JP7556306B2 - Prediction system, information processing device, and information processing program - Google Patents

Description

The present disclosure relates to a prediction system that predicts changes that occur in a controlled object, an information processing device that constitutes the prediction system, and an information processing program for realizing the information processing device.

In various production sites, for one reason or another, changes that are different from normal or expected can occur. If such changes can be predicted in advance and some kind of countermeasure can be taken, it is beneficial for maintaining the performance of production equipment and ensuring product quality.

Regarding such advance predictions, for example, Japanese Patent Laid-Open Publication No. 2009-237832 (Patent Document 1) discloses a method for constructing a variable prediction model that can improve the accuracy of demand forecasts for any period or season. The variable prediction model construction method disclosed in Patent Document 1 employs a process in which, using learning data that has been corrected from accumulated time series data, an appropriate prediction model is constructed for each of multiple learning periods of 7 to 70 days, and the modeling accuracy of each learning period is evaluated to select the optimal learning period and prediction model with the highest prediction accuracy.

JP 2009-237832 A

However, in the variable prediction model construction method disclosed in the above-mentioned Patent Document 1, it is necessary to construct an appropriate prediction model for each of multiple learning periods and then evaluate the modeling accuracy of each learning period, which poses the problem of the large amount of work required to select the optimal learning period and prediction model.

The present disclosure aims to provide a method for more easily evaluating predictive models.

A prediction system based on a certain aspect includes a control calculation unit that executes control calculations for controlling a control target, a prediction value acquisition unit that acquires a prediction value by inputting an actual value consisting of one or more state values among state values that the control calculation unit can refer to into a prediction model, a prediction model generation unit that generates a prediction model based on a tree-structured learning algorithm, and a prediction model evaluation unit that evaluates the prediction model. The prediction model evaluation unit includes an analysis unit that analyzes the characteristics of the tree structure of the prediction model, and an evaluation unit that evaluates the prediction model based on the analysis results of the analysis unit.

With this configuration, it is possible to analyze the characteristics of the tree structure and evaluate the prediction model based on the analysis results, making it possible to easily evaluate the prediction model.

Preferably, the prediction model evaluation unit further includes an evaluation setting unit that sets an evaluation type and a reference value for evaluating the prediction model.

This configuration allows you to set the evaluation type and reference value, making it easy to evaluate the prediction model.

Preferably, the evaluation unit displays the evaluation result of the prediction model on the display unit.
According to this configuration, the evaluation result is displayed on the display unit, so that the prediction model can be easily evaluated.

Preferably, the prediction model generation unit updates the prediction model based on the learning samples corresponding to the evaluation results of the prediction model.

With this configuration, the prediction model is updated based on the learning samples that correspond to the evaluation results of the prediction model, thereby improving the accuracy of the prediction model.

An information processing device based on a certain aspect is an information processing device connected to a control device, and the control device includes a control calculation unit that executes control calculations for controlling a control target, and a predicted value acquisition unit that acquires a predicted value by inputting an actual value consisting of one or more state values among state values that the control calculation unit can refer to into a prediction model. The information processing device includes a prediction model generation unit that generates a prediction model based on a tree-structured learning algorithm, and a prediction model evaluation unit that evaluates the prediction model. The prediction model evaluation unit includes an analysis unit that analyzes characteristics of the tree structure of the prediction model, and an evaluation unit that evaluates the prediction model based on the analysis results of the analysis unit.

With this configuration, it is possible to analyze the characteristics of the tree structure and evaluate the prediction model based on the analysis results, making it possible to easily evaluate the prediction model.

An information processing program based on a certain aspect is an information processing program executed by a computer connected to a control device, and the control device includes a control calculation unit that executes control calculations for controlling a control target, and a predicted value acquisition unit that acquires a predicted value by inputting an actual value consisting of one or more state values among state values that the control calculation unit can refer to into a prediction model. The information processing program causes the computer to execute a step of generating a prediction model based on a tree-structured learning algorithm, and a step of evaluating the prediction model. The step of evaluating the prediction model includes a step of analyzing characteristics of the tree structure of the prediction model, and a step of evaluating the prediction model based on the analysis results.

With this configuration, it is possible to analyze the characteristics of the tree structure and evaluate the prediction model based on the analysis results, making it possible to easily evaluate the prediction model.

In this disclosure, "part" and "device" do not simply mean physical means, but also include configurations in which the functions of the "part" and "device" are realized by software. Furthermore, the functions of one "part" and "device" may be realized by two or more physical means or devices, or the functions of two or more "parts" and "devices" may be realized by one physical means or device. Furthermore, "part" and "device" are concepts that can be rephrased as, for example, "means" and "system."

The prediction system, information processing device, and information processing program disclosed herein can more easily evaluate a prediction model.

1 is a schematic diagram illustrating an example of the overall configuration of a prediction system 1 based on an embodiment. FIG. 2 is a schematic diagram showing an application example of the prediction system 1 based on the embodiment. FIG. 2 is a diagram illustrating a control system of the prediction system 1 based on the embodiment. FIG. 2 is a schematic diagram illustrating an example of control based on a prediction result of a prediction control system by a prediction system 1 based on an embodiment. 13 is a flowchart showing the processing steps of a process for generating a prediction model 140 using the prediction system 1 based on an embodiment. FIG. 1 is a diagram illustrating a learning method of a prediction model 140 based on an embodiment. 2 is a block diagram showing an example of a hardware configuration of a control device 100 constituting a prediction system 1 based on an embodiment. FIG. 2 is a block diagram showing an example of a hardware configuration of a support device 200 constituting the prediction system 1 based on the embodiment. FIG. 2 is a block diagram showing an example of the software configuration of a control device 100 and a support device 200 that constitute a prediction system 1 based on an embodiment. FIG. 10 is a block diagram showing an overview of functional modules included in the analysis program 226 shown in FIG. 9. FIG. 13 is a diagram illustrating evaluation of a prediction model based on an embodiment. FIG. 1 is a flow diagram illustrating evaluation of a predictive model based on an embodiment. FIG. 2 is a diagram illustrating a specific example of a tree structure of a prediction model based on an embodiment. FIG. 11 is a diagram illustrating a specific example of a distribution diagram of a prediction model based on an embodiment. FIG. 13 is a diagram illustrating a model evaluation screen based on the embodiment. FIG. 13 is a diagram illustrating another model evaluation screen based on the embodiment. FIG. 1 is a diagram illustrating an application example of a prediction system 1 based on an embodiment. FIG. 11 is a diagram illustrating additional learning of a prediction model based on an embodiment. FIG. 1 is a schematic diagram showing a general configuration of an anomaly detection system 1A according to a modified example of the present embodiment. 1A and 1B are diagrams illustrating a specific example of an anomaly detection system 1A according to a modified example of the present embodiment. FIG. 11 is a conceptual diagram of an evaluation of an anomaly detection system 1A according to a modified example of the present embodiment. 13 is a diagram illustrating additional learning of an anomaly detection model according to a modified example of the present embodiment. FIG.

