JP7580939B2 - Fire Detection Systems - Google Patents

Description

The present invention relates to a fire detection system.

High-ceiling spaces are often found in buildings, exhibition halls, arenas, and the like that have spaces spanning multiple floors. Examples of spaces spanning multiple floors include atriums (internal public open spaces) and stairwells. Regarding fire detectors that detect fires in high-ceiling spaces, Patent Document 1 describes that they are placed on the ceiling or walls.

JP 2019-194826 A

In high-ceiling spaces, fire detectors installed on the ceiling or walls may be placed farther away from the floor than in normal buildings. For this reason, particularly when a fire breaks out near the floor, it may take longer for the fire detector to detect the fire using heat, smoke, brightness temperature, infrared rays, etc. than it would for a fire detector in a normal building.

The present invention has been developed in consideration of the above circumstances, and its object is to enable early fire detection in high-ceiling spaces.

In order to achieve the above object, a fire detection system according to one aspect of the present invention comprises:
an estimation unit that estimates the probability that a combustion flame exists in an indoor space based on an image recognition result of a neural network for an image of the indoor space in a high ceiling area where the ceiling height is higher than that of a low ceiling area;
a notification unit that outputs a fire prediction signal when the probability becomes equal to or exceeds a fire threshold that is a criterion for determining whether a fire has occurred;
Equipped with.

The present invention makes it possible to detect fires early in high-ceiling spaces.

1 is an explanatory diagram showing a schematic configuration of a fire detection system according to an embodiment; FIG. 2 is a block diagram showing the electrical configuration of the monitoring device of FIG. 1 . 3 is a flowchart showing a procedure of processing performed by a preprocessing unit in FIG. 2 . 3A is an explanatory diagram showing the RGB color space of the image signals output by each surveillance camera in FIG. 1 , and FIG. 3B is an explanatory diagram showing the HSV color space of each image signal converted by the pre-processing unit in FIG. 2 . 2A is an explanatory diagram showing an original image of a monitoring area based on an image signal output by the monitoring camera of FIG. 1, and FIG. 2B is an explanatory diagram showing a mask image obtained by binarizing the original image. 5A is an explanatory diagram showing the maximum and minimum values of the XY coordinates identified for the portion corresponding to the combustion flame in the mask image of FIG. 5B, and FIG. 5B is an explanatory diagram showing a rectangular area circumscribing the combustion flame portion set from the XY coordinate values of FIG. 5A. 3 is a flowchart showing a procedure of a process performed by a fire estimation unit virtually constructed on the monitoring device of FIG. 2 . 3 is an explanatory diagram illustrating a schematic procedure of a process for extracting characteristics of a combustion flame performed by the fire estimation unit in FIG. 2 . FIG. 3 is an explanatory diagram showing a neural network used by the fire estimation unit in FIG. 2 to estimate the probability of the existence of a combustion flame; FIG. 5A and 5B are explanatory diagrams showing a fluctuating combustion flame present in the original image of FIG. 5A, and FIG. 5C is an explanatory diagram showing the shape change due to the fluctuating combustion flames of (a) and (b). FIG. 10 is an explanatory diagram showing the configuration of a long-short-term memory unit network used in the rear stage of the intermediate layer of the neural network of FIG. 9 . 3 is a flowchart showing a procedure of a process performed by a learning processing unit in FIG. 2 .

Below, an embodiment of the present invention will be described with reference to the drawings. The same or equivalent parts and components are given the same or equivalent reference numerals throughout the drawings. However, it should be noted that the drawings are schematic and may differ from the actual product. In addition, the drawings may of course include parts with different dimensional relationships and ratios.

The embodiments shown below are merely examples of devices that embody the technical ideas of this invention, and the technical ideas of this invention do not specify the materials, shapes, structures, arrangements, etc. of each component part to those described below. The technical ideas of this invention may be modified in various ways within the scope of the claims.

The fire detection system 1 of this embodiment shown in FIG. 1 detects the outbreak of a fire caused by a combustion flame in an indoor space 7 of a building having a high ceiling section 3 and a low ceiling section 5.

The high ceiling section 3 can be provided, for example, in the entrance area of a building. The space 7 of the high ceiling section 3 spans multiple floors of the building, and the ceiling height of the high ceiling section 3 is higher than the ceiling height of the low ceiling section 5 in which the space 7 is divided by floors.

The fire detection system 1 detects the occurrence of a fire in each alert zone, which is a space with a certain area or less. In the fire detection system 1 of this embodiment, a fire detector 11 and a surveillance camera 13 are placed in each space 7 in the high ceiling section 3 and the low ceiling section 5, each of which corresponds to a single alert zone, in order to detect fires in these spaces 7.

The fire detection system 1 also includes a receiver 15, a monitor panel 17, a monitoring device 19, and a display panel 21 located in the management office 9 within the building, in addition to the fire detectors 11 and surveillance cameras 13 in each space 7. The display panel 21 includes a message display section (not shown). The message display section can display, for example, the status of fire detection by the fire detection system 1.

