Information processing device, information processing method, and program

Abstract

To determine whether performance data showing past operation results is appropriate as learning data.SOLUTION: The present invention includes a simple physical model generation unit that generates a simplified physical model for determining whether performance data is appropriate to be used as learning data for training a learning model, based on the relationship between an error between an instruction value and a current value in the performance data, in which an instruction value, a current value, and an operation direction correspond to each other, a gradient which is the amount of change in the error, and the operation direction in the performance data, indicating past operation results of a shield excavator.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to an information processing device, an information processing method, and a program. [Background technology]

[0002] In the excavation of a shield machine, there is a technology that uses AI (machine learning) to predict the operation of the driving jack that determines the excavation direction. For example, Patent Document 1 discloses a technology that estimates the position coordinates of the force point by using an estimation model generated by machine learning learning data that associates directional data such as an indication value and a current value with the position coordinates of the force point to be excavated. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2019-143385 A Summary of the Invention [Problem to be solved by the invention]

[0004] However, not all performance data showing past operation results are necessarily suitable as learning data. If a large amount of inappropriate performance data is trained, a trained model with low estimation accuracy will result. Therefore, it is better to be able to distinguish between data suitable as learning data and data that is inappropriate among performance data. For this reason, a method for quantitatively determining whether performance data showing past operation results is suitable as learning data has been desired.

[0005] In view of the above-mentioned problems, an object of the present invention is to provide an information processing device, an information processing method, and a program capable of determining whether performance data indicating past operation performance is appropriate as learning data. [Means for solving the problem]

[0006] An information processing device according to one embodiment of the present invention includes a simplified physical model generation unit that generates a simplified physical model for determining whether it is appropriate to use the performance data, which indicates past operating performance of a shield excavator, as learning data for a learning model, based on the relationship between an error between an instruction value and a current value in the performance data, in which an instruction value, a current value and an operation direction are associated, a gradient which is the amount of change in the error, and the operation direction in the performance data.

[0007] An information processing method according to one aspect of the present invention is an information processing method performed by a computer, in which a simple physical model generation unit generates a simplified physical model for determining whether it is appropriate to use the actual data as learning data for training a learning model, based on the relationship between an instruction value and a current value in the actual data, in which an instruction value, a current value, and an operation direction correspond to each other, a gradient which is the amount of change in the error, and the operation direction in the actual data.

[0008] A program according to one embodiment of the present invention is a program that causes a computer to generate a simplified physical model for determining whether it is appropriate to use performance data indicating past operating performance of a shield tunneling machine, in which an instruction value, a current value, and an operating direction are associated, as learning data for a learning model, based on the relationship between an error between an instruction value and a current value in the performance data, which is the amount of change in the error, a gradient, and the operating direction in the performance data. Effect of the Invention

[0009] According to the present invention, it is possible to determine whether or not performance data indicating past operation performance is appropriate as learning data. [Brief description of the drawings]

[0010] [Figure 1A] FIG. 1 is a diagram for explaining a shield excavator 10 according to an embodiment. [Figure 1B]FIG. 1 is a diagram for explaining a shield excavator 10 according to an embodiment. [Diagram 2] 1 is a block diagram showing a configuration of an information processing device 30 according to an embodiment. [Diagram 3] FIG. 4 is a diagram showing an example of information stored in a performance data storage unit 34 according to the embodiment. [Figure 4] 11 is a diagram for explaining a process for generating a simplified physical model ML according to an embodiment. FIG. [Diagram 5] FIG. 2 is a diagram showing an example of a simplified physical model ML according to the embodiment. [Figure 6] FIG. 2 is a diagram showing an example of a simplified physical model ML according to the embodiment. [Figure 7] 4 is a flowchart showing a flow of processing performed by the information processing device 30 according to the embodiment. [Figure 8] 4 is a flowchart showing a flow of processing performed by the information processing device 30 according to the embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0012] <About the Shield Drilling Machine 10> First, a shield tunneling machine 10 according to an embodiment will be described with reference to Fig. 1 (Figs. 1A and 1B). Fig. 1 is a diagram for explaining the shield tunneling machine 10 according to an embodiment.

[0013] FIG. 1A shows a schematic diagram of a shield excavator 10 seen from the side. As shown in FIG. 1A, the shield excavator 10 excavates the ground while constructing the primary lining SG by assembling segment rings with an erector (not shown) at the rear of a cylindrical skin plate 11. The shield excavator 10 has a chamber 12 at the rear of an annular, face plate-type cutter 16 equipped with a cutter bit 15. A plurality of earth pressure gauges D are installed on the side walls of the chamber 12. The earth pressure gauge D measures the pressure (controlled earth pressure) of the mud in the chamber 12. An additive 14 is injected into the chamber 12 from a mud material injection pipe 13. The excavated soil piled up in the chamber 12 is mixed with the additive 14 by a mixing blade (not shown) and converted into mud. The screw conveyor 17 discharges the mud in the chamber 12 to a conveyor 18 through a soil discharge gate G. The conveyor 18 carries the mud discharged from the screw conveyor 17 out of the tunnel via a conveyor 19. A frame M supports the screw conveyor 17 and the conveyors 18 and 19.

