JP2024024586A - Wildlife damage countermeasure device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a wildlife damage countermeasure device that can efficiently drive off wildlife of offending species that attacks an object.

SOLUTION: A wildlife damage countermeasure device (10) uses voice data output by a sound collector (1) that collects sounds in an object region and a classifier generated using noise data that indicates environmental sounds to determine whether the sounds collected by the sound collector includes calls of wildlife of offending species. The wildlife damage countermeasure device includes a voice discrimination and light irradiation part (2) for controlling a light irradiation part to irradiate the object region with light when the sounds collected by the sound collector (1) includes the calls.

SELECTED DRAWING: Figure 1

COPYRIGHT: (C)2024,JPO&INPIT

Description

The present invention relates to techniques for preventing damage to birds and animals.

Various types of damage caused by birds and animals have been reported. For example, approximately 3.02 billion yen (in 2020) was reported as damage to crops caused by birds nationwide. By type of bird, crows (approximately 1.38 billion yen) were the most common, followed by ducks (approximately 510 million yen) (Non-Patent Document 1).

In order to prevent bird invasion and feeding damage to agricultural land, fish farms, etc., bird nets that physically isolate birds from objects are widely used as the most effective measure. However, if the target area is large, such as a farm field, it is not easy to install bird nets that cover the entire area. In addition, there have been some cases of inappropriate management, such as using nets with large mesh sizes, partially damaged nets, or nets being laid with gaps between seams and the ground. In some areas, the death of wild birds in nets has become a serious problem (Non-Patent Document 2).

In addition to bird nets, devices and systems have been devised to scare birds and beasts away from specific areas such as agricultural land. As a device that uses light, a harmful bird and beast extermination device has been proposed that irradiates a field with a laser beam to protect the field from eating damage by ducks at night when planting lotus seeds (Patent Document 1). Furthermore, a device has been proposed that effectively and efficiently intimidates and drives away harmful birds and beasts by emitting gunshot sounds or irradiating laser beams when they approach (Patent Document 2). Patent Document 2 describes that a threatening device is activated when it is determined that a harmful bird or beast is approaching based on images taken by a camera.

Furthermore, in order to support existing pest and beast extermination devices and realize a pest control device whose effectiveness is unlikely to diminish, a sound source recognition and identification device has been devised that can detect and recognize the intrusion and incoming pests using sounds at a specific fixed frequency. (Patent Document 3). In addition, in fields other than bird and beast damage countermeasures, devices and systems have been devised that identify the type of animal based on its sound (Patent Document 4).

Utility model registration No. 3217837 Utility model registration No. 3213836 Utility model registration No. 3219134 Japanese Patent Application Publication No. 2002-304191

Ministry of Agriculture, Forestry and Fisheries “Situation of crop damage caused by wild birds and animals nationwide (2020)” Rural Development Bureau, Rural Policy Department, Bird and Beast Countermeasures and Rural Environment Division, 2021 (online) https: //www.maff.go.jp/j /seisan/tyozyu/higai/hogai_zyoukyou/index.html Haruhiko Asuka, Susumu Ikeno, and Choichi Watanabe, “Status of damage caused by bird netting in lotus fields in Ibaraki Prefecture,” Strix Vol. 27, 113-124, 2011

However, the technique described in Patent Document 1 has a problem in that not only the inflicting bird species but also other non-inflicting wild birds are subject to expulsion. Furthermore, the technique described in Patent Document 2 judges the approach of harmful birds and beasts based on images, so there is a problem that it is difficult to judge whether it is a harmful species at night or the like. Furthermore, the technique described in Patent Document 3 has a problem in that environmental sounds among the sensed surrounding sounds become noise, making it difficult to detect the sounds of the target harmful species of birds and beasts. As described above, the above-mentioned conventional techniques have room for improvement in terms of more efficiently driving away harmful species of birds and beasts.

An object of one aspect of the present invention is to realize a bird and beast damage control device that can more efficiently drive away harmful birds and beasts that cause harm to a target object.

In order to solve the above-mentioned problems, the bird and beast damage control device according to aspect 1 of the present invention is provided using sound data output by a sound collection unit that collects sound in a target area and noise data representing environmental sounds. a discriminating unit that uses a classifier to determine whether the sound collected by the sound collection unit includes the cry of a bird or beast of a perpetrating species; and an irradiation control section that controls the light irradiation section to irradiate the target area with light.

According to the above configuration, a classifier generated using noise data representing environmental sounds is used to determine whether the collected sounds include the calls of the perpetrator species. irradiates the target area with light when the sound of a voice is included. This makes it possible to drive away harmful species of birds and beasts more efficiently.

In the wildlife damage control device according to aspect 2 of the present invention, in aspect 1, the irradiation control section may change the irradiation direction of the light at a predetermined speed or a predetermined angular velocity. According to the above configuration, harmful species that have invaded the target area can be more effectively expelled.

In the bird and beast damage control device according to aspect 3 of the present invention, in the aspect 2, the irradiation control unit estimates the direction of the sound source of the cry, and adjusts the irradiation direction of the light so that the irradiation position of the light is located at the sound source. It may be moved in the direction closer to it. By bringing the irradiation position closer to the sound source, that is, the position of the harmful species that has invaded the target area, the harmful species that have invaded the target area can be more effectively driven away.

The bird and beast damage control device according to aspect 4 of the present invention, in any one of aspects 1 to 3, further includes a power supply section including an independent power source, and the discrimination section and the irradiation control section are configured such that the power supply section It may be operated by supplied power.

By being equipped with a power supply section that includes an independent power source, it can be installed even in cultivated land where there is no power source. Furthermore, it is easy to install and carry, and when there is no longer a need for expulsion, such as when the harvest of agricultural and marine products is completed within the target area, it can be removed and moved to another target area for use.

Aspect 5 of the present invention provides a bird and animal damage control device according to any one of aspects 1 to 4, wherein the recording unit records the sound collected by the sound collection unit and the date and time when the light irradiation unit irradiated the light. You may further have. According to the above configuration, it becomes possible to verify the effect, analyze, retrain the classifier 23, etc. using recorded data.

In the bird and animal damage control device according to aspect 6 of the present invention, in any one of aspects 1 to 5, the target area may be a lotus field where lotus root is cultivated. According to the above configuration, harmful species that invade lotus fields and cause harm can be more effectively expelled.

In the bird and beast damage control device according to aspect 7 of the present invention, in any one of the aspects 1 to 6, the inflicting species of birds and beasts are species that feed at night in the target area (reed duck, reed duck, wigeon, mallard duck, etc.). It may include at least one of the following: spot-billed duck, shovel duck, pintail duck, teal, white-billed duck, coot, gray heron, night heron, sandpiper, etc.). According to the above configuration, the damage caused by the birds can be effectively suppressed.

In the bird and beast damage control device according to aspect 8 of the present invention, in the aspect 7, the target area is a lotus field where lotus root is cultivated, and the damaging species of birds and beasts include at least one of a mallard duck and a coot. But that's fine. According to the above configuration, it is possible to more effectively drive away at least one of mallard ducks and coots that invade the lotus fields.

A bird and beast damage control device according to aspect 9 of the present invention is provided in any one of aspects 1 to 8, wherein the discrimination unit analyzes the audio data and generates a plurality of parameters representing feature amounts of the audio data, The above-mentioned discrimination may be performed based on the position in the feature amount space indicated by the plurality of generated parameters. According to the above configuration, by performing the above discrimination using a plurality of parameters that are the analysis results of audio data, it is possible to accurately discriminate whether it is a harmful species.

In the bird and beast damage control device according to aspect 10 of the present invention, in aspect 9, the plurality of parameters may include linear predictive coefficients calculated by linear predictive coding (LPC).

