JPH04285293A - Neural type optimum gain auto-tuning method of small diameter tunnel robot - Google Patents

Abstract

PURPOSE:To obtain an optimum gain by causing a tunnel robot to learn using a directional control simulation a neural network to which initial deviation and deflection angle are input and from which deviation and deflection angle feedback gains are output. CONSTITUTION:A neural network N is provided to which deviation and deflection angle Y, are input and from which deviation and deflection angle feedback gains Kp, Ka are output. An initial position deviation Y(O) and an initial pitching angle deflection (O) are put in the input layer I of the neural network N and the deviation and deflection angle feedback gains Kp, Ka are output from the output layer O of the network N. These gains are put in a directional control simulation which uses feedback control laws, so that a robot learns the gains. If the difference between the gain which the robot is learning and that which the robot has learned last time is reduced below a predetermined set value as the learning is repeated a fixed number of times, the gains are regarded as having converged and the value of the gain obtained at that time is used as an optimum gain.

Description

[Detailed description of the invention]

0001

[Field of Industrial Application] The present invention relates to a neural-type optimal gain auto-tuning method for feedback direction control of a small diameter tunnel robot, which controls the head angle of the robot tip and corrects the direction while pushing and propelling the robot without soil removal. It is something.

0002

2. Description of the Related Art A conventional technique for determining an optimum gain will be described below. Figure 1 shows the system configuration of the tunnel robot. This system consists of a tunnel robot main body 11 with a head angle correction function, a buried pipe 12, a push pipe device 13 for pushing the buried pipe 12, a hydraulic device 14, and an operation panel 15. The buried pipes 12 are pushed in one by one using hydraulic pressure by a pushing pipe device 13. At this time, the operator 16 sequentially corrects the head angle and performs direction control so as to follow the planned line. 17 is the surface of the earth.

Next, a feedback direction control method for the robot and its simulator will be described. This control method uses a feedback control law in which the next input head angle is the positional deviation and pitching angle deviation of the robot body multiplied by a certain proportional gain. The head angle and pitching angle are defined in FIG. The simulator is composed of a dynamic model related to direction correction [formula (1)] and calculation formulas (2) and (3) for the pitching angle and position of the robot. The direction control simulation is performed as follows. First, the head angle is determined using the control law of equation (4). Next, the head angle is substituted into the dynamic model of equation (1) to calculate the amount of direction correction. Then, equation (2), equation (
3) to calculate the pitching angle and position of the robot.

The dynamic model of this system is expressed as a probability model in which the direction correction angle can be expressed by the sum of a time series term and a probability distribution term of pitching angle changes that approximately represent the head angle and the robot's posture. The parameters an and bn are estimated by the least squares method. Simulator Δθp (k) =a1 Δθp (k-1) +
...+an Δθp (k-n) +bo θh (k)
+b1 θh (k-1
) +...+bn θh (k-n) +e(k)
(1) θp (k) = θp (k
-1) +Δθp (k)
(2)
Y(k) = Y(k-1) +Lsin (θp (
k))
(3) Control law θh (k) = Kp (Yd (k) −Y(k
−1))+Ka (θd (k) −θp (k−1)
) (4) However, Kp: Position deviation feedback gain Ka
: Angle deviation feedback gain

FIG. 3 defines each parameter used in the simulator and control law. The lower trajectory is the planned line, and the upper trajectory is the robot's trajectory. The position of the design line at stroke k is Y
d(k), the slope of the planned line is θd(k), the robot position is Y(k), and the pitching angle of the robot is θp.
(k), pitching angle change amount as Δθp (k),
Let the length of one stroke be L. Furthermore, in the dynamic model of equation (2), e(k) is the residual and n is the order of the model. A block diagram of direction control is shown in FIG.

As a result of examining the effectiveness of this directional control method using the above simulator, we found that if the position deviation feedback gain and the angular deviation feedback gain are appropriately selected,
It was found that good directional control was achieved.

