Information processing device, section selection method, and section selection program

Abstract

To achieve the automatic extraction of a section representing dynamic characteristics suitable for system identification.SOLUTION: An information processing device comprises: a collection unit which collects input / output time-series data related to a system to be controlled; an estimation unit which estimates a probability distribution on model parameters generated from segment data corresponding to each section into which the time-series data is divided; a selection unit which selects a section with small statistical variation in the probability distribution; and an output unit which outputs information related to the section selected by the selection unit.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to an information processing apparatus, an interval selection method, and an interval selection program.

Background Art

[0002] As one of control methods, model predictive control is known. For example, in model predictive control, a model that represents the behavior of a system to be controlled using a transfer function, a state space representation, or a step response is used.

[0003] As one of the methods for generating such a model, there is a method called system identification that performs modeling by a statistical estimation method from time-series data of input and output related to a system to be controlled. For example, in Patent Document 1 and Patent Document 2, techniques of a type that perform system identification by applying a test signal such as an M-sequence to an input to a system to be controlled have been proposed.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Patent Document 2

Non-Patent Documents

[0005]

Non-Patent Document 1

Non-Patent Document 2

Non-Patent Document 3

Summary of the Invention

Problems to be Solved by the Invention

[0006] However, when performing system identification using time-series data collected from the system to be controlled, the task of selecting an interval suitable for system identification is sometimes entrusted to experts such as engineers.

[0007] In one aspect, the present invention aims to provide an information processing apparatus, an interval selection method, and an interval selection program capable of automatically extracting an interval suitable for identifying a system having dynamic characteristics.

Means for Solving the Problems

[0008] An information processing apparatus according to one aspect includes a collection unit that collects time-series data of input / output related to a system to be controlled, an estimation unit that estimates a probability distribution regarding parameters of a model generated from segment data corresponding to an interval for each interval into which the time-series data is divided, a selection unit that selects an interval in which the statistical variation of the probability distribution is small, and an output unit that outputs information regarding the interval selected by the selection unit.

Effects of the Invention

[0009] As one aspect, it is possible to realize automatic extraction of an interval suitable for identifying a system having dynamic characteristics.

Brief Description of the Drawings

[0010]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

[0011] Hereinafter, embodiments for implementing an information processing apparatus, an interval selection method, and an interval selection program according to the present disclosure (hereinafter referred to as "embodiments") will be described with reference to the accompanying drawings. Note that this embodiment merely shows one example or aspect, and the structure, operation, function, property, characteristic, method, use, etc. according to the present disclosure are not limited by such an example.

[0012] <Embodiment 1> <An example of the use case of the information processing apparatus> FIG. 1 is a block diagram showing an example of the functional configuration of the information processing apparatus. The information processing apparatus 10 shown in FIG. 1 provides a system identification function for generating a model for predicting the behavior of a system to be controlled.

[0013] As one aspect, such a system identification function may be applied to model predictive control such as process control in a plant, air conditioning control in a factory or building, so-called MPC (Model Predictive Control).

[0014] For example, the information processing apparatus 10 may be realized in a form integrated into a control device (control equipment) that collects temperature, pressure, flow rate, etc. from instrumentation equipment arranged in a plant and determines multivariable control amounts such as temperature control, pressure control, and flow control.

[0015] In addition, the information processing apparatus 10 may be realized as a computer that collects time-series data of input and output from the above-described controller and provides a model generated by the above-described system identification function. In this case, the information processing apparatus 10 may be realized as a Web server that provides the above-described system identification function on-premises. In addition, the information processing apparatus 10 can also provide the above-described system identification function as a cloud service by being realized as a SaaS (Software as a Service) type application.

[0016] <One aspect of the problem> In the above-described system identification, which section of the time-series data of the input and output regarding the system to be controlled is used has a great influence on the reliability of the model. In particular, when time-series data collected during normal operation or normal driving of a system to be controlled, such as a device or equipment, is used for system identification, system identification becomes more difficult.

[0017] For example, during normal operation or normal driving of a system, control is performed so that the input and output vary as little as possible with the steady state as the target. In this case, cases where information suitable for system identification is not included in the time-series data are likely to occur over a long period of time, so the reliability of the model decreases. Thus, a situation may occur where system identification becomes difficult as the stability of the input and output of the system increases.

[0018] Therefore, by applying the techniques described in Patent Document 1 and Patent Document 2 above, even when the input and output of the system to be controlled are constantly controlled during normal operation, it is conceivable to collect time-series data that enables system identification by intentionally varying the input.

[0019] However, since the technologies described in Patent Document 1 and Patent Document 2 above only perturb the input to the system, they can only be applied to the scenario of collecting time-series data of input and output. Therefore, the technologies described in Patent Document 1 and Patent Document 2 above have aspects that are not applicable to system identification implemented in a scenario where time-series data of input and output has already been collected.

[0020] From such an aspect, when performing system identification from time-series data collected during normal operation, the task of selecting an interval suitable for system identification is entrusted to experts such as engineers. For example, engineers are required to manually select an interval suitable for system identification while looking at the plot of time-series data. Such manual work not only increases labor as the time length of the time-series data increases, but is also difficult to perform by anyone other than those with specialized knowledge such as engineers.