Since the amount of calculation for calculating the coefficients of LPC is small compared to the amount of calculation of, for example, a convolutional neural network (CNN), the hardware can be configured with a simple structure. Therefore, according to the present embodiment, countermeasures against damage from birds and animals can be effectively taken with a simple configuration.

In the bird and beast damage control device according to aspect 11 of the present invention, in aspect 10, the plurality of parameters may include an autocorrelation coefficient of the audio data. By including not only the LPC parameters but also the autocorrelation coefficient in the input of the classifier 23, the injurious species can be determined more accurately with a simple configuration.

In the bird and beast damage control device according to aspect 12 of the present invention, in any one of aspects 9 to 11, the discrimination section performs the discrimination using a machine learning model that receives the plurality of parameters as input and outputs a discrimination result. It's okay. According to the above configuration, by performing the above discrimination using a machine learning model, it is possible to accurately discriminate the perpetrator type.

In the bird and beast damage control device according to aspect 13 of the present invention, in aspect 12, the machine learning model may be a model constructed by random forest learning. According to the above configuration, by performing the above discrimination using a model constructed by random forest learning, it is possible to accurately discriminate the perpetrator species.

Aspect 14 of the bird and beast damage control device of the present invention is the above aspect 12 or 13, wherein the machine learning model uses teacher data in which a voice representing the cry of a pest species of bird or beast is embedded in the voice data representing the environmental sound. It may also be a trained model that has been machine learned.

According to the above configuration, by using the teacher data in which the calls of the perpetrator species are embedded in the audio data representing environmental sounds, even if the sounds collected in the target area include environmental sounds (noise), , it is possible to more accurately identify the calls of inflicting species.

The wildlife damage control device according to aspect 15 of the present invention may further include the sound collection section and the light irradiation section in any one of the aspects 1 to 14. According to the above configuration, a classifier generated using noise data representing environmental sounds is used to determine whether the collected sounds include the calls of the perpetrator species. irradiates the target area with light when the sound of a voice is included. This makes it possible to drive away harmful species of birds and beasts more efficiently.

In the bird and beast damage control device according to aspect 16 of the present invention, in any one of aspects 1 to 15, the light irradiation section may irradiate laser light.

Laser light is effective at driving away birds and beasts, especially at night. Therefore, by irradiating the laser beam, harmful species that have invaded the target area can be more effectively expelled.

A learning device according to aspect 17 of the present invention includes a generation unit that generates learning audio data in which audio data representing the cry of a perpetrating bird or beast is embedded in audio data representing an environmental sound, and a learning device generated by the generation unit. a parameter generation unit that generates a plurality of parameters representing feature quantities by voice analyzing voice data for use in the field, and a label indicating whether or not the plurality of parameters generated by the parameter generation unit and the cry of the perpetrating species are included. A learning unit that uses the learning data to generate a machine learning model that inputs a plurality of parameters representing feature quantities of the audio data and determines whether the audio data includes the cry of a bird or beast of a perpetrating species.

By using the machine learning model generated by the learning device, it is possible to accurately identify the calls of the injurious species.

A program according to aspect 18 of the present invention causes a computer to use sound data output by a sound collection unit that collects sound in a target area and a classifier generated using noise data representing environmental sounds to a determination unit that determines whether the sound collected by the sound unit includes the cry of a bird or beast of an inflicting species; This is a program for functioning as an irradiation control unit that irradiates light onto a target area.

The bird and beast damage control device according to aspect 19 of the present invention may further include a photographing device that photographs the irradiation direction of the light and outputs image data representing the photographed image in the aspect 1 or 2. According to the above configuration, the photographed image can be used for verifying the effectiveness of driving away birds and beasts by the bird and beast damage prevention device.

In the bird and beast damage control device according to aspect 20 of the present invention, in the aspect 19, the discrimination section uses image data of the photographed image photographed by the photographing device in addition to the audio data outputted by the sound collection section. It may be determined whether a harmful species has flown into the target area. According to the above configuration, detection accuracy can be made higher by performing detection using image data in addition to detection of the injurious species using audio.

Aspect 21 of the present invention provides a bird and beast damage control device according to aspect 12, in which the machine learning model is pre-trained with a set of audio data representing environmental sounds, and then generated from audio data including the calls of the perpetrating species. The model may be a trained model that has been retrained using training data that includes parameters that have been trained. According to the above configuration, after pre-training a discrimination model on a large-scale environmental sound dataset, the discrimination model is retrained (fine-tuning) on a dataset for discriminating the calls of the perpetrator species. It is possible to generate a discrimination model with high discrimination accuracy using data (audio data including the calls of the offending species).

In the bird and beast damage control device according to aspect 22 of the present invention, in aspect 12, the machine learning model may be a model whose weight is reduced by at least one of quantization and compression. According to the above configuration, by reducing the weight of the discriminant model through at least one of quantization and compression, the amount of calculation required for model execution can be reduced.

The bird and beast damage control device according to each aspect of the present invention may be realized by a computer, and in this case, the bird and beast damage control device is operated by operating the computer as each part (software element) included in the bird and beast damage control device. A control program for a bird and beast damage prevention device that is realized by a computer, and a computer-readable recording medium on which the program is recorded also fall within the scope of the present invention.

According to one aspect of the present invention, it is possible to more effectively drive away harmful species of birds and animals that cause harm to a target object.

1 is a diagram schematically showing an installed state of the wildlife damage control device according to the first embodiment; FIG. FIG. 3 is an external view of the sound collecting section according to the first embodiment, viewed from the bottom. FIG. 2 is an external view of the outside of the case of the voice identification/light irradiation unit according to the first embodiment, viewed from the bottom. FIG. 3 is a top view of the inside of the case of the voice identification/light irradiation unit according to the first embodiment. FIG. 3 is a diagram showing how components housed inside the case of the voice identification/light irradiation unit according to the first embodiment are connected. FIG. 3 is a schematic diagram of the voice identification/light irradiation unit according to the first embodiment, viewed from above. FIG. 3 is a schematic side view of the voice identification/light irradiation unit according to the first embodiment. FIG. 3 is a diagram showing the configuration of a power supply unit according to the first embodiment. FIG. 2 is a block diagram showing an example of a functional configuration of a voice identification/light irradiation unit according to the first embodiment. This is a graph showing changes in the number of injurious species captured by fixed-point cameras. 1 is a block diagram showing an example of a configuration of a learning device according to a first embodiment. FIG. FIG. 3 is a diagram for explaining a specific example of the teacher data generation process according to the first embodiment. It is a figure which shows an example of the structure of the wildlife damage prevention device based on a modification. It is a figure which shows the structure of the voice identification/light irradiation part based on a modification. It is a figure which shows the structure of the wildlife damage prevention device based on a modification. It is a figure which shows the structure of the wildlife damage prevention device based on a modification. FIG. 7 is a diagram showing a specific example of teacher data used for training a discriminant model according to a modification. FIG. 7 is a diagram illustrating a specific example of generation of a discriminant model and implementation on a computer according to a modification.

[Embodiment 1]
Hereinafter, one embodiment of the present invention will be described in detail.

<Configuration of bird and beast damage control device>
The bird and beast damage control device 10 according to the present embodiment is a device for taking measures against bird and beast damage in a target area. The bird and beast damage control device 10 is an example of the bird and beast damage control system according to the present specification. In the following, a mode in which the bird and animal damage prevention device 10 is used to prevent damage caused by ducks and the like from eating lotus roots at night will be described. In this embodiment, the target area is a lotus field where lotus roots are cultivated. The target area is particularly flooded lotus fields before harvest. For example, in the lotus fields in Ibaraki Prefecture, which produce 54% of the total lotus root shipments nationwide (in 2020), at least 13 species of birds that feed at night have been confirmed (evil duck, reed duck, wigeon, mallard, spot-billed duck, and shovel duck). , pintail duck, teal, white-winged duck, coot, gray heron, night heron, sandpiper), each species has a distinctive vocalization. Among these, two species (mallard and coot) have been confirmed to damage lotus roots growing in mud (Miyuki Mashiko, Yasuhiro Yamaguchi, Yoshiko Yoshida, ``Lotus roots in mud are eaten by ducks''). "Suffering from feeding damage from other species: Confirmation through field tests" Journal of the Ornithological Society of Japan 71, (2), 153-169, 2022). In the present embodiment, the offending species targeted for countermeasures against wildlife damage include at least one of mallard ducks and coots.

