JP7724652B2 - Control system, control method, and program - Google Patents

Description

The present invention relates to a system and method for controlling the excavation direction of an excavating machine, and a program for causing a computer to execute such control.

For tunnel excavation machines (shields) that use the shield method to excavate tunnels, there is a demand for an automatic operation system that can autonomously correct deviations (deviations) from the planned excavation line without relying on manual operation by an experienced operator. To address this, various operation optimization controllers have been developed to minimize deviations by introducing feedback control, where the operating variable is jack pressure and the control variable is deviation.

Implementing such a controller requires formulating the shield's response to the operating variables, and this is done by referring to dynamics theory and empirical rules, for example. Furthermore, if there are any unaddressed elements that cannot be fully described by the formulation, these are addressed by incorporating correction coefficients determined by matching the actual object. This method makes it possible to operate the shield almost autonomously if the planned excavation line is a straight line or a gentle curve.

Recent developments in complex urban underground spaces require the excavation of three-dimensional sharp curves. However, the controllers described above do not adequately correct deviations in sharp curves, requiring the operator to make constant corrections, which works against automation.

There is a known technology that uses neural networks to solve problems that are difficult to formulate by utilizing algorithms that can describe any nonlinear relationships. One proposed technology uses the operation records of skilled operators and their excavation performance as training data, trains a neural network using deep learning, and uses the trained neural network to realize the operations of a skilled operator that respond to excavation conditions (see, for example, Patent Document 1).

Japanese Patent Application Laid-Open No. 2021-17758

The above-mentioned conventional technology can learn features that are relatively frequently performed, outputting the frequency and amount of operation desired by an experienced operator and can be used as guidance for less experienced operators. However, it does not provide any safety measures in the event of inappropriate operation. This poses a problem in that it cannot be actually implemented in a controller and linked to the shield control input to safely and appropriately correct deviations from the shield's planned line.

The present invention has been made in consideration of the above-mentioned problems, and provides a system for controlling the excavation direction of an excavation machine,
a first prediction means for inputting one or more data items indicating the excavation status of the excavation machine and predicting each piece of data after a certain time has elapsed or after the excavation machine has moved a certain distance using a first trained model;
a control means for calculating an operation amount for controlling the excavation direction of the excavation machine based on a control target value of the excavation machine and the input control predicted value;
a second prediction means for receiving as input the calculated operation amount, the predicted value of each data, and the deviation of the excavation direction from the excavation plan line, and predicting a control amount to be input to the control means as a control prediction value using a second trained model;
A control system is provided in which the control means calculates an operation amount using a control prediction value input from the second prediction means during a certain period of time or while moving a certain distance, and repeatedly outputs the calculated operation amount to the second prediction means, thereby optimizing the operation amount.

The present invention makes it possible to safely and appropriately correct deviations from the shield's planned line by linking it to the shield control input.

FIG. 10 is a diagram showing an example of a shield configuration. A diagram showing an example of how deviations are corrected using shield jack pressure. 10A and 10B are diagrams illustrating disturbance factors other than jack pressure of a shield having a center-bending mechanism. 10A and 10B are diagrams illustrating the deviation of an excavation line indicating the excavation direction of a shield having a center bending mechanism from an excavation plan line. FIG. 1 is a block diagram showing an example of the configuration of a system that performs model predictive control. FIG. 10 is a diagram showing an example of parameters indicating the excavation status of a shield. FIG. 10 is a diagram showing an example of the configuration of a controller as a system for controlling the excavation direction of a shield. 1 is a flowchart showing the procedure for carrying out excavation work using a shield.

The control system of the present invention autonomously corrects the deviation of the excavation line, which indicates the excavation direction of the shield, from the excavation plan line in an excavation machine (shield) that excavates a tunnel using the shield tunneling method. Shield tunneling is a construction method in which a shield, which is a steel cylinder, is pushed forward with a jack to excavate the tunnel, and segments, which are tunnel lining divided into steel or concrete blocks, are assembled at the rear of the shield to build the tunnel.

