Plant response estimation device, plant response estimation method, and program
The plant response estimation device accurately estimates plant response models by selecting suitable learning data from operational data, addressing inaccuracies in conventional methods by excluding steady-state or biased data to enhance model parameter convergence and control stability.
Patent Information
- Application Number
- JP2023071309
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-04-25
- Publication Date
- 2024-11-07
- Estimated Expiration
- 2043-04-25
AI Technical Summary
Conventional plant response estimation methods using steady-state or biased operating data lead to inaccurate parameter estimation in plant response models, as these data types lack sufficient information about plant dynamics, causing the estimated parameters to deviate from true values.
A plant response estimation device that selects appropriate learning data by creating a state vector from operational data, determining its suitability for model updates, and updating parameters using a state vector-based method to ensure accurate model parameter estimation.
The device achieves high-accuracy estimation of plant response models by excluding unsuitable operating data, such as steady-state or biased data, thereby improving model parameter convergence and control stability.
Smart Images

Figure 2024157147000001_ABST
Abstract
Description
[Technical field]
[0001] The present disclosure relates to a plant response estimation device, a plant response estimation method, and a program. [Background technology]
[0002] Model predictive control is known as a control method for making a controlled variable of a controlled object follow a target value. In addition, a recursive least squares method (RLS method) and the like are known as a method for estimating parameters of a linear prediction model used in model predictive control. For example, Patent Document 1 describes a technology for estimating parameters of a plant response model by the RLS method using operation data of a plant in operation. [Prior art documents] [Patent documents]
[0003] [Patent Document 1] JP 2023-53573 A Summary of the Invention [Problem to be solved by the invention]
[0004] However, in the conventional technology, the learning data used to update the parameters of the plant response model may be, for example, operation data that contains only constant values in a steady state, which may lead to inappropriate model parameters being estimated. Operation data that contains only constant values or one frequency component lacks information on the dynamics of the plant and is inappropriate as learning data, but such data may easily occur during operation. When learning using such inappropriate operation data, the estimated values of the parameters of the plant response model may deviate from the true values.
[0005] The present disclosure has been made in consideration of the above points, and provides a technique for accurately estimating a response model from operational data of a plant during operation. [Means for solving the problem]
[0006] A plant response estimation device according to one aspect of the present disclosure includes a state vector creation unit that creates a state vector representing a state of a controlled plant based on operation data of the controlled plant, a learning data selection unit that uses the state vector to determine whether or not to use the state vector as learning data, and a parameter update unit that uses the state vector as learning data to update parameters of a function representing a response model of the controlled plant when it is determined that the state vector is to be used as learning data. Effect of the Invention
[0007] A technique is provided for accurately estimating a response model from operational data of a plant during operation. [Brief description of the drawings]
[0008] [Figure 1] 1 is a diagram illustrating an example of a hardware configuration of a plant response estimation apparatus according to a first embodiment. [Diagram 2] 1 is a diagram illustrating an example of a functional configuration of a plant response estimation apparatus according to a first embodiment. [Diagram 3] 10A and 10B are diagrams for explaining an example of an operation of a buffer unit; [Figure 4] 10 is a flowchart illustrating an example of a learning data selection process according to the first embodiment. [Diagram 5] 11 is a flowchart for explaining an example of an enlarged model parameter estimation process according to the first embodiment. [Figure 6] 11 is a diagram for explaining an example of an operation of a parameter conversion unit. FIG. [Figure 7] 11 is a diagram for explaining an example of an operation of a step response calculation unit. FIG. [Figure 8] 5 is a flowchart for explaining an example of a step response calculation process according to the first embodiment. [Figure 9]FIG. 11 is a diagram for explaining an example of updating a state vector. [Figure 10] 13 is a flowchart illustrating an example of a learning data selection process according to the second embodiment. [Figure 11] 13 is a flowchart illustrating an example of a learning data selection process according to the third embodiment. [Figure 12] 13 is a flowchart illustrating an example of a learning data selection process according to the fourth embodiment. [Figure 13] FIG. 4 is a diagram illustrating a step response of a target plant in an embodiment. [Figure 14] FIG. 11 is a diagram (part 1) showing model parameter estimation results and control results when a learning data selection unit is not used. [Figure 15] FIG. 13 is a diagram (part 2) showing model parameter estimation results and control results when the learning data selection unit is not used. [Figure 16] FIG. 13 is a diagram (part 1) showing model parameter estimation results and control results when a learning data selection unit is used. [Figure 17] FIG. 2 is a diagram (part 2) showing model parameter estimation results and control results when the learning data selection unit is used. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0009] Hereinafter, an embodiment of the present invention will be described. In each of the following embodiments, a plant response estimation device 10 capable of accurately estimating a plant response model from operation data of a plant in operation will be described. Note that "in operation" means that the plant is operating normally and performing normal operation, and may be called, for example, online, operating, or in operation. Also, a plant is an industrial facility composed of one or more machines, devices, equipment, etc., and is a control target controlled by model predictive control using a plant response model. Specific examples of plants include, for example, petrochemical plants, food plants, steel plants, and power plants, but these are merely examples and are not limited to these. Also, in general, in the control systems of these plants, input and output data such as the control amount, operation amount, and disturbance amount of the plant are treated as operation data.
