JP2018173342A - Rainfall sensor, rainfall estimation device, rainfall estimation method, and rainfall estimation system - Google Patents

Abstract

A novel and useful rainfall sensor, rainfall estimation device, rainfall estimation method, and rainfall estimation system are provided. Acceleration data is obtained by a rain sensor having a vibrating body that vibrates due to the impact of raindrops, an acceleration sensor provided on the vibrating body, and a transmission interface for transmitting acceleration data from the acceleration sensor. The rain amount estimation device transmits the acceleration data to the frequency domain data for each time period, performs a Fourier transform, extracts the feature data composed of correlation values between the frequency domain data, and the rain sensor. Means for estimating the rainfall from the feature data using the regression equation obtained by learning, and estimates the rainfall from the acceleration data. [Selection] Figure 2

Description

  The present invention relates to a technique for estimating rainfall, and more particularly to a rainfall sensor that detects the impact of raindrops, a rainfall estimation device that uses the rainfall sensor, a rainfall estimation method, and a rainfall estimation system.

  In recent years, due to the effects of global warming, summer temperatures have risen and local heavy rain has increased. In order to minimize the sudden increase in rivers and landslides caused by heavy rain, a dense rainfall observation network is required.

  A tipping-type rain gauge is widely used as a rain gauge. This rain gauge has a funnel-shaped rainwater supply section having an opening of a predetermined size, and the rainwater is guided to one of a pair of tipping overfalls, and when the amount of rainwater accumulates, the overturning taps fall down and drain. Precipitation is measured by detecting the overturning action with a reed switch and measuring the number of falls (see, for example, Patent Document 1).

JP 2003-84077 A

  However, in the overturning rain gauge of Patent Document 1, foreign matters such as fallen leaves and insects are likely to be mixed from the opening, and the rainwater supply unit is likely to be blocked. Moreover, in order to remove a foreign material, a user must remove and maintenance is required. When installing a rain gauge in a mountainous area, a valley line, or a slope, it is often difficult for a user to access the rain gauge. Moreover, since the overturning rod has a movable member, there exists a possibility that durability may be inferior. In addition, it is difficult to reduce the manufacturing cost because there are many mechanical parts and complicated.

  The present invention has been made to overcome the above drawbacks, and provides a new and useful rainfall sensor, rainfall estimation device, rainfall estimation method, and rainfall estimation system.

  According to one aspect of the present invention, means for receiving acceleration data of vibration generated in a vibrating body due to raindrop impact from a rain sensor, means for Fourier transforming the acceleration data into frequency domain data for each time period, Means for extracting feature data comprising correlation values between data in the frequency domain, and means for estimating rainfall from the feature data using a regression equation obtained by machine learning for the rainfall sensor. An estimation device is provided.

  According to the above aspect, the acceleration value of vibration generated in the vibrating body due to the impact of raindrops is Fourier-transformed to obtain the correlation value of the frequency domain data, which is used as feature data. This makes it possible to extract the characteristics of acceleration data with high accuracy. Thereby, since the identification accuracy of acceleration data is improved, the accuracy of rainfall estimation is improved. Moreover, since the rain amount estimation uses acceleration data by raindrop impact, it can be performed in a short time. Thereby, even when a large amount of rain falls in a very short time, it is possible to estimate the rainfall quickly.

  According to another aspect of the present invention, the amount of rain comprising a vibrating body that vibrates due to the impact of raindrops, an acceleration sensor provided on the vibrating body, and a transmission unit that transmits acceleration data from the acceleration sensor. A sensor is provided.

  According to the above aspect, since the rain sensor detects the acceleration of the vibration of the vibrating body that vibrates due to the impact of raindrops, it does not require a structure for directly measuring the rainfall, so it is compact, low cost, and maintenance Good properties.

  According to another aspect of the present invention, a step of receiving acceleration data of vibration generated in the vibrating body by the impact of raindrops from a rain sensor, and a step of Fourier transforming the acceleration data into frequency domain data for each time period; Extracting feature data comprising correlation values between data in the frequency domain, and estimating rainfall from the feature data using a regression equation obtained by machine learning for the rainfall sensor. A rainfall estimation method is provided.

  According to the above aspect, the acceleration value of vibration generated in the vibrating body due to the impact of raindrops is Fourier-transformed to obtain the correlation value of the frequency domain data, which is used as feature data. This makes it possible to extract the characteristics of acceleration data with high accuracy. Thereby, since the identification accuracy of acceleration data is improved, the accuracy of rainfall estimation is improved.

  According to another aspect of the present invention, there is provided a rainfall estimation system comprising a rainfall sensor and a rainfall estimation device, wherein the rain sensor is provided on the vibrating body that vibrates due to the impact of raindrops and the vibrating body. And a transmission unit for transmitting acceleration data from the acceleration sensor, wherein the rainfall estimation device receives the acceleration data from the rainfall sensor, and the acceleration data for each time period. Using the regression equation obtained by machine learning for the rain sensor, the means for performing Fourier transform on the frequency domain data, the means for extracting the feature data consisting of correlation values between the frequency domain data, Means for estimating rainfall from data is provided.

  According to the above aspect, since the rain sensor is easy to install at low cost and has good maintainability, it is possible to install many rain sensors at high density in an area where landslides are likely to occur due to rainfall. Furthermore, the rainfall estimation by the rainfall estimation device can be accurately performed in a short time. Even if a large amount of rain falls in a very short time, the rainfall can be estimated quickly.

  In the present description, claims, abstract and drawings, the term “rainfall” refers to the amount of rain or precipitation that falls or falls per unit time, for example, rainfall per second, rainfall per minute, It is rainfall per hour.

It is a block diagram which shows the structure of the rainfall estimation system which concerns on one Embodiment of this invention. It is a block diagram which shows the structure of the rainfall sensor which concerns on one Embodiment of this invention. It is a schematic sectional drawing which shows the example of the rainfall sensor which concerns on one Embodiment of this invention. It is a schematic sectional drawing which shows the other example of the rainfall sensor which concerns on one Embodiment of this invention. It is a schematic sectional drawing which shows the other example of the rainfall sensor which concerns on one Embodiment of this invention. It is a block diagram which shows the structure of the rainfall estimation apparatus which concerns on one Embodiment of this invention. It is a block diagram which shows the function structure of the rainfall estimation apparatus which concerns on one Embodiment of this invention. It is a figure which shows the arrangement | sequence of the data of the frequency domain in a feature data extraction part. It is a figure which shows the example of the mask pattern for calculating | requiring the correlation value in a feature data extraction part. It is a flowchart of the rainfall estimation method which concerns on one Embodiment of this invention. It is a flowchart which shows the method of calculating | requiring the regression equation used with the rainfall estimation method which concerns on one Embodiment of this invention. It is a flowchart which shows the other method of calculating | requiring the regression type used with the rainfall estimation method which concerns on one Embodiment of this invention. It is a figure which shows the Example of the rainfall estimation system which concerns on one Embodiment of this invention.

  Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In addition, the same code | symbol is attached | subjected about the element which is common between drawings, and the repetition of detailed description of the element is abbreviate | omitted.

[Rainfall estimation system]
FIG. 1 is a block diagram illustrating a configuration of a rainfall estimation system according to an embodiment of the present invention.