In this embodiment, the automatic fire alarm system can be configured by a known automatic fire alarm system consisting of at least the fire detector 11, the receiver 15, other equipment transmitters, indicator lights, and local sound devices (none of which are shown), together with the surveillance camera 13, the monitor panel 17, and the display panel 21. Also, the receiver 15, the monitor panel 17, the monitoring device 19, and the display panel 21 can each be configured as an independent device. Alternatively, some or all of the receiver 15, the monitor panel 17, the monitoring device 19, and the display panel 21 can be configured as a single device.

Each fire detector 11 detects a fire in each space 7 in the high ceiling section 3 and the low ceiling section 5. According to the installation standards for automatic fire alarm systems, if the fire detector 11 is a flame detector, it is required to be installed so that it can monitor each part of the warning area at a height of 1.2 m from the floor. If the fire detector 11 is a spot-type heat detector or smoke detector, it is required to be installed so that it can monitor the floor. Therefore, each fire detector 11 can be installed on the ceiling or a wall near the ceiling of each space 7 that meets the above-mentioned installation standards. Also, the fire detector 11 can be a type that detects a fire by detecting infrared rays of a specific wavelength emitted by a combustion flame through carbon dioxide resonance radiation, and outputs a fire detection signal.

The wavelength of infrared light detected by the fire detector 11 can be, for example, in the 4 μm range. Infrared light with a wavelength in the 4 μm range is generally known as infrared light emitted by combustion flames through carbon dioxide resonance radiation. The fire detection signals output by the fire detectors 11 in each space 7, i.e., each security zone, are individually transmitted via communication line 23 to a receiver 15 in the management office 9 within the building.

Each surveillance camera 13 captures full-color images of the shooting range including each space 7 in the corresponding high ceiling section 3 and low ceiling section 5.

The surveillance cameras 13 in the low ceiling section 5, which has a low ceiling height, are installed on the ceiling of the low ceiling section 5 or on a wall near the ceiling, just like the fire detectors 11. The surveillance cameras 13 in the high ceiling section 3, which has a higher ceiling height than the low ceiling section 5, are installed on a wall at a specified height closer to the floor than the ceiling of the high ceiling section 3.

When the surveillance camera 13 in the high ceiling section 3 is installed on the ceiling or on a wall near the ceiling, it captures an image of the space 7 at an angle looking down from above onto the floor of the space 7 in the high ceiling section 3. With an image taken at this angle, for example, if a fire breaks out near the floor, it is difficult to recognize changes in shape due to the flickering of the burning flames of the fire in the image.

Therefore, the surveillance camera 13 in the high ceiling section 3 is installed on the wall surface at a predetermined height that is lower than the height of the ceiling or near the ceiling of the high ceiling section 3. By installing the surveillance camera 13 at this height, for example, if a fire breaks out near the floor surface, the surveillance camera 13 can capture an image of the space 7 in the high ceiling section 3 from an angle that makes it easy to recognize the shape and fluctuation of the burning flames occurring in the space 7 in the high ceiling section 3 on the image. The installation position of the surveillance camera 13 in the high ceiling section 3 corresponds to an imaging point at a predetermined height for capturing an image of the space 7 in the high ceiling section 3.

The image signals output by each of the surveillance cameras 13 in the high ceiling section 3 and the low ceiling section 5 are individually transmitted to the monitor panel 17 in the management office 9 within the building via the monitor line 25.

The fire detector 11 is not necessarily installed at a higher position than the surveillance camera 13. Some types of fire detectors 11 are installed closer to the floor. When using this type of fire detector 11, the fire detector 11 may be installed on a wall at a lower position than each of the surveillance cameras 13 in the high ceiling section 3 and low ceiling section 5.

In addition, if each space 7 in the high ceiling section 3 and the low ceiling section 5 is larger than the area permitted to be treated as a single security zone, each space 7 can be divided into multiple security zones, and fire detectors 11 and surveillance cameras 13 can be installed in each security zone.

When the receiver 15 receives a fire detection signal from each fire detector 11, it can display the occurrence of a fire using an indicator (not shown) or the like. For example, this indicator can be provided for each space 7 corresponding to each fire detector 11, i.e., for each alert area.

When an indicator is provided for each space 7, the receiver 15 can display the occurrence of a fire in each space 7 on an indicator corresponding to the fire detector 11 that output the received fire detection signal. In other words, the receiver 15 can notify the user of the occurrence of a fire and the space 7 in which the fire has occurred by displaying the occurrence of a fire on an indicator.

Furthermore, when the receiver 15 receives a fire detection signal from each fire detector 11 or a fire presumption signal from the monitoring device 19 described below, it outputs a fire report signal. The fire report signal is a signal that notifies other equipment that a fire has occurred and the space 7 in which the fire has occurred. The fire report signal output by the receiver 15 can generally be output to a report destination. Examples of report destinations include a central monitoring facility that manages all building equipment, smoke control and exhaust equipment for buildings with high ceilings 3 and low ceilings 5, air conditioning equipment, and fire extinguishing equipment that protects each space 7.