[0010] Depending on the operating conditions of the plant, for example, the steady state may continue or the values of the operating data may be biased toward the upper or lower limit. For this reason, if the parameters of the plant response model are updated using operating data in the steady state or operating data biased toward the upper or lower limit as learning data, the parameters may converge due to the operating data, and as a result, the estimated values of the parameters may deviate from the true values. In this way, operating data in the steady state or operating data biased toward the upper or lower limit can be said to be inappropriate as learning data for the plant response model. Learning data refers to data used to update the parameters of the plant response model.
[0011] Therefore, in the following embodiments, inappropriate operation data such as operation data in a steady state or operation data biased toward upper and lower limits is not used as learning data, and other operation data is used as learning data to update the parameters of the plant response model. This enables the plant response estimation device 10 according to the following embodiments to estimate a plant response model with high accuracy.
[0012] [First embodiment] First, the first embodiment will be described.
[0013] <Hardware Configuration Example of Plant Response Estimation Apparatus 10> An example of a hardware configuration of a plant response estimation device 10 according to the first embodiment is shown in Fig. 1. As shown in Fig. 1, the plant response estimation device 10 according to the first embodiment has an input device 11, a display device 12, an external I / F 13, a communication I / F 14, a processor 15, and a memory device 16. Each of these hardware components is connected to each other via a bus 17 so as to be able to communicate with each other.
[0014] The input device 11 is, for example, a keyboard, a mouse, a touch panel, various physical buttons, etc. The display device 12 is, for example, a display, a display panel, etc. Note that the plant response estimation device 10 does not necessarily have to include at least one of the input device 11 and the display device 12, for example.
[0015] The external I / F 13 is an interface with an external device such as a recording medium 13a. Examples of the recording medium 13a include a CD (Compact Disc), a DVD (Digital Versatile Disk), an SD memory card (Secure Digital memory card), and a USB (Universal Serial Bus) memory card.
[0016] The communication I / F 14 is an interface for connecting the plant response estimation device 10 to a communication network. The processor 15 is, for example, various types of arithmetic devices such as a CPU (Central Processing Unit). The memory device 16 is, for example, various types of storage devices such as an SSD (Solid State Drive), a RAM (Random Access Memory), a ROM (Read Only Memory), and a flash memory.
[0017] 1 is merely an example, and the plant response estimation device 10 may have other hardware configurations. For example, the plant response estimation device 10 may have multiple processors 15 and multiple memory devices 16, or may have various types of hardware other than the hardware shown in the figure.
[0018] <Example of Functional Configuration of Plant Response Estimation Device 10> An example of a functional configuration of a plant response estimation device 10 according to the first embodiment is shown in Fig. 2. As shown in Fig. 2, the plant response estimation device 10 according to the first embodiment has a model parameter estimation unit 101 and a step response calculation unit 102. These units are realized, for example, by processing in which one or more programs installed in the plant response estimation device 10 are executed by a processor 15 or the like.
[0019] The model parameter estimation unit 101 receives observed values of operation data (controlled variable y, manipulated variable u, and disturbance v) from the plant to be controlled (or devices such as sensors that measure the operating state of the plant) for each sampling period Δ, and estimates model parameters of the plant response model in accordance with the sampling period Δ. Note that the operation data may be called, for example, measurement data or observation data.
[0020] Here, if time is t, the controlled variable y, the manipulated variable u, and the disturbance v are expressed as y(t), u(t), and v(t), respectively. k (k is an index representing the sampling time, and is incremented periodically.) k+1 -t k =Δ. Note that k can be an integer greater than or equal to 0, and the sampling time t k In the following, unless otherwise specified, each sampling time t k Let us identify with its index k and also denote it as y(k), u(k), and v(k).
[0021] In the following, the plant response model is defined as the plant response function S with parameters θ.θ (t), the parameter θ is called the model parameter. The plant response model includes various plant response functions S θ However, in the following, a polynomial model such as the ARMAX model is mainly assumed. When the plant response model is an ARMAX model, the model parameter θ becomes the coefficient of the ARMAX model. It goes without saying that various models other than the ARMAX model can be used as the plant response model.
