JP6756889B1 - Vortex detector, vortex detection method, program and trained model - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a vortex detection device, a vortex detection method, a program and a trained model capable of identifying an atmospheric vortex related to a gust more quickly and accurately. SOLUTION: The vortex detection device is a measurement data in a predetermined range including each vortex candidate region having a spatial distribution of the movement speed corresponding to the vortex of the atmosphere, which is extracted from the measurement data of the movement speed of the atmosphere by the Doppler measuring device. Is provided as input data, and a determination unit for determining whether or not the measurement data in a predetermined range includes the vortex is provided by using the trained model learned from the training data including the measurement data of the vortex. The determination unit inputs data to the trained model in which each of the measurement data in a predetermined range is arranged in a two-dimensional matrix with respect to the distance value in the radial direction and the angle value in the azimuth direction with respect to the position of the Doppler measuring device. Judgment is made as. [Selection diagram] Fig. 2

Description

The present invention relates to vortex detectors, vortex detection methods, programs and trained models.

Strong winds (gusts) that occur suddenly due to weather conditions often generate extremely strong pressure and adversely affect various means of transportation such as automobiles, railroads, ships, and airplanes, and objects on the ground such as houses. Therefore, in order to quickly deal with the occurrence of a gust, there is a demand for predicting and detecting the gust in real time.

The generation of gusts is known to correlate with the large-scale eddy structures that occur in the atmosphere. Patent Document 1 discloses a technique of using a small Doppler radar to monitor only a low elevation angle at high speed and detect an atmospheric vortex that causes a gust with high resolution. In Patent Document 2, the cloud top altitude and precipitation intensity are obtained from the radar measurement data in the extraction area based on the tornado and lightning occurrence prediction information by the Japan Meteorological Agency, the cumulonimbus clouds where the tornado may occur are identified, and the movement of the cumulonimbus clouds. A technique for determining a gust warning area in a course prediction range estimated according to speed is disclosed.

JP-A-2010-249550 JP-A-2018-63219

However, in the conventional technique, erroneous determination or detection omission often occurs according to the determination standard, and there is a problem that the gust cannot be identified quickly and accurately.

An object of the present invention is to provide a vortex detection device, a vortex detection method, a program, and a trained model capable of identifying an atmospheric vortex related to a gust more quickly and accurately.

In order to achieve the above object, the vortex detection device of the present invention
Input data of the vortex, which is extracted from the measurement data of the moving speed of the atmosphere by the Doppler measuring device, and includes the measurement data in a predetermined range including the vortex candidate regions having the spatial distribution of the moving speed corresponding to the vortex of the atmosphere. A determination unit for determining whether or not the measurement data in the predetermined range includes a vortex is provided by using a trained model learned from the training data including the measurement data.
The determination unit uses the trained model as data in which each of the measurement data in the predetermined range is arranged in a two-dimensional matrix with respect to the distance value in the radial direction and the angle value in the azimuth direction with respect to the position of the Doppler measuring device. It was decided to perform the above determination as input data to.
Therefore, the vortex detection accuracy can be improved by the two-step evaluation. Further, since the discrimination is not performed by a numerically uniform standard, it is possible to achieve both suppression of erroneous detection and suppression of detection omission in a well-balanced manner. Since it is easy to determine whether the boundary surface is along the radial direction (R direction) with respect to the position of the Doppler measuring device, the convergence and divergence of the atmosphere, which is another physical state where the boundary is not along the R direction (actually). Can be accompanied by two or three-dimensional movements), and erroneous detection based on one-dimensional velocity data can be suppressed. In addition, the boundary where the velocity is reversed always extends along the R direction, which facilitates learning. Further, since the time and effort of coordinate conversion is not required in the real-time processing, the processing can be reduced and speeded up, which makes it easy to secure the real-time property.

Also, preferably
Among the vortex candidate regions determined to contain vortices, a specific portion for specifying a region satisfying a predetermined condition is provided.
The determination unit calculates the probability of being a vortex using the trained model.
The specific unit excludes the vortex candidate regions determined to be vortices within a predetermined proximity range by the determination unit other than those having the maximum probability.
Therefore, even when the same vortex is specified as a vortex candidate in a plurality of portions, the optimum vortex can be selected. As a result, a more accurate value can be obtained when estimating the moving speed of the specified vortex.

Also, preferably
Among the vortex candidate regions determined to contain vortices, the specific unit specifies those measured more than a predetermined number of times as vortices.
Therefore, it is possible to appropriately perform the gust generation prediction, and on the other hand, unnecessary processing can be reduced.

Also, preferably
The predetermined range is defined so that the center of the vortex in the vortex candidate region is located at the center.
Therefore, it is possible to improve the determination accuracy of a vortex whose contour and center position are unclear.

Also, preferably
When the rotation direction of the vortex in the vortex candidate region is opposite to the predetermined reference rotation direction, the input adjusting unit is provided by inverting the sign of the velocity in each of the measurement data in the predetermined range to obtain the input data. I will do it.
Therefore, consistent vortex detection can be easily performed based on a unified standard regardless of the rotation direction of the vortex. In addition, it is not necessary to prepare learning data according to the rotation direction of the vortex, which reduces the time and effort.

Also, preferably
Of the measurement data in the predetermined range, the correction data in which the missing part of the measurement is filled with a predetermined value is decomposed into singular values, and among the singular values in the obtained singular value matrix, other than a predetermined number of values from the largest value side. Is changed to zero, and then the correction data is returned to be used as the input data.
Therefore, it is possible to suppress a decrease in detection accuracy due to an unnatural velocity distribution and suppress an omission of detection of atmospheric vortices.

Also, preferably
The trained model uses a convolutional neural network as an algorithm.
The determination unit includes the distance between the center position of the vortex in the vortex candidate region and the Doppler measuring device in the feature amount coupled by the fully connected layer.
Therefore, it is possible to more reliably discriminate a vortex having a shape according to the distance due to the measurement in polar coordinates and the arrangement of each of the above measurement data.

