Method for generating a slope anomaly detection model, slope monitoring system, slope monitoring device, slope monitoring server, and computer-readable program.

Abstract

This enables high-precision detection of slope anomalies with a simple hardware configuration. [Solution] The method for generating a slope anomaly detection model includes a preparation step of preparing an information processing model that transforms an input image and outputs the transformed image; a learning step (S11~S15) in which a monitoring image of a normal slope is input to the information processing model as training data, a learning process is performed to identify the transformation parameters that suppress the difference between the input image and the output image of the information processing model, and the range of possible differences is identified as the normal range; and a saving step (S16) in which the learned information processing model with the identified normal range and parameters is saved as a slope anomaly detection model. The learning step and the saving step are performed for each of the multiple areas created by dividing the slope, thereby generating the slope anomaly detection model for each area.

Description

Technical Field

[0001] The present invention relates to a method for generating a slope abnormality detection model for detecting slope abnormalities, a slope monitoring system using the same, a slope monitoring device used therein, a slope monitoring server, and a computer-readable program.

Background Art

[0002] Patent Document 1 discloses a method for monitoring a collapsing slope. The monitoring method of Patent Document 1 measures the actual variation value of the collapsing slope, takes a photograph of the collapsing slope, and extracts the photographed images before and after that time when the actually measured variation value is an abnormal value. Patent Document 2 discloses a slope monitoring system. The monitoring system of Patent Document 2 includes an immovable column and a movable column connected by a wire, which are installed on a slope, an inclination angle sensor for detecting the inclination of the lower column part, and a hinge part for connecting the lower end of the upper column and the upper end of the lower column part. Patent Document 3 discloses a slope monitoring device. The slope monitoring device of Patent Document 3 is composed of a display meter installed on a slope for visualizing an indication value, a CCD camera for observing the indication value, and an image processing system for processing an image transferred from the CCD camera.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Patent Document 2

Patent Document 3

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, the technologies disclosed in Patent Documents 1, 2, and 3 have several problems: they require the installation of measuring instruments or markers on the slope, resulting in high installation costs; they cannot be applied to steep slopes where installation is difficult; and if the measuring instruments or markers are dislodged by wind and rain, slope monitoring becomes impossible. In particular, the method in Patent Document 1 is susceptible to measurement errors due to vegetation swaying caused by wind and rain or the behavior of wild animals; and it always requires expert knowledge to determine the criteria for judgment based on the output of the measuring instruments and to visually confirm the images.

[0005] Therefore, the present invention aims to provide a method for generating a slope anomaly detection model that enables high-precision detection of slope anomalies with a simple hardware configuration, a slope monitoring system using the same, a slope monitoring device used therein, a slope monitoring server, and a computer-readable program. [Means for solving the problem]

[0006] A method for generating a slope anomaly detection model according to the present invention includes a preparation step of preparing an information processing model that transforms an input image and outputs a transformed image; a learning step of inputting a monitoring image of a slope under normal conditions as training data to the information processing model, performing a learning process to identify transformation parameters that suppress the difference between the input image and the output image of the information processing model, and identifying the range of possible differences as the normal range; and a saving step of saving the trained information processing model with the identified normal range and parameters as a slope anomaly detection model, wherein the learning step and the saving step are performed for each of a plurality of areas formed by dividing the slope, thereby generating the slope anomaly detection model for each area.

[0007] In any method for generating a slope anomaly detection model according to the present invention, the training data for each area may include at least 12 months' worth of monitoring images for each area.

[0008] In any method for generating a slope anomaly detection model according to the present invention, the information processing model may be an autoencoder that includes an encoder for reducing the dimensionality of the feature vector of the input image and a decoder for outputting a reconstructed image recovered from the dimensionality-reduced feature vector as the output image.

[0009] A slope management system according to the present invention is a slope management system that uses slope anomaly detection models generated for each area by any of the slope anomaly detection model generation methods according to the present invention, and is characterized by comprising: a model management unit that manages the slope anomaly detection models for each area; an image acquisition unit that acquires monitoring images of each area of ​​the slope; a determination unit that inputs the monitoring images of each area to the slope anomaly detection model for each area and determines whether or not the differences of all areas fall within the normal range; and a notification unit that issues an alarm if there is an area where the differences do not fall within the normal range.

[0010] A slope management system according to the present invention may include a weather data acquisition unit that acquires weather data published by a weather server via a communication network, and a normal range adjustment unit that adjusts the size of the normal range according to the acquired weather data.

[0011] In any of the slope management systems according to the present invention, the acquisition of the monitoring image and the determination may be performed at a frequency of at least once every 20 seconds.

[0012] In any slope management system according to the present invention, the system may include a storage unit that stores a history of the monitoring images acquired by the image acquisition unit, and an update unit that updates the slope anomaly detection model by retraining the information processing model using the history of the monitoring images.

[0013] In any slope management system according to the present invention, the update may be performed at a frequency of at least once every 12 months.

[0014] A slope monitoring device according to the present invention is a slope monitoring device used in any slope management system according to the present invention, characterized by comprising the model management unit, the image acquisition unit, and the discrimination unit.

[0015] A slope monitoring server according to the present invention is a slope monitoring server used in any slope monitoring system according to the present invention, and is characterized by comprising the notification unit.

[0016] A computer-readable program according to the present invention is characterized by causing a computer to function as any slope monitoring device according to the present invention.