The embodiment will be described in detail with reference to the drawings. Note that the same or equivalent parts in the drawings will be given the same reference numerals and their description will not be repeated.

<A. Application Examples>
First, an example of a situation to which this embodiment is applied will be described.

The main aspects of a control system having a predictive function based on an embodiment will be described. In the following description, the focus will be primarily on the predictive function possessed by the control system, and therefore the entire control system will also be referred to as the "predictive system."

FIG. 1 is a schematic diagram showing an example of the overall configuration of a prediction system 1 based on an embodiment. Referring to FIG. 1, the prediction system 1 based on an embodiment includes, as main components, a control device 100 that controls a control target, and a support device 200 that is connected to the control device 100.

The control device 100 may be embodied as a type of computer, such as a programmable controller (PLC). The control device 100 may be connected to a field device group 10 via a field bus 2, and may be connected to one or more display devices 400 via a field bus 4. Furthermore, the control device 100 may be connected to an upper server 300 via an upper network 6. Note that the upper server 300 and the display device 400 are optional components and are not required components of the prediction system 1.

The control device 100 has a control logic (hereinafter also referred to as a "PLC engine") that executes various calculations to control equipment and machinery. In addition to the PLC engine, the control device 100 has a collection function that collects data (hereinafter also referred to as "input data") that is measured by the field device group 10 and transferred to the control device 100. Furthermore, the control device 100 also has a prediction function that predicts future changes over time based on the collected input data.

Specifically, a time series database (hereinafter also referred to as "TSDB (Time Series Data Base)") 130 implemented in the control device 100 provides a collection function, and a prediction model 140 implemented in the control device 100 provides a monitoring function. Details of the TSDB 130 and the prediction model 140 will be described later.

It is preferable that field bus 2 and field bus 4 adopt an industrial communication protocol. Known examples of such communication protocols include EtherCAT (registered trademark), EtherNet/IP (registered trademark), DeviceNet (registered trademark), and CompoNet (registered trademark).

The field device group 10 includes devices that collect input data from the controlled object or the manufacturing equipment or production line related to the control (hereinafter collectively referred to as "field"). Assumed devices for collecting such input data include input relays and various sensors. The field device group 10 further includes devices that perform some kind of action on the field based on commands (hereinafter also referred to as "output data") generated by the control device 100. Assumed devices for performing some kind of action on such a field include output relays, contactors, servo drivers and servo motors, and any other actuators. These field device group 10 exchange data including input data and output data with the control device 100 via the field bus 2.

In the example configuration shown in FIG. 1, the field device group 10 includes a remote I/O (Input/Output) device 12, a relay group 14, an image sensor 18 and a camera 20, a servo driver 22 and a servo motor 24.

The remote I/O device 12 includes a communication section that communicates via the field bus 2, and an input/output section (hereinafter also referred to as an "I/O unit") for collecting input data and outputting output data. Input data and output data are exchanged between the control device 100 and the field via such an I/O unit. Figure 1 shows an example in which digital signals are exchanged as input data and output data via a group of relays 14.

The I/O unit may be directly connected to the field bus 2. FIG. 1 shows an example in which the I/O unit 16 is directly connected to the field bus 2.

The image sensor 18 performs image measurement processing such as pattern matching on the image data captured by the camera 20 and transmits the processing results to the control device 100.

The servo driver 22 drives the servo motor 24 according to output data (e.g., position commands, etc.) from the control device 100.

As described above, data is exchanged between the control device 100 and the field device group 10 via the field bus 2, and this exchanged data is updated at very short intervals of the order of several hundred microseconds to several tens of milliseconds. Note that this process of updating exchanged data is sometimes referred to as "I/O refresh processing."

The display device 400, which is connected to the control device 100 via the field bus 4, receives operations from the user, transmits commands to the control device 100 in response to the user operations, and graphically displays the results of calculations performed by the control device 100.

The upper server 300 is connected to the control device 100 via the upper network 6, and exchanges necessary data with the control device 100. A general-purpose protocol such as Ethernet (registered trademark) may be implemented in the upper network 6.

The support device 200 is an information processing device (an example of a computer) that supports the preparations required for the control device 100 to control the control target. Specifically, the support device 200 provides a development environment (program creation and editing tools, parsers, compilers, etc.) for user programs executed by the control device 100, a setting environment for setting parameters (configurations) for the control device 100 and various devices connected to the control device 100, a function for sending the generated user program to the control device 100, a function for modifying and changing the user program executed on the control device 100 online, etc.

Furthermore, the support device 200 based on the embodiment has a function for supporting the generation and optimization of the prediction model 140 implemented in the control device 100. In other words, the support device 200 has a prediction model generation unit that predetermines the prediction model 140. Details of these functions will be described later.