  Referring to FIG. 1, a rainfall estimation system 100 according to an embodiment of the present invention includes a plurality of rainfall sensors 10 to 12 and a rainfall estimation device 50. The rain sensors 10 to 12 transmit acceleration data of vibration generated in the vibrating body due to the impact of raindrops to the rain estimation device 50 during rain. The rainfall sensors 10 to 12 are installed not only in residential areas and commercial areas, but also in areas where disasters are likely to occur due to rainfall, such as mountainous areas, valleys, and volcanic ash deposits. The rainfall estimation device 50 receives acceleration data from the rainfall sensors 10 to 12 and estimates the rainfall. The rainfall estimation device 50 can transmit the estimated rainfall information to a disaster prevention system such as a fire department or a home using the Internet or a local communication network. As will be described later, since the rain sensors 10 to 12 are inexpensive and easy to install and have good maintainability, it is possible to install a large number of rain sensors in an area where landslides are likely to occur due to rainfall. Furthermore, the rainfall estimation by the rainfall estimation device 50 can be performed in a short time. Therefore, even when a large amount of rain falls in a very short time, the rainfall can be estimated quickly, and the rainfall estimation system 100 can promptly notify the estimated rainfall.

[Rainfall sensor]
FIG. 2 is a block diagram showing a configuration of a rainfall sensor according to an embodiment of the present invention.

  Referring to FIG. 2, a rain sensor 10 according to an embodiment of the present invention includes a vibrating body 20 that vibrates in response to a raindrop impact, and vibrations of the vibrating body 20 provided in the vibrating body 20. , The CPU 31 to which acceleration data from the acceleration sensor 30 is supplied, the transmission interface 33 to which the acceleration data and the rain sensor identification information are supplied, and the transmission interface 33. And a memory 32 connected to the CPU 31 and a battery 35 for supplying power to these electronic components.

  The to-be-vibrated body 20 receives raindrops when it rains, and vibrates due to an impact at the time of the collision. When the raindrops come straight from the sky, the vibrating body 20 can receive the raindrops if it has a flat plate-like surface perpendicular thereto. However, when the wind is strong, rain may fall obliquely or laterally from the sky, so in order to receive the impact of raindrops even during such rain, it is preferably a convex body, It is preferably a hemispherical shape (including a shape obtained by bisecting a true sphere, a shape having a part of a spherical surface, and a dome shape) and a pseudospherical shape. The convex body includes a shape obtained by cutting out a part of a polyhedron. The vibrating body 20 is preferably made of metal or plastic in that it can appropriately convert the impact of raindrops into vibration. The vibrating body 20 is preferably thin enough to vibrate by the impact of raindrops.

  The acceleration sensor 30 is provided on the vibrating body 20 and detects the acceleration of vibration of the vibrating body 20. The acceleration sensor 30 is, for example, a three-dimensional sensor, and can detect acceleration in three axis directions. The acceleration sensor 30 is, for example, a MEMS sensor, and the size is, for example, 5 millimeters long × 5 millimeters wide × 1.5 millimeters thick. The acceleration sensor 30 can output the detected acceleration as analog data or digital data. Here, it is assumed that digitized acceleration data is output. When analog data is output, an A / D converter may be connected to convert it into digital data and output. The acceleration data is output with a frequency of several tens of kHz, for example. The sensitivity of the acceleration sensor is, for example, from several mg to several tens of mg per bit (g represents gravitational acceleration), and can be output in 16 bits.

A CPU (central processing unit) 31 can appropriately select a known MPU (microprocessor), and is connected to the acceleration sensor 30, the memory 32, and the transmission interface 33, respectively. The CPU 31 is connected to the acceleration sensor 30 by, for example, a serial interface, I ² C, or SPI. The CPU 31 receives acceleration data from the acceleration sensor 30 and stores it in a memory or sends it to a transmission interface. At that time, not only acceleration data but also auxiliary information including identification information of the rainfall sensor may be added.

  The memory 32 is a random access memory (RAM) or a read only memory (ROM), and may be an independent chip or a memory included in the CPU 31. The memory 32 may be used for temporarily storing acceleration data or may be used for other purposes.

  The transmission interface 33 transmits acceleration data, auxiliary information, and the like to the rainfall estimation device 50 via the antenna. The transmission interface 33 is not particularly limited. For example, a wireless chip or a wireless module having a transmission output of 920 MHz and a transmission output of 0.01 W or less or 0.02 W or less can be used. Note that wired transmission may be performed according to the installation environment of the rainfall sensor 10 or the like. The transmission interface 33 may be used for wired communication. The antenna 34 is not particularly limited, but when the radio frequency is 920 MHz, a rod antenna or a wire antenna can be used.

  The battery 35 supplies power to each electronic component of the rainfall sensor 10. By using the battery 35, the rain sensor 10 can be installed along a steep valley or mountain where it is difficult or expensive to install a power line. The battery 35 may be connected to a photovoltaic power generation panel (not shown) to supply power. Of course, as long as it can be connected to the power line, the battery 35 may be supplied with power, and in that case, the battery 35 may be omitted.

  Since the rain sensor 10 does not directly measure the volume of rainwater, it does not require a container such as a bag for storing rainwater and a mechanism for measuring the volume, which has been required in the past. This eliminates the need for complicated mechanical parts, is compact, low cost, and has good maintainability.

  3-5 is a schematic sectional drawing which shows the example of the rainfall sensor which concerns on one Embodiment of this invention.

  Referring to FIG. 3, the rain sensor 13 generally includes a body to be vibrated 21, an acceleration sensor 30 provided on the body to be vibrated 21, an electronic board 36, a battery 35, and a body that supports the body 21 to be vibrated. And a support column 39 for fixing the rainfall sensor 13 to the ground. In FIG. 3, the antenna and the memory shown in FIG. 2 are not shown.

  The vibrating body 21 has a hemispherical shape or a dome shape. The acceleration sensor 30 is provided on the inner surface of the vibrating body 21 and is fixed via a vibration transmitting body 37. The vibration transmitting body 37 is, for example, a metal plate, a plastic plate, an adhesive layer, or the like, and these may be combined.

  The acceleration sensor 30 is disposed at the zenith of the vibrating body 21 in FIG. By arranging the acceleration sensor 30 at the zenith, the vibration of the vibrating body 21 can be detected efficiently. The acceleration sensor may be disposed at the zenith or may be disposed at another part of the vibrating body 21.

  The acceleration sensor 30 is provided on the electronic board 36 so as to be capable of signal communication with the CPU 31, the transmission interface 33, and the memory 32. The electronic substrate 36 is not particularly limited as long as these electronic components can be connected, but is preferably lightweight. Thereby, the influence on the vibration of the to-be-vibrated body 21 can be reduced.

  A battery 35 is connected to the electronic board 36 via a cable 40 to supply power. The voltage (for example, 3.5 volts) and current to be supplied to the battery 35 are determined in accordance with the specifications of the CPU and acceleration sensor. The smaller the power consumption of the entire circuit, the better.

  The main body 38 is a disk-shaped flat plate and is not particularly limited as long as the mechanical strength and waterproofness are sufficient. The support column 39 is not particularly limited as long as it is a shape that can be easily embedded.

  Since the vibration sensor 20 has a hemispherical shape, the rainfall sensor 13 can accurately receive not only rainfall from above but also rainfall from oblique and lateral directions, so that the accuracy of rainfall estimation is improved. To do. Further, since the vibrating body 20 has a hemispherical shape, vibrations not only in the vertical (vertical) direction but also in a plane perpendicular to the vertical direction, that is, in the horizontal direction are likely to occur. The acceleration sensor 30 can estimate the rain amount more accurately by measuring the acceleration of the vibration in the vertical direction and the two horizontal directions. Furthermore, since the acceleration sensor 30 is provided on the inner surface of the vibrating body, water wetting due to rain can be prevented.

