JP2024070098A - Information processing system and program - Google Patents

Abstract

[Problem] To provide a system, method, and program that make it possible to practically simulate the behavior of an excavating machine. [Solution] An information processing system 30 is a system that reproduces the behavior of an excavating machine, and includes a prediction unit 40 that uses control information output from a trained model used to control the excavating machine and adjustable parameters to predict the operation results of the excavating machine when controlled by the control information, an acquisition unit 41 that acquires the operation results of the excavating machine when actually controlled by the control information, a verification unit 42 that verifies the predicted operation results based on the acquired operation information, and an adjustment unit 43 that adjusts the parameters in accordance with the verification results. [Selected Figure] Figure 5

Description

The present invention relates to an information processing system that reproduces the behavior of an excavation machine and a program for causing a computer to execute the reproduction process.

There is a known technology that uses a trained model learned by machine learning to automate the control of excavation machines such as shield machines (hereinafter simply referred to as shield) (see, for example, Patent Document 1). When shield excavation control is performed using a trained model, the safety and quality of the construction work will decrease if there are abnormal values or missing values in the data input to the trained model. Therefore, in the above technology, as a criterion for evaluating the risk of prediction failure, it is determined whether all of the measurement values are within the learning data range of the measurement values in the teacher data used to train the learning model (whether they are in an interpolated state), and only if the evaluation is successful, the setting value of the control data for the operation of the shield is estimated using the trained model.

When training a learning model, if data whose measured values deviate from the specified values is used as training data, learning to make correct estimations cannot be performed, and the learning model may not estimate appropriate operation settings that correspond to the measured values.

Therefore, a technology has been proposed that determines whether or not to use the data to be judged, which includes operation performance data and situation measurement data, as learning data for a learning model that estimates the setting values for the operation of an excavation machine, based on the degree of deviation between the situation measurement data that measures the excavation situation and the instruction value that is the excavation target (see, for example, Patent Document 2).

JP 2019-143388 A JP 2019-143389 A

By combining the two conventional technologies described above, it is possible to improve the validity and accuracy of automating shield excavation control using a trained model.

In real-world excavation work, there are many cases where conditions such as the alignment and geology change significantly during construction, and the specifications of excavation machinery differ for each work. For this reason, the training data that a single company can collect is data with diverse backgrounds. With such training data, there is no prospect of improving the generalization performance of the learning model, and many of the results fail the prediction failure risk assessment.

To manage the risk of prediction failure, it is necessary to perform simulations in which the output of a trained model is input into an actual device. However, simulators that theoretically reproduce shield behavior to date have been too complex and unrealistic to be put into practical use, as they involve complex models that include many unknown parameters.

The present invention has been made in consideration of the above-mentioned problems, and provides an information processing system that reproduces behavior of an excavation machine,
A prediction means for predicting an operation result of the excavation machine when controlled by the control information, using control information output from a trained model used for controlling the excavation machine and adjustable parameters;
an acquisition means for acquiring an operation result of the excavation machine when actually controlled by the control information;
A verification means for verifying a predicted operation result based on the acquired operation information;
and an adjustment means for adjusting the parameter in accordance with the verification result.

The present invention makes it possible to practically simulate the behavior of an excavation machine.

FIG. 2 is a diagram showing an example of the configuration of a shield as an example of an excavation machine. A diagram illustrating how a shield turns a sharp curve. FIG. 1 is a diagram for explaining data assimilation. FIG. 1 is a diagram showing an example of a hardware configuration of an information processing system. FIG. 2 is a functional block diagram showing an example of a functional configuration of the information processing system. 4 is a flowchart showing an example of shield excavation control. 11 is a flowchart showing an example of a process for estimating the operation result of a shield. A table summarizing the variables involved in the simulation. FIG. 13 is a diagram showing an example of a probability distribution.

Figure 1 shows an example of the configuration of a shield as an example of an excavation machine. An excavation machine is a machine that excavates soil, rocks, natural ground, etc., and those that excavate the entire cross section of a tunnel include a shield, a tunnel boring machine, and a free cross section excavation machine. The excavation machine may be any machine that can excavate the entire cross section of a tunnel, but hereinafter, the excavation machine will be described as a shield.