[0017] A computer-readable program according to the present invention is characterized by causing a computer to function as any slope monitoring server according to the present invention. [Effects of the Invention]

[0018] One method for generating a slope anomaly detection model according to the present invention involves performing a training process on an information processing model using monitoring images of a normal slope as training data for each area of ​​the slope, thereby generating a slope anomaly detection model for each area. The slope anomaly detection models generated in this way for each area are capable of reliably detecting slope anomalies accompanied by displacement in at least one area, and have the effect of being less susceptible to the influence of environmental changes other than slope anomalies.

[0019] In this context, "slope anomalies" refer to slope failures that involve the displacement of at least a portion of the slope, and include landslides, cliff collapses, mudslides, rockfalls, and debris flows. "Environmental changes other than slope anomalies" include changes in the position of the sun depending on the season or time of day, phases of the moon, changes in weather, seasonal changes in plants (plant growth, wilting, leaf fall), and the presence of wild animals in photographs.

[0020] One slope monitoring system according to the present invention detects the presence or absence of slope abnormalities based on whether the difference data when the monitoring image of the slope is input into the slope abnormality detection model is within the normal range. Therefore, if a monitoring camera that captures the slope in the field of view as hardware and a computer that runs the slope abnormality detection model are prepared, there is no need to prepare measuring instruments or signs, nor to install them on the slope.

[0021] According to one slope monitoring device, slope monitoring server, and computer-readable program according to the present invention, one slope monitoring system according to the present invention can be realized. As a result, there is an effect that it is possible to realize a method for generating a slope abnormality detection model that enables detection of slope abnormalities with high accuracy with a simple hardware configuration, a slope monitoring system using the same, a slope monitoring device used therein, a slope monitoring server, and a computer-readable program.

Brief Description of the Drawings

[0022] [Figure 1] FIG. 1 is a schematic configuration diagram of a slope monitoring system according to an embodiment. [Figure 2] FIG. 2 is a configuration diagram of a slope monitoring system when the number of slopes is "1". [Figure 3] FIG. 3 is a diagram for explaining a plurality of areas formed by dividing one slope. [Figure 4] FIG. 4 is an explanatory diagram for explaining the operation schedule of the slope monitoring system. [Figure 5] FIG. 5 is a flowchart of the generation process of the slope abnormality detection model. [Figure 6] FIG. 6 is a flowchart of the monitoring process by the slope monitoring device. [Figure 7] FIG. 7 is a flowchart of the monitoring process by the slope monitoring server. [Figure 8] FIG. 8 is a flowchart of the adjustment process by the slope monitoring server. [Figure 9]Figure 9 is a schematic diagram illustrating the effects of the slope monitoring system according to the embodiment. [Modes for carrying out the invention]

[0023] 1. Outline configuration of the slope monitoring system Hereinafter, an embodiment of the slope monitoring system according to the present invention will be described. Figure 1 is a schematic diagram of the slope monitoring system according to the embodiment. The slope monitoring system 1 shown in Figure 1 is a slope management system using a slope anomaly detection model, and comprises a plurality of monitoring cameras 11,11,11,... that individually photograph different slopes 4-1, 4-2, 4-3,..., a plurality of slope monitoring devices 10,10,10,... that are individually connected to the plurality of monitoring cameras 11,11,11,..., a communication network 2 such as the Internet, a slope monitoring server 100 located in a wide-area monitoring center, and a weather server 3 managed by the national or local government. Of these, the plurality of slope monitoring devices 10,10,10,..., the slope monitoring server 100, and the weather server 3 are connected to the communication network 2. Each of the slopes 4-1, 4-2, 4-3,... that are the target of monitoring by this slope monitoring system 1 is, for example, one of the following (1) to (3). (1) A slope where the surface of the ground, such as rocks and strata, is exposed. (2) A slope covered with grass and trees, with the bare ground visible. (3) Slopes on which landslide prevention measures have been implemented.

[0024] 2. Configuration of the slope monitoring system To simplify the explanation, the configuration of the slope monitoring system 1 will be described in detail below, assuming that the number of slopes to be monitored is "1". Figure 2 is a diagram of the configuration of the slope monitoring system when the number of slopes is "1". As shown in Figure 2, the slope monitoring system 1 includes a monitoring camera 11, a slope monitoring device 10, a slope monitoring server 100, a network 2, and a weather server 3.

[0025] 2-1. Surveillance cameras The surveillance camera 11 is a video camera capable of recording video or time-lapse photography, and can repeatedly generate 4K color surveillance images at a frame rate of, for example, 30fps. The sensitivity of the surveillance camera 11 is set to be able to handle not only daytime brightness but also nighttime brightness. However, the frame rate of the surveillance camera 11 is not limited to 30fps; it is sufficient if it can generate surveillance images at a frequency of at least one frame every 20 seconds. Also, if the field of view of the surveillance camera 11 cannot cover the entire slope to be monitored, the entire slope can be covered by the combined field of view of multiple surveillance cameras 11, 11 (two in Figure 3), for example, as shown in Figure 3. In this embodiment, the slope monitoring system 1 individually determines the presence or absence of abnormalities in each of the multiple areas (50 in this case) A1 to A50 that are created by dividing a single slope into multiple areas. The size of each of the areas A1 to A50 on the slope is approximately several meters wide by several meters long, and the size and shape are common among the areas A1 to A50 on the slope. Incidentally, since there is some distortion in the slope as seen in the actual monitoring image, the size and shape are not necessarily common among the areas in the monitoring image that correspond to each area on the slope.