  Referring to FIG. 4, the rain sensor 14 generally includes a body to be vibrated 22, an acceleration sensor 30 provided on the body to be vibrated 22, an electronic board 36, a battery 35, and a body that supports the body to be vibrated 22. Part 38 and a column part 39 for fixing the rainfall sensor 14 to the ground, and has substantially the same configuration as the rainfall sensor 13 shown in FIG.

  In the rainfall sensor 14, the vibrating body 22 has a pseudo-spherical shape. This has the effect of the rainfall sensor 13 shown in FIG. 3 and can also receive raindrops from below that occur when installed on a mountainous area or an inclined valley slope, so that it is possible to estimate the rainfall more accurately. is there.

  Referring to FIG. 5, the rain sensor 15 includes a roughly flat plate-like vibrating body 23, an acceleration sensor 30 provided on the lower surface of the vibrating body 23, an electronic board 36, a battery 35, and a vibrating body. 23, a support member 41 and a main body 38 for supporting the head 23, and a support column 39 for fixing the rainfall sensor 15 to the ground, and has substantially the same configuration as the rainfall sensor 13 shown in FIG. 3.

  The support member 41 may support the vibrating body 23 on both sides, may have a frame shape that supports the vibrating body 23 along the periphery thereof, and is not particularly limited. By adopting the frame-shaped support member 41, rainwater can be prevented from entering the electronic board or the battery, and a separate waterproof cover need not be provided. For the support member 41, the same material as that of the main body portion 38 can be selected.

  In the rain sensor 15, the vibrating body 23 has a flat plate shape. Thereby, it is possible to reliably receive raindrops from above. Moreover, the to-be-vibrated body 23 may be a photovoltaic power generation panel. By charging a part of the electric power generated by the solar power generation panel to the battery 35, the function of the rainfall sensor 15 can be maintained for a long time.

  As described above, the rain sensors 10 to 15 according to the present embodiment detect and measure the acceleration of the vibration of the vibrating body due to the impact of raindrops using the acceleration sensor 30. Since it does not require a container such as a bag for storing rainwater and a mechanism for measuring the volume, which are necessary in the past, there is no need for complicated mechanical parts, and it is compact, low-cost, and easy to maintain. Furthermore, by making the shape of the vibrating body 20 convex, hemispherical, or pseudo-spherical, it is possible to receive not only rain from above, but also rain from oblique and lateral directions, and the accuracy of rainfall estimation. Will improve. Moreover, since it is hard to accumulate fallen leaves and a sandy substance by these shapes, the influence can be suppressed. For these reasons, even if it is not easy for the user to access the rain gauge, once it is installed, maintenance work can be saved.

[Rainfall estimation device]
FIG. 6 is a block diagram showing a configuration of a rainfall estimation device according to an embodiment of the present invention.

  Referring to FIG. 6, a rainfall estimation device 50 according to an embodiment of the present invention includes a reception interface 52 that receives acceleration data, identification information, and the like wirelessly transmitted from a rainfall sensor via an antenna 51, a CPU 53, A memory 54, a user interface 55, a display 56, a transmission / reception interface 58 that transmits estimated rainfall information to the outside via a server or the Internet, and receives information such as weather, and a bus 59 that connects these signals so as to enable signal communication. And have.

  The reception interface 52 extracts acceleration data, identification information, and the like from the radio signal supplied from the antenna 51 and transmits the data to the CPU 53. The reception interface 52 is, for example, a reception chip or a reception module that can receive a radio wave having a frequency of 920 MHz.

  A CPU (central processing unit, processor) 53 can appropriately select a known MPU (microprocessor). The CPU 53 has a functional unit that controls the rainfall estimation device 50 and performs each process for rainfall estimation described below. Further, the CPU 53 may include a function unit that obtains a regression equation for estimating the rainfall with respect to the rainfall sensor.

  The memory 54 is a RAM (Random Access Memory) and / or a ROM (Read Only Memory), and may be an independent chip or a memory included in the CPU 53. The memory 54 may be used for temporarily storing acceleration data, and stores various data for estimating the rainfall.

  The user interface 55 is an interface for a user operation device, and is connected to an input keyboard (not shown), an operation mouse (not shown), and the like.

  The display 56 is not particularly limited, and a known display can be used. The estimated rainfall obtained by the rainfall estimation device 50, the estimated rainfall for each area indicating the installation location and the estimated rainfall of the rainfall sensors 10 to 15, The accumulated estimated rainfall or the like can be displayed, or information for operating the rainfall estimation device 50 can be displayed.

  The transmission / reception interface 58 can provide estimated rainfall information to the outside via communication with a server storing the estimated rainfall, the Internet or other communication lines, and can receive weather data and the like from the outside.

  Next, the function part for the rainfall estimation implement | achieved by CPU53 of the rainfall estimation apparatus 50 is demonstrated.

  FIG. 7 is a block diagram showing a functional configuration of a rainfall estimation apparatus according to an embodiment of the present invention.

  Referring to FIG. 7, each function component 60 to 65 of CPU 53 is a function realized by a program and data stored in a memory in CPU 53 or an independent memory (hereinafter collectively referred to as “memory 54”). It is. The CPU 53 includes a Fourier transform unit 60, a feature data extraction unit 61, and a rainfall amount estimation unit 62. Further, the CPU 53 may include a filter processing unit 63, a rainfall determination unit 64, and a regression analysis unit 65. The memory 54 stores acceleration data 66 temporarily stored from the rainfall sensor, regression data 68 for rainfall estimation, feature data 69 obtained by machine learning during non-rainfall, and the like.

The Fourier transform unit 60 cuts out acceleration data from the reception interface 52 or the memory 66 in a predetermined time period, and Fourier-transforms the cut out acceleration data into a frequency domain (so-called short-time Fourier transform). The extracted acceleration data is digitized (for example, sampling frequency 100 Hz). When the acceleration data is an analog signal, A / D conversion is performed. Cutting may use a window function. As the window function, known window functions such as a rectangular window, a Gaussian window, a Hann window, and a Hamming window can be used. For example, rectangular windows are set so as not to overlap each other with a window width of 1.28 seconds, for example. Thus, acceleration data for each time period corresponding to the window function is obtained. The Fourier transform unit 60 Fourier transforms the extracted acceleration data for each time period into the frequency domain. Thus, every time period, i.e., for each discrete time, data F ^k _t in the frequency domain is obtained. Data F ^k _t of the frequency domain, when using the complex Fourier transform is expressed by Equation 1 below.

ここで、Ａは振幅、θは位相を表す。添え字のｋは周波数軸上のインデックスであり、ｋ＝１〜ｎの自然数をとり、ｔは、上述した時間期間毎に付される時間軸上のインデックスであり、ｔ＝ｔ、ｔ＋１、・・・、ｔ＋ｍをとり、ｍは自然数である。

Here, A represents amplitude and θ represents phase. The subscript k is an index on the frequency axis, takes a natural number of k = 1 to n, t is an index on the time axis given for each time period described above, and t = t, t + 1,. .., t + m, where m is a natural number.

The feature data extraction unit 61 includes feature data X ^{i, k} _t composed of correlation values between the data F ^k _t (k = 1 to n, t = t, t + 1,..., T + m) obtained by Equation 1 above. To extract.

