Information processing apparatus, interval selection method, and computer-readable storage medium

Abstract

The invention provides an information processing apparatus, a section selection method, and a computer readable storage medium, which can realize automatic extraction of sections suitable for system identification and representing dynamic characteristics. An information processing apparatus includes: a collection unit that collects time-series data of input / output relating to a system to be controlled; an estimation unit that estimates, for each segment obtained by dividing the time-series data, a probability distribution relating to a parameter of a model generated from segment data corresponding to the segment; a selection unit that selects an interval in which the statistical deviation of the probability distribution is small; and an output unit that outputs information relating to the section selected by the selection unit.

Description

Technical Field

[0001] The present invention relates to an information processing apparatus, an interval selection method, and a computer-readable storage medium. Background Art

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

[0003] As one of methods for generating such a model, there is a method called system identification, which models using a statistical estimation method based on time series data of input / output related to a system to be controlled. For example, in Patent Document 1 and Patent Document 2, a technique of a type that applies a test signal such as an M sequence to an input to a system to be controlled for system identification is proposed.

[0004] [Prior Art Documents]

[0005] [Patent Documents]

[0006] [Patent Document 1] Japanese Patent Application Laid-Open No. 2012-528392

[0007] [Patent Document 2] U.S. Patent No. 7209793

[0008] [Non-Patent Documents]

[0009] [Non-Patent Document 1] L. Ljung: System Identification - Theory for the User - 2nd Edition, Prentice Hall (1999)

[0010] [Non-Patent Document 2] N. Chopin, O. Papaspiliopoulos, “An Introduction to Sequential Monte Carlo”, Springer (2020)

[0011] [Non-Patent Document 3] Yamauchi, Watanabe, Nishiguchi: Successive Update of Process Model Based on Evaluation of Parameter Posterior Distribution, Proceedings of the 65th Joint Conference on Automatic Control (2022) Summary of the Invention

[0012] [Problems to be Solved by the Invention]

[0013] However, in the case of system identification using time series data collected from a system to be controlled, there is an aspect of entrusting the task of selecting an interval suitable for system identification to an expert such as an engineer.

[0014] In one aspect, an object of the present invention is to provide an information processing apparatus, an interval selection method, and a computer-readable storage medium that can automatically extract an interval suitable for system identification having dynamic characteristics.

[0015] [Technical means for solving the problem]

[0016] An information processing apparatus according to an embodiment 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 related to parameters of a model generated from segmented data corresponding to the interval for each interval obtained by segmenting the time series data; a selection unit that selects an interval with a small statistical deviation of the probability distribution; and an output unit that outputs information related to the interval selected by the selection unit.

[0017] [Effects of the invention]

[0018] As one aspect, it is possible to automatically extract an interval suitable for system identification having dynamic characteristics. Description of the drawings

[0019] Figure 1 It is a block diagram showing a functional configuration example of the information processing apparatus.

[0020] Figure 2 It is a schematic diagram showing an example of interval segmentation.

[0021] Figure 3 It is a graph showing an example of the estimation result of the probability distribution.

[0022] Figure 4 It is a graph showing an example of the clustering result.

[0023] Figure 5 It is a graph showing an example of the probability distribution of the parameters.

[0024] Figure 6 It is a graph showing an example of the probability distribution of the parameters.

[0025] Figure 7 It is a flowchart showing the process of system identification processing.

[0026] Figure 8 It is a diagram showing an example of the hardware configuration.

[0027] [Explanation of symbols]

[0028] 10: Information processing device

[0029] 11: Data collection unit

[0030] 12: Interval division unit

[0031] 13: Distribution estimation unit

[0032] 14: Distribution evaluation unit

[0033] 15: Interval selection unit

[0034] 16: Output unit

[0035] 17: Identification unit Detailed implementation mode

[0036] Hereinafter, a mode (hereinafter referred to as "implementation mode") for implementing an information processing device, an interval selection method, and an interval selection program of the present disclosure will be described with reference to the accompanying drawings. In addition, the implementation mode only represents an example or aspect, and the structure, operation, function, nature, characteristics, method, use, etc. of the present disclosure are not limited by such an illustration.

[0037] <Embodiment 1>

[0038] <An example of the usage of the information processing device>

[0039] Figure 1 It is a block diagram showing an example of the functional structure of the information processing device. Figure 1 The shown information processing device 10 provides a system identification function, and the system identification function generates a model for predicting the behavior of a system to be controlled.

[0040] As an aspect, such a system identification function can be applied to model predictive control (MPC for short) such as process control in a plant and air conditioning control in a factory or building.

[0041] For example, the information processing device 10 can be implemented in such a form that it is integrally assembled in a control device (control machine) that collects temperature, pressure, flow rate, etc. from measuring machines configured in a plant to determine multi-variable control amounts such as temperature control, pressure control, and flow control.

[0042] In addition, the information processing apparatus 10 can also be implemented as a computer that collects the time series data of input / output from the above-described controller and provides the model generated by the system identification function. In this case, the information processing apparatus 10 can be implemented as a web server that provides the system identification function on-premise. In addition, the information processing apparatus 10 can also be implemented as a software as a service (SaaS)-type application, thereby providing the system identification function in the form of a cloud service.