[0026] 2-2. Slope monitoring device The configuration of the slope monitoring device 10 will now be described with reference to Figure 2. As shown in Figure 2, the slope monitoring device 10 includes a control unit 12 such as a processor, a storage unit 13 such as memory, and a communication unit 14 such as a communication interface, and these components are housed inside a sealed enclosure called a monitoring box. The control unit 12 functions as an image acquisition unit 121, a model management unit 122, and a discrimination unit 123. The image acquisition unit 121 performs processes such as acquiring surveillance images of the entire slope generated by the surveillance camera 11, and acquiring surveillance images of each area of ​​the slope from the surveillance images. For example, the image acquisition unit 121 can acquire surveillance images of each area by dividing the surveillance image of the entire slope into a predetermined division pattern. The processes of acquiring surveillance images of the slope and acquiring surveillance images of each area are performed at a predetermined frequency (here, once every 20 seconds). The model management unit 122 executes a process to manage the area-specific slope anomaly detection models M1 to M50 stored in the memory unit 13.

[0027] The discrimination unit 123 individually inputs the monitoring images from areas A1 to A50 to the slope anomaly detection models M1 to M50 for each area, and acquires the difference data (here, the mean absolute error MAEj) between the input image and the output image of the slope anomaly detection model Mj for each area. The discrimination unit 123 also performs a process to determine whether each of the difference data MAE1 to MAE50 for areas A1 to A50 falls within the normal range (here, whether it falls below the upper limit Tj of the normal range). If there is even one area where the difference data MAEj does not fall below the upper limit Tj, the discrimination unit 123 transmits a message indicating that an anomaly has occurred and the area number of the area in question to the slope monitoring server 100 via the communication unit 14 and network 2. The processing of the discrimination unit 123 is performed at a predetermined frequency (here, once every 20 seconds).

[0028] The memory unit 13 stores the slope monitoring program 131, slope anomaly detection models M1 to M50 for each area, upper limit values ​​T1 to T50 for the normal range for each area, and so on.

[0029] The slope monitoring program 131 is a program that causes the control unit 12 to function as a slope monitoring device.

[0030] The slope anomaly detection model Mj is an information processing model used to detect the presence or absence of anomalies in the j-th area Aj. Details of the slope anomaly detection model Mj will be described later.

[0031] The upper limit of the normal range Tj is the upper limit of the normal range of MAEj, which is the difference data between the input and output images of the j-th slope anomaly detection model Mj. Details of the upper limit of the normal range Tj will be described later.

[0032] The communication unit 14, under the direction of the control unit 12, sequentially transmits data such as monitoring images of the entire slope, monitoring images for each area, and differential data MAEj for each area to the slope monitoring server 100 via the communication network 2. This data is transmitted at a predetermined frequency (in this case, once every 20 seconds).

[0033] 2-3. Slope monitoring server The slope monitoring server 100 will now be described with reference to Figure 2. As shown in Figure 2, the slope monitoring server 100 includes a control unit 15 such as a processor, a storage unit 16 such as memory, and a communication unit 17 such as a communication interface.

[0034] The control unit 15 functions as a weather data acquisition unit 151, a notification unit 153, a normal range adjustment unit 152, and a learning / update unit 155. The weather data acquisition unit 151 promptly acquires the latest weather data that the weather server 3 publishes and updates at a predetermined frequency via the communication network 2 and the communication unit 17. The weather data acquired from the weather server 3 is, for example, precipitation data for each 1km mesh, which the weather server 3 publishes and updates every 10 minutes. Precipitation is the amount of rainfall per hour expressed in water depth (mm). The frequency of acquisition of cumulative rainfall by the weather data acquisition unit 151 is set to, for example, the same frequency as the precipitation update frequency by the weather server 3. The weather data acquisition unit 151 then performs a process to calculate the cumulative rainfall on the monitored slope based on the precipitation amount for each 1km mesh acquired from the weather server 3. Cumulative rainfall is the total amount of rain from the start of the rain to the present, and if there is no rain for a predetermined period (for example, 6 hours or 12 hours), this value is reset to zero. Furthermore, if the slope does not fit into a single mesh but spans multiple meshes (i.e., the slope belongs to multiple meshes), the weather data acquisition unit 151 calculates the cumulative rainfall of the slope by weighted average of the cumulative rainfall of the multiple meshes. The weight of each mesh in this weighted average is set to be larger for meshes that occupy a larger proportion of the slope.

[0035] When the notification unit 153 receives a slope monitoring image from the slope monitoring device 10 via the communication network 2 and the communication unit 17, it executes a process to display the monitoring image in real time on a display (not shown). Furthermore, when the notification unit 153 receives area-specific differential data MAEj (j=1~50), it executes a process to overlay the differential data MAEj for each area Aj as an anomaly severity value onto the corresponding area on the display. For example, the notification unit 153 highlights the outline of an area with a more prominent color the higher the differential data MAEj (anomaly severity value). Additionally, when the notification unit 153 receives notification from the slope monitoring device 10 via the communication network 2 and the communication unit 17 that an anomaly has occurred and the area number of the relevant area, it executes a process to issue an alarm, such as highlighting the relevant area on the display with the most prominent color, and also executes a process to send an emergency email to a pre-registered email address. The alarm can be issued not only by displaying it on the display, but also by sounding it from a speaker (not shown).

