Fire detection method

Abstract

By observing the light emitted from the monitoring area, it enables highly accurate and reliable fire detection. [Solution] The fire detection system includes a fire detection unit 18 that detects a fire in a monitored area based on a first fire judgment result obtained by performing a predetermined fire judgment process on observation data collected by observing an object under observation in the monitored area, and a second fire judgment result calculated by performing a predetermined fire judgment process based on machine learning on the observation data; an observation unit 12 that observes the emission of light from the monitored area; a first processing unit 14 that obtains a first fire judgment result by performing a predetermined fire judgment process on observation data observed and collected by the observation unit 12; and a second processing unit 16 that calculates a second fire judgment result by performing a predetermined fire judgment process based on machine learning on the observation data, thereby obtaining the first and second fire judgment results based on observation data in first, second and third wavelength bands.

Description

Technical Field

[0001] The present invention relates to a fire detection method for observing an observation target in a monitoring area to detect a fire.

Background Art

[0002] As a fire detection system for observing an observation target in a monitoring area to detect a fire, for example, there is a flame detection device that observes light radiated from inside a road tunnel for automobiles and detects a fire when it is determined that light peculiar to a flame has occurred.

[0003] Such a flame detection device observes the radiation of light from a monitoring area and detects a fire by observing light (infrared energy) radiated from a flame accompanying the occurrence of a fire (Patent Documents 1 to 3). For example, a two-wavelength type flame detection device observes light near 4.5 μm, which is a CO2 resonance radiation wavelength band radiated from a flame, and light near, for example, 5.0 μm outside the CO2 resonance radiation band, and the relative ratio of the observation data of the light in these two wavelength bands is a value indicating the occurrence of a flame, and when the observed waveform of the light in the CO2 resonance radiation band has a flickering frequency component peculiar to a flame, it is determined as a fire, and for example, a fire signal is transmitted to a disaster prevention receiving panel to cause a fire alarm operation.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Patent Document 2

Patent Document 3

Patent Document 4

Patent Document 5

Summary of the Invention

Problems to be Solved by the Invention

[0005] The flame detection devices installed inside the tunnel are normally exposed to various wavelengths and intensities of light radiation from sources such as the headlights of vehicles passing on the road, flashing or blinking warning lights of emergency vehicles, lighting fixtures inside the tunnel, sunlight if installed near the tunnel entrance / exit, and workers (human bodies) during maintenance and inspection. However, they are equipped with an identification function to distinguish between this ambient light and the light emitted from flames associated with a fire, thereby enhancing their fire (in this case, fire flame) identification performance.

[0006] Furthermore, including such identification functions, the fire detection system can, for example, self-diagnose whether it is functioning correctly, and if it is diagnosed as not functioning correctly, it outputs a signal indicating a malfunction or failure and notifies the public from the fire prevention receiver panel.

[0007] However, even when diagnosed as normal, it was difficult to completely eliminate the possibility that ambient light similar to the light emitted from a fire flame could be mistaken for a fire, resulting in a false fire alarm being issued from the fire alarm control panel.

[0008] In such cases, until it is confirmed that it is not a fire alarm, entry prohibition warnings must be issued using alarm display boards or similar equipment to prohibit vehicles from passing through the tunnel. A manager must then go to the site (where the flame detection device that was judged to be a fire is installed) to confirm, which takes time and effort before tunnel passage can be resumed and can have significant consequences such as causing traffic congestion.

[0009] In addition to these "false fire alarms," ​​there are also "false alarms" that occur when the flame detection device malfunctions, such as by mistakenly identifying a fire due to internal factors, and a fire alarm is issued even though there is no fire. False alarms have the same impact as false fire alarms.

[0010] One cause of false alarms is electrical spontaneous noise generated due to component failure or deterioration in the electrical circuitry that makes up the flame detection device. In addition, the generation of false fire signals due to faulty wiring or insulation deterioration associated with the flame detection device has also been reported. While such failures, deterioration, and malfunctions are diagnosed during the self-diagnosis of the flame detection device, if a circuit failure or deterioration occurs during the interval between diagnoses, which are performed at predetermined intervals, that is, if the circuit was normal in the previous diagnosis but a failure or deterioration occurs during the period until the next diagnosis, the resulting spontaneous noise may spread to the light receiving circuit, be processed as an input light signal, and be mistaken for light emitted from a flame.

[0011] For convenience, in the following explanation, we may not distinguish between false alarms and non-fire alarms, and instead refer to both collectively as "false alarms."

[0012] The present invention aims to provide a fire detection method with improved fire identification performance compared to conventional methods, thereby reducing false alarms and non-fire alarms. [Means for solving the problem]

[0013] (Fire detection system 1) The present invention is a fire detection system, The system is characterized by detecting a fire based on a first fire judgment result obtained by performing a predetermined fire judgment process on observation data of the target being monitored in the monitoring area, and a second fire judgment result calculated by performing a predetermined fire judgment process based on machine learning on the observation data.

[0014] Here, "observation data" refers to time-series data obtained at predetermined intervals by sampling and A / D converting the observed value signals (analog signals) obtained from observations of the observed object in the monitoring area at a predetermined frequency. Furthermore, "fire judgment processing" refers to an algorithm (essentially a computer program) for determining whether or not predetermined conditions for determining a fire are met, and "first fire judgment result" is the result of that judgment.

[0015] Furthermore, the "fire judgment process based on machine learning processing" is an algorithm (essentially a computer program) for predicting whether the degree of fire likelihood is high or low, and the "second fire judgment result" is the result of quantitatively representing and evaluating the prediction result and making a fire judgment based on this. Here, "machine learning" refers to inputting unknown data into a mathematical model constructed from a dataset that includes both inputs and outputs and predicting the output. In relation to the "machine learning processing" of the present invention, this means inputting new observation data into a mathematical model (learning model unit) constructed (learned) from learning data that includes both previously collected observation data and the second fire judgment result based on said previously collected observation data, deriving a prediction result of the degree of fire likelihood corresponding to the new observation data, and obtaining a second fire judgment result based on this. The machine learning in this embodiment will be explained in detail later.

