阅读说明：本技术 控制装置 (Control device ) 是由 井崎胜敏 桥本诚司 于 2019-04-26 设计创作，主要内容包括：本发明提供一种即便对于死区时间系统,神经网络也能不受死区时间的影响而学习,并且具有改善对于指令输入的过渡特性的能力的控制装置。控制装置(1)具备：对包含死区时间单元的控制对象进行控制的反馈控制器(10)；包含死区时间单元并输出针对输入的期望的响应波形的规范模型部(20)；以及输出与反馈控制器的输出相加并输入控制对象的学习型控制器(30),以使控制对象的输出与规范模型部的输出的误差由于来自学习型控制器的输出的变化而成为最小或预定的阈值以下的方式进行学习。(The invention provides a control device which can learn a neural network without being affected by a dead time even in a dead time system and has the capability of improving transient characteristics for command input. A control device (1) is provided with: a feedback controller (10) for controlling a control target including a dead time unit; a normative model unit (20) that includes a dead time unit and outputs a desired response waveform for an input; and a learning controller (30) for adding the output of the feedback controller to the output of the control object and inputting the sum to the control object, wherein the learning controller learns that the error between the output of the control object and the output of the normative model part is minimum or less than a predetermined threshold value due to the change of the output from the learning controller.)

1. A control device is provided with:

a feedback controller that controls a control object including a dead time unit;

a normative model unit including a dead time unit and outputting an expected response waveform to an input; and

and a learning controller that adds an output of the feedback controller to an output of the control target and inputs the sum to the control target, and performs learning so that an error between the output of the control target and the output of the normative model unit becomes a minimum or a predetermined threshold value or less due to a change in the output from the learning controller.

2. The control apparatus according to claim 1, wherein the learning-type controller is a neural network controller that learns using a neural network.

3. The control device according to claim 2, wherein the neural network controller uses an error between the output of the control target and the output of the normative model unit as a teaching signal of a neural network, and performs learning by using the neural network so that the error becomes a minimum or a predetermined threshold or less.

4. The control device according to claim 1, wherein the dead time of the normative model section is set to be the same as or to the same degree as the dead time of the control target.

5. A control device applied to a control system for controlling a control target by a pre-designed feedback controller, the control device comprising:

a normative model unit including a dead time unit and outputting an expected response waveform to an input; and

and a learning controller that adds an output of the feedback controller to an output of the control target and inputs the sum to the control target, and performs learning so that an error between the output of the control target and the output of the normative model unit becomes a minimum or a predetermined threshold value or less due to a change in the output from the learning controller.

Technical Field

The present invention relates to a control device, and more particularly to a control device that controls a control target including a dead time.

Background

As a method of using a neural network for feedback control, a feedback error learning method and system using an inverse system of a control target are known. Fig. 2 shows a block diagram of the feedback error learning system. In this method, the neural network controller 110 performs learning with the output xc of the feedback controller as a teaching signal so that xc becomes 0 as the learning proceeds. Thus, learning and control are performed so that the error e becomes 0 and the output y becomes the target value yd. Thus, after learning, the controller used is transferred from the feedback controller 120 to the neural network controller 110. As a result, the control system 100 is replaced from the feedback structure to the feedforward structure.

In addition, as a method of introducing a normative model into a control system using a neural network, for example, the following method is disclosed. Patent document 1 discloses a control device in which an output of a feedback control unit and a normative model of a time-series data signal that outputs an ideal expected response based on a steering amount signal are input to a neural network unit. Patent document 2 discloses a structure in which a feedback controller itself is a neural network learning type controller. In addition, patent document 3 discloses a control device in which an estimation device is configured by a neural network having a nonlinear function approximation capability and incorporated as a compensator configuration unit.

Documents of the prior art

Patent document

Patent document 1: japanese laid-open patent publication No. H07-277286

Patent document 2: japanese laid-open patent publication No. H06-035510

Patent document 3: japanese patent laid-open publication No. 04-264602.

Disclosure of Invention

Problems to be solved by the invention

In the system shown in fig. 2, the responsiveness of the output response waveform of the repetitive step command may not improve for each step, that is, over time. This is considered to be because the neural network may not be able to learn in a state where there is no response (output from the control target) due to the dead time of the control target even if there is an input signal to the control target.