The shield 10 is a tunnel boring machine that cuts away soil in front of it, holds down the excavation surface that is in danger of collapsing, and discharges an amount of excavated soil outside the tunnel that balances the amount cut away. Behind the shield, tunnel lining blocks called segments are assembled into a ring shape to complete the tunnel structure.

The shield 10 has a cutter head 11 at its tip, which is equipped with multiple cutting bits arranged cylindrically or radially on a rotatable, roughly circular face plate, in order to cut away soil and sand in front of it. The shield 10 also has a shield jack 12, which is pressed against the assembled segments 13 and moves forward by extending the shield jack 12. Multiple shield jacks 12 are arranged at specified intervals along the inner circumference of the skin plate 14, which is a steel outer tube that constitutes the shield 10.

The development of complex urban underground spaces requires the excavation of three-dimensional sharp curves. For this reason, the shield 10 is equipped with a bending mechanism such as a bending jack 15 that allows the front body 14a and the rear body 14b to be connected by a hinge and bend in the middle, with the skin plate 14 divided into two parts, a front body 14a and a rear body 14b, to accommodate sharp curves. Like the shield jack 12, multiple bending jacks 15 are arranged at predetermined intervals along the inner circumference of the front body 14a of the skin plate 14.

The shield 10 is provided with a chamber 16 immediately behind the cutter head 11 as a mixing chamber for applying earth pressure and water pressure to the excavated surface while plastically fluidizing the excavated soil so that it can be discharged. The excavated soil in the chamber 16 is removed by a screw conveyor 17 and transported outside the mine by a belt conveyor 18, where it is disposed of as discharged soil.

The shield 10 is equipped with an erector 19 for assembling the segments 13 inside the rear body 14b. The erector 19 grasps the segments 13 and transports them to a specified position to install them. The shield 10 also has a backfill injection device 20 that injects injection material into the back of the segments 13 (between the segments 13 and the excavated tunnel wall).

The shield 10 moves forward by extending the shield jacks 12. Also, by changing the extension length (stroke) of each of the shield jacks 12 arranged along the inner circumference of the shield 10 and skin plate 14, and by changing the stroke of each of the bending jacks 15 arranged along the inner circumference of the skin plate 14 as necessary, the shield 10 moves forward while bending in a specified direction. The direction of forward movement is the direction toward the horizontal azimuth and the angle (pitching) of the vertical direction relative to the horizontal.

Figure 2 is a diagram illustrating how a shield with a bending mechanism turns a sharp curve. When the shield is bent by the bending jack 15, the angle formed by the center line passing through the center of the cross section of the front body 14a and the center line passing through the center of the cross section of the rear body 14b is the bending angle α. When the shield 10 turns a sharp curve by the bending jack 15, the orientation of the rear body 14b and the bending angle α are manipulated, and the shield 10 turns while adjusting its posture so that it moves along the excavation direction that reduces the deviation between the shield 10 and the planned line.

When making a sharp turn, for example, a bus or truck needs a space larger than the vehicle width to allow for wheel difference and overhang from the center of rotation. The shield 10 also needs a similar space that is expanded from the excavation diameter of the straight section in order to make a turn, but it cannot make the turn unless it creates that space by excavating extra space itself.

Therefore, a copy cutter 21 is provided on the side of the cutter head 11 provided on the front body 14a of the shield 10 as a protrusion that can protrude in the radial direction of the tunnel. The copy cutter 21 is a device for creating a space in which the shield 10 can be curved.

When the shield 10 makes a sharp curve, the bending jack 15 on the inside of the sharp curve is shortened and the bending jack 15 on the outside is extended, and the shield 10 curves in a roughly V-shaped bend. To do this, it is necessary to predict the space that the shield 10 can assume and excavate (over-excavate) when the copy cutter 21 is in that position.

The overexcavated space is not left hollow, but is filled immediately after excavation with an improvement agent with appropriate plastic flow properties. If it is left hollow, there is a risk of the excavation wall collapsing, and it becomes difficult for the shield 10 to adequately withstand the reaction force and friction from the excavation wall, making it difficult to operate the shield 10.

The entry and exit of the copy cutter 21 is controlled continuously in response to the rotation angle of the cutter head 11. This control can be performed, for example, using an equation that shows the relationship between the protruding length of the copy cutter 21 and a preset circumferential angle.