[0016] (Fire detection system 2) Furthermore, the present invention relates to a fire detection system, The system is characterized by detecting a fire based on a first fire judgment result obtained by performing a predetermined fire judgment process on observation data collected by observing light radiation from a monitoring area, and a second fire judgment result calculated by performing a predetermined machine learning-based fire judgment process on the observation data.

[0017] (Fire detection system 3) Furthermore, the present invention relates to a fire detection system, An observation unit that observes light emission from the monitoring area, A first processing unit performs a predetermined fire determination process on observation data collected by the observation unit to obtain a first fire determination result, A second processing unit performs a fire judgment process based on predetermined machine learning processing on the observation data to calculate a second fire judgment result, A fire detection unit detects a fire in the monitored area based on the first fire determination result and the second fire determination result, It is characterized by having the following features.

[0018] (Second fire judgment result) The second fire determination result is a value within the range from 0 when there is no fire to 1 when there is a fire. That is, the second fire determination result based on the observation data during a fire is 1 (or 100%), and the second fire determination result based on the observation data when there is no fire is 0 (or 0%). According to the observation data, the second fire determination result takes a value within the range from 0 to 1 (or from 0% to 100%).

[0019] (Machine learning) The second processing unit includes a machine learning unit that inputs observation data and outputs a second fire determination result. The machine learning unit pre-inputs, as learning data in advance, the observation data during normal monitoring with at least the case where the second fire determination result is 0 regarded as the correct answer (teacher), and performs machine learning, and inputs the newly observed observation data after the machine learning and outputs the second fire determination result.

[0020] (Fire determination 1 based on the first observation data and the second observation data) The observation unit collects the first observation data of the light in the first wavelength band radiated from the flame along with the fire, and the second observation data of the light outside the first wavelength band, and the first processing unit obtains a first fire determination result based on the first observation data and the second observation data. The second processing unit calculates a second fire determination result based on the first observation data and the second observation data.

[0021] (Fire determination 2 based on the first observation data and the second observation data) The observation unit collects the first observation data of the light in the first wavelength band radiated from the flame along with the fire, and the second observation data of the light outside the first wavelength band, and the first processing unit obtains a first fire determination result based on the ratio between the first integral value for each predetermined period of the first observation data and the second integral value for each predetermined period of the second observation data and the first integral value. The second processing unit calculates the second fire determination result based on the first integral value and ratio generated by the first processing unit.

[0022] (Fire detected) The fire detection unit detects a fire when the first fire determination result meets the predetermined conditions and the second fire determination result is equal to or exceeds a predetermined value.

[0023] (Fire warning signs detected) The fire detection unit is, If the first fire assessment result meets the prescribed conditions, and the second fire assessment result is less than or equal to the prescribed value, or If the first fire assessment result does not meet the prescribed conditions, and the second fire assessment result is equal to or exceeds a prescribed value, This will be classified as a fire precursor detection.

[0024] (Placement of fire detection system) Furthermore, the fire detection system of the present invention is A flame detection device equipped with an observation unit, a first processing unit, and a fire detection unit, A fire detection device is connected to the disaster prevention receiving panel, which is equipped with a second processing unit, It is characterized by having the following features.

[0025] (Fire detection method 1) The present invention is a fire detection method, The system is characterized by detecting a fire based on a first fire judgment result obtained by performing a predetermined fire judgment process on observation data of the target being monitored in the monitoring area, and a second fire judgment result calculated by performing a predetermined fire judgment process based on machine learning on the observation data.

[0026] (Fire detection method 2) The present invention is a fire detection method, The system is characterized by detecting a fire based on a first fire judgment result obtained by performing a predetermined fire judgment process on observation data collected by observing light radiation from a monitoring area, and a second fire judgment result calculated by performing a predetermined machine learning-based fire judgment process on the observation data.

[0027] (Fire detection method 3) Furthermore, the present invention relates to a fire detection method, The observation unit observes the emission of light from the monitoring area. The first processing unit performs a predetermined fire determination process on the observation data collected by the observation unit to obtain the first fire determination result. The second processing unit performs a fire judgment process based on predetermined machine learning processing on the observation data to calculate the second fire judgment result. The fire detection unit detects a fire in the monitored area based on the first fire determination result and the second fire determination result. It is characterized by the above. Other features are the same as those of a fire detection device. [Effects of the Invention]

[0028] (Effectiveness of fire detection systems) According to the fire detection system of the present invention, by detecting a fire based on a first fire judgment result obtained by performing a predetermined fire judgment process on observation data collected by observing light radiation from a monitoring area, and a second fire judgment result calculated by performing a predetermined machine learning-based fire judgment process on the observation data, the fire identification performance can be improved compared to conventional systems, and the problem of false fire alarms, where a fire is mistakenly identified and reported when there is no fire, can be prevented.

[0029] (Effect of the second fire assessment result) Furthermore, since the second fire determination result is a value ranging from 0 (no fire) to 1 (fire), it allows for a quantitative understanding of the likelihood of a fire occurring, enabling reliable fire detection.

[0030] (The effects of machine learning) The second processing unit includes a machine learning unit that takes observation data as input and outputs a second fire judgment result. The machine learning unit pre-inputs observation data from normal monitoring, where the correct answer (training data) is at least when the second fire judgment result is 0, as training data and performs machine learning. After machine learning, it inputs newly observed data and outputs a second fire judgment result. This eliminates the need to artificially devise and determine the conditions for calculating the second fire judgment result, and allows for the output of a second fire judgment result with high accuracy even with unknown observation data through machine learning. Since this training data is observation data from the environment in which the flame detection device is actually installed, if there are phenomena specific to the installation location that differ from general phenomena, it is possible to calculate a second fire judgment result with even higher accuracy that is specific to the installation location.

[0031] Furthermore, machine learning is performed by pre-inputting "observation data during a fire" (where a second fire judgment result of 1 is considered correct) and "observation data during normal monitoring" (where a second fire judgment result of 0 is considered correct) as training data. However, since fires occur very infrequently and are exceptional, it is difficult to collect "observation data during a fire" as training data. On the other hand, "observation data during normal monitoring" is observation data observed during normal monitoring when no fire is occurring, and can be easily and readily collected in large quantities as training data through the operation of the fire detection system. Therefore, by using the "observation data during normal monitoring," which can be collected in large quantities, as training data for machine learning, the accuracy of the learning machine that takes observation data as input and outputs the second fire judgment result can be improved.