Next, an application example of the control device 100 included in the prediction system 1 will be described.
2 is a schematic diagram showing an application example of the prediction system 1 based on the embodiment. In FIG. 2, an example of a production facility including a press machine 30 is shown.

Referring to FIG. 2, the press machine 30 receives the workpiece 31 and places it on a support table 34 provided on a base 33. The workpiece 31 is then compressed by a pressing plate 35 provided at the tip of a drive shaft 36 driven by a motor 37 to produce an intermediate product 32.

In the press machine 30, it is assumed that defects may occur in the intermediate product 32 due to unexpected fluctuations in factors. Therefore, whether or not there is a defect in the intermediate product 32 is determined by inspection using an inspection machine located downstream of the press machine 30, or by visual inspection or random inspection by an inspector. If it is determined that there is a defect, the target values, etc. will be adjusted each time.

As such, in a normal manufacturing process, in order to maintain and improve the yield rate of the intermediate product 32, it is necessary to adjust the target value each time, but even if it is designed in advance from various perspectives, it is difficult to respond to all factor fluctuations.

In response to this, by using the prediction model 140 based on the embodiment to predict the state of the intermediate product 32 (i.e., the quality after processing), the control by the control device 100 can be corrected before a defect actually occurs. By utilizing such advance prediction of unwanted occurrence, it is possible to reduce the labor required for each adjustment of target values, etc., and prevent defects from occurring in the intermediate product 32.

Figure 3 is a diagram illustrating a control system by the prediction system 1 based on an embodiment. Referring to Figure 3, in this example, the controlled object is a press machine 30. The setting values of the device are input to the existing control system to generate a target value. An operation amount for the controlled object is set by feedback control based on the control amount and the target value of the press machine 30. In addition, a prediction control system is provided for the existing control system. Specifically, a predictor is provided that calculates a target variable for an explanatory variable. One example of a predictor is a prediction model 140. A correction value is calculated by a corrector based on the target value of the quality characteristic of the press machine 30 and the objective variable. The correction value is added to the operation amount.

Figure 4 is a schematic diagram showing an example of control based on the prediction results of a predictive control system by prediction system 1 based on an embodiment.

Figure 4 (A) shows the planned value (command) of the pressing position of the press machine at a certain point in time and the actual pressing position (actual value) of the press machine 30. The target value indicates the desired thickness of the intermediate product 32 after processing.

Referring to FIG. 4(B), at a certain point in time, the future pressing position (predicted value) of the press machine 30 is calculated based on the information up to that point (which may include actual values), and the operation amount for the press machine 30 is corrected by a correction value based on the calculated predicted value.

In the press machine 30 shown in Figures 2 to 4, the pressing position of the press machine 30 is predicted and the control amount is corrected based on the prediction result, so that it is not necessary for the operator to adjust the target value each time, for example, in response to variations in the hardness of the workpiece 31. As a result, it is possible to suppress the occurrence of defective products due to unexpected fluctuations in factors, and it is possible to stabilize the quality even if there is some variation in the workpiece 31.

The data used for prediction (actual values or observed values) and the data to be predicted may be partially or completely the same, or may be completely different.

The prediction system 1 based on the embodiment provides a function for appropriately generating the prediction model 140. The function for appropriately generating the prediction model 140 may typically be implemented in the support device 200.

B. Overview of Predictive Model Generation and Operation
Next, an overview of the generation and operation of the prediction model 140 using the prediction system 1 based on the embodiment will be described.

Figure 5 is a flowchart showing the processing steps of the generation process of the prediction model 140 using the prediction system 1 based on the embodiment. Each step shown in Figure 5 is typically realized by the processor 202 of the support device 200 executing a program (such as the analysis program 226 and the OS 228).

Referring to FIG. 5, the support device 200 acquires time series data of actual values stored in the TSDB 130 (step S1). Next, the support device 200 accepts the setting of the prediction target interval from the acquired time series data of actual values (step S2).

The support device 200 selects learning samples to be used in generating a prediction model for predicting changes in the prediction target interval set in step S2 (step S3). In step S3, which of the multiple types of data will be used for learning is selected.

The support device 200 performs machine learning based on the selected learning samples (step S4).

The support device 200 generates a prediction model 140 by machine learning (step S6). In this example, the prediction model 140 is configured by a decision tree. The decision tree is configured to output a pressing position (predicted value) for a feature amount at a certain point in time. The learning method for this decision tree may be CLS (Concept Learning System), ID3 (Iterative Dichotomiser 3), C4.5, etc.

In this specification, "sample" refers to a data string of a predetermined length used as training data for a predicted value to be output from the prediction model 140. The "sample" basically uses the time series data (raw data) to be predicted, but when the prediction target is a feature extracted from the time series data, the feature may be used. The term "sample" focuses on the processing unit when processing multiple data, and does not particularly limit the content of the data contained therein.

In this specification, data referenced to calculate or determine any predicted value is also referred to as "explanatory variables." Any predicted value is calculated or determined using one or more "explanatory variables." Therefore, the training samples will be associated in some way with data that may be candidates for "explanatory variables."

In this specification, "feature amount" is a term that encompasses information contained in the time series data to be processed, and may include, for example, the maximum value, minimum value, median value, average value, standard deviation, variance, etc., of the target time series data. Note that "feature amount" may also include the target time series data itself.

Figure 6 is a diagram illustrating a learning method for a prediction model 140 based on an embodiment. As shown in Figure 6 (A), time series data stored in TSDB 130 is shown. A prediction interval is set for the time series data, and a prediction value, which is the objective variable to be predicted, is estimated for feature quantities at two points in the prediction interval, time t and t-1. As shown in Figure 6 (B), a case is shown in which the learning process is performed using the time series data with a shift of a unit time.

By setting the prediction model 140 generated by the above processing procedure in the control device 100, operation as shown in Figures 2 to 4 becomes possible.

C. Hardware Configuration Example
Next, a hardware configuration example of main devices constituting the prediction system 1 based on the embodiment will be described.