[0007] Next, the optimal feedback gain was determined using the conventional method as follows. An example will be explained using data from the Okayama area where the N value is 0 to 2. Here, the N value is a numerical value representing the hardness and compactness of the soil, and the larger the value, the harder the soil is. Figure 5 shows Kp = 0.01(d
3 shows the deviation absolute value integral value characteristic of Ka-transient response when fixed as (eg/mm). However, the residual e(K) is approximated by a normal distribution with an average value of 0deg and a standard deviation of 0.13deg, and the initial position and angle are 500mm and 0d, respectively.
eg. Further, the planned line was a horizontal line with an initial value of 0 mm. From this figure, Ka
can be found. In this case, Ka=1.5. Next, in the same way as above, increase Kp from 0.01 to 10
(deg/mm) to find the minimum deviation absolute value integral value. Using these results, the Kp-minimum deviation absolute value integral value characteristic shown in FIG. 6 can be determined. From this figure, Kp at which the deviation absolute value integral value is minimum is 0.
07 (deg/mm), and Ka at that time is Kp
When Ka is 0.07 (deg/mm), it is determined to be 7.8 (dimensionless) based on the deviation absolute value integral value characteristic of the transient response. Therefore, the optimal gain for Okayama is Kp =0.
07 (deg/mm), Ka = 7.8 (dimensionless). FIG. 7 shows the results of a directional control simulation using these optimal Ka and Kp. The initial position is 500
mm, and the planned line was a horizontal line at a position of 0 mm. As shown in the figure, it can be seen that good control is being performed.

[0008]

[Problem to be Solved by the Invention] As described above, the optimal gain can be found using the conventional method, but as mentioned above, the problem is that it takes a lot of effort and time to search for the gain by trial and error. there were.

SUMMARY OF THE INVENTION An object of the present invention is to provide a neural-type optimal gain auto-tuning method for a small-diameter tunnel robot that overcomes the problems of the prior art optimal gain search methods, which require a great deal of effort and time.

0010

[Means for Solving the Problems] In order to solve the above problems, the present invention uses a direction control simulation to train a neural network whose inputs are initial deviation and declination, and whose outputs are deviation and declination feedback gain. It is characterized by obtaining the optimum gain.

[0011]

[Operation] By the means described above, the present invention auto-tunes the optimum gain using the learning ability of the neural network, so it does not require much effort and time is significantly shortened compared to the conventional method.

0012

DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

FIG. 8 shows a configuration diagram of a neural network for obtaining an optimal gain. This neural network consists of three layers: an input layer, a middle layer, and an output layer. Each layer is a collection of units, and units from adjacent layers are all connected. Through this connection, values are transmitted from each unit in the previous layer to the units in the subsequent layer. At this time, there is a weighting coefficient unique to the connection, and a value obtained by multiplying the output value of the previous layer unit by this weight is input to the subsequent layer unit. The subsequent layer unit calculates the sum of the values from all units in the previous layer, performs a nonlinear transformation defined as an output function, and then outputs the value to the next layer. The number of units in the input layer is two, and the initial position deviation and initial pitching angle deviation are input to the input layer units. The number of units in the middle layer is set to an arbitrary number. The output layer unit has two layers: a position deviation feedback gain Kp and a pitching angle deviation Ka, which are input to the feedback control law.

FIG. 9 shows a flowchart of optimal gain auto-tuning of feedback gains Kp and Ka. The specific process of auto-tuning is described below. (1) Enter arbitrary data for the minimum deviation and minimum declination as initial settings.

(2) Put the gain output from the output layer into a feedback control law simulator, perform a simulation, and calculate the sum of the positional deviation of the planned line and the robot body at each stroke at that time, and the pitching angle deviation. . This is defined as the total deviation error and total declination error. Also, the total deviation error when the total deviation error becomes the minimum during learning,
The total declination error is stored as the minimum total deviation and the minimum total declination, respectively.

(3) Calculate the absolute value of the difference between the total deviation error and the minimum total deviation, and between the total argument error and the minimum total argument, and use these as the deviation coupling coefficient correction error and the argument coupling coefficient correction error. . Here, these values are large as they are, so they are set to 50 and 50, respectively.
It's divided by 5. Further, the maximum value of each coupling coefficient correction error was set to 10 in both cases.

(4) Next, the gain K output from the output layer
The gain obtained by adding or subtracting an appropriate value to p and Ka (Kp ±α
, Ka ±β), the small diameter tunnel robot is simulated four times, and the sign of the addition and subtraction of the gains (Kp ±α, Ka ±β) when the minimum total deviation error is achieved is:
A sign is assigned to each coupling coefficient correction error obtained in process (3), and these are again defined as a deviation coupling coefficient correction error and an argument coupling coefficient correction error.