[0022] Furthermore, in the following, a model in which an offset term (constant term) representing the steady state of the controlled plant is added to the plant response model represented by the ARMAX model will be referred to as an augmented plant response model.
[0023] The step response calculation unit 102 uses the model parameters (hereinafter, also referred to as model parameter estimates) θ estimated by the model parameter estimation unit 101 to calculate an estimate of the step response of the plant at a given time t (hereinafter, also referred to as step response estimates) S θ (t) is calculated. The step response is the controlled variable when a step signal is applied to the plant as a manipulated variable.
[0024] However, the step response calculation unit 102 is one of the methods of using the model parameter estimate value θ, and the plant response estimation device 10 according to this embodiment may be configured to have only the model parameter estimation unit 101. Furthermore, the plant response estimation device 10 according to this embodiment may be configured to have a functional unit that calculates a plant response other than the step response. For example, it may be configured to have an "impulse response calculation unit" that calculates an impulse response, or a "ramp response calculation unit" that calculates a ramp response. In addition to these, it is applicable to various other usage methods within the scope that does not deviate from the gist of this embodiment.
[0025] Here, the model parameter estimation unit 101 includes a buffer unit 111 , a state vector conversion unit 112 , a learning data selection unit 113 , a sequential estimation calculation unit 114 , and a parameter conversion unit 115 .
[0026] The buffer unit 111 accumulates (buffers) the control amount y(k), the operation amount u(k), and the disturbance v(k) in a certain predetermined period in the memory device 16. The state vector conversion unit 112 resamples the control amount y(k), the operation amount u(k), and the disturbance v(k) buffered in the memory device 16, and creates a state vector ξ(k) including the resampled control amount y(k), the operation amount u(k), and the disturbance v(k). The learning data selection unit 113 uses the state vector ξ(k) to determine whether or not the resampled control amount y(k), the operation amount u(k), and the disturbance v(k) are used as learning data, and sets an update flag to ON if they are used as learning data, and sets the update flag to OFF if they are not used as learning data. If the update flag is ON, the sequential estimation calculation unit 114 uses the state vector ξ(k) to estimate the expanded model parameters, which are model parameters of the expanded plant response model. The parameter conversion unit 115 converts the estimated values of the augmented model parameters into model parameter estimates θ. Note that a state vector is a term used in the field of control, and is a vector that represents the state of a controlled object or a system including the controlled object.
[0027] <Operation of the buffer unit 111> Sampling time t k The control amount buffer in is Y(k), the manipulated variable buffer is U(k), and the disturbance buffer is V(k), and each of these buffers is represented by the following vectors.
[0028]
number
[0029] At this time, when the buffer unit 111 receives the control amount y(k), the operation amount u(k), and the disturbance v(k), as shown in FIG. 3, it updates the control amount buffer Y(k) from the control amount y(k) and the control amount buffer Y(k-1), updates the operation amount buffer U(k) from the operation amount u(k) and the operation amount buffer U(k-1), and updates the disturbance buffer V(k) from the disturbance v(k) and the disturbance buffer V(k-1).
[0030] Specifically, the buffer unit 111 deletes y(k-B1-1) stored in the control amount buffer Y(k-1) and then stores the newly observed control amount y(k), thereby updating the control amount buffer Y(k) to store the control amounts from y(k-B1) to y(k). Similarly, the buffer unit 111 deletes the control amount u(k-B2-1) stored in the control amount buffer U(k-1) and then stores the newly observed control amount u(k), thereby updating the control amount buffer U(k) to store the control amounts from u(k-B2) to u(k). Similarly, the buffer unit 111 deletes the disturbance v(k-B3-1) stored in the disturbance buffer V(k-1) and then stores the newly observed disturbance v(k), thereby updating the disturbance buffer V(k) to store the disturbances from v(k-B3) to v(k).
[0031] It is assumed that the control amount buffer Y(0), the manipulated variable buffer U(0), and the disturbance buffer V(0) are properly initialized (for example, all initialized to 0, etc.). In addition, the above description is given assuming that there is a disturbance v, but if there is no disturbance v, the disturbance buffer V(k) may not be necessary.
[0032] <Operation of state vector conversion unit 112> The resampling period is set to D. At this time, when the control amount buffer Y(k), the manipulated variable buffer U(k), and the disturbance buffer V(k) are updated, the state vector conversion unit 112 performs resampling from each of these buffers Y(k), U(k), and V(k) at the resampling period D to obtain the following resampled control amount vector Y D (k), resampling operation vector U D (k) and the resampled disturbance vector V D (k) is created respectively.