[0043] <One aspect of the problem>

[0044] In the system identification, there is an aspect in which which interval in the time series data of the input / output related to the system to be controlled significantly affects the reliability of the model. In particular, in the case where the time series data collected during the normal operation or normal operation of the system to be controlled, such as a device or equipment, is used for system identification, the system identification becomes more difficult.

[0045] For example, during the normal operation or normal operation of the system, control is performed to target a steady state and minimize changes in input / output. In this case, it is easy to occur that there is no information suitable for system identification in the time series data for a long time, so the reliability of the model is reduced. In this way, a situation may occur in which system identification becomes difficult as the stability of the input / output of the system increases.

[0046] Due to the above situation, considering applying the techniques described in Patent Document 1 or Patent Document 2, even when the input / output of the system to be controlled is controlled to be constant during normal operation, the input is intentionally changed to collect time series data that can be used for system identification.

[0047] However, the techniques described in Patent Document 1 and Patent Document 2 are only techniques for disturbing the input to the system, so they can only be applied to the scenario of collecting time series data of input / output. Therefore, the techniques described in Patent Document 1 and Patent Document 2 have an aspect that they cannot be applied to system identification implemented in the scenario after the time series data of input / output has been collected.

[0048] In terms of such an aspect, when performing system identification based on the time series data collected during normal operation, it will result in entrusting the task of selecting an interval suitable for system identification to an expert such as an engineer. For example, for an engineer, it is necessary to view the plot of the time series data while Figure 1Manual operations are performed to select intervals suitable for system identification. Such manual operations become more labor-intensive as the time length of the time-series data increases, and it is difficult to perform such manual operations except for those with professional knowledge such as engineers.

[0049] <One aspect of the problem-solving approach>

[0050] Therefore, in the present embodiment, as part of the system identification function, an interval selection function is provided, and the interval selection function realizes the automatic extraction of intervals suitable for system identification with dynamic characteristics.

[0051] Such an interval selection function can be realized only with the following viewpoints. That is, a novel viewpoint that the statistical deviation of the parameters of the model for predicting the operation of the system to be controlled changes according to whether it is an interval suitable for system identification with dynamic characteristics is flexibly used as a problem-solving approach. Here, the interval suitable for system identification with dynamic characteristics refers to the following interval: in this interval, a causal relationship including dynamic characteristics can be confirmed between the time-series data of the input to the system and the time-series data of the output from the system, and if the input / output time-series data of this interval is used to identify the system, a highly reliable model can be expected to be obtained.

[0052] In system identification, both the obtained model and the parameters of the model are estimated values, so there remains uncertainty. One of the reasons for the uncertainty is the aspect that the time-series data used in system identification contains external interference or noise, and the aspect that the device or equipment itself to be identified as a system can change. Therefore, the parameters of the model obtained through system identification can be regarded as variables following a certain probability distribution.

[0053] Even if the devices or equipment to be identified as a system are the same, this probability distribution will change according to which interval of the time-series data is used. Moreover, the interval suitable for system identification is smaller than the interval with a large statistical deviation of the probability distribution.

[0054] Usually, in system identification, for the purpose of obtaining an estimated value of a model or its parameters, the probability distribution is not taken into account, but the probability distribution of the parameters can be estimated by using the sequential Monte Carlo method or the like.

[0055] Due to the above situation, in the interval selection function, the time-series data is divided into multiple intervals, and the intervals with a small statistical deviation of the parameter distribution of the model are selected. On the other hand, the intervals with a large statistical deviation are excluded, thereby selecting intervals suitable for system identification with dynamic characteristics.

[0056] Therefore, through the interval selection function of the present embodiment, it is possible to automatically extract an interval suitable for system identification with dynamic characteristics. Thus, as one aspect, it is possible to reduce the labor, such as the man-hours or time of experts such as engineers, for selecting and exporting an interval suitable for system identification from the collected time series data.

[0057] <Structure of Information Processing Apparatus>

[0058] Next, a functional structure example of the information processing apparatus 10 of the present embodiment will be described. In Figure 1 this, the blocks associated with the system identification function possessed by the information processing apparatus 10 are patterned.

[0059] Here, the interval selection function may be provided as part of the system identification function and packaged into the system identification function. However, it is not limited to this, and the interval selection function may also be provided independently of the system identification function. In this case, the hardware resources for providing each of the system identification function and the interval selection function, such as physical devices or physical components, may be different.

[0060] As Figure 1 shown, the information processing apparatus 10 includes a data collection unit 11, an interval division unit 12, a distribution estimation unit 13, a distribution evaluation unit 14, an interval selection unit 15, an output unit 16, and an identification unit 17. In addition, in Figure 1 this, only the functional units associated with the system identification function are excerpted and shown, and it does not prevent the information processing apparatus 10 from including functional units other than those shown in the figure.

