JP7376660B1 - fire detection device - Google Patents

Abstract

An object of the present invention is to detect with higher accuracy whether or not a flame has broken out based on sound data collected in a monitoring area. [Solution] A microphone that collects sounds generated in a monitoring area as acoustic data, a frequency analysis section that calculates a frequency spectrum by frequency analysis of the acoustic data picked up by the microphone, and a frequency spectrum that is calculated by the frequency analysis section. It is determined whether 1/f fluctuation characteristics are included in the frequency band below the standing wave with respect to the frequency spectrum, and if it is determined that 1/f fluctuation characteristics are included, flames occur in the monitoring area. It also includes a flame detection section that detects when something has happened. [Selection diagram] Figure 1

Description

The present disclosure relates to a fire detection device that detects the occurrence of flame in a monitoring area based on acoustic data.

Various fire detectors are used to detect the occurrence of a flame in a monitored area. Specific types include, for example, a differential spot sensor, a constant temperature spot sensor, and a photoelectric spot sensor.

There is also a fire detection device that detects the sound generated when something burns during a fire (for example, see Non-Patent Document 1). Non-Patent Document 1 discloses a method in which a combustion experiment is conducted to capture the characteristics of sound during combustion, and based on frequency analysis of the sound, a fire is distinguished from a normal state, and a fire is determined.

In particular, in Non-Patent Document 1, rather than chasing a specific frequency and level, a certain frequency range is specified and the presence or absence of a combustion phenomenon is identified based on the time change of the integral value of the overall power (level). It is stated that the method to do so is appropriate.

“About frequency analysis of combustion sound (2nd report)”, Fire Science Research Institute Bulletin No. 31 (1994)

However, the conventional technology has the following problems.
Non-Patent Document 1 introduces the concept of temporal changes in the integral value of the sound power spectrum as a measure to more reliably separate background noise and combustion sound.

In general, when detecting a fire, it is not enough to simply distinguish between background noise and combustion sounds; it is necessary to more reliably distinguish between combustion sounds and other noises that occur within the monitoring area to prevent false alarms. It is desirable to suppress this.

In other words, in order to put fire detection based on acoustic data into practical use, it is necessary to find features that can quantitatively separate combustion sounds generated during combustion from sounds other than combustion sounds, without misidentifying fires. It is important to improve detection accuracy.

The present disclosure has been made to solve the above-mentioned problems, and it is possible to detect with higher accuracy whether or not a flame has occurred based on acoustic data collected in a monitoring area. The purpose is to obtain a fire detection device.

A fire detection device according to the present disclosure includes a microphone that collects sounds generated in a monitoring area as acoustic data, a frequency analysis unit that calculates a frequency spectrum by frequency-analyzing the acoustic data collected by the microphone, and It is determined whether or not the frequency spectrum calculated by the analysis unit includes 1/f fluctuation characteristics in the frequency band below the standing wave in the monitoring area, and it is determined that the 1/f fluctuation characteristics are included. The system is equipped with a flame detection unit that detects the occurrence of flame in the monitoring area.

The fire detection device according to the present disclosure also includes a microphone that collects sounds generated in a monitoring area as acoustic data, and a frequency analysis unit that calculates a frequency spectrum by frequency-analyzing the acoustic data collected by the microphone. , Determine whether or not 1/f fluctuation characteristics are included in the frequency band below the standing wave in the monitoring area with respect to the frequency spectrum calculated by the frequency analysis unit, and determine that 1/f fluctuation characteristics are included. In this case, a feature amount calculation unit that calculates the level error of the frequency spectrum with respect to the 1/f fluctuation characteristic and outputs the calculated level error as a feature amount used for determining the occurrence of a flame, and a feature amount calculation unit that generates the flame to be detected in advance. Based on the frequency analysis results by the frequency analysis section, the various acoustic data collected when the fire occurred, as well as the various acoustic data collected when there was no flame, including false alarm factors. A model creation unit creates a machine learning model to determine whether a flame has occurred by using the feature quantities calculated by the feature quantity calculation unit as parameters, and a feature quantity calculation unit when monitoring the monitoring area. The apparatus includes a flame detection section that uses the feature quantities calculated by the section as input to a machine learning model and determines whether or not a flame has occurred.

According to the present disclosure, it is possible to obtain a fire detection device that can detect with higher accuracy whether or not a flame has occurred based on acoustic data collected in a monitoring area.

FIG. 1 is an explanatory diagram showing the overall configuration of a fire detection device according to Embodiment 1 of the present disclosure. FIG. 2 is an explanatory diagram showing an example of a monitoring area when a verification experiment regarding flame detection processing by the fire detection device according to Embodiment 1 of the present disclosure was conducted. FIG. 2 is an explanatory diagram regarding a flame detection method using the fire detection device according to Embodiment 1 of the present disclosure. FIG. 4 is an explanatory diagram showing a frequency spectrum in a monitoring area of a different size from that in FIG. 3 regarding the flame detection method by the fire detection device according to Embodiment 1 of the present disclosure. 5 is an explanatory diagram showing a comparison between the background noise in FIG. 3 and the background noise in FIG. 4 regarding the flame detection method using the fire detection device according to Embodiment 1 of the present disclosure. FIG. In the fire detection device according to Embodiment 1 of the present disclosure, it is a comparison diagram showing frequency spectra regarding three types of acoustic data: background noise, flame, and ventilation fan. In the fire detection device according to Embodiment 1 of the present disclosure, it is a comparison diagram showing frequency spectra regarding three types of acoustic data: background noise, flame, and ventilation fan. FIG. 3 is a diagram showing an average level error calculated by the flame detection unit according to Embodiment 1 of the present disclosure. 2 is a flowchart showing a series of processes related to a fire detection method executed in the fire detection device according to Embodiment 1 of the present disclosure. FIG. 2 is an explanatory diagram showing the overall configuration of a fire detection device according to Embodiment 2 of the present disclosure. 7 is a flowchart showing a series of processes related to a fire detection method executed in a fire detection device according to Embodiment 2 of the present disclosure. FIG. 7 is a diagram showing frequency spectra in which 1/f fluctuation characteristics under various environmental conditions were verified in Embodiment 3 of the present disclosure. FIG. 7 is an explanatory diagram showing a comparison between a spectrogram based on acoustic data when a flame occurs and a spectrogram based on acoustic data when a door opens/closes in Embodiment 3 of the present disclosure. FIG. 9 is an explanatory diagram showing a comparison between a frequency spectrum based on acoustic data when a flame occurs and a frequency spectrum based on acoustic data when a door opens/closes in Embodiment 3 of the present disclosure. FIG. 12 is an explanatory diagram showing a method for identifying combustion sound and door opening/closing sound based on the level transition of a peak frequency in a low frequency range in Embodiment 3 of the present disclosure. 12 is a flowchart showing a series of processes related to a fire detection method executed in a fire detection device according to Embodiment 3 of the present disclosure. 12 is a flowchart showing a series of processes related to a fire detection method executed in a fire detection device according to Embodiment 4 of the present disclosure.