The shield will be briefly described with reference to Figure 1. The shield 10 has a cutter head 11 at its tip, which is equipped with multiple bits for cutting the natural ground, etc. The cutter head 11 rotates in a fixed direction and cuts the natural ground, etc., using the multiple bits attached to its tip. The cutter head 11 is equipped with a copy cutter, which is a bit used when constructing curved sections of a tunnel. The copy cutter, also known as a copy, is stored on the side of the cutter head 11 when constructing straight sections of a tunnel, and protrudes from that side to perform overbreaking when constructing curved sections that require overbreaking.

The shield 10 is equipped with a chamber 12, located immediately behind the cutter head 11, which is a space filled with soil and sand to prevent the excavation surface (face) in the direction of excavation by the shield 10 from collapsing. Additives such as muddy water and air bubbles are injected into the excavated soil to improve its fluidity and water-stopping properties, prevent the excavated soil from adhering within the chamber 12, and enable smooth excavation while maintaining the stability of the face. The soil and sand filled in the chamber 12 is removed by a screw conveyor 13 and transported to the outside of the mine by a belt conveyor 14, where it is disposed of as discharged soil.

The shield 10 is equipped with an erector 16 for assembling the segments 15. The erector 16 holds the segments 15, which have been divided into blocks, and assembles the segments 15 into a ring shape at a predetermined position. The shield 10 is equipped with multiple shield jacks 17. The multiple shield jacks 17 push the assembled segment rings, moving the shield 10 forward. The shield jacks 17 are hydraulically operated, and the extension length (jack stroke) of the shield jacks can be changed, for example, by the jack pressure. The shield jacks 17 are installed in a retracted state when the segment rings are assembled, and the shield 10 is moved forward by extending the jacks.

The shield 10 comprises a skin plate 18, which is the outer shell of the shield 10 and acts as a part that resists earth pressure and water pressure, and a backfill injection device 19 that injects grout into the back of the segment 15 (between the segment 15 and the excavated tunnel wall).

Next, referring to Figure 2, an example of deviation correction using jack pressure will be described as the basis for operating the shield 10. Multiple shield jacks 17 are arranged approximately circumferentially within the approximately circular excavation cross section, and by setting the jack pressure of these multiple shield jacks 17 to be unevenly distributed, a moment can be applied to the shield 10 to change the orientation (face orientation) of the face plate equipped with multiple bits of the cutter head 11 at the tip of the shield 10.

In the example shown in Figure 2, the shield 10 excavates along the excavation plan line 20 while the segment ring 21 is assembled at the rear, and the jacking direction is changed by distributing the jack pressure unevenly and changing the jack stroke on the left and right sides of the excavation plan line 20. The excavation plan line 20 is a line that indicates the designed route for excavating the tunnel, and the jacking direction indicates the excavation direction. In this example, the excavation direction has deviated (deviation) from the excavation plan line 20, and the deviation needs to be corrected.

For example, if there is a deviation to the left of the excavation plan line 20, the face orientation can be corrected to align with the excavation plan line 20 by changing the angle of the face orientation in the opposite direction, to the right. Therefore, by controlling the jack pressure of the shield jack 17 on the left side of the shield 10 to be large and the jack pressure of the shield jack 17 on the right side to be small, a moment can be generated that changes the face orientation of the shield 10 from the left side to align with the central excavation plan line 20.

However, with the recent development of complex urban underground spaces, multiple tunnels and underground structures have already been built, making it necessary to build a tunnel in a way that avoids these. When building a tunnel in such an underground space, a three-dimensional sharp curve must be excavated. With the shield 10 structure shown in Figure 2, even if the jack pressure is set to be unevenly distributed and the facing direction is changed, the angle is small, making it impossible to excavate a sharp curve.

To achieve a face orientation that matches a sharp curve, a center-bending shield is used, as shown in Figure 3, in which the fuselage of the shield 10 is divided into two, front and rear, and the front fuselage 22 and rear fuselage 23 are connected by a hinge, providing a center-bending mechanism that allows the shield to bend midway. The center-bending shield changes its face orientation to match a sharp curve, not only using multiple shield jacks 17 but also the center-bending mechanism.

When making a sharp curve, for example in the case of trucks or buses, a space greater than the vehicle width is required to allow for wheel difference and overhang from the center of rotation. Shield 10 also requires a similar space expanded from the excavation diameter of the straight section in order to make a curve, but is restricted in that it cannot make a curve unless it excavates and creates that space itself.