[0061] The data collection unit 11 is a processing unit that collects time series data of input / output related to a system to be controlled. For example, some control device (not shown), such as the control device, is connected to the device or equipment to be controlled, so that the input / output data of the device or equipment can be obtained from the control device. By collecting the input / output data in real time or in batches to obtain a database or file storing a log of the input / output data, it is possible to collect time series data of the input / output. In addition, the data collection unit 11 may also directly collect input / output data from the device or equipment to be controlled and save it in a storage unit (not shown), and use it as time series data. Furthermore, data preprocessing such as trend removal, filtering, deviation value removal, and missing value interpolation may be applied to the time series data of the input / output.

[0062] The interval division unit 12 is a processing unit that divides the time series data of the input / output into a plurality of intervals. Hereinafter, the partial data corresponding to each interval in the entire time series data may sometimes be described as "segmented data".

[0063] Figure 2 This is a schematic diagram showing an example of interval segmentation. As Figure 2 shown, the interval segmentation unit 12 can divide the entire input / output time series data into a total of N intervals, namely interval 1, interval 2, interval 3, …, interval N. For example, the interval segmentation unit 12 can set the time lengths of the respective intervals at equal intervals. Furthermore, the interval segmentation unit 12 can cause the adjacent intervals before and after to overlap at an arbitrary ratio, such as 50%. For example, Figure 2 shows an example where the intervals before and after overlap at a ratio of 50%. In this case, the moment at the central position of interval 1 coincides with the moment at the start position of interval 2, and the moment at the central position of interval 2 coincides with the moment at the start position of interval 3, and the moment at the central position of interval N-1 coincides with the moment at the start position of interval N.

[0064] Here, the time length of the interval can be set according to the following criteria. Suppose the interval selection function of the present embodiment is applied to the system identification of a dynamic model. Therefore, the interval length can be set to at least the degree capable of estimating the dynamic characteristics of the system identification object, thereby suppressing the case where an overly short time length is set. For example, when the settling time of the step response of the system identification object is known, the interval length can be set to the same degree as the settling time. On the other hand, when the interval length is too long, the possibility that the divided intervals contain periods unsuitable for system identification becomes high, and there is an aspect that it becomes difficult to derive only the intervals suitable for system identification. Therefore, the interval length can be set to be, for example, about 10 times the integer multiple of the settling time as the upper limit and set to be below the upper limit.

[0065] In addition, the purpose of the interval selection function is not system identification itself, but to select intervals suitable for system identification. Therefore, the settling time referred to when setting the interval length can also be of low precision. In addition, when the settling time is unknown, the interval length can be explored by changing the interval length in units of 2 times to 1 minute, 2 minutes, 4 minutes, … and attempting interval selection.

[0066] In addition, in Figure 2 , an example of overlapping adjacent intervals before and after is given, but it is not necessary to overlap the intervals before and after. For example, the processing amount or processing time can be reduced as the overlapping ratio in the intervals before and after is decreased. Furthermore, in Figure 2 , an example of setting the lengths of the respective intervals at equal intervals is given, but it is not limited thereto, and different interval lengths can also be set for each interval.

[0067] The distribution estimation unit 13 is a processing unit that estimates the probability distribution related to the parameters of the model generated from the segmented data of each interval.

[0068] As one aspect, the models generated by the distribution estimation unit 13 and the models generated by the recognition unit 17 described later have different purposes of use, and thus can be different models. Therefore, from the aspect of distinguishing the labels of the two models, the model used by the distribution estimation unit 13 in the distribution estimation for interval selection is sometimes described as the "model for interval selection". On the other hand, the model used by the recognition unit 17 described later in system recognition is sometimes described as the "model for recognition". Furthermore, in the case where the individuals of the model for interval selection and the model for recognition do not need to be distinguished, or in the case where it is clear which of the two models is being referred to, it is sometimes simply referred to as the "model".

[0069] The structure of such a model for interval selection can select any mathematical model such as state space representation or transfer function type. For example, the transfer function type can include an Auto-Regressive with eXogenous (ARX) model, an Auto-Regressive Moving Average (ARMAX) model, an Output Error (OE) model, a Box-Jenkins (BJ) model, etc. In addition to this, it is also possible to use a non-parametric model such as a Finite Impulse Response (FIR) model or a step response, but a parametric model that models with a small number of parameters is more effective in reducing the processing time.

[0070] For example, as one of the methods for directly performing Bayesian estimation of the probability distribution of the parameters of a dynamic model, Sequential Monte Carlo (SMC) 2 using the sequential Monte Carlo method can be cited. The SMC2 has been disclosed in various documents such as Non-Patent Document 2. By using the state space representation as the model structure of the model for interval selection and applying the SMC2 to the segmented data corresponding to each interval, it is possible to estimate the probability distribution, more precisely the posterior distribution, of the parameters in each interval.

[0071] In SMC, a set of candidate values of the state variable x is prepared according to the prior distribution p(x0), and each candidate value is used as an initial value. By comparing the predicted value set of the output variable y obtained by solving the state equation with the observed value of y, the set of candidate values is corrected. By repeatedly performing the prediction, observation, and correction, the distribution of the state variable x is approximated by the set of realized values. If such a state variable x can be used with the input / output sequence t t t t t ​​​​When formulating a state-space representation having a value in a parameter θ to be estimated, the SMC can be used to sequentially estimate the posterior distribution p(θ|y) in the estimation of the parameter θ. In SMC2, particle filtering of the state-space representation is performed for each particle of the parameter θ, and prediction, observation, and correction of the output variable y are repeatedly performed, whereby the posterior distribution p(θ|y) of the parameter θ can be sequentially estimated. t As for the advantages of SMC2, it can be cited that the probability distribution of the parameter is arbitrary and no special assumptions are made.

