Slope collapse prediction system

Abstract

This system provides a slope collapse prediction system that uses artificial intelligence (AI) technology to predict the coordinate displacement at the next time point, determines whether the predicted displacement is normal or abnormal, and calculates the reliability of that determination. [Solution] The input data is the standardized signed "displacement difference distance" for the most recent four time points, which is the distance of the XYZ displacement difference between adjacent time points for the XYZ coordinate displacement, with a sign corresponding to the increase / decrease in that XYZ coordinate displacement distance. The input data is the standardized signed "displacement difference distance" for the next time point (ΔΔ * The first AI unit, which outputs (tn+1)), is constructed using a three-layer neural network. Next, the predicted value of the coordinate displacement at the next time step (Δ * The second AI unit, which takes the difference data for the most recent four time points between (tn+1) and the measured value (ΔΔ(tn+1)) as input data and outputs the label value of the winning unit, is constructed using a pre-trained supervised self-organizing map.

Description

[Technical Field]

[0001] The present invention relates to a slope collapse prediction system, and more specifically, to a slope collapse prediction system that can predict signs of a slope collapse using artificial intelligence (AI) technology, determine whether the predicted displacement is normal or abnormal, and calculate the reliability of that determination. [Background technology]

[0002] In recent years, disaster prediction using so-called AI has been implemented. For example, there are disaster prediction systems (learning models) that predict future measurement data from past measurement data at a certain location, and there are disaster prediction systems that improve the prediction accuracy of the system by using machine learning on a vast amount of measurement data acquired in the past (see, for example, Patent Document 1).

[0003] The above disaster prediction system is composed of a three-layer neural network that takes historical information on the ground surface shape and precipitation of past landslide-affected areas (target areas) as input data and outputs the risk of landslide occurrence. A trained model is generated by machine learning a vast amount of measurement data (training data) acquired in the past, and this trained model is used to predict the risk of landslide occurrence in the target area. The training data consists of pairs of input data and training data. The input data includes, for example, the slope angle of the land in the target area, the type of land surface, the area of ​​funnel-shaped land, precipitation, cumulative precipitation, and predicted precipitation, while the training data represents a level indicating the risk of disaster occurrence (see, for example,

[0077] -

[0079] of Patent Document 1).

[0004] On the other hand, a technique is known that uses a self-organizing map (hereinafter also called "SOM") to cluster road slopes and evaluate the soundness of newly inspected slopes by considering the characteristics of slopes that have been damaged in the past (see, for example, Non-Patent Literature 1). SOM is an AI technology proposed in 1982 by Teuvo Kohonen of Helsinki University of Technology (Finland) and positioned as a type of neural network. The SOM algorithm finds a unit with a coupling weight vector that best matches the input vector and defines this as the "winner unit". Next, only the "winner unit" and units located in its vicinity update their coupling weight vectors in a direction that approaches the input vector. By repeating this process for other input vectors, the coupling weight vectors of the competing layer are optimized to contain information that represents the given data set. In this way, the units of the competing layer that were initially unrelated become self-organized by the given data set. [Prior art documents] [Patent Documents]

[0005] [Patent Document 1] Patent No. 6806396 [Non-patent literature]

[0006] [Non-Patent Document 1] Ayaka Okado et al., "A Proposal for Road Slope Maintenance and Management Method Using Self-Organizing Maps (SOM)," Proceedings of the 66th Annual Scientific Conference of the Japan Society of Civil Engineers (FY2011). [Overview of the project] [Problems that the invention aims to solve]

[0007] In the disaster prediction system described in Patent Document 1 above, when input data (land slope angle of the target area, land surface type, funnel-shaped land area, precipitation, cumulative precipitation, and predicted precipitation) is input to the learning model, output data indicating the probability of classifying the area into one of four levels of danger is output.

[0008] Then, using a loss function, the output data is compared with the training data, and the weights in the trained model are adjusted, updating the parameters of the trained model. By repeating this process, a trained model is generated. In other words, the trained model acts as a classifier that classifies the input data into one of four risk levels, from level 1 to level 4. Level 1 indicates a state where there is no possibility of disaster occurring, or the possibility is extremely low. As the level increases, the probability of disaster occurring increases.

[0009] However, even if the risk level output by a trained model is, for example, Level 1 (meaning the probability of the slope not collapsing is extremely low), a collapse can still occur. If a collapse occurs, people who hesitated to evacuate because they trusted the Level 1 risk level will suffer the consequences of the landslide. In other words, there is a problem in that the risk level output by a trained model is only a rough guideline.

[0010] Therefore, the present invention has been made in view of the problems of the above-mentioned prior art, and its purpose is to provide a slope collapse prediction system that can predict signs of slope collapse using artificial intelligence (AI) technology, determine whether the predicted result is normal or abnormal displacement, and calculate the reliability of that determination. [Means for solving the problem]

[0011] To achieve the above objective, the slope collapse prediction system according to the present invention predicts the coordinate displacement (ΔΔΔ) at the next time step based on the time step XYZ coordinate displacement (ΔX(tn), ΔY(tn), ΔZ(tn)) of the target part (2) measured by the surveying instrument (1). *A slope collapse prediction system (100) comprising a computer (4) that predicts the coordinate displacement (Δ * A first artificial intelligence unit (41) predicts (tn+1) and is configured based on a trained supervised self-organizing map (SOM) that has been machine-learned using predetermined training data, and the coordinate displacement (ΔΔΔ) of the next time step. * The system is characterized by having a second artificial intelligence unit (42) capable of determining whether (tn+1)) is a normal displacement or an abnormal displacement and calculating the reliability of the said determination.

[0012] In the above configuration, the computer (4) has a first artificial intelligence unit (41) and a second artificial intelligence unit (42) with different functions, so the coordinate displacement (ΔΔΔ) at the next time step * In addition to predicting (tn+1), we also predict the coordinate displacement (ΔΔ * It becomes possible to determine whether (tn+1)) is a normal displacement (no slope collapse occurs) or an abnormal displacement (the slope is likely to collapse), and to calculate the reliability of that normal / abnormal displacement determination.

[0013] A second feature of the slope collapse prediction system according to the present invention is that the training data for machine learning of the second artificial intelligence unit (42) consists of "no slope collapse" training data, which consists of actual coordinate displacement values ​​obtained in the past when no slope collapse occurred, and "slope collapse-like" training data created based on the "no slope collapse" training data.

[0014] In the above configuration, input data with similar coordinate displacement characteristics will result in output units that are close to each other on the self-organizing map (SOM), while input data with different coordinate displacement characteristics will result in output units that are far apart on the self-organizing map (SOM). In other words, on the self-organizing map (SOM) after machine learning is complete, high-frequency output units updated with training data where "slope collapse does not occur" will be placed farther away from high-frequency output units updated with training data where "slope collapse occurs".

[0015] This allows us to determine whether the predicted coordinate displacement at the next time step is normal (closer to the frequently updated output unit with training data where "slope collapse does not occur") or abnormal (closer to the frequently updated output unit with training data where "slope collapse occurs") by determining the placement of the mapping destination (winner unit) on the "Self-Organizing Map (SOM) after machine learning is complete".