[0033]
number
[0034] In general, larger values of N, M, and L enable more diverse expressions and are expected to enable more accurate prediction of the plant response; however, larger computational resources and memory capacity are required to estimate the model parameter θ.
[0035] Then, the state vector conversion unit 112 converts the resampling control amount vector Y D (k), resampling operation vector U D (k) and the resampled disturbance vector V D Create a state vector ξ(k) from (k) as follows:
[0036]
number
[0037] It is assumed that the state vector ξ(0) is initialized to an appropriate value (for example, initialized to a zero vector, etc.). In addition, the above description was given assuming that there is a disturbance v, but if there is no disturbance v, the resampled disturbance vector V D (k) is not necessary.
[0038] <Operation of the learning data selection unit 113> When the state vector ξ(k) is created, the learning data selection unit 113 uses this state vector ξ(k) to calculate the resampling control amount vector Y D (k), resampling operation vector U D (k) and the resampled disturbance vector V D After judging whether or not (k) is to be used as learning data, the update flag is turned ON or OFF depending on the result of the judgment, thereby selecting (the observed values of) the driving data to be used as learning data.
[0039] Hereinafter, a process for selecting driving data to be used as learning data for a certain k (learning data selection process according to the first embodiment) will be described with reference to Fig. 4. However, if k = 0, the update flag is set to ON and the process is terminated.
[0040] Step S101: First, the learning data selection unit 113 calculates the norm d(k) of the difference between the state vector ξ(k) and the state vector ξ(k-1) one time before. That is, the learning data selection unit 113 calculates d(k)=||ξ(k)-ξ(k-1)|| p Here, p is either p=1, p=2 or p=∞, ||·||1 is the L1 norm, ||·||2 is the L2 norm, ||·|| ∞ and represent the L∞ norm. Hereinafter, d(k) is also referred to as the vector difference norm.
[0041] Step S102: Next, the learning data selection unit 113 determines whether or not the vector difference norm d(k) is equal to or greater than a preset threshold value γ.
[0042] If it is determined that d(k)≧γ is satisfied (YES in step S102), the learning data selection unit 113 proceeds to step S103. On the other hand, if it is not determined that d(k)≧γ is satisfied (NO in step S102), the learning data selection unit 113 proceeds to step S104.
[0043] Step S103: The learning data selection unit 113 turns on the update flag. Here, the update flag is the resampling operation amount vector U D (k) and the resampled disturbance vector V D This is a flag that indicates whether or not to use (k) as learning data. When the update flag is ON, the resampling operation vector U D (k) and the resampled disturbance vector V D (k) is used as learning data to update the augmented model parameters. On the other hand, if the update flag is OFF, the resampling operation vector U D (k) and the resampled disturbance vector V D (k) is not used as training data, and the augmented model parameters are not updated.
[0044] Step S104: The learning data selection unit 113 turns off the update flag.
[0045] <Operation of sequential estimation calculation unit 114> In the following, the augmented model parameters are e When the learning data selection unit 113 sets the update flag to ON or OFF, the recursive estimation calculation unit 114 calculates the expanded model parameters θ using the state vector ξ(k). e That is, the sequential estimation calculation unit 114 estimates the value of k For each, the expanded model parameters θ e Estimate the value of .
[0046] Below, for a certain k, the augmented model parameters θ e The process of estimating the value of (enlarged model parameter estimation process according to the first embodiment) will be described with reference to FIG. k The expanded model parameter θ e The estimate of θ e This is represented as (k).
[0047] Step S201: First, the sequential estimation calculation unit 114 determines whether the update flag is ON.
[0048] When it is determined that the update flag is ON (YES in step S201), the sequential estimation calculation unit 114 proceeds to step S202. On the other hand, when it is not determined that the update flag is ON (NO in step S201), the sequential estimation calculation unit 114 ends the enlarged model parameter estimation process. In other words, when the update flag is OFF, the sequential estimation calculation unit 114 calculates the enlarged model parameter θ e The process ends without estimating or updating the value of .
[0049] Step S202: The sequential estimation calculation unit 114 calculates the enlarged model parameter θ e and the covariance matrix P are initialized. Here, examples of cases in which it is determined to initialize include the first calculation (i.e., when k=1) and when an initialization instruction is given by a user or the like.
[0050] Expanded model parameter θ e When it is determined that the covariance matrix P should be initialized (YES in step S202), the sequential estimation calculation unit 114 proceeds to step S203. e If it is not determined that covariance matrix P should be initialized (NO in step S202), the recursive estimation calculation unit 114 proceeds to step S204.