[0072] FIG. is an example showing the estimation result of the probability distribution. In Figure 3 As an example, an example of the probability distribution of the parameters of the model for interval selection estimated by SMC2 is shown. For example, in SMC2, the probability distribution of the parameter is estimated in a form of approximating the distribution by using a set of samples. By plotting such samples as a graph with the vertical axis being the probability and the horizontal axis being the value of the parameter, the histogram shown in Figure 3 can be generated. Figure 3 is shown.

[0073] In addition, by using the prediction error method to obtain the parameters of the dynamic model that minimize the sum of the squares of the prediction errors in each interval and their confidence intervals, the probability distribution of the parameters can also be approximately estimated. This method can be applied to the model structure to which the prediction error method can be applied, and the estimation method including the confidence interval has been disclosed in Non-Patent Document 1. It should be noted that the said estimation method assumes that the probability distribution of the error of the parameter follows a normal distribution with a mean of 0. Therefore, when the said assumption is not actually satisfied, the reliability of the estimated distribution is reduced.

[0074] In addition, if the parameters of the dynamic model are sequentially estimated by using the sequential least squares method in each interval, the distribution of the estimated values of the parameters can be obtained. Therefore, this can also be used instead of estimating the probability distribution. It should be noted that the distribution of the estimated values of the parameters is obtained by using the said method. Therefore, when the estimation based on the sequential least squares method is not performed well, a reliable estimation cannot be achieved.

[0075] The distribution evaluation unit 14 is a processing unit that evaluates the probability distribution of the parameters of the model for interval selection. For example, for each interval, the distribution evaluation unit 14 calculates an evaluation value for evaluating the statistical deviation as a statistical deviation index according to the probability distribution of the parameters of the model for interval selection estimated by the distribution estimation unit 13.

[0076] As an example of such a statistical deviation index, standard deviation with the variance of the probability distribution at the head can be cited, etc. In addition, in the case of approximating the probability distribution estimated by the distribution estimation unit 13 with a set of samples, the distribution evaluation unit 14 can also calculate the difference between the maximum value and the minimum value as a statistical deviation index. In this case, from the aspect of avoiding the influence of samples that have become deviated values, as an example, after excluding values of a prescribed ratio, for example 5%, from the upper and lower positions of the probability distribution in the set of samples, the difference between the maximum value and the minimum value can be calculated as a statistical deviation index. In addition, as the statistical deviation index, in addition to using variance, standard deviation, or the difference between the maximum value and the minimum value, other statistical values representing the statistical deviation of the probability distribution can also be used.

[0077] The interval selection unit 15 is a processing unit that selects an interval with a small statistical deviation of the probability distribution.

[0078] In 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 multiple data into groups called clusters. In clustering, the similarity or pattern between data is focused on, and based on this, the data is classified into multiple clusters. Here, the number of parameters of the interval selection model is set to m, and the number of intervals is set to N. The statistical deviation index of the m-th parameter in the t-th interval (t = 1, 2,..., N) is represented by V m (t), and the one obtained by aggregating them into a single vector is represented by V(t). If their relationship is expressed in a mathematical formula, it is as shown in the following formula. "T" in the following formula represents the transpose of a matrix.

[0079] V(t) = [V1(t) V2(t)... V m (t)]T, t = 1, 2,..., N

[0080] Regarding the V(t) as one data, applying clustering to N quantities of data, and thus classifying them into multiple clusters. As one of such clustering methods, the k-means method etc. can be cited.

[0081] Figure 4 is a diagram showing an example of the clustering result. In Figure 4 , the clustering result in the case where the number of parameters is two, namely parameter 1 and parameter 2, and the statistical deviation index is variance, is shown. Furthermore, Figure 4 in the horizontal axis of the chart shown refers to the variance of parameter 1, and the vertical axis refers to the variance of parameter 2. Furthermore, Figure 4 the points plotted in correspond to one V(t) each and correspond to a certain interval t.

[0082] As Figure 4As shown, N intervals are classified into three clusters C1 to C3. In this way, the clusters can be classified into cluster C3 with relatively small statistical deviation, and clusters C1 and C2 with relatively large statistical deviation. In this case, the interval selection unit 15 can select the intervals belonging to cluster C3 with relatively small statistical deviation, and on the other hand, exclude the intervals belonging to clusters C1 and C2 other than these.

