Control device and control method

Abstract

The control device (10) includes: a model building unit (11) that builds a model of the simulated control object (20); a problem subdivision unit (12) that subdivides the model built by the model building unit (11); a control strategy calculation unit (13) that uses the model subdivided by the problem subdivision unit (12) to predict the future state of the control object (20) and calculates the control strategy of the control object (20) based on the predicted future state; and an operation instruction generation unit (14) that generates operation instructions for the control object (20) based on the control strategy calculated by the control strategy calculation unit (13).

Description

Technical Field

[0001] This invention relates to a control device and a control method.

[0002] This application is based on the main priority of Japanese Patent Application No. 2020-190354 filed on November 16, 2020, the contents of which are incorporated herein by reference. Background Technology

[0003] In the fields of power generation or industrial plants, control devices incorporating machine learning are employed to optimize plant operation. As an example, Patent Document 1 discloses a control device comprising a future state prediction device capable of rapidly predicting the state of a controlled object and its surrounding environment after an infinite number of time steps or steps within a predefined finite state space, using a probability density distribution. The future state prediction device includes: a future state prediction calculation unit that performs equivalent series calculations using a model simulating the future state of the controlled object and its surrounding environment in the form of a probability density distribution; and a control rule calculation unit that calculates the operational quantity of the controlled object using the result of the future state prediction calculation unit predicting the state of the controlled object and its surrounding environment after an infinite number of time steps or steps.

[0004] Existing technical documents

[0005] Patent documents

[0006] Patent Document 1: JP 2019-159876 Summary of the Invention

[0007] When the control device disclosed in Patent Document 1 is applied to a controlled object such as a factory, calculations are required for the future state prediction calculation unit to predict the future state. However, the states of the controlled object and its surrounding environment are already defined using probability density distributions. If the states of the controlled object and its surrounding environment are further defined in detail, the calculations for predicting the future state require more memory. As a result, the memory of the control device may be insufficient, potentially leading to a problem where the future state cannot be predicted.

[0008] Therefore, the object of the present invention is to provide a control device and control method that can reduce the memory used for predicting future states.

[0009] The control device of the present invention is characterized by comprising: a model building unit that builds a model simulating a controlled object; a subdivision unit that subdivides the model built by the model building unit; a control strategy calculation unit that uses the model subdivided by the subdivision unit to predict the future state of the controlled object and calculates a control strategy for the controlled object based on the predicted future state; and an operation instruction generation unit that generates operation instructions for the controlled object based on the control strategy calculated by the control strategy calculation unit.

[0010] In the control device of the present invention, since a subdivision unit is provided to subdivide the model constructed by the model building unit, the future state of the controlled object can be predicted based on the subdivision of the model, thereby reducing the memory used for predicting the future state.

[0011] Invention Effects

[0012] According to the present invention, it is possible to reduce the memory required for predicting future states. Attached Figure Description

[0013] Figure 1 This is a schematic diagram illustrating the configuration of the control device and associated devices of the embodiment.

[0014] Figure 2 It is a flowchart illustrating the control method of the control device.

[0015] Figure 3A It is a diagram illustrating the format of the data stored in the measurement signal database.

[0016] Figure 3B It is a diagram illustrating the format of the data stored in the measurement signal database.

[0017] Figure 4A This diagram illustrates the actions of the model building section and the problem subdivision section.

[0018] Figure 4B This diagram illustrates the actions of the model building section and the problem subdivision section.

[0019] Figure 4C This diagram illustrates the actions of the model building section and the problem subdivision section.

[0020] Figure 5A This is a diagram illustrating the state transition probability matrix.

[0021] Figure 5B This is a diagram illustrating the state transition probability matrix.

[0022] Figure 5C This is a diagram illustrating the state transition probability matrix.

[0023] Figure 5D This is a diagram illustrating the state transition probability matrix.

[0024] Figure 6A This is a diagram illustrating the actions of the control strategy calculation unit.

[0025] Figure 6B This is a diagram illustrating the actions of the control strategy calculation unit.

[0026] Figure 6C This is a diagram illustrating the actions of the control strategy calculation unit.