Figure 8 is a diagram showing an arrangement of the data F ^k _t in the frequency domain in the feature data extraction unit. According to FIG. 8, arranged over the k = 1 to n along the frequency axis complex data F ^k _t for the time period t, them along the time axis t = t, t + 1, ··· , T + m. Then, of which focuses on one of the complex data F ^k _t, and one complex data F ^k _t that the target, the complex data F ^k _t least one complex data F ^k is selected by the given mask based on ^' _A correlation value is calculated from _{t ′} . The correlation value is expressed by Equation 2. Here, an example of obtaining a first-order correlation value is shown.

ここで、（ｋ、ｔ）におけるデータＦ^k _tに対する（ｋ’、ｔ’）におけるデータＦ^k’ _t’はマスクmask_iによって指定される。

Here, the data F ^{k ′} _{t ′} in (k ′, t ′) with respect to the data F ^k _t in (k, t) is specified by the mask mask _i .

FIG. 9 is a diagram illustrating an example of a mask pattern for obtaining a correlation value in the feature data extraction unit 61. Referring to FIG. 9, corresponding to the center of the 3 × 3 is F ^k _t surrounded by a black frame, with respect to the frequency axis, k-1, k, k + 1, with respect to the time axis, t-1, t , T + 1 are taken into account. Here, assuming that the mask is only the first order, (1) to (5) shown in FIG. 9 are considered. (1), F ^k _t and product of the complex conjugate, i.e. represents the mask for obtaining the element size, (2) and (3) each time domain, the correlation values in the frequency domain (respectively, F ^k _t and correlation between F ^k _{t + 1,} represents a mask for obtaining a correlation value) between F ^k _t and _{^{F k + 1 t, (4}} ) and (5) at a region between time and frequency domains Represents a mask for obtaining a correlation value (correlation value between F ^k _t and F ^{k + 1} _{t + 1} , correlation value between F ^k _t and F ^k−1 _{t + 1} , respectively). Correlation value of the left side of formula (2) by application of these masks X ^{i, k} _t is calculated. Here, five correlation values of i = 1 to 5 are calculated for one set (k, t). Incidentally, the correlation value X ^{i, k} _t is not only primary, may of course be to calculate the correlation values correlation also taken into account in the second or higher order.

In this way, the feature data extraction unit 61 calculates the correlation value X ^{i, k} _t of the data F ^k _t (k = 1 to n, t = t, t + 1,..., T + m), and the feature data To extract. The correlation values X ^{i, k} _t characteristic data X ^i, a ^k _t, which feature vector and an element can be obtained.

Correlation value X ⁱ along frequency ^axis, in order to compress the number of ^k _t, may perform the filtering process, CPU 53 may have a filter processing unit 63 as shown in FIG. The filter processing unit 63 may use, for example, distribution smoothing filter processing, mel filter processing, or the like. In the filter processing, the correlation value X ^{i, k} _t (k = 1 to n, t = t, t + 1,..., T + m) is multiplied by the weight w ^j _k, and the sum is obtained with respect to k. It is expressed as Equation 3.

ここで、ｊはフィルタのバンクを表す。ｔにおける相関値の数は、周波数領域の要素数がマスク数とフィルタのバンク数とを乗算した数に低減（圧縮）される。これにより、相関値Ｘ^i,j _t（Ｘはハット付き）を用いることで、回帰式の要素数が低減される。また、フィルタ処理を雨滴の衝撃による被振動体の固有振動数に関連して重みｗ^j _kを選択することで、より効果的に特徴データを算出できる。また、このような特徴データを要素とした特徴量ベクトルが得られる。

Here, j represents a bank of filters. The number of correlation values at t is reduced (compressed) to the number obtained by multiplying the number of frequency domain elements by the number of masks and the number of filter banks. Thereby, the number of elements of the regression equation is reduced by using the correlation value X ^{i, j} _t (X is with a hat). Further, the feature data can be calculated more effectively by selecting the weight w ^j _k in relation to the natural frequency of the vibrating body due to the impact of raindrops in the filtering process. Further, a feature amount vector having such feature data as an element is obtained.

Returning to Figure 7, rainfall estimation unit 62 uses the obtained regression equation by machine learning, to estimate the rainfall from the feature data X ^{i, k} _t. The regression equation is obtained by machine learning in association with each rainfall sensor. The rain amount estimation unit 62 calls the rain rate estimation regression data 68 corresponding to the rain amount sensor from the memory 54 and estimates the rain amount. Note that the feature data (feature quantity vector) used in the regression equation is determined when the regression equation is obtained.

  In addition, although it demonstrates in the process which calculates | requires the regression equation mentioned later, when feature data is compressed and a regression equation is calculated | required based on the compression feature-value vector which makes the compressed feature data an element, feature data The extraction unit 61 extracts feature data corresponding to each element of the compressed feature vector. And the rain amount estimation part 62 estimates rain amount from the extracted feature data using the regression formula corresponding to a compression feature-value vector. As a result, the number of feature data is greatly reduced, so that the calculation cost can be reduced and the rainfall can be estimated efficiently.

The rain determination unit 64 performs machine learning using the non-rainy feature data in the environment where the rainfall sensors 10 to 15 are installed, and whether or not it is raining at the time of actual measurement based on the following environmental feature data Determine. Specifically, the rainfall estimation device 50 receives acceleration data measured during non-rainfall in an environment where the rain sensors 10 to 15 are installed, and the Fourier transform unit 60 and the feature data extraction unit 61 perform frequency domain data F. Extract feature data consisting of ^k _t correlation values. The processes in the Fourier transform unit 60 and the feature data extraction unit 61 are the same as those described above. A principal component analysis is performed on the feature data to obtain n eigenvalues and eigenvectors. Among these eigenvectors, a space spanned by eigenvectors up to a dimension where the cumulative contribution rate becomes a predetermined rate (for example, 0.99) is defined as a normal subspace. Feature data (referred to as “environment feature data”) corresponding to the elements of this eigenvector is acquired in advance and stored in the memory. The rain determining unit 64 calculates a normal subspace spanned by environmental feature data, obtains a degree of deviation of the partial space spanned by the feature data obtained during actual measurement from the normal subspace, that is, a deviation d, and the deviation degree When d is less than or equal to the threshold value, it is determined that it is not raining. The degree of deviation d is represented by the distance (so-called Glasman distance) to the normal partial space of the partial space spanned by the characteristic data obtained during actual measurement. The rain determination unit 64 is provided as necessary in consideration of the environment of the rainfall sensors 10 to 15 and the like.

  The regression analysis unit 65 obtains a regression equation that represents the relationship between the estimated rainfall and the feature data by machine learning using the training data.

Training data is obtained as follows. Using rainfall or artificial rain, acceleration data is acquired by the rainfall sensors 10 to 15 and the rainfall is measured. The acceleration data received by the rainfall estimation device 50 extracts feature data comprising a correlation value of the data F ^k _t in the frequency domain by the Fourier transform unit 60 and the feature data extraction unit 61. The processes in the Fourier transform unit 60 and the feature data extraction unit 61 are the same as those described above. Here, as an example, 250 feature data are extracted, that is, a 250-dimensional feature vector is formed. The training data is a set of this feature data and measured rainfall data.