The monitor panel 17 can receive image signals from each surveillance camera 13 and display images of each space 7 on a monitor (not shown) or the like based on the received image signals. The monitor screen can be divided, for example, into multiple areas corresponding to each surveillance camera 13. When the monitor screen is divided into multiple areas corresponding to each surveillance camera 13, the monitor panel 17 can display, for example, in each area of the monitor, images of each space 7 captured by the corresponding surveillance camera 13 in real time.

The monitor board 17 can also distribute the image signals received from each surveillance camera 13 to multiple signals using a distributor (not shown). When distributing image signals to multiple signals, the monitor board 17 can, for example, supply one of the distributed image signals to a monitor and output the other to the surveillance device 19.

Image signals from each surveillance camera 13 are input to the surveillance device 19 via the monitor panel 17. The surveillance device 19 estimates the probability that a burning flame is present in the captured image of the space 7, separately from the detection of a fire in the space 7 by the fire detector 11.

2, the monitoring device 19 can be configured by a personal computer having a CPU 27, a ROM 29, a RAM 31, and an external storage device 33. The CPU 27 can virtually implement a pre-processing unit 35, a fire estimation unit 37, a post-processing unit 39, and a learning processing unit 41 on the monitoring device 19 by executing a program stored in the external storage device 33, for example.

Depending on the results of the processing performed by the fire estimation unit 37 for each image signal received from each surveillance camera 13, the monitoring device 19 can output a fire estimation signal for each space 7 corresponding to each surveillance camera 13 from the post-processing unit 39 to the receiver 15.

The pre-processing unit 35 performs the processing shown in FIG. 3 on the image signals from each surveillance camera 13 input via the monitor panel 17 in FIG. 1. First, the pre-processing unit 35 converts the color space of the input image signals from the surveillance cameras 13 from RGB shown in FIG. 4(a) to HSV shown in FIG. 4(b) (step S11).

HSV is a color space expressed by three components: hue, saturation (saturation/chroma), and value (value/brightness). The color space of the image signal handled in subsequent processing does not necessarily have to be HSV. However, in this embodiment, taking advantage of the ability to expand hue in a circular fashion, image signals in the HSV color space are handled in subsequent processing.

When converting the color space, noise components may be removed from the converted image signal. The noise components are, for example, prominent components that occur in each pixel value of the image signal as the color space is converted. The noise components can be removed, for example, by filtering the image signal using a Gaussian filter to suppress prominent pixel value changes between adjacent pixels.

Next, as shown in FIG. 3, the pre-processing unit 35 generates a mask image shown in FIG. 5(b) for extracting the combustion flame portion from the original image shown in FIG. 5(a) captured by the surveillance camera 13 shown in FIG. 1 (step S13). The mask image can be generated by binarizing the image signal using threshold values set for each of the H, S, and V values. The threshold values for each of the H, S, and V values can be set to values suitable for extracting the combustion flame obtained through experiments, for example.

Then, as shown in FIG. 3, the pre-processing unit 35 extracts the portion corresponding to the combustion flame in the mask image in FIG. 5(b) (step S15). At that time, the pre-processing unit 35, for example, identifies the maximum and minimum values in the XY coordinates of the combustion flame portion in the mask image as shown in FIG. 6(a). Then, as shown in FIG. 6(b), the pre-processing unit 35 sets a rectangular area circumscribing the combustion flame portion by two straight lines in the X-axis direction and two straight lines in the Y-axis direction that pass through each of the set coordinate values.

When multiple rectangular regions can be set for the mask image, the pre-processing unit 35 may, for example, automatically select the rectangular region with the largest area. Alternatively, one rectangular region may be selected from the multiple rectangular regions through a user confirmation operation. Furthermore, when there is no combustion flame portion that can be extracted from the mask image, the pre-processing unit 35 may omit each step from step S15 onwards and end the process shown in FIG. 3.

The pre-processing unit 35 also enlarges or reduces the image portion of the rectangular area in FIG. 6(b) in the original image in FIG. 5(a) captured by the surveillance camera 13 in FIG. 1, as shown in FIG. 3, to generate an estimation image with a size and aspect ratio suitable for processing by the fire estimation unit 37 (step S17). Then, the series of processes ends.

If the size and aspect ratio of the rectangular image do not match those of the estimation image even after enlargement or reduction, the pre-processing unit 35 enlarges or reduces the image portion of the rectangular image to a size that fits within the size of the estimation image. In this case, the pre-processing unit 35 can add dummy pixel values by zero padding to parts of the enlarged or reduced image where there are no pixels corresponding to pixels of the estimation image due to a mismatch in aspect ratio.

The pre-processing unit 35 executes the process according to the flowchart of FIG. 3 for each frame of the image signal input from each surveillance camera 13 in FIG. 1, and generates a video of the estimated image of each space 7 to be input to the fire estimation unit 37. The video of the estimated image can have a number of frames that is longer than the period of fluctuations specific to a combustion flame, for example.

The fire estimation unit 37 performs the process shown in FIG. 7 on the video of the estimation image of each space 7 generated by the pre-processing unit 35. First, the fire estimation unit 37 converts the color space from RGB shown in FIG. 4(a) to HSV shown in FIG. 4(b) for each frame of the video of the estimation image generated by the pre-processing unit 35 (step S31).