[0021] <One aspect of the problem-solving approach> Therefore, in the present embodiment, as part of the above-described system identification function, a section selection function is provided that realizes automatic extraction of an interval suitable for identification of a system having dynamic characteristics.

[0022] Such a section selection function can be realized only with the following points in mind. That is, a new point of view that the statistical variation of the parameters of a model that predicts the behavior of the system to be controlled changes depending on whether or not it is an interval suitable for identification of a system having dynamic characteristics is utilized as a problem-solving approach. Here, an interval suitable for identification of a system having dynamic characteristics means an interval in which a causal relationship including dynamic characteristics is recognized between the time-series data of the input to the system and the time-series data of the output from the system, and it is expected that a highly reliable model can be obtained by identifying the system using the input-output time-series data of the interval.

[0023] Since both the model obtained through system identification and the parameters of the model are estimated values, uncertainty remains. This uncertainty is caused, for example, by the fact that the time-series data used in system identification contains disturbances and noise, and the device or equipment itself that is the target of system identification can vary. Therefore, the parameters of the model obtained through system identification can be regarded as variables following a certain probability distribution.

[0024] Such a probability distribution changes depending on which section of the time-series data is used, even if the device or equipment targeted for system identification is the same. And the section suitable for system identification is the one with a smaller statistical variation of the probability distribution than the section with a large statistical variation.

[0025] Normally, in system identification, the purpose is to obtain one estimated value of the model and its parameters, so the probability distribution is not focused on. However, by using the sequential Monte Carlo method or the like, the probability distribution of the parameters can be estimated.

[0026] Therefore, in the above interval selection function, the time-series data is divided into a plurality of intervals, and while selecting the intervals with a small statistical variation of the parameter distribution of the model and excluding the intervals with a large statistical variation, an interval suitable for identifying a system with dynamic characteristics is selected.

[0027] Therefore, according to the interval selection function according to this embodiment, it becomes possible to realize automatic extraction of an interval suitable for identifying a system with dynamic characteristics. As a result, on the one hand, the labor of selecting and extracting an interval suitable for system identification from the collected time-series data, for example, the time and effort of experts such as engineers, is reduced.

[0028] <Configuration of the information processing device> Next, a functional configuration example of the information processing device 10 according to this embodiment will be described. In FIG. 1, the blocks related to the system identification function of the information processing device 10 are schematically shown.

[0029] Here, the above-described section selection function may be provided in a packaged form as part of the above-described system identification function. Not limited to this, the above-described section selection function may be provided independently of the above-described system identification function. In this case, the hardware resources for providing each of the above-described system identification function and the above-described section selection function, such as physical devices and physical elements, may be different.

[0030] As shown in FIG. 1, the information processing apparatus 10 includes a data collection unit 11, a section division unit 12, a distribution estimation unit 13, a distribution evaluation unit 14, a section selection unit 15, an output unit 16, and an identification unit 17. Note that FIG. 1 only shows an excerpt of the functional units related to the above-described system identification function, and does not prevent the information processing apparatus 10 from including functional units other than those shown.

[0031] The data collection unit 11 is a processing unit that collects time-series data of input / output related to the system to be controlled. For example, since some control device (not shown), such as the above-described control device, is connected to the device or facility to be controlled, input / output data of the device or facility can be acquired from this control device. By collecting this input / output data in real time or in batches and obtaining a database or file in which logs of the input / output data are accumulated, time-series data of input / output can be collected. Also, the data collection unit 11 may directly collect input / output data from the device or facility to be controlled and store it in a storage unit (not shown), and use it as time-series data. Note that data preprocessing such as trend removal, filtering, outlier removal, and missing value interpolation may be applied to the time-series data of input / output.

[0032] The section division unit 12 is a processing unit that divides the time-series data of input / output into a plurality of sections. Hereinafter, the partial data corresponding to each section in the entire time-series data may sometimes be referred to as "segment data".

[0033] FIG. 2 is a schematic diagram showing an example of section division. As shown in FIG. 2, the section division unit 12 can divide the entire input / output time-series data into a total of N sections, namely section 1, section 2, section 3, ···, section N. For example, the section division unit 12 can set the time lengths of the respective sections at equal intervals. Further, the section division unit 12 can overlap adjacent sections before and after at an arbitrary ratio, for example, 50%. For example, FIG. 2 shows an example in which the adjacent sections before and after are overlapped at a ratio of 50%. In this case, the time at the central position of section 1 coincides with the time at the start position of section 2, the time at the central position of section 2 coincides with the time at the start position of section 3, and the time at the central position of section N-1 coincides with the time at the start position of section N.

[0034] Here, the time length of the section can be set according to the following criteria. The section selection function according to the present embodiment is assumed to be applied to the system identification of the dynamic model. Therefore, a section length that can at least estimate the dynamic characteristics of the system identification target may be set, and it is possible to suppress the setting of an excessively short time length. For example, when the settling time of the step response of the system identification target is known, the section length can be set to be about the same as the settling time. On the other hand, when the section length is excessively long, the possibility that the divided sections include a period unsuitable for system identification increases, and there is an aspect that it becomes difficult to extract only the sections suitable for system identification. Therefore, the section length can be set to be an integer multiple of the above-mentioned settling time, for example, about 10 times as the upper limit, and can be set below the upper limit.