  The regression analysis unit 65 performs regression analysis of the feature data and the actually measured rainfall using this training data, and obtains a regression equation for estimating the rainfall from the feature data. The regression analysis is a multiple regression analysis, and it can be assumed that the relationship between rainfall and feature data is linear or nonlinear. For example, assuming that the relationship is linear, that is, a linear function y = f (x), linear multiple regression analysis is performed. Here, y is the actually measured rainfall, and x is a feature vector (for example, 250-dimensional feature vector) whose elements are feature data (for example, 250 feature data). A regression equation having coefficients corresponding to 250-dimensional elements is obtained by multiple regression analysis. For this calculation, a general method for minimizing the square error can be used. For this calculation, numerical solution methods such as L-1 regularization (LASSO (Least Absolute Shrinkage and Selection Operator)) and RIDGE (ridge) regularization can be used, which makes it less susceptible to measurement errors. You can find a solution.

  When it is assumed that the relationship between the rainfall and the feature data is nonlinear, for example, a regression equation is obtained by performing multiple regression analysis using a known method such as support vector regression or neural network.

Note that the number of feature data used in the regression analysis unit 65 can be compressed (reduced). For example, as pre-processing of the above-described multiple regression analysis, principal component analysis is performed using rainfall data measured for a short time in the training data and feature data (for example, n feature data) corresponding thereto. As a result of the principal component analysis, n eigenvalues and eigenvectors are obtained. Among the eigenvalues, a predetermined number, for example, five eigenvalues (λ ₁ ,..., Λ 5) are selected in descending order to determine a new feature vector. A compressed feature quantity vector x ′ composed of this predetermined number of new feature data expressed by Expression 4 is used.

この圧縮特徴量ベクトルｘ’と、実測した雨量とを用いて回帰分析部６５において重回帰分析を行って回帰式を求める。これにより、２５０次元の特徴量ベクトルに対してきわめて少ない５次元の特徴量ベクトルを用いることで重回帰分析の計算量を軽減することができる。さらに、重回帰分析で次元数が多い場合に生じ易い、学習が発散する「多重共線性」の問題を回避できる。

The regression analysis unit 65 performs a multiple regression analysis using this compressed feature vector x ′ and the actually measured rainfall to obtain a regression equation. Thereby, the calculation amount of the multiple regression analysis can be reduced by using an extremely small 5-dimensional feature quantity vector with respect to the 250-dimensional feature quantity vector. Furthermore, it is possible to avoid the problem of “multi-collinearity”, which is likely to occur when multiple dimensions are used in multiple regression analysis and the learning diverges.

  In the above description, the regression analysis unit 65 is executed as a function of the CPU 53 of the rainfall estimation device 50. However, the regression analysis unit 65 is mounted on a device different from the rainfall estimation device 50, for example, a computer. The training data may be subjected to regression analysis to derive a regression equation, and the regression equation information may be stored in the memory of the rainfall estimation device. This apparatus has the configuration shown in FIG. 6 (however, the transmission / reception interface 58 may not be provided), and it is sufficient that at least the regression analysis unit 65 is included in the functional configuration of the CPU 53 shown in FIG.

  Moreover, since different acceleration data can be obtained even in the same rain condition depending on the type of the rain sensors 13 to 15, for example, depending on the shape of the vibrating bodies 20 to 23, an apparatus for obtaining a regression equation for the rain sensors 10 to 15 is prepared. Also good. This device has the configuration shown in FIG. 6 (however, the transmission / reception interface 58 may not be provided), and at least the Fourier transform unit 60, the feature data extraction unit 61, and the regression among the functional configurations of the CPU 53 shown in FIG. It suffices to have the analysis unit 65, and may have a filter processing unit 63 as necessary. By executing the function of the regression analysis unit 65 described above, regression equations corresponding to the rainfall sensors 13 to 15 are obtained.

As described above, the Fourier transform unit 60 performs complex Fourier transform, and the frequency domain data F ^k _t represented by the above Equation 1 is a complex number, and the extracted feature data X ^{i, k} _t is It is a complex number. Data F ^k _t of the complex number, includes the size (the amplitude) and phase, the amount of information for two times greater than the actual number of data enables more reliable rainfall estimation.

On the other hand, as a modification, the Fourier transform unit 60 may use real Fourier transform. Thus, since the data F ^k _t in the frequency domain a real number, characteristic data X ^{i, k} _t is a real number. Thus the number of feature data X ^{i, k _t,} i.e. the dimension of the feature vector can be reduced than in the case of complex numbers. Thereby, since the computational complexity of the Fourier-transform part 60, the feature data extraction part 61, and the rainfall estimation part 62 can be reduced, the power consumption for it can be reduced. In such a case, a function for estimating the rainfall can be implemented in the rain sensor 10. In the CPU 31 of the rainfall sensor 10 shown in FIG. 2, the Fourier transform unit 60, the feature data extraction unit 61, and the rainfall estimation unit 62 (including the filter processing unit 63 as necessary) of the CPU 53 of the rainfall estimation device shown in FIG. Processing may be executed. In this case, the memory 32 illustrated in FIG. 2 may store the rain rate estimation regression data 68 stored in the memory 54 illustrated in FIG. 7 and the feature data 69 obtained by machine learning during non-rainfall. . The rainfall sensor 10 can transmit the estimated rainfall to the rainfall estimation device 50 at predetermined time intervals, for example. Thereby, the amount of information to be transmitted can be reduced as compared with the case of transmitting acceleration data, and the rain amount estimating apparatus 50 can communicate with more rain amount sensors 10 to 15. Even when using a complex Fourier transform in Fourier transform unit 60, when determining the correlation value of the complex data F ^k _t in the frequency domain, as a correlation value only the magnitude of the complex data F ^k _t, the actual number of feature data X ^{i, k} _t may be obtained.

  According to the rain amount estimating apparatus 50 according to the present embodiment, the acceleration data of the vibration generated in the vibrating body due to the impact of raindrops is Fourier-transformed to obtain the correlation value of the frequency domain data, which is used as the feature data. Thereby, it is possible to extract the characteristics of the acceleration data with high accuracy, and in the case of complex data, the phase is also taken into consideration, so that it is possible to extract with higher accuracy. Thereby, since the identification accuracy of acceleration data is improved, the accuracy of estimating the rainfall is improved. Furthermore, since the characteristic data is obtained by the same method when obtaining the regression equation, the accuracy of estimating the rainfall is further improved.

  In addition, according to the rain amount estimating apparatus 50 according to the present embodiment, the rain amount estimation can be performed in a short time since the acceleration data based on the impact of the raindrop is used. Thereby, even when a large amount of rain falls in a very short time, it is possible to estimate the rainfall quickly.

[Rainfall estimation method]
FIG. 10 is a flowchart of a rainfall estimation method according to an embodiment of the present invention. Hereinafter, the rain amount estimation method according to the embodiment of the present invention will be described with reference to FIG. 10 and with reference to FIG. 2 and FIG. 6 as necessary.

  First, the rain sensor detects the acceleration of the vibration of the vibrating body due to the impact of raindrops and transmits it as acceleration data (S100). Specifically, in the rain sensor 10 shown in FIG. 2, the acceleration sensor 30 detects the acceleration of vibration of the vibrating body 20 that has received the impact of raindrops, and is transmitted as digitized acceleration data under the control of the CPU 31. Wireless transmission is performed from the interface 33 via the antenna 34, or wired transmission using a cable or the like.

  Next, the rainfall estimation device receives acceleration data from the rainfall sensor (S110). Specifically, in the rain estimation device 50 shown in FIG. 6, in the case of wireless communication, the reception interface 52 receives acceleration data from the rain sensor 10 from the antenna 51 or by wired communication (not shown). At this time, the identification information of the rainfall sensor 10 may be received. The received acceleration data is temporarily stored in the memory 54 together with the identification information or sent to the CPU 53.