Here, in order to prevent the delay in learning of the neural network due to the delay in the output response caused by the dead time, a method may be considered in which a normative model that can obtain a desired response is used, the normative model is made to have the dead time, and the neural network is learned so that the actual output follows the output of the normative model. However, for example, the methods using the normative model as disclosed in patent documents 1 to 3 have the following problems.

First, the method disclosed in patent document 1 is basically the same as the conventional feedback error learning method, and even if the dead time is added to the normative model, the delay of the controlled object is further delayed. Thus, in the method disclosed in patent document 1, the learning delay cannot be improved.

In the method disclosed in patent document 2, a learning delay can be avoided if a dead time is included in the normative model. However, a model of the control object is required at an initial design stage of the neural network controller. Thus, the design of the controller is complicated and model errors may occur. In addition, it is necessary to compensate all the compensation targets such as target value response, disturbance, and fluctuation with a neural network controller. Therefore, it is difficult to design/adjust the controller for each compensation object, and it becomes complicated to correct the compensator by learning. The method disclosed in patent document 3 also has the same problem as patent document 2.

The above-described methods are control methods that focus on the following of the normative model in a system having no dead time or an influence of the dead time can be ignored, and do not focus on performance improvement such as transient characteristic improvement in consideration of the dead time. As a result, it is difficult to achieve both the transient response characteristic for the dead time system and further characteristic improvement of the learning effect by the neural network in the above-described methods.

In view of the above, an object of the present invention is to construct a control system that solves the above problems. Further, it is an object of the present invention to provide a control device capable of improving transient characteristics with respect to a command input, in which a neural network can be learned without being affected by a dead time even in a dead time system.

Means for solving the problems

According to the 1 st aspect of the present invention, there is provided a control device including:

a feedback controller that controls a control object including a dead time unit;

a normative model unit including a dead time unit and outputting an expected response waveform to an input; and

and a learning controller that adds an output of the feedback controller to an output of the control target and inputs the sum to the control target, and performs learning so that an error between the output of the control target and the output of the normative model unit becomes a minimum or a predetermined threshold value or less due to a change in the output from the learning controller.

According to the 2 nd aspect of the present invention, there is provided a control device applied to a control system for controlling a control target by a feedback controller designed in advance, the control device including:

a normative model unit including a dead time unit and outputting an expected response waveform to an input; and

and a learning controller that adds an output of the feedback controller to an output of the control target and inputs the sum to the control target, and performs learning so that an error between the output of the control target and the output of the normative model unit becomes a minimum or a predetermined threshold value or less due to a change in the output from the learning controller.

Effects of the invention

According to the present invention, it is possible to provide a control device that has a capability of learning a neural network without being affected by a dead time even in a dead time system and improving transient characteristics with respect to a command input.

Drawings

Fig. 1 is a block diagram of a control system according to the present embodiment.

Fig. 2 is a block diagram of a control system of a comparative example.

Fig. 3 is a repetitive step response waveform in the control system of the comparative example.

Fig. 4 is a comparison graph of overlapping of repetitive step response waveforms in the control system of the comparative example.

Fig. 5 is a repetitive step response waveform in the control system of the present embodiment.

Fig. 6 is a graph showing comparison of overlapping of repetitive step response waveforms in the control system according to the present embodiment.

Detailed Description

Hereinafter, embodiments of the present invention will be described with reference to the drawings.

< overview of the present embodiment >

First, an outline of the present embodiment will be described. The control system (control system) according to the present embodiment adopts a control method in which an output of a control target having a dead time such as a process control system is made to follow an output of a normative model also including the dead time by learning.

The Feedback (FB) controller can use an existing Feedback (FB) controller. The response of the control object is made to follow the output of the normative model with dead time. Therefore, in the neural network controller, an error between an output (actual output) of a control target and an output of a normative model is used as a teaching signal of the neural network, and the neural network is learned so as to minimize the error, for example. In addition, the output of the neural network controller and the output of the feedback controller are added to be used as an input to the control target, thereby controlling the control target.

Description of the present embodiment

Fig. 1 is a block diagram of a control system according to the present embodiment. The control system according to the present embodiment includes a control device 1 that controls a control object 2. The control device 1 includes a feedback controller 10, a normative model unit 20, and a neural network controller 30.