[0035] Note that the purpose of the above section selection function is not system identification itself but the selection of sections suitable for system identification. Therefore, the settling time referred to when setting the section length may be of low accuracy. Further, when the settling time is unknown, the section length may be explored by changing the section length by doubling it, such as 1 minute, 2 minutes, 4 minutes, ···, and trying section selection.

[0036] Also, in FIG. 2, an example of overlapping adjacent sections before and after is given, but the sections before and after do not necessarily have to overlap. For example, the processing amount and processing time can be reduced as the overlapping ratio between the front and rear sections is decreased. Further, in FIG. 2, an example in which the lengths of each section are set at equal intervals is given, but it is not limited to this, and different section lengths may be set for each section.

[0037] The distribution estimation unit 13 is a processing unit that estimates the probability distribution regarding the parameters of the model generated from the segment data of the relevant section for each section.

[0038] As one aspect, since the models generated by the distribution estimation unit 13 and the identification unit 17 described later have different purposes of use, they may be different models. Therefore, from the aspect of distinguishing the labels of these two models, the model used by the distribution estimation unit 13 for distribution estimation for section selection may be described as the "section selection model". On the other hand, the model used by the identification unit 17 described later for system identification may be described as the "identification model". Further, when it is not necessary to distinguish the individuals of the section selection model and the identification model, or when it is obvious which of the two models is being referred to, it may simply be described as the "model".

[0039] For the structure of such a section selection model, an arbitrary mathematical model such as a state space representation or a transfer function type can be selected. For example, the transfer function type may include an ARX (Auto-Regressive with eXogenous) model, an ARMAX (Auto-Regressive Moving Average) model, an OE (Output Error) model, a BJ (Box-Jenkins) model, etc. In addition, it is also possible to use a non-parametric model such as an FIR (Finite Impulse Response) model or a step response, but a parametric model that models with a small number of parameters is more effective in terms of reducing the processing time.

[0040] For example, as one method of directly performing Bayesian estimation of the probability distribution of the parameters of a dynamic model, SMC (Sequential Monte Carlo) 2 using the sequential Monte Carlo method can be mentioned. This SMC 2 is disclosed in various documents such as Non-Patent Document 2. By using a state-space representation as the model structure of the interval selection model and applying SMC 2 to the segment data corresponding to each interval, it is possible to estimate the probability distribution of the parameters in each interval, or more precisely, the posterior distribution.

[0041] In SMC, a group of candidate values of the state variable x is prepared from the prior distribution p(x0), and each candidate value is used as an initial value, and the predicted value group of the output variable y obtained by solving the state equation is compared with the observed value of y to correct the group of candidate values. By repeating these prediction, observation, and correction, the distribution of the state variable x is approximated by a set of realized values. If such a state-space representation in which the state variable x has the parameter θ to be estimated as a value can be formulated using the input-output series, the posterior distribution p(θ|y) can be sequentially estimated using SMC for the estimation of the parameter θ. In SMC 2, for each particle of the parameter θ, the particle filter of the state-space representation is executed, and by repeating the prediction, observation, and correction of the output variable y, the posterior distribution p(θ|y) of the parameter θ can be sequentially estimated. t , t , t , t , t , t and the observed value of y to correct the group of candidate values. By repeating these prediction, observation, and correction, the distribution of the state variable x is approximated by a set of realized values. If such a state-space representation in which the state variable x has the parameter θ to be estimated as a value can be formulated using the input-output series, the posterior distribution p(θ|y) can be sequentially estimated using SMC for the estimation of the parameter θ. In SMC 2, for each particle of the parameter θ, the particle filter of the state-space representation is executed, and by repeating the prediction, observation, and correction of the output variable y, the posterior distribution p(θ|y) of the parameter θ can be sequentially estimated. t and the observed value of y to correct the group of candidate values. By repeating these prediction, observation, and correction, the distribution of the state variable x is approximated by a set of realized values. If such a state-space representation in which the state variable x has the parameter θ to be estimated as a value can be formulated using the input-output series, the posterior distribution p(θ|y) can be sequentially estimated using SMC for the estimation of the parameter θ. In SMC 2, for each particle of the parameter θ, the particle filter of the state-space representation is executed, and by repeating the prediction, observation, and correction of the output variable y, the posterior distribution p(θ|y) of the parameter θ can be sequentially estimated. t and the observed value of y to correct the group of candidate values. By repeating these prediction, observation, and correction, the distribution of the state variable x is approximated by a set of realized values. If such a state-space representation in which the state variable x has the parameter θ to be estimated as a value can be formulated using the input-output series, the posterior distribution p(θ|y) can be sequentially estimated using SMC for the estimation of the parameter θ. In SMC 2, for each particle of the parameter θ, the particle filter of the state-space representation is executed, and by repeating the prediction, observation, and correction of the output variable y, the posterior distribution p(θ|y) of the parameter θ can be sequentially estimated. t and the observed value of y to correct the group of candidate values. By repeating these prediction, observation, and correction, the distribution of the state variable x is approximated by a set of realized values. If such a state-space representation in which the state variable x has the parameter θ to be estimated as a value can be formulated using the input-output series, the posterior distribution p(θ|y) can be sequentially estimated using SMC for the estimation of the parameter θ. In SMC 2, for each particle of the parameter θ, the particle filter of the state-space representation is executed, and by repeating the prediction, observation, and correction of the output variable y, the posterior distribution p(θ|y) of the parameter θ can be sequentially estimated. t has the parameter θ to be estimated as a value can be formulated using the input-output series, the posterior distribution p(θ|y) can be sequentially estimated using SMC for the estimation of the parameter θ. In SMC 2, for each particle of the parameter θ, the particle filter of the state-space representation is executed, and by repeating the prediction, observation, and correction of the output variable y, the posterior distribution p(θ|y) of the parameter θ can be sequentially estimated. t and the observed value of y to correct the group of candidate values. By repeating these prediction, observation, and correction, the distribution of the state variable x is approximated by a set of realized values. If such a state-space representation in which the state variable x has the parameter θ to be estimated as a value can be formulated using the input-output series, the posterior distribution p(θ|y) can be sequentially estimated using SMC for the estimation of the parameter θ. In SMC 2, for each particle of the parameter θ, the particle filter of the state-space representation is executed, and by repeating the prediction, observation, and correction of the output variable y, the posterior distribution p(θ|y) of the parameter θ can be sequentially estimated.