[0032] (Effect of fire detection based on the first and second observation data) Furthermore, the fire detection system collects first observation data of light in the first wavelength band emitted from flames associated with a fire, and second observation data of light outside the first wavelength band. Based on the first and second observation data, it obtains a first fire judgment result and calculates a second fire judgment result. For example, when installed in a tunnel to detect fires, even if it observes light emitted from various sources other than flames, such as the headlights of vehicles traveling on the road, flashing or blinking warning lights of emergency vehicles, lighting fixtures inside the tunnel, sunlight if installed at the tunnel entrance / exit, or workers (human bodies) during maintenance and inspection, it will not be judged as a fire detection. This improves fire identification performance compared to conventional systems and prevents situations where a fire alarm is issued when there is no fire, and tunnel passage is prohibited until it is confirmed that it is not a fire alarm.

[0033] (Effect of fire detection method 2 based on first and second observation data) Furthermore, the fire detection system collects first observation data of light in a first wavelength band emitted from flames associated with a fire, and second observation data of light outside the first wavelength band. Based on the first integral value of the first observation data for predetermined periods and the ratio between the first integral value and the second integral value of the second observation data for predetermined periods, it obtains a first fire judgment result. Simultaneously, it calculates a second fire judgment result based on the first integral value and ratio generated by the first processing unit. In addition to the aforementioned "effect of fire judgment 1 based on first and second observation data," this reduces the number of inputs to the second processing unit that performs machine learning (reduces the number of vector dimensions), simplifies the configuration of the machine learning unit, and reduces the burden of computation.

[0034] (Effect of fire detection) Furthermore, the fire detection unit enables more reliable and accurate fire detection by detecting a fire when the first fire determination result meets predetermined conditions and the second fire determination result is equal to or exceeds a predetermined value.

[0035] (Effects of fire precursor detection) The fire detection unit cannot determine whether there is a fire or not if the first fire determination result meets the predetermined conditions and the second fire determination result is less than or equal to a predetermined value, or if the first fire determination result does not meet the predetermined conditions and the second fire determination result is equal to or greater than or exceeds a predetermined value. In such cases, for example, it detects a fire precursor and issues a fire warning, for example, by outputting a fire warning alarm from the disaster prevention receiver panel, thereby prompting relevant parties to check the site and take appropriate action.

[0036] (Effects of deploying a fire detection system) Furthermore, the fire detection system of the present invention simplifies the configuration of the flame detection device by comprising a flame detection device equipped with an observation unit, a first processing unit, and a fire detection unit, and a disaster prevention receiving panel connected to the flame detection device and equipped with a second processing unit. In other words, since the learning control unit and learning data storage unit of the second processing unit are not essential during normal monitoring after learning (the learning model unit is essential), the present invention can be implemented without significantly changing the configuration of conventional flame detection devices by placing them in a disaster prevention receiving panel with ample installation space.

[0037] (Effectiveness of fire detection methods) Furthermore, the fire detection method of the present invention provides the same effects as the fire detection system described above. [Brief explanation of the drawing]

[0038] [Figure 1] This is an explanatory diagram showing the specific details of an embodiment of a fire detection system. [Figure 2] This is an explanatory diagram showing a more detailed functional configuration of the observation unit and the second processing unit shown in Figure 1. [Figure 3] This is a time chart showing an example of comparing the first and second observed values ​​during normal monitoring. [Figure 4] This is a time chart showing an example of comparing the first and second observation values ​​observed during a fire. [Figure 5] This is an explanatory diagram showing the specific details of other embodiments of the fire detection system. [Figure 6]This is an explanatory diagram showing a more detailed functional configuration of the observation unit and the second processing unit shown in Figure 5. [Modes for carrying out the invention]

[0039] Embodiments of the fire detection system and fire detection method according to the present invention will be described in detail below with reference to the drawings. However, the present invention is not limited to the embodiments described below.

[0040] [Basic Concepts of the Embodiment] The embodiments generally relate to a fire detection system and a fire detection method. Here, "fire detection system" refers to a means of observing an object to be observed in a monitoring area (described later), detecting a fire based on this observation, and outputting a fire detection signal to, for example, a fire prevention receiving panel. This concept includes, for example, fire detectors, fire sensors, fire alarms, etc. These are included in the fire detection system and include flame detection devices (e.g., flame detectors) that detect fires by observing light from flames, smoke detection devices (e.g., smoke detectors) that detect fires by observing smoke, heat detection devices (e.g., heat detectors) that detect fires by observing heat (specifically, for example, temperature), and gas detection devices (e.g., gas detectors) that detect fires by observing gas (specifically, for example, gas concentration).

[0041] Here, "monitoring area" refers to the area that is monitored by the fire detection system, and is a concept that includes outdoor or indoor spaces of a certain size, such as the interior of structures like tunnels, rooms in buildings, corridors, and stairwells.

[0042] Furthermore, the "observed object" is the object that the fire detection system observes within its monitoring area, and the fire detection system detects a fire based on this observed data. The observed object is, for example, something that is generated in conjunction with the physicochemical phenomena of a fire, and specific examples include light, smoke, temperature, and gas. The fire is then detected based on changes in observed light data due to the emission of light associated with the flames of a fire (specifically, for example, a change in the infrared intensity emitted from the monitoring area), changes in observed smoke data due to the generation of smoke associated with the combustion of a fire (specifically, for example, a change in smoke concentration), changes in observed heat data due to the generation of heat due to combustion (specifically, for example, a change in temperature), and changes in observed gas data due to the generation of gas associated with combustion (specifically, for example, a change in gas concentration).

[0043] The fire detection system, as an example, is a fire prevention system consisting of a flame detection device and a fire prevention receiving panel connected thereto, and comprises an observation unit, a first processing unit, a second processing unit, and a fire detection unit. For example, the observation unit is located inside the flame detection device, and the first and second processing units may be located anywhere in the system, but in this embodiment, the first processing unit is located inside the flame detection device and the second processing unit is located inside the fire prevention receiving panel.