Hereinafter, preferred embodiments of the fire detection device of the present disclosure will be described using the drawings.
The present disclosure has a technical feature in that the presence or absence of flame generation is determined by extracting characteristic quantities unique to a flame from the frequency analysis results based on sound data collected in the monitoring area, and suppresses the cause of false alarms. This enables highly accurate flame detection.

Embodiment 1.
FIG. 1 is an explanatory diagram showing the overall configuration of a fire detection device according to Embodiment 1 of the present disclosure. The fire detection device according to the first embodiment includes a microphone 10 and a computer 20.

The microphone 10 is a sound collection device that is installed in a flame monitoring area and collects sounds generated within the monitoring area as acoustic data. Note that a plurality of microphones 10 may be installed within the monitoring area.

When determining the presence or absence of a flame based on acoustic data collected using the microphone 10, attenuation is less likely to occur because the focus is on low-frequency pressure fluctuations. Therefore, for example, if a closed room is used as a monitoring area, combustion can be detected regardless of the location of the fire, and there is an advantage that the microphone 10 can be installed anywhere. Furthermore, the microphone 10 does not need to be installed on the ceiling like a smoke detector, and there is a degree of freedom in determining the installation location.

The computer 20 is a controller that determines whether or not a flame has occurred in the monitoring area by performing arithmetic processing on the acoustic data picked up by the microphone 10, and controls the frequency analysis section 21 and the flame detection section 22. We are prepared.

FIG. 2 is an explanatory diagram showing an example of a monitoring area when a verification experiment regarding flame detection processing by the fire detection device according to Embodiment 1 of the present disclosure was conducted. In Figure 2, a room with a width of 12 m x depth of 11 m is used as a monitoring area, heptane is burned in a 50 cm x 50 cm fire pan in the center of the monitoring area, and acoustic data is collected by installing a microphone 10 at a point 3 m away from the fire source. An example is shown below.

The frequency analysis unit 21 calculates a frequency spectrum by performing frequency analysis on the acoustic data picked up by the microphone 10. The flame detection unit 22 determines whether or not the frequency spectrum calculated by the frequency analysis unit 21 includes a 1/f fluctuation characteristic in a frequency band below a standing wave. Furthermore, if it is determined that the 1/f fluctuation characteristic is included, the flame detection unit 22 detects that a flame has occurred in the monitoring area.

Note that the "frequency band below the standing wave" in the present disclosure means a desired frequency band included within the band below the standing wave. For example, if the standing wave is 14.4 Hz, the frequency band from f1 = 2.1 Hz to f2 = 10.0 Hz can be selected as the "frequency band below the standing wave."

Therefore, specific processing contents by the frequency analysis section 21 and the flame detection section 22 will be explained in detail using FIGS. 3 to 8. FIG. 3 is an explanatory diagram regarding a flame detection method using the fire detection device according to Embodiment 1 of the present disclosure.

In Figure 3, acoustic data corresponding to background noise collected within the monitoring area without flame generation, and acoustic data collected within the monitoring area with heptane flame generated, respectively. , the results of calculating frequency spectra by the frequency analysis section 21 are shown in comparison.

In FIG. 3, the part indicated by the dotted ellipse is the area of standing waves. Here, the term "standing wave" refers to a wave generated in a sound field in a closed space at a resonant frequency where the intensity of the radiated sound is easily amplified depending on its shape and size.

The approximate value of this standing wave can be calculated according to the size of the monitoring area, and in the monitoring area of the size shown in FIG. 2, the frequency of the standing wave is 14.4 Hz. In FIG. 3, the frequency band below the standing wave is shown as a rectangular frame.

Furthermore, the standing waves can also be calculated by performing impulse measurements in advance at a location that will be the monitoring area. Furthermore, when there are multiple standing waves, by employing the lowest standing wave, fire detection processing can be performed more accurately.

Focusing on the frequency band below the standing wave, the frequency spectrum when a flame is generated is a power spectrum that approximates the 1/f fluctuation characteristic passing through the peak frequency. On the other hand, the frequency spectrum of background noise does not have a power spectrum that approximates the 1/f fluctuation characteristic passing through the peak frequency.

Furthermore, when focusing on the standing wave region, which is an elliptical portion, there is no major difference between the frequency spectrum when a flame occurs and the frequency spectrum when background noise occurs, although there is a difference in level. Therefore, if a frequency spectrum having a power spectrum similar to the 1/f fluctuation characteristic passing through the peak frequency is obtained in the frequency band below the standing wave, it can be detected that a flame has occurred.

FIG. 4 is an explanatory diagram showing a frequency spectrum in a monitoring area of a different size from that in FIG. 3 regarding the flame detection method by the fire detection device according to Embodiment 1 of the present disclosure. Although FIG. 3 shows a case where the width of the monitoring area is 12 m, FIG. 4 shows a case where the width of the monitoring area is 14 m.

In FIG. 4, when focusing on the frequency band below the standing wave, the frequency spectrum when a flame is generated is a power spectrum that approximates the 1/f fluctuation characteristic passing through the peak frequency. On the other hand, the frequency spectrum of background noise does not have a power spectrum that approximates the 1/f fluctuation characteristic passing through the peak frequency.