[0016] A third feature of the slope collapse prediction system according to the present invention is that the training data used for machine learning of the second artificial intelligence unit (42) includes training data consisting of label values ​​that are positively and negatively conjugate to each other, corresponding to the normal displacement / abnormal displacement.

[0017] In the above configuration, the label values ​​of output units in the self-organizing map (SOM) during machine learning are also updated according to the proximity (distance) of the coordinate displacement between the output unit and the training data. In other words, input data with coordinate displacement characteristics similar to the training data are placed near output units that are frequently updated by the training data in the trained supervised self-organizing map (SOM), and their label values ​​also approximate the label values ​​of output units that are frequently updated by the training data. This demonstrates that by finding the label value of the winning unit that best approximates the input data in the trained supervised self-organizing map (SOM), it is possible to calculate a quantitative degree of closeness (confidence) to how well the input data approximates the training data.

[0018] A fourth feature of the slope collapse prediction system according to the present invention is that the determination of normal displacement / abnormal displacement is made based on the label values ​​(L) of all output units on the trained supervised self-organizing map (SOM). j,k This is based on the average value (={Lmax+Lmin} / 2) of the maximum value (Lmax) and minimum value (Lmin) within ).

[0019] In the above configuration, the maximum label value (Lmax) of the output unit corresponds to the maximum frequency updated based on training data where "slope collapse does not occur." In other words, the maximum label value (Lmax) of the output unit indicates the output unit that best approximates the coordinate displacement characteristics of the training data where "slope collapse does not occur." On the other hand, the minimum label value (Lmin) of the output unit corresponds to the maximum frequency updated based on training data where "slope collapse occurs." In other words, the minimum label value (Lmin) of the output unit indicates the output unit that best approximates the coordinate displacement characteristics of the training data where "slope collapse occurs."

[0020] Thus, the label value of the output unit indicates the proximity to the characteristics of the coordinate displacement of the training data. That is, the label value of the output unit corresponds to the characteristics of the coordinate displacement of the training data. The label value of the training data consists of positive and negative conjugate label values, and it is considered that the average value of the label values (= zero) can be used as a threshold for determining normal displacement / abnormal displacement. Therefore, regarding the label value of the output unit in the trained supervised self-organizing map (SOM) as well, by using the average value of the maximum value (Lmax) and the minimum value (Lmin) among the label values of all output units ({Lmax + Lmin} / 2) as the threshold for determining normal displacement / abnormal displacement, it is considered that the determination of normal displacement / abnormal displacement can be made. In fact, by setting the threshold to the above average value, certain results were obtained for the determination of normal displacement / abnormal displacement.

[0021] The fifth feature of the slope collapse prediction system according to the present invention is that the reliability regarding the determination is based on the label value (L j*,k* ) of the winner unit when the input data ({δ(tn - 3), δ(tn - 2), δ(tn - 1), δ(tn)}) related to the predicted coordinate displacement for a plurality of time points is given to the "trained supervised self-organizing map (SOM)", as well as the maximum value (Lmax) or the minimum value (Lmin) and the average value ({Lmax + Lmin} / 2).

[0022] In the above configuration, the difference between the label value (L j*,k* ) of the winner unit and the average value (threshold L th ) relatively indicates how closely the input data approximates the training data. On the other hand, the difference between the maximum value (Lmax) or the minimum value (Lmin) of the label value of the output unit and the average value (threshold L th ) indicates the absolute value when the input data most closely approximates the training data. Therefore, the "absolute value of the difference between the label value (L j*,k* ) of the winner unit and the average value (threshold L th )" and the "difference between the maximum value (Lmax) or the minimum value (Lmin) of the label value of the output unit and the average value (threshold L thThe ratio of the absolute value of the difference between the input data and the training data quantitatively indicates the degree of approximation of the input data to the training data.

[0023] The sixth feature of the slope collapse prediction system according to the present invention is that the input data related to the coordinate displacement ({δ(tn-3), δ(tn-2), δ(tn-1), δ(tn)}) consists of the most recent measured values ​​(ΔΔ(tn+1)) and the predicted values ​​(ΔΔ * It is composed of the difference data of (tn+1).

[0024] In the above configuration, the predicted coordinate displacement is evaluated based on the difference between the predicted value and the measured value, and as a time-series trend, making it possible to detect abnormal displacements.

[0025] The seventh feature of the slope collapse prediction system according to the present invention is the predicted coordinate displacement (ΔΔΔ) at the next time step. * (tn+1)) is a predicted value for the next time step of the signed "distance of displacement difference" (ΔΔ(tn), ΔΔ(tn)') of the XYZ displacement difference (ΔΔX(tn), ΔΔY(tn), ΔΔZ(tn)) per adjacent time interval for the XYZ coordinate displacement (ΔX(tn), ΔY(tn), ΔY(tn)) for each time step, wherein the sign is positive (+) if the distance of the XYZ coordinate displacement (|Δ(tn)|) is greater than the distance of the previous XYZ coordinate displacement (|Δ(tn-1)|), and negative (-) if the distance of the XYZ coordinate displacement (|Δ(tn)|) is less than the distance of the previous XYZ coordinate displacement (|Δ(tn-1)|).

[0026] In the above configuration, the coordinate displacement (ΔΔX(tn), ΔΔY(tn), ΔΔZ(tn)) and its sign (increase / decrease of the distance of the displacement difference relative to the distance of the immediately preceding displacement difference) are used to determine the coordinate displacement (ΔΔ * Because it predicts the coordinate displacement, it becomes possible to accurately detect the trend of coordinate displacement. Furthermore, when using the signed "displacement difference distance" (ΔΔ(tn)') which is standardized by the mean (μ) and standard deviation (σ), the amount of computation required for data processing is significantly reduced. [Effects of the Invention]

[0027] According to the slope collapse prediction system of the present invention, it is possible to predict signs of a slope collapse using artificial intelligence (AI) technology and calculate the reliability of the prediction results. [Brief explanation of the drawing]

[0028] [Figure 1] This is an explanatory diagram showing a slope collapse prediction system relating to one embodiment of the present invention. [Figure 2] This is an explanatory diagram showing the input and output of data for a slope collapse prediction system according to one embodiment of the present invention. [Figure 3] This is an explanatory diagram showing the structure of the first AI part according to the present invention. [Figure 4] This is an explanatory diagram showing the structure of the second AI unit according to the present invention (during operation). [Figure 5] This is an explanatory diagram showing an example of time-series data for displacement difference and signed "distance of displacement difference" calculated by the data processing device according to the present invention. [Figure 6] This is a flowchart illustrating data processing by the data processing device according to the present invention. [Figure 7] This is an explanatory diagram showing training data in which no collapse occurs among the training data for the trained supervised self-organizing map of the second AI unit according to the present invention. [Figure 8] This is an explanatory diagram showing collapsed training data within the training data for a trained supervised self-organizing map of the second AI unit according to the present invention. [Figure 9] This is a flowchart illustrating data processing during operation by the first AI unit and the second AI unit according to the present invention. [Figure 10] This is an explanatory diagram showing the structure of the second AI unit according to the present invention (during machine learning). [Modes for carrying out the invention]

[0029] Embodiments of the present invention will be described in detail below with reference to the attached drawings.