[0027] Figure 6D This is a diagram illustrating the actions of the control strategy calculation unit.

[0028] Figure 7 This is an example of memory usage displayed on a display device. Detailed Implementation

[0029] Hereinafter, embodiments of the control device and control method of the present invention will be described with reference to the accompanying drawings.

[0030] Figure 1 This is a schematic configuration diagram showing the control device and associated devices of an embodiment. The control device 10 of this embodiment is a device that uses a model to simulate the overall characteristics of the controlled object 20, calculates a control strategy for the controlled object 20 based on future states after an infinite amount of time or infinite number of steps, and generates operation instructions for the controlled object 20 according to the calculated control strategy. For example... Figure 1 As shown, the control device 10 is connected to the controlled object 20 and the external device 30, respectively. The controlled object 20 and the external device 30 correspond to the aforementioned associated devices.

[0031] The controlled object 20 is composed, for example, of a machine 21 constituting a power plant or similar structure, and a machine control unit 22 that controls the machine 21. Although not shown, the machine 21 is equipped with sensors that acquire operating data and image data of the machine 21. The operating data and image data acquired by the sensors are output to the machine control unit 22. Based on this operating data and image data, the machine control unit 22 generates various operating commands, outputs the generated operating commands to the machine 21, and performs various controls on the machine 21.

[0032] External device 30 may be, for example, a computer device (computer), equipped with an external input device 31 having a keyboard 311 and a mouse 312, and a display device 32 having a monitor capable of displaying images or data. Besides a computer, external device 30 may also be a portable terminal such as a tablet computer, smartphone, or laptop PC.

[0033] In this embodiment, the control device 10 and the controlled object 20, as well as the control device 10 and the external device 30, are configured to communicate via a network. Specifically, the external device 30 sends an instruction input via the external input device 31 as an external input signal 1 to the control device 10. The controlled object 20 sends the machine 21's operation data, image data, and operation commands generated by the machine control unit 22 as measurement signals 2 to the control device 10. Then, the control device 10 performs various processes based on the received external input signal 1 and measurement signal 2, further generates operation commands 6, and sends the generated operation commands 6 to the machine control unit 22 of the controlled object 20 and the display device 32 of the external device 30, respectively.

[0034] The control device 10 is, for example, a microcomputer composed of a CPU (Central Processing Unit) for performing calculations, ROM (Read Only Memory) as secondary storage for storing programs used for calculations, and RAM (Random Access Memory) as temporary storage for storing calculation progress or temporary control variables. It performs various processes such as calculations or decisions by executing the stored programs. Furthermore, the programs here can be configured to be sent to the control device 10 via a network.

[0035] The control device 10 mainly comprises a model building unit 11, a problem subdivision unit 12, a control strategy calculation unit 13, an operation instruction generation unit 14, a measurement signal database 15, and a processing result database 16. Figure 1 In this context, the database is represented by DB. Furthermore, the control device 10, serving as an interface for external connections, includes an external input interface 17 and an external output interface 18. The control device 10 communicates with the controlled object 20 and the external device 30 via these interfaces.

[0036] The model building unit 11 constructs a model of the simulated control object 20. More specifically, the model building unit 11 constructs a simulation model of the control object 20 by generating model data of the overall characteristics of the simulated control object 20. In addition, the model building unit 11 outputs the generated model data to the problem subdivision unit 12 and saves it in the processing result database 16.

[0037] The problem subdivision unit 12 is equivalent to the "subdivision unit" described in the claims, and it subdivides the model constructed by the model construction unit 11. More specifically, the problem subdivision unit 12 subdivides the model data generated by the model construction unit 11, outputs the subdivided model data to the control strategy calculation unit 13, and stores it in the processing result database 16.

[0038] The control strategy calculation unit 13 uses the model subdivided by the problem subdivision unit 12 to predict the future state of the controlled object 20, and calculates the control strategy for the controlled object 20 based on the predicted future state. More specifically, the control strategy calculation unit 13 predicts all future states after an infinite time or an infinite number of steps based on the model data subdivided by the problem subdivision unit 12 and the external input signal 1 received via the external input interface 17, and calculates the control strategy for the controlled object 20 from the predicted future states. Furthermore, the control strategy calculation unit 13 stores the calculated control strategy in the processing result database 16. In addition, the control strategy here refers to the processing of generating operation instructions for the controlled object 20, which will be explained in detail later.