[0036] Furthermore, even if an alarm is issued indicating an abnormality in a certain area Aj, if, for example, an operator or expert confirms that there is actually no abnormality in that area Aj, identification data (correction data) indicating that area Aj was normal can be input to the slope monitoring server 100. This identification data (correction data) is stored in the storage unit 16 together with the differential data MAEj of area Aj.

[0037] The normal range adjustment unit 152 performs a process to adjust the upper limit values ​​T1 to T50 of the normal range for each area according to the cumulative rainfall on the slope determined by the weather data acquisition unit 151. Specifically, the normal range adjustment unit 152 performs a process to lower the upper limit values ​​TD1 to D50 of the normal range as the cumulative rainfall on the slope increases (details will be described later). In this way, by setting the upper limit values ​​TD1 to D50 of the normal range lower as the cumulative rainfall on the slope increases, the range considered normal becomes narrower, thus reducing the possibility of overlooking slight slope abnormalities that may occur when the cumulative rainfall on the slope is high.

[0038] The learning and updating unit 155 performs a preparation procedure (storing in the storage unit 16) to prepare an information processing model 162 that transforms an input image (=input image) and outputs a transformed image (=output image), and a learning process that inputs multiple monitoring images of area Aj of the slope under normal conditions as training data into the information processing model 162 and identifies the parameters of the transformation so that the difference data MAEj between the input image and the output image of the information processing model 162 is suppressed. Through this learning process, the range that the difference data MAEj can take (the maximum value of the difference data MAEj) is identified as the upper limit Tj of the normal range of the difference data MAEj. As a result, a trained information processing model 162 with the upper limit Tj of the normal range and the parameters identified is generated. The learning and updating unit 155 then performs a saving procedure that transmits the trained information processing model 162 with the identified parameters to the slope monitoring device 10 as the slope anomaly detection model Mj for area Aj, and also transmits the upper limit Tj of the normal range to the slope monitoring device 10. As a result, the slope anomaly detection model Mj and the upper limit value Tj of the normal range for area Aj are stored in the memory unit 13 of the slope monitoring device 10, with the area number j associated with them. The learning and updating unit 155 generates and stores the slope anomaly detection model Mj and the upper limit value Tj of the normal range for each area by executing the above learning and saving procedures for each area of ​​the slope. With the slope anomaly detection models M1 to M50 and the upper limit values ​​T1 to T50 of the normal range for each area generated and stored in this way, it is possible to reliably detect slope anomalies accompanied by displacement in at least one area, and moreover, it is less susceptible to the influence of environmental changes other than slope anomalies. In this context, "slope anomaly" refers to a slope failure accompanied by displacement of at least a portion of the slope, and includes landslides, cliff collapses, mudslides, rockfalls, and debris flows. Furthermore, "environmental changes other than slope anomalies" include changes in the position of the sun depending on the season or time of day, phases of the moon, changes in weather, seasonal changes in plants (plant growth, wilting, leaf fall), and the presence of wild animals in photographs.

[0039] Furthermore, the monitoring images used as training data are, for example, monitoring images acquired by the monitoring camera 11 from the slope over at least the past 12 months. Here, we assume that no slope anomalies occurred for at least 12 months from the start of monitoring image acquisition, and that the slope monitoring system 1 is put into full operation 12 months after the start of monitoring image acquisition (see Figure 4 described later).

[0040] Furthermore, the learning and updating unit 155 performs a process (update process) to update the area-specific slope anomaly detection models M1 to M50 by retraining the information processing model 162 at an appropriate timing after the start of full operation of the slope monitoring system 1 (see Figure 4 described later). This update process uses the history of monitoring images for each area 163 (described later) and the history of differential data for each area 164 (described later) that have been accumulated in the storage unit 16 up to that point. Such an update process is basically the same as the slope anomaly detection model generation process at the start of operation, differing only in the sampling period of the training data. By repeating such an update process during full operation, the accuracy of slope anomaly detection can be gradually improved while continuing full operation. It is desirable that this update process be performed at least once every 12 months.

[0041] Returning to Figure 2, the memory unit 16 stores the slope monitoring program 161, the information processing model 162, the history of monitoring images 163, the history of differential data 164, and so on. The slope monitoring program 161 stored in the memory unit 16 is a program that causes the control unit 15 to function as a slope monitoring server 100. The information processing model 162 stored in the memory unit 16 is an information processing model such as a neural network model (details of the types of information processing models will be described later). The area-specific monitoring image history 163 stored in the memory unit 16 is the history of area-specific monitoring images captured by the image acquisition unit 121. The history 164 of the area-specific differential data stored in the memory unit 16 is the history of each of the area-specific differential data MAE1 to MAE50 acquired by the discrimination unit 123. In addition, the memory unit 16 stores identification data indicating whether or not the area Aj differential data MAEj at each point in time was within the normal range, and identification data (the correction data mentioned above) indicating whether or not an abnormality actually occurred in area Aj at each point in time.

[0042] 3. Operation Schedule The following describes the operation schedule of the slope monitoring system 1. Figure 4 is an explanatory diagram illustrating the operation schedule of the slope monitoring system. As shown in Figure 4, the operation schedule includes step S1, which starts acquiring monitoring images; step S2, which starts trial operation; step S3, which starts full operation; and step S4, which updates the slope anomaly detection model. Each step will be described below.

