WO2022038804A1

WO2022038804A1 - Diagnostic device and parameter adjustment method

Info

Publication number: WO2022038804A1
Application number: PCT/JP2021/004723
Authority: WO
Inventors: 惟杉田; 智昭蛭田; 嘉成堀; 孝朗関合; 拓実太田
Original assignee: 株式会社日立製作所
Priority date: 2020-08-21
Filing date: 2021-02-09
Publication date: 2022-02-24
Also published as: JP7469991B2; JP2022035686A

Abstract

This diagnostic device comprises a condition input unit for receiving input of a diagnostic purpose of a diagnostic model for diagnosing the state of a subject to be diagnosed, and data conditions for time series data to be used for creating the diagnostic model, and a parameter adjustment unit for automatically adjusting a parameter of the diagnostic model that is created, on the basis of the diagnostic purpose and the data conditions input to the condition input unit, in accordance with a modeling method including at least one of a parameter item defining a parameter for adjusting the diagnostic model, an item of the range of use of the time series data, and an item of evaluation index for evaluating the diagnostic model, and that is evaluated by the evaluation index.

Description

Diagnostic device and parameter adjustment method

The present invention relates to a diagnostic device and a parameter adjustment method.

In the process equipment and equipment used in industrial plants, measuring instruments that measure temperature, pressure, flow rate, etc. are installed in each part in order to monitor the operating status of these equipment. Conventionally, for stable operation of equipment, diagnostic equipment continuously acquires and monitors the measured values measured by these measuring instruments, and issues an alarm when the measured values exceed a certain threshold value. Has dealt with.

In recent years, there have been an increasing number of examples of efforts to deal with equipment before it breaks down by catching signs of abnormality before equipment malfunctions. In this case, the diagnostic device collects the measured values acquired by each measuring instrument and performs large-scale processing by an algorithm such as machine learning to identify the operating state of equipment that is different from normal, resulting in an abnormality. It is expected to catch the signs of.

For example, a diagnostic device that predicts an abnormality in equipment using Adaptive Resonance Theory (hereinafter referred to as ART) has been known. ART is a theory based on a clustering method that classifies multidimensional time-series data into categories (clusters) according to their similarity. When constructing a diagnostic model capable of predicting an abnormality by using a method such as ART, it is necessary to set a resolution parameter or the like that determines the resolution of the data according to the data to be diagnosed. If this parameter is not set properly, there is a possibility that an abnormality in the equipment cannot be found and an erroneous detection that the equipment is in a normal state but is detected as an abnormality may occur. The technique disclosed in Patent Document 1 is known as a technique for setting such a parameter.

Patent Document 1 states that "a diagnostic device has a preprocessing means for generating preprocessed data by executing preprocessing for classifying time-series data based on setting parameters, and the preprocessed data is similar to the data. A classification means that generates classification results classified according to the classification results, a classification result evaluation means that evaluates the classification results according to the time-dependent change pattern of the feature amount of the classification results, and a classification result evaluation means so that the evaluation results are desired values. It is provided with a setting value adjusting means for adjusting a setting parameter used in the preprocessing means. " Further, Patent Document 1 describes that "parameters suitable for diagnosis can be automatically adjusted according to the conditions of the acquired time series data".

JP-A-2017-117834

When classifying time-series data using methods such as ART as in the past, it was necessary to adjust a large number of parameters such as resolution parameters, data items of data used for diagnosis, and data normalization range. .. However, it has been difficult for the user to determine which parameter should be adjusted to obtain the desired result.

In the technique disclosed in Patent Document 1, it was possible to define the evaluation value of the classification result according to whether or not the obtained time series data includes the data at the time of abnormality, and to optimize the related parameters. However, when the diagnostic device builds a diagnostic model, it must consider various data conditions such as whether or not the given data contains abnormal data, as well as various data conditions such as how long the data period to be acquired. Must be. If the diagnostic equipment tries to adjust the parameters in a batch according to such data conditions, it takes a huge amount of calculation time to construct the diagnostic model, and it is necessary to determine the evaluation index of the model that is indispensable for automation in the first place. Sometimes I couldn't.