[0039] The operation instruction generation unit 14 generates operation instructions for the controlled object 20 based on the control strategy calculated by the control strategy calculation unit 13. More specifically, the operation instruction generation unit 14 acquires the control strategy calculated by the control strategy calculation unit 13 and stored in the processing result database 16, and generates operation instructions for the controlled object 20 according to the acquired control strategy. Furthermore, the operation instruction generation unit 14 sends the generated operation instructions to the machine control unit 22 of the controlled object 20 and the display device 32 of the external device 30 via the external output interface 18.

[0040] The measurement signal database 15 receives and stores the measurement signal 2 sent from the control object 20 via the external input interface 17. The measurement signal 2 includes operational data and image data of the control object 20. Correspondingly, the measurement signal database 15 includes an operation database 151 that stores the operational data of the control object 20, and an image database 152 that stores the image data of the control object 20. Furthermore, the data included in the measurement signal 2 is not limited to operational data and image data.

[0041] Electronic information is stored in the operational database 151 and the image database 152, and is stored in a manner commonly referred to as electronic files (electronic data). Furthermore, these databases may be located external to the control device 10 and be able to connect to the control device 10 via a network.

[0042] The following is for reference Figures 2 to 6D The control method of the control device 10 (i.e., the operation of the control device 10) will be explained. In the following description, firstly, based on... Figure 2 The overall flow of the operation of the control device 10 is explained. Subsequently, based on... Figures 3A to 6D Please provide detailed explanations.

[0043] Figure 2 This is a flowchart illustrating the control method of the control device. For example... Figure 2As shown, firstly, in step S100, the control device 10 receives the measurement signal 2 sent from the controlled object 20 via the external input interface 17, and stores the received measurement signal 2 in the measurement signal database 15. The measurement signal 2 includes time-series operation data and image data related to the machine 21 of the controlled object 20. Thus, the control device 10 acquires time-series operation data and image data related to the machine 21. Then, the operation data is stored in the operation database 151, and the image data is stored in the image database 152.

[0044] Figure 3A and Figure 3B It is a diagram illustrating the format of the data stored in the measurement signal database. Figure 3A This is an example of operational data stored in operational database 151. Figure 3B This is an example of image data stored in image database 152. For example... Figure 3A As shown, the operating database 151 stores time-series data for each data item (item A, item B, item C, etc.) at regular intervals. Item A is, for example, temperature, item B is flow rate, and item C is pressure. Additionally, as... Figure 3B As shown, the image database 152 stores, for example, the temperature distribution measured at a certain profile of the machine 21 at intervals of sampling. Furthermore, the operating data and image data of the controlled object 20 can be displayed on the display device 32.

[0045] Additionally, in step S100, the control device 10 also receives the external input signal 1 sent from the external device 30 via the external input interface 17, and acquires the data contained in the received external input signal 1. The external input signal 1 received via the external input interface 17 is output to the control strategy calculation unit 13.

[0046] In step S101, following step S100, the control device 10 determines whether to update the control strategy based on pre-defined conditions. If it is determined that the control strategy should be updated, the process proceeds to step S102. On the other hand, if it is determined that the control strategy should not be updated, the process proceeds to step S105. Here, as a pre-defined condition, for example, when a certain period of measurement data (e.g., one week) has been accumulated, it is considered whether the characteristics of the model created by the model building unit 11 are consistent with the characteristics of the controlled object 20. If they are consistent, it is determined that the control strategy should not be updated; if they are inconsistent, it is determined that the control strategy should be updated. Furthermore, the pre-defined conditions are not limited to this content and can be arbitrarily set.

[0047] In step S102, the model building unit 11 acquires the measurement signal 2 stored in the measurement signal database 15 and generates model data 3 based on the acquired measurement signal 2. The generated model data 3 is output to the problem subdivision unit 12 and stored in the processing result database 16.