[0030] Figure 1 is an explanatory diagram showing a slope collapse prediction system 100 according to one embodiment of the present invention.

[0031] This slope collapse prediction system 100 uses artificial intelligence (hereinafter also referred to as "AI") to predict the coordinate changes (predicted values) of the slope 50 at the next time step that are likely to occur in the very near future, based on the trend (actual values) of coordinate changes of each part of the slope 50. The system is configured to allow the AI ​​to determine whether the predicted coordinate changes are normal or abnormal displacements and to evaluate the reliability of that determination. Here, "normal displacement" refers to a coordinate change in which a collapse is not expected to occur, and "abnormal displacement" refers to a coordinate change in which a collapse is expected to occur. As will be described in more detail later, the AI ​​consists of a first AI unit 41 (Figure 2(b)) that predicts coordinate changes in the very near future (for example, the next time step) from actual coordinate change values ​​(for example, displacement data from the last four time steps), and a second AI unit 42 (Figure 2(b)) that determines whether the predicted coordinate change at the next time step is a normal or abnormal displacement and calculates the reliability of that determination.

[0032] Furthermore, if the reliability calculated by the second AI unit 42 (Figure 2(b)) exceeds a preset threshold (for example, 0.5 = 50%), the system is configured to alert people in the vicinity via the alarm device 5. In this embodiment, the object being measured is a slope 50 of the land, but this slope collapse prediction system 100 can be applied not only to slopes 50 but also to surface displacement of civil engineering structures such as tunnels.

[0033] Furthermore, the term "coordinate change" or "coordinate displacement" here refers to the signed "distance of the displacement difference." Incidentally, "displacement difference" refers to the difference in displacement of the coordinates (X(t), Y(t), Z(t)) at times t and t-1 (ΔX(t)-ΔX(t-1), ΔY(t)-ΔY(t-1), ΔZ(t)-ΔZ(t-1)), and the "distance" of this displacement difference is the square root of the sum of the squares of each component of the displacement difference (=√{(ΔX(t)-ΔX(t-1))). 2+(ΔY(t)-ΔY(t-1)) 2 +(ΔZ(t)-ΔZ(t-1)) 2}) means. Also, "signed" means the distance of displacement (=√{ΔX(t) 2 +ΔY(t) 2 +ΔZ(t) 2 The sign "-" is used when the displacement is less than the distance of the previous displacement. For details, please refer to Figure 3 and see below.

[0034] The configuration of the slope collapse prediction system 100 includes a surveying instrument 1 having a distance measurement function using optical waves and an angle measurement function using an encoder, a plurality of reflective prisms 2 that serve as reference points or points to be measured, a data processing device 3 that calculates coordinate displacement data (ΔΔ(tn)) for the points to be measured (reflective prisms 2) measured by the surveying instrument 1, and coordinate displacement data (ΔΔ(tn)) for the very near future from the calculated coordinate displacement data (ΔΔ(tn)). * (tn+1)) is predicted, and the predicted coordinate displacement data (ΔΔΔ * The system comprises a server 4 that determines whether the displacement (tn+1) is normal or abnormal and calculates the reliability of that determination, an alarm device 5 that emits a loud alarm sound or an abnormal voice message when the slope 50 is in a dangerous condition, and a wireless communication network 6 that enables bidirectional communication between the data processing device 3 and the server 4. Each component will be further described below.

[0035] The surveying instrument 1 irradiates the reflective prism 2 with measurement light (e.g., laser light) and receives the reflected light, thereby obtaining the straight-line distance (SD) from the surveying instrument 1 to the reflective prism 2, the vertical angle (V) which is the angle between the straight-line distance (SD) and the vertical direction, and the horizontal angle (H) which is the angle between the horizontal direction (e.g., true north direction) and the horizontal distance (HD) obtained when the straight-line distance (SD) is projected onto a horizontal plane. Therefore, the Z coordinate of the reflective prism 2 can be accurately measured from the straight-line distance (SD) and the vertical angle (V), and the (X, Y) coordinate of the reflective prism 2 can be accurately measured from the horizontal distance (HD) and the horizontal angle (H). In this way, the surveying instrument 1 can accurately measure the three-dimensional (X, Y, Z) coordinate of the reflective prism 2 by irradiating the reflective prism 2 with measurement light (e.g., laser light) and receiving the reflected light.

[0036] Furthermore, the surveying instrument 1 has a two-way wireless communication function (for example, Wi-Fi® or Bluetooth®) and is configured to wirelessly transmit coordinate data of each part of the slope to the data processing device 3.

[0037] The reflective prism 2 can be a so-called zero-constant prism, for example, one that does not require correction for the measurement values ​​obtained from the surveying instrument 1 and the reflective prism 2. Other types of reflective prisms can also be used.

[0038] The data processing device 3 has a bidirectional wireless communication function (for example, Wi-Fi® or Bluetooth®) and is configured to receive coordinate data of the reflecting prism 2 measured by the surveying instrument 1 via wireless communication. The coordinate data received by the data processing device 3 is temporarily stored in the memory area of ​​the data processing device 3, and the data processing program performs processes such as converting binary data to XYZ coordinate values, displacement processing to calculate the displacement for each coordinate, displacement difference processing to calculate the difference for each displacement, and displacement difference distance calculation processing to calculate the distance of the displacement difference.

[0039] Furthermore, the data processing device 3 is configured to transmit the calculated coordinate displacement data to the server 4 using the wireless communication network 6, and to receive from the server 4 the result of determining whether the predicted coordinate displacement for the next time step is normal or abnormal, as well as the reliability of that determination.

[0040] Furthermore, the data processing device 3 is configured to send an email to the administrator if the reliability of the determination result of whether the predicted coordinate displacement is normal or abnormal exceeds a predetermined threshold.

[0041] The wireless communication network 6 is, for example, a wireless communication network provided by a mobile phone company, such as LTE, 4G, or 5G, and is a bidirectional telecommunications network that can connect to the Internet. Alternatively, a wired bidirectional telecommunications network using optical fiber may be used instead of or in combination with the wireless communication network 6.

[0042] Server 4 can use either its own server or a third-party server (e.g., a cloud server). Based on the coordinate displacement data of the slope 50 transmitted from the data processing device 3, Server 4 predicts the coordinate displacement at the next time step, determines whether the predicted coordinate displacement is normal or abnormal, and calculates the confidence level of that determination.

[0043] Figure 2 is an explanatory diagram showing the data input and output of a slope collapse prediction system 100 according to one embodiment of the present invention. Figure 2(a) shows the data input and output for the data processing device 3, and Figure 2(b) shows the data input and output for the server 4.

[0044] As shown in Figure 2(a), the data processing device 3 receives displacement data (X displacement, Y displacement, Z displacement) for the slope 50 from the surveying instrument 1. Based on this displacement data (X displacement, Y displacement, Z displacement), the data processing device 3 creates coordinate displacement data (signed "distance of displacement difference") and transmits the created coordinate displacement data to the server 4.