[0042] The advantage of SMC 2 is that the probability distribution of the parameters is arbitrary and no particular assumptions are made. Figure 3 is a diagram showing an example of the estimation result of the probability distribution. In Figure 3, as an example, an example of the probability distribution of the parameters of the interval selection model estimated by SMC 2 is shown. For example, in SMC 2, the probability distribution of the parameters is estimated in a form that approximates the distribution by a collection of samples. By plotting such samples on a graph with the vertical axis representing the probability and the horizontal axis representing the values of the parameters, the histogram shown in Figure 3 can be generated.

[0043] In addition, by obtaining the parameters of the dynamic model that minimize the sum of the squared prediction errors in each section and their confidence intervals by the prediction error method, it is also possible to approximately estimate the probability distribution of the parameters. This method is applicable to model structures to which the prediction error method is applicable, and the estimation method including the confidence interval is disclosed in Non-Patent Document 1. It should be noted that since this estimation method assumes that the probability distribution of the parameter error follows a normal distribution with a mean of 0, if this assumption is not actually satisfied, the reliability of the estimated distribution will decrease.

[0044] Also, if the parameters of the dynamic model are sequentially estimated using the sequential least squares method in each section, a distribution of the estimated values of the parameters can be obtained, and thus it can be used to replace the estimated probability distribution. Since what is obtained by this method is the distribution of the estimated values of the parameters, it should be noted that if the estimation by the sequential least squares method does not work well, a reliable estimation cannot be achieved.

[0045] The distribution evaluation unit 14 is a processing unit that evaluates the probability distribution of the parameters of the section selection model. For example, the distribution evaluation unit 14 calculates, for each section, an evaluation value that evaluates the statistical variation from the probability distribution of the parameters of the section selection model estimated by the distribution estimation unit 13 as a statistical variation index.

[0046] Examples of such statistical variation indices include, first, the variance of the probability distribution, the standard deviation, etc. In addition, when the probability distribution estimated by the distribution estimation unit 13 is approximated by a collection of samples, the distribution evaluation unit 14 can also calculate the difference between the maximum value and the minimum value as a statistical variation index. In this case, from the aspect of avoiding the influence of samples that are outliers, as an example, a predetermined ratio, for example, 5% of the values from the top and bottom of the probability distribution in the collection of samples may be excluded, and then the difference between the maximum value and the minimum value may be calculated as a statistical variation index. Note that as the statistical variation index, other statistical values representing the statistical variation of the probability distribution can be used in addition to the variance, the standard deviation, and the difference between the maximum value and the minimum value.

[0047] The interval selection unit 15 is a processing unit that selects an interval with a small statistical variation in the probability distribution.

[0048] For such interval selection, as an example, a determination method using clustering can be applied. Clustering is a type of unsupervised learning and is a method of classifying a plurality of data into groups called clusters. In clustering, attention is paid to the similarity and patterns between data, and based on this, the data is classified into a plurality of clusters. Here, let the number of parameters of the interval selection model be m and the number of intervals be N. Let the statistical variation index of the m-th parameter in the t-th interval (t = 1, 2, ···, N) be V m (t), and the vector obtained by combining them into one vector is represented by V(t). Representing these relationships by mathematical formulas gives the following equation. In the following equation, "T" represents the transpose of a matrix. V(t)=[V1(t)V2(t)…V m (t)]T,t = 1, 2, ···, N

[0049] Regarding this V(t) as one piece of data, by applying clustering to N pieces of data, it is classified into a plurality of clusters. As one such clustering method, the k-means method and the like can be mentioned.

[0050] FIG. 4 is a diagram showing an example of the clustering result. In FIG. 4, the number of parameters is two, namely parameter 1 and parameter 2, and the clustering result is shown when the statistical variation index is the variance. Further, the horizontal axis of the graph shown in FIG. 4 indicates the variance of parameter 1, and the vertical axis indicates the variance of parameter 2. Further, each of the points plotted in FIG. 4 corresponds to one V(t) and corresponds to a certain interval t.