Next, the rainfall estimation device Fourier-transforms the acceleration data into frequency domain data for each time period (S120). Specifically, performs clipping acceleration data at a predetermined period of time, the acceleration data cut out by Fourier transform to the frequency domain to obtain data F ^k _t in the frequency domain is expressed by Equation 1 described above. The window function may be used for extraction. As the window function, the known window functions described above can be used. Thus, acceleration data for each time period corresponding to the window function is obtained. Fourier transform to the time period for each of the acceleration data in each time period to the frequency domain, that is, obtained for each discrete time, the data F ^k _t in the frequency domain. When using the complex Fourier transform, to obtain a complex data F ^k _t expressed by Equation 1 below.

Next, the rainfall estimation device extracts feature data composed of correlation values between frequency domain data (S130). Specifically, it extracts the data _{^{F k t (k = 1~n,}} t = t, t + 1, ···, t + m) , wherein data X ⁱ consisting of a correlation value ^{between, k} _t. Specifically, as shown in FIG. 8, the complex data F ^k _t are arranged over the k = 1 to n along the frequency axis with respect to time period t, them along the time axis t = t , T + 1,..., T + m. Then, of which focuses one of the complex data F ^k _t (in a single complex data F ^k _t that the target, at least one complex data F ^k is selected by the given mask based on the complex data _^'t' those calculated from the. correlation values X ^{i, k} _t is calculated by equation 2 described above. wherein the correlation value X ^{i, k} _t is the characteristic data X ^{i, k _t,} that with this element A quantity vector is obtained.

Incidentally, the correlation value X ⁱ along the frequency ^axis, in order to compress the number of ^k _t, may perform filtering. For example, distributed smoothing filter processing, Mel filter processing, or the like may be used as the filter processing. In the filter processing, the correlation value X ^{i, k} _t (k = 1 to n, t = t, t + 1,..., T + m) is multiplied by the weight w ^j _k, and the sum is obtained with respect to k. It is expressed as Equation 3 above. The correlation value X ^{i, j} _t (X with a hat) obtained by this is feature data, and by using this, the number of elements of the regression equation is reduced.

Next, the rainfall estimation device estimates the rainfall using the regression equation obtained by machine learning for the rainfall sensor (S140). Specifically, a regression equation obtained by machine learning in association with each rainfall sensor is called from the memory, and feature data X ^{i, k} _t (or filtered feature data X ^{i, j} _t (X Substituting a hat))) to obtain the estimated rainfall. From the above, the rainfall amount can be estimated.

  The regression equation used in the step of estimating rainfall is a regression for estimating feature data from feature data by obtaining feature data from measured rainfall and acceleration data acquired at that time, and performing regression analysis by machine learning. Find the formula. This will be specifically described below.

  FIG. 11 is a flowchart illustrating a method for obtaining a regression equation used in the rainfall estimation method according to an embodiment of the present invention. Hereinafter, a method for obtaining a regression equation will be described with reference to FIG. 11 and with reference to FIGS. 2 and 6 as necessary.

  First, acceleration data is acquired by a rainfall sensor used for rainfall estimation using rainfall or artificial rainfall, and rainfall is actually measured (S200). Specifically, acceleration data is acquired by using the rainfall sensor 10 shown in FIG. At the same time, the rainfall is measured by a known method.

Next, the acceleration data for each time period is Fourier transformed (S210). Specifically, acceleration data is acquired from a rainfall sensor, and the acceleration data is cut out in a predetermined time period in the same manner as in step 120 described above, and the extracted acceleration data is Fourier-transformed into the frequency domain. obtain data F ^k _t in the frequency domain. Decide whether the Fourier transform of step 210 is a complex Fourier transform or a real Fourier transform according to whether the feature data of complex number is used or the feature data of real number is used in the rainfall estimation method using the regression equation. Also good.

Next, a correlation value between frequency domain data is calculated to extract feature data (S220). Specifically, to obtain the feature data X ^{i, k} _t data F ^k _t as in step 130 described above. In addition, in order to compress the number of correlation values X ^{i, k} _t along the frequency axis, a filtering process may be performed, and the characteristic data X ^{i, j} _t (X is a hat) is obtained by the filtering process described above. May be obtained.

Next, a regression equation is obtained by machine learning from the feature data and the actually measured rainfall (S230). Specifically, regression analysis of feature data X ^{i, k} _t (or feature data X ^{i, j} _t (X is with a hat)) and measured rainfall is performed, and a regression equation for estimating rainfall from the feature data is obtained. Ask. By the same process as the regression analysis unit described above, a regression equation for the rainfall sensor can be obtained from the feature vector or the compressed feature vector.

  The number of feature data used in S230 of FIG. 11 is compressed (reduced) to determine new feature data, and regression is performed by machine learning from the new feature data (new feature vector) and the measured rainfall. An expression can also be obtained.

  FIG. 12 is a flowchart showing another method for obtaining a regression equation used in the rainfall estimation method according to an embodiment of the present invention. Hereinafter, a method for obtaining the regression equation will be described with reference to FIG.

  Rainfall measured at every predetermined time, for example, every short time, and acceleration data corresponding thereto are acquired (S300). The obtained acceleration data is subjected to Fourier transform (S310) of acceleration data for each time period, and a correlation value between frequency domain data is extracted to extract feature data (S320). Similar to S220, feature data (eg, n feature data) is extracted.

  Next, principal component analysis is performed using the n feature data and the rainfall measured in S300. As a result of the principal component analysis, n eigenvalues and eigenvectors are obtained. Among the eigenvalues, a predetermined number, e.g., 5 eigenvalues are selected in descending order and determined as new feature data, and the compressed feature vector x consisting of this predetermined number of new feature data is expressed by the above-described equation 4. '(S330).

  Next, a regression equation is obtained by performing multiple regression analysis using the compressed feature vector x 'obtained in S330 and the rainfall measured in S300 (S340). Thereby, the calculation amount of the multiple regression analysis can be reduced by using an extremely small 5-dimensional feature quantity vector with respect to the 250-dimensional feature quantity vector. Furthermore, it is possible to avoid the problem of “multi-collinearity”, which is likely to occur when multiple dimensions are used in multiple regression analysis and the learning diverges.

  According to the rainfall estimation method according to the present embodiment, the acceleration data of vibration generated in the vibration sensor of the rainfall sensor due to the impact of raindrops is Fourier-transformed to obtain the correlation value of the frequency domain data, which is used as feature data. This makes it possible to extract the characteristics of acceleration data with high accuracy. Thereby, since the identification accuracy of acceleration data is improved, the accuracy of rainfall estimation is improved.

[Example]
The estimated rainfall of the artificial rainfall was obtained by the rainfall estimation device according to the embodiment of the present invention using the rainfall sensor having the hemispherical shape of the vibrating body shown in FIG. At this time, the Fourier transform unit shown in FIG. 7 uses complex Fourier transform, and the regression analysis unit for obtaining the regression equation compresses the feature data in advance by principal component analysis and selects 21 feature data. In order to verify the estimated rainfall, the rainfall was measured while acquiring acceleration data from rainfall. The actually measured rainfall was 4 conditions of 20 mm, 33 mm, 48 mm, and 66 mm. For each condition, the rainfall amount was estimated four times from the acceleration data for 5 seconds. In addition, these are converted into rainfall per hour.