Although such excavation control of the shield 10 can be performed manually, in recent years it can be performed automatically using a trained machine learning model. A machine learning model takes measurement values (observation values) of monitoring items that monitor the behavior of the shield 10 as input data, and outputs control information used to control the excavation of the shield 10. In order to predict control information from such input data, the model is given training data that is a pair of input data and control information obtained by manual control, and is trained.

The essence of machine learning is to make predictions (interpolated predictions) within the range learned from training data. Therefore, a trained model (a trained model) must not be used outside the range it was learned within (extrapolated use). This is because unintended extrapolation poses the risk of prediction failure.

The training data collected by a single company has a diverse background, since the specifications of the shield 10 vary depending on the ground being excavated, and there are many cases where conditions such as the line shape and geology change significantly even during construction. For example, even when excavating the same mountain, the geological conditions can vary greatly because there may be large rocks in the way or groundwater running through it.

A trained model finds patterns and trends from input data and makes predictions for unknown data. The training data collected by manually controlling the shield 10 is data that varies greatly depending on geological and other conditions, so even if the training data is used to train a learning model, there are insufficient examples of the patterns and trends that should be covered, and the learning results do not converge to a sufficient level, making it unable to handle unknown data. As a result, we are faced with a situation where there is no prospect of improving generalization ability, which is the ability to handle unknown data.

As a result, there is always a lack of training data, which results in many failures in prediction failure risk assessments, and the learning models that have been put to good use in excavation control. If attempts are made to make use of the learning models, additional learning must be repeated to make corrections, resulting in poor productivity.

To manage the risk of prediction failure, a simulation is required to test what would happen if the output of the trained model were actually input into the shield 10 and operated. For the simulation, a simulator is used as a device or program that describes natural laws and empirical rules based on previous knowledge in mathematical formulas and returns responses to virtual operations that mimic those of the actual machine.

However, since the available knowledge is limited, it is not possible to create a simulator that reproduces all real-world phenomena. Therefore, conventionally, a simulator is considered practical if it can realistically reproduce the phenomenon of interest. For example, a flight simulator that allows the pilot to change the inclination of the aircraft by raising or lowering the nose by operating the control stick is known as a practical simulator.

For the shield 10, there is a demand for simulators to perform pre-evaluations when excavating unknown sharp curves. To simulate the behavior of the shield 10, it is necessary to create a situation similar to that behavior using a mathematical model, etc., but such a model is dominated by contributions from interactions with the surrounding ground, such as friction and pressure, and is composed of ground with many unknown elements, so it is a complex model that includes a large number of unknown parameters. Models that include a large number of complex unknown parameters are difficult to handle and difficult to reproduce, and therefore have not yet been put to practical use.

Therefore, we have given up on reproducing models that contain a large number of unknown parameters, and instead adopt a method called assimilation simulation, in which adjustable parameters are introduced and, if the results cannot be reproduced through simulation, the parameters are adjusted to match the actual observational results.

Now, with reference to Figure 3, the concept of assimilation will be explained. The horizontal axis represents elapsed time, and the vertical axis represents predicted values predicted by simulation and observed values obtained by operating the actual machine, with the unknown true state indicated by the dotted line. If a simulator that roughly reproduces the true state is prepared and a simulation is performed from the initial value (a) indicated by the double circle, a prediction as indicated by the solid line is obtained, and the predicted value (b) indicated by the white circle at elapsed time t deviates significantly from the observed value (c) indicated by the black circle. Note that the observed value (c) also deviates from the true state indicated by the dashed line, but this is due to an observation error.

If the simulation were to continue based on the predicted value (b) in the prediction at elapsed time t, it is predicted that the result would deviate even more from the observed value. Therefore, parameters are adjusted based on the observed value (c) at elapsed time t, the simulator model is modified to reduce the deviation at elapsed time t, and the modified model is obtained, i.e., data assimilation is performed. When data assimilation is performed, the predicted value (b) is corrected to the analytical value (d) indicated by the double circle, resulting in the result shown by the solid line from the analytical value (d) to the predicted value (e) indicated by the white circle, which is closer to the true state indicated by the dashed line.