[0051] Step S203: The sequential estimation calculation unit 114 calculates the sampling time t k The index k of is initialized to k=1, and θ e Initialize P(0)=θ0 and P(0)=I, where θ0 is a preset initial value of the enlarged model parameters, and I is a preset arbitrary matrix (for example, a unit matrix).
[0052] Step S204: The sequential estimation calculation unit 114 calculates the expanded model parameter estimate θ e (k-1), the state vector ξ(k), and the control amount y(k) are used to calculate the prediction error ε(k). The prediction error ε(k) is, for example, ε(k)=y(k)-ξ(k). Τ θ e (k-1), where T represents the transpose.
[0053] Step S205: The recursive estimation calculation unit 114 updates the covariance matrix P(k−1) as follows to obtain the covariance matrix P(k):
[0054]
number
[0055] First, it can be used as a forgetting coefficient for forgetting past driving data by setting 0<λ≦1. The closer λ is to 0, the more rapidly the influence of past driving data is reduced, and the closer λ is to 1, the more the influence of past driving data is retained.
[0056] Second, it can be used as a coefficient to accelerate parameter updates by setting λ > 1. The larger the value of λ, the faster the parameter convergence can be.
[0057] Step S206: Then, the sequential estimation calculation unit 114 calculates the extended model parameter estimate value θ e (k-1) is updated by the following to obtain the augmented model parameter estimates θ e Obtain (k).
[0058]
number
[0059] <Operation of the parameter conversion unit 115> Augmented model parameter estimates θ e (k) is expressed as follows:
[0060]
number
[0061] At this time, the parameter conversion unit 115 converts the enlarged model parameter estimate value θ e Once (k) is obtained, as shown in Figure 6, C By converting the vector into one that excludes (k), we obtain the model parameter estimate θ(k). This gives us the following model parameter estimate:
[0062]
number
[0063] <Operation of step response calculation unit 102> In the following, we consider the model parameter estimates used to estimate the step response (i.e., the plant response function S θ The model parameter estimate value θ set to k (θ=θ(k)) is also referred to as the “model parameter setting value θ.” For simplicity, in the following, a unit step signal is assumed as the step signal.
[0064] As shown in FIG. 7, when a model parameter setting value θ and a time t are given, the step response calculation unit 102 calculates a unit step response S at time t after the time t has elapsed from the initial time 0. θ (t) is calculated. The unit step response is the response when a unit step signal is applied as the manipulated variable u (that is, the output of the plant response model of the controlled plant).
[0065] Given a certain time t and a certain model parameter setting θ, the step response S θ The process of calculating (t) (step response calculation process according to the first embodiment) will be described with reference to Fig. 8. In the following, the state vector φ(k') at index k' is represented as follows.
[0066]
number
[0067] Step S301: The step response calculation unit 102 initializes an index indicating a time used only in this process as τ=0 and k′=0, and also initializes a state vector φ(0) as follows:
[0068]
number
[0069] Step S302: The step response calculation unit 102 calculates y(k')=φ(k'). Τ The controlled variable prediction value y(k') is calculated based on θ.
[0070] Step S303: The step response calculation unit 102 updates the state vector φ(k') to the state vector φ(k'+1) at the next index k'+1 by using the controlled variable predicted value y(k'). At this time, as shown in FIG. 9, the step response calculation unit 102 sets the controlled variable predicted value y(k') calculated in the above step S302 to y(k') of the state vector φ(k'+1), and sets the same value as the state vector φ(k') to y(k'-N+1) to y(k'-1). In addition, the step response calculation unit 102 sets 1 to u(k'+1) of the state vector φ(k'+1), and sets the same value as the state vector φ(k') to u(k'-M+1) to u(k'). In addition, the step response calculation unit 102 sets 0 to v(k'+1) of the state vector φ(k'+1), and sets the same value as the state vector φ(k').
[0071] Step S304: The step response calculation unit 102 updates the time τ to τ+Δ, and updates the index k′ to k′+1.
[0072] Step S305: The step response calculation unit 102 judges whether or not τ≧t holds. If it is not judged that τ≧t holds (NO in step S305), the step response calculation unit 102 returns to step S302. As a result, steps S302 to S304 are repeatedly executed until τ≧t holds.
[0073] On the other hand, if it is determined that τ≧t (YES in step S305), the step response calculation unit 102 ends the step response calculation process. As a result, the finally calculated controlled variable prediction value y(k′) is the unit step response S θ (t) (i.e., S θ (t)=y(k) is obtained as the unit step response of the plant response model.
