JPH0293708A - Controller - Google Patents

Abstract

PURPOSE:To improve the processing speed, ecomization, simplification and reliability of a control device by using an analog electric circuit (neuro- controller) simulating a neural circuit as the control device. CONSTITUTION:A state value (c) of a subject 204 to be controlled is guided to a comparator 205 and an error value (e) with a setting value (r) is inputted to a control device 202. A device 203 is composed of neuro-controllers 102-107 and the signal (e) is passed through plural delay devices 102 and guided through a sampler 103 to input distributors (a1-a8) 104. When the delay time of the delay device 102 is wholly defined as a DELTAt, the error value signal of a present time t0 is inputted to the distributor a1, the error value signal of a time t-1 is inputted to the a1 and the error value signal of a time t-(i-1) is inputted to the ai. Each reading is executed in a constant period. The reading value of the ai is multiplied by weight wi and guided to an input adding and output device 104 and an output value is integrated by an integrator 107. Then, an integrated result is outputted as a control value (h). Thus, an actuator 206 controls a physical input to the subject 204.

Description

[Detailed Description of the Invention] [Object of the Invention] (Industrial Application Field) The present invention relates to individual devices controlled by actuators that operate using electric motors, electromagnetic force, pneumatic pressure, hydraulic pressure, water pressure, etc. It is used in a control device applied to a plant system constituted by such equipment.

(Prior Art) This type of control device has conventionally been realized using an analog computing device or a digital computing device (for example, a computer) having functions such as proportionality, integration, and differentiation.

(Problems to be Solved by the Invention) An object of the present invention is to make a control device that satisfies almost the same specifications regarding input and output as conventional control devices, faster, more economical, simpler, and more reliable. It is.

[Structure of the invention]

(Means for Solving the Problems) FIG. 2 shows a general connection relationship between the control device 202 and the controlled object 204. The state value C of the controlled object 204 is guided to the comparator 205 via a line 203, and the difference e from the manually input setting value r is guided to the control device 202 via the line 201.
Enter.

(Operation) The control device 202 calculates the control value and converts it to the line 108.
is input to the actuator 206 by. Actuator 206 controls physical inputs (not shown in FIG. 2) to the controlled object.

In FIG. 2, the problem of the present invention is solved by realizing the control table vg, 202 by an analog electric circuit simulating a neural circuit (such a control device is hereinafter referred to as a neurocontroller). do.

(Example) The principle of the neurocontroller will be explained using FIG. 1. In Fig. 1, the symbols 102, 103, 104°1
The part consisting of elements 05.106 and up to 107 is the neurocontroller. The error value signal e shown in FIG. 2 is input from line 101. This signal passes through a plurality of delay devices 102 and from the output of each delay device to the sampler 102.
03 to the input distributor Cas~a11. ) 104. This signal is simultaneously guided to the learning device [110] and the model control device 111. The number of delay devices 102, their delay times, and the number of input distributors 104 are different from each other in the controlled object 20.
It is determined as appropriate depending on the properties of 4.

Assume now that the delay times of the delay device 102 are all Δt. In Figures 1 and 3, if the current time is to and time 1-o is n·Δt before the current time, the current error value signal is input to the input distributor a□, and the input distributor a□ is input to the input distributor a□. The error value signal of -□ is input. Similarly, input distributor a
An error value signal of time (-N, -) is input to l.

FIG. 3 shows an example of the correspondence among the differential values of the set value signal, error value signal, and control value signal at each time.

Reading to the input distributor is performed every T at a constant period.

The value of 6 guns is determined as appropriate depending on the properties of the controlled object 204. The values read by the distributor are weighted by a weighter 105, as shown in FIG. Example sentence - ba, al
The value read by is the weight V of the weighter at the exit of al
The resultant signal is multiplied by 4 and sent to the input adder and output unit 104. All of the input signals that have entered the input addition and output device are cumulatively added and output.

In FIG. 1, the input adder and output device bt+bz+ba+
A weighting device is further provided at the exit of b4, leading to the input addition and output device C1. An integrator 107 integrates the input addition and output values of the output device C, and the result is used as a control value.

In FIG. 1, a set of input distributors aila21...a,
Input adder and output device, b2. The set of bl, b, and the set of input adders and output units C are respectively called layers. FIG. 1 happens to show a three-layer neurocontroller. This number of layers can be arbitrarily selected. A four-layer structure often exhibits the best characteristics.

The wiring paths between layers are determined according to the characteristics of the applied control object, and do not necessarily have to be as shown in FIG. Generally, the layers are completely connected (each device 104 or 106 in one layer is connected one-to-one with all the devices 104 or 106 in the adjacent layer) as shown in FIG. As a result of learning, a method is adopted in which, as a result of learning, the weights corresponding to unnecessary partial connections automatically become 0. In addition, a neurocontroller in which each device 104 or 106 in a layer is completely connected to each other in addition to one layer can also be adopted.

In FIG. 1, the input adder and the output from the output device C□ pass through the mode switch 109, and in the normal mode, the output is
08 and is input to the actuator 206 in FIG. When the mode switch 109 is set to the learning mode, this path is cut off, and the output of the model control device 111 is input to the 7-couple unit 206 in FIG. In the learning mode, the learning device changes the weight values of the weighter.

The line 101 in FIG. 1 is constantly given an error value signal e. The sampler 103 is operated at each cycle ΔT described above.

While the 6 circuits are closed and opened as shown at time 401 in FIG. 4, the input distributor reads the values of the error value signal shown in FIG. The output from output device c1 is calculated at time 402 in FIG. This output is the time t of the control value signal. This corresponds to the value in .