Next, the fire estimation unit 37 uses a neural network to generate feature maps of the shape and color of each combustion flame for each frame of the H, S, and V components of the estimation image video converted into the HSV color space, i.e., from the estimation image (step S33).

The above-mentioned processes performed by the fire estimation unit 37 in steps S31 and S33 can be performed according to the procedure shown in FIG. 8. In particular, the fire estimation unit 37 can perform the process of generating the feature map in step S33 in FIG. 7 for each of the H, S, and V components by, for example, a combination of a convolution process using a filter and a compression process by pooling.

Then, the fire estimation unit 37 uses a neural network to estimate the probability that a combustion flame is present in the video of the estimation image based on the feature maps for each of the H, S, and V components generated in step S33 (step S35). Then, the series of processes ends.

The process performed by the fire estimation unit 37 in step S35 can use, for example, the neural network in FIG. 9. This neural network has multiple intermediate layers between the input layer and the output layer. The first part of the intermediate layers is a layer for recognizing the shape, color, etc. of objects present in the estimation image from a feature map. The second part of the intermediate layers is a layer for recognizing movement corresponding to the fluctuation of a burning flame from the change between frames in the shape and color of the object recognized in the first part.

In other words, fluctuations occur in a combustion flame. Fluctuations are changes in the strength of a combustion flame over time. For example, if the fluctuations occurring in a combustion flame are "1/f fluctuations," when the changes in the strength of a combustion flame are decomposed into a frequency spectrum, the strength (amplitude) P of the change in strength at each frequency changes with respect to the change in frequency f in the relationship P = a/f (a is a positive constant).

Therefore, when a fire breaks out in space 7, the probability that a fluctuating combustion flame exists in space 7 can be estimated based on the image recognition results of video images of space 7 taken by surveillance camera 13 over a specified period of time. The neural network of the fire estimation unit 37 estimates the probability that a fluctuating combustion flame exists in space 7 using the images taken by surveillance camera 13.

The input layer in FIG. 9 receives the values of each square in a column of the feature map generated by the fire estimation unit 37 for each of the H, S, and V components in step S33 in FIG. 7. In the front part of the intermediate layer in FIG. 9, the fire estimation unit 37 can recognize the shape, color, etc. of objects present in the estimation image by using a neural network formed by repeating a fully connected layer. Note that a non-fully connected layer or a dropout layer may be present in the front part of the intermediate layer.

In the latter part of the intermediate layer in FIG. 9, for example, when a combustion flame F as shown in FIG. 10(a) is present in each frame of the estimation image, the fire estimation unit 37 can recognize whether the combustion flame F is a fluctuating combustion flame F. In the output layer in FIG. 9, values indicating the probability that a fluctuating combustion flame F exists or does not exist in the video of the estimation image are output as real numbers between 0 and 1. Note that in this embodiment, the sum of the two values output to the output layer is always 1.

For example, a network of long short-term memory units (LSTM) can be used in the latter part of the intermediate layer in FIG. 9. The long short-term memory unit can perform processing to recognize the fluctuating combustion flame F using the values of the feature map generated by the fire estimation unit 37 from each estimation image from several frames ago to the present. The long short-term memory unit can be configured, for example, as shown in the explanatory diagram in FIG. 11.

In FIG. 11, for ease of explanation, the long- and short-term memory units that transmit information for the input xt of the current frame t are shown with solid lines, and the long- and short-term memory units that obtain the output ht-1 of the previous frame t-1 are shown with dashed lines. The long- and short-term memory units shown with dashed lines have the same configuration as the long- and short-term memory units shown with solid lines.

First, the long-term and short-term memory units shown by solid lines sum the input xt for the current frame t and the short-term output ht-1 for the previous frame t-1, and distribute this to the three gates for forgetting, input, and output, and to the hyperbolic tangent tanh.

The forget gate uses a sigmoid function σ to set a value ft as a bias for the long-term output ct-1 of the previous frame t-1, based on the sum of the short-term output ht-1 of the previous frame t-1 and the input xt for the current frame t. The value ft is a real value between 0 and 1. The value ft is multiplied by the long-term output ct-1 of the previous frame t-1.

In the input gate, a value it is set using a sigmoid function σ to determine whether to update the short-term output ht-1 or the input xt of the previous frame t-1. The value it is either the short-term output ht-1 or the input xt, or the sum of both. In addition, a vector for the value it, i.e., a value zt including a positive and negative sign, is set using the hyperbolic tangent tanh. The value zt is a real value between -1 and 1. The value zt is multiplied by the value it, and then added to the product ft x (ct-1) of the value ft and the long-term output ct-1 of the previous frame t-1.

The value ft×(ct-1)+it×zt, which is the sum of the product ft×(ct-1) of the number ft and the long-term output ct-1 of the previous frame t-1 and the product it×zt of the number it and the number zt, becomes the long-term output ct of the current frame t. The value ft×(ct-1)+it×zt becomes the input to the hyperbolic tangent tanh for the output gate, which is provided separately from the input gate. The hyperbolic tangent tanh for the output frame is used to set a vector for the short-term output ht of the current frame t, i.e., a numerical value including positive and negative signs.