Therefore, the copy 24 performs over-excavation to allow the shield 10 to assume the posture necessary to make the curve and to secure the space that allows it to make the curve. When the shield 10 advances to the curve, the copy 24 has already completed the necessary over-excavation there, and the shield 10 can assume the required posture within the created extra space by manipulating the jack stroke of the shield jack 17.

As a result of the above-mentioned over-excavation, as shown in Figure 3, the gap between the shield 10 and the natural ground wall when making a sharp curve can become larger in some places than in straight sections. This gap is filled with excavated soil that has been plastically fluidized by the excavation and stirring by the shield 10, or with a modifier with plastic flowability, in order to prevent the collapse of the natural ground wall and to provide the shield 10 with appropriate natural ground reaction and friction forces. For example, it is empirically known that in sharp curves, a good curve is more easily obtained by intentionally controlling the side of the shield to hit the outside of the curve.

Referring to Figure 4, we will explain the interaction of the shield 10 with the ground. By controlling the excavation direction of the shield 10, it is thought that the shield 10 will move in the direction of the face, which is geometrically determined by the face orientation, bend angle, etc.

Here, if the traveling direction of the shield 10 is straight line _L1 and the orientation of the rear trunk 23 of the shield 10 when it is excavating on straight line _L1 is set to an angle of 0, then the orientation of the rear trunk 23 when it curves will be angle _θ1 . Furthermore, the angle formed by a line _L2 perpendicular to the orientation of the rear trunk 23 (the moving direction of the rear trunk 23) and a line _L3 perpendicular to the orientation of the front trunk 22 (the moving direction of the front trunk 22) is the bending angle _θ2 . Then, the angle formed by the straight line _L1 and the line _L3 is the excavation direction _θ3 . Therefore, theoretically, the excavation direction _θ3 is equal to the sum of the orientation _θ1 of the rear trunk 23 and the bending angle _θ2 .

However, in reality, the ground to be excavated is not uniform in quality, and hard rock is present, resulting in complex ground reaction and frictional forces, which causes a deviation from the advancing direction known as the "relief and warpage effect." This deviation from the advancing direction is expressed as an angle Δθ with respect to the direction parallel to the line _L1 . Because of this deviation, the excavation direction _θ3 is calculated by adding Δθ to the face direction _θ1 of the rear body 23 and the center bend angle _θ2 .

When using a bending shield to excavate a sharp curve, it is necessary to consider not only the jacking pressure of the shield jack 17, but also the bending mechanism, relief and warping effects, etc. Therefore, controlling the jack pressure alone is not enough to properly correct deviations.

The controller that corrects the deviation in the shield 10 shown in Figure 2 minimizes the deviation by performing feedback control with the manipulated variable being the jack pressure and the controlled variable being the deviation. In doing so, it is necessary to formulate the shield's response to the manipulated variable, and this is done with reference to dynamics theory and empirical rules, for example.

However, when implementing a controller in the center-bending shield shown in Figure 3 to correct deviations, it is necessary to consider not only the jack pressure but also the effects of mechanical interactions such as the center-bending mechanism and the relief and warping effects, which makes the dynamic behavior complex and difficult to formulate.

To address this issue, the formulation of the deviation response to the shield's operating input, which is necessary for implementing the controller, is achieved using a neural network trained through deep learning, rather than relying on dynamics theory or empirical rules. A neural network is a combination of mathematical models of the nerve cells (neurons) that make up the neural circuits of the brain. In this case, a recurrent neural network (RNN) can be used as the neural network to handle deviations that change over time.

An RNN has three types of layers: an input layer, a hidden layer, and an output layer. The input layer receives data at each time, and the output layer outputs the results at each time. The hidden layer of an RNN has a path that allows the output from the hidden layer at any time t to be the input for the next time t+1. In other words, at time t+1, the hidden layer receives as input the input from the input layer and the output from the previous time t via the path. For this reason, an RNN needs to retain the output from the hidden layer from the previous time, and uses long short-term memory (LSTM) as the algorithm (model) for this. Using LSTM makes it possible to make inferences based on the context of phenomena that progress over time.