Next, using the error propagation learning rule (back propagation), the deviation coupling coefficient correction error obtained earlier,
The amount of correction of the neural network's coupling coefficient is determined by the argument coupling coefficient correction error, and outputs based on the changed coupling coefficients, that is, gains Kp and Ka are output.

However, if the total deviation error of the four simulation results is larger than the total deviation error of the simulation result due to the gain output from the output layer, the deviation coupling coefficient correction error and the argument coupling coefficient correction error are Both are set to 0.

(5) During a certain number of learning times, if the difference between the gains obtained by the K-th learning and the K+1-th learning is less than a certain set value, it is assumed that the gain has converged and the gain value at that time is used as the optimal gain. shall be. Until then (2)
Repeat the process from (4).

However, during process execution, the gains Kp,
When Ka is too large and the head angle becomes very large and saturated, the total deviation error becomes constant and the gain is hardly corrected because the movable range of the head angle is ±1.5 degrees. At that time, the head angle is within the movable range ±
A deviation coupling coefficient correction error and an argument coupling coefficient correction error are provided to reduce the gains Kp and Ka to within 1.5 degrees. FIG. 10 shows a block diagram of autotuning. The learning rule for this neural network is the backpropagation learning rule.

Next, an example of actual auto-tuning will be shown. In the following simulation, the input to the neural network was set to an initial deviation of 50 mm and an initial deviation of 0 degrees. When using the Sayama district simulation model, the gain Kp output from the output layer during neural learning is
FIG. 11 shows how the total deviation error calculated by simulation using Ka and these gains changes depending on the number of learning times. Looking at this figure, it can be seen that the gains Kp and Ka at the initial stage of learning fluctuate considerably, and the total deviation error is steadily decreasing. Furthermore, the gain is almost stable after about 600 learning times, and it can be said that it has converged without much change even after repeated learning times. Furthermore, the simulation results during learning in the Okayama area are shown in FIGS. 12 to 17. From these figures, it can be seen that as learning progresses, a gain that provides better control is output. Figure 18 shows the results of a simulation with an initial position of 50 mm using the Okayama area model using the weights after learning.
Shown below. It can be seen that after learning, the gain for optimal control is output.

FIG. 20 shows an initial deviation of 50 mm and an initial deviation angle of 0 d.
eg, the optimum gains Kp, Ka, and the total deviation error obtained by changing the respective gains in small steps, and the optimum gains Kp, Ka, and the total deviation error, obtained by the neural network of the present invention. This result shows that the neural network obtains almost the optimal gain through learning. In previous neural learning, an initial argument of 0deg was used as an input, but the system learns well even when the initial argument is other than 0. Figure 19 shows the simulation results of the model in the Okayama area when the initial deviation after learning is 500 mm and the initial declination angle is 1 degree. As you can see from the figure, it is well controlled.

[0024]

[Effects of the Invention] As explained above, according to the present invention, the neural network whose input is the initial deviation and declination, and whose output is the deviation and declination feedback gain is automatically optimized by learning using direction control simulation. Since a gain can be obtained, this method requires less effort and time compared to the conventional method and is much shorter than the conventional method.

[Brief explanation of the drawing]

FIG. 1 is a configuration diagram showing the system configuration of a tunnel robot.

FIG. 2 is an explanatory diagram for explaining definitions of a head angle and a pitching angle change amount.

FIG. 3 is an explanatory diagram for explaining the definition of each parameter.

FIG. 4 is a block diagram of direction control.

FIG. 5 is a characteristic diagram showing the deviation absolute value integral characteristic of Ka-transient response when Kp is fixed at 0.01 (deg/mm).

FIG. 6 is a characteristic diagram showing Kp-minimum deviation absolute value integral value characteristics.

FIG. 7 is a characteristic diagram showing directional control simulation results when using optimal Ka and Kp.

FIG. 8 is a configuration diagram of a neural network.

FIG. 9 is a flowchart of optimal gain autotuning.

FIG. 10 is a block diagram of autotuning.

FIG. 11 is a characteristic diagram showing the relationship between the total deviation error and the number of learning times.