The feedback controller 10 controls the control object 2 according to a predetermined target value yd regarding the output of the control object 2. For example, the feedback controller 10 inputs a difference e between a predetermined target value (also referred to as SV) yd and an output (also referred to as a measured value, PV) of the control target 2, performs a predetermined control calculation, and outputs an operation amount (1 st operation amount) to the control target 2. The feedback controller 10 operates as a main controller, for example. For example, the feedback controller 10 is a controller for operating the output of the control target 2 in accordance with a desired design, assuming that there is no modeling error and no disturbance. As the feedback controller 10, for example, a PID controller or the like that can be automatically designed by auto tuning or the like can be used. In addition, the feedback controller 10 may also use an I-PD controller that suppresses the overshoot, so that the rise from the target value is improved by the neural network controller 30.

The normative model unit 20 includes a dead time (dead time unit) and outputs a desired response waveform to an input. The specification model unit 20 receives the target value yd as input. The input/output relationship of the normative model unit 20 can be expressed by, for example, a first-order lag system including a dead time unit, but the relationship is not limited to this, and may be an appropriate relationship including a dead time unit. The dead time of the normative model unit 20 may be set to the same time as the dead time of the control object 2, for example. The dead time of the normative model unit 20 may be about the same as the dead time of the control target 2. The same degree may be a degree to which the responsiveness of the output of the control object 2 can be improved, for example, by the neural network controller 30. The dead time of the controlled object 2 may be rounded by a predetermined number of bits, that is, may be within a predetermined allowable error range. As an example, the dead time of the normative model unit 20 may be in a range of about ± 10% or about ± 30% with respect to the dead time of the control target 2. The error ey between the output of the normative model unit 20 including the dead time and the output of the control object 2 is supplied to the neural network controller 30 as a teaching signal.

The output (the 2 nd operation amount) of the neural network controller 30 is added to the output (the 1 st operation amount) of the feedback controller 10 and is input to the control target 2. The neural network controller 30 uses the neural network to compare the output of the controlled object 2 with that of the normative model unit 20 by the change (adjustment) of the output of the neural network controller 30The output error ey is learned so as to be the minimum or lower than a predetermined threshold value. For example, the neural network controller 30 learns by the steepest descent method and the reflection propagation method so that the square error ey²And minimum. As input signals, the neural network controller 30 is input with a target value yd and an output y of a control target. The neural network controller 30 supplies an input signal and an output corresponding to the learning result. Further, as described above, the output xN from the neural network controller 30 is added to the output of the feedback controller 10 to obtain the manipulated variable x, and is input to the control target 2. In this way, the output xN of the neural network controller 30 and the output of the feedback controller 10 are added and input to the control object 2, so that the roles of the feedback controller 10 and the neural network controller 30 can be separated. In addition, as an input signal, the neural network controller 30 may also be input with the error ey.

In addition, the neural network has inputs and outputs and one or more intermediate layers. The middle layer is composed of a plurality of nodes. The structure of the neural network can adopt an appropriate structure, and a known learning method can be adopted as the learning method for the neural network.

Further, the control device 1 may include: a differentiator 11 for determining a difference ey between the output y of the controlled object 2 and the output of the normative model unit 20; an adder 12 that adds an output of the feedback controller 10 and an output of the neural network controller 30; and a differentiator 13 for determining a difference e between the target value yd and the output y of the controlled object 2.

The normative model unit 20 and the neural network controller 30 may be mounted on a digital device having a Processing unit such as a cpu (central Processing unit) and a dsp (digital Signal processor), and a storage unit such as a memory. The processing unit and the storage unit of the normative model unit 20 and the neural network controller 30 may be shared, or may be separate. In addition, the neural network controller 30 may have a plurality of processing units and execute at least a part of the processing in parallel.

(Effect)

According to the control device of the present embodiment, for example, the following effects are obtained. However, the control device of the present embodiment is not necessarily limited to a device that exhibits all of the following effects.

As the feedback controller 10, a controller that can be designed by auto-tuning or the like can be used. Therefore, in designing the feedback controller 10, a model of the control object 2 is not necessary. In addition, since the model of the control target 2 is not required in designing the neural network controller 30, the model is not necessary in designing each controller of the control device 1.