To briefly explain RNN, the intermediate layer takes as input the linear sum of the input data at time t+1 and the output data from the intermediate layer at the previous time t, each with its own weight, and outputs a value converted using an activation function such as a sigmoid function or ReLU (ramp function) as the output data of the intermediate layer. In deep learning, training data known as teacher data is used to adjust the weights used in the intermediate layer.

Neural networks are capable of describing any nonlinear relationship, making them capable of solving problems that are difficult to formulate. However, neural networks can output without error even when their inference deviates from interpolation based on learning experience and becomes an unknown extrapolation. This can lead to problems due to unexpected output, such as reckless operations that a human would not implicitly perform. This makes it impossible to guarantee safety.

Model predictive control (MPC) is known as a method for quickly optimizing control output. MPC is a widely used method that is said to enable high-performance control in complex systems, and is capable of controlling while observing safety and other constraints. MPC is a technology whose key is prediction and optimization. In a shield control model, MPC can be used if the data x _n (n is the time step) indicating the model's status and the data x _n+1 indicating the model's status at the next time step according to the control input u _n and the control input y can be described in the form shown in Figure 5.

When controlling the excavation direction of the shield, the control target can be a deviation (difference between the excavation plan line and the excavation line) of 0, and the control input can be the shield's face orientation. Typical data items that indicate the model's status are shown in Figure 6. These data items indicate the status of the shield, and can be measured and calculated using various sensors and computers, and obtained in chronological order.

The excavation deviation shown in Figure 6 is the deviation (deviation) of the excavation line from the excavation design line, the excavation displacement is the relative movement of the shield 10 based on an arbitrary point in time, and the excavation orientation indicates the direction in which the shield 10 is facing. Pitching and rolling indicate the attitude angle of the shield 10. The bending angle indicates the angle formed by the center line of the front body 22 and the center line of the rear body 23 of the shield 10, which has a bending mechanism. The jack stroke indicates how many millimeters the jack has extended, and the stroke difference (difference in jack stroke) indicates, for example, the difference in extension between the left and right jacks.

Jack pressure indicates the force pushing the jack, and jack speed indicates the speed at which the jack is pushed per second (the speed at which the jack is extended). Copy cutter stroke indicates the amount of copy protrusion from the side of the cutter head 11. Cutter torque indicates the amount of torque applied to the cutter head 11 due to its rotation. Horizontal and vertical moments indicate the moments applied in the horizontal and vertical directions by unevenly distributing the jack pressure to curve the shield 10. Total thrust indicates how much the shield 10 is pushing as a whole. Cutter rotation direction indicates the direction in which the cutter head 11 rotates. Note that these items are merely examples and may include other items. Other items may include the type of soil to be excavated, the depth from the ground surface, the type of additive, etc.

Data for these items can be acquired at regular intervals or after a certain distance has been moved. For example, data can be acquired for each cycle, with one cycle being the time until excavation is stopped each time the segment rings are assembled. Note that the face orientation also changes depending on the bending angle, but since the bending angle is not adjusted during excavation, it is not considered an operation amount. For this reason, the bending angle is included in the data items that indicate the status of the shield 10.

Figure 7 shows an example of the configuration of a controller as a control system. The controller has a hardware configuration similar to that of a typical PC (Personal Computer), including a CPU, and realizes various functions by executing programs using the CPU. The controller includes, as functional means, a first prediction means, a control means, and a second prediction means.

The first prediction means is implemented as a first neural network 30 that receives as input one or more pieces of data u _e indicating the excavation status of the shield 10 and predicts each piece of data at the next time step using a first trained model. The data u _e is field data indicating the status of the shield 10 shown in FIG. 6 .

The control means utilizes MPC and is implemented as an optimization control unit 31, which corresponds to the optimizer shown in Fig. 5, and calculates an operation amount u' for controlling the excavation direction of the shield 10 based on the control target value _yr (deviation = 0) of the shield 10 and the input control prediction value _ym . The operation amount u' is optimized using, for example, the receding horizon method.

The receding horizon method is a technique for determining control inputs at each time step by solving an optimization problem for the interval up to the next time step. The optimization problem is the problem of minimizing deviations under given constraints, and this method determines the control inputs that minimize the deviations, i.e., the optimized manipulated variables.

