Learning model generation device

Abstract

To adequately capture a probabilistic characteristic of a process behavior with a minimum requirement of an amount of information, and generate a learning model which is highly accurate, safe, and suitable for controlling a target process without instability due to missing data.SOLUTION: A learning model generation device that combines plural signals to generate a state transition probability matrix for a defined multidimensional state includes an arithmetic processing unit for: setting an initial value for a lower limit value, an upper limit value, and a division number for each dimension of the multidimensional state; calculating a learning data sufficiency rate, which is a total value of a transition probability to a state near a distribution center of a transition destination, and is defined as one or more pre-transition states within a predetermined distance from a target pre-transition state or a transition destination state with a maximum transition probability; and decreasing the division number when the calculated learning data sufficiency rate is below a defined range, and increasing the division number when the calculated learning data sufficiency rate exceeds the defined range.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a learning model generation device that is applied to process control of a plant or the like. [Background technology]

[0002] There are high expectations for the application of artificial intelligence (hereinafter referred to as AI) technology to process control in industrial plants, etc. Because process control is mission-critical, the introduction of AI technology requires explainability, and in this respect, methods based on state transition probability models are suitable. An example of such a method is the method described in Patent Document 1. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2019-159876 Summary of the Invention [Problem to be solved by the invention]

[0004] In order to utilize a state transition probability model for process control, there are challenges specific to the control application, and it is necessary to properly grasp the characteristics of the process behavior and construct a learning model. For example, it is extremely important for practical purposes to set the signals (which will often be multiple) that make up the state, and to define the upper and lower limits and number of divisions (or division width) of the signal values.

[0005] Furthermore, to evaluate the validity of a state transition probability model generated by learning, it is necessary to grasp the trends in the transition probabilities between each state divided into meshes in multidimensional space and evaluate their validity. However, it is difficult to visualize transition probabilities in multidimensional space, and analyzing these trends requires knowledge and experience.

[0006] Thus, in order to build an appropriate learning model that can be used for control, it has traditionally been necessary for experts with the knowledge and experience to recognize the issues mentioned above and to spend time repeatedly analyzing and adjusting the signals that make up the state, changing the upper and lower limits and division widths of each signal value, and making repeated adjustments through trial and error. This work is difficult to automate, and manual work based on specialized knowledge has been unavoidable. [Means for solving the problem]

[0007] The present invention includes multiple means for solving at least some of the above problems, and an example thereof is as follows: That is, a learning model generation device for generating a state transition probability matrix of a multidimensional state defined by combining multiple signals, the learning model generation device having a processing unit that sets initial values ​​for a lower limit value, an upper limit value, and a division number for each dimension of the multidimensional state, calculates a learning data fulfillment rate, which is the sum of the transition probability to a destination state near the distribution center, which is defined as being within a predetermined distance from the destination state or the destination state with the maximum transition probability, for one or more pre-transition states, and reduces the division number if the calculated learning data fulfillment rate is below a specified range, and corrects the division number to be increased if the calculated learning data fulfillment rate is above the specified range. [Effects of the Invention]

[0008] It is possible to generate a learning model that is suitable for controlling the target process, which is highly accurate, accurately captures the probabilistic characteristics of process behavior with the minimum amount of information required, and is safe and free from instability due to missing data.

[0009] Problems, configurations, and effects other than those described above will become apparent from the following description of the preferred embodiments of the invention. [Brief explanation of the drawings]

[0010] [Figure 1] 10 is an example of a state transition probability matrix in the present embodiment. [Figure 2] 1 shows the overall configuration of a control system according to the present embodiment. [Figure 3] 10 is a processing flow example of mesh definition in this embodiment (embodiment 1). [Figure 4] 10 is a processing flow example for correcting the number of divisions in this embodiment (embodiment 1). [Figure 5] 10 is a processing flow example of mesh definition in this embodiment (embodiment 2). [Figure 6] 10 is a process flow example for correcting the number of divisions in this embodiment (embodiment 2). [Figure 7] 10 is a processing flow example of mesh definition in this embodiment (embodiment 3). [Figure 8] 10 is a diagram showing an example of a processing flow of mesh definition in this embodiment (embodiment 4). [Figure 9] 10 is a processing flow example for correcting the upper and lower limit values ​​of dimension k in this embodiment (embodiment 4). [Figure 10] 10 is a flowchart illustrating an example of a mesh definition process in this embodiment (embodiment 5). [Figure 11] 10 is a processing flow example for correcting the upper and lower limit values ​​of dimension k in this embodiment (embodiment 5). [Figure 12] 10 is a flowchart illustrating an example of a mesh definition process in this embodiment (embodiment 6). [Figure 13] 10 is a processing flow example for correcting the number of divisions for each dimension in this embodiment (embodiment 7). [Figure 14] 10 is a processing flow example for generating candidates for the number of divisions in this embodiment (embodiment 7). [Figure 15] 13 is an example of a list (screen) of candidates for the number of divisions in this embodiment (embodiment 7). [Figure 16] 10 is an example of a screen display of a learning data fulfillment rate and the like in this embodiment (embodiment 7). [Figure 17] 10 is an example of a screen display of the number of vertices etc. in this embodiment (embodiment 2). DETAILED DESCRIPTION OF THE INVENTION

[0011] First, let us explain the state transition probability model. In a state transition probability model, the state of a plant process (hereafter referred to as "state" for simplicity) is expressed by combining the measurement values ​​required to represent the target process. A lower limit, upper limit, and number of divisions (or division width) are determined for each measurement signal, and a multidimensional space with as many dimensions as the number of signals (each signal's numerical axis corresponds to one dimension) is divided into regions (hereafter referred to as "mesh division") for each signal. For example, if four measurement signals are selected to make up a state and the number of divisions between the upper and lower limit values ​​of each signal is set to 20, the number of regions (meshes) into which the four-dimensional space consisting of the four signals is divided, i.e., the number of states, will be 20 x 20 x 20 x 20 = 160,000, depending on the combination of signal values.

[0012] In process control based on a state transition probability model, a state transition probability matrix, which is a probabilistic model of the likelihood of transitions occurring between all states defined as above (between each region divided into meshes in a multidimensional space consisting of a large number of signals) in a predetermined time period, is used to calculate control command values.

[0013] A state transition probability matrix is ​​a matrix in which each state defined by dividing the multidimensional space into regions as described above is assigned a serial number, and the probability of a state transition from the i-th to the j-th state in a predetermined cycle is assigned to the i-th row, j-th column element. Figure 1 shows an example of a state transition probability matrix. The state transition probability matrix is ​​constructed using operational data and simulation data of the target process. This construction process is hereinafter referred to as learning.

[0014] When using the state transition probability matrix for control, for example, if the current state deviates from the control target value, the state to which it is desirable to transition next in order to change the process from the current state toward a state corresponding to the control target value (hereinafter referred to as the target state) is calculated using the state transition probability matrix, and a control command value for realizing such a state transition is calculated.

[0015] The state transition probability matrix thus functions as a learning model, and in this application, this will be referred to as a learning model as appropriate. As shown in Patent Document 1, among learning models, there is also a form of a matrix in which the state transition probability matrix is ​​a sum of a series with a damping coefficient, and this is also included.

[0016] In this embodiment, determining the upper and lower limits and the number of divisions for the signal values ​​of each dimension that make up the state is called mesh definition. Here, the signal refers not only to a signal that is directly measured, but also to a variable calculated based on the measured signal. The difference between the upper and lower limit values ​​of each dimension is called the upper / lower limit width, and the upper / lower limit width divided by the number of divisions is called the division width.