FIG. 1 is a diagram schematically showing the installed state of the wildlife damage control device 10. As shown in FIG. In FIG. 1, a wildlife damage control device 10 is installed near cultivated land (lotus fields before harvest) where damage is to be prevented. Please note that the lotus field is flooded and lotus roots are grown in the mud, making it impossible to enter. Attach it to a single pipe, etc. that was used as a support when it was installed. At that time, in order to avoid enveloping the injurious species that were expelled by the light irradiation of the bird and beast damage prevention device 10 when they fly away, we selected a place without a bird net and installed it after coordinating with farmers etc. conduct.

As shown in FIG. 1, the bird and beast damage prevention device 10 includes, as a basic configuration, a sound collection section 1, a voice identification/light irradiation section 2, and a power supply section 3 that supplies power to these sections. The power supply section 3 may use a household power source, but in cultivated areas where there is no power source, a plastic case 33 that houses the solar panel 31 and a battery that stores the electric power generated by the solar panel 31 is used. Equipped with Hereinafter, as shown by the arrows in FIG. 1, the vertical direction of the wildlife damage prevention device 10 will be defined.

(Sound collection section)
FIG. 2 is an external view of the sound collecting section 1 viewed from the bottom. The sound collection unit 1 collects sound in the target area. In the example of FIG. 2, a rain shield plate 12 having a diameter twice or more larger than that of the microphone array 11 is attached downward to the top of the microphone array 11 to provide rain protection. The rain shield plate is, for example, a plastic plate, and more specifically, as an example, is a flower pot stand plate. Furthermore, the microphone array 11 and the rain shield plate 12 are covered with a cover 13. The cover 13 is a rainproof/windproof cloth cover. The cover 13 is preferably made of a thin nylon or polyurethane material that has excellent elasticity and water absorption and quick drying properties. The microphone array 11 is connected to the voice identification/light irradiation unit 2 via a USB cable 14 . After fixing these to the sound collecting part support rod 15, the sound collecting part 1 is fixed to the pillar using a sound collecting part fixing tool 16 so as to be parallel to the ground.

The microphone array 11 includes four microphone arrays that can estimate the direction of sound from four directions. The microphone array 11 collects sounds within a range of about 100 meters around it. The four microphone holes of the microphone array 11 are covered with a breathable waterproof seal. When a single microphone is used, the accuracy of detecting the calls of the offending species is greatly reduced because the collected sounds are significantly modified by the long-wave dynamics of the background. On the other hand, according to the present embodiment, since the control chip of the microphone array 11 produces sounds with noise reduction and echoes removed, it is possible to improve the accuracy of cries detection.

(Voice identification/light irradiation part)
FIG. 3 is an external view of the outside of the case of the voice identification/light irradiation unit 2 viewed from below. The voice identification/light irradiation unit 2 is an example of the wildlife damage prevention device according to the present specification. The case 211 is a waterproof case made of plastic. A hemispherical dome 212 made of acrylic is provided at the bottom of the case 211. The dome 212 is a transparent member that accommodates the parts of the voice identification/light irradiation section 2. A dome 212 protrudes from the bottom of the case 211, and a laser housing 221 and connection wires 222, 223 are housed within the dome 212. After fixing the case 211 to the support rod 213, the voice identification/light irradiation unit 2 is fixed to the pillar using the fixture 214 so that the lower surface of the voice identification/light irradiation unit 2 is parallel to the ground.

FIG. 4 is a diagram showing the inside of the case 211 viewed from above, and FIG. 5 is a diagram showing how components housed inside the case 211 are connected. A computer 227 is used to detect the cry. The computer 227 is, for example, a single board computer, and includes, for example, a processor such as a CPU (central processing unit) and a memory. The computer 227 has sufficient CPU processing power to process audio input every 0.5 seconds, for example, and fragments the input sound using a signal-to-noise ratio, etc., and then generates a cry for the fragmented sound. Perform detection. The expansion board 228 is an expansion board for controlling the stepping motor and laser. An expansion board 228 added to the computer 227 allows the voice recognition/light irradiation unit 2 to be controlled via general-purpose input/output. Note that the computer 227 is not limited to a computer including a processor and a memory, and may be a microcontroller, for example.

In addition, in order to make effectiveness verification more efficient, it is necessary to correctly obtain and store the date and time when the calls of the offending species were detected and the date and time when light irradiation was performed. Since an Internet connection is required for the computer 227 to correct the time by itself, by providing a GPS receiver 224 instead, the time can be corrected at any time even in places without an Internet connection, and the correct time can be recorded. In addition, the sounds collected when the calls of the offending species are detected can be stored in the computer 227. In other words, the wildlife damage control device 10 includes a computer 227 that records the sound collected by the sound collection unit 1 and the date and time when the sound identification/light irradiation unit 2 irradiated light. The computer 227 is an example of a recording unit according to this specification.

FIG. 6 is a schematic diagram of the voice identification/light irradiation section 2 viewed from the top, and FIG. 7 is a schematic diagram of the voice identification/light irradiation section 2 viewed from the side (a surface perpendicular to the top surface). The light beam is emitted through the dome 212 by a laser module 226 that includes a marking laser that emits a green laser beam in an approximately 90 degree fan. The laser module 226 is an example of a light irradiation unit according to this specification. The light beam emitted by the laser module 226 is, for example, class 1, class 1M, class 2, or class 2M. The light beam emitted by the laser module 226 reaches approximately 50 meters away at night. Further, by attaching a stepping motor 225 with a panning function for moving the laser module 226 inside the case 211, the light beam can be rotated 360 degrees. The laser module 226 can be covered with a laser housing 221 made of aluminum material to facilitate dissipation of its heat.

However, the laser light emitted by the laser module 226 is not limited to green, and the laser module 226 may emit laser light of other wavelengths. Further, the shape of the laser light emitted by the laser module 226 is not limited to a fan shape, and the laser module 226 may emit laser light in other shapes.

Pan/tilt control is required when using point-shaped laser beams to irradiate the entire target area, and plastic pan/tilt units can easily become hot and damaged, whereas fan-shaped laser beams are directed vertically. Setting it to irradiate eliminates the need for a tilt function, so it can be handled using only a pan function using an inexpensive stepping motor 225, reducing power consumption and allowing stable operation over a long period of time.

The stepping motor 225 is connected to a motor driver 229 and the motor driver 229 is connected to an expansion board 228 by connecting lines, and the expansion board 228 is connected to the appropriate general purpose input/output of the computer 227.

(Power supply part)
FIG. 8 is a diagram showing the configuration of the power supply section 3. As shown in FIG. The power supply unit 3 supplies power to the wildlife damage control device 10 in cultivated land or the like where there is no power supply. Electric power generated by the solar panel 31 is stored in a battery 32 via a solar charge controller 35, and is supplied to the voice identification/light irradiation unit 2 via a connecting line 34 via a DC/DC converter 36. In other words, the wildlife damage prevention device 10 has a power supply unit 3 including a solar cell, and the voice identification/light irradiation unit 2 operates with the power supplied by the power supply unit 3.