[0083] Figure 5 and Figure 6 is a diagram showing an example of the probability distribution of parameters. Figure 5 and Figure 6 In the diagrams shown, the horizontal axis represents the value of the parameter, and the vertical axis represents the probability. For example, in Figure 5 it shows the probability distribution of the parameters of the interval selection model corresponding to the intervals belonging to cluster C1 or cluster C2 with relatively large statistical deviation among the clusters obtained from the clustering results shown by Figure 4 On the other hand, in Figure 6 it shows the probability distribution of the parameters of the interval selection model corresponding to the intervals belonging to cluster C3 with relatively small statistical deviation among the clusters obtained from the clustering results shown by Figure 4

[0084] In the case where the statistical deviation of the probability distribution of the parameters is large as shown in Figure 5 , the intervals can be recognized as having low reliability and not suitable for system identification. On the other hand, in the case where the statistical deviation of the probability distribution of the parameters is small as shown in Figure 6 , the intervals can be recognized as having high reliability and suitable for system identification. According to this situation, by selecting the intervals belonging to cluster C3 with relatively small statistical deviation among the N intervals, the intervals suitable for system identification can be selected.

[0085] In addition, the method of classifying into intervals with relatively small statistical deviation and intervals with relatively large statistical deviation is not limited to clustering, and other unsupervised learning methods can also be used. The advantage of using such unsupervised learning methods is that even if there is no prior knowledge about the distribution or magnitude of the statistical deviation, the relative magnitude relationship of the statistical deviation can be used to achieve interval selection.

[0086] Furthermore, a method of selecting an interval belonging to a cluster with a relatively small statistical deviation is exemplified here, but the method of selecting an interval is not limited to the method based on clusters. For example, the N intervals can be sorted in ascending order of statistical deviation, and the upper specified number of intervals, such as the upper 10 intervals, or intervals satisfying the upper specified ratio, such as 25%, can be selected. In addition to this, intervals can also be selected sequentially starting from the intervals with small statistical deviation until the sum of the lengths of the intervals selected by the interval selection unit 15 reaches a specified ratio, such as 25%, of the entire time series data collected by the data collection unit 11. These methods are common in that they prioritize intervals with relatively small statistical deviations and select one or more intervals. Any of these methods is based on relative magnitudes for selection, so at least the interval with the smallest statistical deviation can be selected.

[0087] In contrast, intervals can also be selected based on absolute magnitudes rather than relative magnitudes of statistical deviations. For example, in the case where knowledge about the distribution or magnitude of statistical deviations is available in advance, a method of setting a threshold based on the knowledge and selecting intervals with statistical deviations below the threshold can be considered. When using such an interval selection method based on the absolute magnitude of statistical deviations, it is sometimes the case that the statistical deviations in all N intervals are greater than the set threshold, so all intervals are excluded and the number of intervals selected by the interval selection unit becomes 0. Thus, this embodiment does not exclude the case where the number of intervals to be selected by the interval selection unit is 0 because all intervals are excluded.

[0088] In addition, a method of preferentially selecting one or more intervals with relatively small statistical deviations and a method of selecting intervals based on absolute magnitudes can be appropriately combined to select intervals.

[0089] The output unit 16 is a processing unit that outputs information on the intervals selected by the interval selection unit 15. As an example of such an output destination, the identification unit 17 described later can be cited. For example, the output unit 16 can output, for each interval selected by the interval selection unit 15, the range of the index indicating the interval, or the range of the date and time corresponding to the interval. Not limited to such identification information of intervals, the output unit 16 can obtain the segmented data corresponding to the interval from the data collection unit 11 for each interval selected by the interval selection unit 15 and output it. At this time, when the intervals selected by the interval selection unit 15 are separated into a plurality of discontinuous intervals, it is desirable to output in a form that makes it possible to know where the intervals selected by the interval selection unit 15 are separated. The reason for this is that if the segmented data of a plurality of discontinuous intervals are connected and output, discontinuous points that originally do not exist will be generated in the time series data, which will have an adverse effect on system identification. If information related to the discontinuous points is output together, a countermeasure of avoiding the discontinuous points for identification can be taken, so the adverse effect can be avoided.

[0090] As an example of another output destination, a display unit or a printing unit (not shown) can be cited. For example, the output unit 16 can display a table graph drawn from the interval selected by the interval selection unit 15 and the time series data collected by the data collection unit 11. According to such a display, it is possible to visualize where in the time series data is selected and where is not selected. Therefore, an engineer performing system identification can determine whether the automatically selected interval is appropriate. In addition, it can be set that when it is determined that the interval selected by the interval selection unit 15 is inappropriate, the threshold used by the interval selection unit 15 can be changed via a user interface such as a graphical user interface (GUI) or a command-line interface (CLI). Thus, as one aspect, if an interval that should be selected is not selected, the threshold can be increased so that an interval with a greater statistical deviation is also selected. As another aspect, if an interval that should not be selected is selected, the threshold can be decreased so that only an interval with a smaller statistical deviation is selected. In addition to this, a user interface for separately specifying a part of the automatically selected interval and excluding it, or separately specifying a part of the automatically excluded interval and selecting it can also be provided, so that the interval selected by the interval selection unit 15 can be freely changed according to the judgment of the engineer.

