JPH07187072A - Automatic control method for rov - Google Patents

Abstract

PURPOSE:To predict a periodical disturbance and properly control ROV on the basis of the prediction by generating an error signal on the basis of an error between the state of ROV and a target value for adjusting the control parameter of a neuro controller. CONSTITUTION:The diving state of a rope-operated unmanned diving vehicle (ROV) 20 such as a depth is detected with sensor and, then, the first manipulated variable UPID is determined with a PID controller 21 on the basis of the error of an output value from the sensor. Thereafter, the first manipulated variable UPID is used to operate ROV 20 and the motion characteristics thereof in a disturbance due to the operation is learned via a neuro prediction model 22, thereby modelling the characteristics. Then, an error signal is calculated with the model 22 on the basis of a deviation between the speed and acceleration of ROV and the target values. Furthermore, the control parameter of a neuro controller 23 is adjusted on the basis of the error signal, thereby determining the second manipulated variable Unn. Also, the first and second manipulated variables UPID and Unn are summed up for use as the final manipulated variable of ROV. As a result, ROV can be properly controlled even in a disturbance.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an automatic control method for an ROV (roped unmanned submersible), and more particularly to an automatic ROV control method capable of appropriately controlling the position of the ROV against periodic disturbances and the like. It is a thing.

[0002]

2. Description of the Related Art ROV (Remotely Operated Vechicle)
Is a cable-operated unmanned submersible used for underwater work and surveys.

Normally, the ROV is manually operated by an operator on the mother ship, but it is difficult to operate because it is operated remotely. In order to reduce the operator's burden on the operation of the ROV, an automatic control method based on the classical control theory is installed.

In this conventional automatic control method, it is necessary to predict the environment encountered by the ROV and adjust the parameters.

As the automatic control method, the PI shown in FIG. 8 is used.
D There is an automatic control method, for example, the target value depth is PI
When input to the D controller 40, the error between the current depth and the target depth from the ROV 42 is calculated, and an appropriate coefficient a
Is a manipulated variable 1, a value obtained by multiplying the current ROV speed by an appropriate coefficient b is a manipulated variable 2, and the target depth and the ROV depth at each time from the time when the control is started to the present. The value obtained by multiplying the error of (1) by an appropriate coefficient c from the control start time to the present is taken as the manipulated variable 3, and these manipulated variables 1, 2, 3 are added together to obtain the manipulated variable after a minute time from the present. PID control is performed.

[0006]

However, since this control determines the manipulated variable from the past ROV state,
ROV environment underwater (environment subject to wave force, tidal current, etc.), especially when a periodic disturbance is applied
Cannot be controlled to a constant depth. That is, the environment that the ROV encounters in water is unknown, and in the conventional automatic control method, the coefficients a, b, and c are set so that the ROV is under a certain condition, and a periodic disturbance is generated. With the accompanying control, it is difficult to properly set the coefficients a, b, and c, and it is also difficult to adjust the coefficients. Even if the coefficients a, b, and c are appropriately adjusted for a certain disturbance,
Disturbances in the natural environment change in various ways, and it is very difficult to accurately predict them. Therefore, it has been difficult to construct a control method adapted to all environments by the conventional automatic control method.

Therefore, an object of the present invention is to solve the above problems, predict a periodic disturbance, and then based on the prediction, ROV.
An object of the present invention is to provide an automatic ROV control method capable of appropriately controlling the above.

[0008]

In order to achieve the above object, the present invention provides a method for automatically controlling a ROV equipped with a thruster for diving underwater, in which a sensor detects a diving state such as the depth of the ROV. , The first operation amount is determined by the PID controller from the error of the output value of the sensor with respect to the target value such as depth, and the RO is determined by the first operation amount.
V is operated, ROV motion characteristics in disturbance based on the operation are learned and modeled by a neuro prediction model, and then an error signal between the actual ROV velocity and acceleration and a target value is calculated, and the error signal is calculated. A second operation amount is determined by adjusting the control parameter of the neuro controller by the neural network, and the first operation amount and the second operation amount are added to obtain the final operation amount of the ROV. , The neuro controller receives a waveform modeling the disturbance from the oscillator network, and after the neuro prediction model learns the ROV motion characteristics during the disturbance, the second manipulated variable is determined by the waveform of the oscillator network. It was done like this.

[0009]

According to the above construction, first, the motion characteristic of the ROV in the disturbance is determined by the deviation of the speed and the acceleration which are the outputs of the prediction model based on the control by the PID controller and the actual ROV state. The predictive model learns, and the predictive model is used to generate an error signal from the error between the state of the ROV and the target value, whereby the neuro controller learns and adjusts its control parameter, thereby determining the ROV for the disturbance. It can be controlled accurately. The oscillator network inputs the disturbance into the neuro controller as it learns.