FIG. 5 is an explanatory diagram showing a comparison between the background noise in FIG. 3 and the background noise in FIG. 4 regarding the flame detection method by the fire detection device according to Embodiment 1 of the present disclosure. Specifically, FIG. 5 shows a comparison of frequency spectra related to background noise in a case where a room with a width of 14 m is used as a monitoring area and a case where a room with a width of 12 m is used as a monitoring area.

In both frequency spectra, when focusing on the frequency band below the standing wave, it can be seen that the error with the 1/f fluctuation characteristic passing through the peak frequency is large. Therefore, by quantitatively calculating the error between the 1/f fluctuation characteristic passing through the peak frequency and the frequency spectrum in the frequency band below the standing wave, background noise and flames can be compared without depending on the size of the monitoring area. It can be seen that it is possible to identify the

Next, we will show the verification results regarding the sound of ventilation fans as a cause of false alarms. FIG. 6 is a comparison diagram showing frequency spectra regarding three types of acoustic data of background noise, flame, and ventilation fan in the fire detection device according to Embodiment 1 of the present disclosure. As shown in FIG. 6, it can be seen that in the frequency spectrum based on the acoustic data from the ventilation fan, the error with the 1/f fluctuation characteristic passing through the peak frequency is large in the frequency band below the standing wave.

FIG. 7 is a comparison diagram showing frequency spectra regarding three types of acoustic data of background noise, flame, and ventilation fan in the fire detection device according to Embodiment 1 of the present disclosure. Specifically, in FIG. 7(A), in the frequency band from f1 = 2.1 Hz to f2 = 10.0 Hz as the frequency band below the standing wave, 1 which passes through the peak frequency for each of background noise, flame, and ventilation fan. /f fluctuation characteristics.

Further, FIG. 7B shows a 1/f fluctuation characteristic of the flame passing through the peak frequency in a frequency band from f1=2.1 Hz to f2=20.0 Hz, which is a frequency band including standing waves.

As shown in FIG. 7(A), in the frequency band below the standing wave, the frequency spectrum related to flames approximates the 1/f fluctuation characteristic, but the frequency spectra related to background noise and ventilation fans are similar to 1/f fluctuation characteristics. It does not approximate the f fluctuation characteristics.

Furthermore, as shown in Fig. 7(B), in the frequency band including standing waves, the peak frequency exists in the frequency band of 10.0 Hz or higher in the frequency spectrum related to the flame, so the 1/f fluctuation characteristic It can be seen that it no longer approximates.

Therefore, in the first embodiment, the 1/f fluctuation characteristic is calculated from the frequency spectrum, and an average level error is introduced as an index for quantitatively specifying the degree of approximation between the frequency spectrum and the 1/f fluctuation characteristic. Therefore, a method for calculating the 1/f fluctuation characteristic and the average level error will be specifically explained using mathematical formulas.

The flame detection unit 22 uses the frequency spectrum calculated by the frequency analysis unit 21 to calculate the 1/f fluctuation characteristic Y(f) [dB] based on the following equation (1).

Each symbol in the above formula (1) means the following.
f: Frequency [Hz]
fmax: Peak frequency in the frequency band below the standing wave of the flame frequency spectrum [Hz]
Pmax: Power of flame peak frequency fmax α: Weighting coefficient preset as a real value between 1 and 2

Note that, as the weighting coefficient α becomes larger, the level attenuation gradient of the 1/f fluctuation characteristic becomes steeper as the frequency becomes higher. The weighting coefficient α can also be set to an appropriate value depending on the size of the monitoring area and the measurement results of flames, background noise, and noise occurring in the monitoring area as a frequency band below the standing wave.

Furthermore, the flame detection unit 22 uses the calculated fluctuation characteristic Y(f) and the frequency spectrum calculated by the frequency analysis unit 21 to calculate the average level error Lave [dB] based on the following formula (2). .

Each symbol in the above formula (2) means the following.
X(f): Value of frequency spectrum at frequency f Y(f): Value of 1/f fluctuation characteristic at frequency f f1: Peak frequency fmax in the frequency band below the standing wave, or a value lower than the peak frequency fmax f2 :Frequency of the standing wave or a frequency about 1 to 2 Hz lower than the standing wave

FIG. 8 is a diagram showing the average level error Lave calculated by the flame detection unit 22 according to the first embodiment of the present disclosure. The average level errors for each of "background noise", "flame", and "ventilation fan" shown on the left side of FIG. 8 are as follows: f1= After calculating the 1/f fluctuation characteristic Y(f) using the above formula (1) for the range defined by 2.1Hz to f2 = 10.0Hz, the average level error Lave is calculated using the above formula (2). It corresponds to the result obtained.

In addition, the average level error regarding "flame (conventional)" shown on the right side of FIG. 8 is in the frequency spectrum shown in FIG. Corresponds to the result of calculating the average level error Lave using the above equation (2) after calculating the 1/f fluctuation characteristic Y(f) using the above equation (1) for the range specified by 20.0 Hz. are doing.

As shown in FIG. 8, by determining the average level error Lave limited to the frequency band below the standing wave, for example, by setting the determination threshold = 4 dB, it is determined whether the average level error Lave is 4 dB or less. By doing so, it is possible to distinguish between background noise and ventilation fans and detect the occurrence of flames. However, even for a flame, the average level error Lave determined in a frequency band including standing waves is larger than the determination threshold, resulting in a result that does not approximate the 1/f fluctuation characteristic.

Therefore, the flame detection unit 22 detects the peak frequency in the frequency band below the standing wave, calculates the 1/f fluctuation characteristic that passes through the peak frequency, and further calculates the average level error to detect the detection target. flame can be detected with high precision.

Next, a series of processes executed in the fire detection device according to the first embodiment will be described using a flowchart. FIG. 9 is a flowchart showing a series of processes related to the fire detection method executed in the fire detection device according to Embodiment 1 of the present disclosure.