[0017] Furthermore, within the space in which a transition can occur from a pre-transition state in a multidimensional space, a range of a predetermined distance from the pre-transition state or the transition destination state with the maximum transition probability is called the vicinity of the distribution center of the transition destination (or within the distance near the distribution center). Furthermore, the sum of the transition probabilities to the states near the distribution center of the transition destination, with the pre-transition state as the target, is called the learning data sufficiency rate.

[0018] As mentioned above, in order to utilize such a state transition probability model for process control, it is extremely important from a practical standpoint to properly capture the characteristics of the process behavior and define the mesh.

[0019] For example, if the division is too coarse, the transition probability between states will not be dispersed, and a learning model will be generated in which the state often does not transition over time (i.e., remains in the same state). A learning model with almost no state transitions like this is not suitable for the purpose of control, which is to use state transition patterns to bring the process state closer to the control target.

[0020] On the other hand, if the division is too fine, the number of defined states increases, posing the problem of requiring a large amount of training data for comprehensive learning. Furthermore, if the training data is insufficient, problems arise due to the generation of many states for which no state transition history exists. For such states for which no transition history exists, the probability of transitioning from other states to that state is zero, and the probability of transitioning from that state to other states is also zero. However, during actual plant control, such process states are quite likely to occur, and in such cases, there is no transition pattern to approach the control target, making it impossible to calculate how the process state should transition.

[0021] As another example of a case where the division is too fine, if the state definition is divided finer than the resolution of the measurement values ​​transmitted within the control system, a learning model will be generated in which states with zero and non-zero state transition probabilities alternate along the division of that dimension of the learning model. In this case, the following two problems arise.

[0022] First, the information that the state transition probability is zero merely represents the missing data caused by the state division, and does not represent the probability distribution of the actual behavior of the process. Therefore, control using such a learning model is at high risk of unstable operation.

[0023] Second, in actual control, if the range of change in the state transition is too small, the optimal transition pattern of the process state calculated using such a finely divided state transition probability model may be smaller than the step size (resolution) of the input signal used to calculate the control command within the control system. In such cases, even if the optimal state transition pattern is calculated, it will not be reflected in the actual control operation, resulting in the problem that the control command value will remain constant and become uncontrollable.

[0024] Care must also be taken when setting the upper and lower limits of the domain of each signal that makes up the state.

[0025] For example, if the upper and lower limits of the domain are set broadly to encompass the range of events that the process can experience, many undefined state segments will be created near the upper and lower limits of the domain, where no state transition records exist. As mentioned above, when a state with undefined transition records occurs during actual control, it is impossible to calculate how the process state should transition. Furthermore, if the domain is set broadly, the number of divisions must be increased according to the breadth of the domain. This increases the amount of memory required to hold a learning model with multidimensional state definitions, and the amount of data required to perform learning becomes enormous. In addition, the computational load for calculating the optimal state transition pattern increases.

[0026] Conversely, if the upper and lower limits of the domain are narrowed too much, a very small portion of the distribution range will be expanded and extracted, potentially resulting in a model with a transition probability that is nearly uniform over a wide range within the domain.If the state transition pattern of the acquired model is nearly uniform, it will essentially be the same as determining the transition pattern by randomly assigning numbers, and therefore cannot be used effectively for the purpose of control, which is to control the controlled object to a target value by following states that are statistically likely to transition.

[0027] In view of such problems, this embodiment provides a method for appropriate mesh definition in a state transition probability model.

[0028] <Overall system configuration> First, we will explain where the present invention is applied in process control. As shown in Figure 2, the target control system is a control system that includes a controlled process 1000 having a sensor 1001 that measures the value of a controlled variable of the process (a target signal to be controlled to a desired value) and a control terminal 1002 that manipulates the input to the process so that the controlled variable approaches a control target value, a controller 2000 that receives the measurement signal acquired by the sensor 1001 and outputs a command value of the manipulated variable (hereinafter referred to as the manipulated variable for simplicity) to the control terminal 1002, and a learning device 3000 that builds and stores information on the past behavior of the process 1000, which is used by the controller 2000 to calculate the manipulated variable, based on past measurement values ​​acquired by the sensor 1001 of the target process 1000.

[0029] Here, the controlled process 1000 is, for example, various machines in a factory (incinerators, reactors, heat exchangers, etc.), and the controlled quantity of the process is a quantity that is intended to follow a target value, such as the temperature of a reactor.

[0030] In such a learning device 3000, the measurement values ​​of the operating data of the target process 1000 acquired by the sensor 1001 are received at an input terminal 3001 provided in the device 3000 and transmitted to a learning model generation unit 3010 provided in the device 3000.

[0031] In the learning model generation unit 3010, based on the accumulated measurement values, the behavior characteristics of the controlled process 1000 (in this embodiment, a state transition probability model) are constructed and stored as a learning model 3020. The learning model generation unit has an arithmetic processing unit (processor) (not shown) and a memory in which programs for various processes are stored, and the processor executes these programs to realize desired functions.

[0032] The constructed and held state transition probability model is then output to the control device 2000 from an output terminal 3002 provided in the learning device 3000 .

[0033] The control device 2000 calculates a command value of the manipulated variable to be transmitted to the operation terminal 1002 of the process 1000 based on the state transition probability model received in this way and real-time measurement information about the process 1000 received from the sensor 1001 .

[0034] 2, the learning device 3000 and the control device 2000 are shown separately, but this is done for ease of understanding, and the learning device 3000 may be incorporated into the control device 2000. Furthermore, the learning model generation unit 3010 itself may be configured using a personal computer or the like, and may be realized as an independent device (learning model generation device).

[0035] Of the above control system configurations, the following examples relate to the learning model generation unit 3010 that executes the construction of the learning model 3020. That is, in the following examples, several processing flows for mesh definition executed by the learning model generation unit 3010 are explained below. These relate to a method for determining the number of divisions for dividing the numerical axis of each dimension into meshes by referring to the distribution of transition probabilities to all states within the definition range, starting from one state before the transition.

[0036] In the following description, when describing processing by a program, the program, functional units, etc. may be described as the main components, but the main hardware components are a processor or an information processing device (computer) configured to include the processor, etc. The information processing device executes processing according to a program read into memory using resources such as memory and a communication interface as appropriate. While FIG. 2 shows an example of a CPU as the processor, a GPU (Graphical Processing Unit) or the like may also be used. Furthermore, processing to realize a function is not limited to software program processing, but can also be implemented using a dedicated circuit. The dedicated circuit may be a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), or the like. [Example]

[0037] 3 shows the processing flow for mesh definition in this embodiment, starting from one state before a transition. In this flow, upper and lower limits for each dimension are set in step S101, then an initial value for the number of divisions for each dimension is set in step S102, and the number of divisions for each dimension is corrected based on the training data sufficiency rate in step S104. The processing contents of steps S101, S102, and S104 are described in detail below.

[0038] <Step S101> In general, from an engineering perspective, specific methods for setting the upper and lower limits of each dimension in step S101 include: 1) setting the maximum and minimum range based on the definition of the measurement signal, and 2) using the maximum and minimum range based on actual performance. Furthermore, from the perspective of using a state transition probability matrix for control as in this embodiment, 3) narrowing the statistical distribution range of the signal value to a range with a relatively high occurrence frequency is useful.

[0039] Suitable ranges include, for example, a range of 1 to 3 standard deviations above and below the mean value (±1σ, ±2σ, ±3σ), or a statistical 90% distribution range (a range where the cumulative probability is between the bottom 5% and the top 5%).

[0040] The reason for this is that narrowing the upper and lower limit values ​​to a range with a high occurrence frequency, as in 3) above, effectively avoids the two problems described below. The first problem that can be avoided here is that if the upper and lower limit values ​​are set as the maximum and minimum range based on the definition or actual performance of the measurement signal, many undefined state categories with no state transition performance results will arise near the upper and lower limits, making it impossible to calculate how the process state should transition when this state occurs during actual control. The second problem that can be avoided is that if upper and lower limits are set in the same way, the amount of memory required for multidimensional state definitions will increase, resulting in a higher calculation load for calculating the optimal state transition pattern.