Also, preferably
An extraction unit for extracting the vortex candidate region from the measurement data by the Doppler measuring device based on the fitting to a predetermined vortex structure model is provided.
Therefore, it is possible to reduce the processing load of the entire process for detecting the atmospheric vortex, facilitate real-time performance, improve the detection accuracy, and suppress detection omission and erroneous detection.

Further, the vortex detection method of the present invention
Input data of the vortex, which is extracted from the measurement data of the moving speed of the atmosphere by the Doppler measuring device, and includes the measurement data in a predetermined range including the vortex candidate regions having the spatial distribution of the moving speed corresponding to the vortex of the atmosphere. A determination step of determining whether or not the measurement data in the predetermined range includes a vortex by using the trained model trained by the training data including the measurement data is included.
In the determination step, the trained model is obtained by arranging the measurement data in the predetermined range in a two-dimensional matrix with respect to the distance value in the radial direction and the angle value in the azimuth direction with respect to the position of the Doppler measuring device. The determination is made as input data to.
Therefore, the velocity boundary is unified into a pattern extending in the R direction, false detection and the like can be suppressed, and the detection accuracy can be improved. In addition, learning can be facilitated, processing can be reduced, and speed can be increased.

Also, preferably
The predetermined range is defined so that the center of the vortex in the vortex candidate region is located at the center.
Therefore, it is possible to improve the determination accuracy of a vortex whose outline or the like is unclear.

Also, preferably
When the rotation direction of the vortex in the vortex candidate region is opposite to the predetermined reference rotation direction, an input adjustment step is included in which the sign of the velocity is inverted in each of the measurement data in the predetermined range to obtain the input data. I will do it.
Therefore, consistent vortex detection can be easily performed based on a unified standard regardless of the rotation direction of the vortex. In addition, it is not necessary to prepare learning data according to the rotation direction of the vortex, which reduces the time and effort.

Also, preferably
Of the measurement data in the predetermined range, the correction data in which the missing part of the measurement is filled with a predetermined value is decomposed into singular values, and among the singular values in the obtained singular value matrix, other than a predetermined number of values from the largest value side. Is set to zero, and then the correction data is returned to include the defect completion step which is used as the input data.
Therefore, it is possible to suppress a decrease in detection accuracy due to an unnatural velocity distribution and suppress an omission of detection of atmospheric vortices.

Also, preferably
The trained model uses a convolutional neural network as an algorithm.
In the determination step, the feature amount coupled by the fully connected layer includes the distance between the center position of the vortex in the vortex candidate region and the Doppler measuring device.
Therefore, it is possible to more reliably discriminate a vortex having a shape according to the distance due to the measurement in polar coordinates and the arrangement of each of the above measurement data.

In addition, the program of the present invention
Computer,
Measurement of vortex by inputting measurement data in a predetermined range including each vortex candidate region having a spatial distribution of the movement speed corresponding to the vortex of the atmosphere, extracted from the measurement data of the movement speed of the atmosphere by the Doppler measuring device. Using the trained model trained by the training data including the data, the trained model is operated as a judgment means for outputting a judgment as to whether or not the measurement data in the predetermined range includes a vortex, and the judgment means is operated in the predetermined range. Each of the measurement data of the above is made by using the data arranged in a two-dimensional matrix for the distance value in the radial direction and the angle value in the azimuth angle direction with respect to the position of the Doppler measuring device.
Therefore, the velocity boundary is unified into a pattern extending in the R direction, false detection and the like can be suppressed, and the detection accuracy can be improved. In addition, learning can be facilitated, processing can be reduced, and speed can be increased.

Moreover, the trained model of the present invention
The predetermined range of measurement data including the vortex candidate regions having the spatial distribution of the moving speed corresponding to the vortex of the atmosphere extracted from the measurement data of the moving speed of the atmosphere by the Doppler measuring device is input. It is a trained model that outputs a judgment as to whether or not the measurement data of the range contains a vortex.
The training data in the predetermined range including the measurement data of the vortex is trained by using the algorithm of the machine learning model.
The learning is performed using the data in which each of the measurement data in the predetermined range is arranged in a two-dimensional matrix with respect to the distance value in the radial direction and the angle value in the azimuth direction with respect to the position of the Doppler measuring device. ..
Therefore, the velocity boundary is unified into a pattern extending in the R direction, false detection and the like can be suppressed, and the detection accuracy can be improved. In addition, learning can be facilitated, processing can be reduced, and speed can be increased.

Also, preferably
It is assumed that the center of the vortex is learned by the training data located in the center.
Therefore, it is possible to improve the determination accuracy of a vortex whose outline or the like is unclear.

Also, preferably
When the rotation direction of the vortex in the training data is opposite to the predetermined reference rotation direction, the sign of each velocity of the training data is inverted and input for learning.
Therefore, consistent vortex detection can be easily performed based on a unified standard regardless of the rotation direction of the vortex. In addition, it is not necessary to prepare learning data according to the rotation direction of the vortex, which reduces the time and effort.

Also, preferably
The correction data in which the missing part of the measurement is filled with a predetermined value in the training data is decomposed into singular values, and among the singular values in the obtained singular value matrix, the values other than the predetermined number from the larger value are set to zero. Then, it is decided that the learning is performed by the learning data obtained by returning to the correction data.
Therefore, it is possible to suppress a decrease in detection accuracy due to an unnatural velocity distribution and suppress an omission of detection of atmospheric vortices.

Also, preferably
It has a convolutional layer and a fully connected layer using a convolutional neural network as the algorithm.
In the fully connected layer, each feature amount obtained in the convolution layer and the distance between the center position of the vortex in the vortex candidate region and the Doppler measuring device are input as values to be fully combined.
Therefore, it is possible to more reliably discriminate a vortex having a shape according to the distance due to the measurement in polar coordinates and the arrangement of each of the above measurement data.