First, in step S901, the frequency analysis unit 21 performs sound collection processing of acoustic data generated within the monitoring area via the microphone 10 installed in the monitoring area. Next, in step S902, the frequency analysis unit 21 calculates a frequency spectrum by performing frequency analysis processing on the acoustic data picked up by the microphone 10.

Next, in step S903, the flame detection unit 22 executes 1/f fluctuation characteristic detection processing in the frequency band below the standing wave. Specifically, the flame detection unit 22 calculates the peak frequency Pmax and also calculates the 1/f fluctuation characteristic using the above equation (1).

Next, in step S904, the flame detection unit 22 calculates a level error between the frequency spectrum calculated in step S902 and the 1/f fluctuation characteristic calculated in step S903. As an example of the level error calculation process, the flame detection unit 22 can use the average level error Lave calculated using the above equation (2) as the level error.

Note that the level error is not limited to the above equation (2), but can be expressed as [Y(f) - It is also possible to use other index values, such as calculating the mean squared error.

Next, in step S905, the flame detection unit 22 determines whether the level error calculated in step S904 is less than or equal to a preset determination threshold. Then, if the level error is below the determination threshold, the flame detection unit 22 determines that the frequency spectrum in the frequency band below the standing wave includes a 1/f fluctuation characteristic and that a flame has occurred, The process advances to step S906.

On the other hand, if the level error exceeds the determination threshold, the flame detection unit 22 determines that the frequency spectrum in the frequency band below the standing wave does not include the 1/f fluctuation characteristic and that no flame is generated. Then, the series of processing ends.

If the process advances to step S906, the flame detection unit 22 executes flame detection notification processing. For example, the flame detection unit 22 sends a fire signal to the fire receiver to notify that it has detected the occurrence of a flame, thereby activating fire extinguishing equipment, activating the linked operation of smoke prevention equipment such as shutters, and emergency broadcasting equipment. It is possible to carry out transfers of information to etc.

As described above, according to Embodiment 1, the feature amount peculiar to a flame is calculated from the frequency analysis result based on the acoustic data collected in the monitoring area, and it is quantified whether or not a flame has occurred in the monitoring area. can be judged accurately. Specifically, by focusing on the frequency band below the standing wave, calculating the 1/f fluctuation characteristics, and using the level error of the frequency spectrum with respect to the 1/f fluctuation characteristics as a feature quantity, it is possible to determine whether a flame has occurred. This can be identified as a false alarm factor and detected with higher accuracy.

In particular, the fire detection device according to Embodiment 1 has a technical feature in that it determines whether or not a flame has occurred through arithmetic processing based on acoustic data collected within the monitoring area. You can get the following effect.

Effect 1: There is a degree of freedom in the installation position of the microphone, making it easy to add a fire monitoring function as a retrofit or temporarily. That is, fire detection processing in various monitoring areas can be realized with a relatively simple configuration.

Effect 2: Fire detection processing is performed by focusing on the acoustic data generated from the combustion itself, rather than monitoring the smoke generated as a secondary effect of a fire, so fires can be detected more quickly.

Effect 3: By combining the fire detection method according to the present disclosure and the fire detection method using other sensors, false alarm factors can be suppressed and whether or not a flame has occurred can be detected with higher accuracy. For example, when the possibility of a fire is detected by another fire detection method, if the fire detection method of the present disclosure is further applied and the 1/f fluctuation characteristic is not detected, it is determined that the source is a false alarm. can be judged.

Embodiment 2.
In the first embodiment described above, in the frequency band below the standing wave, the level error of the frequency spectrum with respect to the 1/f fluctuation characteristic is calculated as a feature amount, and when the feature amount becomes less than a preset determination threshold, a flame is detected. We have explained the case of detecting the occurrence of. On the other hand, in the second embodiment, a case will be described in which the calculated feature amount is input to a learning model to determine whether or not a flame has occurred.

FIG. 10 is an explanatory diagram showing the overall configuration of a fire detection device according to Embodiment 2 of the present disclosure. The fire detection device according to the second embodiment includes a microphone 10 and a computer 20.

The microphone 10 is installed in the flame monitoring area and collects sounds generated within the monitoring area as acoustic data. The computer 20 is a controller that determines whether a flame has occurred in the monitoring area by performing arithmetic processing on the acoustic data picked up by the microphone 10, and includes a frequency analysis section 21 and a flame detection section 22. It is equipped with

Comparing the configuration shown in FIG. 10 in the second embodiment with the configuration shown in FIG. The difference is that it has a support vector machine 24. Therefore, the differences between Embodiment 1 and Embodiment 2 will be mainly described below.

In the first embodiment, the flame detection unit 22 performs all of the following functions 1 to 3.
Function 1: A function of calculating 1/f fluctuation characteristics in the frequency band below the standing wave with respect to the frequency spectrum calculated by the frequency analysis section 21.
Function 2: A function that calculates the level error of the frequency spectrum for the 1/f fluctuation characteristic as a feature quantity.
Function 3: Compare the feature amount with a preset determination threshold, and if the feature amount is less than the determination threshold, the frequency spectrum calculated by the frequency analysis unit 21 is 1/1 in the frequency band below the standing wave. This function includes f-fluctuation characteristics and outputs a judgment result indicating that a flame has occurred in the monitoring area.

On the other hand, in the second embodiment, the feature amount calculating section 23 performs the portion up to calculating the feature amount using function 1 and function 2. Furthermore, in the support vector machine 24 in the second embodiment, a learning model that uses feature quantities as input parameters is created in advance, and the feature quantities calculated by the feature quantity calculation unit 23 are input to the learning model. Through learning, it is determined whether a flame has occurred.

That is, in the second embodiment, instead of setting a determination threshold value in advance, a learning model is created using the feature amount as an input parameter to determine whether or not a flame has occurred. Therefore, a supplementary explanation of the support vector machine 24 will be given below.

The support vector machine 24 is one of the identification methods using supervised learning, and is a known learning model. In the second embodiment, a learning model is created based on the calculation results of feature amounts in various situations in order to identify whether or not a flame is occurring.