[0051] As shown in FIG. 4, the N intervals are classified into three clusters C1 to C3. In this way, it is possible to classify them into cluster C3 with relatively small statistical variation and clusters C1 and C2 with relatively large statistical variation. In this case, while the interval selection unit 15 selects intervals belonging to cluster C3 with relatively small statistical variation, it can exclude intervals belonging to the other clusters C1 and C2.

[0052] FIGS. 5 and 6 are diagrams (1) and (2) showing an example of the probability distribution of parameters. The horizontal axis of the graphs shown in FIGS. 5 and 6 indicates the values of the parameters, and the vertical axis indicates the probability. For example, FIG. 5 shows the probability distribution of the parameters of the interval selection model corresponding to the intervals belonging to clusters C1 and C2 with relatively large statistical variation among the clusters obtained by the clustering result shown in FIG. 4. On the other hand, FIG. 6 shows the probability distribution of the parameters of the interval selection model corresponding to the intervals belonging to cluster C3 with relatively small statistical variation among the clusters obtained by the clustering result shown in FIG. 4.

[0053] As shown in FIG. 5, when the statistical variation of the probability distribution of the parameters is large, it can be identified as an interval with low reliability and not suitable for system identification. On the other hand, as shown in FIG. 6, when the statistical variation of the probability distribution of the parameters is small, it can be identified as an interval with high reliability and suitable for system identification. From this, by selecting the intervals belonging to cluster C3 with relatively small statistical variation among the N intervals, it is possible to select intervals suitable for system identification.

[0054] Note that the method of classifying intervals with relatively small and large statistical variations is not limited to clustering, and other unsupervised learning methods may be used. The advantage of using such an unsupervised learning method is that interval selection can be realized by utilizing the relative magnitude relationship of statistical variations without prior knowledge of the distribution and magnitude of statistical variations.

[0055] Furthermore, here, as an example, a method of selecting an interval belonging to a cluster with relatively small statistical variation has been given, but the method of selecting an interval is not limited to those based on clusters. For example, N intervals can be sorted in ascending order of statistical variation, and intervals that satisfy a predetermined number in the upper part, for example, the upper 10, or a predetermined ratio in the upper part, for example, 25%, can be selected. In addition, until the total length of the intervals selected by the interval selection unit 15 reaches a predetermined ratio, for example, 25%, with respect to the entire time series data collected by the data collection unit 11, intervals with small statistical variation may be selected in order. These methods are common in that one or more intervals with relatively small statistical variation are preferentially selected. Since any method is selected based on relative magnitude, at least the interval with the smallest statistical variation is selected.

[0056] On the other hand, it is also possible to select an interval based on absolute magnitude rather than relative magnitude of statistical variation. For example, when there is prior knowledge about the distribution and magnitude of statistical variation, a method of setting a threshold based on that knowledge and selecting an interval whose statistical variation is below the threshold can be considered. When using such an interval selection method based on the absolute magnitude of statistical variation, since the statistical variation in all N intervals is larger than the set threshold, all intervals may be excluded and the number of intervals selected by the interval selection unit may become 0. Thus, this embodiment does not preclude the interval selection unit from excluding all intervals and the number of selected intervals becoming 0.

[0057] Also, a method of preferentially selecting one or more intervals with relatively small statistical variation and a method of selecting an interval based on absolute magnitude may be appropriately combined to select an interval.

[0058] The output unit 16 is a processing unit that outputs information on the section selected by the section selection unit 15. As an example of such an output destination, the identification unit 17 described later can be mentioned. For example, the output unit 16 can output the range of indexes indicating the section for each section selected by the section selection unit 15, or the range of date and time corresponding to the section. Not limited to such identification information of the section, the output unit 16 can acquire and output the segment data corresponding to the section for each section selected by the section selection unit 15 from the data collection unit 11. In this case, when the sections selected by the section selection unit 15 are divided into a plurality of non-consecutive sections, it is desirable to output in a form that shows where the sections selected by the section selection unit 15 are divided. This is because when the segment data of a plurality of non-consecutive sections are connected and output, discontinuous points that do not originally exist in the time-series data are generated, which has an adverse effect on system identification. If information on the discontinuous points is also output, it is possible to take measures such as avoiding the discontinuous points for identification, so that the adverse effect can be avoided.

[0059] As an example of another output destination, there are a display unit, a printing unit, etc. (not shown). For example, the output unit 16 can display a chart in which the section selected by the section selection unit 15 and the time-series data collected by the data collection unit 11 are plotted. According to such a display, it is possible to visualize which part of the time-series data is selected and which is not, so an engineer who performs system identification can automatically determine whether the selected section is appropriate. Further, when it is determined that the section selected by the section selection unit 15 is inappropriate, the threshold value used by the section selection unit 15 may be changed via a user interface such as a GUI or a CLI. As a result, on the one hand, if the section to be selected is not selected, it becomes possible to increase the threshold value so that a section with a larger statistical variation is also selected. On the other hand, if a section that should not be selected is selected, it becomes possible to decrease the threshold value so that only a section with a smaller statistical variation is selected. In addition, a user interface may be provided that allows an engineer to freely change the section selected by the section selection unit 15 by individually designating and excluding a part of the automatically selected section or individually designating and selecting a part of the section automatically excluded.