[0045] As shown in Figure 2(b), the coordinate displacement data (standardized, coded "distance of displacement difference") received from the data processing device 3 is used by the first AI unit 41 to predict the coordinate displacement for the next time step. The coordinate displacement for the next time step predicted by the first AI unit 41 is transmitted to the second AI unit 42, where the difference between the predicted value and the measured value is calculated. Based on this difference between the predicted value and the measured value, the second AI unit 42 determines whether the predicted coordinate displacement for the next time step is a normal or abnormal displacement, and calculates the confidence level of that determination.

[0046] Returning to Figure 2(a), the determination result of normal / abnormal displacement by the second AI unit 42, and the calculation of the reliability of the determination result, are transmitted to the data processing unit 3. Upon receiving the normal / abnormal displacement determination result and the reliability of the determination result, the data processing unit 3 will activate the alarm device 5 and issue an alarm if the determination result is abnormal displacement and the reliability is above a predetermined threshold (e.g., 0.5). At the same time, it will send an email (a message stating that there is a risk of collapse and that residents living near the slope 50 and the manager of the slope 50 should evacuate immediately) to the residents living nearby.

[0047] Figure 3 is an explanatory diagram showing the structure of the first AI section 41 according to the present invention. This first AI unit 41 has a three-layer neural network structure consisting of an input layer, an intermediate layer, and an output layer. The input layer takes the input data {ΔΔ(tn)'} of standardized signed "displacement difference distance" for the most recent time tn and assigns a predetermined weight W Δ Multiply by and bias b Δ The value obtained by adding {ΔΔ(tn)' × W} Δ +b Δ It consists of a unit that outputs} to the next intermediate layer.

[0048] Note that the weight W Δ The data size depends on the data size of the intermediate layer values ​​(hereinafter referred to as "intermediate values") M(tn). In this embodiment, the data size of the intermediate values ​​M(tn) corresponds to, for example, a 1 × 512 matrix. Therefore, the weight W Δ and bias bΔ The data size of both corresponds to a 1x512 matrix.

[0049] The hidden layer applies a predetermined weight W to the median value M(tn) at time tn, compared to the median value M(tn-1) at the immediately preceding time tn-1. M Multiply by and bias b M Adding {M(tn-1)×W} M +b M The intermediate layer uses its intermediate value M(tn) and the output from the previous input layer {ΔΔ(tn)' × W}. Δ +b Δ It consists of units that output the result of adding}. Weight W M The data size of depends on the data size of the median value M(tn), and is equivalent to a 512x512 matrix. Similarly, bias b M The data size corresponds to a 1x512 matrix.

[0050] In other words, the hidden layer is configured to use a processing function f to regressively use the intermediate value M(tn-1) at the previous time tn-1 to calculate the intermediate value M(tn) at time tn as shown in Equation 1 below. Equation 1: M(tn)=f(W Δ ,b Δ ,ΔΔ(tn)',W M ,b M ,M(tn-1))

[0051] Furthermore, this first AI unit 41 takes the time-series data {ΔΔ(tn-3)', ΔΔ(tn-2)', ΔΔ(tn-1)', ΔΔ(tn)'} of standardized signed "displacement difference distances" for the four most recent time points (tn-3, tn-2, tn-1, tn) as input and predicts the standardized signed "displacement difference distance" ΔΔ for the next time point. *This outputs (tn+1). At first glance, it appears that equation 1 above uses only one input data ΔΔ(tn)' at time tn, but as shown in equations 2 to 4 below, by regressively using the intermediate values ​​M(tn-3), M(tn-2), and M(tn-1) at times tn-3, tn-2, and tn-1, the first AI unit 41 essentially takes as input the standardized signed "displacement difference distance" time series data {ΔΔ(tn-3)', ΔΔ(tn-2)', ΔΔ(tn-1)', ΔΔ(tn)'} for the four most recent time points (tn-3, tn-2, tn-1, tn). Equation 2: M(tn-1)=f(W Δ ,b Δ ,ΔΔ(tn-1)',W M ,b M ,M(tn-2)) Equation 3: M(tn-2)=f(W Δ ,b Δ ,ΔΔ(tn-2)',W M ,b M ,M(tn-3)) Equation 4: M(tn-3)=f(W Δ ,b Δ ,ΔΔ(tn-3)',W M ,b M ,0))

[0052] Furthermore, the processing function f in this embodiment assigns a predetermined weight W to the input data {ΔΔ(tn)'}. Δ Multiply by and bias b Δ The result of adding this to the previously given intermediate value M(tn-1), and a predetermined weight W M Multiply by and bias b M The processing function outputs the sum of the result obtained by adding the weights W. However, the processing function f is not limited to this. The processing function f will differ depending on the content of the processing. Furthermore, the weights W Δ ,W M Data size, and bias b Δ ,b M The data size will also vary depending on the data size of the intermediate value M(tn).

[0053] The output layer calculates the predicted value ΔΔΔ, which is the standardized signed "distance of displacement difference" for the next time step, by multiplying the intermediate value M(tn) at time tn output from the hidden layer by a weight W and adding a bias b. * The output is set to (tn+1). The data size of the weight W depends on the data size of the intermediate value M(tn). In this embodiment, the data size of the intermediate value M(tn) corresponds to a 1×512 matrix, so the data size of the weight W corresponds to a 512×1 matrix. Similarly, the data size of the bias b corresponds to a 1×1 matrix.

[0054] Weight W Δ ,W M ,W and bias b Δ ,b M For each value of b, the square of the difference between the predicted value and the measured value of the standardized signed "distance of displacement difference" (=[ΔΔ * (tn+1)-ΔΔ(tn+1)'] 2 Each component is optimized by machine learning to the value that minimizes the cost function C, which is defined by the sum of the values ​​of the components. As an optimization method, for example, the known backpropagation method can be applied. The training data for machine learning consists of previously acquired displacement data where collapse does not occur (standardized signed "displacement difference distance" data). This training data will be described later with reference to Figure 7.

[0055] Figure 4 is an explanatory diagram showing the structure (during operation) of the second AI unit 42 according to the present invention. This second AI unit 42 is composed of a self-organizing map (hereinafter also referred to as "SOM") whose input layer is time-series data {δ(tn-3), δ(tn-2), δ(tn-1), δ(tn)} of the difference between predicted and measured coordinate displacements for the most recent four time points, and whose output layer is grid points (j,k) on a plane. Each grid point (j,k) corresponds to a data section {V1} of the time-series data of the difference between predicted and measured values ​​in the input layer. j,k , V2 j,k , V3 j,k V4 j,k} and the label value L corresponding to normal / abnormal displacement. j,kIt has the following properties. Note that each grid point (j,k) is sometimes referred to as an output unit.

[0056] Incidentally, this self-organizing map is a pre-trained supervised self-organizing map. Therefore, the data section {V1 j,k , V2 j,k , V3 j,k V4 j,k} and label value L j,k This value is a pre-trained value learned from the training data. The learning process using this training data will be described later with reference to Figure 7. Also, although the self-organizing map in Figure 4 is shown as a two-dimensional SOM for explanatory purposes, it is preferable to use a torus-type SOM in practice, which connects the "top and bottom edges" and the "left and right edges" respectively, and can ensure a uniform learning amount (neighborhood NR) regardless of the position of the winning unit.

[0057] Here, we consider a mapping F that maps the input data of the input layer (time series data of the difference between predicted and measured values ​​for 4 time points {δ(tn-3), δ(tn-2), δ(tn-1), δ(tn)}) to the grid points (j,k).