Constraints are conditions that restrict the amount of operation. The amount of operation is determined by, for example, the jack stroke of the shield jack 17, but it is not possible to operate it beyond the rated value of the jack stroke. Therefore, the rated value of the jack stroke is the constraint. Note that this is just one example, and the constraint is not limited to the rated value of the jack stroke.

The optimal control variable u determined in this way is input to the shield 10 and output to the second neural network 32 as a control variable for further optimization up to the next time step.

The second prediction means utilizes MPC and is therefore implemented as a second neural network 32 corresponding to the model shown in Fig. 5. The second neural network 32 receives as input the manipulated variable u' calculated by the optimization control unit 31, the predicted value u _e ' of each data output from the first neural network 30, and the deviation y _p of the excavation line indicating the excavation direction of the shield 10 with respect to the excavation plan line, and predicts a control value to be input to the optimization control unit 31 using the second trained model, and outputs it as a control predicted value y _m .

In this way, by realizing MPC using the optimization control unit 31 and second neural network 32, it is possible to prevent unexpected operation amounts from being input to the shield 10, thereby ensuring safety. Furthermore, by using a neural network for prediction, it becomes possible to perform control that takes into account numerous disturbance factors other than jack pressure.

The trained model of the manipulated variable is implemented as a prediction model, divided into two parts: a first neural network 30 and a second neural network 32. The first neural network 30 inputs data indicating the current situation of the shield 10 at each time step for future prediction and operates to predict each data item at the next time step. The first neural network 30 can be the same network within the search loop for the optimal manipulated variable of the MPC, which is performed at each time step, and is therefore placed outside the search loop. The second neural network 32 predicts the response of the controlled variable using the manipulated variable as a parameter based on the data predicted by the first neural network 30. The second neural network 32 must be placed within the search loop for the optimal manipulated variable of the MPC. Both the first and second neural networks 30, 32 are constructed as RNNs with long short-term memory (LSTM) because they are intended to be networks that can take into account the change history of time-series data.

One configuration is to absorb the first neural network 30 into the second neural network 32 and omit the first neural network 30. In this configuration, the second neural network 32 alone predicts both the data for the next time step and the response of the controlled variable with the manipulated variable as a parameter. The MPC's search loop for the optimal manipulated variable, which is performed for each time step, is executed many times in total and requires a fast response. With a hardware configuration similar to that of a typical PC (Personal Computer), the MPC's execution performance is low, making this configuration undesirable.

As shown in Figure 7, the execution performance of the MPC can be improved by implementing two separate neural networks: a first neural network 30 that is input once in the search loop for the optimal operation amount of each MPC and is sufficient for prediction, and a second neural network 32 that requires prediction for each individual search for the optimal operation amount of the MPC.

The procedure for carrying out excavation work using a shield will be explained with reference to Figure 8. Starting from step 100, in step 101 an LSTM algorithm is prepared. The LSTM algorithm is an algorithm incorporated into a predictive model. In step 102, data is reconstructed from data collected in previous work, in step 103 test excavation (manual operation) is performed with the shield, and in step 104 training data is collected. The training data is actually obtained by manual operation, etc., and is obtained as actual operating variables such as jack pressure, parameters indicating the excavation status, and deviations from the excavation plan line. Steps 101 to 103 may be performed in parallel or in a different order.

In step 105, specific data is extracted from the collected training data and prepared as data for constructing sharp curves in the shield. In step 106, the prepared data is divided into two groups: training data and test data. Data can be divided by randomly selecting data from the prepared data and dividing it into two groups. Training data is data used to train the LSTM, and test data is data used to test the trained LSTM and verify the effectiveness of the learning.

In step 107, the LSTM is trained using the training data, and the trained LSTM is tested using the test data. When training the LSTM, the weights are adjusted to the optimum. In step 108, the prediction accuracy of the test results is evaluated. A threshold is set to determine how accurately the test results are predicted, and the accuracy is evaluated based on whether the threshold is met or exceeded. The threshold is, for example, 80%, and if the measured or calculated results for the input data are reproduced 80% or more, the prediction can be evaluated as being correct. Note that the threshold is not limited to 80%. If the prediction accuracy does not reach the threshold, return to step 105, prepare data again, and train and test the LSTM.