The present invention has been made in view of such a situation, and an object thereof is to be able to easily adjust parameters necessary for constructing a diagnostic model and to be able to construct an appropriate diagnostic model.

The diagnostic apparatus according to the present invention has a condition input unit for inputting a diagnostic purpose of a diagnostic model for diagnosing the state of a diagnosis target and a condition input unit for inputting data conditions of time-series data used for constructing the diagnostic model, and an input diagnostic purpose. And includes at least one of a parameter item in which parameters for adjusting the diagnostic model are defined based on data conditions, an item in the range of use of time-series data, and an evaluation index item for evaluating the diagnostic model. It is equipped with a parameter adjustment unit that automatically adjusts the parameters of the diagnostic model that is constructed according to the modeling method and evaluated by the evaluation index.

According to the present invention, it is possible to easily adjust the parameters necessary for constructing the diagnostic model according to the diagnostic purpose and the data conditions, so that it is possible to construct an appropriate diagnostic model.
Issues, configurations and effects other than those described above will be clarified by the following description of the embodiments.

It is a block diagram which shows the structural example of the diagnostic apparatus which concerns on embodiment of this invention. It is a figure which shows the relationship between the precondition input which concerns on embodiment of this invention, and the modeling method determination. It is a figure which shows the structural example of the condition setting screen which concerns on embodiment of this invention. It is a figure explaining the operation of ART which concerns on embodiment of this invention. It is a figure which clustered two kinds of measured values which concerns on embodiment of this invention. It is a graph which shows the change of the degree of abnormality which concerns on embodiment of this invention. It is a figure which shows the 1st display example of the diagnostic model which concerns on embodiment of this invention. It is a figure which shows the 2nd display example of the diagnostic model which concerns on embodiment of this invention. It is a contour figure of the evaluation index which concerns on embodiment of this invention. It is a flowchart which shows the processing example of the diagnostic apparatus which concerns on 1st Embodiment of this invention. It is a figure which shows the display example of the screen which the user which concerns on 1st Embodiment of this invention specifies a data period. It is a figure which shows the relationship between the precondition input which concerns on 4th Embodiment of this invention, and the modeling method determination. It is a block diagram which shows the structural example of the diagnostic apparatus which concerns on 5th Embodiment of this invention. It is a figure which shows the evaluation index of the diagnostic model and the number of clusters for each search point which concerns on 5th Embodiment of this invention. It is a block diagram which shows the structural example of the diagnostic apparatus which concerns on 6th Embodiment of this invention. It is a block diagram which shows the hardware configuration example of the computer which concerns on embodiment of this invention.

Hereinafter, embodiments for carrying out the present invention will be described with reference to the attached drawings. In the present specification and the drawings, components having substantially the same function or configuration are designated by the same reference numerals, and duplicate description will be omitted. Although the present invention is described assuming equipment constituting a general industrial plant, the present invention is not limited thereto, and can be used for industrial machinery and equipment in other fields.

[Explanation of configuration and operation common to each embodiment]
First, the configuration and operation common to each embodiment will be described. Here, when the embodiment is not specified, it is also referred to as "the present embodiment". The equipment handled in this embodiment is assumed to be, for example, a heat exchanger, a turbine, an engine, a pump, a fan, a stirring tank, a rotating equipment, or the like. In a general industrial plant, these multiple devices are connected by piping, and raw materials and intermediate products are transferred and continuously processed to manufacture final products. In addition, each device may function in a state of being combined into one facility.

A large number of measuring instruments are installed in these devices. Therefore, a diagnostic device is used that monitors and diagnoses the operating state of process equipment and equipment (an example of a diagnosis target) used in an industrial plant and the like, and detects signs of abnormalities occurring in the equipment and equipment. The diagnostic device generates time-series data of the object to be measured by the measuring instrument by acquiring the measured values continuously output for each measuring instrument. The time-series data generated for each measurement target is generally called process data, and is stored in the database in time-series via, for example, a programmable logic controller (PLC: Programmable Logic Controller) or a distributed control system (DCS: Distributed Control System). Stored.