According to the present invention, there is an effect that the atmospheric vortex related to the gust can be specified more quickly and accurately.

It is a block diagram which shows the functional structure of the processing apparatus of this embodiment. It is a figure which shows the processing procedure of vortex detection. It is a figure which shows the example of the measurement data arranged in the orthogonal coordinate system. It is a figure which shows the measurement data of the vortex candidate which filled the defective part with zero, and the corrected measurement data. It is a figure explaining the movement prediction of a vortex. It is a flowchart which shows the control procedure of a vortex detection control process. It is a flowchart which shows the control procedure of the generation process of a trained model.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
FIG. 1 is a block diagram showing a functional configuration of the processing device 1 of the present embodiment.
The processing device 1 (vortex detection device) is a computer that performs various processes for detecting a vortex based on the measurement data by the Doppler radar 4 (Doppler measuring device). The processing device 1 includes a control unit 11, a storage unit 12, a display unit 13, an operation reception unit 14, an input / output interface 15, and the like.

The control unit 11 includes a processor such as a CPU (Central Processing Unit) and controls various operations of the processing device 1.

The storage unit 12 provides the CPU with a memory space for work, and stores temporary data, setting data, and various programs. The working memory space and temporary data are allocated to RAM (Random Access Memory), and the setting data and programs are stored in non-volatile memory (including HDD (hard disk drive)) such as flash memory.

The program 121 stored in the storage unit 12 includes a vortex detection control program. The vortex detection control program includes a trained model 1210. Further, the storage unit 12 stores vortex pattern data 122, which is a setting related to the model of the vortex structure.

The display unit 13 performs various displays on the digital display screen (display) based on the control of the control unit 11. The displayed contents may include the position and history of the vortex structure specified by the vortex detection control program.

The operation reception unit 14 has a pointing device such as a keyboard and / or a touch panel, receives an input operation from the outside, and outputs a control signal to the control unit 11. The operation reception unit 14 is used, for example, when manually starting or terminating various programs, or when switching the display contents by the display unit 13.

The input / output interface 15 includes a communication connection terminal and a driver with an external device, for example, a LAN (Local Area Network) card and the like. Here, the external device 2 and the meteorological data server 3 can be connected via a LAN (and the Internet). Further, the input / output interface 15 may include a connection terminal for a portable recording medium such as various optical discs.

The external device 2 stores learning data 21 for generating and improving the trained model 1210, and the control unit 11 can acquire and read these data. Here, the learning data 21 is supervised data, and data indicating a correct answer is associated and held.

The meteorological data server 3 inputs the measurement data by the Doppler radar 4 in substantially real time and distributes it to the outside together with the past historical data.

Next, the vortex detection of the present embodiment will be described.
The vortex detection is performed using the measurement data by the Doppler radar 4. The Doppler radar 4 receives a reflected wave due to raindrops of a predetermined electromagnetic wave (microwave or the like) emitted, and the raindrops in the radial direction (R direction; the value of the distance from the Doppler radar 4) of the Doppler radar 4, that is, The spatial distribution of the moving speed of the atmosphere (wind speed VR) is acquired based on the Doppler effect. The electromagnetic wave transmission / reception direction of the Doppler radar 4 goes around in a predetermined cycle (for example, 30 seconds) while rotating and scanning in the azimuth direction (direction A; the value of the horizontal angle difference with respect to the predetermined reference direction), and the Doppler radar 4 makes a round in this cycle. Wind velocity distribution data in all directions around No. 4 is obtained.

If there is a vortex structure in the measurement data, the movement speed of the straight line (line segment) in the R direction including the center of the vortex is opposite to that of the Doppler radar direction (-R direction) (minimum velocity) (+ R direction). ) Part (maximum velocity) is adjacent to the shear structure (maximum / minimum velocity pair structure arranged in the A direction).

FIG. 2 is a diagram showing a processing procedure for vortex detection by the processing device 1 of the present embodiment.
First, the maximum / minimum pair structure corresponding to the vortex structure is detected from the radar measurement data, and vortex candidates are extracted (PR1; extraction unit). Input the measurement data of a predetermined range including the extracted vortex candidate region into a trained model (here, a convolutional neural network; CNN, is used as an algorithm), and use this, that is, in the convolutional layer. After convolution is performed to extract the feature amount (feature map is generated), the feature amount is fully connected in the fully connected layer, and the probability of being a vortex is calculated (PR2; determination unit).

When a plurality of vortices are detected in close proximity, those judged to be duplicates are deleted (PR3). Further, in the measurement data obtained for each rotation scan cycle, it is determined whether or not the detection is performed a predetermined number of times (for example, three times) or more within the reference time (PR4). PR3 and PR4 are included in the specific part of this embodiment. Then, the detection result of the vortex is confirmed, and the moving speed is estimated (PR5).

The detection of vortex candidates (PR1) from radar measurement data is performed by fitting the velocity distribution using a vortex model (a vortex structure model such as the Rankine vortex model, which is well known in the past) stored in the vortex pattern data 122, as in the conventional case. Is done by. The two-dimensional data of the wind speed VR (scalar value) is treated in the same manner as the pixel value of the image data (here, with a sign). At this time, instead of converting the measurement data into a Cartesian coordinate system or the like according to the actual spatial distribution (for example, x = r · cоs θ, y = r · sin θ), the Cartesian coordinate system for the A direction and the R direction (for example) For example, the data is treated as being arranged in a two-dimensional matrix with x = r and y = θ).

FIG. 3 is a diagram showing an example of measurement data arranged in this Cartesian coordinate system. Here, the shade of velocity is adjusted in units of appropriate values. By this arrangement, the boundary of the velocity portion in the opposite direction is unified so as to extend in the R direction. That is, since phenomena such as divergence and convergence in which the boundary in the velocity direction extends in the A direction are easily excluded, the pattern to be recognized is more easily limited, which leads to certainty of recognition and simplification of the learning process. Moreover, since the labor of coordinate conversion is saved, the processing itself becomes easy.