[0041] <Step S102> The initial value of the number of divisions for each dimension in step S102 can be set efficiently by using at least one of the following two methods.

[0042] <First Method of Step S102> The first method for executing step S102 is to set the initial value of the number of divisions to a small value if priority is given to keeping the amount of memory required to hold the learning model (hereinafter referred to as consumed memory amount) small, and to set the initial value of the number of divisions to a large value if priority is given to the accuracy of the learning model.

[0043] In this way, it is possible to efficiently initialize the mesh definition for efficiently generating an appropriate learning model according to the characteristics, requirements, and constraints of the entire target process, including the learning and control system, such as the characteristics of the target process, the requirements for control, constraints on the amount of memory available for this learning model generation means, and constraints on the amount of memory available for the control system that performs control using the learning model generated by this learning model generation means.

[0044] When prioritizing reduction in memory consumption of the learning model, it is appropriate to set the initial value of the number of divisions to at least 3. This provides at least three divisions: a standard value, a large value, and a small value, enabling a minimum control operation to transition the plant process state in any direction (i.e., increase, decrease, maintain the value, or change in any direction along the axis of the signal dimension).

[0045] When prioritizing the accuracy of the learning model, it is preferable to select the initial value of the number of divisions from a range of about 20 to 500. Setting it in this way satisfies the required accuracy requirements for the controlled amount of the target process, which will be described below, and also makes it possible to avoid an explosive increase in the memory requirements of a learning model defined by a multidimensional mesh (for example, if the number of divisions for each dimension is set to Ndiv and the total number of dimensions is Ndim, then the total number of meshes will increase exponentially as the dimensions increase, as Ndiv^Ndim).

[0046] <Second Method of Step S102> The second method for executing step S102 is to set the initial value of the division number large for signals constituting the state dimension whose values ​​are to be made to coincide with the control target (hereinafter referred to as controlled variables), and to set the initial value of the division number small for signals that are factors causing fluctuations in the controlled variables (hereinafter referred to as explanatory variables).

[0047] Specific examples of suitable initial values ​​for the number of divisions are approximately 20 to 100 for the dimensions of the controlled variables and approximately 3 to 20 for the dimensions of the explanatory variables. In this way, for the dimensions of the controlled variables, the division width of the state definition is made finer to allow for more precise state transitions when control is executed, thereby improving control accuracy, while for the explanatory variables, since it is sufficient to know the trend rather than the quantity to be directly controlled, the amount of memory consumed by the learning model can be reduced by making the division width coarser than the controlled variables (i.e., making the number of divisions smaller), making it possible to initially set a mesh definition that strikes a good balance between the accuracy of the learning model and the amount of memory consumed.

[0048] Regarding the method for determining the initial value of the division number for the dimension of the controlled variable in this second method, it is also possible to perform the following systematic calculation. The following will be described separately for the dimension of the controlled variable and the dimension of the target variable.

[0049] First, regarding the dimension of the controlled quantity, 1) in the case of constant value control where it is required to control the controlled quantity so that it becomes a constant value, the initial value of the number of divisions should be set to n (n is a value equal to or greater than 1) times the upper and lower limit range of the state of the dimension of the controlled quantity divided by the required accuracy (a value converted into the number of divisions, which indicates to what extent the standard deviation of the controlled quantity from the target value is to be suppressed; in other words, to what extent the number of divisions is to be suppressed). By dividing the quantity finer than the required accuracy, it is possible to achieve sufficiently good accuracy when executing control.

[0050] 2) When variable target value control is required, in which the target value of the controlled quantity changes over time, as is often seen in batch processes, the initial value of the division number should be n times the value obtained by dividing the range between the upper and lower limit values ​​of the target value of the dimension of the controlled quantity by the required accuracy.

[0051] In these methods 1) and 2), the value of "n" is preferably approximately 1 to 5, and the larger the value of "n", the higher the resolution of control for the required accuracy can be.

[0052] Furthermore, with regard to the dimension of the explanatory variables, if the degree of influence of each explanatory variable on the controlled quantity is large, the initial value of the number of divisions should be set relatively large, and if the degree of influence is small, the initial value of the number of divisions should be set small. For example, if the degree of influence of each explanatory variable on the controlled quantity is quantified using a known method for calculating the contribution of the explanatory variable, it is possible to efficiently set the number of divisions according to the degree of influence on the actual data. As such a degree of contribution, for example, the standard partial regression coefficient obtained by performing multiple regression analysis using the explanatory variables with the controlled quantity as the response variable can be used.

[0053] Using this systematic calculation method, the initial value of the division number of the controlled variable dimension can be accurately and automatically determined according to the required control accuracy, and the initial value of the division number of each dimension is appropriately set according to its contribution to the controlled variable, thereby reducing the memory consumption of the learning model without adversely affecting the accuracy of the model. Furthermore, by taking into account the n-fold calculation, it is ensured that the state is defined with a resolution equal to or finer than the target accuracy, and the amplitude of fluctuations in the controlled variable around the target value during control execution is effectively suppressed, resulting in high control performance.

[0054] <Step S104> In step S104, one representative state during plant operation is selected, and the number of divisions is corrected for each dimension based on the learning data sufficiency rate. Here, the representative state is typically an average state that occurs frequently, or a state in which the controlled variable satisfies the control target. The processing flow in step S104 is shown in Figure 4. This flow diagram shows the flow (steps S401 to S405) for correcting the number of divisions for the kth dimension (hereinafter referred to as dimension k) that constitutes the state.

[0055] (Step S401) First, the learning data fulfillment rate is calculated for the initial value set in advance in step S102 (FIG. 3) for the number of divisions of dimension k. A specific method for calculating the learning data fulfillment rate will be described in detail below with reference to Equations 1, 2, 3, and 5.

[0056] The learning data sufficiency rate (hereinafter represented by the mathematical symbol Pprox(S(i))) is the sum of the transition probabilities to the state near the distribution center of the destination state for the pre-transition state S(i). Here, "near the distribution center of the destination state" refers to a range within a predetermined distance from the pre-transition state or the destination state with the maximum transition probability, within the space in the multidimensional space where transitions can occur from the pre-transition state.

[0057] The calculation of the learning data sufficiency rate Pprox(S(i)) strictly in accordance with this definition can be carried out according to Equation (1) below.

[0058]

Equation

[0059] In Equation (1), k is the dimension of the state, Ndim is the number of state dimensions, m_k (in this embodiment, the subscript is represented by an underscore as appropriate) is the number of the division section of the k-th dimension, Nprox_k is the number of divisions set along the axis in the vicinity of the distribution center of the k-th dimension (for example, if Nprox3 = 4, the range of the division width shifted up and down 4 units from the center of the distribution in the 3rd dimension is set as the range of the vicinity of the center), and P_trans(i, j_pmax) represents the transition probability to the transition destination j with the maximum transition probability among the transition destinations j that can transition from state i.

[0060] Also,.offset(m_1, m_2, m_3, ···, m_k, ···, m_ndim) on the right side of P_trans(i, j_pmax) represents the transition probability to the state where the 1st dimension is shifted by the number of divisions m_1, the 2nd dimension is shifted by the number of divisions m_2, ···, the k-th dimension is shifted by the number of divisions m_k, ···, and the Ndim-th dimension is shifted by the number of divisions m_ndim, centered on the state j_max with the maximum transition probability.