Because there is no prospect of improving the generalization performance of the learning model used for excavation control of the shield 10, once a certain level of generalization performance has been obtained, the learned model is input into the excavation control, and the results of this input are verified by a simulation incorporating the above-mentioned data assimilation. Depending on whether the simulation results are acceptable or not, it can be determined whether to use automatic excavation control using the learned model or manual control. This makes it possible to manage the risk of prediction failure under automatic excavation control using a learning model.

Therefore, the present invention provides an information processing system capable of performing a simulation incorporating the above-mentioned data assimilation. Note that the information processing system may be the same as the one that performs learning of the learning model used for excavation control of the shield 10, prediction using the learned model, etc.

Figure 4 is a diagram showing an example of the hardware configuration of the information processing system 30. The information processing system 30 can be configured using a general computer. Therefore, the information processing system 30 can adopt a hardware configuration similar to that of a general computer. The information processing system 30 includes, as hardware, a CPU (Central Processing Unit) 31, a ROM (Read Only Memory) 32, a RAM (Random Access Memory) 33, a HDD (Hard Disk Drive) 34, an external device I/F 35, an input/output I/F 36, a display device 37, and an input device 38.

The CPU 31 controls the entire system and executes various applications. The applications may include a simulator that performs the above simulation. The ROM 32 stores a basic input/output system (BIOS) that loads the operating system (OS) when the system is started and controls input/output for peripheral devices, and firmware that controls circuits inside the system such as the HDD 34. The RAM 33 is used as the main memory and provides a working area for the CPU 31. The HDD 34 stores various applications and the OS executed by the CPU 31, various setting information, various data, etc. Here, the HDD 34 is used, but is not limited to this and may be a storage device such as an SSD (Solid State Drive).

The external device I/F 35 connects the system to external devices such as an operation panel and various environmental measuring devices (sensors), and controls communication with the external devices. The display device 37 is a device for displaying information, such as a liquid crystal display or an organic EL (Electro Luminescence) display. The input device 38 is a device for inputting information using a keyboard or mouse, selecting applications, and issuing instructions to run applications. The input/output I/F 36 controls the output of information to the display device 37 and the input of information from the input device 38. In this example, the display device 37 and the input device 38 have been described as separate devices, but this is not limited to the above, and a touch panel equipped with the functions of both the display device 37 and the input device 38 may be used.

The information processing system 30 may also include other circuits, such as a communication I/F that connects to a network such as the Internet and controls communication with communication devices on the network, and a short-range wireless communication circuit that enables short-range wireless communication using Bluetooth (registered trademark) or the like.

The information processing system 30 may be composed of one device or program, or may be composed of two or more devices or programs.

Figure 5 is a functional block diagram showing an example of the functional configuration of information processing system 30. Each function of information processing system 30 is realized by CPU 31 executing a program such as an application provided in information processing system 30. Note that some or all of each function of information processing system 30 may be configured with hardware such as a circuit.

The information processing system 30 includes, as functional units, a prediction unit 40, an acquisition unit 41, a verification unit 42, and an adjustment unit 43. The prediction unit 40 uses control information output from a trained model used for excavation control of the shield 10 and adjustable parameters to predict the operation result of the shield 10 when excavation control is performed using the control information. A simulator is used to predict the operation result of the shield 10. The simulator is created as a mathematical model, such as a formula including adjustable parameters.

The trained model is a model trained using training data. The training data is data that associates the amount of operation of the shield 10 actually operated by a person with the observed values that measure the results of that operation. Examples of the amount of operation include the bending angle α, the driving force of the shield jack 12, the rotational torque of the cutter head 11, and the stroke of the copy cutter 21. Examples of the observed values include pitching, which indicates the orientation of the shield 10 and the vertical rotation angle. Therefore, by measuring the orientation and pitching of the shield 10 using a sensor or the like and inputting the results into the trained model, the planned line of the tunnel, etc. can be taken into account, the amount of operation, etc. can be output as control information, and the control information can be set in the shield 10, thereby automating the excavation control of the shield 10.

A learning model is constructed so that it can best reproduce the training data through training, but there is a risk of overfitting, where the model only matches the training data. Overfitting occurs when there is little training data or when the model is too complex. Therefore, test data separate from the training data is prepared, and it is checked whether the trained model can reproduce the separate test data (the generalization performance of the model).