Further, in the present embodiment, the name of the data acquired by each measuring instrument is referred to as "signal". The content of the signal may be analog data such as temperature, pressure, flow rate, and current value, but is not limited to this, and includes digital data indicating an ON / OFF signal on the control sequence of the equipment.

Prior to the description of the present embodiment, the definitions of each term will be described.
Abnormality is defined as a state in which the equipment is not operating stably. Here, the state in which stable operation is not possible is a case where the operating performance of the equipment and the quality of the product do not fall within the initially specified range. In addition, even if it is within the specified range, the situation where the index indicating performance and quality gradually changes toward the outside of the specified range is also regarded as abnormal.

A diagnostic model is defined as diagnosing an abnormality or normality with respect to given process data. Generally, diagnostic models are divided into physical models and machine learning models.
A physical model is a mathematical expression of a physical phenomenon. Since the physical relationship of each variable is clear, the physical model is characterized by high reliability and explanation. Also, if there is a physical model to explain the phenomenon that occurred in a certain diagnosis target, and if a phenomenon similar to this phenomenon also occurs in the diagnosis target of other similar projects, it will be the diagnosis target of other similar projects. Has the advantage that the same physical model can be deployed. However, there is a drawback that the creation of a diagnostic model is complicated and it is necessary to acquire basic data for various operating conditions in advance.

On the other hand, the machine learning model has the disadvantage that it is necessary to acquire a large amount of data, but it has the advantage that it is possible to construct a diagnostic model relatively easily without examining complicated physical phenomena. For the above reasons, in recent years, the introduction of diagnostic devices based on machine learning models has been increasing as an abnormality detection model for plants and equipment (hereinafter referred to as a diagnostic model). The present invention describes the following embodiments assuming a machine learning model, but the present invention is not necessarily limited to this, and a physical model may be used.

<Configuration example of diagnostic device>
Here, a configuration example of the diagnostic device according to the present embodiment will be described.
FIG. 1 is a block diagram showing a configuration example of the diagnostic apparatus 100.
The diagnostic device 100 includes a condition input unit 1, a condition determination unit 2, an evaluation index determination unit 3, a data acquisition unit 4, a measurement value database 5, a model construction unit 6, an abnormality determination unit 7, an evaluation index calculation unit 8, and a parameter adjustment unit. 9 and a result output unit 10 are provided.

The diagnostic device 100 inputs the measured values of the measuring instruments installed in the equipment or the equipment, and the plurality of measured values are stored in the measured value database 5 as time-series measured value data (so-called raw data). The measured values include not only the data acquired from the measuring instruments directly installed in the equipment or equipment, but also the data estimated and processed by the secondary processing, the set values set by the operator on the control panel, PLC, and DCS, and A digital signal such as an ON / OFF signal may be included. Here, these secondary processing data (estimated values), set values, and digital signal data are also referred to as measured values.

In addition, the measured value database 5 stores not only the measured values of the measuring instruments installed in the relevant equipment or equipment, but also data such as the outside air temperature and the composition of each fluid and raw material. These data may be measured values by a measuring instrument, but may be data related to set values, coefficients, or conditions predetermined by the process. For example, in the case of a heat exchanger, the set value is the outlet temperature, outlet pressure, or control target value of the process fluid, the coefficient is the heat transfer coefficient, and the condition is the heat exchange between gas and liquid. , Fluid composition or load, etc.