[0014] FIG. 1B shows a conceptual diagram explaining a propulsion jack that propels the shield excavator 10. As shown in FIG. 1B, the shield excavator 10 is provided with a propulsion jack 20 (hereinafter, sometimes simply referred to as a "jack"), which will be described later. A plurality of propulsion jacks 20 are provided so as to surround the inner periphery of the skin plate 11. The propulsion jacks 20 are disposed between the skin plate 11 and the segment ring. The excavation direction and excavation speed of the shield excavator 10 are controlled by extending the propulsion jacks 20 by hydraulic control or the like. In the example shown in this figure, an example is shown in which 22 propulsion jacks 20 are provided in the shield excavator 10, but this number is not limited.

[0015] The position coordinates of the point of force that the shield excavator 10 applies to the excavation surface are set depending on which of the propulsion jacks 20 is driven (extended). The direction in which the shield excavator 10 propels is set by applying propulsion pressure to a position corresponding to the point of force that is set depending on the driving of the propulsion jack 20. In other words, the excavation direction is set by the jack pattern that indicates which of the propulsion jacks 20 is driven.

[0016] In excavation work using the shield excavator 10, the operator follows instructions from staff members or the like (or instructions prepared in advance) and operates the propulsion jacks 20 while monitoring data from various measuring instruments. The operator operates the propulsion jacks 20 by selecting whether or not to use each of the propulsion jacks 20, thereby controlling the excavation direction of the shield excavator 10.

[0017] Except for special circumstances, the operator operates the propulsion jack 20, the bending mechanism (mechanism for dividing the shield body into a front body and a rear body and bending it with the bending jack) if provided, and the copy cutter (a cutter bit used for correcting curved sections and meandering), etc., so that the current value determined based on data from various measuring instruments approaches the excavation instruction value (indicated value). Therefore, the error between the current value and the indicated value related to the directional control of the shield excavator 10 is an important indicator of the accuracy of the operation of the propulsion jack 20.

[0018] On the other hand, in a special situation, for example, in order to perform some kind of control different from directional control (such as ensuring clearance), the operator may operate the propulsion jack 20 in a direction that causes the current value to deviate from the command value.

[0019] When generating a learned model that estimates directional control, it is possible to use past operating records of the operator actually operating the propulsion jack 20 during excavation work as learning data to be trained by the learning model. However, as described above, the performance data indicating past operation results may contain a mixture of operations that have been performed in a direction that reduces the error between the current value and the specified value and operations that increase the error between the current value and the specified value. Depending on the relationship with the subject to be estimated by the trained model, the data may contain data that is inappropriate for learning. For example, actual data showing operations that increase the error between the indicated value and the current value in order to secure clearance is inappropriate as learning data for a learning model that estimates the excavation direction.Furthermore, it is inappropriate to have a learning model for estimating horizontal operations learn horizontal operations that occur as a result of jacking an operator to excavate vertically. If a large amount of inappropriate performance data is used for training, there is a high possibility that the estimation model will have low estimation accuracy. For this reason, a method for evaluating the suitability of performance data as training data when using the data has been demanded.

[0020] To address this issue, in this embodiment, a simplified physical model ML (see FIGS. 5 and 6) is generated that can calculate an index for evaluating the suitability of performance data as learning data. This makes it possible to determine whether the performance data to be used as learning data is suitable for use as learning data. For example, it becomes possible to understand to what extent the performance data to be used as learning data, among performance data of drillings performed at different sites or geologies, deviates in suitability as learning data compared to other data. In addition, a process for improving the quality of the performance data as learning data can be performed on the performance data according to the index calculated from the simplified physical model ML. As a process for improving the quality of the learning data, for example, there is a process for removing inappropriate data. By performing such a process, it is possible to generate high-quality learning data. By training high-quality learning data, it is possible to generate a trained model with high estimation accuracy. In this embodiment, the information processing device 30 generates the simplified physical model ML.

[0021] <Regarding the information processing device 30> Fig. 2 is a block diagram showing a configuration of an information processing device 30 according to an embodiment. As shown in Fig. 2, the information processing device 30 includes, for example, an operation data acquisition unit 31, a direction data acquisition unit 32, a simple physical model generation unit 33, a performance data storage unit 34, a simple physical model storage unit 35, a learning data storage unit 36, a learning data generation unit 37, and an output unit 38.

[0022] The operation data acquisition unit 31 acquires operation data related to operations as past operation records. The operation data acquisition unit 31 acquires data indicating the operation of the propulsion jack 20, for example, information indicating a jacking pattern, as the operation data. For example, the operation data acquisition unit 31 acquires information indicating the jacking pattern operated by the operator as the operation data. The operation data acquisition unit 31 may acquire information indicating the force point position corresponding to the jacking pattern as the operation data.

[0023] The direction data acquisition unit 32 acquires direction data indicating the excavation direction as a past operation record. The direction data acquisition unit 32 acquires, for example, a left / right jack stroke difference current value (one example of a current value) and a left / right jack stroke difference indication value (one example of an indication value) as the direction data. Any method may be adopted as a method for measuring the left / right jack stroke difference current value. For example, a measurement value measured by a measuring device such as a gyrocompass installed on the shield excavator 10 can be used as the left / right jack stroke difference current value. On the other hand, since the left / right jack stroke difference indication value is predetermined based on the excavation instruction, the left / right jack stroke difference indication value can be acquired from the excavation instruction.