[0058] The mapping F is a mapping that satisfies the following condition (1). (1) The distance d between the data portion Xp = {δ(tn-3), δ(tn-2), δ(tn-1), δ(tn)} of the input data and the grid point (j,k). j,k Calculate the distance d for each of the following intervals: |Xp-(j,k)| over 0≦j≦N and 0≦k≦N. j,k The minimum value d is when |Xp-(j,k)| is found. j*,k* The lattice point (j*,k*) in this case is the winning unit.

[0059] Next, (2) the label values ​​L of all lattice points (j,k) j,k The maximum value Lmax and minimum value Lmin are searched for, and a threshold L is used to determine whether the displacement is normal or abnormal. th Calculate this threshold L. th The calculation of this will be explained later, referring to Figure 7.

[0060] And (3) this threshold L th The system determines whether the displacement is normal or abnormal, and assigns the label value L to the winning unit. j*,k* The reliability of the determination is calculated using this method. The determination of normal / abnormal displacement, and the calculation of the reliability of that determination, will be described later with reference to Figure 7.

[0061] Incidentally, the training data used during learning consists of time series data {δ(tn-3), δ(tn-2), δ(tn-1), δ(tn)} of the difference between predicted and actual coordinate displacement values ​​for the most recent four time points, and labels {1: normal, -1: abnormal} (teaching data). If the time series data for the most recent four time points belongs to the vast amount of [displacement data where collapse does not occur] acquired in the past, the label of the training data will be {1}. On the other hand, if the time series data for the most recent four time points belongs to [displacement data that resembles a collapse], the label of the training data will be {-1}.

[0062] This "displacement data resembling collapse" can be created by statistically calculating outliers from a vast amount of "displacement data where collapse does not occur." Specifically, the mean μ and standard deviation σ are calculated for the standardized signed "displacement difference distance" (Figure 5) of the "displacement data where collapse does not occur," and values ​​below [μ-3σ] or above [μ+3σ] can be considered as "displacement data resembling collapse." Therefore, the self-organizing map shown in Figure 4 (the above mapping F) is machine-learned using training data consisting of a vast amount of "displacement data where collapse does not occur" and "displacement data resembling collapse."

[0063] Figure 5 is an explanatory diagram showing an example of time-series data for displacement difference and signed "distance of displacement difference" calculated by the data processing device 3 according to the present invention. For the sake of explanation, time-series data of displacement measured by the surveying instrument 1 is also shown. Furthermore, only the time-series data for displacement, displacement difference, and signed "distance of displacement difference" from time t1 to t20 are shown as examples.

[0064] Regarding displacement, it is measured for each of the XYZ coordinates. For example, the "X displacement" at time tn can be defined as {X coordinate X(tn) at time tn} - {initial value of X coordinate X(0)} ≡ ΔX(tn). Similarly, the "Y displacement" at time tn can be defined as {Y coordinate Y(tn) at time tn} - {initial value of Y coordinate Y(0)} ≡ ΔY(tn). Similarly, the "Z displacement" at time tn can be defined as {Z coordinate Z(tn) at time tn} - {initial value of Z coordinate Z(0)} ≡ ΔZ(tn).

[0065] Regarding the displacement difference, it is calculated for each of the XYZ displacements. For example, the "X displacement difference" at time tn can be calculated as {X displacement ΔX(tn) at time tn} - {X displacement ΔX(tn - 1) at time tn - 1} ≡ ΔΔX(tn). Similarly, the "Y displacement difference" at time tn can be calculated as {Y displacement ΔY(tn) at time tn} - {Y displacement ΔY(tn - 1) at time tn - 1} ≡ ΔΔY(tn). Similarly, the "Z displacement difference" at time tn can be calculated as {Z displacement ΔZ(tn) at time tn} - {Z displacement ΔZ(tn - 1) at time tn - 1} ≡ ΔΔZ(tn). Note that since n ≥ 2, the X displacement difference, Y displacement difference, and Z displacement difference at time t1 cannot be calculated. Therefore, the "distance of displacement difference" at time t1 described later cannot be calculated either.

[0066] Next, the "distance of displacement difference" at time tn is the square root of the sum of the squares of the X displacement difference, Y displacement difference, and Z displacement difference, that is, √{(X displacement difference ΔΔX(tn)) 2 +(Y displacement difference ΔΔY(tn)) 2 +(Z displacement difference ΔΔZ(tn)) 2} ≡ |ΔΔ(tn)| and can be calculated. Similarly, the "distance of displacement" at time tn is √{(X displacement ΔX(tn)) 2 +(Y displacement ΔY(tn)) 2 +(Z displacement ΔZ(tn)) 2} ≡ |Δ(tn)| and can be calculated.

[0067] Furthermore, the sign of the "distance of the displacement difference" at time tn can be the same as the sign of the difference in the "distance of displacement" between time tn and time tn-1. For example, the difference in the "distance of displacement" between time t2 and time t1 is 9.3 - 10.4 = -1.1 < 0. Therefore, the sign of the "distance of the displacement difference" at time t2 is negative. Consequently, the signed "distance of the displacement difference" ΔΔ(t2) at time t2 is -1.3.

[0068] On the other hand, the difference in the "displacement distance" between time t3 and time t2 is 12.4 - 9.3 = 3.1 > 0. Therefore, the sign of the "displacement difference distance" at time t3 is positive. In the case of a positive sign, nothing is added, so the signed "displacement difference distance" ΔΔ(t3) at time t3 remains 3.3. The sign of the "displacement difference distance" after time t3 can also be determined from the sign of the difference in the "displacement distance" in the same way as above.

[0069] Note that for this signed "distance of displacement difference," the mean μ=0 and the variance σ 2 It is desirable that the distribution be standardized by a mean μ and standard deviation σ so that it follows a standard normal distribution of = 1. This standardization will be described later, referring to step S5 in Figure 6.

[0070] Standardized, signed "displacement difference distance" time series data (hereinafter also referred to as "actual displacement time series data") is transmitted from the data processing device 3 to the server 4. Based on the actual displacement time series data, the server 4 predicts the coordinate displacement (standardized, signed "displacement difference distance") for the next time step, determines whether the predicted coordinate displacement is normal or abnormal, and calculates the confidence level of the determination result.

[0071] In summary, the data processing performed by the data processing device 3 is as follows:

[0072] Figure 6 is a flowchart illustrating data processing by the data processing device 3 according to the present invention. First, as step S1, displacement data (ΔX(tn), ΔY(tn), ΔZ(tn)) at a measurement target point for each predetermined time (times t1, t2, ···) is captured from the measuring instrument 1. Here, the X displacement ΔX(tn) at time tn is calculated by ΔX(tn)=X(tn)-X(t0). Similarly, the Y displacement ΔY(tn) and Z displacement ΔZ(tn) at time tn are calculated by ΔY(tn)=Y(tn)-Y(t0) and ΔZ(tn)=Z(tn)-Z(t0), respectively.