The measurement data M of the region of a predetermined size including the region of the vortex candidate extracted by the fitting is input to the trained model 1210, and the probability of having a vortex structure is output (PR2). The measurement data M in which the Doppler velocity (scalar value) is arranged two-dimensionally can be handled in the same manner as the image data, and CNN determines whether or not it is a vortex by using pattern recognition. CNN is a stage (convolution layer) in which a feature map is generated by a convolution process and an extraction of a feature amount is performed (a pooling process may be included), and a stage in which a probability of being a vortex is calculated by a total combination of the feature amounts. It is divided into (fully bonded layer).

The convolution and pooling processes are the same as in the prior art, and the convolution of the number of steps according to the trained model 1210 is repeated using the coefficient of the filter determined by learning. However, in the measurement data M by the Doppler radar 4, there is a problem that the reflection from the position where there is no raindrop cannot be obtained due to the characteristic of acquiring the reflection from the raindrop, and the data is often lost. On the other hand, in the present embodiment, the missing data is filled in (complemented) and then input to the trained model 1210.

At this time, if the holes are simply filled with zero values or the like, the velocity distribution becomes unnatural and the accuracy of vortex detection is lowered. Therefore, here, singular value decomposition (M = USV ^VT ) is performed on the correction data in which the coordinates of the data loss in the measurement data M are once filled with a predetermined value such as a zero value. U and V are unitary matrices, and S is a singular value matrix. Although not particularly limited, when U and V are square matrices, the diagonal components of the singular value matrix are singular values. Each singular value corresponds to the feature quantity of the spatial change of the value in the measurement data M.

Among the singular values in the singular value matrix S, a matrix Sa is generated in which a predetermined number (for example, 5 or the like) is left from the side with the larger value and the other values are changed to zero. Then, by performing a matrix operation ^{Ma = U · Sa · V T} , the input of the corrected measurement data Ma that have been converted to the interpolation value that reflects the main trend of the spatial change (correction data returned) to the trained model 1210 Let it be data.

FIG. 4 is a diagram showing the region measurement data of the vortex candidate in which the defective portion is filled with zeros and the corrected measurement data. In the data filled with the predetermined values (fixed values) shown in FIG. 4 (a), unnatural parts with no spatial change of numerical values are mixed in some areas in the region where the Doppler velocity is high, but the correction shown in FIG. 4 (b). In the completed measurement data, these parts are complemented (interpolated) with values according to the tendency of the surrounding velocity change.

Further, as shown in FIG. 4, a region having a predetermined size is defined so that the center of the vortex candidate extracted by the processing PR1 becomes the center of the measurement data M. The center of the vortex is defined as the midpoint between the maximum velocity position and the minimum velocity position.

Further, the vortex is not necessarily limited to either clockwise (clockwise) or counterclockwise (counterclockwise). However, within the range in which the presence or absence of the vortex is determined, there is no large difference in shape depending on the direction of rotation. Therefore, in the present embodiment, when the rotation direction of the vortex obtained by fitting is different from the predetermined reference rotation direction (for example, counterclockwise) (in the opposite direction), the sign of the velocity within the region of the predetermined size Is reversed to reverse the direction of rotation. As a result, in the trained model 1210, it is sufficient if only the vortex rotating in the reference rotation direction can be recognized, and the learning content is simplified.

Comparing the probability of being a vortex calculated by the above PR2 process with a predetermined reference value, a vortex candidate smaller than the reference value is considered not to be a vortex.

Further, in the present embodiment, the trained model 1210 uses CNN as an algorithm. At this time, not only the feature amount obtained by folding the input data but also the distance between the center position of the target region (vortex candidate) and the Doppler radar 4 is included in the feature amount and fully combined. As described above, since the Doppler radar 4 radiates electromagnetic waves radially from the transmitter / receiver, the shape of the vortex measured tends to become coarser as the distance from the Doppler radar 4 increases. In addition, when the measurement data is arranged in a two-dimensional matrix in the R direction and the A direction, the distance in the A direction in the actual space changes according to the distance value in the R direction. The shape (aspect ratio) changes. Therefore, by including the distance in the feature amount as described above, it is possible to more reliably detect a vortex having a shape corresponding to the distance.

Further, when a plurality of vortex candidates are extracted in a range smaller than the size of the vortex (a predetermined proximity range), the same vortex may be extracted in duplicate. In the present embodiment, when a plurality of vortex candidates are extracted within a predetermined distance, those other than those having the maximum probability of being vortices are deleted (excluded) from the vortex candidates (PR3). In the image recognition of the vortex, the outline is unclear as compared with the image of a normal object or character, and the center position is indirectly determined from the maximum velocity position and the minimum velocity position as described above. On the other hand, as described above, the center position of the vortex candidate is set as the center of a predetermined region, and as described later, the center of the vortex in each learning data 21 is indicated by the training data 21 at the center of the region (N × N matrix). By matching the diagonal intersection coordinates of the vortex, the vortex candidate closer to the true center of the vortex is more likely to be a vortex, so that the best candidate can be appropriately left based on the probability at the time of duplicate extraction. It will be possible.

Further, since it is sufficient to detect vortices that occur continuously to some extent, here, the second detection is performed within a predetermined time from the previous detection, and the third detection (more than a predetermined number of times) within a predetermined time from the second detection. ) Is detected as a continuous vortex (PR4). The predetermined time may be, for example, several times the rotation scan cycle of the Doppler radar 4, that is, it is not always limited to the case where it is continuously detected every time.