[0024] The simple physical model generation unit 33 generates the simple physical model ML. The simple physical model generation unit 33 stores the created simple physical model ML in the simple physical model storage unit . The simplified physical model ML is a feature space including a first axis corresponding to the error and a second axis corresponding to the gradient (see FIG. 5). The error here is a value corresponding to the difference between the indicated value and the current value. The gradient is the amount of change in the error per unit time. In this feature space, the performance data is plotted according to the relationship between the error, gradient, and operation direction in the performance data (see FIG. 6). Based on the distribution state of the performance data plotted in the feature space, it is possible to determine whether the performance data is appropriate as learning data. The performance data indicates the past operation results of the shield tunneling machine 10, and is performance data in which an instruction value, a current value, and an operation direction are associated with each other.

[0025] First, the simple physical model generation unit 33 generates performance data using the operation data and the direction data. The simple physical model generation unit 33 acquires operation data from the operation data acquisition unit 31, and extracts data indicating time-series changes in the operation direction from the acquired operation data. The simple physical model generation unit 33 also acquires direction data from the direction data acquisition unit 32, and extracts data indicating time-series changes in the instruction value and the current value from the acquired direction data. The simple physical model generation unit 33 generates performance data by associating the instruction value, the current value, and the operation direction along a time series using the extracted data. The simple physical model generation unit 33 stores the generated performance data in the performance data storage unit 34.

[0026] FIG. 3 is a diagram showing an example of information stored in the performance data storage unit 34 according to the embodiment. As shown in FIG. 3, the performance data includes data corresponding to each of the items of, for example, time t, instruction value s(t), current value a(t), error d(t), gradient Δd(t), operation presence / absence, and operation direction. Time t indicates a time in the time series change of the performance data. Instruction value s(t) indicates instruction value s corresponding to time t. Current value a(t) indicates current value a corresponding to time t. Error d(t) indicates the difference between instruction value s(t) and current value a(t) at time t. In the example of this figure, d(t)=a(t)-s(t). The gradient indicates the amount of change per unit time of error d(t) at time t. The gradient may be at least an index showing the relationship between time and the amount of change in error. For example, the gradient may be calculated by using time-averaged error data. The operation presence / absence indicates whether or not a jack operation was performed at time t. Here, an example is shown in which "1" is set when a jack operation is performed, and "0" is set when a jack operation is not performed. The operation direction is shown as the operation direction when an operation is performed. Here, either the left or right operation direction in the horizontal direction based on the excavation direction is shown. The operation direction may also be either the up or down operation direction based on the excavation direction. Furthermore, the azimuth angle may be shown together with the operation direction.

[0027] Next, the simple physical model generation unit 33 generates a simple physical model ML using the performance data. A method for generating the simple physical model ML will be described with reference to Fig. 4 to Fig. 6. Fig. 4 is a diagram for explaining the process of generating the simple physical model ML. Fig. 5 and Fig. 6 are diagrams showing examples of the simple physical model ML.

[0028] Figure 4 shows that the direction of the appropriate jacking operation (the operation to correct the position of the jacking force point) is determined by the relationship between the error and the gradient. The vertical axis of Figure 4 represents the error, and the horizontal axis represents the time. The square marks represent the error at the time corresponding to the time on the horizontal axis. The arrow marks indicate the direction of the gradient (the gradient of the error). The line type of the arrow marks indicates what type of jacking operation is appropriate.

[0029] The dotted arrow in Figure 4 indicates that the appropriate jacking operation is one that turns the error positive (+). In other words, it shows a situation in which the jacking operation needs to be corrected in the direction that moves the error positive (+). For example, the arrow mark with the symbol L1 indicates that the error indicated by the square mark is negative and its gradient is also negative. In this case, if the jack operation is not corrected so that the error turns positive (+), the error will gradually increase in the negative direction. Therefore, it is necessary to correct the jack operation in the direction that moves the error to positive (+). Also, the arrow mark with the symbol L2 indicates that the error indicated by the square mark is negative and its gradient is zero. In this case, unless you correct the jack operation so that the error turns positive (+), the error will not decrease. Therefore, you need to correct the jack operation in the direction that moves the error to positive (+). Also, the arrow mark L4 indicates that the error indicated by the square mark is zero and its gradient is negative. In this case, if the jack operation is not corrected so that the error turns positive (+), the error will increase in the negative direction. Therefore, it is necessary to correct the jack operation in the direction that moves the error to positive (+).

[0030] Also, the arrow mark L32 indicates that the error indicated by the square mark is negative and its gradient is positive. In this case, the error gradually approaches zero, but the gradient is too gentle. In this case, unless the jack operation is corrected so that the error turns positive (+), it will take a long time for the error to become zero. For this reason, it is necessary to correct the jack operation in the direction that moves the error to positive (+).

[0031] Also, the arrow mark L73 indicates that the error indicated by the square mark is positive and its gradient is negative. In this case, the error gradually approaches zero, but the gradient is too steep. In this case, unless the jack operation is corrected so that the error turns positive (+), the error will pass zero and increase in the negative direction immediately. For this reason, it is necessary to correct the jack operation in the direction that moves the error to positive (+).