[0044] Furthermore, the fire detection system, for example, detects fires in a monitoring area within a road tunnel and is a component of a tunnel disaster prevention system. In other words, a tunnel disaster prevention system (tunnel disaster prevention equipment) is a system that includes the fire detection system of this application.

[0045] The "observation unit" is responsible for observing the emission of light from the monitoring area. Here, "emission of light from the monitoring area" is a concept that includes, for example, the light emitted from flames during a fire, but also includes, in addition to flames, the emission of light of various wavelengths and intensities from, for example, the headlights of passing vehicles, flashing or blinking warning lights of emergency vehicles, lighting fixtures inside tunnels, sunlight when installed at the entrances and exits of tunnels, and the light emitted by workers (human bodies) during maintenance and inspections.

[0046] Furthermore, "observing light emission" means, for example, observing the light in a first wavelength band specific to the flame emitted from the flame to collect a first observation value, and observing the light in a second wavelength band other than the first wavelength band specific to the flame to collect a second observation value. These values ​​are then sampled at a predetermined frequency and A / D converted to generate first and second observation data, which are time-series data at predetermined time intervals. In the following explanation, the first and second observation values, which are analog signals, are shown as lowercase x1 and x2, and the first and second observation data are shown as uppercase X1 and X2.

[0047] Furthermore, the "first processing unit" determines whether or not there is a fire based on the observation data collected by the observation unit. In this embodiment, it determines whether or not there is a fire based on the first fire determination result obtained by performing a predetermined fire determination process on the observation data.

[0048] Furthermore, the "second processing unit" quantitatively calculates the degree of fire likelihood based on observation data collected by the observation unit. In this embodiment, the degree of fire likelihood is quantitatively determined from the second fire judgment result calculated by performing a fire judgment process based on predetermined machine learning processing on the observation data. Here, the second fire judgment result based on observation data when there is a fire is 1 (or 100%), and the second fire judgment result based on observation data when there is no fire is 0 (or 0%), and the second fire judgment result ranges from 0 to 1 (or 0% to 100%) depending on the observation data.

[0049] In this embodiment, the second processing unit is a machine learning unit that calculates and outputs a second fire judgment result by inputting observed values ​​and performing binary classification. Here, "machine learning unit" refers to an algorithm that generates a learning model unit that learns the relationship between input and output from pre-provided supervised learning data (data in which the correct output for a given input is known), and after learning, inputs unknown data into the learning model unit to estimate the output. This concept includes, for example, well-known multilayer (deep) neural networks (DNN), random forests, support vector machines (SVC), etc.

[0050] In this embodiment, the machine learning unit pre-collects and performs machine learning on "observation data during a fire" where the second fire judgment result is 1 is considered correct (teacher), and "observation data during normal monitoring" where the second fire judgment result is 0 is considered correct (teacher). After machine learning, it inputs newly observed data and outputs the second fire judgment result.

[0051] Here, "fire observation data," which is one of the training data sets, is observation data observed when a fire occurs in the monitoring area and the first processing unit determines that "it is a fire." Since the frequency of fire occurrences is extremely low and can be said to be almost zero, it is difficult or impossible to collect it as training data. In contrast, "normal monitoring observation data," which is another training data set, is observation data observed during normal monitoring when no fire has occurred and the first processing unit determines that "it is not a fire." This data can be easily and readily collected in large quantities, and in this embodiment, the accuracy of the machine learning unit that takes observation data as input and outputs the second fire judgment result is improved by performing machine learning using this "normal monitoring observation data."

[0052] Furthermore, the fire detection system detects a fire based on a first fire determination result from the first processing unit and a second fire determination result from the second processing unit. In this embodiment, a fire is detected when the first fire determination result satisfies predetermined conditions (determined to be a fire) and the second fire determination result is equal to or greater than a predetermined value or exceeds a predetermined value. In addition, if the first fire determination result satisfies predetermined conditions (determined to be a fire) and the second fire determination result is less than or equal to a predetermined value, or if the first fire determination result does not satisfy predetermined conditions (determined not to be a fire) and the second fire determination result is equal to or greater than a predetermined value or exceeds a predetermined value, for example, a fire precursor detection is performed.

[0053] In other embodiments of the fire detection system, the first processing unit obtains a first fire determination result based on the first integral value of the first observation data for each predetermined period and the ratio between the first integral value and the second integral value of the second observation data for each predetermined period, and the second processing unit calculates a second fire determination result based on the first integral value and ratio generated by the first processing unit.

[0054] The following describes a specific embodiment. In the embodiment shown below, the "monitoring area" is the "internal space of the tunnel," the observation unit, first processing unit, and fire detection unit of the "fire detection system" are provided in the "flame detection device," the second processing unit is provided in the disaster prevention receiving panel, the "observation data observing the emission of light from the monitoring area" consists of "first observation data of light in the first wavelength band emitted from the flames due to the fire" and "second observation data of light in the second wavelength band other than the first wavelength band," and the "machine learning unit of the second processing unit" is a "multilayer neural network."

[0055] [Specific details of the embodiment] Next, we will describe the specific details of the embodiment. The details will be explained in the following sections. a. Fire detection system b. Observation Department c. First Processing Unit d. Second Processing Unit d1. Machine Learning Department d2. Learning Model Section d3. Learning Control Unit d4. Collection of training data d5. Machine Learning d6. Fire detection unit e. Other embodiments of fire detection systems e1. Second Processing Unit e2. Machine Learning Department e3. Learning Model Section e4. Learning Control Unit e5. Collection of training data e6. Machine Learning e7. Effects of reducing the number of inputs e8. Fire detection unit f. Modified versions of the present invention

[0056] [a. Fire detection system] First, the embodiment of the fire detection system will be described in more detail. As shown in Figure 1, the fire detection system 10(10-1) consists of an observation unit 12, a first processing unit 14, a second processing unit 16, and a fire detection unit 18.

[0057] [b. Observation Department] The observation unit 12 observes the emission of light from the monitoring area, for example, the space inside the tunnel, and outputs observation data. Its configuration and functions are arbitrary, but in this embodiment, it observes light in a first wavelength band around 4.5 μm emitted from flames associated with a fire and outputs first observation data X1, and also observes light in a second wavelength band other than flames, for example, around 5.0 μm, and outputs second observation data X2.