[0010]

An embodiment of the present invention will be described in detail below with reference to the accompanying drawings.

First, the structure of an ROV (roped unmanned submersible vehicle) according to the present invention will be described with reference to FIG. In FIG.
(A) shows a side view and (b) shows a bottom view.

Although the submersible body 10 is omitted in the drawing,
The cable from the mother ship is connected. The submersible body 10 includes three thrusters 1 that propel the body 10 in the vertical direction.
1 and two thrusters 12 for propelling forward and backward and a thruster 13 for changing the direction to the left and right are provided, and these are respectively driven by a forward / reverse rotation motor 14.

At the upper front side of the submersible body 10, the body 10
A depth sensor 15 is provided to detect the depth of the. Further, the submersible body 10 is provided with a pinger 16 that emits a sound in the water, and the sound produced by the pinger 16 is measured at three points on the mother ship side on the sea to detect the position of the submersible body 10.

The submersible body 10 is equipped with, for example, a TV camera, and various underwater operations can be performed by remote control on the mother ship side while watching the TV camera.

When this underwater work is performed in a shallow water area, the diving vehicle body 10 is subject to periodic up-and-down movements due to the rocking of waves and lateral and lateral rolls as disturbances. Need to be kept stationary during work.

An automatic control method for the submersible body 10 will be described.

First, in order to keep the submersible body 10 stationary with respect to external disturbances, the thrusters 1 are so arranged as to cancel the external disturbances.
Although it is sufficient to control the forward and reverse rotation directions of 1 to 13 and the number of rotations thereof, it is difficult for the classical control method to predict these disturbances and automatically adjust and control the system parameters. Therefore, the present invention is a control system using a control system based on a neural network having a function (learning function) of automatically adjusting system parameters with respect to a target output and a control system based on classical control theory in parallel. Adaptive Neural-net Controlle
r) is applied to motion control of ROV to configure an automatic control system.

FIG. 1 is a block diagram of the control method of the present invention, in which 20 is an ROV to be controlled and 21 is a PI.
D controller, 22 is a neuro prediction model, 23 is a neuro controller, 24 is an oscillator network,
25, 26 and 27 are adders, and 28 is disturbance.

First, the PID controller 21 receives the deviation signal s between the output y of the ROV 20 added to the adder 25 and the target value o, and uses the deviation and the differentiation and integration of the deviation to generate a first manipulated variable. The _UPID is output to the adder 25.

The neuro prediction model 22 is a neural network that models the motion characteristics of the ROV in the environment (disturbance) encountered by the ROV 20. This network has the same output (speed and acceleration) as the ROV response when the same input as the operation amount given to the ROV is given.
An error signal for adjusting the internal parameters so as to operate is generated from the deviation of the ROV response and the target value using this network. That is, the neuro prediction model 2
2 is the first operation amount U _PID of the PID controller 21.
And the second manipulated variable U _nn of the neuro controller 21 are added by the adder 26, the final manipulated variable U _o is input, and the output y of the ROV 20 is input. First, the response of the ROV 20 to the first manipulated variable _PID is input. RO with learning
The response of the V20 in the disturbance 28 is predicted, and the velocity and the acceleration based on the prediction model are output to the adder 27.
The ROV in the disturbance is adjusted by adjusting the coefficient inside the prediction model based on the deviation of the actual ROV 20 and the acceleration door input to
Model (learn) the motion characteristics of. The prediction model produces an error signal from the deviation between the actual ROV velocity and acceleration and the target value.

The neuro-controller 23 is a neural network that produces a controller that can optimally control the ROV in the environment encountered by the ROV. Using the error signal E generated by the neuro prediction model 22, the internal parameters are automatically adjusted (learned), and the optimum controller for controlling the ROV thrusters 11 to 13 is generated.

Like the neuro prediction model 22, the oscillator network 24 learns the disturbance 28 in advance and also learns the frequency component of the disturbance 28, and R
When the OV 20 is optimally controlled, the learned frequency component is generated and output 29 to the neuro controller 23.

The basic control operation will be described below in order.

(1) Control by PID controller 21 Since minute and random values are given as initial values to the parameters of the neuro prediction model network and the controller network, each network has only a minute output before learning.

The operation amount of the ROV 20 to the thrusters 11 to 13 is a value (final operation amount) obtained by adding the output of the controller network (second operation amount) and the output of the PID controller (first operation amount). . Therefore, the PID controller 21 controls the ROV 20 before the controller network has sufficiently learned.

(2) Neuro Prediction Model 22 While the PID controller 21 is controlling the ROV 20 in the disturbance, the prediction model network has a relationship between the manipulated variable to the ROV thruster and the ROV response (that is, ROV depth, velocity, acceleration). Etc.) to learn.