At the output gate, a sigmoid function σ is used to set a value ot, which is used as a bias to define the portion to be output from the sum of the short-term output ht-1 and the input xt of the previous frame t-1. The value ot is a real value whose maximum value is the sum of the short-term output ht-1 and the input xt of the previous frame t-1. The value ot is multiplied by the value set by the hyperbolic tangent tanh for the output frame. This multiplied value becomes the short-term output ht of the current frame t.

In the long- and short-term memory unit described above, the sigmoid functions σ of the three gates for forgetting, input, and output, and the hyperbolic tangent tanh for the input gate are the subjects of learning.

The long- and short-term memory unit described above allows the fire estimation unit 37 in FIG. 2 to estimate the probability that a combustion flame F with long-term dependency fluctuations exists in the video of the estimation image, while preventing the number of solutions found by information processing from increasing and making it impossible to find an optimal solution. For example, when the combustion flame F shown in the hatched portion of FIG. 10(a) changes over time to the shape shown in the hatched portion of FIG. 10(b), the fire estimation unit 37 can estimate the probability that the combustion flame F is fluctuating. At this time, the fire estimation unit 37 estimates the probability that a combustion flame F with fluctuating exists by estimating the probability that the change in the external shape of the combustion flame F shown by the arrow in FIG. 10(c), which shows the combustion flame F before and after the shape change, corresponds to fluctuating.

Then, in step S35 of FIG. 7, the fire estimation unit 37 acquires the value indicating the probability that a burning flame F exists from the two real values output from the output layer of the neural network in FIG. 9 as a fire probability value P. The acquired fire probability value P becomes a value indicating the estimation result by the fire estimation unit 37 of the probability that a burning flame F exists in the video of the estimation image.

When the fire probability value P for each space 7 acquired by the fire estimation unit 37 is equal to or greater than the fire threshold, which is the criterion for determining whether a fire has occurred in the space 7, the post-processing unit 39 outputs a fire estimation signal for the space 7 in which the fire probability value P is equal to or greater than the fire threshold to the receiver 15.

The receiver 15 that receives the fire estimation signal from the post-processing unit 39 can display on an indicator the occurrence of a fire and the location of the fire, etc., based on which space 7 the received fire estimation signal corresponds to. In addition, the receiver 15 that receives the fire estimation signal outputs a fire notification signal to the notification destination, just as when it receives a fire detection signal from the fire detector 11.

The learning processing unit 41 executes the process shown in FIG. 12 for the neural network that executes the procedure shown in FIG. 8 and the neural network in FIG. 9 of the fire estimation unit 37 corresponding to the fire detector 11 and surveillance camera 13 in each space 7. First, the learning processing unit 41 checks whether a learning request for the neural network has been input (step S51).

The learning request may be input by a user operation. Alternatively, the learning request may be input automatically when the consistency between the fire occurrence state in the space 7 or the output state of the fire detection signal from the fire detector 11 and the fire probability value P estimated by the fire estimation unit 37 falls to a level below a standard.

Here, the fire detector 11 is manufactured so that it will not fail to output a fire detection signal when a fire has occurred, even if it may erroneously issue a non-fire alarm when there is no fire. For this reason, when the fire detector 11 is not outputting a fire detection signal, it means that no fire has occurred in the space 7. Therefore, if the fire detector 11 does not output a fire detection signal even after a certain amount of time has passed after the fire probability value P estimated by the fire estimation unit 37 becomes equal to or greater than the fire threshold, it is considered that the estimation result of the fire probability value P by the fire estimation unit 37 is inappropriate.

Conversely, if the fire estimation unit 37 continues to estimate a fire probability value P that is below the fire threshold even though the fire detector 11 is outputting a fire detection signal, the estimation result of the fire probability value P by the fire estimation unit 37 is considered to be inappropriate.

When such a situation occurs, it is necessary to execute a learning process for the neural network of the fire estimation unit 37. Therefore, for example, when the frequency with which the above-mentioned situation occurs reaches a certain level or more, a learning request generated by logic may be input to the learning processing unit 41, or a learning request may be input to the learning processing unit 41 by a user operation.

The learning data used to train the neural network can be, for example, video footage taken by the surveillance cameras 13 in the high ceiling section 3 and low ceiling section 5 of each space 7 when normal heptane (n-heptane) or alcohol starts to burn in the pan. The substance burned in the pan may be a flammable substance other than normal heptane (n-heptane) or alcohol.

The learning data used to train the neural network may be, in addition to the above images, a video of a combustion test in a space (not shown) of a fire test facility that simulates each space 7 in the high ceiling section 3 and the low ceiling section 5. In this video, the moment when the combustion of the flammable material in the fire pan is started in the location that represents each space 7 in the fire test facility is captured by a camera in the fire test facility that represents the surveillance camera 13 in each space 7 in the high ceiling section 3 and the low ceiling section 5.

By using the above-mentioned video as learning data, the neural network of the fire estimation unit 37 can efficiently learn the characteristics of the shape change due to the fluctuation of the burning flame F at the early stage of the outbreak of a fire in each space 7 of the high ceiling section 3 and the low ceiling section 5. Therefore, the accuracy with which the fire estimation unit 37 after learning estimates a fire probability value P equal to or greater than the fire threshold at the early stage of the outbreak of a fire can be increased by the learning of the neural network.