[0073] Next, as step S2, the X displacement difference, Y displacement difference, and Z displacement difference for each predetermined time at the measurement target point are calculated respectively. As described above, for the X displacement difference at time tn, it can be calculated by ΔΔX(tn)=ΔX(tn)-ΔX(tn - 1). Similarly, for the Y displacement difference at time tn, it can be calculated by ΔΔY(tn)=ΔY(tn)-ΔY(tn - 1). Similarly, for the Z displacement difference at time tn, it can be calculated by ΔΔZ(tn)=ΔZ(tn)-ΔZ(tn - 1).

[0074] Next, as step S3, the distance of the displacement difference for each predetermined time at the measurement target point is calculated. As described above, for the distance of the displacement difference at time tn, it can be calculated by √{ΔΔX(tn) 2 +ΔΔY(tn) 2 +ΔΔZ(tn) 2}.

[0075] Next, in step S4, the signed "displacement difference distance" at the point under measurement for each predetermined time interval is calculated. As described above, the signing is based on the "displacement distance" at time tn. That is, if the "displacement distance" at time tn is less than the "displacement distance" immediately before it, i.e., |Δ(tn)|-|Δ(tn-1)|<0, the sign attached to the "displacement difference distance" at time tn will be negative. On the other hand, if the "distance of change" at time tn is more than the "displacement distance" immediately before it, i.e., |Δ(tn)|-|Δ(tn-1)|≧0, the sign attached to the "displacement difference distance" at time tn will remain unchanged.

[0076] Next, in step S5, the standardized signed "displacement difference distance" at the point under measurement for each predetermined time interval is calculated. Here, "standardization" means that the signed "displacement difference distance" data ΔΔ(tn) is transformed into ΔΔ(tn)' = (ΔΔ(tn)-μ) / σ (where ΔΔ(tn)'=0 when σ=0) by using the mean μ and standard deviation σ of the data ΔΔ(tn) for multiple recent time intervals, for example, 64 time intervals. For the standardized data ΔΔ(tn)', the mean μ=0 and variance σ 2 It will follow a standard normal distribution with =1.

[0077] The mean μ and standard deviation σ at time tn (n≧3) are given as the mean μ and standard deviation σ from time t3 to time tn. However, at time t2, the standard deviation σ is zero, so ΔΔ(tn)'=0.

[0078] In step S6, time-series data of standardized, signed "displacement difference distance" at the point under measurement for each predetermined time interval is sent to server 4. The transmitted standardized, signed "displacement difference distance" data will be used to predict the coordinate displacement at the point under measurement for the next time interval.

[0079] Figure 7 is an explanatory diagram showing training data in which no collapse occurs among the training data for the trained supervised self-organizing map of the second AI unit 42 according to the present invention.

[0080] Predicted coordinate displacement at the next time step tn+1 (ΔΔ * (tn+1)) is calculated by the first AI unit 41 of server 4 (Figure 2(b)). The first AI unit 41 calculates the predicted value of the coordinate displacement at time tn+1 (ΔΔΔ * To output (tn+1), use the time series data of coordinate displacements from the four most recent time points (tn-3, tn-2, tn-1, tn) {ΔΔ(tn-3)', ΔΔ(tn-2)', ΔΔ(tn-1)', ΔΔ(tn)'} as input data. Coordinate displacement is a standardized signed "distance of displacement difference". For example, using the time series data of coordinate displacements from time t2 to time t5 {ΔΔ(t2)', ΔΔ(t3)', ΔΔ(t4)', ΔΔ(t5)'}, the standardized signed "distance of displacement difference" ΔΔ at time t6 is obtained. * We predict (t6). Note that standardized signed "displacement difference distance" data is available from time t2 onwards. Therefore, n must satisfy n-3≧2 (n≧5). This means that coordinate displacement can be predicted from time t5.

[0081] The second AI unit 42 predicts the standardized signed "distance of displacement difference" ΔΔ * The difference δ(tn+1) between (tn+1) and the measured value ΔΔ(tn+1)' is calculated. Then, during machine learning, the second AI unit 42 takes the time series data {δ(tn-3), δ(tn-2), δ(tn-1), δ(tn)} of the difference between the predicted and measured coordinate displacements for the most recent four time points (tn-3, tn-2, tn-1, tn) as input data, and the labels {1: normal, -1: abnormal} as training data, and calculates the winning unit (j*, k*) using a self-organizing map, and the data section {V1 j,k , V2 j,k , V3 j,k V4 j,k} and label value L j,k Update.

[0082] For example, when the input data is time-series data {0.7, -0.5, -0.7, 0.3} of the difference between predicted and measured coordinate displacement values ​​at times t6 to t9, and the training data is label {1}, the winning unit (j*,k*) is the grid point (0,0). Therefore, the data section {V1} for the output unit within the grid point (0,0) and its neighboring region NR is j,k , V2 j,k , V3 j,k V4 j,k} and label value L j,k The following will be updated. Similarly, when the time series data {-0.5, -0.7, 0.3, 0.8} of the difference between predicted and measured coordinate displacement values ​​at times t7 to t10 is used as input data and the label {1} ​​is used as training data, the winning unit (j*,k*) is the grid point (9,10). Therefore, the data section {V1} for the output unit within the grid point (9,10) and its neighboring region NR is obtained. j,k , V2 j,k , V3 j,k V4 j,k} and label value L j,k This will be updated.

[0083] By the way, machine learning using training data for self-organizing maps involves calculating the winner unit (j*,k*) and the data section {V1} for the winner unit and its neighboring region NR, which contains output units. j,k , V2 j,k , V3 j,k V4 j,k} and label value L j,k This is up to the point of updating the data. The determination of normal / abnormal displacement for the winning unit (j*,k*) and the calculation of the reliability of that determination will be performed during actual operation (measurement). Specifically, the label value L of the winning unit. j*,k* The threshold L th If the above conditions are met, the winning unit (j*,k*) is determined to have a normal displacement. Meanwhile, the label value L of the winning unit j*,k* The threshold L th If the value is less than the threshold L, the winning unit (j*,k*) is determined to have an abnormal displacement. thRegarding this, the label value L of all output units in the output layer j,k From this, obtain the maximum value Lmax and the minimum value Lmin, and the threshold L th = {maximum value Lmax + minimum value Lmin} / 2, where L is the threshold value. th It is possible to define this.

[0084] Here, assuming that the training data shown in Figure 7 is time-series data relating to coordinate displacement during operation, we will explain the determination of normal / abnormal displacement for the winning unit (j*,k*), using the label value L of all output units. j,k The maximum value Lmax is 0.978. On the other hand, the label value L of all output units j,k The minimum value Lmin is -0.889. The threshold L in this case is th This is calculated as 0.0445 (={0.978-0.889} / 2). Threshold L th If we set = 0.0445, the label value L will be the winning unit (j*,k*) at the grid point (0,0). j*,k* The threshold L is 0.978, and 0.978 ≥ threshold L th Since (=0.0445), the grid point (0,0) of the winning unit (j*,k*) will be determined to be a normal displacement.

[0085] On the other hand, when the input data is time-series data {0.5, 0.9, -2.1, 2.7} of the difference between predicted and measured coordinate displacement values ​​from time t15 to time t18, the winning unit (j*,k*) is the grid point (1,11). The label value L is set to the grid point (1,11) as the winning unit (j*,k*). j*,k* The threshold L is -0.889, and -0.889 < threshold L th Since (=0.0445), the lattice point (1,11) of the winning unit (j*,k*) will be determined to be an abnormal displacement.