[0061] Also, the vertical bar notation _(d < Dprox) at the end of the equation represents conditional sum processing for determining whether the distance d indicating how far the state shifted by.offset(m_1, m_2, m_3, ···, m_k, ···, m_ndim) from the distribution center after the transition is within the predetermined range of Dprox, and only including the states within the predetermined specified distance Dprox in the sum.

[0062] Here, the distance d in Equation (1) can be calculated as the Euclidean distance in the multi-dimensional space as shown in Equation (2) below as the simplest definition

[0063]

number

[0064] However, the distance d is not limited to equation 2, and other known formulas for calculating distance in a multidimensional space, such as the Mahalanobis distance, may be used.

[0065] Furthermore, the predetermined specified distance Dprox for the distance d is preferably at least 1 and up to approximately 110. A smaller specified distance is suitable when the number of divisions is relatively small, and a larger specified distance is suitable when the number of divisions is relatively large.

[0066] On the other hand, the calculation of the training data sufficiency rate Pprox(S(i)) using the above-mentioned formula 1 requires an enormous amount of calculation because there are an enormous number of combinations of patterns for shifting each dimension from the distribution center position in the part written as .offset(m_1,m_2,m_3,...,m_k,...,m_ndim) in the formula, and it is necessary to determine whether the distance from the distribution center for each pattern is within a specified range. This results in the problem that it takes a long time to generate a training model.

[0067] As a calculation method to solve this, the following equation 3 can be used.

[0068]

number

[0069] This calculation method calculates the sum of transition probabilities changed within a predetermined range for each dimension. Since it does not involve shifting from the distribution center in all dimensional directions as in Equation 1, nor does it calculate and determine distances as in Equation 2, it is possible to speed up calculations and reduce the time required to generate a learning model. With this method, the values ​​of each axial component of the learning data sufficiency rate (described later) can be efficiently calculated using Equation 5 and shown as a breakdown for each axial direction.

[0070] Although the method using Equation 3 has been described here, Equation 4 below may be used instead.

[0071]

number

[0072] The difference between Equation 3 and Equation 4 is how the distribution center is defined: Equation 3 defines the distribution center as the distribution center after a transition, while Equation 4 defines the distribution center as the distribution center before a transition. Equation 3 faithfully reflects the distribution characteristics of the transition and is highly accurate, and is suitable for cases where the state transition probability matrix is ​​one in which the state transition rarely remains in the same state (i.e., the values ​​of the diagonal elements of the state transition probability matrix are not extremely large).

[0073] On the other hand, jpmax in the formula (the number of the transition destination state at which the transition probability is maximized) requires fixing i for P(i,j), changing j to all states, and comparing the transition probabilities to identify the j at which the transition probability is maximized, which results in a higher calculation load than Formula 4. Formula 4 is suitable for cases where the state transition probability matrix is ​​such that the transitions remain at the previous state and move little by little to neighboring states (in other words, when the values ​​of the diagonal components of the state transition probability matrix are large and the values ​​in the vicinity are smaller, and the center of the distribution tends to concentrate near the diagonal).

[0074] Furthermore, since it is not necessary to repeat the comparative calculation for identifying the transition destination with the maximum probability by changing j to obtain jpmax as in Equation 3, Equation 4 is simple and requires a small amount of calculation.

[0075] (Step S402) 4, next, in step S402, it is determined whether the learning data sufficiency rate calculated in step S401 is below the lower limit of a predetermined desirable range (hereinafter referred to as the specified range). A specific example of a suitable specified range is approximately 0.4 to 0.8.

[0076] If the training data sufficiency rate is too low, there is a risk that only a narrow range of transition probability distribution characteristics will be captured. Conversely, if the training data sufficiency rate is too high, states with low occurrence probabilities distributed over a wide space will be included, which will result in the inclusion of a lot of information that is unnecessary for the purpose of control, which is to efficiently transition states, and the amount of calculation required to find the optimal state transition destination may become unnecessarily large.

[0077] (Step S403) If the determination result in step S402 is true, the number of divisions is reduced by one unit in step S403, the learning data sufficiency rate for that number of divisions is calculated, and the calculated learning data sufficiency rate is determined again in step S402. This process is repeated until the determination in step S402 is false. One unit of the number of divisions is a value, such as 1 or 2, that represents how much the number is changed when increasing or decreasing the number of divisions to adjust it.

[0078] (Step S404) If the determination in step S402 is false, then in step S404 it is determined whether the current learning data sufficiency rate exceeds a specified range. If the determination result is true, then in step S405 the number of divisions is increased by one unit, the learning data sufficiency rate for that number of divisions is calculated, and the calculated learning data sufficiency rate is determined again in step S404. This process is repeated until the determination in step S404 is false.

[0079] If the determination in step S404 is false, the correction of the division number for dimension k is completed. The above process is executed for each dimension constituting the state, and the division numbers for all dimensions are determined.

[0080] In this example, first, in steps S402 and S403, it is determined whether the learning data sufficiency rate is below the specified range and the division number adjustment process is performed, and then, in steps S404 and S405, it is determined whether the learning data sufficiency rate exceeds the specified range and the division number adjustment process is performed, but the order of these steps may be reversed (i.e., steps S404 and S405 may be performed first, and then steps S402 and S403 may be performed).

[0081] As described above, according to this embodiment, the number of mesh divisions for each dimension is determined by giving priority to the learning data fulfillment rate, and thus it is possible to appropriately capture the center and its surroundings of the distribution of the most frequently occurring probabilistic state transition patterns of the target process, thereby efficiently generating a learning model that effectively simulates the probabilistic characteristics of the state transition patterns.

[0082] By using such a learning model for control, it becomes possible to connect and trace state transition patterns that are statistically likely to occur in order to bring the current state of the process closer to the control target, thereby enabling operation that efficiently approaches and maintains the control target. [Example]

[0083] FIG. 5 shows another example of the processing flow for mesh definition when one state before a transition is used as the starting point. In this flow, the upper and lower limits of each dimension are set in step S101, and then the initial value of the number of divisions for each dimension is set in step S102. This is the same as the first embodiment (FIG. 3), but differs in that step S103 is added instead of step S104 in FIG. 3. In step S103, the number of divisions for each dimension is corrected based on the number of vertices of the shape of the probability distribution along the axis of each dimension. The processing of step S103 will be explained below with reference to FIG. 6. This flow diagram shows the flow for correcting the number of divisions for the kth dimension (hereinafter, dimension k) that constitutes the i-th state (hereinafter, state i) in the state definition.

[0084] (Step S601) First, when step S103 is executed, the number of vertices in the distribution shape of the state transition probability from state i to the destination state j for the number of divisions at that time (i.e., the initial value of the number of divisions set in step S102) is calculated in step S601.

[0085] Here, the distribution shape of the state transition probability is the shape of the probability distribution obtained by plotting the transition probability from state i to state j, with the transition probability of the destination state j defined along the axis of dimension k (the horizontal axis is from the upper limit to the lower limit, and the vertical axis is the transition probability from state i to state j). Also, the number of vertices is the total number of vertices on the upper and lower sides when this distribution shape has multiple peaks.

[0086] For example, a unimodal bell-shaped curve has one vertex, a bimodal curve generally has three vertices, including the peaks and valleys, and an n-modal curve generally has (2n-1) vertices, including the peaks and valleys. (Note that below, the number of vertices in the distribution shape of such state transition probabilities may be referred to as the number of vertices in the probability distribution shape or the number of vertices in the probability distribution, as appropriate.)

[0087] (Step S602) Next, in step S602, it is determined whether the number of vertices of the probability distribution calculated in this way is below the lower limit of a predetermined desirable range (hereinafter referred to as a specified range). A suitable example of the specified range is approximately 1 to 5.