[0060] The identification unit 17 is a processing unit that executes system identification. For example, the identification unit 17 generates an identification model from the segment data of each section selected by the section selection unit 15 for each section. The system identification here may be a method different from the method used when the distribution estimation unit 13 estimates the probability distribution of the parameters of the section selection model, and it is not necessary to use the same method. For example, in the distribution estimation unit 13, a state space representation model is selected as the section selection model, and the probability distribution of the parameters is estimated using SMC2, while in the identification unit 17, a BJ model is selected as the identification model, and the prediction error method is applied to execute system identification by the least squares method, and such a proper use can be made.

[0061] <Flow of processing> Next, the processing flow of the information processing apparatus 10 according to the present embodiment will be described. FIG. 7 is a flowchart showing the procedure of the system identification process.

[0062] As shown in FIG. 7, when the data collection unit 11 collects the input / output time-series data regarding the system to be controlled (step S101), the section division unit 12 divides the input / output time-series data collected in step S101 into a plurality of sections (step S102).

[0063] Thereafter, loop process 1 that repeats the processes of step S103 below and step S104 below is executed for the number of times corresponding to the number N of the sections divided in step S102. Although FIG. 7 shows an example in which the processes of step S103 below and step S104 below are repeated, the processes of step S103 below and step S104 below may be executed in parallel for each of the N sections.

[0064] That is, the distribution estimation unit 13 estimates the probability distribution regarding the parameters of the section selection model generated from the segment data of the t-th section (step S103). Subsequently, the distribution evaluation unit 14 calculates, as a statistical variation index, an evaluation value for evaluating the statistical variation from the probability distribution of the parameters of the section selection model estimated in step S103 (step S104).

[0065] By repeating such loop process 1, the statistical variation is calculated for each of the N sections.

[0066] Then, the section selection unit 15 selects the sections belonging to the cluster with relatively small statistical variation calculated in step S104 (step S105). Note that, in step S105, the sections belonging to the cluster with relatively large statistical variation are excluded.

[0067] Subsequently, the output unit 16 outputs the information of the section selected in step S105 to the identification unit 17 (step S106). Then, the identification unit 17 executes system identification to generate an identification model from the segment data corresponding to the section output in step S106 (step S107), and ends the process.

[0068] Note that the process shown in FIG. 7 can be appropriately omitted when possible. For example, in a certain section divided by the section division unit 12, if the time-series data (especially the input time-series data) hardly changes and is substantially constant, it is obvious that this section will not be suitable for identification. In such a case, the processes of the distribution estimation unit 13 and the distribution evaluation unit 14 may be omitted, and the section selection unit 15 may exclude the section on the grounds that there is no change in the time-series data. By omitting the processing of such obvious sections in this way, the effect of shortening the processing time can be obtained.

[0069] <One aspect of the effect> As described above, the information processing apparatus 10 according to the present embodiment estimates the probability distribution of the parameters of the model generated from the segment data corresponding to each section into which the input / output time-series data related to the system to be controlled is divided, and selects a section with a small statistical variation in the probability distribution. Therefore, according to the information processing apparatus 10 according to the present embodiment, it becomes possible to realize automatic extraction of a section suitable for identification of a system having dynamic characteristics. As a result, as one aspect, the labor of selecting and extracting a section suitable for system identification from the collected time-series data, for example, the labor and time of experts such as engineers, are reduced.

[0070] <Other embodiments> Now, although the embodiments of the present disclosure have been described so far, various applications are possible, and furthermore, in addition to the above-described embodiments, it may be implemented in various different forms.

[0071] <Application example 1: Use of a low-order model> As an application example 1 of the above embodiment, the structure of the interval selection model used in the distribution estimation unit 13 can be a low-order model, for example, a first-order lag system. This realizes a reduction in the amount of calculation and thus a shortening of the processing time.

[0072] When expressing the model of the first-order lag system by a transfer function, it becomes the following formula (1). As shown in the following formula (1), the model of the first-order lag system is represented by two parameters, a gain G and a time constant T.

[0073]

Equation

[0074] Hereinafter, taking the case where there are two inputs and one output for system identification as an example, as disclosed in Non-Patent Document 3, the number of inputs may be three or more. Here, the k-th input is u k (t), its Laplace transform is U k (s), the output is y(t), its Laplace transform is Y(s), and the gain and time constant of the transfer function from the k-th input to the output are G k and T k respectively. In this case, the transfer function representation of the interval selection model is as shown in formula (2).

[0075]

Equation

[0076] From the aspect of estimating the probability distribution of the parameters of the interval selection model by SMC2, the transfer function representation of the interval selection model is converted into a discrete-time state-space representation. Such a conversion can be realized by the following formula (3) and the following formula (4). Here, for simplicity of explanation, an example with a sampling interval of 1 is given, but the sampling interval can be arbitrarily changed.