(Functional configuration of voice identification/light irradiation unit)
FIG. 9 is a block diagram showing an example of the functional configuration of the voice identification/light irradiation section 2. As shown in FIG. In the figure, the voice identification/light irradiation section 2 includes a discrimination section 21 and an irradiation control section 22. The determination unit 21 and the irradiation control unit 22 are realized, for example, by a processor of the computer 227 reading and executing a computer program stored in a memory.

(Discrimination part)
The determination unit 21 uses the audio data output by the sound collection unit 1 and the classifiers 23_1 and 23_2 to determine whether the sounds collected by the sound collection unit 1 include the cries of a bird or beast of a perpetrating species. do. The audio data to be determined by the determining unit 21 in one determination process is, for example, audio data with a time length of 0.5 seconds. In this case, the determination unit 21 repeatedly performs the determination process on the audio data having a time length of 0.5 seconds over the operating period of the wildlife damage control device 10.

(classifier)
The classifier 23_1 is a classifier for determining whether the sound data obtained by the discriminating unit 21 includes the call of a coot. This is a classifier for determining whether cries are included. Below, if there is no need to distinguish between the classifier 23_1 and the classifier 23_2, they will be referred to as the "classifier 23." The classifier 23 is a trained model constructed by supervised machine learning. In this embodiment, the classifier 23 is a trained model constructed using a random forest method. The learning data used for learning by the classifier 23 includes noise data representing environmental sounds. In other words, the classifier 23 is a classifier generated using noise data representing environmental sounds and audio data of the target species.

(classifier input)
The input of the classifier 23 includes, for example, a plurality of parameters representing feature amounts of the audio data. Here, the plurality of parameters include, for example, 16 linear predictive coefficients calculated by linear predictive coding (LPC) and 17 coefficients, including an autocorrelation coefficient. According to LPC, one piece of audio data can be expressed as a polynomial, so multiple parameters can be obtained from the audio data. However, the plurality of parameters representing the feature amount are not limited to these, and may include other parameters representing the feature amount of the audio data.

(classifier output)
Further, the output of the classifier includes, for example, a determination result as to whether or not the audio data includes the cry of the perpetrator species. In other words, the discrimination unit 21 performs the above discrimination using a machine learning model that receives a plurality of parameters as input and outputs a discrimination result.

(Specific example of discrimination processing by the discrimination unit)
As an example, the determining unit 21 performs the determining process according to the following procedure. First, the determination unit 21 analyzes data that includes sounds other than noise among audio data having a time length of 0.5 seconds, and generates a plurality of parameters representing feature amounts of the audio data. More specifically, the determination unit 21 calculates 16 linear prediction coefficients from the audio data using LPC. Further, the determining unit 21 calculates an autocorrelation coefficient of the audio data.

The discriminator 21 inputs the calculated 17 parameters (16 linear prediction functions and 1 autocorrelation coefficient) to the classifier 23 and obtains the output of the classifier 23, thereby determining the target for discrimination. Determine whether the audio data contains the calls of the invading species.

(irradiation control section)
The irradiation control unit 22 controls laser light irradiation based on the determination result of the determination unit 21. That is, when the sound collected by the sound collection unit 1 includes the cry, the irradiation control unit 22 controls the laser module 226 to irradiate the target area with light. In this embodiment, the irradiation control unit 22 performs light irradiation without determining a direction. For example, the voice identification/light irradiation unit 2 controls the stepping motor 225 to change the direction of light irradiation at a predetermined speed or a predetermined angular velocity, and gradually moves the light irradiation position in the target area.

The laser light emitted by the voice identification/light irradiation unit 2 moves slowly in the target area and illuminates withered stems and leaves of agricultural products grown in the target area, the water surface, and the like. Injury species of birds and animals that have invaded the target area will be alert to their surroundings due to the irradiated laser light, the shaking of the water surface illuminated by the laser light, and the shadows of crops caused by the laser light irradiation, and will be able to move out of the target area. Moving. This makes it possible to drive away harmful species that have invaded the target area.

At this time, the voice identification/light irradiation unit 2 may gradually move the light irradiation position in a direction closer to the injurious species. In this case, the voice identification/light irradiation unit 2 estimates the direction of the inflicting species (the direction of the sound source of the sound) based on the signal output by the microphone array 11, and determines the direction of the light irradiation so that the light irradiation position is located at the sound source. It may be moved in the direction closer to it.

Further, the audio identification/light irradiation unit 2 records the audio collected by the sound collection unit 1 and the date and time when the audio identification/light irradiation unit 2 irradiated the light. The recorded data is used, for example, for verification and analysis of the expulsion effect by the wildlife damage control device 10, relearning of the classifier 23, and the like.

The applicant installed the bird and animal damage control device 10 according to the present embodiment in a lotus field and actually conducted observation. FIG. 10 is a graph showing changes in the number of individuals of the offending species (mall duck) as seen by a fixed point camera photographing the target area. In the graph of FIG. 10, the horizontal axis indicates the date and the number of individuals of the offending species captured by the fixed point camera. In this example, the wildlife damage control device 10 was installed in the target area at date and time t11. As is clear from a comparison between before and after the installation, after the installation of the bird and animal damage control device 10, the number of individuals of the harmful species captured by the fixed-point camera decreased.

<Effects of the bird and beast damage control device 10>
As explained above, the bird and beast damage control device 10 according to the present embodiment uses the classifier 23 generated using the audio data output by the sound collection unit 1 that collects sounds in the target area and the noise data representing environmental sounds. A discriminating unit 21 determines whether the sound collected by the sound collecting unit 1 includes the cry of a bird or beast of an infringing species, and if the sound collected by the sound collecting unit 1 includes the above-mentioned cry, It has an irradiation control unit 22 that irradiates light onto a target area. By using the classifier 23 generated using noise data representing environmental sounds, it is possible to more accurately discriminate the harmful species, thereby making it possible to drive away the harmful species more efficiently.

Furthermore, the bird and animal damage control device 10 according to the present embodiment identifies the harmful species flying into the target area by the sound of their cries (voices), and temporarily irradiates the light beam only when the presence of the harmful species is detected. This makes it possible to efficiently expel harmful species by targeting them. Since it is possible to narrow down the target to harmful species, it is possible to reduce damage caused by harmful species while coexisting with other bird species that use cultivated land as their habitat.

In addition, the bird and animal damage control device 10 according to the present embodiment does not constantly or periodically irradiate a light beam to drive away a harmful species, but irradiates a light beam at the timing when the cry of a harmful species is detected. do. This makes it possible to avoid the injurious species from becoming habituated to the threat, thereby making it possible to maintain the repelling effect for a long period of time.

Furthermore, the light rays emitted by the wildlife damage control device 10 can reach up to about 50 meters away at night and can rotate 360° horizontally. Therefore, by installing it in the center of the target area where you want to prevent damage, you can set the effective repelling range within a radius of about 50 meters. In addition, the bird and animal damage prevention device 10 can collect cries within a range of about 100 meters around the cries using the microphone array 11 used in the sound collection unit 1, and can automatically identify the cries. Even when the incoming or presence of a harmful species is detected outside the effective repelling range by the above-mentioned light irradiation, by temporarily irradiating the light beam, it has the effect of preventing intrusion into the effective repelling range.

In addition, although the bird and beast damage prevention device 10 operates stably using a household power source, it only operates at one time during the night, and the power consumption is as low as about 100 Wh per day, and it can also operate with power supplied from solar cells. It can be installed in cultivated land etc. where there is no power supply. Furthermore, it is easy to install and carry, and when harvesting is finished within the target area and there is no longer a need for expulsion, it can be removed and moved to another target area for use.