[0048] In step S103, following step S102, the problem subdivision unit 12 subdivides the model data 3 generated by the model construction unit 11 to generate subdivided model data 4. The generated subdivided model data 4 is output to the control strategy calculation unit 13 and stored in the processing result database 16. Furthermore, in subsequent uses... Figures 4A to 5D This section describes the details of steps S102 related to model building and steps S103 related to problem subdivision.

[0049] In step S104, following step S103, the control strategy calculation unit 13 calculates the control strategy based on the reward function included in the external input signal 1 received via the external input interface 17 and the subdivision model data 4 generated by the problem subdivision unit 12, thereby generating a control strategy signal 5. The generated control strategy signal 5 is stored in the processing result database 16. This will be used later. Figures 6A to 6D This section describes the details of step S104, which involves calculating the control strategy.

[0050] In step S105, the model building unit 11 generates a state ID7 based on the measurement signal 2 at the current moment. The generated state ID7 is output to the operation command generation unit 14.

[0051] In step S106, following step S105, the operation instruction generation unit 14 acquires the control strategy signal 5 stored in the processing result database 16, and uses the acquired control strategy signal 5 and the status ID 7 generated by the model building unit 11 to generate an operation instruction 6 for the controlled object 20. The generated operation instruction 6 is sent to the machine control unit 22 of the controlled object 20 and the display device 32 of the external device 30 via the external output interface 18. Furthermore, for the method of generating operation instructions based on the control strategy signal and the status ID, already known techniques can be used, and their details are omitted.

[0052] In the controlled object 20, the machine control unit 22 controls the machine 21 according to the sent operation command 6. In this way, by using the control device 10, the measured values ​​of temperature, flow rate, and pressure of the machine 21, as well as the temperature distribution, can be controlled to an optimal state. On the other hand, in the external device 30, the display device 32 displays the content of the sent operation command 6 and images such as trend graphs on a monitor. By viewing the content displayed on the monitor, the operator can confirm the content of the operation command 6.

[0053] In step S107, following step S106, the control device 10 determines whether to terminate control. If it is determined that control has not terminated (in other words, control continues), the process returns to step S100. On the other hand, if it is determined that control has terminated, the series of processes ends.

[0054] Next, use Figures 4A to 5D The steps S102 related to model building and S103 related to problem subdivision are explained in detail.

[0055] In this embodiment, an example is given of the model building unit 11 building a model relative to the control object 20 having the machine 21 and the machine control unit 22. However, as an example of the control object 20, it can be any mechanical or biological action, natural or physical phenomenon, chemical reaction, change in money or price, change in consumer needs, etc., and is not limited to the example described here.

[0056] Furthermore, in this embodiment, the input to the model is the state of the simulated object (i.e., the controlled object) and influencing factors such as time elapsed, operations, and external disturbances, and the output of the model is the state of the simulated object after being affected by the influencing factors. As for the form of the model constructed using the model construction unit 11, it is considered to represent a neural network, a radial basis function network, or a matrix expressing the weights of a neural network and a radial basis function network, or a state transition probability matrix, but is not limited to these matrices.

[0057] Regarding the method for constructing a model using a state transition probability matrix, the content disclosed in Patent Document 1 can be used, for example. That is, the data is discretized from the constant measurement data of the simulated object (in this embodiment, the data included in the measurement signal 2 of the control object 20) by referring to a table defining state IDs. However, data discretization can also be performed using data clustering techniques such as vector quantization or adaptive resonance theory. Therefore, by using these techniques in a way that corresponds to the state of the control object, a suitable simulation model of the control object 20 can be constructed. Furthermore, at this time, the model construction unit 11 discretizes at least one of the operating data and image data included in the measurement signal 2 using a reference table, adaptive resonance theory, or vector quantization, thus requiring only the definition of the state transition matrix's state. Additionally, when discretizing the data, it is preferable that the deviation of the reward falling into a single state ID is very small.