[0058] To explain in more detail, in this embodiment, the observation unit 12 is provided in the flame detection device 20 installed inside the tunnel, as shown in Figure 2. The observation unit 12 comprises a translucent window 25, a first observation unit 12a, a second observation unit 12b, and an observation data generation unit 13.

[0059] The first observation unit 12a receives light in the first wavelength band (around 4.5 μm) characteristic of flames, which has been selectively transmitted (passed through) by an optical wavelength bandpass filter from the light emitted from the tunnel space using the sensor unit 22a, converts it into an electrical signal, performs predetermined processing such as amplification by the amplification processing unit 24a to obtain the first observation value x1, and outputs it to the observation data generation unit 13.

[0060] The second observation unit 12b receives light in a second wavelength band (for example, around 5.0 μm) that is different from the first wavelength band, selected and transmitted (passed through) by an optical wavelength bandpass filter from the light emitted from the tunnel space by the sensor unit 22b, converts it into an electrical signal, performs predetermined processing such as amplification by the amplification processing unit 24b to obtain a second observation value x2, and outputs it to the observation data generation unit 13. The amplification processing units 24a and 24b are equipped with a preamplifier, a frequency filter that selectively passes a predetermined frequency band including the flame flicker frequency, and a main amplifier, etc.

[0061] Figure 3 shows an example comparing the first observed value x1 and the second observed value x2 observed during normal monitoring. Each shows the waveform change of the observed value relative to the midpoint level over a predetermined unit time T, for example, T = 2 seconds, which is observed as time-series data. During normal monitoring when no fire is occurring, the amount of light specific to the flame that passes through the optical wavelength bandpass filter in the first wavelength band is small, so the amplitude change of the first observed value x1 is small. On the other hand, the amount of light due to factors other than the flame that passes through the optical wavelength bandpass filter in the second wavelength band is relatively large, so the amplitude change of the second observed value x2 is larger than that of the first observed value x1.

[0062] Figure 4 shows an example comparing the first observed value x1 and the second observed value x2 observed during a fire. The amplitude change of the first observed value x1 is large because a large amount of light specific to the flame is transmitted through the optical wavelength bandpass filter in the first wavelength band. On the other hand, the amount of light due to factors other than the flame is relatively small when transmitted through the optical wavelength bandpass filter in the second wavelength band, so the amplitude change of the second observed value x2 is smaller than that of the first observed value x1.

[0063] The observation data generation unit 13 takes the first observation value x1 and the second observation value x2, which are analog signals output from the first observation unit 12a and the second observation unit 12b, and samples them at a predetermined sampling frequency, for example, 128 Hz, every T=2 seconds as shown in Figure 3 or Figure 4. By performing A / D conversion, it obtains 256 time-series data points, which are then read and stored as the first observation data X1 and the second observation data X2. In the following description, "first observation data X1" and "second observation data X2" refer to these "256 time-series data points."

[0064] [c. First Processing Unit] The first processing unit 14 performs a predetermined fire judgment process on the first observation data X1 and second observation data X2, which are time-series data stored in the observation data generation unit 13 every T=2 seconds, obtains the first fire judgment result E1, and outputs it to the fire detection unit 18. Its configuration and function are arbitrary, but in this embodiment, as shown in Figure 2, it is provided in the flame detection device 20 and consists of a computer circuit equipped with a CPU, memory, and various input / output ports.

[0065] The observation data generation unit 13, which is provided in the observation unit 12, may also be provided in the first processing unit 14. In this case, the second processing unit 16 will acquire the first observation data X1 and the second observation data X2 from the first processing unit 14. The fire judgment conditions set in the fire judgment process in the first processing unit 14 are arbitrary, but for example, the following two-stage fire judgment conditions may be set.

[0066] The first stage of fire detection is, for example, when the integral value σX1 of the first observation data X1 is equal to or exceeds a predetermined threshold, the relative ratio (σX1 / σX2) of the integral value σX1 of the first observation data X1 and the integral value σX2 of the second observation data X2 is calculated, and if this relative ratio exceeds a predetermined threshold, the first stage of fire detection is deemed to be satisfied, and the process proceeds to the next second stage.

[0067] The second stage of fire determination involves analyzing the first observation data X1 using a Fast Fourier Transform (FFT) to calculate, for example, the relative ratio (σFL / σFH) of the integral value σFL of the low-frequency component (relative intensity) below 4 Hz and below 8 Hz to the integral value σFH of the high-frequency component (relative intensity) above 4 Hz and below 8 Hz. If this relative ratio is equal to or exceeds a predetermined threshold, the second stage of fire determination is deemed satisfied, a fire is determined, and the first fire determination result E1=1 is output. If the first processing unit 14 does not satisfy the first or second stage of fire determination, it determines that there is no fire and outputs the first fire determination result E1=0.

[0068] [d. Second Processing Unit] The second processing unit 16 performs a fire judgment process based on predetermined machine learning processing on the first observation data X1 and second observation data X2, which are time-series data stored in the observation data generation unit 13 every T=2 seconds, calculates a second fire judgment result E2 and outputs it to the fire detection unit 18. Its configuration and functions are arbitrary, but in this embodiment it is composed of a computer circuit equipped with a CPU, memory and various input / output ports. For example, as shown in the second processing unit 16 of Figure 2, it is located on the fire prevention receiving panel 40 side rather than the flame detection device 20 side and is equipped with a machine learning unit 26.

[0069] (d1. Machine Learning Department) The machine learning unit 26 takes observation data (in this case, first observation data X1 and second observation data X2) as input and outputs a second fire judgment result E2. Its function, configuration, and type are arbitrary, but in this embodiment it consists of a learning model unit 28, a learning control unit 30, and a learning data storage unit 32.

[0070] (d2. Learning Model Section) The learning model unit 28 is machine-trained using supervised training data (data where the output is the correct answer for a given input). Its configuration and function are arbitrary, but for example, it is a multilayer neural network that performs binary classification. As is well known, a multilayer neural network for binary classification consists of an input layer, multiple hidden layers, and an output layer, and for example, a sigmoid function is used as the activation function of the output layer. In this embodiment, it is a multilayer neural network for binary classification that classifies the degree of fire Y1 and the degree of non-fire Y2.