Further, the discrimination unit 21 of the bird and beast damage prevention device 10 analyzes the audio data to generate a plurality of parameters representing the feature amount of the audio data, and based on the position in the feature amount space indicated by the generated plurality of parameters, Perform the above determination. As a method for determining the type of perpetrator from audio data, for example, a method using a convolutional neural network (CNN) can be considered. More specifically, for example, a classifier may be generated by training using a spectrogram of a sound as training data. However, the method using CNN has a problem in that the amount of calculation increases and the processing load increases. Another problem is that the device becomes hot due to the large amount of calculations.

In contrast, according to the present embodiment, by performing the determination using parameters such as LPC parameters and autocorrelation coefficients, it is possible to reduce the amount of calculation and reduce the processing load. Further, the configuration of the bird and beast damage prevention device 10 can be simplified, and the device can be made inexpensive.

<Configuration of learning device>
Next, the learning device 4 that constructs the classifiers 23_1 and 23_2 will be explained with reference to the drawings. The learning unit 413 creates classifiers 23_1 and 23_2 for each of the coot and mallard using random forest.

FIG. 11 is a block diagram showing the configuration of the learning device 4. As shown in FIG. The learning device 4 is, for example, a general-purpose computer. The learning device 4 includes a control section 41, a storage section 42, a communication section 43, and an input/output section 44. The learning device 4 may be a different device from the wildlife damage prevention device 10, or the learning device 4 and the wildlife damage prevention device 10 may be configured as an integrated device.

(Communication Department)
The communication unit 43 communicates with a device external to the learning device 4 via a communication line. Although the specific configuration of the communication line does not limit this exemplary embodiment, examples of the communication line include a wireless LAN (local area network), wired LAN, WAN (wide area network), public line network, and mobile data. communication network, or a combination of these. The communication unit 43 transmits data supplied from the control unit 41 to other devices, and supplies data received from other devices to the control unit 41. Note that the communication unit 43 is not an essential configuration in the learning device 4 according to the present specification, and the learning device 4 may have a configuration that does not include the communication unit 43.

(input/output section)
The input/output unit 44 is connected to input/output devices such as a keyboard, a mouse, a display, a printer, and a touch panel. The input/output unit 40A receives input of various information to the learning device 4 from connected input devices. Further, the input/output section 44 outputs various information to the connected output device under the control of the control section 41. Examples of the input/output unit 44 include an interface such as a USB (universal serial bus).

(Control unit/Teacher data generation unit)
The control unit 41 includes a teacher data generation unit 411, a parameter generation unit 412, and a learning unit 413. The teacher data generation unit 411 is an example of a generation unit according to this specification. The teacher data generation unit 411 generates audio data for learning of the classifier 23. Note that for the learning of the classifier 23, a set of a plurality of parameters calculated from the audio data is used instead of the audio data itself.

The teacher data generation unit 411 extracts audio data including two types of calls, a coot and a mallard, from the sample sound data. Here, the sample sound data is audio data with a time length of 0.5 seconds, and as an example, data of a call actually recorded in a lotus field, or data of a call of a coot or mallard published on the web. It is sound data. Both coot and mallard species are known to have multiple patterns of calls, but in this embodiment, the teacher data generation unit 411 uses, as an example, a typical call pattern that is most frequently heard in lotus fields in winter. Extract. Typical calls for mallards include a repetitive ``GwaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaA'', a coot has a short ``aaaaaaaaaaaaaaaaaaaaaaaaa''. Here, in a mallard duck, one ``guh'', which breaks down the repeated phrase, takes about 0.3 seconds, and in a coot, one ``kuh'' is much shorter than a single phrase in a mallard duck. Therefore, in order to be able to detect the entire or part of a characteristic voice phrase, the length of the sound to be extracted is set to 0.5 seconds, and the teacher data generation unit 411 continuously samples samples for 0.5 seconds.

The teacher data generation unit 411 uses the audio data extracted from the sample sound data to generate teacher data in which audio data representing the cries of the offending species of birds and beasts is embedded in audio data representing environmental sounds. The teacher data is an example of learning audio data according to this specification.

FIG. 12 is a diagram for explaining a specific example of the teacher data generation process performed by the teacher data generator 411. The teacher data generation unit 411 generates, as teacher data, audio data that includes both a signal (the cry of an infringing species) and noise (environmental sound), and audio data that does not include a signal and includes only noise. In the following, audio data including signals will also be referred to as "signal data." In addition, audio data that does not include a signal is also called "noise data."

In the example of FIG. 12, the teacher data generation unit 411 performs the following processes (i) to (iv). (i) First, the teacher data generation unit 411 detects a signal syllable (a syllable including the cry of the perpetrator species) in the sample sound data by setting the signal-to-noise ratio threshold to 5. Note that the signal-to-noise ratio threshold is arbitrary and may be any other value.

The teacher data generation unit 411 saves each detected syllable in a wav format file. If a plurality of signal syllables are detected from one sample sound, the teacher data generation unit 411 saves each of the plurality of signal syllables in one wav file. Since we are targeting mallards and coots this time, the maximum length of the signal is 0.3 seconds. When a signal is included multiple times in one sample, the teacher data generation unit 411 stores WAV files corresponding to the number of times. In the example of FIG. 12, the teacher data generation unit 411 generates signal data s11, s12, and s21, which are WAV files of signal syllables, from sample sound data s1 and s2. At this point, the administrator of the bird and animal damage prevention device 10 confirms the sound of the saved wav file, and discards the sound that does not contain the sound of the inflicting species.

(ii) Next, the teacher data generation unit 411 collects and connects the remaining regions (portions that do not include the syllables of the signal) of the sample sound data to generate noise data representing noise (environmental sound). In the example of FIG. 12, the teacher data generation unit 411 generates noise data n11 and n12 from sample sound data s1 and s2.

(iii) Next, the teacher data generation unit 411 measures the time length of the signal data generated in (i) above. Note that, as described above, one signal has a maximum duration of 0.3 seconds.

(iv) The teacher data generation unit 411 randomly selects any one of the plurality of noise data generated in (ii) above, and creates two copy data of the selected noise data. The teacher data generation unit 411 uses one of the two generated copy data as noise teacher data as it is. Further, as the teacher data, noise data is generated by removing the data for the time measured in (iii) from the other copy data at an arbitrary position.

(v) In addition, the teacher data generation unit 411 embeds the signal data of (iii) above in an arbitrary position of the noise data generated in (iv) above, thereby generating teacher data including a signal (the cry of the offending species). generate. The teacher data generated by the teacher data generation unit 411 is used for learning of the classifier 23. In other words, the classifier 23 can be said to be a model that has been machine-learned using teacher data in which sounds representing the cries of birds and animals of the offending species are embedded in audio data representing environmental sounds.

In the example of FIG. 12, in (iv) and (v) above, the teacher data generation unit 411 generates copy data n11_1 and n11_2 of the noise data n11. Further, the teacher data generation unit 411 generates copy data n21_1 and n21_2 of the noise data n21. The copy data n11_2 and n21_2 serve as noise teacher data. On the other hand, signal data s21 and s11 are embedded in arbitrary positions in the copy data n11_1 and n21_2, respectively, and serve as training data for the signal. As described above, the teacher data generation unit 411 generates teacher data including signals and teacher data not including signals.

(Parameter generation part)
The parameter generation unit 412 performs voice analysis on the teacher data generated by the teacher data generation unit 411, and generates a plurality of parameters representing voice feature amounts. In this embodiment, we will use a classical method used to identify human voices based on the physical characteristics of the vocal system as a strategy for performing voice analysis as numerical data with time series rather than images. apply. One of them is LPC. Since LPC can express one sound as a polynomial, multiple parameters can be obtained. By using random forest in a hyperspace with the same number of dimensions as the obtained parameters, it is possible to distinguish between signals and noise.