(c1: Example of hardware configuration of control device 100)
Fig. 7 is a block diagram showing an example of a hardware configuration of a control device 100 constituting the prediction system 1 based on the embodiment. Referring to Fig. 6, the control device 100 includes a processor 102 such as a CPU (Central Processing Unit) or an MPU (Micro-Processing Unit), a chipset 104, a main memory device 106, a secondary memory device 108, a host network controller 110, a USB (Universal Serial Bus) controller 112, a memory card interface 114, an internal bus controller 122, field bus controllers 118 and 120, and I/O units 124-1, 124-2, ....

The processor 102 reads various programs stored in the secondary storage device 108, deploys them in the main storage device 106, and executes them to realize the PLC engine 150 and the prediction model 140. The chipset 104 controls data transmission between the processor 102 and each component.

In addition to the system program for implementing the PLC engine 150, the secondary storage device 108 stores a user program that is executed using the PLC engine 150. Furthermore, the secondary storage device 108 also stores a program for implementing the prediction model 140.

The upper network controller 110 controls data exchange with another device via the upper network 6. The USB controller 112 controls data exchange with the support device 200 via a USB connection.

The memory card interface 114 is configured to allow the memory card 116 to be attached and detached, and is capable of writing data to the memory card 116 and reading various data (such as user programs and trace data) from the memory card 116.

The internal bus controller 122 is an interface that exchanges data with the I/O units 124-1, 124-2, etc. mounted on the control device 100.

Fieldbus controller 118 controls data exchange with other devices via fieldbus 2. Similarly, fieldbus controller 120 controls data exchange with other devices via fieldbus 4.

Although FIG. 7 shows an example of a configuration in which the processor 102 executes a program to provide the necessary functions, some or all of these provided functions may be implemented using a dedicated hardware circuit (e.g., an ASIC (Application Specific Integrated Circuit) or an FPGA (Field-Programmable Gate Array)). Alternatively, the main part of the control device 100 may be realized using hardware based on a general-purpose architecture (e.g., an industrial PC based on a general-purpose PC). In this case, virtualization technology may be used to run multiple OSs (operating systems) with different uses in parallel, and necessary applications may be executed on each OS.

(c2: Hardware configuration example of support device 200)
The support device 200 according to the embodiment is realized, for example, by executing a program using hardware based on a general-purpose architecture (for example, a general-purpose personal computer).

FIG. 8 is a block diagram showing an example of the hardware configuration of a support device 200 constituting a prediction system 1 based on an embodiment. Referring to FIG. 8, the support device 200 includes a processor 202 such as a CPU or MPU, an optical drive 204, a main memory device 206, a secondary memory device 208, a USB controller 212, a higher-level network controller 214, an input unit 216, and a display unit 218. These components are connected via a bus 220.

The processor 202 reads various programs stored in the secondary storage device 208, deploys them in the main storage device 206, and executes them to realize various processes including the model generation process described below.

The secondary storage device 208 is composed of, for example, a hard disk drive (HDD) or a flash solid state drive (SSD). The secondary storage device 208 typically stores a development program 222 for creating user programs to be executed in the support device 200, debugging the created user programs, defining the system configuration, setting various parameters, etc., a PLC interface program 224 for exchanging data related to the prediction function with the control device 100, an analysis program 226 for generating the prediction model 140, etc., and an OS 228. The secondary storage device 208 may store necessary programs other than the programs shown in FIG. 8.

The support device 200 has an optical drive 204, and the program stored in a recording medium 205 (e.g., an optical recording medium such as a DVD (Digital Versatile Disc)) that non-transiently stores a computer-readable program is read and installed in a secondary storage device 208, etc.

The various programs executed by the support device 200 may be installed via a computer-readable recording medium 205, or may be installed by downloading from any server on the network. In addition, the functions provided by the support device 200 based on the embodiment may be realized by using some of the modules provided by the OS.

The USB controller 212 controls data exchange with the control device 100 via a USB connection. The upper network controller 214 controls data exchange with another device via any network.

The input unit 216 is made up of a keyboard, mouse, etc., and accepts user operations. The display unit 218 is made up of a display, various indicators, a printer, etc., and outputs processing results from the processor 202, etc.

Figure 8 shows an example of a configuration in which the necessary functions are provided by the processor 202 executing a program, but some or all of these provided functions may be implemented using dedicated hardware circuits (e.g., ASICs or FPGAs, etc.).

<D. Software configuration example/functional configuration example>
Next, examples of the software configuration and the functional configuration of the control device 100 and the support device 200 that configure the prediction system 1 based on the embodiment will be described.

Figure 9 is a block diagram showing an example of the software configuration of the control device 100 and the support device 200 that constitute the prediction system 1 based on the embodiment. Referring to Figure 9, the control device 100 includes, as its main functional components, a TSDB 130 and a prediction model 140 in addition to a PLC engine 150.

The PLC engine 150 sequentially interprets the user program 154 and executes the specified control calculation. The PLC engine 150 manages state values collected from the field in the form of variables 152, which are updated at a predetermined interval. The PLC engine 150 may be realized by the processor 102 of the control device 100 executing a system program.

In this specification, "state value" includes input values collected from the field, command values output to the field, and system state values and internal values managed within the control device 100. In the control device 100 based on the embodiment, "state values" are referred to in the form of "variables," so for convenience in the following description, the term "variable" is used to include "state values." Note that the technical scope of the present disclosure is not limited to configurations in which "state values" are referred to in the form of "variables."

The user program 154 includes predicted value acquisition code 156, error evaluation code 158, additional learning code 160, TSDB write code 162, and control operation code 164.

The predicted value acquisition code 156 realizes a predicted value acquisition unit that acquires a predicted value by inputting actual values consisting of one or more state values among the state values that the control operation code 164 can refer to into the prediction model 140. More specifically, the predicted value acquisition code 156 includes an instruction to acquire a predicted value by acquiring the necessary actual values managed as variables 152 and inputting them into the prediction model 140.