[0032] The solid arrow in Figure 4 indicates that the appropriate jacking operation is one that turns the error negative (-). In other words, it shows a situation in which the jacking operation needs to be corrected in the direction that moves the error negative (-). For example, the arrow mark L9 indicates that the error indicated by the square mark is positive and its gradient is also positive. In this case, if the jack operation is not corrected so that the error turns negative (-), the error will gradually increase in the positive direction. Therefore, it is necessary to correct the jack operation in the direction that moves the error to negative (-). Also, the arrow mark L8 indicates that the error indicated by the square mark is positive and its gradient is zero. In this case, unless you correct the jack operation so that the error turns negative (-), the error will not decrease. Therefore, you need to correct the jack operation in the direction that moves the error to negative (-). Also, the arrow mark of L6 indicates that the error indicated by the square mark is zero and its gradient is positive. In this case, if the jack operation is not corrected so that the error turns negative (-), the error will gradually increase in the positive direction. Therefore, it is necessary to correct the jack operation in the direction that moves the error to negative (-).

[0033] Also, the arrow mark L72 indicates that the error indicated by the square mark is positive and its gradient is negative. In this case, the error gradually approaches zero, but the gradient change is too gradual. In this case, unless the jack operation is corrected so that the error turns negative (-), it will take a long time for the error to become zero. For this reason, it is necessary to correct the jack operation in the direction that moves the error to negative (-).

[0034] Also, the arrow mark L33 indicates that the error indicated by the square mark is negative and its gradient is positive. In this case, the error gradually approaches zero, but the gradient change is too steep. In this case, unless the jack operation is corrected so that the error turns negative (-), the error will increase in the positive direction as soon as it passes zero. For this reason, it is necessary to correct the jack operation in the direction that moves the error to negative (-).

[0035] The dashed arrow in Figure 4 indicates that the relationship between the error and the slope is appropriate and there is no need to correct the jacking operation. For example, the arrow mark of code L31 indicates that the error indicated by the square mark is negative and its gradient is positive. In this case, the error gradually approaches zero. The gradient change is also appropriate. Therefore, there is no need to correct the jack operation. For example, the arrow mark of code L71 indicates that the error indicated by the square mark is positive and its gradient is negative. In this case, the error gradually approaches zero. The gradient change is also appropriate. Therefore, there is no need to correct the jack operation. For example, the arrow mark of the symbol L5 indicates that the error indicated by the square mark is zero and the gradient is also zero. In this case, the error remains zero. Therefore, there is no need to correct the jack operation.

[0036] In this way, it is possible to estimate the appropriate jack operation at a certain point in time based on the relationship between the error at that point in time and its gradient. By utilizing this property, the simplified physical model generation unit 33 generates a feature space that can indicate the appropriate jack operation direction as a region.

[0037] An example of the simplified physical model ML is shown in Fig. 5. The simplified physical model ML here is shown as a feature space in which the vertical axis (first axis) is the gradient and the horizontal axis (second axis) is the error.

[0038] In the feature space, the first region, which is painted with a polka dot pattern, is the region where the jack operation needs to be corrected in the direction in which the error moves to the plus (+) side. This first region corresponds to the dotted arrow mark in Figure 3. For example, region E1 corresponds to code L1 in FIG. 4 and is a region where the error is negative and the slope is also negative. Region E2 corresponds to code L2 in FIG. 4 and is a region where the error is negative and the slope is zero. Region E4 corresponds to code L4 in FIG. 4 and is a region where the error is zero and the slope is negative. Region E32 corresponds to code L32 in FIG. 4 and is a region where the error is negative and the slope is positive but too gentle. Region E73 corresponds to code L73 in FIG. 4 and is a region where the error is positive and the slope is negative but too steep.

[0039] In the feature space, the second region, which is painted with a horizontal line pattern, is the region where the jack operation needs to be corrected in the direction in which the error moves negative (-). This second region corresponds to the solid arrow mark in Figure 3. For example, region E9 corresponds to code L9 in FIG. 4 and is a region where the error is positive and the slope is positive. Region E8 corresponds to code L8 in FIG. 4 and is a region where the error is positive and the slope is zero. Region E6 corresponds to code L6 in FIG. 4 and is a region where the error is zero and the slope is positive. Region E72 corresponds to code L72 in FIG. 4 and is a region where the error is positive and the slope is negative but too gentle. Region E33 corresponds to code L33 in FIG. 4 and is a region where the error is negative and the slope is positive but too steep.

[0040] In the feature space, the third region of the white pattern is a region where the jack operation does not need to be corrected. This third region corresponds to the dashed arrow mark in Figure 4. For example, region E31 corresponds to code L31 in Fig. 4, and is a region where the error is negative and the gradient is positive with an appropriate slope. Region E71 corresponds to code L71 in Fig. 4, and is a region where the error is positive and the gradient is negative with an appropriate slope. Region E5 corresponds to code L5 in Fig. 4, and is a region where the error is zero and the gradient is also zero.

[0041] In this way, by generating a feature space with the error on the horizontal axis and the gradient on the vertical axis, it is possible to classify the appropriate jack operation direction at each point in time into regions. This makes it possible to visually express the appropriate jack operation direction in relation to the error and gradient in an easy-to-understand manner.

[0042] In addition, the areas surrounded by dashed lines drawn parallel to the horizontal and vertical axes in FIG. 5 (areas E2, E8, E4, and E6) are areas where both the error and gradient can be considered to be 0 (zero).