[0071] The trained model unit 28 receives the first observation data X1 (X11, X12, .... X1n) and the second observation data X2 (X21, X22, .... X2n) for each unit time observed by the observation unit 12 in parallel, calculates the degree of fire Y1 and the degree of non-fire Y2, and outputs the degree of fire Y1 as the second fire judgment result E2. The second fire judgment result E2 takes a value in the range of 0 to 1, with a minimum value of 0 when no fire has occurred (normal monitoring) and a maximum value of 1 when a fire has occurred.

[0072] (d3. Learning Control Unit) The learning control unit 30 uses supervised learning data to train the learning model unit 28. Its functions and configuration are arbitrary, but in this embodiment, it is composed of a computer circuit equipped with a CPU, memory, and various input / output ports, and has the functions of collecting learning data and training the learning model unit 28.

[0073] (d4. Collection of training data) The learning control unit 30 collects first observation data X1 and second observation data X2 during a fire, and combines them to create first learning data, for example, training data (Y1,Y2)=(1,0) where Y1=1 is the correct answer for the degree of fire and Y2=0 is the correct answer for the degree of non-fire. (Y1,Y2)(X1,X2)=(1,0)(X1,X2) The first training data, as written above, is generated and stored in the training data storage unit 32. However, the occurrence of fires inside tunnels is extremely exceptional and can be said to be almost zero, and for this reason, it is often difficult or impossible to collect the first training data in reality.

[0074] Furthermore, the learning control unit 30 collects first observation data X1 and second observation data X2 during normal monitoring when no fire has occurred, and combines them to create second learning data, for example, training data (Y1,Y2)=(0,1) where Y1=0 is the correct answer for the degree of fire during normal monitoring and Y2=1 is the correct answer for the degree of non-fire. (Y1,Y2)(X1,X2)=(0,1)(X1,X2) A second set of training data is generated and stored in the training data storage unit 32. This second set of training data can be easily and in large quantities during normal monitoring, and in this embodiment, training data essentially means the second set of training data.

[0075] The learning control unit 30 collects learning data (second learning data) for a predetermined period, for example, 1 to 3 months, preferably 1 year, after the start of operation of the fire detection system 10 (10-1). If the second processing unit 16, like the first processing unit 14, samples and reads first observation data X1 and second observation data X2 for a unit time T = 2 seconds at 128 Hz, it can collect a large amount of learning data (second learning data), such as 43,200 data points in 1 day (24 hours) and 1,296,000 data points in 1 month (30 days).

[0076] (d5. Machine Learning) After the training data collection period has elapsed, the learning control unit 30 reads the training data (second training data) from the training data storage unit 32 and performs machine learning on the learning model unit 28. The method of this machine learning is arbitrary, but for example, the learning control unit 30 sets the structure of the multilayer neural network constituting the learning model unit 28, such as the number of nodes in the input layer (e.g., 256 nodes), the number of stages and nodes in the hidden layer, and the number of nodes in the output layer (e.g., 2 nodes), as well as the weights of each node, based on the user's operation. Subsequently, it reads the training data (second training data), sets the first and second observation data (X1, X2) as inputs to the learning model unit 28, and sets the correct answer (0, 1) as the output (Y1, Y2), and performs machine learning using a known backpropagation method.

[0077] The machine learning procedure is arbitrary, but for example, it would follow these steps. (a) Divide the learning data into training data and validation data. Use 70-80% as training data and the remaining 20-30% as validation data. (i) Divide the training data into multiple blocks, for example, blocks of 10,000 data points. (c) Divide the verification data into multiple blocks, for example, blocks of 1,000 data points. (e) The training data is read out in blocks and set in the learning model unit 28 to perform machine learning. (e) The verification data is read in block units and input into the trained learning model unit 28 to perform verification and determine the accuracy rate. The accuracy rate is the average value of the degree of non-fire Y2 output after inputting the verification data. (k) Repeat steps (d) and (e) as long as the accuracy rate increases. If the accuracy rate saturates or begins to decrease during the repetition, the learning process is terminated due to overfitting. If the accuracy rate falls below a predetermined value, for example 80%, the learning model unit 28 is rebuilt and the process is restarted from the beginning because the accuracy is too low.

[0078] Alternatively, the functions of the learning control unit 30 shown in Figure 2 may be provided as functions of a dedicated machine learning tool on a separate computer device or server, etc., separated from the second processing unit 16. The machine learning tool can then perform machine learning on the learning model unit 28 in the same manner as described above, and the trained learning model unit 28 can be implemented (placed) in the second processing unit 16.

[0079] (d6. Fire detection unit) The fire detection unit 18 detects a fire based on the first fire judgment result E1 from the first processing unit 14 and the second fire judgment result E2 from the second processing unit 16. For example, if the first fire judgment result E1 from the first processing unit 14 satisfies predetermined conditions for determining a fire (determined to be a fire), and the second fire judgment result E2 from the second processing unit 16 is a predetermined value, for example, 0.6 or higher, then a fire is detected, and a fire detection signal E3 is transmitted to the disaster prevention receiving panel 40 (for example, the control unit 42). Upon receiving the fire detection signal E3, the disaster prevention receiving panel 40 (for example, the alarm unit 44) outputs a fire alarm to notify the public, and also issues an entry prohibition alarm using an alarm display board or similar equipment to prohibit vehicles from passing through the tunnel, and a management officer goes to the site to confirm the fire and take action.

[0080] Furthermore, the fire detection unit 18 is (1) When the first fire determination result E1 by the first processing unit 14 satisfies the predetermined conditions, and the second fire determination result E2 by the second processing unit 16 is less than a predetermined value, for example, 0.6, (2) When the first fire determination result E1 by the first processing unit 14 satisfies the predetermined conditions, and the second fire determination result E2 by the second processing unit 16 is a predetermined value, for example, 0.6 or more, Upon detecting a fire precursor, a fire precursor detection signal E4 is transmitted to the fire prevention receiving panel (e.g., control unit 42). The fire prevention receiving panel 40 (e.g., alarm unit 44), upon receiving the fire precursor detection signal E4, outputs a fire warning alert to draw attention, enabling the person in charge of management to go to the site, confirm the fire, and take action.