The above-mentioned learning data can be stored in a learning data storage section (not shown) of the external storage device 33, for example, in association with a target value of the fire probability value P at the time of fire occurrence.

If a learning request has not been input to the learning processing unit 41 (NO in step S51), the series of processes is terminated. If a learning request has been input to the learning processing unit 41, the learning process of the neural network of the fire estimation unit 37 is executed using the learning data stored in the learning data storage unit of the external storage device 33 (step S53).

This learning process can be performed, for example, using the backpropagation method. In this case, the learning processing unit 41 reads each piece of learning data from the external storage device 33. The learning processing unit 41 then calculates the difference between the fire probability value P of each piece of learning data and the ideal value, and inputs this as a teacher signal to the output layer of the neural network.

Then, when the learning processing unit 41 has executed the learning process of step S53 using all the learning data in the external storage device 33 (YES in step S55), the series of processes ends.

The learning data in the external storage device 33 can be used when performing follow-up training of the neural network after the operation of the fire detection system 1. The learning data in the external storage device 33 may also be used for initial training of the neural network before the operation of the fire detection system 1.

As is clear from the above description, in this embodiment, the fire estimation unit 37 of the monitoring device 19 realizes an estimation unit that estimates the probability that a burning flame F exists in the space 7. In addition, the post-processing unit 39 of the monitoring device 19 realizes a notification unit that outputs a fire estimation signal to the receiver 15 when the fire probability value P becomes equal to or greater than the fire threshold, which is the criterion for determining whether a fire has occurred. Note that the fire threshold can be a real number between 0 and 1.

Furthermore, in this embodiment, the receiver 15 realizes a fire notification unit that outputs a fire notification signal when at least one of a fire detection signal from the fire detector 11 and a fire estimation signal from the post-processing unit 39 of the monitoring device 19 is input.

In the fire detection system 1 of this embodiment described above, the fire estimation unit 37 of the monitoring device 19 estimates a fire probability value P by performing image recognition on video images of the space 7 in the high ceiling section 3 and low ceiling section 5 captured by the monitoring camera 13. If the fire probability value P is equal to or greater than the fire threshold, the monitoring device 19 outputs a fire estimation signal to the receiver 15, assuming that a fire has occurred in the space 7 in the high ceiling section 3 or low ceiling section 5 corresponding to the monitoring camera 13. The receiver 15 that receives the fire estimation signal outputs a fire notification signal to the notification destination, even if it does not receive a fire detection signal from the fire detector 11 in the space 7 where the fire probability value P is equal to or greater than the fire threshold.

For this reason, even if the fire detector 11 in the high ceiling section 3 takes a long time to detect infrared rays of a specific wavelength caused by a fire, if there is a high possibility of a fire breaking out, the receiver 15 can output a fire notification signal to the notification destination before receiving a fire detection signal from the fire detector 11. This makes it possible to detect a fire early, especially in the space 7 of the high ceiling section 3 where the ceiling height is high.

Even in the low ceiling area 5, if it takes a long time for the fire detector 11 to detect infrared rays of a specific wavelength, if there is a high possibility of a fire occurring, the receiver 15 can output a fire notification signal to the monitoring device 19 as the notification destination before receiving a fire detection signal from the fire detector 11.

In addition, as the image of the space 7 in the high ceiling section 3 and the low ceiling section 5, a thermal distribution image obtained by thermography may be used instead of an image captured by the surveillance camera 13. In this case, since the thermal distribution image is a monochrome image and has only a single gradation component, it is possible to binarize the pixel signal using a threshold value without converting the color space. Therefore, the preprocessing performed by the preprocessing unit 35 on the thermal distribution image can be one that omits the processing of step S11 in FIG. 3.

In addition, in the fire detection system 1 of this embodiment, the detector that outputs the fire detection signal is an infrared fire detector 11. However, the detector is not limited to one that detects fires using only infrared rays, and may be one that detects fires using, for example, ultraviolet rays, both ultraviolet and infrared rays, heat, smoke, brightness temperature, etc.

Furthermore, in the fire detection system 1 of this embodiment, a network of long- and short-term memory units suitable for processing to recognize a fluctuating combustion flame F is used in the neural network of the fire estimation unit 37.

However, the neural network of the fire estimation unit 37 does not have to be capable of recognizing the fluctuations of the combustion flame F, so long as it can recognize the combustion flame F from the image captured by the surveillance camera 13. For example, the neural network of the fire estimation unit 37 may be configured only with a network such as a Convolutional Neural Network (CNN) or a Recurrent Neural Network (RNN). In that case, the learning data for the neural network may be still images instead of video.

In addition, in the fire detection system 1 of this embodiment, surveillance cameras 13 are installed in both the high ceiling section 3 and the low ceiling section 5, and images of each space 7 are captured and used by the fire estimation unit 37 to estimate the fire probability value P.