In step 109, an MPC algorithm is prepared to be incorporated into the predictive model together with the LSTM. In step 110, the LSTM and MPC algorithm are incorporated into the predictive model. In step 111, constraints and control target values are set, and implementation into the controller is completed. This makes it possible to perform MPC control during excavation work.

In step 112, actual excavation work begins and the MPC is started. In step 113, the MPC performs prediction and optimization processing, and in step 114, the optimized operation variables are output. In step 115, the shield 10 is controlled based on the output operation variables, and the deviation of the shield 10 is corrected. The shield 10 obtains the deviation as a control variable, and in step 116, it receives input of parameters, control variables, and operation variables indicating the current excavation status, and returns to step 113, where the prediction and optimization processing is performed again.

In step 117, additional learning data is collected in the shield 10, and in step 118, it is determined whether to end excavation. If excavation is not to be ended, the process returns to step 105. This allows learning and testing to be performed again using the additional learning data, improving prediction accuracy. If excavation is to be ended in step 118, the process proceeds to step 119, where MPC is ended and excavation is terminated.

As explained above, the present invention makes it possible to provide a system, method, and program that can correct deviations in sharp curves, taking into account not only jack pressure but also the effects of mechanical interactions such as the bending mechanism, relief and warping effects.

The control system, control method, and program of the present invention have been described in detail above with reference to the embodiments shown in the drawings. However, the present invention is not limited to the above-described embodiments and can be modified within the scope of what one skilled in the art can conceive, including other embodiments, additions, modifications, and deletions. Any modification that achieves the functions and effects of the present invention is within the scope of the present invention.

10...Shield 11...Cutter head 12...Chamber 13...Screw conveyor 14...Belt conveyor 15...Segment 16...Erector 17...Shield jack 18...Skin plate 19...Backfill injection device 20...Drilling planning line 21...Segment ring 22...Front body 23...Rear body 24...Copy 30...First neural network 31...Optimization control unit 32...Second neural network

Claims (7)

A system for controlling an excavation direction of an excavation machine, comprising:
a first prediction means for inputting one or more data items indicating the excavation status of the excavation machine and predicting each piece of data after a certain time has elapsed or after the excavation machine has moved a certain distance using a first trained model;
a control means for calculating an operation amount for controlling the excavation direction of the excavation machine based on a control target value of the excavation machine and an input control prediction value;
a second prediction means for receiving as input the calculated manipulated variable, the predicted value of each piece of data, and the deviation of the excavation direction from the excavation plan line, and predicting a control variable to be input to the control means as the predicted control value using a second trained model;
the control means, while moving within the fixed time period or the fixed distance, calculates the manipulated variable using the control prediction value input from the second prediction means, and repeatedly outputs the calculated manipulated variable to the second prediction means, thereby optimizing the manipulated variable. The control system described in claim 1, wherein the control means optimizes the manipulated variable in accordance with set constraints. The control system described in claim 1 or 2, wherein the first prediction means is implemented as a first neural network, the second prediction means is implemented as a second neural network, and the first and second neural networks are recurrent neural networks. 4. The control system of claim 3, wherein the first and second neural networks implement a long short-term memory (LSTM) algorithm as the first and second trained models, respectively . The control system described in any one of claims 1 to 4, wherein the excavation machine is a shield machine in which a front body and a rear body are connected by a bending mechanism, and the data includes a bending angle indicating the angle between the front body and the rear body, with the bending mechanism at the center. 1. A method for controlling a drilling direction of an excavating machine, comprising:
a step in which a computer receives as input one or more data indicating an excavation status of the excavation machine and predicts each piece of data after a certain time has passed or after the excavation machine has moved a certain distance using a first trained model;
a step in which the computer calculates an operation amount for controlling the excavation direction of the excavation machine based on a control target value of the excavation machine and an input control predicted value;
the computer receives as input the calculated operation amount, the predicted value of each piece of data, and the deviation of the excavation direction from the excavation plan line, and predicts a control amount to be input as the predicted control value using a second trained model;
The method, wherein the step of calculating the manipulated variable and the step of predicting the controlled variable are repeatedly performed within the fixed time period or while moving the fixed distance, and the manipulated variable is optimized. A program for causing a computer to execute each step included in the method of claim 6.