The error evaluation code 158 includes instructions to evaluate the error between the predicted value obtained by the predicted value acquisition code 156 and the target value.

The additional learning code 160 includes instructions to additionally learn the prediction model 140, as necessary, depending on the error evaluated by the error evaluation code 158.

TSDB write code 162 obtains predetermined variables from among the variables managed as variables 152 and writes them to memory area 132 of TSDB 130.

The control operation code 164 realizes a control operation unit that executes a control operation for controlling the control object. More specifically, the control operation code 164 executes a control operation for controlling the control object, and corrects the target value used in the control operation as necessary according to the error evaluated by the error evaluation code 158.

TSDB130 has an export module 134 that exports data written to memory area 132 to a support device 200, etc., as necessary.

The predictive model 140 has a reference trajectory 144 .
On the other hand, a development program 222 and an analysis program 226 are installed in the support device 200 .

The development program 222 generates the user program 154 in accordance with user operations and transfers it to the control device 100. The development program 222 also has the function of appropriately modifying the contents of the control operation code 164.

The analysis program 226 corresponds to an information processing program for realizing a prediction model generation unit that predetermines the prediction model 140. More specifically, the analysis program 226 supports the generation of the prediction model 140, and includes a model generation module 2262 and an evaluation module 2264.

The model generation module 2262 realizes the functions required for the process of generating the predictive model 140.

The evaluation module 2264 evaluates the performance of the target predictive model 140 .
Fig. 10 is a block diagram showing an outline of functional modules included in analysis program 226 shown in Fig. 9. Referring to Fig. 10, analysis program 226 of support device 200 includes, as main functional components, a user interface 230, an input/output management module 236, a screen display module 238, a graph library 240, an analysis module 242, and an analysis library 244.

The user interface 230 accepts settings from the user and executes overall processing for providing various information to the user. In a specific implementation form, the user interface 230 has a script engine 232, reads a setting file 234 that includes a script that describes the necessary processing, and executes the set processing.

The input/output management module 236 includes a file input function that reads data from a specified file, a stream input function that receives a data stream, and a file output function that outputs a file containing the generated data.

The screen display module 238 includes a function for generating a model evaluation screen based on the input prediction model, and a function for displaying the evaluation results in response to a user operation. It may also be configured to execute necessary processing by referring to the graph library 240.

The analysis module 242 is a module that realizes the main processing of the analysis program 226, and has a model generation function. Each function included in the analysis module 242 is realized by referring to the analysis library 244.

The analysis library 244 includes libraries for each function included in the analysis module 242 to execute processing. More specifically, the analysis library 244 may have a statistics function, a decision tree function, a time series regression function, a grid search function, a clustering function, an inference speed evaluation function, an accuracy evaluation function, and an anomaly detection function.

E. Evaluation of predictive models
11 is a diagram illustrating the evaluation of a prediction model based on an embodiment. Referring to FIG. 11(A) and (B), two types of decision tree structure prediction models are shown.

The computational process of the prediction model is a search process that follows links from the root node to the leaf nodes. Specifically, it determines whether the input data (explanatory variable X1 or X2) satisfies the branching condition set in the root data, and based on the result of this determination, the search proceeds to the next applicable node.

When analyzing the decision tree structure of the prediction model in Figure 11 (A), it is found that there is a leaf node with a shallow tree depth of "1". Also, when comparing the areas of each region in the distribution map, it is shown that there are cases where relatively large regions exist. Therefore, the prediction model is biased due to the large variance.

When analyzing the decision tree structure of the prediction model in Figure 11 (B), all leaf nodes have a tree depth of "3". Also, when comparing the areas of each region in the distribution map, it is shown that there are no relatively large regions. Therefore, the variance is small, making this a prediction model with little bias.

Therefore, when comparing the prediction model in FIG. 11(A) with the prediction model in FIG. 11(B), it is possible to evaluate that the prediction model in FIG. 11(A) has poor performance and the prediction model in FIG. 11(B) has good performance.

In this embodiment, to evaluate a prediction model of a decision tree structure, the decision tree structure is analyzed to calculate characteristic values of leaf nodes. Specifically, the tree depth and the area of a distribution map corresponding to the leaf node are calculated as characteristic values of the leaf node. The prediction model is evaluated based on whether the calculated characteristic values satisfy a threshold value. The area of each region of the distribution map is the area normalized on each axis.

Fig. 12 is a flow diagram for explaining the evaluation of a prediction model based on an embodiment. Referring to Fig. 12, the support device 200 acquires a prediction model (step S10). Next, the support device 200 analyzes the decision tree structure of the acquired prediction model (step S12). Specifically, the support device 200 calculates the characteristic value of a leaf node of the prediction model. More specifically, the depth of the tree and the area of a distribution map associated with the leaf node are calculated as the characteristic value of the leaf node.

Next, the support device 200 displays a model evaluation screen (step S14).
Then, the process ends (END).

Figure 13 is a diagram illustrating a specific example of a tree structure of a prediction model based on an embodiment. With reference to Figure 13, the tree structure of the prediction model is shown. Specifically, the prediction model estimates output data (objective variables Y0 to Y4) for input data (explanatory variables x1 or x2).

At the root node R0, it is determined whether X1<6. If X1<6 is satisfied at the root node R0, proceed to the leaf node R1. The leaf node R1 corresponds to the objective variable Y0. On the other hand, if 6≦X1 is satisfied at the root node R0, proceed to the intermediate node R5. At the intermediate node R5, it is determined whether X2<10 is satisfied. If X2<10 is satisfied at the intermediate node R5, proceed to the intermediate node R3. If 10≦X2 is satisfied at the intermediate node R5, proceed to the intermediate node R8.

At intermediate node R3, it is determined whether X1<4. If X1<4 is satisfied at intermediate node R3, proceed to leaf node R2. Leaf node R2 corresponds to objective variable Y1. If 4≦X1 is satisfied at intermediate node R3, proceed to leaf node R4. Leaf node R4 corresponds to objective variable Y2.