[0043] The 45-degree line shown in Fig. 5 is an ideal straight line that corresponds to symbols L31 and L71 in Fig. 3 and indicates an appropriate relationship between error and gradient. A region having a certain width centered on this ideal straight line (45-degree line) may be regarded as a region indicating an appropriate relationship between error and gradient. Based on this idea, the regions having a certain width centered on the 45-degree line are regarded as regions E31 and E71.

[0044] The 45-degree line in Figure 5 can be expressed as a linear function that passes through the origin and has a negative slope in the feature space. err ) may be determined based on the tendency of the performance data plotted in the feature space. Alternatively, the scale of the coordinate axes is converted so that the error and gradient in the performance data are on the same scale, and a straight line with a gradient of -1 is defined on the converted coordinate axes. Then, an inverse transformation is performed to return the defined straight line (linear function) to the original coordinate axes. This allows the gradient of the inverse transformed straight line to be calculated based on the actual gradient (a err ) can also be determined. The parameters (angle width from the 45-degree line) that define the range of areas E31 and E71 can be determined from actual performance data. If it is difficult to determine the parameters, the angle width can be assumed to be zero, and areas E31 and E71 can be omitted.

[0045] 6 shows an example in which performance data is plotted in a feature space as a simple physical model ML. First, the simple physical model generation unit 33 determines the position to plot in the feature space based on the relationship between the error and gradient in the performance data. Next, the simple physical model generation unit 33 performs plotting in a different display mode based on the operation direction in the performance data. In this way, the performance data can be classified into appropriate data and inappropriate data (as data to be learned by the learning model) according to the relationship between the error, gradient, and operation direction.

[0046] For example, the simplified physical model generation unit 33 plots performance data in which the jack operation was performed in a positive (+) direction in a first color (e.g., blue), while the simplified physical model generation unit 33 plots performance data in which the jack operation was performed in a negative (-) direction in a second color different from the first color (e.g., red).

[0047] In this case, in the first region (regions E1, E2, E4, E32, and E73) in the feature space, the performance data in which an appropriate operation that reduces the error when the error is negative (or positive close to zero), that is, the jack operation in the direction that moves the error to the plus (+) side, is plotted in a first color (e.g., blue). On the other hand, the performance data in which an inappropriate operation that increases the error in the negative direction, that is, the jack operation in the direction that moves the error to the minus (-) side, is plotted in a second color (e.g., red).

[0048] That is, the performance data plotted in the first region in a first color (e.g., blue) indicates that the jack operation was performed appropriately in a direction that reduces the error, whereas the performance data plotted in the first region in a second color (e.g., red) indicates that the jack operation was performed appropriately in a direction that increases the error.

[0049] In addition, in the second region (regions E9, E8, E6, E72, and E33) in the feature space, when the error is positive (or negative close to zero), the performance data in which the jack is operated appropriately to reduce the error, that is, in which the jack is operated in a direction that moves the error to the negative (-) side, is plotted in a second color (e.g., red). On the other hand, the performance data in which the jack is operated appropriately to increase the error in the positive direction, that is, in which the jack is operated in a direction that moves the error to the positive (+) side, is plotted in a first color (e.g., blue).

[0050] That is, the performance data plotted in the second region in the second color (e.g., red) indicates that the jack operation was performed appropriately in a direction that reduces the error, whereas the performance data plotted in the second region in the first color (e.g., blue) indicates that the jack operation was performed appropriately in a direction that increases the error.

[0051] In this way, by plotting the performance data in the feature space, it is possible to classify the performance data into data that is appropriate for use as learning data and data that is inappropriate. In other words, it is possible to determine whether it is appropriate to use the performance data as learning data based on the distribution state of the performance data plotted in the feature space.

[0052] The simple physical model generation unit 33 calculates an index based on the distribution state of the performance data plotted in the feature space. The index here is an index for determining whether it is appropriate to use the performance data as learning data. For example, the simple physical model generation unit 33 calculates the ratio of inappropriate data contained in the performance data as the index. In this case, a larger index value indicates that the performance data contains a larger amount of inappropriate data (as learning data), and indicates that the quality is too low to be used as learning data. Specifically, the simplified physical model generation unit 33 calculates, as the index, a ratio (first ratio) of the performance data plotted in the first region that is plotted in a second color (e.g., red). The simplified physical model generation unit 33 may also calculate, as the index, a ratio (second ratio) of the performance data plotted in the second region that is plotted in the first color (e.g., blue). The simplified physical model generation unit 33 may also calculate, as the index, an average value of the first ratio and the second ratio.

[0053] Of course, the simplified physical model generating unit 33 may calculate the ratio of appropriate data contained in the performance data as an index. In this case, a larger index value indicates that the performance data contains more appropriate data (as learning data), and thus indicates high quality for use as learning data.

[0054] Furthermore, the simple physical model generation unit 33 may enable the user to visually check the distribution state of the actual data in the simple physical model ML by outputting the actual data plotted in the feature space via the output unit 38. By looking at the distribution state of the actual data, the user can qualitatively determine whether or not the actual data can be used as learning data. For example, the simplified physical model generating unit 33 plots in the feature space each piece of performance data corresponding to excavation performed at different sites and geologies, and displays the distribution of each piece. The user can compare the distributions and determine whether each piece of performance data is suitable for use as learning data, or to what extent a particular piece of performance data deviates from other pieces of performance data.