[0043] Step S1: The slope monitoring device 10 starts acquiring monitoring images. The frequency of acquiring monitoring images is, for example, one frame every 20 seconds. Thereafter, the history 163 of monitoring images for each area is accumulated in the storage unit 16 (Figure 1) of the slope monitoring server 100.

[0044] Step S2: Three months after the execution of Step S1, the slope monitoring server 100 performs a learning process for the area-specific information processing model 162 based on the normal monitoring images (for three months) for each area stored in the memory unit 16. This generates the area-specific slope anomaly detection models M1 to M50, which are stored in the memory unit 13 of the slope monitoring device 10. Subsequently, a trial operation of the slope monitoring system 1 using the area-specific slope anomaly detection models M1 to M50 is started.

[0045] Step S3: Twelve months after the execution of Step S1, the slope monitoring server 100 performs a learning process for the area-specific information processing model 162 based on the normal monitoring images (for 12 months) for each area stored in the memory unit 16. The determination of whether the monitoring images for each area are normal monitoring images is made based on the history of differential data 164 for each area. As a result, area-specific slope anomaly detection models M1 to M50 are generated and overwritten and saved in the memory unit 13 of the slope monitoring device 10. Subsequently, the slope monitoring system 1 is put into full operation using the latest slope anomaly detection models M1 to M50.

[0046] Step S4: Twelve months after the execution of Step S3, the slope monitoring server 100 performs a learning process (update process) of the area-specific information processing model 162 based on the normal monitoring images (24 months' worth) for each area stored in the memory unit 16. The determination of whether the monitoring images for each area are normal monitoring images is performed based on the identification data (correction data mentioned above) stored in the memory unit 16. Therefore, the area-specific slope anomaly detection models M1 to M50 are updated (overwritten) based on the normal monitoring images for each area. Subsequently, the slope monitoring system 1 continues to operate using the updated slope anomaly detection models M1 to M50. This Step S4 is repeated, for example, once every 12 months.

[0047] According to the above operational schedule (Figure 4), trial operation will begin three months after the start of monitoring image acquisition, and full operation will begin twelve months after the start of monitoring image acquisition. In trial operation, slope anomaly detection models M1 to M50, which were generated based on monitoring images from the past three months under normal conditions, will be used, while in full operation, slope anomaly detection models M1 to M50, which were generated based on monitoring images from the past twelve months under normal conditions, will be used. Therefore, trial operation, which has lower monitoring accuracy (accuracy in calculating anomaly values ​​for each area, and accuracy in issuing alarms), will begin nine months earlier than full operation. Such trial operation is more susceptible to seasonal changes than full operation, so it is possible that anomaly values ​​may be calculated to be higher than they actually are depending on the season, or that seasonal changes may be misinterpreted as slope anomalies, leading to false alarms. However, if trial operation begins three months after the installation of the surveillance camera 11 and the slope monitoring device 10, the operator or expert can use the output of the notification unit 153 (abnormality value for each area, whether or not an alarm was issued) as a basis for making decisions regarding slope monitoring during the nine months until full operation begins. Therefore, the usefulness of trial operation is considered to be considerable.

[0048] Furthermore, in the slope monitoring system 1 of this embodiment, the slope anomaly detection models M1 to M50 are updated once every 12 months after the start of operation. As the number of normal monitoring images stored in the memory unit 16 increases over time, the sampling period for the training data is extended each time this update is repeated, and as a result, the monitoring accuracy in operation is improved.

[0049] 4. Types of Information Processing Models The following describes the types of information processing models 162. For example, an autoencoder can be used as the information processing model 162. An autoencoder is a neural network model that includes an encoder (encoder) that reduces the dimensionality of the feature vector of the input image, and a decoder (decoder) that outputs a reconstructed image, which is reconstructed from the feature vector after dimensionality reduction, as the output image. In the training process of this autoencoder, the encoder and decoder parameters that suppress the difference data MAE between the input image and the output image are identified, and the upper limit T of the normal range of the difference data MAE is identified. The trained autoencoder with the upper limit T of the normal range and the identified parameters is then saved as the slope anomaly detection model M. In this embodiment, since it is necessary to generate slope anomaly detection models M1 to M50 for each area, the input of monitoring images to the autoencoder, training of the autoencoder, identification of the upper limit value T, and identification of parameters are performed for each area.

[0050] 5. Generation process for slope anomaly detection model The following describes the process for generating the slope anomaly detection model. Figure 5 is a flowchart of the slope anomaly detection model generation process. The slope anomaly detection model generation process includes, for example, the following steps S11 to S18. These steps S11 to S18 are mainly performed by the learning and updating unit 155.

[0051] Step S11: The learning / update unit 155 sets the area number j to its initial value "1". Step S12: The learning and updating unit 155 extracts multiple monitoring images of area j under normal conditions based on the history of monitoring images 163 for each area and the history of differential data 164 for each area stored in the memory unit 16. Incidentally, at the start of trial operation (Figure 4, step S2), monitoring images of area Aj for the past 3 months are extracted, and at the start of full operation (Figure 4, step S3), monitoring images of area Aj for the past 12 months are extracted.