[0074] [Second embodiment] Next, a second embodiment will be described. In the second embodiment, the learning data selection process is different from that in the first embodiment. Note that matters not specifically mentioned in the second embodiment may be the same as those in the first embodiment.
[0075] Therefore, the process of selecting driving data to be used as learning data for a certain k (learning data selection process according to the second embodiment) will be described below with reference to Fig. 10. However, when k = 0, the update flag is set to ON and the process is terminated.
[0076] Step S401: First, similarly to step S101 in FIG. 4, the learning data selection unit 113 calculates the norm d(k) of the difference between the state vector ξ(k) and the state vector ξ(k-1) one time before, d(k)=||ξ(k)-ξ(k-1)|| p as the vector difference norm.
[0077] Step S402: Next, the learning data selection unit 113 calculates the threshold value γ(k) using the state vector ξ(k). That is, the learning data selection unit 113 calculates the threshold value γ(k) as follows.
[0078]
number
[0079] Step S403: Next, the learning data selection unit 113 determines whether the vector difference norm d(k) is equal to or greater than a threshold value γ(k).
[0080] If it is determined that d(k)≧γ(k) holds (YES in step S403), the learning data selection unit 113 proceeds to step S404. On the other hand, if it is not determined that d(k)≧γ(k) holds (NO in step S403), the learning data selection unit 113 proceeds to step S405.
[0081] Step S404: The learning data selection unit 113 turns on the update flag, similarly to step S103 in FIG.
[0082] Step S405: The learning data selection unit 113 turns off the update flag, similarly to step S104 in FIG.
[0083] [Third embodiment] Next, a third embodiment will be described. In the third embodiment, the learning data selection process is different from that in the first embodiment. Also, the calculation method of the threshold value γ(k) is different from that in the second embodiment. Note that matters not specifically mentioned in the third embodiment may be the same as those in the first embodiment.
[0084] Therefore, hereinafter, a process for selecting driving data to be used as learning data for a certain k (learning data selection process according to the third embodiment) will be described with reference to Fig. 11. However, when k = 0, the update flag is set to ON and the process is terminated.
[0085] Step S501: First, the learning data selection unit 113 determines whether k=1.
[0086] If it is determined that k=1 (YES in step S501), the learning data selection unit 113 proceeds to step S502. On the other hand, if it is not determined that k=1 (NO in step S501), the learning data selection unit 113 proceeds to step S503.
[0087] Step S502: The learning data selection unit 113 selects E ξ (0)=0, S ξ Initialize E(0)=0. ξ (0) is the initial value of the mean norm of the state vector, S ξ (0) is the initial value of the average value of the squared norm of the state vector.
[0088] Step S503: As in step S101 of FIG. 4, the learning data selection unit 113 calculates the norm d(k) of the difference between the state vector ξ(k) and the state vector ξ(k-1)=||ξ(k)-ξ(k-1)|| p as the vector difference norm.
[0089] Step S504: Next, the learning data selection unit 113 calculates the threshold value γ(k) using the state vector ξ(k). That is, the learning data selection unit 113 calculates the threshold value γ(k) as follows.
[0090]
number
[0091] Step S505: Next, the learning data selection unit 113 determines whether the vector difference norm d(k) is equal to or greater than a threshold value γ(k).
[0092] If it is determined that d(k)≧γ(k) holds (YES in step S505), the learning data selection unit 113 proceeds to step S506. On the other hand, if it is not determined that d(k)≧γ(k) holds (NO in step S505), the learning data selection unit 113 proceeds to step S507.
[0093] Step S506: The learning data selection unit 113 turns on the update flag, similarly to step S103 in FIG.
[0094] Step S507: The learning data selection unit 113 turns off the update flag, similarly to step S104 in FIG.
[0095] [Fourth embodiment] Next, a fourth embodiment will be described. In the fourth embodiment, the learning data selection process is different from the first to third embodiments. Note that matters not specifically mentioned in the fourth embodiment may be the same as those in the first embodiment.
[0096] Therefore, hereinafter, a process for selecting driving data to be used as learning data for a certain k (learning data selection process according to the fourth embodiment) will be described with reference to Fig. 12. However, when k = 0, the update flag is set to ON and the process ends.
[0097] Step S601: First, the learning data selection unit 113 determines whether k=1.
[0098] If it is determined that k=1 (YES in step S601), the learning data selection unit 113 proceeds to step S602. On the other hand, if it is not determined that k=1 (NO in step S601), the learning data selection unit 113 proceeds to step S603.