(Contents of condition input unit 1)
Next, the contents of the condition input unit 1 will be described.
The condition input unit 1 used as a component of the diagnostic device 100 has a function for the user to input at least a diagnostic purpose and data conditions to the diagnostic device 100. Therefore, the diagnostic purpose of the diagnostic model for diagnosing the state of the diagnosis target and the data condition of the time series data used for constructing the diagnostic model are input to the condition input unit 1. The time-series data is the measurement value data measured by the measurement unit 2 stored in the measurement value database 5 in time series. The content input through the condition input unit 1 is output to the condition determination unit 2, the evaluation index determination unit 3, and the data acquisition unit 4.

The diagnostic purpose is an abnormal phenomenon that the diagnostic model wants to capture. Through the development of an abnormality diagnostic model, the present inventors have found that an appropriate diagnostic model differs depending on a specific phenomenon of abnormality, even if it is generally called an abnormality of equipment. For example, the user can select these diagnostic purposes depending on whether he / she wants to catch a failure of a device or equipment or whether he / she wants to catch a deteriorated state. A failure is defined as a condition in which the functional unit's ability to perform the requested function is lost. On the other hand, the deteriorated state is defined as a state in which the state is unsteadily changing toward a failure.

The data condition is whether or not the time-series data acquired by the diagnostic apparatus 100 from the measured value database 5 includes abnormal data (abnormal data), and whether the abnormal data is time-series continuous data or discrete. It is to distinguish between data and so on. However, since it is rare that abnormal data can be acquired for plants and equipment, a diagnostic model may be constructed without using the abnormal data. Further, the abnormal data does not necessarily have to be continuous time-series data, and may be measured values at the time of periodic inspection as recorded in the maintenance management ledger. In such a case, the abnormal data becomes discrete data. Further, the visual judgment result at the time of inspection may be used, and in this case, it is referred to as qualitative data as to whether it is normal or abnormal.

(Operation of condition determination unit 2)
The content of the condition input unit 1 affects the operation of the condition determination unit 2. Here, the contents of the condition input unit 1 and an example of the operation of the condition determination unit 2 will be described with reference to FIG.
FIG. 2 is a diagram showing the relationship between the precondition input and the modeling method determination.

The prerequisite input shown in the upper part of FIG. 2 indicates the items of the diagnostic purpose and the data condition input from the condition input unit 1. The purpose of diagnosis includes an item for selecting the purpose, and the data condition includes an item for the presence or absence of abnormal data. The items of purpose selection and presence / absence of abnormal data are set by the user through the screen shown in FIG. 3 to be described later.

As shown in the lower part of FIG. 2, the condition determination unit 2 automatically determines the modeling method based on the diagnostic purpose and data conditions input by the condition input unit 1. Here, the modeling method includes adjustment parameter items, usage data range, and model evaluation index. The adjustment parameter item defines the parameters that adjust the diagnostic model. Further, the usage range (acquisition period, etc.) of the time series data that can be acquired from the measured value database 5 is defined in the usage data range. In addition, an evaluation index for evaluating a diagnostic model is defined in the evaluation index item.

The condition determination unit 2 determines the contents of the parameter items and the items of the usage range of the time series data based on the diagnostic purpose and the data conditions input from the condition input unit 1. For example, the condition determination unit 2 determines the conditions to be presented to the user based on the rules registered in advance in the condition determination unit 2 itself for each item of the modeling method to be set by the user. In this rule, for example, the condition determination unit 2 categorizes the diagnostic purpose and the data condition to some extent, determines the parameter to be changed in the case of a certain data condition for a certain diagnostic purpose, or sets a data set to be input to machine learning. It is stipulated to decide and to decide the necessary information as an evaluation index. The items determined by the condition determination unit 2 affect the operation of each subsequent functional unit. Here, each item included in the modeling method determination shown in the lower part of FIG. 2 will be described.

(Adjustment parameter item)
The item priority determines the priority of the parameter item searched by the parameter adjusting unit 9.
Patterning is to aggregate the search range in a certain parameter item into several patterns, and the search range is reduced by the patterning. For example, as a patterning of the search condition, there is a process of patterning the normalization range. In this process, when converting the input data from 0 to 1 for normalization, it is determined in advance whether the conditions for giving the maximum value and the minimum value are taken from the data of the learning period or the data of the entire acquisition period. Since it is determined, the data acquisition unit 4 does not have to acquire data in an extra range.