[0052] Step S13: The learning and updating unit 155 starts the learning process of the information processing model Mj using the multiple monitoring images of area Aj extracted in step S12 as training data. Step S14: The learning and updating unit 155 identifies parameters for the information processing model Mj such that the difference data MAEj between the input image and the output image of the information processing model Mj is suppressed. Step S15: The learning and updating unit 155 identifies the maximum value of the difference data MAEj of multiple monitoring images in area Aj as the upper limit Tj of the normal range of the difference data MAEj. This completes the learning of the information processing model Mj.

[0053] Step S16: The learning and updating unit 155 stores the learned information processing model Mj, with its identified upper limit Tj of the normal range and parameters, in the storage unit 13 of the slope monitoring device 10 as the slope anomaly detection model Mj for area Aj. The learning and updating unit 155 also stores the identified upper limit Tj as the upper limit Tj of the normal range for area Aj in the storage unit 13 of the slope monitoring device 10. Step S17: The learning / update unit 155 determines whether the area number j has reached the maximum value "50". If it has not reached the maximum value, it proceeds to step S18; if it has reached the maximum value, it terminates the flow. Step S18: The learning / update unit 155 increments the area number j and then proceeds to step S12.

[0054] 6. Monitoring and processing by slope monitoring device The monitoring process by the slope monitoring device 10 will be described below. Figure 6 is a flowchart of the monitoring process by the slope monitoring device. This monitoring process includes, for example, the following steps S21 to S28. These steps S21 to S28 are performed by the image acquisition unit 121, the model management unit 122, or the discrimination unit 123 of the slope monitoring device 10.

[0055] Step S21: The image acquisition unit 121 sets area number j to "1". Step S22: The image acquisition unit 121 acquires the surveillance image of area Aj from the surveillance image acquired by the surveillance camera 11. Step S23: The model management unit 122 inputs the monitoring image of area Aj into the slope anomaly detection model Mj stored in the memory unit 13. Step S24: The discrimination unit 123 calculates the difference data MAEj between the input image and the output image of the slope anomaly detection model Mj.

[0056] Step S25: The discrimination unit 123 determines whether the calculated difference data MAEj exceeds the upper limit Tj stored in the storage unit 13. If it exceeds the limit, the unit proceeds to step S26; otherwise, the unit proceeds to step S27. Step S26: The discrimination unit 123 notifies the server that an abnormality has occurred in area Aj, and then proceeds to step S27.

[0057] Step S27: The image acquisition unit 121 determines whether the area number j has reached the maximum value "50". If it has not reached the maximum value, it proceeds to step S28. If it has reached the maximum value, the flow ends. Step S28: The image acquisition unit 121 increments the area number j and then proceeds to step S22. In the monitoring process described above, the processing for multiple areas A1 to A50 was executed sequentially, but the processing for multiple areas A1 to A50 may also be executed in parallel.

[0058] 7. Monitoring process by the slope monitoring server The monitoring process performed by the slope monitoring server 100 will be described below. Figure 7 is a flowchart of the monitoring process performed by the slope monitoring server. This monitoring process includes, for example, the following steps S31 to S35. These steps S31 to S35 are mainly performed by the notification unit 153 of the slope monitoring server 100.

[0059] Step S31: The notification unit 153 sequentially captures the slope monitoring images and other data transmitted from the slope monitoring device 10. Step S32: The notification unit 153 displays the slope monitoring image received from the slope monitoring device 10 on the display in real time, and also overlays a grid indicating the boundaries of areas A1 to A50 onto the monitoring image. Step S33: The notification unit 153 determines whether or not a message indicating an abnormality and the area number of the relevant area have been transmitted from the slope monitoring device 10. If the message has been transmitted, the unit proceeds to step S34; otherwise, the unit proceeds to step S31.

[0060] Step S34: The notification unit 153 highlights the corresponding area on the display. Step S35: The notification unit 153 issues an alarm via the display and speaker, and sends an emergency email to a pre-registered email address, then proceeds to step S31.

[0061] 8. Normal range adjustment process The normal range adjustment process will be described below. Figure 8 is a flowchart of the adjustment process performed by the slope monitoring server. This adjustment process includes, for example, the following steps S41 to S48. These steps S41 to S48 are mainly performed by the weather data acquisition unit 151 and the normal range adjustment unit 152 of the slope monitoring server 100.

[0062] Step S41: The weather data acquisition unit 151 determines whether or not it is time for the weather server 3 to update the weather data (in this case, every 10 minutes). If it is time, it proceeds to step S42; otherwise, it waits. Step S42: The weather data acquisition unit 151 acquires the latest weather data (in this case, precipitation per 1km mesh) published by the weather server 3. Step S43: The weather data acquisition unit 151 calculates the cumulative rainfall of the monitored slope based on the precipitation amount for each 1km mesh acquired. However, if the slope belongs to multiple meshes, the cumulative rainfall of the slope is calculated as a weighted average of the cumulative rainfall of those multiple meshes, and the weight of each mesh is set to be larger for meshes that the slope occupies a larger proportion of the mesh.

[0063] Step S44: The normal range adjustment unit 152 determines whether the cumulative rainfall on the slope falls below the first threshold U1. If it does (S44YES), it proceeds to step S45 as "low risk," otherwise it proceeds to step S46. Step S45: The normal range adjustment unit 152 sets the upper limits T1 to T50 of the normal range for the difference data in areas A1 to A50 to their original values, and returns to step S41. Here, "original values" refers to the upper limits T1 to T50 identified in the most recent learning or update process. Step S46: The normal range adjustment unit 152 determines whether the cumulative rainfall on the slope falls below a second threshold U2 which is greater than the first threshold U1. If it falls below the second threshold (S46YES), it proceeds to step S47 as "medium risk," and if not (S46NO), it proceeds to step S48 as "high risk."