Furthermore, the parameter generation unit 412 extracts syllables from noise data in the teacher data using the same procedure. If the parameter generation unit 412 succeeds in extracting a syllable and calculating the 17 coefficients, it labels it as noise.

(Study Department)
The learning unit 413 inputs a plurality of parameters representing feature quantities of the audio data using teacher data including a plurality of parameters generated by the parameter generation unit 412 and a label indicating whether or not the cry of the above-mentioned perpetrator species is included. As a result, classifiers 23_1 and 23_2 are generated that determine whether the audio data includes the cry of a bird or beast that is an infringing species.

<Effects of learning devices>
As explained above, the learning device 4 according to the present embodiment includes a teacher data generation unit 411 that generates learning audio data in which audio data representing the cries of the inflicting bird and beast species is embedded in audio data representing environmental sounds; a parameter generation unit 412 that performs audio analysis on the learning audio data generated by the teacher data generation unit 411 to generate a plurality of parameters representing feature quantities; and a label indicating the plurality of parameters generated by the parameter generation unit 412 and the type of perpetrator. and a learning unit 413 that causes the classifier 23 to learn using learning data including the following. By using the classifier 23 generated by the learning device 4, the calls of the injurious species can be discriminated with higher accuracy.

[Modified example]
(Bird and animal damage)
In the above-described embodiment, a case has been described in which the bird and animal damage control device 10 is used to prevent lotus root damage in a lotus field where lotus root is cultivated. The object of the bird and beast damage control device 10 according to the present specification is not limited to the above-mentioned example. The targets of the bird and beast damage control device 10 include feeding damage to agricultural and marine products other than lotus root, noise caused by the cries of birds and beasts, sanitary damage caused by bird and beast feces, and the like. In other words, the bird and animal damage control device 10 according to the present specification is not only effective in reducing feeding damage to agricultural and marine products, but also as a measure to drive away birds such as pigeons and starlings that cause damage to people's lives with their droppings and cries in residential areas. It is applicable.

(Target area)
Further, in the above-described embodiment, a case has been described in which the target area is a flooded lotus field before harvesting, but the target area is not limited to the above-mentioned example. For example, the target area is a field where other agricultural and marine products are cultivated, It may be around facilities such as aquaculture farms or stations.

(perpetrating species)
Further, in the above-described embodiment, a case has been described in which the harmful species are mallard ducks and coots, but the harmful species are not limited to the examples shown in the above-mentioned embodiments. If a survey is conducted in the target area and feeding damage caused by a new species is found, it may be added to the list of damaging species that are subject to countermeasures against wildlife damage. Moreover, the offending species is not limited to birds, but may also be animals such as wild boars.

(wavelength and direction of irradiation light)
In the above-mentioned embodiment, the case where the voice identification/light irradiation unit 2 irradiates green laser light is illustrated, but the light that the voice identification/light irradiation unit 2 irradiates is not limited to the example described above. The voice identification/light irradiation section 2 may irradiate light of other wavelengths. Further, in the above-described embodiment, the voice identification/light irradiation unit 2 changes the irradiation direction of the light beam at a predetermined speed or a predetermined angular velocity, but the method of changing the irradiation direction is not limited to the above-mentioned example. For example, the voice identification/light irradiation unit 2 may have a configuration in which the direction of light irradiation is not changed. In this case, the audio identification/light irradiation unit 2 may, for example, estimate the direction of the sound source of the signal and irradiate light in the direction of the sound source. Further, in the above-described embodiment, the case where the voice identification/light irradiation section 2 irradiates laser light is illustrated, but the light irradiated by the voice identification/light irradiation section 2 is not limited to laser light, and may be any light other than laser light. It may be.

(irradiation pattern)
Furthermore, in the above-described embodiment, the voice identification/light irradiation unit 2 may perform another process for driving away the harmful species when the harmful species does not escape despite irradiation with laser light. For example, the voice identification/light irradiation unit 2 may perform control to change the laser light irradiation pattern if the cry of the injurious species is detected even after the laser light is irradiated. Here, the irradiation pattern includes, for example, at least one of the irradiation direction of the laser beam, the irradiation position, the color and intensity of the laser beam, the blinking pattern, the moving speed in the irradiation direction, and the angular velocity. By irradiating laser light with various irradiation patterns, it is possible to prevent harmful species from becoming accustomed to laser light.

(Power supply part)
In the above-described embodiment, the power supply section 3 includes a solar cell, but is not limited to this. The power supply section 3 may include an independent power source other than a solar cell. Here, an independent power source is a power source that can supply power independently, and includes, for example, a solar cell, a storage battery, a power generator such as a power generator or a wind power generator, a space transmission wireless power supply, or a thermoelectric conversion element. Including the power supply used, etc. Further, the wildlife damage control device 10 is not limited to a device equipped with an independent power source, and may be configured to operate using a household power source, for example.

(classifier)
In the above-described embodiment, a trained model constructed by a random forest method was used as the classifier 23. The classifier 23 is limited to the example shown in the embodiment described above. The classifier 23 may be, for example, a trained model trained by a support vector machine (SVM) technique. In the case of a random forest or a support vector machine, it can also be said that the determination unit 21 determines whether the voice data includes the cry of the injurious species, based on the position in the feature space indicated by a plurality of parameters.

(Data recording)
In the above-described embodiment, the wildlife damage control device 10 recorded the sound collected in the target area and the date and time of light irradiation. In addition to this, a configuration may be adopted in which an infrared camera images the target area, and the image obtained by the imaging is recorded in association with the audio and the date and time of irradiation.

(Relearning)
Further, the data recorded by the wildlife damage prevention device 10 may be used for relearning the classifier 23. For example, when the determining unit 21 determines that the audio data includes the cry of a perpetrator species, the determining unit 21 attaches a signal label to the audio data, and uses the audio data as learning data for relearning the classifier 23. On the other hand, when determining that the audio data does not include the cry of the perpetrator species, the determining unit 21 assigns a noise label to the audio data and uses it as learning data for relearning the classifier 23. In addition, even if the cry of the perpetrator species is detected, if any of the return values of the function that calculates the 17 coefficients is null, the discrimination unit 21 labels it as noise. Learning data for relearning the classifier 23 may be generated.

The learning unit 413 re-trains the classifier 23 using the generated learning data. By relearning the classifier 23, it is possible to improve the classification accuracy of the classifier 23 while using the bird and animal damage control device 10.

(About the sound collection section)
In the above-described embodiment, a configuration in which the sound collection unit 1 includes the microphone array 11 has been described, but the sound collection unit 1 may include a small waterproof monaural microphone instead of the microphone array 11. FIG. 13 is a diagram showing an example of the configuration of a wildlife damage control device 10A according to this modification. In the example of FIG. 13, the bird and beast damage control device 10A includes a sound collection unit 1A instead of the sound collection unit 1 included in the bird and beast damage control device 10 according to the above-described embodiment. The sound collection section 1A includes a small and waterproof monaural microphone, and is installed in the voice identification/light irradiation section 2. In the example of FIG. 13, the sound collection section 1A and the voice identification/light irradiation section 2 may be integrated together and attached to a pillar, etc.; No need to fix. Therefore, the cost of materials and labor during construction can be reduced.

By the way, a microphone array can localize a sound source and has a noise removal function as a secondary feature, so the sound quality is good, whereas a small monaural microphone does not have a noise removal function and has poor sound quality. Therefore, when using the sound collection section 1A including a small monaural microphone, the learning section 413 may perform re-learning on the classifier 23, taking into account the poor sound quality of the collected sound. As an example, the learning unit 413 may retrain the classifier 23 using audio data recorded with a small monaural microphone. In this case, the training data used for relearning includes, for example, a plurality of parameters obtained from audio data collected by a small monaural microphone and a label indicating whether or not the cry of the perpetrator species is included. . In this case, the classifier 23 (machine learning model) converts a pre-trained learned model into a trained model that is retrained using teacher data including parameters generated from audio data collected by a monaural microphone. It can also be said that it is a model.