[0058] The following describes the process of the model construction unit 11 constructing the state transition probability model (step S102). Figure 4AIn this model, s1 to s8 represent the states of the state transition probability model, and lines are used to connect the states that can transition over time Δt. For example, a state s1 can transition to one of the states s1, s2, s3, or s4. Therefore, the model building unit 11 defines the states of the state transition matrix based on information obtained from the measurement signal database 15 (here, measurement signal 2), and models the relationships between the states that can transition over time Δt.

[0059] Figure 4B as well as Figure 4C This diagram illustrates the process of the problem subdivision unit 12 subdividing the model (step S103). Here, the problem subdivision unit 12 divides the model into multiple models by aggregating multiple states into a single cluster. In this invention, "dividing into models" is used to mean subdividing the problem. Furthermore, the problem subdivision unit 12 can subdivide the problem in a manner that ensures the number of states falling into the cluster is within a specified range, or the problem can be arbitrarily subdivided by the operator.

[0060] exist Figure 4B In the example, the problem subdivision unit 12 subdivides the problem into two problems through the synthesis states s1-s4 and s5-s8 (see reference). Figure 4B (Two circles in the middle). Figure 4C In the example, the problem subdivision unit 12 subdivides the problem into two problems by synthesizing states s1-s6 and s7-s8 (see reference). Figure 4C (Two circles in the diagram). In any example, the comprehensive clustering is divided into two, that is, the problem as a whole is divided into two problems for learning.

[0061] Here, the problem subdivision section 12 preferably subdivides the model in a manner that minimizes the number of joints between the comprehensive clusters. For example, in Figure 4B In the example shown, the union number is 1. Figure 4C In the example shown, the union number is 3, and... Figure 4C compared to, Figure 4B The number of associations is less, therefore, as long as like Figure 4B The problem can be subdivided in this way. This is because the fewer the number of joints, the fewer the elements in the state transition matrix, and therefore, the memory required to store the model subdivided by the problem subdivision unit 12 can be reduced accordingly. As a result, the memory required for predicting future states can be reduced.

[0062] Furthermore, regarding the subdivision method based on the problem subdivision section 12, it is acceptable as long as the model as a whole can be divided into multiple parts, and it is not limited to the above-mentioned content.

[0063] Figures 5A to 5DThis is a diagram illustrating the state transition probability matrix. The state transition probability matrix is ​​typically used in models that simulate the motion characteristics or physical phenomena of a controlled object; it is a function (or matrix) that stores the transition probabilities between all states. Figures 5A to 5D In the table, the rows represent the source states si (i = 1, 2, ..., n (here, n = 8)), and the columns represent the destination states sj (j = 1, 2, ..., n (here, n = 8)). The state transition probability P(sj|si) is the probability of transitioning from state si to state sj when the element Tij has undergone a pre-defined increment time Δt (or, steps). Furthermore, when the state transition probability P(sj|si) is independent of time, the influencing factor can be used instead of time to represent the amount or number of steps that interfere with the simulated object.

[0064] Figure 5A This shows a model that simulates the overall characteristics of the controlled object, and is related to... Figure 4A The corresponding state transition probability matrix. Figure 5A In this context, Tij represents the probability of transitioning from states s1 to s8 of the transfer source to states s1 to s8 of the transfer destination. Figures 5B to 5D Models representing problem segmentation (segmentation models), such as Figure 4B The image shows the state transition probability matrix after subdividing the model. That is, Figures 5B to 5D like Figure 4B That is a matrix showing the individual clusters and the relationships between them.

[0065] Then, in Figure 5A The state transition probability matrix has 8×8=64 elements, in contrast, Figures 5B to 5D A total of 4×4 + 4×4 + 2×2 = 36 elements were used. In this way, by subdividing the model to define the state transition probability matrix, the number of elements in the matrix can be reduced, thus reducing the memory required to store the matrix. Therefore, the memory required to store the model subdivided by the problem subdivision unit 12 can be reduced accordingly, and the memory used for predicting future states can be reduced.

[0066] In addition, Figures 4A to 5D The example described is an example with 8 states, but there is no limit to the number of states. The more states there are, the more significant the effect of reducing the memory required for the storage matrix becomes, and the easier it is to apply the control device 10 to the controlled object 20.