[0055] Returning to the explanation of Fig. 2, the learning data generating unit 37 generates learning data. The learning data generating unit 37 generates learning data according to, for example, an index (an index for determining whether it is appropriate to use performance data as learning data) calculated from the simplified physical model ML.

[0056] For example, the learning data generating unit 37 generates learning data by removing data that is inappropriate as learning data from the performance data. Specifically, the learning data generation unit 37 regards as the learning data the actual data plotted in the first area excluding the actual data plotted in the second color (e.g., red), that is, the actual data plotted in the first area in the first color (e.g., blue). Moreover, the learning data generation unit 37 regards as the learning data the actual data plotted in the second area excluding the actual data plotted in the first color (e.g., blue), that is, the actual data plotted in the second area in the second color (e.g., red). The learning data generating unit 37 stores the generated learning data in the learning data storage unit .

[0057] The output unit 38 outputs various information in response to operation input by the user. For example, the output unit 38 outputs a simplified physical model ML in which the performance data is plotted, and displays it on a display (not shown). This allows the user to grasp the distribution state in the performance data, and to determine whether it is appropriate to use the performance data as learning data by comparing it with the distribution state in other performance data. The output unit 38 may also output the learning data generated by the learning data generation unit 37.

[0058] The storage units (including the performance data storage unit 34, the simplified physical model storage unit 35, and the learning data storage unit 36) of the information processing device 30 are configured with storage media such as a hard disk drive (HDD), a flash memory, an electrically erasable programmable read only memory (EEPROM), a random access read / write memory (RAM), a read only memory (ROM), or a combination of these. The information processing device 30 storage units store programs for executing various processes for realizing the functions of the information processing device 30, and temporary data used when executing the various processes. The performance data storage unit 34 stores performance data. The simplified physical model storage unit 35 stores the simplified physical model ML. The learning data storage unit 36 ​​stores learning data.

[0059] In addition, the functional units (operation data acquisition unit 31, directional data acquisition unit 32, simple physical model generation unit 33, learning data generation unit 37, and output unit 38) of the information processing device 30 are realized by having a CPU (Central Processing Unit) and / or GPU (Graphics Processing Unit) that the information processing device 30 has as hardware execute a program.

[0060] (Processing flow performed by information processing device 30) Here, the flow of processing performed by the information processing device 30 will be described with reference to Fig. 7 and Fig. 8. Fig. 7 and Fig. 8 are flowcharts showing the flow of processing performed by the information processing device 30 according to the embodiment. Fig. 7 shows the flow of processing in which the information processing device 30 generates and outputs a simple physical model ML.

[0061] FIG. 7 shows a process flow of the information processing device 30 generating the simple physical model ML. As shown in FIG. 7, first, the operation data acquisition unit 31 of the information processing device 30 acquires operation data (step S10). Next, the direction data acquisition unit 32 acquires direction data (step S11). The simple physical model generation unit 33 generates the simple physical model ML (step S12). The simple physical model generation unit 33 generates performance data using the operation data and the direction data, and generates the simple physical model ML by plotting the generated performance data in a feature space. The simple physical model generation unit 33 stores the generated simple physical model ML in the simple physical model storage unit 35 (step S13). The simple physical model generation unit 33 calculates an index (step S14). The output unit 38 outputs the simple physical model ML in which the performance data is plotted, and the index in the performance data (step S15).

[0062] FIG. 8 shows a process flow of the information processing device 30 generating learning data. As shown in FIG. 8, first, the learning data generating unit 37 acquires the performance data by referring to the performance data storage unit 34 (step S20). The learning data generating unit 37 acquires the simplified physical model ML by referring to the simplified physical model storage unit 35 (step S21). The learning data generating unit 37 generates learning data by removing inappropriate data from the performance data based on the performance data and the simplified physical model ML (step S22). The learning data generating unit 37 stores the generated learning data in the learning data storage unit 36 ​​(step S23). The output unit 38 outputs the learning data by referring to the learning data storage unit 36 ​​in response to an operation input by a user or the like (step S24).

[0063] As described above, the information processing device 30 of the embodiment includes a simple physical model generation unit 33. The simple physical model generation unit 33 generates a simple physical model ML. The simple physical model ML is a model for determining whether it is appropriate to use the performance data as learning data for the learning model to learn. The performance data is data indicating the past operation performance of the shield excavator, and is data in which an instruction value, a current value, and an operation direction are associated with each other. The simple physical model generation unit 33 generates the simple physical model ML based on the relationship between the error, the gradient, and the operation direction in the performance data. The error is the difference between the instruction value and the current value. The gradient is the amount of change in the error. The operation direction is the excavation direction determined by operating the propulsion jack 20 in the shield excavator 10. As a result, the information processing device 30 of the embodiment can generate a simplified physical model ML, and can distinguish between appropriate and inappropriate learning data based on the relationship between the error, gradient, and operation direction in the performance data. Therefore, it is possible to determine whether performance data indicating past operation performance is appropriate as learning data.

[0064] In the information processing device 30 according to the embodiment, the simple physical model generating unit 33 generates a feature space including a first axis corresponding to error and a second axis corresponding to gradient as the simple physical model ML. This makes it possible to distinguish between areas suitable for learning data and areas not suitable for learning data using the feature space. Therefore, it is possible to present whether or not performance data is suitable for learning data in association with the area in the feature space.