However, in the low ceiling section 5, where the ceiling height is lower than in the high ceiling section 3, it is considered that it will not take as long for the fire detector 11 on or near the ceiling to detect infrared rays of a specific wavelength and output a fire detection signal as it does for the fire detector 11 in the high ceiling section 3. Therefore, it may be possible to omit installing a surveillance camera 13 in the low ceiling section 5 for the purpose of early detection of a fire outbreak.

Furthermore, in the fire detection system 1 of this embodiment, the fire detector 11 is placed in the high ceiling section 3 and used in conjunction with the surveillance camera 13. Then, when either a fire detection signal from the fire detector 11 or a fire estimation signal from the post-processing section 39 of the surveillance device 19 is input to the receiver 15, the receiver 15 outputs a fire notification signal to the notification destination.

However, the fire detector 11 and receiver 15 may be omitted, and when the post-processing unit 39 of the monitoring device 19 outputs a fire estimation signal, the output may trigger the start of response to a fire outbreak in the space 7 of the high ceiling unit 3.

The response that is initiated in response to the output of a fire prediction signal can be, for example, a display similar to the display of a fire occurrence by alert area that is performed by the receiver 15 using an indicator (not shown) when a fire detection signal is received from the fire detector 11.

Furthermore, the fire detector 11 and the receiver 15 may not be omitted, and a fire detection system with a different configuration may be used. In this case, for example, when a fire estimation signal from the post-processing unit 39 of the monitoring device 19 is input to the receiver 15, an alarm indicating a fire in the high ceiling section 3 may be output on the indicator of the receiver 15 or on the message display unit of the display panel 21 prior to the fire notification signal. The fire notification signal from the receiver 15 to the notification destination may be output when a fire detection signal from the fire detector 11 is input to the receiver 15 after the alarm is output.

[Invention disclosed by the embodiment and its effects]
The above-described embodiment discloses the invention in the following aspects.

First, as a first aspect of the invention,
a fire detector that outputs a fire detection signal when detecting a fire in an indoor space having a high ceiling, the ceiling of which is higher than that of a low ceiling;
an estimation unit that estimates a probability that a combustion flame exists in the indoor space based on an image recognition result of a neural network for the image of the indoor space;
a notification unit that outputs a fire prediction signal when the probability becomes equal to or exceeds a fire threshold that is a criterion for determining whether a fire has occurred;
a fire notification unit that outputs a fire notification signal to a notification destination of a fire in the indoor space when at least one of the fire detection signal and the fire estimation signal is output;
A fire detection system is disclosed comprising:

According to the first aspect of the invention, the combustion flame that occurs during a fire emits infrared rays of a specific wavelength due to carbon dioxide resonance radiation. Therefore, when a fire breaks out in an indoor space with a high ceiling, a fire detector can, for example, detect infrared rays of a specific wavelength emitted from the combustion flame and detect the outbreak of the fire.

Furthermore, a combustion flame not only changes the state of infrared rays, but can also change the state of other elements, such as ultraviolet rays, both ultraviolet and infrared rays, heat, smoke, and brightness temperature. Therefore, when a fire breaks out in an indoor space with a high ceiling, a fire detector can detect the start of a fire by detecting, for example, changes in ultraviolet rays, both ultraviolet and infrared rays, heat, smoke, brightness temperature, etc., whose states change due to the outbreak of a combustion flame, or by detecting the presence of these elements themselves.

Furthermore, when a fire breaks out in an indoor space with a high ceiling, a burning flame is captured in an image of the indoor space with a high ceiling captured by a surveillance camera. Therefore, the estimation unit can estimate the probability that a burning flame is present in the indoor space with a high ceiling, for example, based on the image recognition results of a neural network for images of the indoor space with a high ceiling taken over a specified period of time.

In high-ceiling areas where the ceiling height is higher than in low-ceiling areas, fire detectors installed on the ceiling or walls may be placed at a greater distance from the floor than in normal buildings. For this reason, particularly when a fire breaks out near the floor, it may take longer for the fire detector to detect the fire using infrared rays, ultraviolet rays, both ultraviolet and infrared rays, heat, smoke, brightness temperature, etc. than it would for a fire detector in a low-ceiling area.

On the other hand, images of indoor spaces with high ceilings show the presence of a burning flame, even if the flame is near the floor, regardless of the height difference between the ceiling or the area near the ceiling. Therefore, the probability of the presence of a burning flame estimated by the estimation unit is high if a fire has broken out in an indoor space with a high ceiling.

Therefore, even if the fire detector does not output a fire detection signal, if the estimation unit estimates the probability that a combustion flame exists in the indoor space of a high ceiling area, it can be inferred that there is a high possibility that a fire has broken out in the high ceiling area.

Therefore, the value of the probability of the existence of a burning flame estimated by the estimation unit is defined as a standard value for the occurrence of a fire by a fire threshold. If the probability estimated by the estimation unit is equal to or greater than the fire threshold, the notification unit outputs a fire estimation signal. When the notification unit outputs a fire estimation signal, even if the fire detector does not output a fire detection signal, it can be assumed that a fire has occurred in the indoor space with a high ceiling, and it can be used as a trigger to start dealing with the occurrence of a fire in the indoor space with a high ceiling. This makes it possible to detect fires early in indoor spaces with high ceilings.