(Usage data range)
The items of the usage data range include a learning period required for constructing a diagnostic model, an evaluation period in which a normal period and an abnormal period are separately set, and an abnormal period in which an abnormality or deterioration occurs in the diagnosis target.
The learning period is a period in which the user is instructed to set a period for giving the measured value data to the model building unit 6 as learning input data in order to build the diagnostic model.
The evaluation period is the same as the learning period defined earlier, but depending on the data conditions, whether or not to set the normal period, and change the instruction to the user to set the normal period and the abnormal period separately. It is a period to determine. In the figure, the item described as the evaluation period (normal) defines the content instructing the user to set the normal period.
The evaluation period (abnormality / deterioration) defines the content of instructing the user to set an abnormal period (described as an evaluation period in the figure) in which an abnormality or deterioration occurs in the diagnosis target.

(Model evaluation index)
The items of the evaluation index include the detection performance indicating the performance of the diagnostic model to detect the abnormality of the diagnosis target.
The detection performance defines the content of instructing the user to set the detection performance as an evaluation index for the diagnostic device 100 to automatically select the most promising diagnostic model. This detection performance is an index showing the performance of the diagnostic model to detect an abnormality to be diagnosed.

On the other hand, in the condition input unit 1, the user may directly specify some elements of the modeling method. When some elements of the modeling method are specified, the contents of the condition input unit 1 affect the evaluation index determination unit 3 and the data acquisition unit 4.

(Operation of evaluation index determination unit 3)
Next, the operation of the evaluation index determination unit 3 will be described with reference to FIG.
The evaluation index determination unit 3 determines a specific evaluation index defined by the item of the evaluation index according to the input contents from the condition input unit 1 including the diagnostic purpose and the data condition to the condition determination unit 2. In the present embodiment, a specific evaluation index is referred to as an evaluation index of a diagnostic model (referred to as a “model evaluation index”). Here, the evaluation index is an index for judging the quality of the constructed diagnostic model, and is indispensable for optimizing the diagnostic model. A good diagnostic model can be such that the abnormality of the equipment to be diagnosed is accurately grasped, the sign of the abnormality can be detected at an early stage, and there are few false alarms. Therefore, as the evaluation index of the diagnostic model, the correct answer rate, the precision rate, the recall rate, the false alarm rate, etc. for the abnormality determination result of the diagnostic model can be mentioned.

Further, the elements of the confusion matrix created to calculate the values such as the correct answer rate may be combined and used as an evaluation index. Further, as the evaluation index, an index of how quickly the diagnostic model was able to detect an abnormality by inserting a time axis may be used. These indicators may be in a trade-off relationship. For example, a diagnostic model that detects an abnormality early tends to have a large false positive rate, and a diagnostic model that detects an abnormality over a long period of time tends to have a poor detection rate. Therefore, as an index, for example, the F value defined by the harmonic mean of the precision rate and the recall rate is adopted, or the value obtained by obtaining the AUC (Area Under the Curve) from the ROC curve (Receiver Operating Characteristic curve) is used. May be good. In addition, the diagnostic model may be regarded as a multi-objective optimization problem, and each index may be combined.

Further, the evaluation index determination unit 3 determines the evaluation index of the diagnostic model in advance according to the diagnostic purpose and the data condition included in the condition input unit 1, and stores it in the evaluation index database (not shown). Unlike the measured value database 5 shown in FIG. 1, the evaluation index database stores a setting file in which the determined evaluation index is written. When the diagnostic purpose and the data condition are input from the condition input unit 1, the evaluation index determination unit 3 can read the evaluation index from the evaluation index database and present the evaluation index to the user.