[0099] Step S602: The learning data selection unit 113 selects E y (0)=0, S y Initialize E(0)=0. y (0) is the initial value of the average controlled variable, S y (0) is the initial value of the average squared value of the controlled variable.
[0100] Step S603: The learning data selection unit 113 uses the state vector ξ(k) to calculate the prediction error (absolute value of the prediction error) d(k). The prediction error d(k) is calculated as follows, in the same way as in step S204 of FIG. 5: d(k)=|y(k)-ξ(k) Τ θ e It is calculated by |(k-1)|.
[0101] Step S604: Next, the learning data selection unit 113 calculates the threshold value γ(k) using the control amount y(k). That is, the learning data selection unit 113 calculates the threshold value γ(k) as follows.
[0102]
number
[0103] Step S605: Next, the learning data selection unit 113 determines whether the prediction error d(k) is equal to or greater than a threshold value γ(k).
[0104] If it is determined that d(k)≧γ(k) holds (YES in step S605), the learning data selection unit 113 proceeds to step S606. On the other hand, if it is not determined that d(k)≧γ(k) holds (NO in step S605), the learning data selection unit 113 proceeds to step S607.
[0105] Step S606: The learning data selection unit 113 turns on the update flag, similarly to step S103 in FIG.
[0106] Step S607: The learning data selection unit 113 turns off the update flag, similarly to step S104 in FIG.
[0107] [Example] An example of the plant response estimation device 10 according to the first embodiment will be described below.
[0108] In this embodiment, the augmented plant response model of the plant to be controlled is expressed by the following equation.
[0109] y(k)=a1y(k-1)+a2y(k-2)+b0u(k)+b1u(k-1)+b2u(k-2)+c0 In other words, the controlled variable y has two-dimensional elements, and the manipulated variable u has three-dimensional elements. It is assumed that there is no disturbance v in this embodiment.
[0110] At this time, the state vector ξ(k) and the augmented model parameters θ e (k) is expressed as follows:
[0111]
number
[0112] Furthermore, in this embodiment, the learning data selection unit 113 uses the L1 norm as the norm when calculating the vector difference norm d(k) and sets the threshold value to γ=0.1. In addition, the step response of the controlled plant in this embodiment is assumed to be the step response shown in FIG.
[0113] Under the above settings, a comparison was made between a case where the learning data selection unit 113 was not used and a case where the learning data selection unit 113 was used.
[0114] Fig. 14 shows the model parameter estimation results and control results when the adjustment coefficient is set to λ=0.999 without using the learning data selection unit 113. Fig. 15 shows the model parameter estimation results and control results when the adjustment coefficient is set to λ=1.1 without using the learning data selection unit 113.
[0115] In both Figures 14 and 15, the controlled variable y and the manipulated variable u are constant from time 0 to 1000, so the model parameters also fluctuate from their initial values near time 0 and then remain constant until time 1000.
[0116] After time 1000, in Fig. 14, the model parameters change, and changes are seen up to around time 4000. On the other hand, in Fig. 15, the model parameters do not change. This is because when the adjustment coefficient λ is large, the covariance matrix is updated quickly and becomes close to zero.
[0117] Regarding the control response, in Fig. 14, the model parameters have changed, so the controlled variable can track the target value. On the other hand, in Fig. 15, the model parameters have not changed from the initially estimated values, so the model is not estimated correctly, and the controlled variable cannot track the target value, resulting in an unstable response.
[0118] Fig. 16 shows the model parameter estimation results and control results when the adjustment coefficient is set to λ=0.999 using the learning data selection unit 113. Fig. 17 shows the model parameter estimation results and control results when the adjustment coefficient is set to λ=1.1 using the learning data selection unit 113.
[0119] In both Figures 16 and 17, the controlled variable y and the manipulated variable u are constant from time 0 to 1000, so the model parameters also fluctuate from their initial values near time 0 and then remain constant until time 1000.
[0120] After time 1000, in Fig. 16, the model parameters change and converge around time 1500. On the other hand, in Fig. 17, unlike Fig. 15, the model parameters change and converge around time 1200. This is because the steady-state operating data from time 0 to 1000 is not used for updating due to the action of the learning data selection unit 113, so the covariance matrix does not converge to zero even in the steady state, and also because the operating data in which the controlled variable y and the manipulated variable u have changed around time 1000 is used for updating, accelerating the parameter update and bringing them closer to the true values, resulting in faster convergence.
[0121] Regarding the control response, the model parameters are estimated relatively correctly in both Figs. 16 and 17, so the controlled variable can track the target value.
[0122] Therefore, when the adjustment coefficient λ is made larger than 1 to hasten the convergence of the model parameters, the function of the learning data selection unit 113 can be said to be able to prevent the parameters from converging on the operating data in the steady state.
