humidity data are saved, when the cumulative number is greater than or equal to 2 times, fire analysis model optimization is triggered, the model uses transfer learning plus focal loss to optimize the classification layer; until the false alarm suppression rate is greater than 35%; the optimized model is continuously monitored and verified;

[0105] When the manual confirmation of a missed alarm is triggered, the scene basic threshold is immediately reduced by 0.1, and the lower limit of the basic threshold is 0.4; the corresponding feature weight in the fire analysis model (obtained according to the data annotation when manually reviewing) is increased by 15%, and the upper limit is 90%, the model incremental retraining is triggered, until the optimization effect reaches the missed alarm suppression rate is greater than 70%, the optimized model continues to be monitored and verified.

[0106] The basic principles of the present application are described above in conjunction with specific embodiments. It should be understood that the above-described specific details of the disclosure are only for the purpose of illustration and understanding, and are not intended to limit the scope of protection, and any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A fire monitoring controller with adaptive alarm threshold adjustment function, characterized in that: include: Data acquisition module, used to obtain visible light and infrared dual-spectrum video streams, as well as ambient temperature and humidity of the monitored area; The data processing module performs scene classification, fire feature analysis, dynamic alarm threshold calculation, and alarm decision-making based on video streams and environmental data; Feedback learning module to optimize fire analysis models based on false positive and / or false negative data; The alarm execution module triggers the sound and light alarm and fire-fighting equipment linkage according to the alarm decision.

2. A fire monitoring controller with adaptive alarm threshold adjustment function according to claim 1, characterized in that: The data processing module performs scene classification, specifically: Perform Canny edge detection on visible light video frames. If the vertical line density is greater than 40 lines / ㎡, it is determined to be a warehouse scene. Analyze infrared thermal images and identify an industrial workshop scene if the surface heat source ratio is greater than 15% or the density of high-temperature equipment is greater than 5 per 100 square meters. The visible light image is segmented in HSV color space. When the proportion of green pixels is greater than 30% and the building density is less than 10%, it is determined to be a wild scene.

3. A fire monitoring controller with adaptive alarm threshold adjustment function according to claim 1, characterized in that: The fire characteristics analysis of the data processing module is specifically as follows: For flame characteristics, in the visible light channel: determine whether the regional flicker frequency is within the 3-18 Hz range, whether the RGB pixel R>G>B ratio is greater than 60%, and whether the edge Fourier descriptor satisfies the circularity of <0.7; in the infrared channel: calculate whether the thermal radiation gradient ΔT is ≥5°C / s; For smoke characteristics, in the visible light channel: analyze whether the slope of the moving target's upward trajectory is greater than 55° and whether the contrast decay rate of the pixels in 5 consecutive frames is greater than 25%; in the infrared channel: track whether the centroid movement rate of the low-temperature mass is greater than 0.3m / s; Input flame characteristics and smoke characteristics into the fire analysis model and output a fire probability value P between 0 and 1. fire , the fire analysis model is MobileNetV3.

4. A fire monitoring controller with adaptive alarm threshold adjustment function as claimed in claims 2 and 3, characterized in that: The dynamic alarm threshold calculation of the data processing module is specifically as follows: Match the basic threshold T according to the scene classification results b , specifically: 0.6 for warehouse scene, 0.7 for industrial workshop scene, and 0.65 for outdoor scene; Environmental compensation: When relative humidity>85%, K env =1.2; When the ambient temperature is greater than 50°C and the scene is an industrial workshop, K env =1.15; when the ambient temperature is greater than 40°C and it is an outdoor scene, K env =0.9; the K env is the environmental compensation coefficient; Dynamic alarm threshold T d :T d =T b ×K env ×(1-0.05×P fire ).

5. A fire monitoring controller with adaptive alarm threshold adjustment function as claimed in claim 3, characterized in that: The smoke feature analysis includes the response to strong wind disturbances, specifically: The horizontal wind speed is calculated by the smoke motion vector of three consecutive frames. When the displacement is greater than 5 pixels / frame, the horizontal wind speed is determined to be greater than 2m / s. When the horizontal wind speed is greater than 2m / s, the visible light channel smoke trajectory slope analysis is turned off, the infrared channel flame detection is strengthened and the ΔT threshold is reduced to 3℃ / s.

6. A fire monitoring controller with adaptive alarm threshold adjustment function as claimed in claim 4, characterized in that: The dynamic alarm threshold is also adjusted according to the day and night mode, specifically: Calculate the average brightness value of the visible light video. When the average brightness value is greater than 180 lux, it is determined to be daytime. K day =1.1; when the average brightness value is less than 30 lux, it is judged to be night time, K day =0.85; when 180lux≥average brightness≥30lux, K day =1.0; the K day is the compensation factor for day and night mode; Dynamic alarm threshold correction: T d =T d ×K day .

7. A fire monitoring controller with adaptive alarm threshold adjustment function according to claim 6, characterized in that: The feedback learning module is specifically used to: After manually marking false alarms, the flame and smoke characteristic values ​​and temperature and humidity data were extracted; When P fire ≥1.5 times T d If the fire is not a fire, it is marked as a serious false alarm, and adversarial samples are generated immediately. The fire analysis model is retrained within 72 hours. Otherwise, it is marked as a common false alarm, and the flame and smoke characteristic data as well as the temperature and humidity data are saved. When the cumulative number of times is ≥ 2, the fire analysis model optimization is triggered; When a missed alarm is confirmed, the basic threshold of the scene is immediately triggered to be reduced by 0.1, and the lower limit of the basic threshold is 0.4; the corresponding feature weight in the fire analysis model is increased by 15%, and the upper limit is 90%.

8. A fire monitoring controller with adaptive alarm threshold adjustment function as claimed in claim 6, characterized in that: The alarm decision is specifically: Level 3 Response: P fire ≥ dynamic alarm threshold, the sound and light alarm is activated and the decibel is greater than 90dB; Secondary response: P fire When the value is ≥1.25 times the dynamic alarm threshold, the sprinkler system will be linked to handle the situation based on the three-level response; First level response: P fire When the value is ≥1.5 times the dynamic threshold, the smoke exhaust fan and emergency broadcast reminder will be started based on the secondary response.

Images