[0088] This is because the probability distribution of the process signal values ​​of a plant is often unimodal (i.e., has one peak) centered on the average value, and even in cases where it is multimodal, it often has around two peaks (i.e., three peaks above and below), and at most the number of peaks is generally within three (i.e., five peaks above and below).

[0089] However, for example, if the target value of the controlled variable of the process is set in a unidirectional manner, such that the higher the better, or the lower the better, the shape of the probability distribution after a transition in the value of dimension k for state i may be monotonically decreasing or monotonically increasing, and may not have a peak, and in such cases the lower limit of the specified range may be 0.

[0090] Furthermore, if there are operational constraints on the plant equipped with the process or if there are multiple different operation modes, the number of peaks in the probability distribution shape will increase depending on the number of constraints and operation modes, and therefore the preferred value of the specified range will be higher than in the above example.

[0091] (Step S603) If the determination result in step S602 is true, the number of divisions is increased by one unit, the number of vertices of the probability distribution at that number of divisions is calculated in step S603, and the calculated number of vertices of the probability distribution is determined again in step S602. This process is repeated until the determination in step S602 is false. Here, one unit of the number of divisions is a value, such as 1 or 2, that represents the number of divisions to be changed by when increasing or decreasing the number of divisions to adjust it.

[0092] (Step S604) If the determination in step S602 is false, it is determined in step S604 whether the number of vertices in the current probability distribution exceeds a specified range.

[0093] (Step S605) If the determination result in step S604 is true, the number of divisions is decreased by one unit, the number of vertices of the probability distribution at that number of divisions is calculated in step S605, and the calculated number of vertices of the probability distribution is determined again in step S604. This process is repeated until the determination in step S604 is false.

[0094] If the determination in step S604 is false, the correction of the division number for dimension k is completed. The above process is executed for each dimension constituting the state, and the division numbers for all dimensions are corrected.

[0095] In this example, it has been explained that first, in steps S602 and S603, it is determined whether the number of vertices in the probability distribution is below a specified range and the number of divisions is corrected, and then, in steps S604 and S605, it is determined whether the number of vertices in the probability distribution is above a specified range and the number of divisions is corrected, but the order of these steps may be reversed.

[0096] Furthermore, in order to visualize the calculation process in each of the above steps, the number of vertices of the transition probability distribution along the calculated division section of each dimension, or information on whether the number of vertices is below, within, or above a specified range, may be displayed on the display means 3003 on a display screen 1700 as shown in FIG. 17 .

[0097] In this way, by determining the number of mesh divisions for each dimension taking into account the number of vertices of the probability distribution shape, the number of divisions for each dimension can be corrected so that a learning model that appropriately captures the probabilistic state transition patterns of the target process can be generated.In addition, it is possible to prevent problems such as an excessive number of divisions consuming more memory than necessary, or the occurrence of a state where the state transition probability is undefined (transition probability is 0), making it impossible to calculate a pattern that transitions the process state to the control target during control, or, conversely, an insufficient number of divisions resulting in a coarse mesh that makes it impossible to learn the process state transition patterns with the precision required for high-precision control, resulting in coarse control operation and poor control performance, thereby building a learning model with sufficient precision for use in control.

[0098] Another aspect of step S103 is that it allows for adjustment of the trade-off between control accuracy and required memory size. Generally, when attempting to increase control accuracy, the mesh division of the learning model becomes finer, and the amount of memory required to hold the learning model becomes enormous. Conversely, when the amount of required memory is reduced, the mesh division of the learning model becomes coarser, and control accuracy decreases. This adjustment, taking into account the trade-off between required memory size and control performance, becomes possible by setting the number of vertices in step S103.

[0099] For example, if the number of vertices is set relatively large (for example, 5 or more, i.e., trimodal or more), the amount of memory required increases, but it becomes possible to build a highly accurate model that captures various patterns of process operation in detail. Conversely, if the number of vertices is set relatively small (for example, 3 or less, i.e., bimodal or unimodal), it becomes possible to ensure control performance using a learning model that roughly captures the distribution trend of occurrence frequency, while reducing the memory consumption required to hold the learning model.

[0100] As described above, according to this embodiment, the number of mesh divisions for each dimension is determined by prioritizing the number of vertices of the probability distribution shape, making it possible to efficiently execute the necessary and sufficient mesh definition to appropriately capture and simulate the probabilistic state transition pattern of the target process. [Example]

[0101] 7 shows another example of the processing flow for mesh definition when one state before a transition is used as the starting point. In this flow, first, upper and lower limits for each dimension are set in step S101, then an initial value for the number of divisions for each dimension is set in step S102, the number of divisions for each dimension is corrected in step S103 based on the number of vertices of the shape of the probability distribution along the axis of each dimension, and the number of divisions for each dimension is corrected in step S104 based on the training data sufficiency rate. Here, the processing contents of steps S101, S102, and S104 are the same as those in Example 1 (FIGS. 3 and 4), and the processing contents of step S103 are the same as those in Example 2 (FIGS. 5 and 6), so their explanations are omitted.

[0102] In this embodiment, the number of vertices of the probability distribution shape is taken into consideration in step S103, and then the training data sufficiency rate is taken into consideration in step S104 to determine the number of mesh divisions for each dimension, thereby obtaining the effects of both steps S103 and S104 described above.

[0103] Furthermore, by narrowing down the number of divisions required to represent the probability distribution characteristics of each dimension to a reasonable range in advance in step S103, the calculation for adjusting the number of divisions in the subsequent step S104, which involves a high calculation load, can be performed efficiently, thereby shortening the overall calculation time required for mesh definition.

[0104] Next, in the following examples, we will explain the mesh definition processing flow executed by the learning model generation means 3010 from a different perspective than the above-mentioned examples. These relate to a method of determining the upper or lower limit value of each dimension by referring to the distribution of transition probabilities to all states within the definition range, starting from one state before the transition. [Example]

[0105] 8 shows an example of the processing flow for mesh definition when a state before a transition is used as the starting point. In this flow, the upper and lower limits of each dimension are set in step S801, the number of divisions for each dimension is set in step S802, and the upper or lower limit of each dimension is corrected based on the training data sufficiency rate in step S806.

[0106] The processing content of step S801 is the same as step S101 in each of the above-mentioned embodiments, and step S802 is also the same as step S102 in each of the above-mentioned embodiments except that in this embodiment, a value that is determined from the beginning rather than an initial value is set, so an explanation will be omitted.

[0107] The processing content of step S806 will be explained below.

[0108] (Step S806) In step S806, one representative state during plant operation is selected, and the upper or lower limit value is corrected for each dimension based on the learning data fulfillment rate. Here, the representative state is, as in the above-mentioned embodiment, an average state that occurs frequently, or a state in which the controlled variable satisfies the control target.

[0109] FIG. 9 shows a processing flow for correcting the upper or lower limit value using the representative state selected in this way. This flow diagram shows a flow for correcting the upper or lower limit value for the kth dimension (hereinafter referred to as dimension k) that constitutes the state. Among these, the processing contents of steps S901, S902, and S904 are the same as steps S401, S402, and S404 in FIG. 4 of the first embodiment. However, a difference is that the processing steps corresponding to steps S403 and S405 in FIG. 4 are steps S903 and S905 in FIG. 9. These steps S903 and S905 will be explained below.

[0110] (Step S903) In step S903, it is determined whether the learning data sufficiency rate calculated in step S901 is below the lower limit of a predetermined desirable range (hereinafter referred to as the specified range) in step S902, and if the determination result is true (below), the upper and lower limit range is reduced by one unit. This includes either decreasing the upper limit by one unit, increasing the lower limit by one unit, or both. The learning data sufficiency rate in this state is then calculated, and the calculated learning data sufficiency rate is again determined in step S902.