(3) Learning of controller network The controller network learns by the error signal E generated after the predictive model network has sufficiently learned the motion characteristics and disturbance of the ROV, and automatically generates the optimum controller.

(4) Control by the controller network 23 When the controller network automatically generates the optimum controller by learning, the value fed back is attenuated. For this reason, the output of the PID controller is attenuated, and the operation amount to the ROV thruster is the sum of the controller network and the PID controller, so that the ROV is finally controlled only by the controller network.

Next, the learning of the neural network will be described in more detail.

The neural network is one of the AI technologies and is an information processing system that digitizes the functions of the neural network of the brain.

The basics of this neural network are shown in FIG.
Will be described.

As shown in FIGS. 3 (a) and 3 (b), a large number of processing elements 30 are combined to form a structure.
Here, a PD in which processing elements are arranged in layers
A P (Parallel Distributed Processing) type neural network was used.

This processing element is represented by
This is an element that performs the processing shown in Expression 2.

[0034]

[Equation 1]

[0035]

[Equation 2]

However, i represents the processing element number of the layer currently under consideration, and j represents the previous processing element number. Further, o represents the output of the processing element, and θ is a variable called offset. f is an arbitrary continuous function, and a sigmoid function is usually used.

The output from each processing element is multiplied by the variable w _ij as shown in the equation 1 to be input to another processing element. This variable w _ij is called a weight. This weight is a parameter to be adjusted inside the neural network.

When an input value is given to the leftmost input layer in the network of FIG. 3, the above processing is sequentially performed through the intermediate layer to the right output layer.

Learning of the neural network (adjustment of weight) is performed by an error backpropagation learning algorithm. The error E between the output o _k from the network and the value (teaching data) Y _k desired to be output to the original network is evaluated by the following equation.

[0040]

[Equation 3]

The weight update amount Δ is set so that this error is reduced.
w _ij is determined by the equation 4 and the weight is updated. (Steepest descent method)

[0042]

[Equation 4]

Here, η is called a learning coefficient, and is a coefficient that determines the update amount of the weight.

Now, referring to FIG. 4, an explanation will be given of the ural network which constitutes the system shown in FIG.

The neuro prediction model network is shown in FIG.
As shown in (a), the first operation amount U _PID of the PID controller 21, the second operation amount U _nn of the neuro controller 23, and position information Zt such as the depth of the ROV 20 are input, and based on these inputs. An error signal is generated to train the neuro controller network of FIG. 4 (b).

Here, the processing performed by the neuro prediction model network to generate the error signal can be obtained by expanding the partial differential term of the above-mentioned expression 4 as shown in the expression 5.

[0047]

[Equation 5]

Here, u is a value obtained by adding u _PID and u _nn .

In Equation 5, the calculation to be performed by the neuro prediction model network is a partial differentiation of o _k with u. This part is called Jacobian.

The part in the neuro prediction model network necessary for calculating the Jacobian is the part that combines the input of the manipulated variable into the ROV thruster and the output of the neuro prediction model network. In order to sufficiently learn the weights of this part to generate a correct error signal, it is necessary to learn that the weights of this part are sufficiently larger than the coupling from the depth to the output. Moreover, it is necessary to configure the neuro prediction model network so that the Jacobian can be obtained sufficiently large.

The output of the neuro prediction model network in FIG. 4 (a) is two, that is, the velocity and the acceleration, in order to learn the neuro prediction model network so that a large Jacobian can be obtained. The basis is described below.

Now, if the transfer function relating to the depth of the ROV is linearized and obtained, it can be expressed as shown in equation 6.

[0053]

[Equation 6]

A Bode diagram of this transfer function is shown in FIG.
Here, K = 0.5 and T = 20.8. Also ROV
FIG. 6 also shows the transfer functions relating to the velocity and acceleration.

From FIGS. 5 and 6, it can be seen that it is more advantageous to learn the predictive model network by the velocity in the low frequency region and the acceleration in the high frequency region in order to obtain a large Jacobian. Here, if the input and output phases are synchronized, the Jacobian is positive and maximum, and if inverted, the Jacobian is negative and minimum. If the phases are shifted by π / 2, the average Jacobian value becomes zero.

This is because the Jacobian becomes 0 when the phase is π / 2 in the Bode diagram, increases to + as the phase decreases, and becomes maximum when the phase is 0. When the phase decreases from 0 to −, the Jacobian starts to decrease, and becomes 0 again when the phase is −π / 2. As the phase is further reduced, the Jacobian decreases to-and becomes minimum at -π.