The acquisition unit 41 acquires the operation results of the shield 10 when excavation is actually controlled based on the control information. The operation results of the shield 10 are observed values measured using the above-mentioned sensors, etc.

The verification unit 42 verifies the operation result predicted by the prediction unit 40 based on the information acquired by the acquisition unit 41, i.e., the observed value. The prediction unit 40 receives the control information output by the trained model as input, performs a simulation, and predicts the direction, etc. measured by the above-mentioned sensors, etc. The verification unit 42 compares the predicted operation result with the observed value, and verifies the validity of the predicted operation result based on whether the difference is equal to or greater than a threshold value.

The adjustment unit 43 adjusts the parameters according to the verification results. If the verification unit 42 determines that the difference between the predicted operation result and the observed value is equal to or greater than a threshold and that the predicted result is invalid, the adjustment unit 43 adjusts the parameters. The adjusted parameters are set in the mathematical model used by the prediction unit 40 for the simulation, and the mathematical model is modified.

The information processing system 30, which includes the acquisition unit 41, verification unit 42, and adjustment unit 43, is a system that incorporates a data assimilation technique, and has the characteristic of being, in principle, capable of tracking environmental changes and discontinuous situations, which is a characteristic of assimilation. Note that in the extreme case where no previous knowledge is incorporated at all, some of the characteristics of assimilation will be lost. However, since simulations can be performed using a formula that is simply a linear combination of the parameters of interest, it has the characteristic of being easier to handle than previous models.

The information processing system 30 may include other functional units, not limited to the prediction unit 40, the acquisition unit 41, the verification unit 42, and the adjustment unit 43. The information processing system 30 may include, as other functional units, a learning unit that learns a learning model using teacher data, a judgment unit that checks generalization performance, a prediction execution unit that executes predictions using a learned model, etc.

In addition, the information processing system 30 may further include, as another functional unit, an evaluation unit that evaluates the result of control based on the control information output from the trained model by determining whether the result is within an acceptable range.

Figure 6 is a flowchart showing the flow of excavation control of the shield 10. Before performing excavation control, a learning model, parameters to be set in the simulation, etc. are prepared. Excavation control of the shield 10 starts at step 100, and in step 101, manual control is performed to obtain training data. In step 102, training data is obtained, and in step 103, the obtained training data is used to train the learning model.

In step 104, test data prepared separately from the training data is used to confirm the generalization performance of the learning model. In other words, it is confirmed whether the control information can be predicted with a predetermined accuracy and whether a predetermined generalization performance is obtained. The prediction accuracy may be any accuracy and can be determined arbitrarily. After confirming the generalization performance, if the prediction is possible with the predetermined accuracy, proceed to step 105; if the prediction is not possible, return to step 102 and repeat the learning.

In step 105, the system is manually controlled, and observation values measured by sensors or the like attached to the shield 10 are used as input, and the learned model is used to predict control information. In step 106, a simulation is performed using the prediction from the learned model as input, the prediction is corrected as necessary, and the prediction or corrected prediction is output.

In step 107, automatic excavation control (predictive control) is performed using the trained model based on the output prediction. Even if predictive control is performed here and the shield 10 moves in an impossible direction, the behavior of the shield 10 is slow, and there is ample time to make corrections. For this reason, predictive control is performed here.

In step 108, the control result of the predictive control is evaluated. If the control result of the predictive control is within the tolerance range, which indicates that the shield 10 is advancing at an acceptable angle with respect to the planned line, the process proceeds to step 109. On the other hand, if it is outside the tolerance range, the process returns to step 101, where manual control is performed to correct the direction in which the shield 10 advances, etc. Then, in step 102, training data is acquired, and in step 103, the learning model is retrained.

In step 109, it is determined whether excavation has ended. If so, the process proceeds to step 110, where excavation control is terminated. If not, the process returns to step 105, where control information for the next time step is predicted.

Figure 7 is a flow chart showing an example of the simulation of step 106 shown in Figure 6. Starting from step 200, in step 201, a simulation is performed using the prediction of the trained model as input. The simulation is performed by inputting the prediction into a mathematical model including adjustable parameters. The simulation virtually controls the shield 10 using the predicted control information, and outputs the results of the operation. In step 202, the control information predicted by the trained model is actually set in the shield 10, the results of the excavation control are obtained as observed values, and the obtained observed values are compared with the results of the simulation. As a result of the comparison, the validity of the prediction is determined depending on whether the difference between the observed values and the results of the simulation is equal to or greater than a threshold value.