In the control system of the present embodiment, learning is performed so that the output of the controlled object 2 follows the output of the normative model unit 20. By providing the normative model unit 20 with the dead time, it is possible to avoid the neural network controller 30 from starting learning by the neural network (i.e., establishing a causal relationship) in a state where there is no output of the control target 2. In addition, in learning of the neural network, the problem of learning in advance in the dead time can be avoided. Thus, there is no need to delay learning of the neural network by only the dead time, and there is no need to intentionally increase the learning period. This can avoid the phenomenon that the neural network controller 30 provides an excessive amount of control input when increasing the output of the control object 2.

The feedback controller 10 is primarily intended to operate in a manner that meets the nominal specifications at the time of design. For example, the feedback controller 10 operates to satisfy the specification of a control device (controller) in the control system, the operation specification of the PID, and the like. On the other hand, the neural network controller 30 functions to make the output of the controlled object 2 follow the output of the normative model unit 20 after learning. Further, in the case where a modeling error or disturbance occurs, the neural network controller 30 compensates for the modeling error or disturbance. When such an error or disturbance occurs, an error occurs between the output of the controlled object 2 and the output of the normative model unit 20, and the neural network controller 30 operates based on the error to compensate for the modeling error or disturbance.

In addition to the above effects, the control device of the present embodiment has the following effects.

Since the control input follows the output of the normative model unit 20, the control input is unlikely to become excessively large even if the neural network is learned by setting and adjusting the normative model unit 20. In other words, the input of the control object 2 can be indirectly adjusted.

The design of the neural network controller 30 does not require a model of the control object. In addition, the feedback controller 10 designed by auto-tuning can be utilized, and thus the control system can be designed without a model.

Even if the learning of the neural network progresses, the feedback control system can be maintained without shifting to the feedforward configuration. For example, when the error between the output of the normative model unit 20 and the output of the controlled object 2 is zero, it corresponds to the operation of only the feedback controller 10.

By using the I-PD configuration in the feedback controller 10, only the responsiveness can be improved without an overshoot as the neural network learning proceeds. For example, the output of the controlled object 2 may be controlled to increase slowly immediately after the start of the control, but the increase may be improved as the learning progresses while suppressing the overshoot. In addition, in the case where learning of the neural network controller 30 is not good or control performance is not improved, or the like, even if the output of the neural network controller 30 is limited or zero, the initial basic performance is ensured by the feedback controller 10.

Since learning is performed in accordance with the output of the normative model unit 20, the present invention is easily applicable to a multiple-input multiple-output (MIMO) system. For example, in a control system that controls multipoint temperatures, uniform control of multipoint (multiple output) temperatures may be performed including a transition state. When applied to a mimo system, the error, the operation amount, and the like include a plurality of elements corresponding to input/output, and can be expressed by a vector, for example.

The control device of the present embodiment can be applied to a control system having a dead time, such as a process control system or a temperature adjustment system. Specific examples thereof include a temperature adjusting/conditioning system, an injection molding machine, and a hot plate. In such a field, design of a feedback controller is generally performed by automatic tuning using ON/OFF (ON/OFF) of a control input without deriving a model of a control object. The present embodiment has an advantage that, by adding and introducing a controller using a neural network to such a conventional control system, the conventional design method not using a model is inherited, and further, the control performance can be improved by learning together with the operation.

(simulation results)

The simulation results and effects of the control system using the control device 1 in the present embodiment will be described by comparison with comparative examples.

First, a response waveform in the comparative example will be described. Fig. 2 is a block diagram of a control system of a comparative example. In the comparative example, the above-described feedback error learning system is adopted as a related art. In this example, the neural network controller 110 performs learning so that xc becomes 0 as the learning proceeds, using the output xc of the feedback controller 120 as a teaching signal. Thus, the control system of the comparative example performs learning and control such that the error e between the target value yd and the controlled object 130 becomes 0 (in other words, the output y is the target value yd). Thus, after learning, the controller used is transferred from the feedback controller 120 to the neural network controller 110. Here, the feedback controller 120 uses a PI controller. In the neural network of the neural network controller 110, the middle layer is two layers, and the number of nodes in each layer is set to 10.

Fig. 3 shows a repetitive step response waveform in the control system of the comparative example. The horizontal axis of fig. 3 represents time. Fig. 3 shows an output response waveform 32 of the control target 130 with respect to the target value (repeated step command) 31 in the upper row, and shows an output (FBA)33 of the feedback controller 120 and an output (NNout)34 of the neural network controller 110 in the lower row. As shown in fig. 3, no improvement in the responsiveness with the passage of time was observed.