As described above, since the Jacobian has the property of changing periodically with respect to the phase, the velocity in the low frequency range
In the high frequency range, only Jacobian of depth (position) and acceleration is generated. Further, since the Jacobian is a partial differentiation of the output of the prediction model by the operation amount to the prediction model, the gain is effective in addition to the phase.

However, considering the gain, it can be seen that the acceleration becomes constant in the high frequency range, whereas the gain decreases as the frequency becomes higher in the depth. For this reason, it is advantageous to use the acceleration rather than the depth in the high frequency region in order to obtain a correct Jacobian.

A feature of the neurocontroller network is that it contains oscillator elements in addition to the depth fed back as an input. The reason why the oscillator must be included in the input is that the neural network having the layer structure cannot automatically generate the periodic output, and therefore the oscillator network 24 inputs the periodic component of the disturbance to the controller 21.

Next, a ROV model (total length 1.5 m, width 0.8 m, height 0.8 m) was formed by the automatic control method of the present invention shown in FIG. 1 and the conventional control method described in FIG. The experimental results of controlling the will be described.

This ROV model has an aerial weight of 403 kg.
So, it becomes almost neutral buoyancy in water. The model has 6 thrusters as described in FIG. 2, but only 3 vertical thrusters were used in the experiment, and the test conditions were a water tank with a depth of 118 mm, ± A wave of 50 mm was generated and the ROV model was controlled so that the target depth was 0.8 m. The coefficients of the PID controllers of both systems are proportional coefficient kp = 10.2 and differential coefficient kd =
0.195 and an integration coefficient ki = 1.28.

FIG. 7 shows the result of the automatic control method of the present invention, which is shown in FIG.
Is a graph showing the result of the conventional system.

First, referring to FIG. 9 of the conventional system,
(A) is the output (operation amount) of the PID controller Upid
Change, (b) model ROV depth change, (c) model R
OV velocity change, (d) shows the water surface change of the water tank.

Similarly, in FIG. 7 of the system of the present invention, (a) changes the output (first operation amount) of the PID controller Upid, (b) changes the depth of the model ROV,
(C) changes in speed of the model ROV, (d) changes in water level of the water tank, (e) changes in operation amount of the neuro controller (second
Operation amount) Unn.

From FIG. 9, the depth of the ROV model in the regular wave is set to the target value (0.8 m) with only the PID controller.
It can be seen that the depth variation of the ROV model is larger than the wave height.

On the other hand, in the system of the present invention, as shown in (e), after the learning amount of the neuro controller is set to about zero for about 18 seconds, the second manipulation amount Unn is gradually changed. It can be seen that the output UPID of the PID controller shown in (a) gradually decreases and the manipulated variable becomes zero in about 120 seconds, and the depth change of the ROV model also disappears as shown in (b).

[0067]

In summary, according to the present invention, the conventional P
In the ID control, it was difficult to control the ROV to be constant with respect to the disturbance, but the control system using the PID and the neural network of the present invention in parallel enables the ROV to be accurately controlled even during the disturbance. It becomes possible.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention.

2A and 2B show details of an ROV in the present invention, in which FIG. 2A is a side view and FIG. 2B is a plan view.

FIG. 3 is a diagram showing details of a neural network and a processing element.

FIG. 4 is a diagram showing a configuration of the neural network of FIG.

FIG. 5 is a diagram showing a Bode diagram of phase with respect to frequency in the control method according to the present invention.

FIG. 6 is a diagram showing a Bode diagram of gain with respect to frequency in the control method according to the present invention.

FIG. 7 is a diagram showing experimental results using the control method of the present invention.

FIG. 8 is a block diagram of a conventional control method.

FIG. 9 is a diagram showing an experimental result using a conventional control method.

[Explanation of symbols]

 20 ROV 21 PID controller 22 Neuro prediction model 23 Neuro controller 24 Oscillator network

Claims (2)

[Claims] 1. An automatic control method for a ROV equipped with a thruster for diving underwater, etc., wherein a sensor detects a diving state such as the depth of the ROV, and an error of the output value of the sensor with respect to a target value such as the depth is detected. The PID controller determines the first manipulated variable and R
The OV is operated, and the ROV motion characteristics in the disturbance based on the operation are learned and modeled by the neuro prediction model, and then the error signal between the actual ROV velocity and acceleration and the target value is calculated, and the error signal is calculated. It is characterized in that the second operation amount is determined by adjusting the control parameter of the neuro controller by the neural network, and the first operation amount and the second operation amount are added to obtain the final operation amount of the ROV. ROV automatic control method. 2. The neuro controller receives a waveform modeling a disturbance from the oscillator network, and after the neuro prediction model learns the motion characteristics of the ROV in the disturbance, the waveform of the oscillator network becomes a second waveform. The automatic ROV control method according to claim 1, wherein an operation amount is determined.