[0111] This process is repeated until the determination result in S902 becomes false. Here, one unit when decreasing the upper limit value or increasing the lower limit value is a value that indicates how much the value is changed.

[0112] (Step S905) If the result of the determination in step S902 is false, it is determined in step S904 whether the current learning data sufficiency rate exceeds the specified range. If the result is true, the upper and lower limits are increased by one unit in step S905, and the learning data sufficiency rate at that time is calculated. It is again determined in step S904 whether the calculated learning data sufficiency rate exceeds the specified range, and this process is repeated until the result of the determination in step S904 becomes false.

[0113] In this example, it has been described that first, in steps S902 and S903, it is determined whether the learning data sufficiency rate is below the specified range, and the upper and lower limits are corrected, and then, in steps S904 and S905, it is determined whether the learning data sufficiency rate exceeds the specified range, and the upper and lower limits are corrected, but the order of these steps may be reversed. That is, steps S904 and S905 may be performed first, and then steps S902 and S903 may be performed.

[0114] According to this embodiment, even if the number of divisions for each dimension is set appropriately to a certain extent, when the upper and lower limit values ​​are set too broadly or too narrowly and the state transition probability matrix does not adequately capture the change characteristics of the process, it is possible to adjust the setting of the upper and lower limit values ​​by taking into account the sufficiency rate of the learning data and adequately capturing the change characteristics near the center of the process distribution. [Example]

[0115] 10 shows an example of the processing flow for mesh definition when one state before a transition is used as the starting point. In this flow, the upper and lower limits of each dimension are set in step S1001, then the number of divisions for each dimension is set in step S1002, and the upper or lower limit of each dimension is corrected in step S1005 based on the number of vertices of the probability distribution shape.

[0116] The execution contents of steps S1001 and S1002 are the same as steps S801 and S802 of the fourth embodiment in Fig. 8, and therefore a description thereof will be omitted. Step S1005 will be described below.

[0117] (Step S1005) In step 5, one representative state during plant operation is selected, and the upper or lower limit value is corrected for each dimension based on the learning data fulfillment rate. Here, the representative state refers to an average state that occurs frequently, or a state in which the controlled variable satisfies the control target, as in the above-described embodiment. Another example of a processing flow for correcting the upper or lower limit value using the representative state selected in this way is shown in Figure 11.

[0118] This flow diagram shows a flow for correcting the upper or lower limit value for the kth dimension (hereinafter referred to as dimension k) that constitutes a state. Among these, the processing contents of steps S1101, S1102, and S1104 are the same as steps S601, S602, and S604 in FIG. 6 of the second embodiment. However, the processing steps corresponding to steps S603 and S605 in FIG. 6 are steps S1103 and S1105 in FIG. 11, respectively. Steps S1103 and S1105 will be described below.

[0119] (Step S1103) In step S1101, the number of vertices of the probability distribution shape is calculated, and in step S1102, it is determined whether it is below the lower limit of a predetermined desirable range (hereinafter referred to as the specified range). If the result is true (below), in step S1103, the number of vertices of the probability distribution shape when the upper and lower limit range is reduced by one unit is calculated, and the calculated number of vertices is again determined in step S1102. This process is repeated until the result of the determination in step S1102 is false.

[0120] (Step S1105) If the result of the determination in step S1102 is false, step S11044 determines whether the number of vertices in the probability distribution shape at that time exceeds a specified range. If the result is true (exceeds), step S1105 calculates the number of vertices in the probability distribution shape when the upper and lower limit ranges are increased by one unit. The calculated number of vertices is then determined again in step S1104. This process is repeated until the result of the determination in step S11044 becomes false.

[0121] In this example, it has been described that first, in steps S1102 and S1103, it is determined whether the number of vertices of the probability distribution shape is below a specified range, and the upper and lower limit ranges are corrected, and then, in steps S1104 and S1105, it is determined whether the number of vertices of the probability distribution shape exceeds a specified range, and the upper and lower limit ranges are corrected, but the order of these steps may be reversed. In other words, steps S1104 and S1105 may be performed first, and then steps S1102 and S1103 may be performed.

[0122] According to this embodiment, even if the number of divisions for each dimension is set appropriately to a certain extent, when the upper and lower limit values ​​are set too broadly or too narrowly and the state transition probability matrix does not adequately capture the change characteristics of the process, by taking into account the number of vertices of the probability distribution, it is possible to appropriately adjust the setting of the upper and lower limit values ​​while suppressing the decrease in accuracy of the learning model caused by dividing the distribution waveform too finely or too coarsely. [Example]

[0123] 12 shows an example of a processing flow for mesh definition when a state before a transition is used as the starting point. In this flow, first, the upper and lower limits of each dimension are set in step S1201, then the number of divisions for each dimension is set in step S1202, the upper or lower limit of each dimension is corrected based on the number of vertices of the probability distribution shape in step S1205, and then the upper or lower limit of each dimension is corrected based on the training data fulfillment rate in step S1206.

[0124] The processing contents of steps S1201 and S1202 are the same as steps S801 or S1001, and S802 or S1002 in Examples 4 and 5, the processing contents of step S1205 are the same as step S1005 (Figure 10) in Example 5, and the processing contents of step S1206 are the same as step S806 (Figure 8) in Example 4.

[0125] According to this embodiment, when the number of divisions in each dimension is set appropriately to a certain extent, but the upper and lower limit values ​​are set too broadly or too narrowly, and the state transition probability matrix does not adequately capture the change characteristics of the process, the upper and lower limit values ​​are first adjusted by taking into account the number of vertices in the probability distribution to prevent a decrease in the accuracy of the learning model caused by dividing the distribution waveform too finely or too coarsely, and then the upper and lower limit values ​​are adjusted by taking into account the learning data fulfillment rate to capture the change characteristics near the center of the process distribution.This procedure allows the range of the probability distribution to be simulated to be adjusted while ensuring the accuracy of the model, and enables efficient execution of highly accurate and flexible mesh definition according to the characteristics of the process to be controlled. [Example]

[0126] In the above-described embodiments, one pre-transition state is determined, and the distribution of transition probabilities to various post-transition states is used as a starting point to determine the number of divisions for each dimension (embodiments 1 to 3) or determine upper and lower limits (embodiments 4 to 6). These examples are effective when the controlled process 1000 is controlled to always follow a constant target value, this target value does not change, fluctuations in the process state are relatively small, and extreme state transitions do not occur.

[0127] However, even if the control target value of the target process 1000 is a constant value, if the set value changes from time to time, or if the control target value is set to change over time (for example, increasing at a constant rate or decreasing at a constant rate), or if the operating state of the process changes frequently and various process states can be taken other than those near the control target value, the method of the above-mentioned embodiment, in which a mesh is defined using only one point as a reference point before a state transition, may not necessarily facilitate high-precision control.

[0128] In such a case, it is desirable to define a mesh by referring to a plurality of representative states that the target system 1000 can take as pre-transition states. However, if mesh definitions are simply repeated from a plurality of pre-transition states as in the above-described embodiment, there will be as many mesh definitions as there are pre-transition states referenced, and not only will it be difficult to determine which mesh definition is the most appropriate, but it will also be difficult to determine whether a selected mesh definition is valid for other pre-transition states.

[0129] In this embodiment, in order to address such issues, a mesh definition flow that can refer to multiple pre-transition states and build a learning model that is highly compatible with a wide range of pre-transition states will be described below.

[0130] In the mesh definition flow of this embodiment, step S104 in FIG. 3 of the first embodiment, in which the number of divisions for each dimension is corrected based on the training data fulfillment rate, is executed in the flow of FIG.