[0064] Step S47: The normal range adjustment unit 152 lowers the upper limits T1 to T50 of the normal range for the differential data in areas A1 to A50 to 0.9 times their original values, and returns to step S41. This increases the likelihood of detecting an anomaly (the likelihood of an alarm being triggered). Step S48: The normal range adjustment unit 152 lowers the upper limits T1 to T50 of the normal range for the differential data in areas A1 to A50 to 0.8 times their original values, and returns to step S41. This further increases the likelihood of detecting an anomaly (the likelihood of an alarm being triggered).

[0065] 9. Effects of the Embodiment As described above, in the slope anomaly detection model generation method according to this embodiment, the learning process of the information processing model 162, which uses monitoring images of the slope under normal conditions as training data, is performed for each area of ​​the slope to generate slope anomaly detection models M1 to M50 for each area. The slope anomaly detection models M1 to M50 for each area generated in this way can reliably detect slope anomalies accompanied by displacement in at least one area, and are less susceptible to the influence of environmental changes other than slope anomalies. Moreover, since the training data for each area includes monitoring images under normal conditions for at least 12 months for each area, the influence of environmental changes that may occur within 12 months can be reliably eliminated.

[0066] Specifically, during the 12-month sampling period for training data prior to full operation, various environmental changes occur, such as changes in the position of the sun, phases of the moon, weather changes, seasonal changes in plants, and the behavior of wild animals. Therefore, the influence of these environmental changes can be reliably eliminated. For example, if we assume that a tree like the one shown in Figure 9(A) grows in a certain area of ​​the slope to be monitored, even if changes occur such as changes in the amount of leaves due to tree growth (Figure 9(B)), changes in time of day (Figure 9(C)), swaying of branches due to wind and rain (Figure 9(D)), or the presence of wild animals in the image (Figure 9(E)), the slope monitoring system 1 of this embodiment will not issue a false alarm. On the other hand, if a change in the position of the tree occurs (Figure 9(F)), the slope monitoring system 1 of this embodiment will detect the abnormality and immediately issue an alarm.

[0067] Furthermore, the slope monitoring system 1 according to this embodiment detects whether or not there is an abnormality in area Aj by checking whether the difference data MAEj obtained when monitoring images of area Aj of the slope are input to the slope abnormality detection model Mj is below the upper limit Tj of the normal range. Therefore, as hardware, if a monitoring camera 11 that has the slope in its field of view and a computer that runs the slope abnormality detection model Mj are prepared, there is no need to set up measuring instruments or markers on the slope. Therefore, according to the slope monitoring system 1 of this embodiment, it is possible to detect slope abnormalities with high accuracy using a simple hardware configuration.

[0068] 10. Variations 10-1. Variations of Information Processing Models As the information processing model 162 described above, for example, a Generative Adversarial Network (GAN) can be used. A Generative Adversarial Network is a neural network model that includes a generator that outputs a fake image generated from an input image as an output image, and a discriminator that determines whether the output image is real or fake. In the training process of this Generative Adversarial Network, the parameters of the generator and discriminator are converged by repeatedly adjusting the parameters of the generator to increase the probability of misclassification and adjusting the parameters of the discriminator to decrease the probability of misclassification. As a result, the parameters of the generator are identified, and the upper limit of the normal range of the difference data between the input image and the output image of the generator is identified. The trained generator with the upper limit of the normal range T and the identified parameters is then saved as the slope anomaly detection model M. In this embodiment, since it is necessary to generate slope anomaly detection models M1 to M50 for each area, the input of monitoring images to the generative adversarial network, training of the generative adversarial network, identification of the upper limit T, and determination of parameters are performed for each area.

[0069] 10-2. Variations of Functional Division The division of functions in the above-described embodiments or modifications is not limited to those described above. Some of the functions of the slope monitoring device 10 may be mounted on the slope monitoring server 100, and some or all of the functions of the slope monitoring server 100 may be mounted on the slope monitoring device 10. For example, in the embodiments or modifications described above, a computer belonging to the slope monitoring system 1 (the control unit 15 of the slope monitoring server 100) performed the process of generating the slope anomaly detection model. However, this process of generating the slope anomaly detection model may be performed by a computer other than the slope monitoring system 1. However, it is preferable that the slope anomaly detection models M1 to M50 be mounted on the slope monitoring device 10 side, as in the embodiments or modified examples described above. This is because it is thought that the delay time from when an anomaly occurs until the anomaly is detected (= the time required to transfer the monitoring image) can be reduced.