[0077]

Equation

[0078] Here, "u" k " refers to the vector obtained by vertically arranging Input 1 and Input 2. "x" k " refers to the state variable (2 rows by 1 column). "v" k " refers to the system noise (2 rows by 1 column). "y" k " refers to the output (scalar). "w" k " refers to the observation noise. "k" refers to the index representing time. It is assumed that each of the system noise and the observation noise follows a normal distribution independently at each time instant.

[0079] Under such settings, the distribution estimation unit 13 estimates the probability distributions of G1, G2, T1, and T2, which are the parameters of the interval selection model, by SMC2. Other than this, the processing is the same as that described in the above embodiment.

[0080] What is obtained by interval selection is not the parameters of the interval selection model itself, but the statistical variation of the probability distribution of the parameters. For example, as an interval suitable for system identification, an interval satisfying the following conditions (a) to (c) can be selected. (a) The time-series data of the input and output vary moderately. (b) The causal relationship between the input and output can be confirmed from the time-series data. (c) There is little disturbance included in the time-series data.

[0081] And, to determine the three conditions of (a) to (c) above, it can be considered sufficient as long as it is a model that can roughly represent the dynamic characteristics between input and output. Even for such a model, due to reasons such as no variation in the time-series data of input and output, unclear causal relationship between input and output, and many disturbances, the phenomenon of large statistical variation in parameters can be captured. With such a perspective, it can be conceived to use a first-order lag system, which is the simplest among the models that can represent dynamic characteristics, as the model for interval selection. Since the first-order lag system has only two parameters, gain and time constant, and can shorten the processing time of the distribution estimation unit 13, according to this Application Example 1, an effect can be obtained that intervals can be automatically selected in a shorter processing time compared to the case of using a higher-order model.

[0082] However, when using a first-order lag system as the model for interval selection, there may be cases where the assumption that the probability distribution of the parameters follows a normal distribution is not satisfied. In this case, it may be appropriate to estimate the probability distribution using a method based on the sequential Monte Carlo method that does not require an assumption for the probability distribution of the parameters. The method using the confidence interval of the least squares method assumes that the probability distribution of the parameters follows a normal distribution, so there is a possibility that the performance will deteriorate in this Application Example 1.

[0083] <Application Example 2: Estimation of Only Gain> As a further Application Example 2 of the above Application Example 1, the part representing the dynamic characteristics of the interval selection model used in the distribution estimation unit 13 is fixed, and the estimation is narrowed down to the probability distribution of the gain. Thereby, a further reduction in the amount of calculation and thus a further shortening of the processing time are achieved.

[0084] For example, as the interval selection model, a model having the structure of the following formula (5) is used. Here, "G" represents the gain, and "D(s)" represents a transfer function with a gain of 1, and they are in a structure of multiplying these two. D(s) represents only the dynamic characteristic part of the object of system identification, and the gain is normalized to 1. The distribution estimation unit 13 performs estimation by narrowing down to the probability distribution of the gain G. D(s) can be set in advance based on knowledge about the object of system identification, etc., and is fixed in the estimation by the distribution estimation unit 13 and is out of the estimation target. Note that the part representing the dynamic characteristics of the object of system identification corresponding to D(s) may be in other expressions as long as it can represent the dynamic characteristics, such as state-space representation. GD(s) ··· (5)

[0085] In this application example 2, it can be applied to a use case where the dynamic characteristics of the object of system identification can be obtained from modeling based on physical laws or computer simulations, and only the part corresponding to the gain is modeled with time-series data collected from the actual object of system identification. In such a use case, since only the gain has a large uncertainty, it is possible to determine whether each interval is suitable for system identification only from the statistical variation of the probability distribution of the gain. Also, by fixing the dynamic characteristic part, the estimation target of the probability distribution is narrowed down to the gain, so the algorithm for estimating the probability distribution can be simplified or made simpler. This is because to estimate the probability distribution of the dynamic characteristic part, past observation values and state histories are required, whereas for the estimation of only the gain, it can be estimated without past history. By this, there is an effect of reducing the processing amount not only by reducing the number of parameters for estimating the probability distribution. Note that this application example 2 can prohibit or limit the application when, for example, the approximate time constant is also unknown and the dynamic characteristics are unknown, or when the variation of the dynamic characteristics is large.

[0086] <Numerical values, etc.> The matters described in the above embodiments, such as specific examples like the number of inputs and outputs of the system to be controlled, the types of interval selection models and identification models, the types of algorithms used for estimating probability distributions, etc., are merely examples and can be changed. Also, the flowcharts described in the embodiments can have their processing order changed within a non - contradictory range.

[0087] <Component> Regarding the processing procedures, control procedures, specific names, and information including various data and parameters shown in the above documents and drawings, they can be arbitrarily changed unless otherwise specified. For example, any one or more of the functional units among the data collection unit 11, interval division unit 12, distribution estimation unit 13, distribution evaluation unit 14, interval selection unit 15, output unit 16, and identification unit 17 that the information processing apparatus 10 has may be configured by separate apparatuses.

[0088] Moreover, each component of each illustrated apparatus is a functional concept and does not necessarily have to be physically configured as shown in the figure. That is, the specific forms of dispersion and integration of each apparatus are not limited to those shown in the figure. In other words, all or part of it can be functionally or physically dispersed and integrated in arbitrary units according to various loads, usage situations, etc. Note that each configuration may be a physical configuration.