[0131] 13, first, in step S1400, at least one predetermined state S(i) (where i = 1, 2, 3, ...) is assumed as a state before the transition, and for each of these states, the division number Ndiv(S(i), k) of each dimension k is calculated. The division numbers for each dimension generated in this way are hereinafter referred to as division number candidates.

[0132] Then, in the following step S1499, the number of divisions for each dimension is finally determined by referring to Ndiv(S(i), k), which is the candidate number of divisions for each dimension for at least one pre-transition state calculated in this way. Here, the at least one or more predetermined states S(i) do not necessarily mean a specific state or all states before the transition, but rather it is sufficient if they include a plurality of representative states among the various states that the process can take. The specific procedures of steps S1400 and S1499 are shown below.

[0133] <Step S1400> A specific example of the processing procedure of step S1400 (FIG. 13) will be described using the processing flow of FIG.

[0134] (Step S1401) First, the state S(i) is set as the state before the transition, and the learning data fulfillment rate Pprox(S(i)) corresponding to this state before the transition is calculated. The method for calculating the learning data fulfillment rate Pprox(S(i)) is the same as that described in step S401 in FIG. 4 of the first embodiment, and is calculated using Equations 1 to 4.

[0135] (Step S1402) Next, it is determined whether the calculated learning data fulfillment rate is below a predetermined range (specified range).

[0136] If the result of the determination in step S1402 is true (below), the process proceeds to a repetitive process consisting of steps S1402a to S1402d described below, where a new dimension k is selected and the division number of that dimension k is corrected, and this process is repeated until the result of the determination in step S1402d is resolved (until the result is no longer below). The repetitive process of steps S1402a to S1402d is as follows.

[0137] First, in step S1402a, the component Pax(S(i), k) along the axis direction of each dimension k for the learning data fulfillment rate Pprox(S(i)) for the state S(i) before the transition is calculated. This component Pax(S(i), k) along each axis direction is calculated using Equation 5. (Hereinafter, this Pax(S(i), k) will be referred to as the k-th dimension component or the k-th axis direction component of the learning data fulfillment rate, as appropriate.)

[0138]

number

[0139] Next, in step S1402b, the dimension k for which the calculated axial component Pax(S(i), k) is the smallest is selected as the target for correcting the number of divisions, and in the subsequent step S1402c, the number of divisions Ndiv(S(i), k) of that dimension is reduced by one unit (here, one unit of the number of divisions is a value that represents the number of divisions to be changed by, for example, 1 or 2, when adjusting the number of divisions by increasing or decreasing it, as in step S403 in Figure 4 of the above-mentioned first embodiment).

[0140] Then, in step S1402d, the learning data sufficiency rate Pprox(S(i)) is calculated based on the reduced number of divisions, and it is determined whether the value is within a predetermined range (the predetermined range is preferably approximately 0.4 to 0.8, as shown in step S402 in FIG. 4 of the first embodiment).

[0141] If the determination result in step S1402d is false, the processes of steps S1402a, S1402b, and S1402c are executed again, and such processes are repeated until the determination result in step S1402d becomes true. If the determination result in step S1402d becomes true, the correction of the division number using the state S(i) before the transition ends, and the process proceeds to the next step S1404.

[0142] Returning to step S1402, if the determination result here is false, the process proceeds to step S1403, where it is determined whether the learning data sufficiency rate at this time exceeds a predetermined range (specified range), and if the determination result is true, the process proceeds to an iterative process consisting of steps S1403a, S1403b, S1403c, and S1403d described in detail below, where a correction process is repeatedly executed to change the number of divisions while changing the target dimension k until the determination result is resolved in S1403d (until the result is no longer exceeded). The iterative process consisting of steps S1403a, S1403b, S1403c, and S1403d is as follows.

[0143] First, in step S1403a, the component Pax(S(i), k) along the axis direction of each dimension k for the learning data fulfillment rate Pprox(S(i)) for the state S(i) before the transition is calculated. This component Pax(S(i), k) along each axis direction is calculated using the above-mentioned (Equation 5).

[0144] Next, in step S1403b, the dimension k for which the calculated axial component Pax(S(i), k) is the largest is selected as the target for correcting the number of divisions, and in the following step S1403c, the number of divisions Ndiv(S(i), k) for that dimension is increased by one unit (here, one unit is the same as in step S1402c).

[0145] In the next step S1403d, the learning data sufficiency rate Pprox(S(i)) is calculated based on the increased number of divisions, and it is determined whether the value no longer exceeds the specified range. If the determination result is false, the processing of steps S1403a, S1403b, and S1403c is executed again.

[0146] This repetitive process is executed until the determination result in step S1403d becomes true, and when the determination result becomes true, the correction of the division number using the pre-transition state S(i) ends, and the process proceeds to the next step S1404.

[0147] Also, if the determination result in step S1403 is false, the process proceeds to the next step S1404.

[0148] In the subsequent step S1404, if the judgment result of the aforementioned step S1402 is true, the number of divisions for each dimension finally calculated in step S1402c is assigned, or if the judgment result of the aforementioned step S1403 is true, the number of divisions for each dimension finally calculated in step S1403c is assigned, or if the judgment results of both the aforementioned steps S1402 and S1403 are false, the number of divisions for each dimension used in step S1401 is assigned to the state S(i) before the transition, and registered as one candidate for the number of divisions for each dimension. In the next step S1405, it is determined whether the process of registering one candidate number of divisions for each dimension corresponding to the pre-transition state S(i) has been completed for all of the predetermined states S(i) (i=1, 2, 3, ...). If the determination result is false, the same process as described above is executed for the next state S(i). If the determination result is true, the creation of candidate number of divisions is completed. This completes the process of step S1400 in Figure 13.

[0149] <Step S1499> In step S1499 (FIG. 13), the number of divisions Ndiv(k) for each dimension is finally determined from within the range of maximum and minimum values ​​of the candidate number of divisions Ndiv(S(i), k) for each dimension calculated using multiple pre-transition states S(i) (i = 1, 2, 3, ...) as described above. Two methods for doing this are described below. These two methods make it possible to efficiently and accurately calculate, from among candidates that have multiple representative pre-transition states, which number of divisions is best overall from the calculation results of the number of divisions that change depending on the pre-transition state.

[0150] <Method 1> The first method of executing step S1499 is to calculate the average value of the division numbers Ndiv(S(i), k) calculated and registered for multiple pre-transition states S(i) (i=1, 2, 3, ...) for each dimension k, round the value to an integer using any known method such as rounding off, rounding up, or rounding down, and set this value as the division number Ndiv(k) for dimension k. The average value here is preferably the mode of the distribution, but is not limited to this and may be any average value according to a known definition such as the median of the distribution or the arithmetic mean.

[0151] The reason why the mode is preferable is that using a division number that has a high statistical frequency results in a higher rate of matching with all of the multiple pre-transition states S(i) (i = 1, 2, 3, ...) used to select the division number than using a division number that does not necessarily have the highest frequency of distribution, such as the arithmetic mean or median.

[0152] <Second Method> A second method for executing step S1499 will be described with reference to the table shown in Fig. 15. As described above, in step S1400 prior to step S1499, when processing is performed according to the flow shown in Fig. 14, one candidate number of divisions along each dimension k is generated for each of the states i before a plurality of state transitions. This candidate changes depending on the initial value of the number of divisions that is set when the processing of step S1401 in Fig. 14 is performed.

[0153] Therefore, if the process of Figure 14 is executed multiple times with different initial values ​​for the number of divisions, at least one candidate number of divisions along each dimension k (in practice, multiple candidates in most cases) will be generated for each of the multiple state i before the state transition.