At intermediate node R8, it is determined whether X1<3. If X1<3 is satisfied at intermediate node R8, the process proceeds to leaf node R6. Leaf node R6 corresponds to objective variable Y3. If 3≦X1 is satisfied at intermediate node R8, the process proceeds to leaf node R7. Leaf node R7 corresponds to objective variable Y4.

Figure 14 is a diagram illustrating a specific example of a distribution diagram of a prediction model based on an embodiment. Referring to Figure 14, the prediction model estimates output data (objective variables Y0 to Y4) for input data (horizontal axis x1, vertical axis x2).

As shown in the distribution diagram, the normalized area of objective variable Y0 is 0.5. The normalized area of objective variable Y1 is 0.16. The normalized area of objective variable Y2 is 0.08. The normalized area of objective variable Y3 is 0.13. The normalized area of objective variable Y4 is 0.13.

FIG. 15 is a diagram illustrating a model evaluation screen based on an embodiment. Referring to FIG. 15, a model evaluation screen 250 is shown. The model evaluation screen is provided with an input field 260 for an input model file of a predictive model. A case is shown in which a user selects file F in which predictive model data is stored in input field 260. By selecting an arbitrary file, it is possible to analyze the characteristics of the tree structure of a predictive model.

The model evaluation screen 250 is provided with a selection item 261 for selecting a judgment mode. As an example, it is possible to select "tree depth" and "normalized area" as selection items. However, this is not limited to this, and it is of course possible to increase the selection items according to various judgment methods. In this example, a case where "tree depth" is selected is shown.

The model evaluation screen 250 is provided with an input field 262 for setting a set threshold. The input field may be switched depending on the judgment mode. As an example, a case is shown in which "2" is set as the set threshold for "tree depth."

The model evaluation screen 250 has tabs 264 and 270 that allow the selection of a "Tree View" and a distribution diagram, respectively, as the analysis results of the tree structure characteristics of the predictive model. An example is shown in which tab 264 is selected.

In this example, a tree structure similar to that described in FIG. 13 is displayed on screen 266. By checking the tree structure, it is possible to easily grasp the characteristics of the prediction model.

Specifically, it is shown that the minimum tree depth is "1", which is below the set threshold of "2". In addition, the tree structure makes it easy to understand the degree of variation.

The model evaluation screen 250 displays a message 268 as the analysis result, stating, "Data in the range 10<x1<20 will contribute to improving the accuracy of the model." By adding data in this range, it is possible to improve the accuracy of the prediction model, as explained in FIG. 11 as an example.

FIG. 16 is a diagram illustrating another model evaluation screen based on an embodiment. Referring to FIG. 16, a model evaluation screen 250 is shown. The model evaluation screen is provided with an input field 260 for an input model file of a predictive model. A case is shown in which a user selects file F in which predictive model data is stored in input field 260. By selecting an arbitrary file, it is possible to analyze the characteristics of the tree structure of the predictive model.

The model evaluation screen 250 is provided with a selection item 261 for selecting a judgment mode. As an example, it is possible to select "tree depth" and "normalized area" as selection items. However, this is not limited to this, and it is of course possible to increase the number of selection items according to various judgment methods. In this example, a case where "normalized area" is selected is shown.

The model evaluation screen 250 is provided with an input field 262 for setting a set threshold. The input field may be switched depending on the judgment mode. As an example, a case is shown in which "0.3" is set as the set threshold for "normalized area."

The model evaluation screen 250 has tabs 264 and 270 that allow the selection of a "Tree View" and a distribution diagram, respectively, as the analysis results of the tree structure characteristics of the predictive model. An example is shown in which tab 270 is selected.

In this example, a distribution diagram similar to that described in FIG. 14 is displayed on screen 266. By checking the distribution diagram, it is possible to easily understand the characteristics of the prediction model.

Specifically, it is shown that the maximum normalized area is "0.5," which is greater than the set threshold of "0.3." In addition, the distribution diagram makes it easy to grasp the degree of variation.

The model evaluation screen 250 displays a message 268 as the analysis result, stating, "Data in the range 10<x1<20 will contribute to improving the accuracy of the model." By adding data in this range, it is possible to improve the accuracy of the prediction model, as explained in FIG. 11 as an example.

In this example, a prediction model that predicts a dependent variable for two explanatory variables has been described, but the same can be applied to a prediction model that predicts a dependent variable for more than two explanatory variables (n≧2).

< F . Specific Examples>
17A and 17B are diagrams for explaining application examples of the prediction system 1 based on the embodiment. Referring to Fig. 17A, a diagram for explaining a pressing result based on the variation of the workpiece material properties of the press machine 30 is shown. Specifically, the pressing position of the press machine 30 is shown to be different depending on whether the workpiece is hard or soft as the workpiece material properties.

Referring to FIG. 17(B), the prediction system 1 calculates the pressing position (predicted value) using the prediction model 140, and corrects the operation amount for the press machine 30 based on the difference between the predicted value and the target position trajectory. As a result, it is possible to suppress the occurrence of defective products due to unexpected fluctuations in factors, and it is possible to stabilize the quality even if there is some variation in the workpiece 31.

In this regard, by improving the accuracy of the prediction model 140, it is possible to achieve greater stability in quality.

Figure 18 is a diagram for explaining additional learning of a prediction model based on an embodiment. Referring to Figure 18 (A), the press results based on the variability of the workpiece material properties of the press machine 30 before additional learning and the tree structure of the prediction model based on the press results are shown. As described above, for example, as a result of analyzing the tree structure of the prediction model, the depth of the tree is shallow because there is little data passing through point "a1". Therefore, by searching for data passing through point "a1", it is possible to further improve the accuracy of the prediction model. The search for data passing through point "a1" may be obtained from data stored in a past database, or may be obtained when the actual machine is operated multiple times.