[0081] [e. Other embodiments of fire detection systems] Figure 5 is an explanatory diagram showing another embodiment of the fire detection system of the present invention, and Figure 6 is an explanatory diagram showing a more detailed functional configuration of the observation unit and the second processing unit of Figure 5.

[0082] The fire detection system 10(10-2) of this embodiment consists of an observation unit 12, a first processing unit 14, a second processing unit 16, and a fire detection unit 18. The observation unit 12, the first processing unit 14, and the fire detection unit 18 are basically the same as in the embodiment of Figure 1, but the second processing unit 16 is input the integral value (first integral value) σX1 of the first observation data X1 obtained by the first processing unit 14, and the relative ratio (σX1 / σX2) of this integral value σX1 and the integral value (second integral value) σX2 of the second observation data X2.

[0083] (e1. Second Processing Unit) The second processing unit 16 receives the integral value σX1 of the first observation data X1 generated by the first processing unit 14 and the relative ratio (σX1 / σX2) between the integral value σX1 and the integral value σX2 of the second observation data X2, executes a fire judgment process based on predetermined machine learning processing, calculates a second fire judgment result E2, and outputs it to the fire detection unit 18. For example, as shown in the second processing unit 16 of Figure 6, it includes a machine learning unit 26.

[0084] (e2. Machine Learning Department) The machine learning unit 26 takes the integral value σX1 and relative ratio (σX1 / σX2) based on the first and second observation data X1 and X2 as input and outputs the second fire judgment result E2. Its function, configuration, and type are arbitrary, but it consists of a learning model unit 28, a learning control unit 30, and a learning data storage unit 32, similar to Figure 2.

[0085] (e3. Learning Model Section) The learning model unit 28 is machine-trained using supervised training data (data where the output is the correct answer for a given input). Its configuration and function are arbitrary, but for example, it is a multilayer neural network that performs binary classification. As is well known, a multilayer neural network for binary classification consists of an input layer, multiple hidden layers, and an output layer, and for example, a sigmoid function is used as the activation function of the output layer. In this embodiment, it is a multilayer neural network for binary classification that classifies the degree of fire Y1 and the degree of non-fire Y2.

[0086] The trained model unit 28 receives the integral value σX1 and relative ratio (σX1 / σX2) obtained from the first and second observation data X1 and X2 for each unit time observed by the observation unit 12 in parallel, calculates the degree of fire Y1 and the degree of non-fire Y2, and outputs the degree of fire Y1 as the second fire judgment result E2. The second fire judgment result E2 takes a value in the range of 0 to 1, with a minimum value of 0 when no fire has occurred (during normal monitoring) and a maximum value of 1 when a fire has occurred.

[0087] (e4. Learning Control Unit) The learning control unit 30 uses supervised learning data to train the learning model unit 28. Its functions and configuration are arbitrary, but in this embodiment, it is composed of a computer circuit equipped with a CPU, memory, and various input / output ports, and has the functions of collecting learning data and training the learning model unit 28.

[0088] (e5. Collection of training data) The learning control unit 30 collects the integral value σX1 and relative ratio (σX1 / σX2) obtained based on the first and second observation data X1 and X2 during a fire, and combines this with training data (Y1,Y2)=(1,0) where Y1=1 is the correct answer for the degree of fire and Y2=0 is the correct answer for the degree of non-fire, to create first learning data, for example, (Y1,Y2)(σX1,σX1 / σX2)=(1,0)(σX1,σX1 / σX2) The first training data, as written above, is generated and stored in the training data storage unit 32. However, the occurrence of fires inside tunnels is extremely exceptional and can be said to be almost zero, and for this reason, it is often difficult or impossible to collect the first training data in reality.

[0089] Furthermore, during normal monitoring when no fire has occurred, the learning control unit 30 collects the integral value σX1 and relative ratio (σX1 / σX2) obtained based on the first and second observation data X1 and X2, and combines this with training data (Y1,Y2)=(0,1) where Y1=0 is the correct answer for the degree of fire during normal monitoring and Y2=1 is the correct answer for the degree of non-fire, to create second learning data, for example, (Y1,Y2)(σX1,σX1 / σX2)=(0,1)(σX1,σX1 / σX2) A second set of training data is generated and stored in the training data storage unit 32. This second set of training data can be easily and in large quantities during normal monitoring, and in this embodiment, training data essentially means the second set of training data.

[0090] (e6. Machine Learning) After the learning data collection period has elapsed, the learning control unit 30 reads the learning data (second learning data) from the learning data storage unit 32 and performs machine learning on the learning model unit 28. The method of this machine learning is arbitrary, but for example, the learning control unit 30 sets the structure of the multilayer neural network constituting the learning model unit 28, such as the number of nodes in the input layer (e.g., 256 nodes), the number of stages and nodes in the hidden layer, and the number of nodes in the output layer (e.g., 2 nodes), as well as the weights of each node, based on the user's operation. Subsequently, it reads the learning data (second learning data), sets the integral value σX1 and the relative ratio (σX1 / σX2) as inputs to the learning model unit 28, and sets the correct answer (0,1) as the output (Y1,Y2), and performs machine learning using a known backpropagation method. The procedure for machine learning is arbitrary, but it is the same as shown in the embodiment of Figure 2.

[0091] (e7: Effects of reducing the number of inputs) In the embodiment shown in Figure 2, the learning model unit 28 processes a 256-dimensional vector input by inputting 256 observation values ​​consisting of the first observation data X1 and the second observation data X2 in parallel. However, in the embodiment shown in Figure 6, the learning model unit 28 is reduced to a 2-dimensional vector input by inputting two feature values ​​consisting of the integral value σX1 and the relative ratio (σX1 / σX2) in parallel. This significantly reduces the configuration and computational processing of the machine learning unit 26, including the computational processing of the learning control unit 30.

[0092] (e8. Fire detection unit) The fire detection unit 18 detects a fire in the same manner as in the embodiment shown in Figure 2, based on the first fire determination result E1 from the first processing unit 14 and the second fire determination result E2 from the second processing unit 16.