[0067] Next, use Figures 6A to 6D The steps related to the control strategy calculation are explained in detail below. As described above, the control strategy calculation unit 13 uses the subdivided model data 4 to predict all future states after an infinite time or an infinite number of steps, and calculates the control strategy signal 5 of the controlled object 20 based on the predicted future states.

[0068] Specifically, firstly, the control strategy calculation unit 13 calculates a decaying state transition matrix using a model of state transition probabilities subdivided by the problem subdivision unit 12, and then predicts the future state. The method for calculating the decaying state transition matrix can be, for example, using the following equation (1). In equation (1), the stored form of the model is assumed to be a state transition probability matrix T.

[0069]

Mathematical Formula 1

[0070] D=T+γT 2 +γT 3 +…+γ ∞-1 T ∞ (1)

[0071] In equation (1), D is the decay-type state transition matrix, γ is the decay rate, and is a constant greater than 0 and less than 1. Additionally, T... k Let D be a function (or matrix) storing the transition probabilities between all states after a time interval of Δt×k. In this way, the decaying state transition matrix D represents the transition probability matrix T from the state transition probability matrix T after a time interval of Δt to the state transition probability matrix T after a time interval of Δt×∞. ∞ The sum of these values ​​also forms a matrix that preserves the statistical proximity between all states. Furthermore, since the weights of states transitioning further into the future decrease, they are multiplied by a decay rate γ corresponding to the elapsed time. Additionally, k here is a discretized positive integer.

[0072] We need to transform the state transition probability matrix T into the state transition probability matrix T. ∞ Equation (1) is difficult to perform in real time. Therefore, equation (1) is transformed into equation (2) below.

[0073]

Mathematical Formula 2

[0074] D = T(E - γT) -1 (2)

[0075] In equation (2), E is the identity matrix. Equation (2) is the equivalent calculation formula to equation (1). This is achieved by transforming the state transition probability matrix T from equation (1) to the state transition probability matrix T... ∞ The calculation of the sum is transformed into the inverse matrix of (E-γT) in equation (2), and the same calculation result as in equation (1) is obtained within a finite time. Here, a pseudo-inverse matrix can also be used when the state transition probability matrix T is not linearly independent. Alternatively, the decaying state transition matrix D can be replaced by a matrix that normalizes the decaying state transition matrix in each row.

[0076] In this way, by setting the model that simulates the behavior of the simulated object as a state transition model, using Tk The state transition probability is calculated after a time interval Δt × k. Additionally, the state transition probability matrix T is obtained from the state transition probability matrix T after a time interval Δt to the state transition probability matrix T after a time interval Δt × ∞. ∞ The sum is weighted based on the decay rate γ using the elapsed time, and the state transition probability after Δt×∞ time is calculated within a finite time.

[0077] Next, the control strategy calculation unit 13 calculates the control strategy based on the reward function included in the external input signal 1. Here, the reward function is a function that expresses the control objective, such as the target position or target velocity, using the form of a function, table, vector, matrix, etc.

[0078] Figure 6A An example is shown when the reward function is in vector form. Figure 6A In this invention, the reward function R is numerically represented by the state ID of each transition source. State s is a discrete space that divides the whole into n ranges (here, n = 8). The transition from the initial state to state s3 is set as the goal. The element value of the vector that becomes the goal sets state s3 to 1, and the other states to 0. Furthermore, in this invention, the element value of the vector or the value of the reward function R is referred to as the reward.

[0079] The control strategy calculation unit 13 calculates the control strategy of the controlled object 20 by calculating the optimal control rule (i.e., the optimal operating quantity) based on the reward function R and the decay-type state transition matrix D calculated above. Figure 6B An example of a control rule is shown. Figure 6B In this context, the operand ID is numerically represented by the state ID of each transfer source. The state s is a discrete space that divides the entire system into n ranges (here, n = 8). The optimal operand ac (c = 1, 2, ..., m) is stored relative to the range of each state. The method for calculating the optimal operand ac will be explained later.

[0080] The following is an example of a method for calculating the optimal control rule. Here, the control strategy calculation unit 13 uses the following three stages to calculate the optimal control rule.