FIG. 5 is a diagram for explaining the movement prediction of the vortex.
The vortex is often accompanied by movement, but since the movement speed is not specified by the first detection, the second detection is, for example, the movement speed normally assumed from the position P1 detected the first time (with the background). Position within the range of Rp1 = Vom × dt1 within the range that can be moved between the measurement interval dt1 (twice the measurement interval dt1 if the measurement is performed twice) at the upper limit Vom of the wind (environmental wind) When it is made in P2, it is judged to be the same vortex. Further, since the moving speed Vp between them is specified by the two detections, the third estimated detection position P3p can be obtained by the moving speed Vp and the moving time until the detection (measurement interval dt1). Variation of expected moving speed Vp around this assumed detection position P3p (It may be a statistical measurement value, for example, it may be a ratio to the speed, or it may be an absolute value). When the third detection is made within the range Rp2 (here, the radius Rp2) estimated according to the above, it is determined that the vortex is the same as the previous two times. The upper limit Vom of the normally assumed movement speed and the upper limit of the variation in the calculated movement speed may be fixedly or variably determined based on past statistical values or the like.

FIG. 6 is a flowchart showing a control procedure by the control unit 11 of the vortex detection control process executed based on the program 121 in the processing device 1 of the present embodiment. This vortex detection control process is called and executed every time the measurement data of each cycle by the Doppler radar 4 is acquired (stored in the storage unit 12) from the meteorological data server 3 for each cycle.

When the vortex detection control process is started, the control unit 11 reads the measurement data from the storage unit 12 (step S101). The control unit 11 detects the minimum / maximum pair structure of the velocities arranged in the A direction corresponding to the vortex for each coordinate of the measurement data (step S102). The control unit 11 performs fitting to the region including one detected pair structure based on the vortex structure model included in the vortex pattern data 122 (step S103). The control unit 11 specifies the degree of matching with the vortex structure, the center position, and the rotation direction of the vortex based on the fitting result (step S104). Further, parameters other than the above, which are not used for specifying the vortex structure, for example, the rotation speed of the vortex and the vorticity may be calculated together as an index indicating the strength of the vortex.

The control unit 11 extracts an image in a predetermined range in which the center position of the detected vortex structure is centered. When the rotation direction of the vortex is different from the reference rotation direction, the control unit 11 inverts the sign of the velocity at each position (step S105; input adjustment unit, input adjustment step). The control unit 11 performs the missing data complementing process as described above (step S106; missing complementing section, missing complementing step).

The control unit 11 inputs an image of a predetermined range extracted and adjusted as necessary into the trained model 1210, and acquires the probability of being a vortex output from the trained model 1210 (step S107; Judgment unit, judgment step, judgment means). The processes of steps S103 to S107 are sequentially performed for all the pairs detected in the process of step S102. The control unit 11 excludes vortices having a probability of being less than a predetermined reference value from the vortex candidates (step S108).

The control unit 11 determines whether or not there is a portion having a plurality of vortex candidates within a predetermined distance range (step S109). If it is determined that there is (“YES” in step S109), the control unit 11 compares the probabilities that the vortex candidates in the distance range are vortices, and excludes the vortex candidates other than the highest one from the vortex candidates (step). S110). Then, the process of the control unit 11 shifts to step S111. When it is determined that there is no portion having a plurality of vortex candidates (“NO” in step S109), the processing of the control unit 11 proceeds to step S111.

When the process shifts to the process of step S111, the control unit 11 has a predetermined distance (described above) from the vortex structure detected within a predetermined reference time (for example, within 4 cycles) from the current vortex detection among the remaining vortex candidates. The distance obtained by multiplying the elapsed time from the first detection by a predetermined reference speed, and the distance calculated by multiplying the speed calculated based on the movement at the time of the second detection and the elapsed time from the second detection) Select the one inside the above and add 1 to the number of detections. Further, if the distance is not within the predetermined distance, the number of detections is set to "1" (step S111).

The control unit 11 determines whether or not the number of detections is equal to or greater than a predetermined reference number (for example, 3 times) for each of the remaining vortex candidates (step S112). The control unit 11 determines that the number of times is equal to or greater than the reference number (“YES” in step S112) as a vortex structure (step S114). For the determined vortex structure, the moving speed (including the direction) is estimated and stored together. The velocity in this case is the velocity in normal space. The control unit 11 holds data with the number of detections as vortex candidates when the number of detections is less than the reference number (“NO” in step S112) (step S113). The control unit 11 calculates the moving speed for the vortex candidate for the second detection, and sets a predetermined reference speed for the vortex candidate for the first detection. Then, the control unit 11 ends the vortex detection control process.

Next, the learning operation of the machine learning model will be described.
The trained model 1210 is generated by learning using measurement data including a large number of vortex structures as training data. As the learning data 21, data for which it is determined in advance whether or not the vortex structure is formed is prepared in the same size as the size at the time of vortex detection. The ratio of vortices to non-vortices may be determined as appropriate. Further, as described above, the roughness of the vortex measurement and the shape on the array may change depending on the distance from the Doppler radar 4, so that the learning data 21 covering the measurement in each distance range is prepared. Good.

Further, in the learning data 21, as shown above, the rotation direction of the vortex is unified. The vortex often rotates in one direction (counterclockwise), but not necessarily in one direction. Further, since the rotation direction is biased, it is difficult to prepare the required number of measurement data of the vortices rotating in each direction, and it takes time and effort to divert the measurement data of other Doppler radars. Since the Doppler radar measures only the velocity in the radial direction, the rotation direction of the vortex can be easily reversed by simply reversing the sign of the measured velocity.

Further, the vortex measurement data used as the learning data 21 is determined so that the center of the vortex is always located at the center of the image region. Unlike normal image recognition, when learning measurement data whose outer boundaries are unclear, the center position is determined in advance and the learning data 21 is given, so that a position other than the original center of the vortex is set in the center for image recognition. It is possible to reduce the probability of having a vortex structure when it is made to do so.

Here, the learning data 21 is generated outside the processing device 1 and held by the external device 2, but is not limited to this.