[0065] Furthermore, in the information processing device 30 of the embodiment, the simple physical model generation unit 33 generates a simple physical model ML in which the operation directions in the performance data are plotted in the feature space. In this way, in the information processing device 30 of the embodiment, performance data can be mapped in the feature space, and the performance data can be classified into appropriate and inappropriate training data based on the distribution state and presented. Therefore, it is possible to present the performance data in a manner that allows the user to easily see whether the performance data is appropriate as training data or not.

[0066] Moreover, in the information processing device 30 of the embodiment, the simple physical model generation unit 33 calculates an index for determining whether it is appropriate to use the performance data as learning data. The simple physical model generation unit 33 calculates, as an index, the proportion of performance data plotted in a first region (an example of a specific region) in the feature space where it is appropriate to correct the operation direction to a direction where the error is plus (+) (an example of a specific direction) where the operation direction is minus (-) (an example of a direction different from the specific direction). This allows the information processing device 30 of the embodiment to quantitatively calculate the degree to which it is appropriate to use the performance data as learning data.

[0067] Moreover, the information processing device 30 of the embodiment further includes a learning data generating unit 37. The learning data generating unit 37 generates learning data. The learning data generating unit 37 generates learning data by excluding the operation direction in a negative (-) direction (an example of a direction different from the specific direction) from the performance data plotted in a first region (an example of a specific region) where it is appropriate to correct the operation direction in a direction where the error is positive (+) (an example of a specific direction). As a result, the information processing device 30 of the embodiment can generate learning data by excluding data inappropriate for learning from the performance data, and can improve the quality of the learning data.

[0068] (Modification of the embodiment) Here, a modified example of the embodiment will be described. In this modified example, re-labeling (data expansion) is performed to artificially increase the amount of learning data.

[0069] The jack operation in the shield tunneling machine 10 is not performed all the time, but is performed irregularly at time intervals. For example, Fig. 3 shows that no operation is performed from time t=0 (zero) to time t=t1-1, operation is performed at time t=t1, then no operation is performed until time t=t2-1, and operation is performed at time t=t2.

[0070] If there is a lot of data in the training data where no operation is performed and the change in the operation direction (change in the point of force) is 0 (zero), there is too little training data that actually shows the relationship between the error and the operation, which can cause overlearning in the learning process and a significant decrease in the estimation accuracy of the trained model. Alternatively, there is a high possibility that a trained model that always estimates that no operation will be performed will be generated.

[0071] For this reason, in this modified example, the performance data is increased artificially. For example, the learning data generating unit 37 generates new performance data (referred to as second performance data) by performing "re-labeling" such that the jack is operated in a time section corresponding to a certain time N before (or after) the time when the operator actually operated the jack in the performance data (referred to as first performance data). This makes it possible to generate N pieces of second performance data from the first performance data based on the actual operation performance, and it is possible to (artificially) increase the performance data by N times.

[0072] However, as described in the above embodiment, the performance data may contain data that is inappropriate for learning. If data augmentation is performed using inappropriate data, the quality of the training data will be reduced.

[0073] To address this issue, in this modified example, data that is inappropriate for learning is excluded from the target of data augmentation. This makes it possible to prevent the quality of the pseudo performance data generated by data augmentation from deteriorating as learning data.

[0074] Specifically, the learning data generating unit 37 first performs data expansion to generate pseudo data regardless of whether the performance data includes inappropriate data. Next, the learning data generating unit 37 maps the pseudo data to the simplified physical model ML. Then, the learning data generating unit 37 generates learning data by removing data that is inappropriate as learning data from the simplified physical model ML to which the pseudo data is mapped. This makes it possible to generate as much pseudo data as possible that is appropriate for use as learning data. Therefore, it is possible to increase the amount of data that satisfies the simplified physical model, i.e., is appropriate for use as learning data, while maintaining the variation of the learning data. Alternatively, the learning data generation unit 37 may determine whether to use the pseudo data as learning data depending on an index calculated from the simple physical model ML (an index for determining whether it is appropriate to use the pseudo data as learning data). For example, when an index calculated from the simplified physical model ML to which the pseudo data is mapped is equal to or greater than a threshold value (a value previously set as a lower limit value appropriate for use as learning data), the learning data generating unit 37 may adopt the pseudo data as learning data, and adopt pseudo data having an index less than the threshold value as learning data. In this case, the pseudo data adopted as learning data may be removed from the data that does not satisfy the simplified physical model, i.e., data that is inappropriate as learning data, and used as learning data. Furthermore, the learning data generation unit 37 may extract data from the performance data in chronological order based on the performance data and the simplified physical model ML, and if the extracted data is appropriate data, generate new performance data in which the jack was operated in a time period corresponding to a fixed time N before (or after) the time when the jack was operated in the data. If the data extracted from the performance data is inappropriate data, the simplified physical model generation unit 33 may not generate new performance data in which the operation time in the data is shifted forward (or after) by a fixed time N.