In other words, the feature values based on the acoustic data when the flame to be detected actually occurs, and the feature values based on the acoustic data when the flame does not occur, including false alarm factors, are used as input parameters to support the detection target. A learning model is created in advance using the vector machine 24.

As a result, the support vector machine 24 in the flame detection unit 22 receives the feature calculated by the feature calculation unit 23 as input and uses machine learning to identify whether or not a flame is occurring within the monitoring area. becomes possible.

By having such a configuration, the fire detection device according to the second embodiment creates a learning model in advance, and uses the feature quantities calculated from the acoustic data collected during monitoring as a learned support vector. By applying this as an input parameter to the machine 24, it is possible to detect with higher accuracy whether or not a flame has occurred within the monitoring area by identifying it as a cause of false alarm.

Next, a series of processes executed in the fire detection device according to the second embodiment will be described using a flowchart. FIG. 11 is a flowchart showing a series of processes related to the fire detection method executed in the fire detection device according to Embodiment 2 of the present disclosure.

First, in step S1101, the frequency analysis unit 21 performs sound collection processing of acoustic data generated within the monitoring area via the microphone 10 installed in the monitoring area. Next, in step S1102, the frequency analysis unit 21 calculates a frequency spectrum by performing frequency analysis processing on the acoustic data picked up by the microphone 10.

Next, in step S1103, the feature value calculation unit 23 in the flame detection unit 22 executes 1/f fluctuation characteristic detection processing in the frequency band below the standing wave. Specifically, the feature value calculation unit 23 calculates the peak frequency Pmax and also calculates the 1/f fluctuation characteristic using the above equation (1).

Next, in step S1104, the feature calculation unit 23 calculates a level error between the frequency spectrum calculated in step S1102 and the 1/f fluctuation characteristic calculated in step S1103. As an example of the level error calculation process, the feature amount calculation unit 23 can use the average level error Lave calculated using the above equation (2) as the level error that is the feature amount.

Note that the level error, which is a feature quantity, is not limited to the above equation (2), but can be expressed as [Y(f)-X( It is also possible to use other index values, such as calculating the average squared error of [f)].

Next, in step S1105, the support vector machine 24 creates a learning model in advance, and executes flame detection processing using the machine learning model by inputting the level error calculated as the feature amount in step S1104.

Next, in step S1106, if the support vector machine 24 determines that a flame has occurred as a result of executing the flame detection process using the machine learning model, the process proceeds to step S1107.

On the other hand, if the support vector machine 24 determines that no flame has occurred as a result of executing the flame detection process using the machine learning model, it ends the series of processes.

If the process advances to step S1107, the flame detection unit 22 executes flame detection notification processing. For example, the flame detection unit 22 sends a fire signal to the fire receiver to notify that it has detected the occurrence of a flame, thereby activating fire extinguishing equipment, activating the linked operation of smoke prevention equipment such as shutters, and emergency broadcasting equipment. It is possible to carry out transfers of information to etc.

As described above, according to Embodiment 2, flame-specific features are calculated from the frequency analysis results based on the acoustic data collected in the monitoring area, and a pre-trained machine learning model is used to calculate the characteristics of the monitoring area. It is possible to quantitatively determine whether or not a flame has occurred inside the chamber. Specifically, by focusing on the frequency band below the standing wave, calculating the 1/f fluctuation characteristics, and using the level error of the frequency spectrum with respect to the 1/f fluctuation characteristics as a feature quantity, it is possible to determine whether a flame has occurred. This can be identified as a false alarm factor and detected with higher accuracy.

In other words, instead of setting the judgment threshold in advance, by creating a learning model, the 1/f fluctuation characteristic is calculated by focusing on the frequency band below the standing wave, and the 1/f fluctuation characteristic is calculated. By using the level error of the frequency spectrum as a feature quantity, it is possible to detect with high precision whether or not a flame has occurred, and similarly to the first embodiment, effects 1 to 3 can be achieved.

Although FIG. 10 shows a configuration in which the flame detection unit 22 includes the feature quantity calculation unit 23 and the support vector machine 24, the feature quantity calculation unit 23 and the support vector machine 24 are not included in the flame detection unit 22. An independent configuration is also possible.

In this case, the feature quantity calculation unit specializes in calculating feature quantities, the support vector machine 24 specializes in pre-generating machine learning models, and the flame detection unit 22 specializes in feature quantity calculation when monitoring the monitoring area. The feature quantity calculated by the calculation unit 23 is used as an input parameter for the machine learning model generated in advance by the support vector machine 24, and the machine learning model is specialized for determining whether or not a flame has occurred.

Embodiment 3.
In the first and second embodiments described above, focusing on the frequency band below the standing wave, the 1/f fluctuation characteristic is calculated, the level error of the frequency spectrum with respect to the 1/f fluctuation characteristic is calculated as a feature amount, and this characteristic is calculated. Flame detection processing using the amount was performed. In contrast, in Embodiment 3, when a characteristic similar to a combustion sound is detected in a specific band (low range), conditions with little temporal level fluctuation are further considered to reduce the cause of false alarms. A case of suppression will be explained.

Note that the overall configuration of the fire detection device according to the third embodiment is the same as the overall configuration shown in FIG. 1 in the first embodiment, and a description thereof will be omitted.

FIG. 12 is a diagram showing frequency spectra in which 1/f fluctuation characteristics under various environmental conditions were verified in Embodiment 3 of the present disclosure. Specifically, FIG. 12 shows waveforms related to the following four types of frequency spectra.

First waveform: Frequency spectrum waveform based on acoustic data collected when a fire occurred in the combustion laboratory.Second waveform: First waveform collected when a fire occurred in the combustion laboratory. Frequency spectrum waveform based on acoustic data different from when the door was opened/closed Third waveform: Frequency spectrum waveform based on acoustic data collected when the door was opened/closed Fourth waveform: The moving ceiling was opened/closed Frequency spectrum waveform based on acoustic data collected at the time

In FIG. 12, 1/f fluctuation characteristics passing through a peak frequency in a specific low frequency band are also displayed for each of the first to fourth waveforms. The sound of opening and closing a door or a moving ceiling has a frequency spectrum in the low range that exhibits fluctuations similar to 1/f fluctuation characteristics, and there is a possibility that the sound may be erroneously detected as the occurrence of flame.