FIG. 7 is a flowchart showing a control procedure of the generation process of the trained model of the present embodiment.
Here, the model generation process will be described as being executed by the processing device 1.

When the model generation process is started, the control unit 11 acquires a list of training data 21 (step S201). The control unit 11 selects one measurement data selected from the acquired list (step S202).

When the rotation direction of the selected measurement data is not a predetermined reference rotation direction, the control unit 11 inverts the sign of each speed to unify the rotation direction to the reference rotation direction (step S203). The rotation direction may be unified to the reference rotation direction in advance. Further, the information on the rotation direction may be included in the list in advance, or may be included in the metadata corresponding to each measurement data. Further, even in the case of data that does not include vortices, if the data is detected as a vortex candidate by fitting, unified processing may be performed based on the rotation direction at the time of detection. In addition, the control unit 11 performs the above-mentioned defect complement processing.

The control unit 11 inputs the selected learning data 21 into the machine learning model (step S204), convolves the preset (designed) steps, and performs feature quantities (which may include pooling). Extraction is performed (step S205). The control unit 11 inputs the distance between the center position of the image region (the center of the vortex in the case of a vortex) and the Doppler radar 4, and fully combines the feature amount and the distance (step S206).

The control unit 11 compares the output result with the (associated) teacher data in the learning data 21, feeds back the deviation (correctness determination error, etc.), and weights the convolution filter coefficient and the total coupling. Correct the coefficient and the like (step S207). For feedback, a usual method, for example, an error backpropagation method, may be used.

The control unit 11 determines whether or not all the vortex measurement data have been selected from the list of acquired vortex measurement data (step S208). If it is determined that there is something that has not been selected (“NO” in step S208), the process of the control unit 11 returns to step S202. When it is determined that all the models have been selected (“YES” in step S208), the control unit 11 determines the model to be learned as a trained model (step S209), and the control unit 11 ends the model generation process. To do.

As described above, the processing device 1 of the present embodiment as the vortex detection device is a vortex candidate having a spatial distribution of the moving speed corresponding to the vortex of the atmosphere extracted from the measurement data of the moving speed of the atmosphere by the Doppler radar 4. Whether or not the measurement data in the predetermined range includes the vortex by using the trained model 1210 trained by the training data 21 including the measurement data of the vortex with the measurement data in the predetermined range including each region as input data. A control unit 11 is provided as a determination unit for determining whether or not. The control unit 11 as a determination unit has a two-dimensional matrix in which each of the measurement data in a predetermined range has a distance value in the radial direction (R direction) and an angle value in the azimuth direction (A direction) with respect to the position of the Doppler radar 4. The data arranged in is determined as input data to the trained model 1210.
In this way, once the vortex candidates are extracted using the prior art, the certainty as a vortex is determined using the trained model 1210, so that it is compared with the case where the determination is performed by a numerically uniform standard. Therefore, the possibility of erroneous detection of vortices and omission of detection can be reduced. Further, since the trained model 1210 is not applied to the entire image from the beginning, an unnecessary load is not applied in the processing. Then, by arranging the measurement data in a two-dimensional matrix using linear coordinate axes extending orthogonally to the R direction and the A direction instead of the distribution in the real space, the boundary where the velocity is inverted is always in the R direction. It will grow along and facilitate learning. Further, since the time and effort of coordinate conversion is not required in the real-time processing, the processing can be reduced and the speed can be increased (speeded up), which makes it easier to secure the real-time property. Furthermore, since it can be easily determined whether or not the boundary surface is along the R direction, the convergence or divergence of the atmosphere, which is another physical state whose boundary is not along the R direction (actually, a two- or three-dimensional movement). Can be easily identified with (may be accompanied by), and erroneous detection by one-dimensional velocity data can be suppressed.

Further, the control unit 11 specifies, as a specific unit, a vortex candidate region determined to include a vortex that satisfies a predetermined condition such as spatial exclusivity. That is, the control unit 11 calculates the probability that the vortex candidate in the target region is a vortex using the trained model 1210 as the determination unit, and one or a plurality of vortices determined to be vortices within a predetermined proximity range as the specific unit. Exclude the vortex candidate regions other than those with the highest probability.
In this way, even when the same vortex is specified as a vortex candidate in a plurality of portions, the optimum vortex can be selected. As a result, a more accurate value can be obtained when estimating the moving speed of the specified vortex.

Further, the control unit 11 specifies, as a specific unit, a vortex candidate region determined to include a vortex, which has been measured a predetermined number of times or more as a vortex. That is, since the one in which the vortex structure continuously exists to some extent is selected and specified, it is possible to appropriately perform the gust generation prediction, and on the other hand, unnecessary processing can be reduced.

Further, the predetermined range is defined so that the center of the vortex in the vortex candidate region is located at the center. Compared to normal image recognition, the outer boundary of the vortex is unclear, the size varies, and the center position is difficult to determine directly. Therefore, even when using CNN, by fixing it in the center, the vortex can be fixed. Judgment accuracy can be improved.

Further, as an input adjusting unit, the control unit 11 reverses the sign of the velocity in each of the measurement data in a predetermined range when the rotation direction of the vortex in the vortex candidate region is opposite to the predetermined reference rotation direction. Use as input data. The direction of rotation of the vortex is not necessarily limited to one direction, but it is almost a target, so the direction of rotation itself does not matter whether or not there is a vortex. Since it can be easily inverted by inverting the code of the one-dimensional velocity data, by inverting and unifying the rotation direction and then calculating the probability of being a vortex, consistent vortex detection can be easily performed based on a unified standard. It will be possible. In addition, it is not necessary to prepare learning data according to the rotation direction of the vortex, which reduces the time and effort.