  FIG. 13 is a diagram illustrating an example of a rainfall estimation system according to an embodiment of the present invention. Referring to FIG. 13, the horizontal axis represents measured rainfall, and the vertical axis represents estimated rainfall. It can be seen that the normal rainfall of 20 mm to the strong rainfall of 66 mm can be estimated well.

  The preferred embodiment of the present invention has been described in detail above, but the present invention is not limited to the specific embodiment, and various modifications and changes can be made within the scope of the present invention described in the claims. Is possible. For example, in the above-described embodiment, the rainfall sensor may be combined with various sensors such as temperature and humidity. The rainfall sensor may be fixed to the ground, or may be installed on the rooftop or side wall of a building. Moreover, the machine learning for obtaining the regression equation may be performed by a computer different from the rainfall estimation device, and is not particularly limited.

  In the above-described embodiment, the rain sensor and the rainfall estimation device are directly connected. However, a repeater may be provided between the rain sensor and the rainfall estimation device. A repeater is installed near a number of rainfall sensors, receives acceleration data, and transmits it to a rainfall estimation device via a communication line such as the Internet.

In addition, the following additional notes are disclosed regarding the above description.
(Supplementary Note 1) Means for receiving acceleration data of vibration generated in a vibrating body due to a raindrop impact from a rain sensor;
Means for Fourier transforming the acceleration data into frequency domain data for each time period;
Means for extracting feature data comprising correlation values between the frequency domain data;
Means for estimating rainfall from the feature data using a regression equation obtained by machine learning for the rain sensor;
A rainfall estimation device comprising:
(Supplementary Note 2) An interface for receiving acceleration data of vibration generated in a vibrating body due to raindrop impact from a rain sensor;
CPU,
A memory for storing regression equations obtained by machine learning;
The CPU includes:
Fourier transforming the acceleration data into frequency domain data for each time period;
Extracting feature data comprising correlation values between the frequency domain data;
Estimating the rainfall from the feature data using a regression equation supplied from the memory to the rainfall sensor;
A rainfall estimation device that executes
(Additional remark 3) The said correlation value arrange | positions the data of the said frequency domain for every said time period on a time axis, pays attention to one of the data, the predetermined data based on the one said focused data, and the data The rainfall estimation apparatus according to appendix 1 or 2, which is calculated from at least one data selected by a mask.
(Additional remark 4) The rain amount estimation apparatus as described in any one of Additional remarks 1-3 further provided with the means or step which produces | generates the feature data compressed by filtering the said correlation value.
(Supplementary Note 5) The regression equation is determined by performing machine learning using the actually measured rainfall and the feature data calculated by the Fourier transform unit and the extraction unit from the acceleration data at that time. The rainfall estimation apparatus as described in any one of -4.
(Supplementary Note 6) The regression equation performs principal component analysis of the actually measured rainfall and the feature data calculated by using the Fourier transform means and the extraction means from the acceleration data for each predetermined time from the rain sensor. To determine eigenvalues, determine a predetermined number of eigenvalues as new feature data in descending order of the eigenvalues, and perform machine learning using the predetermined number of new feature data as feature quantity vectors. 4. The rainfall estimation device according to claim 1.
(Supplementary note 7) The rainfall estimation device according to supplementary note 6, wherein the estimating means or the estimating step estimates the rainfall from the feature data corresponding to the feature vector using the regression equation.
(Additional remark 8) Based on the acceleration data of the vibration which arises in the said to-be-vibrated body at the time of non-raining, the said Fourier-transform means and the means for extraction, or other characteristic data are calculated by the step of Fourier-transform, and the step of extracting Machine learning using other feature data is performed to calculate a subspace spanned by another feature vector having a predetermined cumulative contribution rate, and a deviation degree from the subspace is obtained using the feature data, and the deviation degree The rain amount estimation apparatus according to any one of appendices 1 to 7, further comprising means for determining that it is raining when the value exceeds a predetermined threshold.
(Supplementary Note 9) The Fourier transform is a complex Fourier transform, and the frequency domain data and the feature data are complex numbers.
The rainfall estimation device according to any one of appendices 1 to 8, wherein the means for estimating the rainfall or the estimating step estimates the rainfall by a regression equation for the complex feature data.
(Supplementary Note 10) The Fourier transform is a real Fourier transform, the frequency domain data and the feature data are real numbers,
The rainfall estimation device according to any one of appendices 1 to 8, wherein the means for estimating the rainfall or the estimating step estimates the rainfall by a regression equation for the real feature data.
(Supplementary Note 11) A vibrating body that vibrates due to the impact of raindrops;
An acceleration sensor provided on the vibrating body;
A transmission unit for transmitting acceleration data from the acceleration sensor;
A rain sensor with.
(Additional remark 12) The to-be-vibrated body vibrated by the impact of raindrops,
An acceleration sensor provided on the vibrating body;
A transmission interface for transmitting acceleration data from the acceleration sensor;
A rain sensor with.
(Additional remark 13) The rain sensor of Additional remark 11 or 12 whose said to-be-vibrated body is a convex-shaped body.
(Additional remark 14) The said to-be-vibrated body is a hemisphere or a pseudosphere, The rain sensor of Additional remark 11 or 12.
(Additional remark 15) The said acceleration sensor is the rainfall sensor of Additional remark 13 or 14 provided in the top | zenith of the inner surface of the said to-be-vibrated body.
(Additional remark 16) The rain sensor of Additional remark 11 or 12 whose said to-be-vibrated body is a plate-shaped object.
(Supplementary note 17) The rain sensor according to any one of supplementary notes 11 to 16, wherein the vibrating body is made of a thin metal or plastic.
(Supplementary note 18) The rain sensor according to any one of supplementary notes 11 to 17, wherein the acceleration sensor is a three-axis MEMS sensor.
(Supplementary note 19) means for Fourier transforming the acceleration data into frequency domain data for each time period;
Means for extracting feature data comprising correlation values between the frequency domain data;
Means for estimating rainfall from the feature data using a regression equation obtained by machine learning for the rainfall sensor;
Further comprising
The rainfall sensor according to appendix 11, wherein the transmission unit transmits information of the estimated rainfall instead of the acceleration data.
(Supplementary Note 20) A CPU is further provided.
In the CPU,
Fourier transforming the acceleration data into frequency domain data for each time period;
Extracting feature data comprising correlation values between the frequency domain data;
Using the regression equation obtained by machine learning for the rainfall sensor, estimating the rainfall from the feature data;
And execute
The rainfall sensor according to appendix 12, wherein the transmission interface transmits the information on the estimated rainfall instead of the acceleration data.
(Supplementary Note 21) The Fourier transform is a real Fourier transform, the frequency domain data and the feature data are real numbers,
The rain sensor according to appendix 19 or 20, wherein the means for estimating the rain amount or the step of estimating the rain amount is estimated by a regression equation for the real feature data.
(Additional remark 22) The step which receives the acceleration data of the vibration which arises in a to-be-vibrated body by the impact of a raindrop from a rain sensor,
Fourier transforming the acceleration data into frequency domain data for each time period;
Extracting feature data comprising correlation values between the frequency domain data;
Estimating the rainfall from the feature data using a regression equation obtained by machine learning for the rain sensor;
Rainfall estimation method including.
(Additional remark 23) The rain amount estimation method of Additional remark 22 which further includes the step which a rain sensor detects the acceleration of the vibration of a to-be-vibrated body by the impact of a raindrop, and transmits as acceleration data.
(Supplementary Note 24) The correlation value is obtained by arranging frequency domain data for each time period on a time axis, paying attention to one of the data, and a predetermined mask based on the data and the data. 24. The rainfall estimation method according to appendix 22 or 23, which is calculated from at least one data selected by.
(Supplementary note 25) The rainfall estimation method according to any one of supplementary notes 22 to 24, further including a step of generating feature data compressed by filtering the correlation value.
(Additional remark 26) The step which calculates | requires the said regression equation by performing machine learning using the measured rainfall and the feature data calculated similarly to the step of performing the said Fourier transform from the acceleration data in that case, and the said step of extracting is further included The rain amount estimation method according to any one of Supplementary Notes 22 to 25.
(Supplementary Note 27) Eigenvalues are obtained by performing principal component analysis on the measured rainfall and feature data calculated in the same manner as the Fourier transform step and the extraction step from acceleration data every predetermined time from the rain sensor. And a step of determining a predetermined number of the eigenvalues in the descending order of the eigenvalues as new feature data, and performing machine learning using the predetermined number of new feature data as a feature quantity vector to obtain the regression equation. The rainfall estimation method according to any one of 22 to 25.
(Supplementary note 28) The rainfall estimation method according to supplementary note 27, wherein the estimating step estimates the rainfall from the feature data corresponding to the feature vector using the regression equation.
(Supplementary Note 29) Other feature data is calculated in the same manner as the Fourier transform step and the extraction step based on the acceleration data of vibration generated in the vibrating body during non-rainfall, and the other feature data is used. Machine learning is performed to calculate a subspace spanned by another feature vector having a predetermined cumulative contribution rate, and a deviation degree from the partial space is obtained using the feature data, and the deviation degree exceeds a predetermined threshold value The rain amount estimating method according to any one of appendices 22 to 27, further including a step of determining that it is raining in some cases prior to the estimating step.
(Supplementary Note 30) The Fourier transform is a complex Fourier transform, the frequency domain data and the feature data are complex numbers,
The rainfall estimation method according to any one of appendices 22 to 29, wherein the step of estimating the rainfall is estimated by a regression equation for the complex feature data.
(Supplementary Note 31) The Fourier transform is a real Fourier transform, the frequency domain data and the feature data are real numbers,
The rainfall estimation method according to any one of appendices 22 to 29, wherein the step of estimating the rainfall is estimated by a regression equation for the real feature data.
(Supplementary Note 32) Rain sensor,
A rainfall estimation device;
A rainfall estimation system comprising:
The rainfall sensor
A body to be vibrated by the impact of raindrops;
An acceleration sensor provided on the vibrating body;
A transmission unit for transmitting acceleration data from the acceleration sensor;
Have
The rainfall estimation device
Means for receiving the acceleration data from the rainfall sensor;
Means for Fourier transforming the acceleration data into frequency domain data for each time period;
Means for extracting feature data comprising correlation values between the frequency domain data;
Means for estimating rainfall from the feature data using a regression equation obtained by machine learning for the rain sensor;
Having
The rainfall estimation system.
(Supplementary Note 33) The rainfall sensor according to any one of Supplementary Notes 11 to 21, and
The rainfall estimation device according to any one of appendices 1 to 10,
Rainfall estimation system with