When the mode switch 109 is in the normal mode, this output is directly applied to the controlled object via the actuator 206. When mode switch 109 is in learning mode, this output is given to learning device 110.

In the case of learning mode, it works as follows.

The model control device 111 receives the same error value signal e from the line 101.
and calculate an exemplary control output value.

This device is preconfigured to calculate optimal control values that optimize the evaluation function set by the user. The learning device starts reading the literature (Rumelhart, D.E., J.L., McC.
co-authored by l-elland, rParallel Dist
RibutedProcessingJVol, 1.
pp, 318-362. The MIT Pres.
The algorithm for calculating six weights is to calculate the weights of all weighters 105 by a method called back propagation, which is known by J. S., 1986), and to reset each weight. Although there are several known calculation methods other than the calculation method, the learning device 110 can use any calculation method.

The operations described above are repeated every time ΔT, as shown in FIG.

After the learning is completed, the mode switch 27 is set to normal mode.
When relearning is not necessary at all, the learning device 110 and the model control device 111 may be removed after learning is completed. In this case, the output of the input adder and output device C2 is directly connected to the actuator 206 through the integrator 107.

(Other examples) (a) Other Example 1: Another example is shown in FIG.

In the embodiment described above, the mode switch was considered to be switched manually, but in the embodiment shown in FIG. 5, the mode switch is automatically switched in one period according to the time shown at 402 in FIG. That is, when the neurocontroller operation is ON, the mode is switched to the normal mode, and when the neurocontroller operation is OFF, the mode is switched to the learning mode. When switching to the learning mode, the output is held by the sample value holder 501 and the held value is applied to the actuator 206.

A feature of this embodiment is that it is possible to learn online, that is, to learn while the neurocontroller is in operation.

(b) Other Embodiment 2: The part consisting of elements 102 to 107 shown in FIG. 1 or 5, that is, the part of the neurocontroller, is realized by a microprocessor, and the neurocontroller described above is implemented. This function is implemented by software. In this case, the microprocessor reads the error value signal every time Δt (corresponding to the delay time), stores it, and
The same processing as described above is performed for each time. In this example,
The times Δt and ΔT can be freely changed by software depending on the characteristics and operating conditions of the controlled object during operation or when stopped.

In addition to this, a similar effect can be obtained.

■ Conventional control devices, such as PID (proportional, integral,
It can be realized with a simpler structure than a differential controller or a digital controller using a microprocessor. For this reason, depending on the field of application, high speed, low cost,
High reliability can be expected.

■ By changing the weight values, completely different control characteristics can be provided, so there is high flexibility in application, ease of specification changes, and high possibility of standardization.

〔Effect of the invention〕

In applications where it is possible to create some kind of relationship between the error value signal waveform and the corresponding ideal control value signal waveform, by implementing the present invention, it is possible to create an exemplary control device. Control with equivalent characteristics can be implemented.

Exemplary controllers herein do not necessarily refer to existing controllers. It also includes control that goes beyond existing physical control devices (control devices designed based on control theory), such as control performed intuitively by human hands (for example, steering). By using the present invention, it is possible to make a neurocontroller learn human-like or sensory control that cannot be achieved with a physical control device, thereby realizing human-like control.

[Brief explanation of drawings]

FIG. 1 is a diagram showing the details of the neurocontroller, which is the subject of the present invention, and its relationship with the learning device, model control device, and mode switch used for learning, and FIG. 2 shows the neurocontroller of FIG. 1, Figure 3 is a diagram showing the relationship between the input, output, and the controlled object. Figure 3 is a diagram showing the relationship between the waveform input to the neurocontroller and the waveform that will be output from it. Figure 4 is a diagram showing the relationship between the waveform input to the neurocontroller and the waveform output from the neurocontroller. FIG. 5 is a diagram showing the interrelationship between input/output, internal calculation, and learning time, and is a diagram showing another embodiment. Agent Patent Attorney Nori Ken Yudo Daishimaru Ken Figure 2 Figure 1 Time

Claims (2)

[Claims] (1) A control device that is installed to maintain a controlled object, such as a device, equipment, or industrial process, in a desired operating state by controlling the input of an arbitrary physical quantity. Then, the time-series value of the deviation between the measured value representing the operating state of the controlled object and the desired operating value given in advance is continuously obtained at each sampling interval for the operating device, and the time-series value of the obtained deviation is obtained. Each time we sample, we divide it into multiple groups along the flow of time, add the value obtained by multiplying the time-series value by the weighting coefficient for each group, and then add the value that has stopped for each group again along the flow of time. Divide the group into new groups, perform weight addition for each group, repeat this grouping and weight addition until there is only one group, and calculate the weight calculated for the last group. A control device that uses an additional value to control input to a controlled object at each sampling time. (2) While a controlled object such as a device, equipment, or industrial process that operates by inputting an arbitrary physical quantity is in operation, the output of the control device, that is, the final weighted addition value, is sent to the controlled object. Either disconnect it from the input section or keep it as it is, and if it is disconnected, connect the same input as the control device to the newly installed model control device, and connect its output to the input section of the controlled object to connect it to the input section of the controlled object. The operation of the controlled object is continued, while the input and output of the control device are newly connected to a learning device, and at each sampling time, the learning device connects the input of the control device, the output of the control device, and the model control device. From the time-series values of the deviation from the output, each of the weight values in the control device is calculated using an algorithm such as back propagation in neural network simulation calculation, and the newly calculated weight values are used to calculate the weight values in the control device. A system configured to update historical weight values for.