[0123] As described above, the learning data selection unit 113 stops parameter updating when the operating data does not contain sufficient information, for example, when the steady state continues, and performs parameter updating when the controlled variable y or the manipulated variable u changes. This has the advantage that, for example, a situation in which the parameters inadvertently converge to a local value far from the true value in the steady state is avoided, while the parameters quickly converge to a value closer to the true value than in the case where the learning data is not selected.
[0124] In the above embodiments, the case where the parameters of the plant response model are estimated has been mainly described, but the present invention is not limited thereto. For example, the plant response estimation device 10 according to each of the above embodiments may function as a control device that actually controls the controlled plant by the plant response model in which the parameters are set. In this case, the plant response estimation device 10 according to each of the above embodiments may control the controlled plant by the plant response model in which the parameters are set, using a known control method such as model predictive control. In this way, the plant response estimation device 10 according to each of the above embodiments can also contribute to the tracking ability of control and the stabilization of response by combining it with a known control method such as model predictive control.
[0125] The present invention is not limited to the above-described embodiments specifically disclosed, and various modifications, changes, combinations with known technologies, etc. are possible without departing from the scope of the claims. [Explanation of symbols]
[0126] 10 Plant response estimator 11 Input Devices 12 Display device 13 External I / F 13a Recording media 14 Communication I / F 15 Processors 16 Memory Devices 17 Bus 101 Model parameter estimation unit 102 Step response calculation section 111 Buffer section 112 State Vector Transformation Unit 113 Learning Data Selection Unit 114 Sequential estimation calculation section 115 Parameter conversion section
Claims
1. a state vector creation unit that creates a state vector representing a state of a controlled plant based on operation data of the controlled plant; a learning data selection unit that uses the state vector to determine whether or not the state vector is to be used as learning data; a parameter update unit that updates parameters of a function representing a response model of the controlled plant by using the state vector as learning data when it is determined that the state vector is to be used as learning data; A plant response estimator having
2. The learning data selection unit is Calculating a norm of the difference between a current value and a previous value of the state vector; If the norm is equal to or greater than a predetermined threshold, the state vector is determined to be learning data; 2. The plant response estimation apparatus according to claim 1, wherein when the norm is less than the threshold value, it is determined that the state vector is not to be used as learning data.
3. The learning data selection unit is 3. The plant response estimator according to claim 2, wherein the threshold value is calculated using a variance of the state vector and a predetermined coefficient.
4. The learning data selection unit is calculating an estimate of a controlled variable of the controlled plant using the current value of the state vector and the parameters; calculating a prediction error representing the error between the estimated value of the controlled variable and the current value of the controlled variable; If the prediction error is equal to or greater than a predetermined threshold, the state vector is determined to be training data; 2. The plant response estimating apparatus according to claim 1, wherein when the prediction error is less than the threshold value, it is determined that the state vector is not to be used as learning data.
5. The learning data selection unit is 5. The plant response estimation device according to claim 4, wherein the threshold value is calculated using a variance of the controlled variable and a predetermined coefficient.
6. The parameter update unit is calculating estimates of said parameters by recursive least squares with predetermined adjustment factors; 6. The plant response estimation device according to claim 1, wherein the adjustment coefficient is used for calculating the estimated values of the parameters and for adjusting an update amount when estimating a covariance matrix.
7. The parameter update unit is 7. The plant response estimator according to claim 6, wherein the adjustment coefficient is set to a value greater than 1.
8. a state vector creation step of creating a state vector representing a state of a controlled plant based on operation data of the controlled plant; a learning data selection step of determining whether or not the state vector is to be used as learning data by using the state vector; a parameter updating step of updating parameters of a function representing a response model of the controlled plant by using the state vector as learning data when it is determined that the state vector is to be used as learning data; A computer-implemented plant response estimation method.
9. a state vector creation step of creating a state vector representing a state of a controlled plant based on operation data of the controlled plant; a learning data selection step of determining whether or not the state vector is to be used as learning data by using the state vector; a parameter updating step of updating parameters of a function representing a response model of the controlled plant by using the state vector as learning data when it is determined that the state vector is to be used as learning data; A program that causes a computer to execute the following.
Citation Information
Patent Citations
Method and system for using inference measurement to detect abnormal events in continuous industrial processes
JP2010519636A
Failure prediction method, failure prediction apparatus and failure prediction program
JP2019028565A
Machine learning device, error occurrence prediction device, and control device
JP2022133615A
Plant response estimation device, plant response estimation method and program
JP2023053573A
Control device, control method, and program
JP7115654B1