[0091] The identification unit 17 is a processing unit that performs system identification. For example, for each interval selected by the interval selection unit 15, the identification unit 17 generates an identification model based on the segmented data of the interval. The system identification here can be a method different from the method used when the distribution estimation unit 13 estimates the probability distribution of the parameters of the interval selection model, and the same method may not necessarily be used. For example, the following differential use can be performed: the distribution estimation unit 13 selects a state space representation model as the interval selection model and uses SMC2 to estimate the probability distribution of the parameters. On the other hand, the identification unit 17 selects a BJ model as the identification model and applies the prediction error method to perform system identification by the least squares method.

[0092] <Process flow>

[0093] Next, the process flow of the information processing apparatus 10 of the present embodiment will be described. Figure 7 It is a flowchart showing the process of system identification processing.

[0094] As Figure 7 shown, when the data collection unit 11 collects time series data of input / output related to the system to be controlled (step S101), the interval division unit 12 divides the time series data of input / output collected in step S101 into a plurality of intervals (step S102).

[0095] Then, loop process 1 is executed, and loop process 1 repeatedly performs the processes of the following step S103 and the following step S104 for the number of times corresponding to the number N of the intervals segmented in step S102. In addition, in Figure 7 an example of repeatedly performing the processes of the following step S103 and the following step S104 is given, but the processes of the following step S103 and the following step S104 may also be executed in parallel for each of the N intervals.

[0096] That is, the distribution estimation unit 13 estimates the probability distribution related to the parameters of the interval selection model generated from the segmented data of the t-th interval (step S103). Subsequently, the distribution evaluation unit 14 calculates an evaluation value for evaluating the statistical deviation as a statistical deviation index based on the probability distribution of the parameters of the interval selection model estimated in step S103 (step S104).

[0097] By repeatedly performing such loop process 1, the statistical deviation is calculated for each of the N intervals.

[0098] Then, the interval selection unit 15 selects the intervals belonging to the cluster with relatively small statistical deviation calculated in step S104 (step S105). In addition, in step S105, the intervals belonging to the cluster with relatively large statistical deviation are excluded.

[0099] Subsequently, the output unit 16 outputs the information of the intervals selected in step S105 to the identification unit 17 (step S106). On this basis, the identification unit 17 performs system identification for generating an identification model based on the segmented data corresponding to the intervals output in step S106 (step S107), and ends the process.

[0100] In addition, Figure 7 the processes shown can be appropriately omitted if possible. For example, in a certain interval segmented by the interval segmentation unit 12, if the time series data (especially the input time series data) is substantially constant and has no change, this interval is obviously not suitable for identification. In such a case, the processes of the distribution estimation unit 13 and the distribution evaluation unit 14 can be omitted, and the interval selection unit 15 excludes the interval on the grounds that the time series data has no change. By omitting the processes for obvious intervals in this way, the effect of shortening the processing time can be obtained.

[0101] <An aspect of the effect>

[0102] As described above, for each interval obtained by dividing the time-series data of the input / output of the system to be controlled, the information processing apparatus 10 of the present embodiment estimates the probability distribution of the parameters of the model generated from the segmented data corresponding to the interval, and selects the interval with a small statistical deviation of the probability distribution. Therefore, with the information processing apparatus 10 of the present embodiment, it is possible to automatically extract an interval suitable for system identification having dynamic characteristics. Thus, as one aspect, it is possible to reduce the labor for selecting and deriving an interval suitable for system identification from the collected time-series data, such as the man-hours or time of experts such as engineers.

[0103] <Other Embodiments>

[0104] Incidentally, although the embodiments of the present disclosure have been described so far, various applications are possible. Furthermore, in addition to the embodiments described above, it can be implemented in various different ways.

[0105] <Application Example 1: Use of Low-Order Model>

[0106] As Application Example 1 of the above-described embodiment, the structure of the interval selection model used in the distribution estimation unit 13 can be set to a low-order model, for example, a first-order lag system. Thereby, the amount of calculation is reduced, and furthermore, the processing time is shortened.

[0107] When using a transfer function to represent the model of a first-order lag system, 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, the gain G and the time constant T.

[0108] [Equation 1]

[0109]

[0110] Hereinafter, a case where the number of inputs of the object of system identification is two and the number of outputs is one is taken as an example. However, as disclosed in Non-Patent Document 3, the number of inputs can also be three or more. Here, it is assumed that the k-th input is represented by u k (t), and its Laplace transform is represented by U k (s), the output is represented by y(t), and its Laplace transform is represented by Y(s). The gain and time constant of the transfer function from the k-th input to the output are represented by G k and T k respectively. In this case, the transfer function representation of the interval selection model becomes as shown in formula (2).

[0111] [Equation 2]

[0112]

[0113] From the aspect of estimating the probability distribution of the parameters of the interval selection model using SMC2, the transfer function representation of the interval selection model is transformed into a discrete-time state-space representation. This transformation can be achieved by the following formula (3) and formula (4). In addition, for the sake of simplicity of explanation, an example with a sampling interval of 1 is given here, but the sampling interval can be changed arbitrarily.

[0114] [Equation 3]

[0115]

[0116] [Equation 4]

[0117]

[0118] 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 and 1 column). "v k " refers to the system noise (2 rows and 1 column). "y k " refers to the output (scalar). "w k " refers to the observation noise. "k" refers to the index representing the time. In addition, it is assumed that the system noise and the observation noise each independently follow a normal distribution at each time.