Referring to FIG. 18(B), a case is shown in which another data passing through point "a1" is acquired. By additionally learning this data, the tree structure of the prediction model changes. In this example, it is possible to suppress the variation in the tree depth and further improve the accuracy of the prediction model.

< G . Modifications>
In the above explanation, the prediction system 1 that predicts changes over time has been described, but the present invention can also be applied to an anomaly detection system that detects anomalies occurring in controlled objects, etc.

Fig. 19 is a schematic diagram showing the general configuration of an anomaly detection system 1A according to a modified example of this embodiment. Referring to Fig. 19, the anomaly detection system 1A selects learning samples from raw data 40 acquired from a control target (sample selection 42), and generates an anomaly detection model 44 based on the selected learning samples. Then, an anomaly detection operation 46 is performed using the generated anomaly detection model 44.

The anomaly detection model 44 is intended to detect when the controlled object is in a state different from the normal state, and uses raw data (time series data) collected from the controlled object to generate an anomaly detection model 44 that fits the collected raw data. When unusual raw data is input to the anomaly detection model 44, a value indicating an unusual state is output, making it possible to detect the occurrence of some kind of anomaly in the controlled object. For the learning samples used in such anomaly detection model 44, it is preferable to use raw data with different change patterns, as in step S3 above.

Figure 20 is a diagram illustrating a specific example of an anomaly detection system 1A according to a modified example of this embodiment.

Referring to Figure 20 (A), the waveform patterns of normal, abnormal A, and abnormal B are shown as raw data.

Figure 20 (B) Normal, abnormal A, and abnormal B waveform patterns are classified using the average value and variance value of the waveform as waveform features. Specifically, the vertical axis F1 is the average value of the waveform, and the horizontal axis F2 is the variance value of the waveform.

Figures 20(C) and (D) are diagrams explaining the anomaly detection model 44 when the above classification patterns are selected as samples. The analysis module 242 selects waveform patterns as samples in the same manner as described above to generate the anomaly detection model 44.

Figure 20 (C) shows a distribution diagram of normal, abnormality A, and abnormality B in the anomaly detection model 44. Figure 20 (D) shows the tree structure of the anomaly detection model 44.

The anomaly detection model 44 estimates output data (objective variables "normal", "abnormal A", "abnormal B") for the input data (explanatory variables F1 or F2).

At the root node R10, it is determined whether F2<a. If F2<a is satisfied at the root node R10, the process proceeds to the leaf node R11. The leaf node R11 corresponds to the objective variable "normal". On the other hand, if a≦F2 is satisfied at the root node R10, the process proceeds to the intermediate node R15. At the intermediate node R15, it is determined whether F1<d is satisfied. If F1<d is satisfied at the intermediate node R15, the process proceeds to the intermediate node R13. If d≦F1 is satisfied at the intermediate node R15, the process proceeds to the intermediate node R18.

At intermediate node R13, it is determined whether F2<b. If F2<b is satisfied at intermediate node R13, the process proceeds to leaf node R12. Leaf node R12 corresponds to the objective variable "abnormality B". If b≦F2 is satisfied at intermediate node R13, the process proceeds to leaf node R14. Leaf node R14 corresponds to the objective variable "abnormality A".

At intermediate node R18, it is determined whether F2<c. If F2<c is satisfied at intermediate node R18, the process proceeds to leaf node R16. Leaf node R16 corresponds to the objective variable "abnormality A". If c≦F2 is satisfied at intermediate node R18, the process proceeds to leaf node R17. Leaf node R17 corresponds to the objective variable "abnormality B".

Figure 21 is a conceptual diagram of an evaluation of anomaly detection system 1A according to a modified example of this embodiment. When the decision tree structure of the anomaly detection model 44 in Figures 20 (C) and (D) is analyzed with reference to Figure 21 (A), it can be seen that the tree depth of the "normal" leaf nodes is shallow. In addition, when the distribution diagram is checked, it can be seen that the normalized area is large.

Therefore, by displaying the model evaluation screen according to the above method, it is possible to easily understand the characteristics of the anomaly detection model.

It is also easy to understand what data is needed to improve the accuracy of the anomaly detection model 44. Specifically, the variance value of the waveform F2 < a.

Referring to FIG. 21(B), data on a newly-occurring "abnormality A" is shown. By additionally learning this data, it is possible to improve the accuracy of the anomaly detection model 44.

Figure 22 is a diagram explaining additional learning of an anomaly detection model according to a modified example of this embodiment. Referring to Figures 22 (A) and (B), a case is shown in which anomaly detection model 44 is updated by additional learning of data on a newly occurring "anomaly A."

Figure 22 (A) is a distribution diagram of the updated anomaly detection model 44. Figure 22 (B) is a tree structure of the updated anomaly detection model.

The above describes the case of generating a predictive model based on a tree-structured learning algorithm, such as a decision tree, but it is not limited to decision trees and can be applied to predictive models using other learning algorithms as long as they are tree-structured.

< H . Notes>
The present embodiment as described above includes the following technical idea.

[Configuration 1]
The prediction system (1) includes a control calculation unit (164) that executes control calculation for controlling a control target, a prediction value acquisition unit (156) that acquires a prediction value by inputting an actual value consisting of one or more state values among state values that the control calculation unit can refer to into a prediction model, a prediction model generation unit (2262) that generates a prediction model based on a tree-structured learning algorithm, and a prediction model evaluation unit (2264) that evaluates the prediction model. The prediction model evaluation unit includes an analysis unit (244) that analyzes characteristics of the tree structure of the prediction model, and an evaluation unit (242) that evaluates the prediction model based on the analysis result of the analysis unit.

[Configuration 2]
The prediction model evaluation unit further includes an evaluation setting unit (261, 262) that sets the type of evaluation and a reference value for evaluating the prediction model.

[Configuration 3]
The evaluation unit (238) displays the evaluation results of the prediction model on the display unit.