[0086] Similarly, when the input data is time-series data {0.9, -2.1, 2.7, -2.5} of the difference between predicted and actual coordinate displacement values ​​from time t16 to time t19, the winning unit (j*,k*) is the grid point (3,9). The label value L is assigned to the grid point (3,9) as the winning unit (j*,k*).j*,k* The threshold L is 0.016, and 0.016 < threshold L th Since (=0.0445), the grid point (3,9) of the winning unit (j*,k*) will be determined to be an abnormal displacement.

[0087] Label value L of the winning unit j*,k* The reliability is calculated using the formulas in equation 5 or equation 6 below. The formula for calculating reliability differs depending on whether the winning unit (j*,k*) has a normal or abnormal displacement. (Assuming the winning unit (j*,k*) has normal displacement) Equation 5: Confidence = {(Label value of the winning unit L j*,k* )-(threshold L th )} / {(Maximum value Lmax of the output unit's label value)-(Threshold L th )} × 100[%] Note that the label value L of the winning unit j*,k* The threshold L th If it is equal to , the credibility is 0%. On the other hand, the label value L of the winning unit. j*,k* However, if it is equal to the maximum value Lmax of the output unit's label, the confidence level is 100%.

[0088] (If the winning unit (j*,k*) has an abnormal displacement) Equation 6: Credibility = {(Threshold L th )-(Label value L of the winning unit) j*,k* )} / {(threshold L th ) - (Minimum value of the output unit's label Lmin) × 100 [%] Note that the label value L of the winning unit j*,k* The threshold L th If it is equal to , the credibility is 0%. On the other hand, the label value L of the winning unit. j*,k* However, if it is equal to the minimum value Lmin of the output unit's label, the reliability is 100%.

[0089] For example, in Figure 7, when the input data is the time series data {-0.5, -0.7, 0.3, 0.8} of the difference between predicted and measured coordinate displacement values ​​from time t7 to time t10, the winning unit (j*,k*) is the grid point (9,10). Threshold L th If = 0.0445, the label value L of the winning unit. j*,k* Since the value is 0.216, the grid point (9,10), which is the winning unit (j*,k*), is in normal displacement. Therefore, equation 5 above is applied as the formula for calculating the reliability. The credibility is calculated as follows: Credibility = {0.216 - 0.0445} / {0.978 - 0.0445} × 100 = 18.4%.

[0090] Figure 8 is an explanatory diagram showing collapsed training data within the training data for the trained supervised self-organizing map of the second AI unit 42 according to the present invention.

[0091] In this collapse-like training data, instead of the measured value in the difference δ(tn) between the predicted and measured values ​​shown in Figure 7, random numbers outside the range of [μ-3σ, μ+3σ] when [displacement data where collapse does not occur] (for example, the standardized signed "distance between displacement steps" ΔΔ(tn)' in Figure 5) are represented by a normal distribution with mean μ and standard deviation σ are treated as [displacement data that resembles collapse]. Note that the predicted value ΔΔ of the coordinate displacement at the next time step is used. * For (tn+1), the value calculated by the same first AI unit 41 as in Figure 7 is used.

[0092] Furthermore, for this collapsed training data, the difference between the predicted and actual values ​​at times tn-3, tn-2, and tn-1, as shown in Figure 7, {δ(tn-3), δ(tn-2), δ(tn-1)} is used for the input data during training of the second AI unit 42. On the other hand, for the difference at time tn, the difference between the predicted value and [random numbers outside the ±3σ range] δ"(tn) is used. This is because this collapsed training data assumes a situation where no collapse occurs at times tn-3, tn-2, and tn-1, but the collapse occurs for the first time at time tn. Therefore, the label value as training data during machine learning is always -1.

[0093] Thus, the second AI unit 42 is trained using the training data shown in Figure 7 [where slope collapse does not occur] and the training data shown in Figure 8 [where slope collapse is considered to have occurred].

[0094] Figure 9 is a flowchart showing data processing during operation by the first AI unit 41 and the second AI unit 42 according to the present invention. The second AI unit 42 is pre-trained using the training data shown in Figure 7 [training data in which slope collapse does not occur] and the training data shown in Figure 8 [training data in which slope collapse is considered to have occurred].

[0095] In step SS1, the first AI unit 41 predicts the coordinate displacement ΔΔ for the next time step. * (tn+1) is calculated. As shown in Figure 7 or 8, the first AI unit 41 takes the time-series data of standardized signed "displacement difference distances" for the four most recent time points {ΔΔ(tn-3)', ΔΔ(tn-2)', ΔΔ(tn-1)', ΔΔ(tn)'} as input data and calculates the standardized signed "displacement difference distance" ΔΔ at the next time point. * Calculate (tn+1).

[0096] In step SS2, the second AI unit 42 predicts the coordinate displacement ΔΔ at the next time step. * The difference between (tn+1) and the measured value ΔΔ(tn+1) is δ(tn+1) (=ΔΔ * Calculate (tn+1)-ΔΔ(tn+1)).

[0097] In step SS3, the second AI unit 42 selects a winning unit (j*,k*) using time-series data {δ(tn-3), δ(tn-2), δ(tn-1), δ(tn)} of the difference between predicted and measured coordinate displacements for the most recent four time points as input data. The winning unit (j*,k*) is determined by the distance {d j,k In |0≦j≦N,0≦k≦N}, the distance is at its minimum value d j*,k* The grid points (j*,k*) that satisfy this condition are designated as the winning unit. Note that each grid point (j,k) corresponds to the data section {V1} of the above time series data. j,k , V2 j,k , V3 j,k V4 j,k} and label value L j,k It has.

[0098] In step SS4, the second AI unit 42 determines whether the winning unit (j*,k*) has normal or abnormal displacement. The determination of normal / abnormal displacement is based on the label value L of all output units. j,k We search for the minimum value Lmin and the maximum value Lmax, and set the average of the minimum value Lmin and the maximum value Lmax as the threshold L. th (={minimum value Lmin and maximum value Lmax} / 2). Next, the label value L of the winning unit. j,k The threshold L th If the above conditions are met, the winning unit (j*,k*) is determined to have normal displacement. Determining normal displacement means that a collapse will not occur with a predetermined probability. The magnitude of this probability is the label value L of the winning unit (j*,k*). j*,k* This is represented by the credibility of the winning unit (j*,k*). j*,k* The reliability of this will be discussed later in Step SS5.

[0099] On the other hand, the label value L of the winning unit (j*,k*) j*,k* The threshold L th If the value is less than the specified value, the winning unit (j*,k*) is determined to have an abnormal displacement. An abnormal displacement determination means that a collapse will occur with a predetermined probability. The magnitude of this probability is the label value L of the winning unit (j*,k*).j*,k* It is expressed by its credibility.

[0100] In step SS5, the label value L of the winning unit (j*,k*) j*,k* The reliability is calculated. The formula for calculating the reliability differs depending on the result of the determination of normal / abnormal displacement of the winning unit (j*,k*). If the winning unit (j*,k*) is determined to have normal displacement, the formula for calculating the reliability is {the label value of the winning unit L j*,k* -Threshold L th} / {Label value L of all output units} j,k Maximum value Lmax - threshold L th This becomes} × 100[%].