[0089] Furthermore, each processing function performed by each apparatus can be realized in whole or in any part by a CPU (Central Processing Unit) and a program analyzed and executed by the CPU, or can be realized as hardware by wired logic.

[0090] The above - mentioned components include those that can be easily assumed by those skilled in the art, those that are substantially the same, and those within the so - called equivalent range. Furthermore, the above - described embodiments can be appropriately combined within a range that does not conflict with the processing content.

[0091] In addition, the above-mentioned "section (module, -er suffix, -or suffix)" can be read as a unit, means, circuit, etc. For example, a communication section (communication module), a control section (control module), and a storage section (storage module) can be read as a communication unit, a control unit, and a storage unit, respectively.

[0092] <Hardware> Next, a hardware configuration example of the computer described in the above embodiment will be described. FIG. 8 is a diagram showing a hardware configuration example. As shown in FIG. 8, the information processing apparatus 10 includes a communication apparatus 10a, a storage apparatus 10b, a memory 10c, and a processor 10d. Note that each unit shown in FIG. 8 may be connected to each other by a bus or the like.

[0093] The communication apparatus 10a is a network interface card or the like. The storage apparatus 10b is a storage device such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive). For example, the storage apparatus 10b stores a program for operating the functions shown in FIG. 7, a DB (DataBase), and the like.

[0094] The processor 10d reads a program for executing the same processing as the processing section shown in FIG. 1 from the storage apparatus 10b or the like and expands it in the memory 10c, thereby operating a process for executing the functions described in FIG. 1.

[0095] Such a process realizes the same functions as the processing section included in the information processing apparatus 10. For example, the processor 10d reads a program having the same functions as the data collection section 11, the section division section 12, the distribution estimation section 13, the distribution evaluation section 14, the section selection section 15, the output section 16, the identification section 17, etc. from the storage apparatus 10b or the like. Then, the processor 10d executes a process for executing the same processing as the data collection section 11, the section division section 12, the distribution estimation section 13, the distribution evaluation section 14, the section selection section 15, the output section 16, the identification section 17, etc.

[0096] In this way, the information processing apparatus 10 operates as an information processing apparatus that executes the interval selection method by reading and executing a program. Also, the information processing apparatus 10 can read the program from a recording medium by a medium reading device and realize the same functions as those of the above-described embodiments, Application Example 1, and Application Example 2 by executing the read program. Note that the program in other embodiments is not limited to being executed by the information processing apparatus 10. For example, when another computer or server executes the program, or when they cooperate to execute the program, the above-described interval selection function can be similarly applied.

[0097] The above program can be distributed via a network such as the Internet. Also, the above program can be recorded on an arbitrary recording medium and executed by being read from the recording medium by a computer. For example, the recording medium can be realized by a hard disk, a flexible disk (FD), a CD-ROM, an MO (Magneto-Optical disk), a DVD (Digital Versatile Disc), or the like.

[0098] As described above, some of the embodiments have been described in detail with reference to the drawings. However, these are merely examples, and each embodiment can be implemented in other forms that are variously modified and improved based on the knowledge of those skilled in the art, including the aspects described in the column of the disclosure of the invention.

Explanation of Reference Numerals

[0099] 10 Information processing apparatus 11 Data collection unit 12 Interval division unit 13 Distribution estimation unit 14 Distribution evaluation unit 15 Interval selection unit 16 Output unit 17 Identification unit

Claims

1. A collection unit that collects time-series data of input / output related to a system to be controlled, an estimation unit that estimates a probability distribution regarding parameters of a first model generated from segment data corresponding to each interval into which the time-series data is divided, a selection unit that selects an interval in which the statistical variation of the probability distribution is small, an output unit that outputs information regarding the interval selected by the selection unit, and an information processing apparatus characterized by comprising these components.

2. The information processing apparatus according to claim 1, wherein the selection unit includes a process of clustering the intervals based on the statistical variation of the probability distribution and selecting one or more intervals included in a cluster having a relatively small statistical variation among the clusters obtained as a result of the clustering.

3. The information processing apparatus according to claim 1, further comprising an identification unit that generates a second model used for identifying the system based on segment data corresponding to the interval output by the output unit among the time-series data collected by the collection unit.

4. The information processing apparatus according to claim 3, wherein the second model has a structure different from that of the first model.

5. The information processing apparatus according to claim 1, wherein the first model has a structure of a transfer function representation of a first-order lag system.

6. The first model has a model structure that multiplies a gain and a dynamic characteristic part of the system, and the estimation unit fixes the dynamic characteristic part and narrows down to the gain to perform estimation of the probability distribution.

7. Collect time-series data of input / output related to a system to be controlled, estimate a probability distribution regarding parameters of a first model generated from segment data corresponding to each interval into which the time-series data is divided, select an interval in which the statistical variation of the probability distribution is small, output information regarding the selected interval, and a section selection method characterized in that a computer executes the process.

8. Collect time-series data of input / output related to a system to be controlled, estimate a probability distribution regarding parameters of a first model generated from segment data corresponding to each interval into which the time-series data is divided, select an interval in which the statistical variation of the probability distribution is small, output information regarding the selected interval, and a section selection program characterized in that the process is executed by a computer.

Images