[Configuration 4]
The prediction model generation unit updates the prediction model based on the learning sample corresponding to the evaluation result of the prediction model.

[Configuration 5]
The information processing device (200) is an information processing device connected to a control device (100), and the control device includes a control calculation unit (164) that executes control calculation for controlling a control target, and a predicted value acquisition unit (156) that acquires a predicted value by inputting an actual value consisting of one or more state values among state values that the control calculation unit can refer to into a prediction model. The information processing device includes a prediction model generation unit (2262) that generates a prediction model based on a tree-structured learning algorithm, and a prediction model evaluation unit (2264) that evaluates the prediction model. The prediction model evaluation unit includes an analysis unit (244) that analyzes the characteristics of the tree structure of the prediction model, and an evaluation unit (242) that evaluates the prediction model based on the analysis result of the analysis unit.

[Configuration 6]
The information processing program is executed by a computer connected to a control device (100), and the control device includes a control calculation unit (164) that executes control calculations for controlling a control target, and a predicted value acquisition unit (156) that acquires a predicted value by inputting an actual value consisting of one or more state values among state values that the control calculation unit can refer to into a prediction model. The information processing program causes the computer to execute a step (S6) of generating a prediction model based on a tree-structured learning algorithm, and a step (S12) of evaluating the prediction model. The step of evaluating the prediction model includes a step of analyzing characteristics of the tree structure of the prediction model, and a step of evaluating the prediction model based on the analysis results.

K. Advantages
In the prediction system based on the embodiment, a prediction model can be easily evaluated, so that a prediction model suitable for actual operation can be easily generated.

The embodiments disclosed herein should be considered to be illustrative and not restrictive in all respects. The scope of the present disclosure is indicated by the claims, not by the above description, and is intended to include all modifications within the meaning and scope of the claims.

1 Prediction system, 1A Anomaly detection system, 2, 4 Field bus, 6 Upper network, 10 Field device group, 14 Relay group, 100 Control device, 102, 202 Processor, 104 Chip set, 106, 206 Main memory, 108, 208 Secondary memory, 110, 214 Upper network controller, 112, 212 Controller, 114 Memory card interface, 116 Memory card, 118, 120 Field bus controller, 122 Internal bus controller, 132 Memory area, 134 Export module, 140 Prediction model, 144 Reference trajectory, 150 Engine, 152 Variables, 154 User program, 156 Prediction value acquisition code, 158 Error evaluation code, 160 Additional learning code, 162 Write code, 164 Control operation code, 200 Support device, 204 Optical drive, 205 Recording medium, 216 input unit, 218 display unit, 220 bus, 222 development program, 224 interface program, 226 analysis program, 230 user interface, 232 script engine, 234 configuration file, 236 input/output management module, 238 screen display module, 240 graph library, 242 analysis module, 244 analysis library, 250 model evaluation screen, 260, 262 input field, 261 selection item, 264, 270 tab, 266 screen, 268 message, 300 host server, 400 display device, 2262 model generation module, 2264 evaluation module.

Claims (6)

A control calculation unit that executes a control calculation for controlling a control target;
a prediction value acquisition unit that acquires a prediction value by inputting an actual value consisting of one or more state values among state values that the control calculation unit can refer to into a prediction model;
a prediction model generation unit that generates the prediction model based on a tree-structured learning algorithm;
A prediction model evaluation unit that evaluates the prediction model,
The prediction model evaluation unit,
an analysis unit that analyzes characteristics of a tree structure of the prediction model;
an evaluation unit that evaluates the prediction model based on an analysis result of the analysis unit ;
the analysis unit calculates areas of distribution diagrams of a plurality of leaf nodes as characteristic values of the plurality of leaf nodes in the tree structure of the prediction model,
A prediction system, wherein the evaluation unit evaluates the prediction model based on the calculated area of the distribution diagram of the multiple leaf nodes . The prediction system according to claim 1 , wherein the prediction model evaluation unit further includes an evaluation setting unit that sets a type of evaluation and a reference value for evaluating the prediction model. The prediction system according to claim 1 , wherein the evaluation unit displays the evaluation result of the prediction model on a display unit. The prediction system according to claim 1 , wherein the prediction model generation unit updates the prediction model based on a learning sample corresponding to an evaluation result of the prediction model. An information processing device connected to a control device, the control device comprising: a control calculation unit that executes control calculations for controlling a control target; and a predicted value acquisition unit that acquires a predicted value by inputting an actual value consisting of one or more state values among state values that the control calculation unit can refer to into a prediction model;
The information processing device includes:
a prediction model generation unit that generates the prediction model based on a tree-structured learning algorithm;
A prediction model evaluation unit that evaluates the prediction model,
The prediction model evaluation unit,
an analysis unit that analyzes characteristics of a tree structure of the prediction model;
an evaluation unit that evaluates the prediction model based on an analysis result of the analysis unit ;
the analysis unit calculates areas of distribution diagrams of a plurality of leaf nodes as characteristic values of the plurality of leaf nodes in the tree structure of the prediction model,
The evaluation unit evaluates the prediction model based on the calculated area of a distribution diagram of the multiple leaf nodes . An information processing program executed by a computer connected to a control device, the control device comprising: a control calculation unit that executes control calculations for controlling a control target; and a predicted value acquisition unit that acquires a predicted value by inputting an actual value consisting of one or more state values among state values that the control calculation unit can refer to into a prediction model;
The information processing program is configured to:
generating the predictive model based on a tree-structured learning algorithm;
and evaluating the predictive model;
The step of evaluating the predictive model comprises:
analyzing a property of the tree structure of the predictive model;
and evaluating the predictive model based on the analysis results ;
the analyzing step includes a step of calculating areas of a distribution diagram of a plurality of leaf nodes as characteristic values of the plurality of leaf nodes in the tree structure of the prediction model,
An information processing program, wherein the evaluating step includes a step of evaluating the prediction model based on the calculated area of a distribution diagram of the plurality of leaf nodes .

Images