FIG. 12 is a characteristic diagram showing simulation results during learning in the Okayama area with 0 learning times.

FIG. 13 is a characteristic diagram showing simulation results during learning in the Okayama area with 10 learning times.

FIG. 14 is a characteristic diagram showing simulation results during learning in the Okayama area with 20 learnings.

FIG. 15 is a characteristic diagram showing simulation results during learning in the Okayama area with 30 learnings.

FIG. 16 is a characteristic diagram showing simulation results during learning in the Okayama area with 40 learnings.

FIG. 17 is a characteristic diagram showing simulation results during learning in the Okayama area with 50 learnings.

FIG. 18 is a characteristic diagram showing simulation results at an initial position of 50 mm in a model for the Okayama area using weights after learning.

[Figure 19] Initial deviation after learning 500mm, initial deviation angle 1d
It is a characteristic diagram showing simulation results of a model in the Okayama area in the case of eg.

FIG. 20: Optimal gain Kp according to the conventional method and the method of the present invention
, Ka, and a comparison of total deviation errors.

[Explanation of symbols]

11... Tunnel robot main body, 12... Buried pipe, 13... Push pipe device, 14... Hydraulic device, 15... Operation panel.

Claims (1)

[Claims] Claim 1: An input layer, at least one intermediate layer, and an output layer, each layer being a set of units,
The initial position deviation and the initial pitching angle deviation are input to the input layer, the position deviation feedback gain Kp and the pitching angle deviation Ka are output to the output layer, and the units in the adjacent layers are all connected, and each connection has its own characteristics. The output value of the previous layer unit is multiplied by this weight and the value is input to the subsequent layer unit, which calculates the sum of the values from all units of the previous layer and defines it as an output function. Using a neural network that outputs the value to the next layer after applying nonlinear transformation to A neural-type optimal gain auto-tuning method for a small-diameter tunnel robot, characterized by obtaining an optimal position deviation feedback gain Kp and an optimal pitching angle deviation Ka according to the law. Process (1) Input arbitrary data for the minimum deviation and minimum declination as initial settings. (2) Enter the set initial position deviation and initial pitching angle deviation into the input layer of the neural network, input the gain output from the output layer into a direction control simulator using a feedback control law, perform a simulation, and then Calculate the total positional deviation between the planned line and the robot body during the stroke, and the pitching angle deviation. This is defined as the total deviation error and total declination error. Further, the total deviation error and total declination error when the total deviation error becomes the minimum during learning are stored as the minimum total deviation and minimum total declination, respectively. (3) Calculate the absolute value of the difference between the total deviation error and the minimum total deviation, and between the total argument error and the minimum total argument, and use these as the deviation coupling coefficient correction error and the argument coupling coefficient correction error. If these values are large, divide them by a certain constant. (4) Next, the gains Kp and Ka output from the output layer
Gain (Kp ±α, Ka ±
β) to simulate a small diameter tunnel robot 4
The gain (Kp
±α, Ka ±β) are added and subtracted by the process (
A sign is assigned to each coupling coefficient correction error obtained in step 3), and these are again referred to as a deviation coupling coefficient correction error and an argument coupling coefficient correction error. Next, using the error propagation learning rule (back propagation), the amount of correction of the neural network's coupling coefficient is determined by the deviation coupling coefficient correction error and the deviation coupling coefficient correction error obtained earlier, and the output based on the changed coupling coefficient,
That is, gains Kp and Ka are produced. However, if the total deviation error of the four simulation results is larger than the total deviation error of the simulation result due to the gain output from the output layer, both the deviation coupling coefficient correction error and the argument coupling coefficient correction error are 0. do. (5) If the difference between the gains obtained by the K-th learning and the K+1-th learning is less than or equal to a certain set value during a certain number of learning times, it is assumed that convergence has been achieved, and the gain value at that time is set as the optimal gain. Until then, processes (2) to (4) are repeated. However, during process execution, if the gains Kp and Ka are too large and the head angle becomes very large and saturated, the movable range of the head angle will be ±1.5.
Because of this, the total deviation error becomes constant and there are times when the gain is hardly corrected. At that time, the gains Kp and Ka are set so that the head angle is within the movable range of ±1.5 degrees.
Provides a deviation coupling coefficient correction error and an argument coupling correction error that reduce the deviation.