[0093] [f. Variations of the present invention] Embodiments of the fire detection system and fire detection method according to the present invention will be described in detail.

[0094] (First processing unit of the 3-wavelength system) The observation unit 12 described above uses a so-called two-wavelength method as an example, but a so-called three-wavelength method may also be used. In this case, the observation unit 12 detects a first observation value x1 of light in the first wavelength band (around 4.5 μm) characteristic of flames, a second observation value x2 of light in the second wavelength band other than the first wavelength band (for example, around 5.0 μm), and a third observation value x3 of light in the third wavelength band other than the first and second wavelength bands (for example, around 2.3 μm).

[0095] When the observation unit 12 uses a three-wavelength system, the fire detection process in the first processing unit 14 determines a fire by setting three stages of fire detection conditions. The first stage fire detection condition is the same as the first stage fire detection condition of the two-wavelength system described above. For the second stage fire detection condition, the relative ratio (σX1 / σX3) of the integral values ​​of the first observation data X1 and the third observation data X3 is calculated, and if this relative ratio exceeds a predetermined threshold, the second stage fire detection condition is considered satisfied, and the process proceeds to the next third stage. The third stage fire detection condition is the same as the second stage fire detection condition of the two-wavelength system described above.

[0096] Furthermore, if the observation unit 12 uses a three-wavelength system, the machine learning processing in the second processing unit 16 is configured to input the first observation data X1, the second data X2, and the third observation data X3 to the learning machine unit 26 to calculate the degree of fire Y1 and the degree of non-fire Y2. The first learning data during a fire and the second learning data during normal monitoring are collected and machine-learned. After that, the newly observed observation data X1, X2, and X3 are input to classify the degree of fire Y1 and the degree of non-fire Y2, and the degree of fire Y1 is output as the second fire judgment result E2.

[0097] Furthermore, in another embodiment where the observation unit 12 uses a three-wavelength system, the machine learning processing in the second processing unit 16 is configured such that the learning machine unit 26 is input to calculate the degree of fire Y1 and the degree of non-fire Y2. This learning machine unit is then trained by inputting the integral value σX1 of the first observation data X1, the relative ratio (σX1 / σX2) between the integral value σx1 and the integral value σX2 of the second data X2, and the relative ratio (σX1 / σX3) between the integral value σx1 and the integral value σX3 of the third data X3. The first learning data during a fire and the second learning data during normal monitoring are collected and machine learning is performed. After that, the newly observed observation data X1, X2, and X3 are input to classify the degree of fire Y1 and the degree of non-fire Y2, and the degree of fire Y1 is output as the second fire judgment result E2.

[0098] (Machine Learning Department) In the above embodiment, the machine learning unit 26 of the second processing unit 16 outputs the degree of fire Y1 as the second fire judgment result E2 to the fire detection unit 18. However, instead, the degree of non-fire Y2 may be output as the second fire judgment result E2. In this case, the fire detection unit 18 detects a fire when the predetermined conditions for the first processing unit 14 to determine it is a fire are met (it determines it is a fire), and the second fire judgment result E2 from the second processing unit 16 is a predetermined value, for example, 0.4 or less.

[0099] (Fire alarm system) The above embodiment uses tunnel fire prevention equipment as an example, but is not limited to this and is arbitrary; for example, it may be applied to a fire alarm system that monitors fires by connecting a detector to a receiver. In a fire alarm system, the fire detection system of this embodiment may, for example, have a flame detection device installed in a room of a building that is the monitoring area, or it may be installed outdoors to monitor arson, etc.

[0100] (others) Furthermore, the present invention includes appropriate modifications that do not impair its purpose and advantages, and is not limited by the numerical values ​​shown in the above embodiments. [Explanation of Symbols]

[0101] 10(10-1), 10(10-2): Fire detection system 12: Observation Department 12a: First Observation Unit 12b: Second Observation Unit 13: Observation Data Generation Unit 14: First Processing Unit 16: Second Processing Unit 18: Fire detection area 20: Flame detection device 22a, 22b: Sensor section 24a, 24b: Amplification processing unit 25: Translucent window 26: Machine Learning Department 28: Learning Model Department 30: Learning Control Unit 32: Learning data storage unit 40: Disaster Prevention Receiving Panel 42: Control Unit 44: Alarm section

Claims

1. A fire detection device positioned in the monitoring area observes the target and collects observation data, then performs a predetermined fire judgment process on the collected data to obtain a first fire judgment result. A machine learning unit located in a different location from the aforementioned monitoring area performs a fire judgment process based on predetermined machine learning processing on the observation data to calculate a second fire judgment result. A fire detection method characterized in that the fire detection device detects a fire in the monitored area based on the first fire determination result and the second fire determination result.

2. A fire detection method according to claim 1, The fire detection method is characterized in that the machine learning unit is provided in a fire prevention receiving panel to which the fire detection device is connected.

3. A first processing unit located in the monitoring area observes the target of observation and performs a predetermined fire judgment process on the collected observation data to obtain a first fire judgment result. A second processing unit, located in a different location from the aforementioned monitoring area, performs a fire determination process based on predetermined machine learning processing on the observation data to calculate a second fire determination result. A fire detection method characterized by detecting a fire in the monitored area based on the first fire determination result and the second fire determination result using a fire detection device.

4. A fire detection method according to claim 3, The fire detection method is characterized in that the second processing unit is provided in a fire prevention receiving panel to which the fire detection device is connected.

5. A fire detection method according to any one of claims 1 to 4, The fire detection method is characterized in that the second fire determination result is a value in the range from 0 for cases of fire to 1 for cases of non-fire.

6. A fire detection method according to claim 1 or 2, The aforementioned machine learning unit, At least the case where the second fire judgment result is 1 is considered correct, and the observation data during normal monitoring is input in advance as training data and machine learning is performed. A fire detection method characterized by inputting newly observed data after the machine learning process and outputting the second fire determination result.

7. A fire detection method according to claim 5 or 6, The fire detection device is characterized in that it detects a fire when the first fire determination result satisfies predetermined conditions and the second fire determination result is less than or equal to a predetermined value.

Images