[0081] Phase 1: First, calculate and save each state s and set the state s as the objective using the reward function R. goal The state value function is a function of the proximity (or a statistical indicator representing the ease of transfer). In this invention, this function is called the state value function V. In addition, the state value function V can also be stored in the form of a table, vector, matrix, etc., in addition to being a function. In this invention, there is no limitation on the storage form. The following equation (3) shows an example of the method for calculating the state value function V.

[0082]

Mathematical Expression 3

[0083] V=DR tr (3)

[0084] As shown in equation (3), the state value function V is R, which is the transpose of the decaying state transition matrix D and the reward function R. tr The product of. For example, the state value function V becomes Figure 6C This is an n-dimensional (here, n=8) vector. The easier it is to move towards the state s that becomes the goal. goal The higher the transition state, the higher the element value of the state value function V. In this invention, this element value is referred to as value. Furthermore, the state value function V of this invention is equivalent to the definition and value of the state value function in reinforcement learning.

[0085] Phase 2: Next, using the state value function V, among the states sj of the destination that can be transitioned from the source state si, calculate the state s that is most likely to become the destination for each state si of the source. goal The transition state sj*. Equation (4) below shows an example of how to calculate state sj*.

[0086]

Mathematical Expression 4

[0087] sj*=argmax(V(sj)T(si,sj)) (4) Here, T(si,sj) refers to the element values ​​of the si row and sj column in the state transition probability matrix T. Figure 6D An example of the calculation result of equation (4) is shown. Figure 6D In this context, the state ID of the transfer destination is expressed for each transfer source's state ID.

[0088] Phase 3: In the final phase, the operation amount *a* required to transition from each state *si* of the transfer source to the state *sj* obtained in Phase 2 is calculated. The operation amount *a* can be calculated by finding the inverse model (a model that takes the states *si* and *sj** of the transfer source as inputs and outputs the corresponding operation amount *a*). The calculation results of Phase 3 are, for example... Figure 6B This is the control rule shown.

[0089] In this way, by using the above equation (3) to calculate the value, it is possible to evaluate the value of s in each state. goal The ease of transition is determined by using the above equation (4) to specifically (determine) the state that is most likely to transition to s after a time interval of Δt. goal The state sj* is transitioned to using the inverse model-specific operation a for transitioning to state sj*.

[0090] The control device 10 according to this embodiment includes a problem subdivision unit 12 that subdivides the model constructed by the model construction unit 11. Therefore, by predicting future states based on model subdivision, the memory used for predicting future states can be reduced. More specifically, the memory used for model construction (e.g., the memory storing the model constructed by the model construction unit 11 and the memory storing the model subdivided by the problem subdivision unit 12) can be reduced by the subdivision of the problem subdivision unit 12 compared to the case of constructing a conventional state transition probability model. As a result, the effect of reducing the memory used by the control device 10 can be expected.

[0091] To facilitate confirmation of the aforementioned memory reduction effect, in this embodiment, the display device 32 preferably also displays the memory usage status.

[0092] Figure 7 This is an example of how the memory usage status of a display device is shown. For example... Figure 7 As shown, the display device 32 displays items such as "Potential memory used" (i.e., memory that the control device 10 can use), "Memory used for constructing the state transition probability model," and "Memory used for constructing the problem subdivision model." Specific display results can be, for example, the usage capacity of each memory or a ratio relative to the overall capacity.

[0093] Therefore, the usage status of each memory can be easily monitored through the display device 32, and the effect of memory reduction can be easily confirmed. In addition, the "memory used for constructing the state transition probability model" here refers to the memory that stores the model constructed by the model construction unit 11, and the "memory used for constructing the model after problem subdivision" refers to the memory that stores the model subdivided by the problem subdivision unit 12.

[0094] Alternatively, the display device 32 can arbitrarily display the results obtained by the control device 10 of this embodiment, such as the operation result of the information compression means, the number of clusters, the number of comprehensive clusters, the compression efficiency, the number of joints, and the amount of memory reduction.

[0095] The embodiments of the present invention have been described in detail above, but the present invention is not limited to the above embodiments. Various design changes can be made within the spirit of the present invention without departing from the scope of protection.