Further, the control unit 11, as a defect complementing unit, decomposes the correction data in which the measurement defect portion of the measurement data in a predetermined range is filled with a predetermined value into singular values, and among the singular values in the obtained singular value matrix. After changing the values other than a predetermined number of values to zero from the side with the largest value, return to the correction data and use it as input data. In this way, the measurement data of the Doppler radar 4, which is prone to defects depending on the presence or absence of raindrops, is appropriately complemented according to the surrounding tendency, resulting in an unnatural velocity distribution due to fixed values or linear interpolation. It is possible to suppress a decrease in detection accuracy and suppress omission of detection of atmospheric vortices.

Further, the trained model 1210 uses a convolutional neural network as an algorithm, and the control unit 11 uses the convolutional neural network as a determination unit as a determination unit for the center position of the vortex in the vortex candidate region and the Doppler measurement device for the feature quantity connected by the fully connected layer. Include the distance to. In the measurement of Doppler radar 4, the spatial resolution of the measurement becomes coarser according to the distance from the transmission / reception position of the electromagnetic wave. Therefore, by including this distance in the feature amount, the vortex of the shape according to the distance can be more reliably discriminated. can do. Further, as described above, in the Cartesian coordinate system in the R direction and the A direction, the distance in the real space in the A direction changes according to the distance in the R direction, so that the aspect ratio of the vortex also changes according to the distance. Can be done. Therefore, the accuracy of discriminating the vortex having a shape that changes according to the distance is improved including such an influence.

Further, as an extraction unit, the control unit 11 extracts a vortex candidate region from the measurement data by the Doppler radar 4 based on the fitting to the predetermined vortex structure model related to the vortex pattern data 122. In this way, the region where there is a possibility of vortex is first extracted according to the plausibility of the spatial distribution of velocity as compared with the model, so the region that requires easy and high-speed more accurate processing with less load processing is limited. be able to. Therefore, it is possible to reduce the processing load of the entire process for detecting the atmospheric vortex, facilitate real-time performance, improve the detection accuracy, and suppress detection omission and erroneous detection.

Further, by the vortex detection method using each process performed by the processing device 1, it is possible to more appropriately suppress erroneous detection and detection omission while reducing the processing load and increasing the speed.

Further, by performing each of the above processes as software based on the program 121 including the trained model 1210 installed in the computer, it is not necessary to prepare special hardware equipment, and it is more practical at an appropriate cost. Vortex detection can be performed.

Further, by calculating the probability of being a vortex using the trained model 1210 and determining whether or not it is a vortex, it is possible to more effectively suppress erroneous detection and detection omission according to the conventional determination criteria. it can.

The present invention is not limited to the above embodiment, and various modifications can be made.
For example, in the above embodiment, as a method of extracting the target to be input to the trained model 1210, the region of the vortex (maximum / minimum velocity pair) is detected by fitting, but other conventional methods for detecting the vortex. Method may be used.

Further, in the above embodiment, the distance from the Doppler radar 4 is added to the feature quantity that is fully coupled in the fully coupled layer, but the distance value is not directly added, but the pixel in the R direction. The distance may be expressed by a number or the like. Alternatively, the actual distance per pixel in the A direction may be added. Also, the distance value does not have to be added as a feature quantity.

Further, in the above embodiment, the processing is performed by unifying the rotation direction of the vortex, but the processing may be performed separately. In this case, for example, in the learning data 21, the vortex of rotation in one direction may be inverted and used again as the learning data.

Further, the complement processing of the data missing portion using the singular value decomposition as shown in the above embodiment is one of the preferable examples, and other complement processing may be performed. Further, even when the singular value decomposition is used, the number of singular values that are not set to zero is not limited to five cases.

Further, in the above embodiment, the center of the vortex (vortex candidate region) is selected so as to be aligned with the center of the predetermined range, and is aligned with the center of the region indicated by the learning data 21, but it is not strictly the center. It may be adjusted to a predetermined position other than the center. Alternatively, it does not have to be aligned at all.

Further, in the above embodiment, it has been described that CNN is used as an algorithm of the machine learning model to extract features by convolution (and pooling depending on the design) and to calculate the probability by full coupling. The algorithm of the model is not limited to this, and other algorithms may be used in part or in whole.

Further, in the above embodiment, the measurement data of the Doppler radar has been described as an example, but the measurement data of the Doppler lidar may be used.

Further, in the above embodiment, the trained model 1210 is generated in the processing device 1 based on the learning data 21 of the external device 2, but it is completely generated externally and then via the input / output interface 15. The learning data 21 may also be generated in the processing device 1.

Further, in the above-described embodiment, the storage unit 12 composed of a flash memory, an HDD, or the like has been described as an example as a computer-readable medium of the program 121 related to the image formation control process including the setting of the recording medium or the like. Not limited to. As other computer-readable media, other non-volatile memories and portable recording media such as CD-ROMs and DVD discs can be applied. A carrier wave is also applied to the present invention as a medium for providing data of a program according to the present invention via a communication line.
In addition, specific details such as the configuration, processing contents, and procedure shown in the above embodiment can be appropriately changed without departing from the spirit of the present invention.

Further, by constructing a system that detects a gust in real time using the vortex detection device, the vortex detection method, the program, and the learned model of the present invention and predicts the course and approach time of the gust, the damage caused by the gust can be caused. It will be possible to provide warning information and regulatory information to the residents and transportation of the expected area at an early stage.