[0119] Under this setting, the distribution estimation unit 13 uses SMC2 to estimate the probability distributions of G1, G2, T1, and T2, which are the parameters of the interval selection model. Other than that, the processing is the same as that described in the above embodiment.

[0120] In interval selection, the statistical deviation of the probability distribution of the parameters rather than the parameters themselves of the interval selection model is obtained. For example, as an interval suitable for system identification, an interval that satisfies the following conditions (a) to (c) can be selected.

[0121] (a) The time series data of the input / output changes moderately

[0122] (b) The causal relationship between the input / output can be confirmed based on the time series data

[0123] (c) There is little external interference in the time series data

[0124] Moreover, when determining the above three conditions (a) to (c), any model that can generally represent the dynamic characteristics between input and output can be considered sufficient. Even for such a model, due to reasons such as no changes in the time series data of input / output, unclear causal relationship between input / output, and many external interferences, a phenomenon where the statistical deviation of the parameters becomes larger can be captured. Because of this perspective, it is possible to think of using the simplest first-order lag system among the models that can represent dynamic characteristics as the model for interval selection. The parameters of the first-order lag system are only two, namely the gain and the time constant, which can shorten the processing time of the distribution estimation unit 13. Therefore, through this Application Example 1, an effect of automatically selecting an interval with a shorter processing time compared to the case of using a higher-order model can be obtained.

[0125] However, when using the first-order lag system as the model for interval selection, sometimes the assumption that the probability distribution of the parameters follows a normal distribution is not satisfied. In this case, a method based on the sequential Monte Carlo method that does not require an assumption about the probability distribution of the parameters can also be used to estimate the probability distribution. The method of the confidence interval using the least squares method assumes that the probability distribution of the parameters follows a normal distribution, so the performance may decrease in this Application Example 1.

[0126] <Application Example 2: Estimating Only the Gain>

[0127] As another Application Example 2 of the above Application Example 1, the part representing the dynamic characteristics of the model for interval selection used in the distribution estimation unit 13 is fixed, and the estimation of the probability distribution focused on the gain is performed. Thereby, a further reduction in the amount of calculation is achieved, and furthermore, a further shortening of the processing time is achieved.

[0128] For example, a model having the following structure of Equation (5) can be used as the model for interval selection. Here, "G" represents the gain, and in addition, "D(s)" represents a transfer function with a gain of 1, presenting a structure where the two are multiplied. D(s) only represents the part of the dynamic characteristics of the object to be system-identified, and the gain is normalized to 1. The distribution estimation unit 13 performs the estimation focusing on the probability distribution of the gain G. D(s) can be set in advance based on knowledge related to the object to be system-identified and is fixed in the estimation using the distribution estimation unit 13 and is not an estimation object. In addition, the part representing the dynamic characteristics of the object to be system-identified corresponding to D(s) can also be other representations as long as it is a part that can represent dynamic characteristics such as state space representation.

[0129] GD(s)…(5)

[0130] This Application Example 2 can be applied to the following use cases, etc.: The dynamic characteristics of the system identification object can be obtained from physical law-based modeling or computer simulation, and only the part equivalent to the gain is modeled using the time series data collected from the actually identified object of the system. In such a use case, only the gain has a large uncertainty. Therefore, by using only the statistical deviation of the probability distribution of the gain, it is possible to determine whether each interval is suitable for system identification. In addition, by fixing the part of the dynamic characteristics, the object of estimating the probability distribution is concentrated on the gain, so that the algorithm for estimating the probability distribution can be simplified or made more concise. The reason is that when estimating the probability distribution of the part of the dynamic characteristics, the past observation values or the history of the state are required. In contrast, if only the gain is estimated, the estimation can be performed even without the past history. Due to the above situation, not only the number of parameters for estimating the probability distribution is reduced, but also the processing amount is reduced. In addition, for example, when the rough time constant is unknown and the dynamic characteristics are unknown, or when the dynamic characteristics change greatly, the application of this Application Example 2 can be prohibited or restricted.

[0131] <Numerical values, etc.>

[0132] The matters described in the above embodiments, such as specific examples of the number of inputs or outputs of the system to be controlled, the types of interval selection models or identification models, the types of algorithms used in the estimation of probability distributions, etc. are only examples and can be changed. In addition, the processing order of the flowcharts described in the embodiments can also be changed within a non-contradictory range.

[0133] <Constituent components>

[0134] Except for the specially described cases, the information including processing procedures, control procedures, specific names, various data or parameters shown in the above text or drawings can be arbitrarily changed. For example, any one or more of the functional parts such as 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 of the information processing device 10 can also be constituted by different devices.

[0135] In addition, each constituent component of each device shown in the figure is a functional concept, and it is not necessarily physically constituted as shown in the figure. That is, the specific form of variance or integration of each device is not limited to the form shown in the figure. That is, all or part of it can be variated and integrated functionally or physically in any unit according to various loads or usage conditions, etc. In addition, each structure can also be a physical structure.