In addition, as an invention according to a second aspect, a fire detection system is disclosed that further includes a fire detector that outputs a fire detection signal when a fire is detected in the indoor space, and a fire notification unit that outputs a fire notification signal to a notification destination of a fire in the indoor space when at least one of the fire detection signal and the fire presumption signal is output.

According to the invention according to the second aspect, when at least one of the following occurs: the fire detector in the high ceiling outputs a fire detection signal, and the notification unit outputs a fire estimation signal based on the estimation result of the estimation unit, the fire notification unit outputs a fire notification signal to the destination of the fire notification. Therefore, the occurrence of a fire in an indoor space in a high ceiling can be more reliably detected by detecting the element itself or a change in the state of an element caused by a burning flame, and detecting the burning flame by an image.

Furthermore, as a third aspect of the invention, a fire detection system is disclosed in which the image is an image of the indoor space taken from a shooting point at a predetermined height lower than the ceiling height of the high ceiling section, and the fire detector is installed on the ceiling of the high ceiling section.

According to the invention of the third aspect, when an indoor space is photographed from the ceiling of a high ceiling or near the ceiling, an image of the indoor space is taken at an angle looking down on the floor of the indoor space in the high ceiling from above. With an image taken at this angle, for example, if a fire breaks out near the floor of the high ceiling, it is difficult to recognize changes in shape due to the flickering of the burning flames of the fire in the image.

Therefore, images of the indoor space in a high ceiling are captured from a shooting point at a specified height that is lower than the height of the ceiling or near the ceiling of the high ceiling. By capturing images of the indoor space in a high ceiling from this height, even if a fire breaks out near the floor of the high ceiling, for example, it is possible to capture images of the indoor space in a high ceiling from an angle that makes it easy to recognize the shape and flicker of the burning flames of the fire on the image.

In addition, as a fourth aspect of the invention, a neural network is disclosed in which the neural network has a long short-term memory unit (LSTM) network in the intermediate layer that performs image recognition of fluctuations in a combustion flame over a predetermined period of time.

According to the fourth aspect of the invention, a fluctuating combustion flame present in an indoor space with a high ceiling is extracted in a long-term/short-term memory unit in the intermediate layer of a neural network using the most recent images from a specified period of time in the past. Therefore, a fluctuating combustion flame, which is a time-varying element, can be extracted with high accuracy by image recognition and used to estimate the probability of the existence of a combustion flame by the estimation unit.

In addition, as a fifth aspect of the invention, a fire detection system is disclosed in which the neural network has a network that uses as learning data images taken of the indoor space or a space simulating the indoor space at the time when the combustion of the flammable material in the fire pan begins.

According to the invention of the fifth aspect, the neural network of the fire estimation unit can be efficiently trained on the characteristics of the shape change caused by the fluctuation of the burning flame at the early stage of the outbreak of a fire in an indoor space with a high ceiling. Therefore, the accuracy with which the trained fire estimation unit estimates the probability of a fire exceeding the fire threshold at the early stage of the outbreak of a fire in an indoor space with a high ceiling can be increased by training the neural network.

Here, the combustible substance may be normal heptane (n-heptane) or alcohol.

1 Fire detection system 3 High ceiling area 5 Low ceiling area 7 Space 9 Management office 11 Fire detector (detector)
13 Surveillance camera 15 Receiver (fire alarm unit)
17 Monitor panel 19 Monitoring device 21 Display panel 23 Communication line 25 Monitor line 27 CPU
29 ROM
31 RAM
33 External storage device 35 Pre-processing unit 37 Fire estimation unit 39 Post-processing unit 41 Learning processing unit F Combustion flame P Fire probability value

Claims (4)

a pre-processing unit that converts a color space of an image signal of an indoor space in a high-ceilinged area, which has a higher ceiling height than a low-ceilinged area, from RGB to HSV, binarizes the image signal using image signal thresholds set for each of the H, S, and V values, thereby generating a mask image, and enlarges or reduces an image portion of a rectangular area circumscribing a burning flame portion in the mask image, thereby generating an estimation image of a size and aspect ratio suitable for image recognition processing using a neural network of the fire estimation unit;
an estimation unit that estimates a probability that a combustion flame exists in the indoor space based on an image recognition result of the neural network for the estimation image;
a notification unit that outputs a fire prediction signal when the probability becomes equal to or exceeds a fire threshold that is a criterion for determining whether a fire has occurred;
A fire detection system comprising: The fire detection system of claim 1 further comprises a fire detector that outputs a fire detection signal when it detects a fire in the indoor space, and a fire notification unit that outputs a fire notification signal to a notification destination of a fire in the indoor space when at least one of the fire detection signal and the fire presumption signal is output. 3. The fire detection system of claim 2, wherein the estimation image is an image of the indoor space photographed from a shooting point at a predetermined height lower than the ceiling height of the high ceiling section, and the fire detector is installed on a ceiling of the high ceiling section. The fire detection system according to any one of claims 1 to 3, wherein the neural network has a long short-term memory (LSTM) network in an intermediate layer that performs image recognition of fluctuations in a combustion flame over a predetermined period of time.

Images