[0070] 10-3. Variations of Numbers In the above-described embodiment or modified version of the slope monitoring system 1, the number of slope areas was set to 50, but a number other than 50 may be used. However, regardless of the number, it is desirable that the area on the slope be divided into several meters x several meters in size. Furthermore, in the slope monitoring system 1 of the above-described embodiment or modified example, the timing of the start of trial operation was set to 3 months after the start of acquisition of monitoring images, the timing of the start of full operation was set to 9 months after the start of trial operation, and the timing of the first update was set to 12 months after the start of full operation. However, these timings can be changed as appropriate. Furthermore, in the above-described embodiment or modified version of the slope monitoring system 1, the update processing frequency was set to once every 12 months. However, if it is necessary to maximize the accuracy of monitoring, the update processing frequency may be increased to more than once every 12 months. Conversely, if it is not necessary to maximize the accuracy of monitoring, the update processing frequency may be decreased to less than once every 12 months. Furthermore, in the normal range adjustment process described above (Figure 8), the upper limit T was reduced to 0.9 times when the risk level was "medium" and to 0.8 times when the risk level was "high". However, these multipliers can be changed as appropriate. Furthermore, in the normal range adjustment process described above (Figure 8), the risk level was set to three stages: "low," "medium," and "high," but it may also be set to two or four or more stages. Furthermore, in the embodiments or modifications described above, the mean absolute error (MAE) was used as the difference data between the input image and the output image, but another index that reflects the difference between the input image and the output image may be used. Furthermore, in the slope monitoring system 1 of the above-described embodiment or modified version, the weather data acquired from the weather server 3 was precipitation data for each 1km mesh, which is published and updated every 10 minutes. However, other weather data capable of determining the cumulative rainfall on the slope may also be used. Furthermore, in the slope monitoring system 1 of the above-described embodiment or modification, cumulative rainfall on the slope was used as meteorological data to reflect the size of the normal range (size of the upper limit T1 to T50), but other meteorological data indicating the possibility of landslides occurring on the slope may also be used.

[0071] 10-4. Others The present invention is not limited to the embodiments or modifications described above, and can be materialized by appropriately modifying the components without departing from the spirit of the invention. Furthermore, various inventions can be formed by appropriately combining the multiple components disclosed in each embodiment. [Explanation of symbols]

[0072] 1. Slope monitoring system 10. Slope monitoring device 100 Slope Monitoring Server 11 Surveillance cameras 12 Control Unit 121 Image acquisition unit 122 Model Management Department 123 Discrimination part 13 Storage section 131 Slope Monitoring Program 14 Communications Department 15 Control Unit 151 Weather Data Acquisition Unit 152 Normal range adjustment unit 153 Notification Department 155 Learning and Update Section 16 Memory section 161 Slope Monitoring Program 162 Information Processing Models 163 Surveillance Image History 164 History of differential data 17 Communication section such as communication interface 2. Communication Network 3 Weather Server 4-1~4-3 Slope Areas A1-A50 T1~T50 Upper limit of the normal range M1-M50 Slope Anomaly Detection Model

Claims

1. Preparation steps for preparing an information processing model that transforms an input image and outputs the transformed image, A learning procedure that inputs a normal slope monitoring image as training data into the information processing model, performs a learning process to identify the transformation parameters that suppress the difference between the input and output images of the information processing model, and identifies the range of possible differences as the normal range. The procedure includes saving the trained information processing model, in which the normal range and parameters have been identified, as a slope anomaly detection model. By performing the learning procedure and the saving procedure for each of the multiple areas created by dividing the slope, the slope anomaly detection model is generated for each area. A method for generating a slope anomaly detection model characterized by the above.

2. In the method for generating a slope anomaly detection model according to claim 1, The aforementioned training data for each area includes at least 12 months' worth of surveillance images for each area. A method for generating a slope anomaly detection model characterized by the above.

3. In the method for generating a slope anomaly detection model according to claim 1, The aforementioned information processing model is This autoencoder includes an encoder that reduces the dimensionality of the feature vector of the input image, and a decoder that outputs a reconstructed image, which is reconstructed from the dimensionality-reduced feature vector, as the output image. A method for generating a slope anomaly detection model characterized by the above.

4. A slope management system using a slope anomaly detection model generated for each area by the method for generating a slope anomaly detection model described in claim 1, A model management unit manages the aforementioned slope anomaly detection models for each area, An image acquisition unit that acquires monitoring images of each area of ​​the slope, A determination unit inputs the monitoring images of each area into the slope anomaly detection model for each area and determines whether the differences in all areas fall within the normal range. A notification unit that issues an alarm if there is an area where the difference does not fall within the normal range, A slope monitoring system characterized by comprising the following features.

5. In the slope monitoring system according to claim 4, A weather data acquisition unit that takes in weather data published by a weather server via a communication network, A normal range adjustment unit adjusts the size of the normal range according to the acquired weather data. A slope monitoring system characterized by comprising the following features.

6. In the slope monitoring system according to claim 4, The acquisition of the monitoring image and the determination are performed at a frequency of at least once every 20 seconds. A slope monitoring system characterized by the following features.

7. In the slope monitoring system according to claim 4, A storage unit that stores the history of the monitoring images acquired by the image acquisition unit, An update unit updates the slope anomaly detection model by retraining the information processing model using the history of the aforementioned monitoring images. A slope monitoring system characterized by comprising the following features.

8. In the slope monitoring system according to claim 7, The aforementioned updates are performed at least once every 12 months. A slope monitoring system characterized by the following features.

9. A slope monitoring device used in the slope monitoring system described in claim 4, The model management unit, the image acquisition unit, and the discrimination unit are all included. A slope monitoring device characterized by the following features.

10. A slope monitoring server used in the slope monitoring system described in claim 4, A slope monitoring server characterized by comprising the notification unit described above.

11. A computer-readable program, A computer-readable program characterized by causing a computer to function as a slope monitoring device according to claim 9.

12. A computer-readable program, A computer-readable program characterized by causing a computer to function as a slope monitoring server as described in claim 10.

Images