[0136] Furthermore, all or any part of the processing functions performed by each device can be implemented by a central processing unit (CPU) and a program parsed and executed by the CPU, or can be implemented as hardware based on wired logic.

[0137] The constituent components include components that can be easily conceived by those skilled in the art, substantially identical components, and components within the so-called equivalent scope. Furthermore, the above-described embodiments can be appropriately combined within the range where the processing contents do not conflict.

[0138] In addition, the above-mentioned "module (-er suffix, -or suffix)" can be renamed as a unit, component, circuit, etc. For example, a communication module, a control module, and a storage module can be renamed as a communication unit, a control unit, and a storage unit, respectively.

[0139] <Hardware>

[0140] Next, a hardware structure example of the computer described in the above embodiment will be described. Figure 8 It is a diagram showing a hardware structure example. As Figure 8 shown, the information processing device 10 includes a communication device 10a, a storage device 10b, a memory 10c, and a processor 10d. In addition, Figure 8 each of the parts shown can be connected to each other via a bus or the like.

[0141] The communication device 10a is a network interface card or the like. The storage device 10b is a storage device such as a hard disk drive (HDD) or a solid state drive (SSD). For example, the storage device 10b stores a program or a database (DB) that runs the Figure 7 functions shown.

[0142] The processor 10d reads out and executes a program that performs the same processing as the processing unit shown in Figure 1 from the storage device 10b or the like and expands it into the memory 10c, thereby running a process that executes the Figure 1 functions described in

[0143] This kind of process implements the same functions as the processing unit of the information processing device 10. For example, the processor 10d reads out a program having the same functions as the data collection unit 11, the interval division unit 12, the distribution estimation unit 13, the distribution evaluation unit 14, the interval selection unit 15, the output unit 16, the recognition unit 17, etc. from the storage device 10b or the like. Then, the processor 10d executes the following process: the process executes the same processing as the data collection unit 11, the interval division unit 12, the distribution estimation unit 13, the distribution evaluation unit 14, the interval selection unit 15, the output unit 16, the recognition unit 17, etc.

[0144] In this way, the information processing device 10 operates as an information processing device that executes the interval selection method by reading out and executing a program. In addition, the information processing device 10 can also achieve the same functions as the above-described embodiment, application example 1, and application example 2 by reading out the program from the recording medium by the medium reading device and executing the read program. Furthermore, the programs mentioned in the other embodiments are not limited to being executed by the information processing device 10. For example, in the case where the program is executed by other computers or servers, or in the case where they cooperate to execute the program, the interval selection function can be similarly applied.

[0145] The program can be distributed via a network such as the Internet. In addition, the program can be executed by being recorded on an arbitrary recording medium and read out from the recording medium by a computer. For example, the recording medium can be implemented by a hard disk, a flexible disk (FD), a compact disc read-only memory (CD-ROM), a magneto-optical disk (MO), a digital versatile disc (DVD), etc.

[0146] As described above, several embodiments have been described in detail based on the drawings, but these are examples, and each embodiment can be implemented in the form described in the disclosure of the invention and other forms obtained by various deformations and improvements based on the knowledge of those skilled in the art.

Claims

1. An information processing device, characterized in that: include: A collecting unit that collects time series data of input / output related to a system to be controlled; an estimating unit that estimates, for each interval into which the time series data is segmented, a probability distribution related to a parameter of a first model generated based on segmented data corresponding to the interval; A selection unit, selecting an interval with a small statistical deviation of the probability distribution; as well as The output unit outputs information related to the section selected by the selection unit.

2. The information processing device according to claim 1, wherein: The selection unit includes a process of clustering the intervals based on the statistical deviation of the probability distribution and selecting one or more intervals included in a cluster having a relatively small statistical deviation, among clusters obtained as a result of the clustering. 3 . The information processing device according to claim 1 , further comprising an identification unit that generates a second model for identifying the system based on segmented data corresponding to the interval output by the output unit in the time series data collected by the collection unit.

4. The information processing device according to claim 3, wherein: The second model has a different structure than the first model.

5. The information processing device according to claim 1, wherein: The first model has a structure represented by a transfer function of a first-order lag system.

6. The information processing device according to claim 5, wherein: The first model has a model structure in which a gain is multiplied by a dynamic characteristic portion of the system. The estimation unit fixes the dynamic characteristic portion and estimates the probability distribution focusing on the gain.

7. A method for selecting an interval, characterized in that: The following processing is performed by the computer: Collecting time series data of input / output related to the system to be controlled; For each interval into which the time series data is segmented, estimating a probability distribution related to a parameter of a first model generated according to the segmented data corresponding to the interval; Selecting an interval with a small statistical deviation of the probability distribution; as well as Outputs information about the selected interval.

8. A computer-readable storage medium storing an interval selection program, characterized in that: Make the computer execute the following processing: Collecting time series data of input / output related to the system to be controlled; For each interval into which the time series data is segmented, estimating a probability distribution related to a parameter of a first model generated according to the segmented data corresponding to the interval; Selecting an interval with a small statistical deviation of the probability distribution; as well as Outputs information about the selected interval.

Images