[0075] As described above, in the information processing device 30 according to the modified example of the present embodiment, the learning data generating unit 37 generates data-extended learning data. The learning data generating unit 37 generates pseudo learning data by data-extending achievement data in which the operation direction in achievement data plotted in a first region (an example of a specific region) in the feature space is a plus (+) direction (an example of the same direction as the specific direction) where it is appropriate to correct the operation direction in a direction in which the error is plus (+) (an example of a specific direction). As a result, in the information processing device 30 according to the modified example of the embodiment, data extension can be performed without deteriorating the quality of the learning data, and high-quality learning data can be increased in a pseudo manner.

[0076] In the above-described embodiment, the current value of the left and right jack stroke difference and the indicated value of the left and right jack stroke difference are used as directional data, but the present invention is not limited to this. Any data indicating at least the direction in which the shield excavator 10 advances can be used as directional data. Examples of directional data include cutter torque, cutter speed, thrust pressure, thrust speed, thrust speed instruction deviation value, controlled earth pressure, face earth pressure average value instruction range outside dummy, agitator torque, screw speed, primary screw pressure, secondary screw pressure, No. 1 copy stroke, No. 1 copy stroke instruction value difference, No. 1 copy position, No. 1 copy position instruction deviation value, pitching, pitching instruction value difference (pitching angle error), rolling, rolling It is possible to use the following: angle indication difference, up and down bend angle, up and down bend angle indication difference, left and right bend angle, left and right bend angle indication difference, S / M front body direction, S / M front body direction indication difference, S / M rear body direction, S / M rear body direction indication difference, planned route horizontal deviation (horizontal deviation: control point), planned route vertical deviation (vertical deviation: control point), direction (azimuth angle error: control point), direction indication difference (control point), planned route direction (control point), pitch (control point), planned route pitch (control point), etc. These direction data can be used as explanatory variables for the learning data to be trained by the learning model.

[0077] The information processing device 30 in the above-mentioned embodiment may be realized in whole or in part by a computer. In that case, a program for realizing this function may be recorded in a computer-readable recording medium, and the program recorded in the recording medium may be read into a computer system and executed to realize the function. The term "computer system" as used herein includes hardware such as an OS and peripheral devices. The term "computer-readable recording medium" refers to portable media such as flexible disks, optical magnetic disks, ROMs, and CD-ROMs, and storage devices such as hard disks built into a computer system. The term "computer-readable recording medium" may also include a medium that dynamically holds a program for a short period of time, such as a communication line when a program is transmitted via a network such as the Internet or a communication line such as a telephone line, and a medium that holds a program for a certain period of time, such as a volatile memory inside a computer system that is a server or client in that case. The above-mentioned program may be a program for realizing a part of the above-mentioned function, or may be a program that can realize the above-mentioned function in combination with a program already recorded in the computer system, or may be a program that is realized using a programmable logic device such as an FPGA.

[0078] Although an embodiment of the present invention has been described in detail above with reference to the drawings, the specific configuration is not limited to this embodiment, and designs that do not deviate from the gist of the present invention are also included. [Explanation of symbols]

[0079] 10... shield tunneling machine, 20... propulsion jack, 30... information processing device, 31... operation data acquisition unit, 32... direction data acquisition unit, 33... simple physical model generation unit, 37... learning data generation unit, 38... output unit

Claims

1. a simplified physical model generation unit that generates a simplified physical model for determining whether or not it is appropriate to use actual data indicating past operation results of the shield excavator, the simplified physical model being based on a relationship between an error between an instruction value and a current value in the actual data in which an instruction value, a current value, and an operation direction are associated with each other, a gradient which is a change in the error, and the operation direction in the actual data, as learning data for a learning model; An information processing device comprising:

2. the simplified physical model generation unit generates, as the simplified physical model, a feature space including a first axis corresponding to the error and a second axis corresponding to the gradient. The information processing device according to claim 1 .

3. the simplified physical model generation unit generates the simplified physical model by plotting the operation direction in the performance data in the feature amount space. The information processing device according to claim 2 .

4. the simplified physical model generation unit calculates, as an index for determining whether it is appropriate to use the achievement data as the learning data, a ratio of the operation directions different from the specific direction among the achievement data plotted in a specific region in the feature space where it is appropriate to correct the operation direction to the specific direction. The information processing device according to claim 3 .

5. a learning data generating unit that generates the learning data by excluding operation directions different from the specific direction from the achievement data plotted in a specific region in the feature space where it is appropriate to correct the operation direction to the specific direction, The information processing device according to claim 3 .

6. a learning data generating unit that generates the learning data by data-extending the achievement data in which the operation direction in the achievement data plotted in a specific region in the feature space where it is appropriate to correct the operation direction to a specific direction is the same as the specific direction, The information processing device according to claim 3 .

7. An information processing method performed by a computer, comprising: the simplified physical model generation unit generates a simplified physical model for determining whether it is appropriate to use the performance data as learning data for training a learning model, based on the relationship between an error between an instruction value and a current value in the performance data, which indicates past operation performance of the shield excavator and in which an instruction value, a current value, and an operation direction are associated with each other, a gradient which is a change in the error, and the operation direction in the performance data; Information processing methods.

8. On the computer, Generate a simplified physical model for determining whether it is appropriate to use the performance data as learning data for a learning model, based on the relationship between an error between an instruction value and a current value in performance data showing past operation performance of a shield machine, the error being a gradient which is a change in the error, and the operation direction in the performance data, in which an instruction value, a current value, and an operation direction are associated with each other; program.

Images