10-15 Rain sensor 20-23 Vibrated body 30 Acceleration sensor 31, 53 CPU
32, 54 Memory 33 Transmission interface 34, 51 Antenna 50 Rain amount estimation device 52 Reception interface 60 Fourier transform unit 61 Feature data extraction unit 62 Rain amount estimation unit 63 Filter processing unit 64 Rain judgment unit 65 Regression analysis unit 100 Rain amount estimation system

Claims (13)

Means for receiving acceleration data of vibration generated in the vibrating body due to raindrop impact from a rain sensor;
Means for Fourier transforming the acceleration data into frequency domain data for each time period;
Means for extracting feature data comprising correlation values between the frequency domain data;
Means for estimating rainfall from the feature data using a regression equation obtained by machine learning for the rain sensor;
A rainfall estimation device comprising:   The regression equation obtains an eigenvalue by performing principal component analysis of the measured rainfall and the feature data calculated using the Fourier transform means and the extraction means from acceleration data every predetermined time from the rain sensor. The rain amount estimation according to claim 1, wherein a predetermined number of the eigenvalues are determined as new feature data in descending order of the eigenvalues, and machine learning is performed using the predetermined number of new feature data as a feature amount vector. apparatus.   The rainfall estimation device according to claim 2, wherein the estimating means estimates the rainfall from the feature data corresponding to the feature vector using the regression equation. A body to be vibrated by the impact of raindrops;
An acceleration sensor provided on the vibrating body;
A transmission unit for transmitting acceleration data from the acceleration sensor;
A rain sensor with.   The rain sensor according to claim 4, wherein the vibrating body is a convex body.   The rain sensor according to claim 4, wherein the vibrating body is hemispherical or pseudospherical. Means for Fourier transforming the acceleration data into frequency domain data for each time period;
Means for extracting feature data comprising correlation values between the frequency domain data;
Means for estimating rainfall from the feature data using a regression equation obtained by machine learning for the rainfall sensor;
Further comprising
The rainfall sensor according to any one of claims 4 to 6, wherein the transmission unit transmits the information on the estimated rainfall instead of the acceleration data. Receiving acceleration data of vibration generated in the vibrating body due to raindrop impact from the rain sensor;
Fourier transforming the acceleration data into frequency domain data for each time period;
Extracting feature data comprising correlation values between the frequency domain data;
Estimating the rainfall from the feature data using a regression equation obtained by machine learning for the rain sensor;
Rainfall estimation method including.   The rain amount estimating method according to claim 8, further comprising a step of the rain sensor detecting acceleration of vibration of the vibrating body due to a raindrop impact and transmitting the acceleration as acceleration data.   The eigenvalue is obtained by performing principal component analysis of the actually measured rain amount and the feature data calculated in the same manner as the step of Fourier transform and the extraction step from the acceleration data every predetermined time from the rain sensor, and the eigenvalue 10. The method further comprises the step of determining a predetermined number of the eigenvalues in the descending order as new feature data, performing machine learning using the predetermined number of new feature data as a feature quantity vector, and obtaining the regression equation. The described rainfall estimation method.   The rainfall estimation method according to claim 10, wherein the estimating step estimates the rainfall from the feature data corresponding to the feature vector using the regression equation.   Other feature data is calculated in the same manner as the Fourier transform step and the extraction step based on acceleration data of vibration generated in the vibrating body during non-rainfall, and machine learning is performed using the other feature data. Calculating a partial space spanned by another feature vector having a predetermined cumulative contribution rate, obtaining a deviation degree from the partial space using the feature data, and if the deviation degree exceeds a predetermined threshold, The rainfall estimation method according to any one of claims 8 to 11, further comprising a step of determining that there is a prior to the step of estimating. A rain sensor,
A rainfall estimation device;
A rainfall estimation system comprising:
The rainfall sensor
A body to be vibrated by the impact of raindrops;
An acceleration sensor provided on the vibrating body;
A transmission unit for transmitting acceleration data from the acceleration sensor;
Have
The rainfall estimation device
Means for receiving the acceleration data from the rainfall sensor;
Means for Fourier transforming the acceleration data into frequency domain data for each time period;
Means for extracting feature data comprising correlation values between the frequency domain data;
Means for estimating rainfall from the feature data using a regression equation obtained by machine learning for the rain sensor;
Having
The rainfall estimation system.

Images