Fig. 4 is a comparison graph of overlapping of repetitive step response waveforms in the control system of the comparative example. The horizontal axis of fig. 4 is time. Fig. 4 shows a waveform 41 in which responses (step responses) to step commands in a plurality of positive directions are displayed in an overlapping manner in the upper row, and a waveform 43 in which responses (step responses) to step commands in a plurality of negative directions are displayed in an overlapping manner in the lower row. More specifically, the upper row and the lower row in fig. 4 show, in addition, step response waveforms (shown by thin lines, broken lines, and thick lines, respectively) for the 1 st, 5 th, and 10 th step commands superimposed on the repeated step commands 31 shown in fig. 3, with the rise or fall of each step command being time 0. As a reference example, ideal response waveforms 42 and 44 are shown by broken lines. Even when fig. 4 is observed, the response waveforms are substantially superimposed, and the improvement of the responsiveness in each step is not observed.

On the other hand, fig. 5 and 6 show simulation results in the control system of the present embodiment as an example. Fig. 5 is a repetitive step response waveform in the control system of the present embodiment. Fig. 6 is a graph showing comparison of overlapping of repetitive step response waveforms in the control system according to the present embodiment.

The control object 2 and the feedback controller 10 are configured to be the same as the control object 130 and the feedback controller 120 of the comparative example shown in fig. 2, respectively. In the neural network of the neural network controller 30, the intermediate layer is 2 layers, the number of nodes is 10, and the structure is the same as that of the neural network controller 110.

The horizontal axis of fig. 5 is time. Fig. 5 shows an output response waveform 52 of the controlled object 2 with respect to a target value (repetitive step command) 51 in the upper row and an output (FBA)53 of the feedback controller 10 and an output (NNout)54 of the neural network controller 30 in the lower row, as in fig. 3.

The horizontal axis of fig. 6 is time. Fig. 6 shows waveforms 61 to 63 in the upper row, which are displayed by superimposing responses (step responses) to a plurality of step commands in the positive direction, and waveforms 65 to 67 in the lower row, which are displayed by superimposing responses (step responses) to a plurality of step commands in the negative direction, as in fig. 4. More specifically, in the upper row and the lower row of fig. 6, for the repeated step command 51 shown in fig. 5, the step response waveforms 61 and 65 for the 1 st step command, the step response waveforms 62 and 66 for the 5 th step command, and the step response waveforms 63 and 67 for the 10 th step command are superimposed with the rise of each step command as time 0. In addition, as a reference example, ideal response waveforms (for example, outputs of the specification model unit 20) 64 and 68 are shown by broken lines.

By repeating the step response, it can be confirmed that the overshoot from the target value is reduced together with the positive and negative responses, and the settling time is also increased, following the output of the normative model. In addition, from the lower row of fig. 5, it can be confirmed that the output (NNout)54 of the neural network controller 30 increases due to the repeated step response. This indicates that the learning of the neural network controller 30 is performed in such a manner that the output signal y follows the normative model output.

(others)

In the above embodiment, the neural network controller 30 performs learning using the neural network, but learning may be performed using a function other than the neural network. That is, the neural network controller 30 may also be a learning type controller. In addition, a 2 nd control device having a configuration not including the feedback controller 10 may be provided among the control devices 1. For example, the control system may be configured by applying a control device including the normative model unit 20 and the neural network controller 30 to a control system that controls a control target by using a previously designed conventional feedback controller.

The above-described configurations and processes can be realized by a computer having a processing unit and a storage unit. The processing unit executes the processing of each configuration. The storage unit stores a program executed by the processing unit. The above-described processing may be realized as a control method executed by the processing unit. The present invention can be realized by a program or a program medium containing instructions for causing a processing unit to execute the above-described processing, a computer-readable recording medium and a non-transitory recording medium storing the program, and the like.

[ industrial applicability ]

The control device and the control system according to the present embodiment can be applied to, for example, a control system that controls a control target having a dead time. As one example, it may be applicable to a process control system or a temperature regulation system. More specific examples include a temperature control/air conditioning system, an injection molding machine, and a hot plate.

[ description of reference numerals ]

1 a control device; 2 controlling the object; 10 a feedback controller; 20 a normative model part; 30 a neural network controller; 51 target value (repeat step command); 52 outputting a response waveform; 53 feedback controller output (FBA); 54 output of the neural network controller (NNout).