Therefore, in Embodiment 3, by further considering conditions with little temporal level fluctuation, erroneous detection of opening/closing sounds of doors, etc. is suppressed, and this will be explained in detail using drawings. . In the following, a case where a flame occurs and a case where a door opening/closing sound occurs will be specifically explained.

FIG. 13 is an explanatory diagram showing a comparison between a spectrogram based on acoustic data when a flame occurs and a spectrogram based on acoustic data when a door opens/closes in Embodiment 3 of the present disclosure.

Specifically, Figure 13 (A) is a spectrogram obtained when the sound of heptane combustion was collected in a laboratory, and Figure 13 (B) is a spectrogram obtained when the sound of a door opening and closing was collected. It is a spectrogram. In the spectrograms shown in FIGS. 13A and 13B, the vertical axis is frequency on a logarithmic scale, the horizontal axis is time, and the temporal change in frequency components is expressed as a continuation of brightness.

FIG. 14 is an explanatory diagram showing a comparison between a frequency spectrum based on acoustic data when a flame occurs and a frequency spectrum based on acoustic data when a door opens/closes in Embodiment 3 of the present disclosure. be. FIG. 14 also shows the 1/f fluctuation characteristics calculated in the low frequency range of 10 Hz or less.

As shown in Figure 14, there is no big difference in the frequency spectrum in the low range between the frequency spectrum caused by flame and the frequency spectrum caused by door opening/closing, and fire detection processing is performed based only on whether or not they have 1/f fluctuation characteristics. If you try to do this, there is a risk that a fire will be falsely detected even when the door is opened or closed.

On the other hand, as shown in FIG. 13, the spectrogram due to flame and the spectrogram due to door opening/closing are significantly different in the temporal changes in frequency components in the low range of 64 Hz or less. Therefore, in the third embodiment, focusing on such a difference, the door opening/closing sound and the combustion sound are quantitatively distinguished by determining the level transition over time regarding the peak frequency in the low range. .

FIG. 15 is an explanatory diagram showing a method for identifying combustion sound and door opening/closing sound based on a level transition of a peak frequency in a low frequency range in Embodiment 3 of the present disclosure. FIG. 15(A) shows the level transition of the peak frequency in the low range regarding the combustion sound shown in FIG. 13(A) above and the door opening/closing sound shown in FIG. 13(B) above. The figure shows the time course of the peak frequency level at .

In addition to extracting the level transition from the spectrogram, the level transition can also be handled by restricting the band using a bandpass filter so that only a desired low band passes, and then extracting the time-series waveform of the peak frequency.

FIG. 15A illustrates a case in which moving average processing is performed for each second on time-series data of peak frequencies, and level changes over 120 seconds are determined. The peak frequencies of door opening/closing sounds and combustion sounds occur at extremely low frequencies (below 1 to 10 Hz). For this reason, if the resolution is increased too much, it becomes difficult to extract characteristic fluctuations. Therefore, in the third embodiment, after extracting the level transition, moving average processing is performed every second, for example, to make it easier to extract the level fluctuation.

The level transition regarding the combustion sound corresponds to the solid line in FIG. 15(A), and the level transition regarding the door opening/closing sound corresponds to the solid line in FIG. 15(A). Comparing the two waveforms, it can be seen that the door opening/closing sound has a greater variation in level transition than the combustion sound.

FIG. 15(B) shows a level fluctuation amount calculated as an index value indicating variation in level transition by calculating the standard deviation from the level transition waveform of FIG. 15(A). As shown in FIGS. 15(A) and 15(B), the amount of level fluctuation calculated based on the level transition of the peak frequency in the low range is used as an index value, and if this index value is less than or equal to the preset judgment fluctuation amount, If so, it is determined that a flame has broken out in the monitored area, thereby eliminating false positive factors such as the sound of doors opening and closing.

By calculating the fluctuation range of the noise source in advance, it is possible to set an appropriate value as the determination fluctuation amount. For example, in the example of FIG. 15(B), the determination variation amount can be set to 6 dB.

Further, not only when the temporal level fluctuation in the low frequency range is less than the determination fluctuation amount, but also when the fluctuation amount gradually increases, it can be determined that a flame has occurred, rather than being a cause of erroneous detection.

Note that the low range in the third embodiment can be set as a frequency band below the standing waves used in the first and second embodiments, but it can also be set as a region including standing waves. I can do it. For example, if the standing wave is 14.4 Hz, a frequency band from f1 = 2.1 Hz to f2 = 15.0 Hz can be selected as the "low-range specific frequency band".

Next, a series of processes executed in the fire detection device according to the third embodiment will be described using a flowchart. FIG. 16 is a flowchart showing a series of processes related to the fire detection method executed in the fire detection device according to Embodiment 3 of the present disclosure.

First, in step S1601, the frequency analysis unit 21 performs sound collection processing of acoustic data generated within the monitoring area via the microphone 10 installed in the monitoring area. Next, in step S1602, the frequency analysis unit 21 calculates a frequency spectrum by performing frequency analysis processing on the acoustic data picked up by the microphone 10.

Next, in step S1603, the flame detection unit 22 executes 1/f fluctuation characteristic detection processing in a specific low frequency band. Specifically, the flame detection unit 22 calculates the peak frequency Pmax and also calculates the 1/f fluctuation characteristic using the above equation (1).

Next, in step S1604, the flame detection unit 22 calculates a level error between the frequency spectrum calculated in step S1602 and the 1/f fluctuation characteristic calculated in step S1603. As an example of the level error calculation process, the flame detection unit 22 can use the average level error Lave calculated using the above equation (2) as the level error.

Note that the level error is not limited to the above equation (2), but can be expressed as [Y(f) - It is also possible to use other index values, such as calculating the mean squared error.