[0101] On the other hand, if the winning unit is determined to have abnormal displacement, the formula for calculating the credibility is {threshold L th - Label value L of the winning unit j*,k*} / {threshold L th - Label value L for all output units j,k The minimum value is Lmin × 100[%].

[0102] As described above, the slope collapse prediction system 100 of the present invention has been explained with reference to the drawings, but the present invention is not limited to the above. That is, various modifications, changes, and additions can be made without departing from the technical scope of the present invention. For example, the X displacement ΔX(tn), Y displacement ΔY(tn), and Z displacement ΔZ(tn) of the slope 50 are measured by the surveying instrument 1 and the reflecting prism 2, but it is also possible to use a receiver with a built-in antenna that can receive positioning signals from a satellite positioning system such as GNSS at a predetermined timing instead of the surveying instrument 1 and the reflecting prism 2. In this case, a coordinate transformation is required to convert the GNSS coordinate system to the local coordinate system. This coordinate transformation can be performed by the data processing device 3.

[0103] Furthermore, when the first AI unit 41 predicts the coordinate displacement for the next time step, it uses time-series data of signed "displacement difference distances" for the most recent four time steps as input data. However, the input data may also be standardized time-series data of signed "displacement difference distances" for five or more time steps or three or fewer time steps.

[0104] Similarly, when the second AI unit 42 calculates the winning unit, it uses time-series data of the difference between predicted and measured values ​​for the coordinate displacement over the most recent four time points as input data. However, the input data may also be time-series data of the difference between predicted and measured values ​​for the coordinate displacement over the most recent five time points or three time points or less.

[0105] Furthermore, for the input data used by the first AI unit 41 when predicting the coordinate displacement at the next time step, it is also possible to use the signed "distance of displacement difference" ΔΔ(tn) before standardization. [Explanation of Symbols]

[0106] 1 Surveying instrument 2. Reflecting prism (target section) 3. Data processing equipment (computer) 3a Monitor screen 4 servers 41 1st AI Department 42 2nd AI Department 5 Alarm device 6. Wireless communication network (two-way telecommunications network) 50 slopes 100 Slope Collapse Prediction System ΔX(tn) X-coordinate displacement ΔY(tn) Y-coordinate displacement ΔZ(tn) Z-coordinate displacement |Δ(tn)| Distance of coordinate displacement ΔΔX(tn) X displacement difference ΔΔY(tn) Y displacement difference ΔΔZ(tn) Z displacement difference |ΔΔ(tn)| Distance of displacement difference ΔΔ(tn) Signed "distance of displacement difference" ΔΔ(tn)' Standardized signed "distance of displacement difference" ΔΔ * (tn+1) Predicted value of the standardized signed "distance of displacement difference" for the next time step. δ(tn) is the difference between the predicted and measured values ​​of the standardized signed "displacement difference distance" for the next time step. W Δ Weights in the input layer b Δ Bias in the Input Layer W M Weights in the middle layer b M Bias in the Middle Class W input layer weights b. Bias in the input layer f is a function related to the regression processing of the hidden layer. (j,k) Grid points of the output unit on the SOM (j*,k*) Grid points of the winner unit on the SOM L j,k Output unit label value L j*,k* Winning unit label value Lmax: Maximum label value for all output units Lmin: Minimum label value for all output units L th Threshold for determining normal / abnormal displacement for the winning unit

Claims

1. Based on the time-dependent XYZ coordinate displacement (ΔX(tn), ΔY(tn), ΔZ(tn)) of the target unit (2) measured by the surveying instrument (1), the coordinate displacement (ΔΔ ＊ A slope collapse prediction system (100) equipped with a computer (4) that predicts (tn+1), The computer (4) is configured based on a trained three-layer neural network that has been machine-learned using predetermined training data, and calculates the coordinate displacement (ΔΔ) at the next time step. ＊ The first artificial intelligence unit (41) predicts (tn+1), It is constructed based on a trained supervised self-organizing map (SOM) that has been machine-learned using predetermined training data, and the coordinate displacement (ΔΔ) at the next time step. ＊ The system includes a second artificial intelligence unit (42) capable of determining whether (tn+1) is a normal or abnormal displacement and calculating the reliability of the determination. A slope collapse prediction system characterized by the following features.

2. In the slope collapse prediction system described in claim 1, The training data for the second artificial intelligence unit (42) during machine learning consists of "no slope collapse" training data, which is composed of actual coordinate displacement values ​​obtained in the past when no slope collapse occurred, and "slope collapse-like" training data, which is created based on the "no slope collapse" training data. A slope collapse prediction system characterized by the following features.

3. In the slope collapse prediction system described in claim 1, The training data for machine learning performed on the second artificial intelligence unit (42) includes training data consisting of mutually positive and negative conjugate label values ​​corresponding to the normal displacement / abnormal displacement. A slope collapse prediction system characterized by the following features.

4. In the slope collapse prediction system described in claim 1, The determination of normal / abnormal displacement is made by the label values ​​(L) of all output units on the trained supervised self-organizing map (SOM). j,k This is based on the average value (= {Lmax + Lmin} / 2) of the maximum value (Lmax) and minimum value (Lmin) within ). A slope collapse prediction system characterized by the following features.

5. In the slope collapse prediction system described in claim 4, The confidence level of the above determination is determined by the label value (L) of the winning unit when the input data relating to the predicted coordinate displacement for multiple time minutes ({δ(tn-3), δ(tn-2), δ(tn-1), δ(tn)}) is given to the "trained supervised self-organizing map (SOM)". j*,k* ), and calculated based on the maximum value (Lmax) or the minimum value (Lmin) and the average value (= {Lmax + Lmin} / 2) A slope collapse prediction system characterized by the following features.

6. In the slope collapse prediction system described in claim 5, The input data related to the coordinate displacement ({δ(tn-3), δ(tn-2), δ(tn-1), δ(tn)}) consists of the most recent measured values ​​(ΔΔ(tn+1)) and predicted values ​​(ΔΔ ＊ It is composed of the difference data of (tn+1). A slope collapse prediction system characterized by the following features.

7. In the slope collapse prediction system described in claim 1, The coordinate displacement (ΔΔ) at the next time step ＊ (tn+1)) is the predicted value for the next time step of the signed "distance of displacement difference" (ΔΔ(tn), ΔΔ(tn)') of the XYZ displacement difference (ΔΔX(tn), ΔΔY(tn), ΔΔZ(tn)) per adjacent time interval for the XYZ coordinate displacement (ΔX(tn), ΔY(tn), ΔY(tn)) for each time step, Regarding the sign, if the distance of the XYZ coordinate displacement (|Δ(tn)|) is greater than the distance of the previous XYZ coordinate displacement (|Δ(tn-1)|), it is considered positive (+), and if the distance of the XYZ coordinate displacement (|Δ(tn)|) is less than the distance of the previous XYZ coordinate displacement (|Δ(tn-1)|), it is considered negative (-). A slope collapse prediction system characterized by the following features.

Images