By retraining the classifier 23 using voice data recorded with a microphone with poor sound quality, it is possible to maintain the same level of cry recognition ability as when using a microphone with good sound quality. As a result, according to the configuration according to this modification, a monaural microphone which is inexpensive and has poor sound quality can be used in the bird and beast damage prevention device, and the versatility of the bird and beast damage prevention device can be expanded.

(About voice recognition/light irradiation unit power control)
In the embodiment described above, the wildlife damage control device 10 may further include a power supply control unit that controls the power supply of the determination unit 21 and the irradiation control unit 22. FIG. 14 is a diagram showing an example of the configuration of the voice identification/light irradiation section 2A according to this modification. In the example of FIG. 14, the voice identification/light irradiation unit 2A includes a power source in addition to the discrimination unit 21, the irradiation control unit 22, the classifier 23_1, and the classifier 23_2 included in the voice recognition/light irradiation unit 2 of the embodiment described above. A control section 24 is provided. The power supply control unit 24 turns off the power to the voice identification/light irradiation unit 2 during the day, and activates it only at night when countermeasures are required. According to this configuration, power consumption can be suppressed compared to the case where the power supply control of the voice identification/light irradiation unit 2 is not performed, and stable long-term operation can be achieved with a 30W solar panel, for example.

(Addition of image acquisition section)
In the above-described embodiment, the voice identification/light irradiation unit 2 may be linked with a photographing device (such as a network camera) in order to collect images used for verifying the expulsion effect and the like. That is, the wildlife damage control device may further include a photographing device that photographs the direction of light irradiation and outputs image data representing the photographed image. In the present disclosure, captured images include moving images and still images. Furthermore, in the present disclosure, a still image may be a single still image captured in response to one light irradiation, or may be a still image captured continuously within a predetermined time period including one light irradiation. It may also be a plurality of still images. The plurality of still images is a still image sequence in which each still image is connected in chronological order. Whether a moving image or a still image is selected as the captured image can be appropriately selected depending on the band or communication speed of the network in which the wildlife damage prevention device 10A is installed, the use of the captured image, and the like. The photographed image taken by the photographing device may be used for verifying the repelling effect, etc., or may be used for detecting the arrival of harmful species. In the latter case, the determining unit 21 determines whether a harmful species has flown into the target area, using image data of a photographed image photographed by the photographing device 5 in addition to the audio data output by the sound collecting unit 1. In the former case, it is necessary to understand the reaction of the harmful species to the irradiated light, and it is not necessary to process the captured images in real time, so it is preferable to use a moving image as the captured image. In the latter case, since it is not necessary to understand the reaction of the harmful species to light, a still image may be used as the captured image. However, if the computer 227 that realizes the determination unit 21 by executing a program has sufficient processing power, a moving image may be used as the captured image.

FIG. 15 is a diagram illustrating a configuration example of a wildlife damage control device according to this modification. The bird and beast damage prevention device 10B illustrated in FIG. Be prepared. The photographing device 5 is, for example, a network camera. In this example, the voice identification/light irradiation unit 2 and the photographing device 5 are separate bodies, and are each attached to a pillar 6 separately. The photographing device 5 oscillates in conjunction with the oscillation and light irradiation of the voice identification/light irradiation unit 2 and records before and after the irradiation by the voice identification/light irradiation unit 2 . The recorded images allow you to efficiently check the chase scene. Further, effectiveness verification is not essential for farmers who are users of the bird and beast damage control device 10B. Therefore, if it is more important to suppress the manufacturing cost of the bird and beast damage control device 10B than to verify the effectiveness of the bird and beast damage control device 10B, it is preferable to configure the bird and beast damage control device 10 in which the photographing device 5 is omitted. . In addition, in one aspect of the present invention, the photographing device 5 can be added to the bird and beast damage control device 10A shown in FIG. 13 instead of the bird and beast damage control device 10B.

Moreover, FIG. 16 is a diagram showing another configuration example of the wildlife damage control device according to this modification. In the wildlife damage prevention device 10C illustrated in FIG. 16, a laser housing 221C is attached to a photographing device 5C. The laser housing 221C is similar to the laser housing 221 described in the above embodiment. In this case, the laser housing 221C also oscillates using the oscillating function of the imaging device 5C to emit light. As a result, parts necessary for swinging the laser housing 221C can be omitted, and material costs can be reduced.

When using a photographed image to detect an incoming harmful species, the audio identification/light irradiation unit 2 uses, for example, the detection results of the harmful species by sound using the classifier 23 and the image data of the captured image captured by the photographing device 5. It may be determined whether a harmful species has arrived using both the detection results obtained by analyzing the . More specifically, for example, when a harmful species is detected by sound using the classifier 23 and an incoming harmful species is detected by image analysis, the audio identification/light irradiation unit 2 detects a harmful species in the target area. Laser light may be irradiated.

In addition, the voice identification/light irradiation unit 2 inputs, for example, parameters obtained from the voice data collected by the sound collection unit 1 and image data photographed by the photographing device 5 to the classifier 23 (machine learning model). Based on the output obtained, it may be determined whether a harmful species has arrived. In this case, the teacher data used for the learning of the classifier 23 includes, for example, parameters obtained from audio data, image data taken by a photographing device, and a label indicating whether a harmful species has arrived. In addition, the input of the classifier 23 includes, for example, parameters obtained from the audio data collected by the sound collection unit 1 and the image data captured by the imaging device 5, and the output of the classifier 23 includes, for example, the perpetrator species. Contains information indicating whether or not the object has flown into the target area.

When detecting the arrival (presence) of harmful species, detection accuracy can be increased by performing detection using images in addition to audio detection.

(About the discrimination part)
An AI (Artificial Intelligence) method may be used for the algorithm of the determining unit 21 that determines whether the collected sounds include the cries of the perpetrator species. As an example of a discriminant model using AI, an example will be described in which a trained model generated using a convolutional neural network (CNN) is used as the classifier 23.

FIG. 17 is a diagram showing a specific example of teacher data used for training the discriminant model according to this modification. In the example of FIG. 17, the classifier 23C, which is a discrimination model, is trained to input a 0.96-second audio fragment converted into a mel spectrogram and output an inference result as to whether the audio is the cry of a perpetrator species. It's a model. For example, YAMNet is used as the CNN architecture, but is not limited thereto. In this example, the training data used for the learning of the classifier 23C is a parameter representing the characteristics of the audio data obtained by analyzing the audio data, and a label (e.g. , "coot"). In the learning phase, the learning unit 413 inputs parameters included in the teacher data to the classifier 23C (S101), and the classifier 23C infers a label (S102). The learning unit 413 trains the classifier 23C by updating the model parameters of the classifier 23C so as to reduce the error of inference by the classifier 23C.

When using a discrimination model such as the one illustrated in Figure 17 to discriminate the calls of invading species, challenges include (i) difficulty in training with a small amount of training data, and (ii) large amount of calculations required to run the model. It will be done. In this modification, the above problem is solved by using (i) pre-training with a large-scale environmental sound dataset and (ii) reducing the weight of the model by quantization and compression.

FIG. 18 is a diagram showing a specific example of generation of a discriminant model and implementation in the computer 227 according to this modification. First, in order to obtain good performance with a small amount of training data, the learning unit 413 pre-trains a discriminant model on an existing large-scale environmental sound dataset (S201). For example, the AudioSet-YouTube (registered trademark) corpus is used for pre-training. After that, the learning unit 413 retrains (fine-tunes) the learned model using the data set for identifying the calls of the perpetrator species (S202). The teacher data used for retraining includes, for example, a plurality of parameters obtained by analyzing audio data that includes the calls of the perpetrator species, and a label indicating whether or not the calls of the perpetrator species are included. In other words, the machine learning model used by the discrimination unit 21 is pre-trained with a set of audio data representing environmental sounds, and then retrained using training data that includes parameters generated from audio data that includes the calls of the perpetrator species. This is a trained model.