Next, in step S1605, the flame detection unit 22 determines whether the level error calculated in step S1604 is less than or equal to a preset level threshold. Then, if the level error is less than or equal to the level threshold, the flame detection unit 22 proceeds to step S1606.

On the other hand, if the level error exceeds the level threshold, the flame detection unit 22 determines that the frequency spectrum in the specific low frequency band does not include the 1/f fluctuation characteristic and that no flame is generated. Then, the series of processing ends.

If the process proceeds to step S1606, the flame detection unit 22 executes a variation calculation process in order to further calculate the variation in level transition over time regarding the peak frequency in the low range.

Next, in step S1607, the flame detection unit 22 determines whether the variation amount calculated in step S1606 is less than or equal to a preset determination variation amount. Then, if the variation amount is less than or equal to the determined variation amount, the flame detection unit 22 determines that a flame has occurred, and proceeds to step S1608.

On the other hand, if the amount of variation exceeds the determined amount of variation, the flame detection unit 22 determines that no flame is generated, and ends the series of processes.

If the process advances to step S1608, the flame detection unit 22 executes flame detection notification processing. For example, the flame detection unit 22 sends a fire signal to the fire receiver to notify that it has detected the occurrence of a flame, thereby activating fire extinguishing equipment, activating the linked operation of smoke prevention equipment such as shutters, and emergency broadcasting equipment. It is possible to carry out transfers of information to etc.

As described above, according to Embodiment 3, the feature amount peculiar to a flame is calculated from the frequency analysis result based on the acoustic data collected in the monitoring area, and it is quantified whether or not a flame has occurred within the monitoring area. can be judged accurately. Specifically, we calculated the 1/f fluctuation characteristic in the low frequency range, used the level error of the frequency spectrum with respect to the 1/f fluctuation characteristic as the first feature, and focused on the level transition of the peak frequency in the low frequency range. By using the amount of level fluctuation as the second feature quantity, it is possible to detect with higher accuracy whether or not a flame has occurred by distinguishing it from a false alarm factor such as the sound of opening and closing a door.

In particular, the fire detection device according to the third embodiment can obtain the effects 1 to 3 described in the first and second embodiments, as well as the following unique effect 4.
Effect 4: Regarding acoustic data, by combining fire detection processing based on frequency analysis and fire detection processing based on time-series data analysis, it is possible to detect frequencies similar to 1/f fluctuation characteristics, such as the sound of doors and windows opening and closing, and wind noise. Erroneous detection of sounds having a spectrum can be suppressed.

Embodiment 4.
In the third embodiment described above, the flame occurs when the level error, which is the first feature amount, is less than or equal to the level threshold value, and the amount of variation, which is the second feature amount, is less than or equal to the determined variation amount. We have explained the case of detecting. In contrast, in the fourth embodiment, a case will be described in which the calculated first feature amount and second feature amount are input to a learning model to determine whether or not a flame has occurred. The differences between Embodiment 3 and Embodiment 4 will be mainly described below.

Note that the overall configuration of the fire detection device according to the fourth embodiment is the same as the overall configuration shown in FIG. 10 in the second embodiment, and a description thereof will be omitted.

In the third embodiment, the flame detection unit 22 performs all of the following functions 1 to 4.
Function 1: A function to calculate 1/f fluctuation characteristics in a specific low frequency band with respect to the frequency spectrum calculated by the frequency analysis section 21.
Function 2: A function that calculates the level error of the frequency spectrum with respect to the 1/f fluctuation characteristic as the first feature quantity.
Function 3: Compare the first feature amount and a preset level threshold value, and if the first feature amount is less than or equal to the level threshold value, further perform a process of calculating the amount of variation in level transition in the low range, A function that calculates the amount of variation as a second feature amount.
Function 4: A function that outputs a determination result that a flame has occurred in the monitoring area when the second feature amount is less than or equal to the determination variation amount.

On the other hand, in the fourth embodiment, the part that calculates the first feature amount by function 1 and function 2, and the part that calculates the second feature amount of function 3 are transferred to the feature amount calculation unit 23. It is being carried out in Furthermore, in the support vector machine 24 according to the fourth embodiment, a learning model is created in advance using the first feature amount and the second feature amount as input parameters, and the feature amount calculation unit 23 Using the calculated first feature amount and second feature amount as input, machine learning is used to determine whether a flame has occurred.

That is, in the fourth embodiment, instead of setting the level threshold and the determination variation amount in advance, by creating a learning model using the first feature amount and the second feature amount as input parameters, Determining whether or not a flame has occurred. Therefore, a supplementary explanation of the support vector machine 24 will be given below.

The support vector machine 24 is one of the identification methods using supervised learning, and is a known learning model. In the fourth embodiment, a learning model is created based on the first feature amount and the second calculation result in various situations in order to identify whether or not a flame is occurring.

That is, the first feature amount and the second feature amount are based on the acoustic data when the flame to be detected actually occurs, and the first feature amount is based on the acoustic data when no flame occurs, including false alarm factors. A learning model is created in advance by the support vector machine 24 using the feature amount and the second feature amount as input parameters.

As a result, the support vector machine 24 in the flame detection unit 22 inputs the first feature amount and the second feature amount calculated by the feature amount calculation unit 23, and determines whether a flame is occurring within the monitoring area. It becomes possible to identify this using machine learning.

By having such a configuration, the fire detection device according to the fourth embodiment creates a learning model in advance, and uses the first feature quantity and the second feature quantity calculated from the acoustic data collected during monitoring. By applying the feature quantity as an input parameter to the trained support vector machine 24, it is possible to detect with higher accuracy whether or not a flame has occurred within the monitoring area by identifying it as a cause of false alarm.

Next, a series of processes executed in the fire detection device according to the second embodiment will be described using a flowchart. FIG. 17 is a flowchart showing a series of processes related to the fire detection method executed in the fire detection device according to Embodiment 4 of the present disclosure.