[0154] Of the multiple candidates for the number of divisions (number of divisions for each dimension k for state i) generated in this way, the maximum and minimum numbers of divisions are listed in table 1500. The overlapping portion of the minimum and maximum ranges in the vertical direction 1501 shown in table 1500 is narrowed down, and the number of divisions for each dimension is selected from within that maximum and minimum range. Table 1500 is generated by the learning model generation unit 3010 and stored in a memory unit (not shown), etc., and the selection process 1501 and selection result 1502 are displayed on the display means 3003 so that the user can view it. It is also possible to display only the final selection result 1502.

[0155] Furthermore, in order to visualize the calculation process and results of each step described above, a table 1600 shown in FIG. 16 may be displayed on the display means 3003, and the portion of the transition destination state within a predetermined distribution center neighborhood distance corresponding to the transition source state may be highlighted, or the learning data fulfillment rate corresponding to each transition source state or information on whether the learning data fulfillment rate is below, within, or above a specified range may be displayed.

[0156] Here, it is not necessary to list all pre-transition states. For example, only representative states within the core range of process operation may be extracted and displayed. Furthermore, the "State Value (by Dimension)" may be displayed as a representative value, such as the lower limit, upper limit, or average value of each mesh. While the "Judgment" for the "Learning Data Sufficiency Rate" is represented by "O" and "X," it may also be displayed by coloring the row in the table, changing the color of the text in the table, or making it bold. Furthermore, the max, avg, and min for the "Sufficiency Rate" in the "Decision Result" display the maximum, average, and minimum values ​​in the "Sufficiency Rate" list in the top Ns rows. However, for simplicity, only the average value may be displayed. Furthermore, the intermediate progress may be omitted, and only the "Decision Result" in the bottom row may be displayed. These variations in display format are also applicable to Figure 17.

[0157] It should be noted that the above-described embodiments may be combined as appropriate within the scope of the present invention, or some of the embodiments may be replaced with other embodiments by selecting appropriate items, etc. For example, the screen display examples shown in Figures 15 to 17 are similarly applicable to at least the items that can be displayed in the other embodiments. [Explanation of symbols]

[0158] 1000: Controlled process 1001: Sensor 1002: Operation end 2000: Control device 3000: Learning device 3001: Input terminal 3002: Output terminal 3003:Display means 3010: Learning model generation unit 3020: Learning Model

Claims

1. A learning model generation device that generates a state transition probability matrix of a multidimensional state defined by combining a plurality of signals, setting initial values ​​of a lower limit value, an upper limit value, and a division number for each dimension of the multidimensional state; calculating a learning data sufficiency rate, which is the total value of the transition probability to a state near the distribution center of the destination state, which is defined as being within a predetermined distance from the destination state or the destination state with the maximum transition probability, for one or more pre-transition states; a calculation processing unit that corrects the number of divisions so as to decrease the number of divisions when the calculated learning data fulfillment rate is below a specified range, and to increase the number of divisions when the calculated learning data fulfillment rate exceeds the specified range; A learning model generation device.

2. The learning model generation device according to claim 1, The arithmetic processing unit A learning model generation device characterized in that the initial values ​​of the lower and upper limits of each dimension are set within the range of the statistical distribution range of the signal values ​​that has a high occurrence frequency.

3. The learning model generation device according to claim 2, The arithmetic processing unit A learning model generation device characterized in that, among the signals that make up each dimension, the initial value of the number of divisions is set large for signals whose values ​​are to be matched to a control target, and the initial value of the number of divisions is set small for signals that are factors that cause fluctuations in the controlled quantity.

4. The learning model generation device according to claim 2, The arithmetic processing unit A learning model generation device characterized in that, when the controlled quantity is controlled to a constant value, the initial value of the number of divisions is set to n times the value obtained by dividing the range between the upper and lower limit values ​​of the state of the dimension of the controlled quantity by the required accuracy, and when the controlled quantity is controlled to a variable target value, the initial value of the number of divisions is set to n times (n is a value greater than or equal to 1).

5. A learning model generation device that generates a state transition probability matrix of a multidimensional state defined by combining a plurality of signals, setting initial values ​​of a lower limit value, an upper limit value, and a division number for each dimension of the multidimensional state; a calculation processing unit that corrects the number of divisions so that the number of vertices of the transition probability distribution along the division sections of each dimension is increased when the number of vertices falls below a specified range, and the number of divisions is decreased when the number of vertices exceeds a specified range; A learning model generation device.

6. The learning model generation device according to claim 5, The arithmetic processing unit calculating a learning data sufficiency rate, which is the total value of the transition probability to a state near the distribution center of the destination state, which is defined as being within a predetermined distance from the destination state or the destination state with the maximum transition probability, for one or more pre-transition states; A learning model generation device characterized in that the number of divisions is reduced when the calculated learning data fulfillment rate is below a specified range, and the number of divisions is increased when the calculated learning data fulfillment rate exceeds the specified range.

7. A learning model generation device that generates a state transition probability matrix of a multidimensional state defined by combining a plurality of signals, setting initial values ​​of a lower limit value, an upper limit value, and a division number for each dimension of the multidimensional state; calculating a learning data sufficiency rate, which is the total value of the transition probability to a state near the distribution center of the destination state, which is defined as being within a predetermined distance from the destination state or the destination state with the maximum transition probability, for one or more pre-transition states; a calculation processing unit that corrects the upper and lower limit range by decreasing the upper and lower limit range, which is the range between the upper limit and the lower limit, when the calculated learning data fulfillment rate is below a specified range, and by increasing the upper and lower limit range when the calculated learning data fulfillment rate exceeds the specified range; A learning model generation device.

8. A learning model generation device that generates a state transition probability matrix of a multidimensional state defined by combining a plurality of signals, setting initial values ​​of a lower limit value, an upper limit value, and a division number for each dimension of the multidimensional state; a calculation processing unit that corrects the number of vertices of the transition probability distribution along the division section of each dimension so as to increase an upper / lower limit width, which is the width between the upper limit value and the lower limit value, when the number of vertices of the transition probability distribution along the division section of each dimension falls below a specified range, and to decrease the upper / lower limit width when the number of vertices exceeds the specified range; A learning model generation device.

9. The learning model generation device according to claim 5, The arithmetic processing unit calculating a learning data sufficiency rate, which is the total value of the transition probability to a state near the distribution center of the destination state, which is defined as being within a predetermined distance from the destination state or the destination state with the maximum transition probability, for one or more pre-transition states; When the calculated learning data fulfillment rate is below a specified range, the upper and lower limit range, which is the range between the upper limit value and the lower limit value, is reduced, and when the calculated learning data fulfillment rate is above the specified range, the upper and lower limit range is increased. A learning model generation device.

10. A learning model generation device that generates a state transition probability matrix of a multidimensional state defined by combining a plurality of signals, setting initial values ​​of a lower limit value, an upper limit value, and a division number for each dimension of the multidimensional state; For a plurality of candidate division numbers generated by correcting the number of divisions for a plurality of pre-transition states, the most frequent value of the plurality of candidate division numbers for each dimension is converted into an integer, or the number of divisions to be finally adopted is determined by using the overlapping portion of the ranges of the minimum and maximum number of divisions for a plurality of pre-transition states. A learning model generation device.

11. The learning model generation device according to any one of claims 1 to 4, a display means for displaying the state transition probability model, and displaying on the display means either a portion of the transition destination state within a predetermined distribution center neighborhood distance corresponding to the transition source state, or information on the learning data fulfillment rate corresponding to each transition source state or whether the learning data fulfillment rate is below a specified range, within a specified range, or above a specified range; A learning model generation device.

12. The learning model generation device according to claim 5, a display means for displaying the state transition probability model, and displaying on the display means the number of vertices of the transition probability distribution along the division section of each dimension, or information on whether the number of vertices is below, within, or above a specified range; A learning model generation device.

Images