1 Processing device 2 External device 21 Learning data 3 Meteorological data server 4 Doppler radar 11 Control unit 12 Storage unit 121 Program 1210 Learned model 122 Swirl pattern data 13 Display unit 14 Operation reception unit 15 Input / output interface

Claims (19)

Input data of the vortex, which is extracted from the measurement data of the moving speed of the atmosphere by the Doppler measuring device, and includes the measurement data in a predetermined range including the vortex candidate regions having the spatial distribution of the moving speed corresponding to the vortex of the atmosphere. A determination unit for determining whether or not the measurement data in the predetermined range includes a vortex is provided by using a trained model learned from the training data including the measurement data.
The determination unit uses the trained model as data in which each of the measurement data in the predetermined range is arranged in a two-dimensional matrix with respect to the distance value in the radial direction and the angle value in the azimuth angle direction with respect to the position of the Doppler measuring device. A vortex detection device characterized in that the determination is performed as input data to. Among the vortex candidate regions determined to contain vortices, a specific portion for specifying a region satisfying a predetermined condition is provided.
The determination unit calculates the probability of being a vortex using the trained model.
The vortex detection device according to claim 1, wherein the specific unit excludes a vortex candidate region determined to be a vortex within a predetermined proximity range by the determination unit other than the one having the maximum probability. .. The vortex detection device according to claim 2, wherein the specific unit identifies a vortex candidate region that has been determined to contain a vortex and has been measured a predetermined number of times or more as a vortex. The vortex detection device according to any one of claims 1 to 3, wherein the predetermined range is defined so that the center of the vortex in the vortex candidate region is located at the center. When the rotation direction of the vortex in the vortex candidate region is opposite to the predetermined reference rotation direction, the input adjusting unit is provided by inverting the sign of the velocity in each of the measurement data in the predetermined range to obtain the input data. The vortex detection device according to any one of claims 1 to 4, wherein the vortex detection device is characterized. Of the measurement data in the predetermined range, the correction data in which the missing part of the measurement is filled with a predetermined value is decomposed into singular values, and among the singular values in the obtained singular value matrix, other than a predetermined number of values from the largest value side. The vortex detection device according to any one of claims 1 to 5, wherein the vortex detection device includes a defect complementing portion for changing the value to zero and then returning the correction data to the input data. The trained model uses a convolutional neural network as an algorithm.
The determination unit according to any one of claims 1 to 6, wherein the feature amount coupled by the fully connected layer includes the distance between the center position of the vortex in the vortex candidate region and the Doppler measuring device. The described vortex detector. The vortex according to any one of claims 1 to 7, further comprising an extraction unit that extracts the vortex candidate region from the measurement data by the Doppler measuring device based on the fitting to a predetermined vortex structure model. Detection device. Input data of the vortex, which is extracted from the measurement data of the moving speed of the atmosphere by the Doppler measuring device, and includes the measurement data in a predetermined range including the vortex candidate regions having the spatial distribution of the moving speed corresponding to the vortex of the atmosphere. A determination step of determining whether or not the measurement data in the predetermined range includes a vortex by using the trained model trained by the training data including the measurement data is included.
In the determination step, the trained model is obtained by arranging the measurement data in the predetermined range in a two-dimensional matrix with respect to the distance value in the radial direction and the angle value in the azimuth angle direction with respect to the position of the Doppler measuring device. A vortex detection method characterized in that the determination is performed as input data to. The vortex detection method according to claim 9, wherein the predetermined range is defined so that the center of the vortex in the vortex candidate region is located at the center. When the rotation direction of the vortex in the vortex candidate region is opposite to the predetermined reference rotation direction, an input adjustment step is included in which the sign of the velocity is inverted in each of the measurement data in the predetermined range to obtain the input data. The vortex detection method according to claim 9 or 10, characterized in that. Of the measurement data in the predetermined range, the correction data in which the missing part of the measurement is filled with a predetermined value is decomposed into singular values, and among the singular values in the obtained singular value matrix, other than a predetermined number of values from the largest value side. The vortex detection method according to any one of claims 9 to 11, wherein the vortex detection method includes a defect completion step in which is set to zero and then returned to the correction data to be used as the input data. The trained model uses a convolutional neural network as an algorithm.
In any one of claims 9 to 12, the determination step includes the distance between the center position of the vortex in the vortex candidate region and the Doppler measuring device in the feature amount coupled in the fully connected layer. The described vortex detection method. Computer,
Measurement of vortex by inputting measurement data in a predetermined range including each vortex candidate region having a spatial distribution of the movement speed corresponding to the vortex of the atmosphere, extracted from the measurement data of the movement speed of the atmosphere by the Doppler measuring device. Using the trained model trained by the training data including the data, the trained model is operated as a judgment means for outputting a judgment as to whether or not the measurement data in the predetermined range includes a vortex, and the judgment means is operated in the predetermined range. Each of the measurement data of the above is a program characterized in that the determination is made using the data arranged in a two-dimensional matrix for the distance value in the radial direction and the angle value in the azimuth angle direction with respect to the position of the Doppler measuring device. The predetermined range of measurement data including the vortex candidate regions having the spatial distribution of the moving speed corresponding to the vortex of the atmosphere extracted from the measurement data of the moving speed of the atmosphere by the Doppler measuring device is input. It is a trained model that outputs a judgment as to whether or not the measurement data of the range contains a vortex.
The training data in the predetermined range including the measurement data of the vortex is trained by using the algorithm of the machine learning model.
The learning is performed using the data in which each of the measurement data in the predetermined range is arranged in a two-dimensional matrix with respect to the distance value in the radial direction and the angle value in the azimuth direction with respect to the position of the Doppler measuring device. A trained model characterized by that. The trained model according to claim 15, wherein the center of the vortex is trained by the training data located at the center. 15 or claim 15, wherein when the rotation direction of the vortex in the training data is opposite to the predetermined reference rotation direction, the learning is performed by inverting the sign of each speed of the training data and inputting the vortex. 16 Learned models. The correction data in which the missing part of the measurement is filled with a predetermined value in the training data is decomposed into singular values, and among the singular values in the obtained singular value matrix, the values other than the predetermined number from the larger value are set to zero. The trained model according to any one of claims 15 to 17, wherein the training is performed by the training data obtained by returning to the correction data. It has a convolutional layer and a fully connected layer using a convolutional neural network as the algorithm.
The fully connected layer is characterized in that each feature amount obtained in the convolution layer and the distance between the center position of the vortex in the vortex candidate region and the Doppler measuring device are input as values to be fully combined. The trained model according to any one of claims 15 to 18.