If the difference is equal to or greater than the threshold and the prediction is determined to be invalid, the process proceeds to step 203, where the prediction is corrected. In other words, the parameters used in the simulation are adjusted to correct the prediction so that the simulation results are closer to the observed values.

If the prediction is determined to be valid in step 202, the prediction determined to be valid is output, and if the prediction is determined to be invalid, a corrected prediction is output in step 203 and the simulation is terminated in step 204.

Next, data assimilation will be described in more detail with examples. In step (1), the variables shown in FIG. 8 are prepared as variables related to the simulation. Note that the variables shown in FIG. 8 are only an example, so only some of these variables may be prepared, or other variables may be prepared, or other variables may be prepared in addition to these variables.

The variables shown in FIG. 8 can be broadly divided into variables that indicate the amount of operation of the shield 10, variables that indicate the operating environment, and variables that indicate the operation effect. The variables used as control information are the variables that indicate the amount of operation and the variables that indicate the operating environment, and the observed values measured by sensors, etc. are variables that indicate the operation effect.

After preparing the variables related to the simulation, in step (2), a formula is prepared that is assumed to hold between the prepared variables. Since it is difficult to construct a theory for the simulator of the shield 10, a fictitious formula such as the following formula 1 can be used. Note that the following formula 1 is a formula that outputs the operation effect when a certain operation amount is applied in a certain operating environment.

After preparing the formulas, in step (3), a parameter θ that changes over time is introduced as coefficients to variables indicating the operation amount and the operation environment, that is, θ _X1H to θ _X10H , θ _F1H to θ _F6H , θ _X1V to θ _X10V , and θ _F1V to θ _F6V , giving the simulator a degree of freedom to follow the actual observation results. Then, the above formula 1 can be rewritten as the following formula 2.

After the parameters are introduced, a simulation is performed as step (4). When operation effects ΔH(t) and ΔV(t) are given at a certain time t, the state after the operation at time t can be expressed as in the following formula 3 using ΔH and ΔV calculated in the same manner at previous times t-1, t-2, ..., 2, 1 and the initial values S _H (0) and S _V (0) of the state after the operation.

Here, the simulator output S(t) is written as Equation 4 below.

In parameter adjustment, the simulator's ability to track reality is evaluated and corrected by comparing the simulator output with the actual observed values, so the expected observation error w is added to the actual observed values in advance.

In the above formula 5, Y(t) (Y has a ^ at the top) is the predicted observation value given by the simulator. For w, for example, a sample from a normal distribution with mean 0 and variance ν is given.

If the simulation results show that the parameter θ needs to be adjusted at time t, then in step (5), θ adjustment (assessment and correction of the simulator's ability to track reality) is performed. The predicted observed value Y(t) (Y has a ^ at the top) given by the simulator, the actual observed value Y(t), the operation amount F(t) thought to be the cause of these, and the operating environment X(t) are prepared as a data set. Here, a value sampled from an appropriate probability distribution is given as the initial value of θ used when calculating Y(t) for the first time when no θ adjustment has yet been performed.

As an appropriate probability distribution p(x), for example, a uniform distribution in which x ranges from -1 to 1 with the same probability of 1/2 can be given, as shown in Figure 9. Note that this is just one example, and is not limiting.

After preparing the data set, θ is adjusted. For the adjustment, a simulator optimization technique called assimilation is used. There are many known algorithms for assimilation, and detailed explanations of each algorithm will be omitted here. In the simulator for the shield 10, for example, a method using a Gaussian kernel can be applied.

In assimilation, the value θ contained in an unsatisfactory formula in the initial state is sequentially corrected in a satisfactory direction. Here, θ is not a fixed, specific numerical value, but a realized value sampled from a certain probability distribution π(θ). For this reason, the actual target to be corrected in assimilation is the probability distribution π(θ). If the probability distribution before correction is called the "prior distribution," then the probability distribution after correction is called the "posterior distribution."

In parameter adjustment, we resample θ from the posterior distribution, return to step (3) above, and perform the calculations for the next time step t = t + 1. This is repeated until we obtain a satisfactory simulator.