[0096] For example, a display unit may be further provided in the control device 10. By displaying the usage status of the "available memory", "memory used for constructing the state transition probability model" and "memory used for constructing the problem subdivision model", as well as the intermediate and final results obtained by the operation of the control device 10, this content can be easily confirmed on the control device 10 side.

[0097] Furthermore, a display step can be added to the control method of the control device 10. This display step shows the usage status of the aforementioned "possible memory," "memory used for constructing the state transition probability model," and "memory used for constructing the problem subdivision model," as well as intermediate and final results obtained through the actions of the control device 10. For example, this display step can be added between step S106, which relates to generating operation instructions, and step S107, which relates to determining the end of control. This allows for monitoring the usage status of each memory or the results of the control device 10's actions.

[0098] Explanation of reference numerals in the attached figures

[0099] 1 External input signal

[0100] 2. Measurement signal

[0101] 5 Control Strategy Signals

[0102] 6 Operation Instructions

[0103] 10 Control Devices

[0104] 11 Model Construction Department

[0105] 12 Problem Subdivision Department

[0106] 13 Control Strategy Calculation Department

[0107] 14 Operation Instruction Generation Unit

[0108] 15 Measurement Signal Database

[0109] 16 Processing Results Database

[0110] 17 External Input Interfaces

[0111] 18 external output interfaces

[0112] 20 Control Objects

[0113] 21 Machines

[0114] 22 Machine Control Department

[0115] 30 External Devices

[0116] 31 External Input Device

[0117] 32 display devices

[0118] 151 Operating Database

[0119] 152 Image Database

[0120] 311 Keyboard

[0121] 312 mouse.

Claims

1. A control device, characterized in that, have: The model construction department is responsible for constructing models of the simulated control objects. The subdivision unit further subdivides the model constructed by the model construction unit; The control strategy calculation unit uses the model subdivided by the subdivision unit to predict the future state of the controlled object, and calculates the control strategy of the controlled object based on the predicted future state; as well as The operation instruction generation unit generates operation instructions for the controlled object based on the control strategy calculated by the control strategy calculation unit. The subdivision unit subdivides the model by dividing the model constructed by the model construction unit into multiple subgroups by aggregating multiple states into a single comprehensive cluster, and by reducing the number of joints that combine the comprehensive clusters.

2. The control device according to claim 1, characterized in that, The model construction unit constructs the model by using the overall characteristics of the controlled object as a state transition matrix. The subdivision unit subdivides the state transition matrix.

3. The control device according to claim 2, characterized in that, The model building unit defines the state of the state transition matrix by discretizing at least one of the operating data and image data of the controlled object using a reference table, adaptive resonance theory, or vector quantization.

4. The control device according to any one of claims 1 to 3, characterized in that, It also includes a display unit that displays at least a memory storing models constructed by the model building unit and a memory storing models subdivided by the subdivision unit.

5. A control method, characterized in that, include: The model building steps involve constructing a model of the simulated control object. The subdivision step further subdivides the model constructed in the model construction step. The control strategy calculation step involves using the model subdivided in the subdivision step to predict the future state of the controlled object, and calculating the control strategy of the controlled object based on the predicted future state. as well as The operation instruction generation step generates operation instructions for the controlled object based on the control strategy calculated in the control strategy calculation step. In the subdivision step, the model is subdivided in a way that the model built in the model building step is divided into multiple subgroups by aggregating multiple states into a single comprehensive cluster, and the number of joints that combine the comprehensive clusters is reduced.

6. The control method according to claim 5, characterized in that, In the model construction step, the characteristics of the controlled object as a whole are used as a state transition matrix to construct the model. In the subdivision step, the state transition matrix is ​​subdivided.

7. The control method according to claim 6, characterized in that, In the model construction step, the state of the state transition matrix is ​​defined by discretizing at least one of the operating data and image data of the controlled object using a reference table, adaptive resonance theory, or vector quantization.

8. The control method according to any one of claims 5 to 7, characterized in that, It also includes a display step, in which at least the memory of the model built in the model building step and the memory of the model subdivided in the subdivision step are displayed.

Images