First, in step S1701, the frequency analysis unit 21 performs sound collection processing of acoustic data generated within the monitoring area via the microphone 10 installed in the monitoring area. Next, in step S1702, the frequency analysis unit 21 calculates a frequency spectrum by performing frequency analysis processing on the acoustic data picked up by the microphone 10.

Next, in step S1703, the feature value calculation unit 23 in the flame detection unit 22 executes a 1/f fluctuation characteristic detection process in a specific low frequency band. Specifically, the feature value calculation unit 23 calculates the peak frequency Pmax and also calculates the 1/f fluctuation characteristic using the above equation (1).

Next, in step S1704, the feature calculation unit 23 calculates a level error between the frequency spectrum calculated in step S1702 and the 1/f fluctuation characteristic calculated in step S1703. As an example of the level error calculation process, the feature amount calculation unit 23 can use the average level error Lave calculated using the above equation (2) as the level error that is the first feature amount.

Note that the level error, which is the first feature, is not limited to the above equation (2), but can be expressed as [Y(f) It is also possible to use other index values, such as calculating the average squared error of [-X(f)].

Next, in step S1705, the flame detection unit 22 executes a variation calculation process to further calculate the variation in the level transition over time with respect to the peak frequency in the low range as a second feature quantity.

Next, in step S1706, the support vector machine 24 creates a learning model in advance, and uses the level error calculated as the first feature amount in step S1704 and the variation amount calculated as the second feature amount in step S1705. With this as input, flame detection processing using a machine learning model is executed.

Next, in step S1707, if the support vector machine 24 determines that a flame has occurred as a result of executing the flame detection process using the machine learning model, the process proceeds to step S1708.

On the other hand, if the support vector machine 24 determines that no flame has occurred as a result of executing the flame detection process using the machine learning model, it ends the series of processes.

If the process advances to step S1708, the flame detection unit 22 executes flame detection notification processing. For example, the flame detection unit 22 sends a fire signal to the fire receiver to notify that it has detected the occurrence of a flame, thereby activating fire extinguishing equipment, activating the linked operation of smoke prevention equipment such as shutters, and emergency broadcasting equipment. It is possible to carry out transfers of information to etc.

As described above, according to Embodiment 4, flame-specific features are calculated from the frequency analysis results based on the acoustic data collected in the monitoring area, and a machine learning model learned in advance is used to calculate the characteristics of the monitoring area. It is possible to quantitatively determine whether or not a flame has occurred inside the chamber. Specifically, by focusing on the frequency band below the standing wave, calculating the 1/f fluctuation characteristics, and using the level error of the frequency spectrum with respect to the 1/f fluctuation characteristics as a feature quantity, it is possible to determine whether a flame has occurred. This can be identified as a false alarm factor and detected with higher accuracy.

In other words, by creating a learning model instead of setting the level threshold value and determination variation amount in advance, it is possible to detect with high accuracy whether or not a flame has occurred, as in the third embodiment. , the same effects 1 to 4 as in the third embodiment can be achieved.

Although FIG. 10 shows a configuration in which the flame detection unit 22 includes the feature quantity calculation unit 23 and the support vector machine 24, the feature quantity calculation unit 23 and the support vector machine 24 are not included in the flame detection unit 22. An independent configuration is also possible.

In this case, the feature quantity calculation unit specializes in calculating the first feature quantity and the second feature quantity, the support vector machine 24 specializes in pre-generating the machine learning model, and the flame detection unit 22 When monitoring the monitoring area, the first feature amount and the second feature amount calculated by the feature amount calculation unit 23 are used as input parameters for the machine learning model generated in advance by the support vector machine 24, and the flame generation model is used as input parameters for the machine learning model generated in advance by the support vector machine 24. It will be specialized in determining whether or not.

In addition, in the above-mentioned Embodiment 2 and Embodiment 4, the case where a support vector machine is used as a model creation unit that creates a machine learning model has been described, but the model creation unit is not limited to this. . Similar effects can also be achieved by configuring the model creation unit using other means such as a neural network to create a machine learning model.

10 microphone, 20 computer, 21 frequency analysis section, 22 flame detection section, 23 feature quantity calculation section, 24 support vector machine (model creation section).

Claims (4)

A microphone that collects sounds generated in the monitoring area as acoustic data,
a frequency analysis unit that calculates a frequency spectrum by frequency-analyzing the acoustic data picked up by the microphone;
Determining whether or not the frequency spectrum calculated by the frequency analysis unit includes a 1/f fluctuation characteristic in a frequency band below a standing wave in the monitoring area, a flame detection unit that detects that a flame has occurred in the monitoring area when it is determined that a flame has occurred in the monitoring area. The fire detection device according to claim 1, wherein the standing wave is specified in advance based on acoustic data collected in the monitoring area, or specified in advance based on the size of the monitoring area. If the 1/f fluctuation characteristic is included in a frequency band below the standing wave, the flame detection section further calculates a level error of the frequency spectrum with respect to the 1/f fluctuation characteristic, and calculates the level error of the frequency spectrum with respect to the 1/f fluctuation characteristic. The fire detection device according to claim 1 or 2, wherein if the error is less than or equal to a preset determination threshold, it is detected that a flame has occurred in the monitoring area. A microphone that collects sounds generated in the monitoring area as acoustic data,
a frequency analysis unit that calculates a frequency spectrum by frequency-analyzing the acoustic data picked up by the microphone;
Determining whether or not the frequency spectrum calculated by the frequency analysis unit includes a 1/f fluctuation characteristic in a frequency band below a standing wave in the monitoring area, If it is determined that there is a flame, a feature amount calculation unit that calculates a level error of the frequency spectrum with respect to the 1/f fluctuation characteristic and outputs the calculated level error as a feature amount used for determining flame occurrence;
The above-mentioned frequency a model creation unit that creates a machine learning model for determining whether or not the flame has occurred by using the feature quantity calculated by the feature quantity calculation unit as a parameter based on the frequency analysis result by the analysis unit; ,
A fire detection device comprising: a flame detection section that determines whether or not a flame has occurred by using the feature amount calculated by the feature amount calculation section as input to the machine learning model when monitoring the monitoring area.