In this way, the parameter θ can be adjusted to correct the predictions, but there may be cases where the correction status of the probability distribution π(θ), which is the actual target of correction, does not converge well by repeating steps (3) to (5) above, leaving one dissatisfied. In such cases, the correction process can be traced to extract the problem, and the problem can be addressed by adjusting the hyperparameters, which are empirical tuning elements included in the assimilation procedure.

As explained above, even in cases where it is difficult to prepare sufficient training data, such as in shield excavation control, by using a learning model that has been trained to a certain extent, it is possible to verify the validity of predictions through simulation, making it possible to manage the risk of prediction failure and expanding the scope of application of excavation control using trained models. In addition, by applying the assimilation method, it becomes possible to practically simulate shields, which have been considered difficult due to their complexity and the inclusion of a large number of unknown parameters.

This system can be used not only as a simulator for testing trained models, but also as a tool for searching for optimal operations when manually operating a shield, for example, or for pre-evaluating the results of such operations.

Thus far, the simulation system, method, and program of the present invention have been described in detail with reference to the embodiments shown in the drawings, but the present invention is not limited to the above-described embodiments, and can be modified within the scope of what a person skilled in the art can imagine, including other embodiments, additions, modifications, deletions, etc., and any aspect is within the scope of the present invention as long as it provides the functions and effects of the present invention.

Reference Signs List 10...Shield 11...Cutter head 12...Shield jack 13...Segment 14...Skin plate 14a...Front body 14b...Rear body 15...Center bending jack 16...Chamber 17...Screw conveyor 18...Belt conveyor 19...Erector 20...Backfill injection device 21...Copy cutter 30...Information processing system 31...CPU
32...ROM
33...RAM
34...HDD
35...External device I/F
36... Input/output I/F
37: Display device 38: Input device 40: Prediction unit 41: Acquisition unit 42: Verification unit 43: Adjustment unit

Claims (8)

An information processing system that reproduces a behavior of an excavation machine,
A prediction means for predicting an operation result of the excavation machine when controlled by the control information, using control information output from a trained model used for controlling the excavation machine and adjustable parameters;
an acquisition means for acquiring an operation result of the excavation machine when actually controlled according to the control information;
a verification means for verifying the predicted operation result based on the acquired operation result;
and adjusting means for adjusting the parameters in accordance with a verification result. The information processing system according to claim 1, wherein the verification means determines the validity of the prediction means based on whether or not a difference between an observed value, which is the obtained operation result, and a predicted value, which is the predicted operation result, is equal to or greater than a threshold value. The information processing system according to claim 2, wherein the prediction means calculates the predicted value using a formula in which the time-varying control information is a variable and the time-varying parameters are coefficients. The information processing system according to claim 2 or 3, wherein the adjustment means adjusts the parameters when the prediction means is determined to be invalid as a result of the verification. The information processing system according to claim 4, wherein the adjustment means adjusts the parameters by modifying a probability distribution from which the parameters are sampled and resampling from the modified probability distribution. The information processing system according to claim 5, wherein the adjustment of the parameters by the adjustment means, the calculation of the predicted value by the prediction means, and the acquisition of the operation results of the excavation machine by the acquisition means are repeated until the verification means determines that the prediction means is valid. A learning means for learning a learning model used for controlling the excavation machine using teacher information;
A determination means for applying test data to the trained model, which is a trained model, to determine whether the trained model has a predetermined generalization performance;
A prediction execution means for executing a prediction of the control information using the trained model when it is determined that the trained model has the predetermined generalization performance;
The information processing system according to claim 1 , further comprising an evaluation means for evaluating the result of control based on the control information output from the trained model by determining whether the result is within an acceptable range. A program for causing a computer to execute a process for reproducing a behavior of an excavation machine,
A step of predicting an operation result of the excavation machine when controlled by the control information, using control information output from a trained model used for controlling the excavation machine and adjustable parameters;
acquiring an operation result of the excavation machine when actually controlled according to the control information;
verifying the predicted operation result based on the obtained operation result;
adjusting the parameters according to the verification results;
A program that executes a step of re-predicting the operation results of the excavation machine using control information output from the learned model and the adjusted parameters.