Next, in order to implement the trained model on the computer 227 with low computing power, the learning unit 413 performs quantization and compression (S203). In quantization, the learning unit 413 reduces the weight of the model by reducing calculation accuracy. In compression, the learning unit 413 converts the model into a more lightweight format. An example of such a format is the TensorFlowLite (tflite) format. In other words, the machine learning model used by the determining unit 21 is a model whose weight has been reduced by at least one of quantization and compression. The lightweight model file obtained by compression is mounted on a computer or the like (S204).

According to the configuration according to this modified example, misclassifications can be greatly reduced even in a real environment using a microphone with large noise and distortion, and execution can be performed with a small amount of calculation compared to a discrimination model using a random forest. Become. Additionally, the compressed discriminant model can be implemented in a microcontroller, leading to miniaturization and improved portability of the device.

In this way, by using the AI method for the algorithm of the discrimination unit 21, the accuracy of the discrimination by the discrimination unit 21 can be increased. Further, according to the above configuration, it is possible to implement a language that can be operated on a microcomputer (small integrated circuit), which leads to miniaturization and improved portability of the present device. The inventor generated a classifier 23 by quantizing and compressing the CNN after retraining using the method described above, and measured the classification rate using the generated classifier 23. As a result, the misclassification rate was 20. %, significantly lowering the rate of misclassification than when using conventional methods.

[Example of implementation using software]
The functions of the bird and beast damage control device 10 and the learning device 4 (hereinafter referred to as "devices") are programs for making a computer function as the devices, and programs for making the computer function as each control block of the device. This can be realized by a program.

In this case, the device includes a computer having at least one control device (for example, a processor) and at least one storage device (for example, a memory) as hardware for executing the program. By executing the above program using this control device and storage device, each function described in each of the above embodiments is realized.

The above program may be recorded on one or more computer-readable recording media instead of being temporary. This recording medium may or may not be included in the above device. In the latter case, the program may be supplied to the device via any transmission medium, wired or wireless.

Further, part or all of the functions of each of the control blocks described above can also be realized by a logic circuit. For example, an integrated circuit in which a logic circuit functioning as each of the control blocks described above is formed is also included in the scope of the present invention.

The present invention is not limited to the embodiments described above, and various modifications can be made within the scope of the claims, and embodiments obtained by appropriately combining technical means disclosed in different embodiments. are also included within the technical scope of the present invention.

10, 10A, 10B, 10C Bird and beast damage control device 1, 1A Sound collection section 11 Microphone array 12 Rain cover 13 Cover 14 USB cable 15 Sound collection section support rod 16 Sound collection section fixture 2, 2A Voice identification/light irradiation section 21 Discrimination section 22 Irradiation control section 23, 23_1, 23_2, 23C Classifier 24 Power supply control section 211 Case 212 Dome 213 Support rod 214 Fixture 221, 221C Laser housing 222 Connection line (stepping motor 225)
223 Connection line (laser module 226)
224 GPS receiver 225 Stepping motor 226 Laser module 227 Computer 228 Expansion board 229 Motor driver 3 Power supply unit 31 Solar panel 32 Battery 33 Case 34 Connection line 35 Solar charge controller 36 DC/DC converter 5, 5C Photography device

Claims (22)

Using the audio data output by the sound collection unit that collects sound in the target area and a classifier generated using noise data representing environmental sounds, the sound collected by the sound collection unit is used to identify the harmful species of birds and animals. a determination unit that determines whether a cry is included;
an irradiation control unit that controls a light irradiation unit to irradiate light to the target area when the sound collected by the sound collection unit includes the cry;
A bird and beast damage control device with The irradiation control unit changes the irradiation direction of the light at a predetermined speed or a predetermined angular velocity.
The bird and animal damage control device according to claim 1. The irradiation control unit estimates the direction of the sound source of the cry, and moves the irradiation direction of the light in a direction in which the irradiation position of the light approaches the sound source.
The bird and beast damage control device according to claim 2. further comprising a power supply unit including a self-contained power supply;
The discrimination unit and the irradiation control unit are operated by power supplied by the power supply unit.
The bird and animal damage control device according to claim 1 or 2. further comprising a recording unit that records the sound collected by the sound collection unit and the date and time when the light irradiation unit irradiated the light;
The bird and animal damage control device according to claim 1 or 2. The target area is a lotus field where lotus root is cultivated.
The bird and animal damage control device according to claim 1 or 2. The harmful species of birds and animals includes at least one of the species that forage at night in the target area.
The bird and animal damage control device according to claim 1 or 2. The target area is a lotus field where lotus root is cultivated,
The harmful species of birds and animals include at least one of a mallard and a coot;
The bird and beast damage control device according to claim 7. The discrimination unit analyzes the audio data and generates a plurality of parameters representing feature amounts of the audio data,
performing the discrimination based on the position in the feature space indicated by the plurality of generated parameters;
The bird and animal damage control device according to claim 1 or 2. The plurality of parameters include linear predictive coefficients calculated by linear predictive coding (LPC),
The bird and beast damage control device according to claim 9. the plurality of parameters include an autocorrelation coefficient of the audio data;
The bird and animal damage control device according to claim 10. The discrimination unit performs the discrimination using a machine learning model that receives the plurality of parameters as input and outputs a discrimination result.
The bird and beast damage control device according to claim 9. The machine learning model is a model constructed by random forest learning,
The bird and animal damage control device according to claim 12. The machine learning model is a trained model that has been machine learned using teacher data in which audio data representing the environmental sounds is embedded with audio representing the cries of bird and beast species.
The bird and animal damage control device according to claim 12. The sound collection section;
the light irradiation section;
The bird and animal damage control device according to claim 1 or 2, further comprising: The light irradiation unit irradiates laser light.
The bird and animal damage control device according to claim 1 or 2. a generation unit that generates learning audio data in which audio data representing the cries of a harmful species of bird or beast is embedded in audio data representing environmental sounds;
a parameter generation unit that generates a plurality of parameters representing feature quantities by voice-analyzing the learning audio data generated by the generation unit;
Using learning data including a plurality of parameters generated by the parameter generation unit and a label indicating whether or not the cry of the perpetrator species is included, a plurality of parameters representing feature quantities of the audio data are input and the audio data is generated. a learning unit that generates a machine learning model that determines whether the song contains the sounds of a harmful species of bird or beast;
learning devices including; computer,
Using the audio data output by the sound collection unit that collects sound in the target area and a classifier generated using noise data representing environmental sounds, the sound collected by the sound collection unit is used to identify the harmful species of birds and animals. a determination unit that determines whether a cry is included;
an irradiation control unit that controls a light irradiation unit to irradiate light to the target area when the sound collected by the sound collection unit includes the cry;
A program to function as further comprising a photographing device that photographs the irradiation direction of the light and outputs image data representing the photographed image;
The bird and animal damage control device according to claim 1 or 2. The determining unit determines whether the harmful species has flown into the target area using image data of a photographed image photographed by the photographing device in addition to audio data output by the sound collecting unit.
The bird and animal damage control device according to claim 19. The machine learning model is a trained model that is pre-trained on a set of audio data representing environmental sounds and then retrained using training data that includes parameters generated from audio data that includes the calls of the perpetrator species. be,
The bird and animal damage control device according to claim 12. The machine learning model is a model that has been reduced in weight by at least one of quantization and compression.
The bird and animal damage control device according to claim 12.

Images