<Condition setting screen for evaluation index of diagnostic model>
As a method of determining the evaluation index, there is a method in which the user selects and determines from the candidates of the evaluation index each time. Therefore, the evaluation index determination unit 3 outputs a list of candidate evaluation indexes to a user interface such as a display device, and presents a screen on which the user can select the evaluation index. Here, a detailed configuration example of the condition setting screen for setting various conditions in the condition input unit 1 will be described with reference to FIG.
FIG. 3 is a diagram showing a configuration example of the condition setting screen 20.

The condition setting screen 20 shows each item of the diagnostic purpose, the data condition, and the model evaluation index that can be selected by the user.
In the item for diagnostic purposes, a check box is displayed in which the user can select at least one of failure detection and deterioration detection.

In the data condition item, a pull-down menu is displayed in which the user can select either with abnormal data or without abnormal data. The data condition item may be provided with a submenu that allows the user to select whether the abnormality data includes only time information or numerical information indicating the degree of abnormality for each time.
In the item of the model evaluation index, a pull-down menu is displayed in which the user can select any of the recall rate, the precision rate, the F value, and the AUC.

Note that some of the items shown on the condition setting screen 20 cannot be applied depending on the diagnostic purpose or data conditions. Therefore, the items of the evaluation index that cannot be applied may be grayed out on the condition setting screen 20 so that the user cannot select them. In addition, the condition setting screen 20 has a function of storing the diagnostic model constructed in the past in a database (not shown) and recommending the evaluation index applied to the diagnostic model to the user for the same diagnostic purpose and data condition. May be added to.

Then, the item selected by the user through the condition setting screen 20 is output from the evaluation index determination unit 3 to the data acquisition unit 4. Returning to FIG. 1, the operation of the data acquisition unit 4 will be described.

(Operation of data acquisition unit 4)
The data acquisition unit 4 acquires time-series data representing the state of the diagnosis target according to the usage range of the time-series data. For example, the data acquisition unit 4 has data item information (referred to as “Tag information”) selected by the user or determined by the condition determination unit 2, and a period for acquiring time-series data (usage data range shown in FIG. 2). ), The time-series data 11 of the target Tag is acquired from the measured value database 5 in which the time-series data of each signal is accumulated. The signal data acquired by the data acquisition unit 4 is later used as input data when constructing a diagnostic model. This input data is used as training data for machine learning and evaluation data for evaluating the performance of the machine learning model. Further, as for the input signal for which the abnormality diagnosis is performed during the operation of the diagnostic apparatus 100, the same Tag information as the Tag information attached at the time of constructing the diagnostic model is selected. The Tag information may be directly specified by the user to the condition input unit 1, or may be regarded as one of the predetermined parameter items.

(Operation of model building unit 6)
The model building unit 6 builds a diagnostic model from the time series data according to the modeling method determined in the processing in the previous stage. At this time, the model building unit 6 executes two processes, a learning mode and a diagnostic mode, to build a diagnostic model. The machine learning algorithm used for model construction is not limited to ART, and can be applied to clustering, classification, regression, and the like. Further, among the clustering algorithms, ART, Mahalanobis Taguchi method (MT method), vector quantization, Kmeans, spectral clustering and the like can be applied.

When executing the learning mode processing, the model building unit 6 constructs a machine learning model by inputting the time-series data of the learning target period acquired by the data acquisition unit 4. Further, when the model building unit 6 executes the processing of the diagnosis mode, the model building unit 6 inputs the time series data of the diagnosis target period into the diagnosis model built in the learning mode and outputs the result. Here, the result can be output as an abnormality degree, for example, a predicted value or a distance on the vector space with respect to the learning data, depending on the type of machine learning algorithm. This degree of abnormality is used in the data determination method of the diagnostic model as described later.

Here, the ART used by the model building unit 6 in the learning mode and the diagnostic mode will be described.
FIG. 4 is a diagram illustrating the operation of ART. The time variation of the two types of measured values A and B is shown on the upper side of FIG. 4, and the cluster No. is shown on the lower side of FIG.

ART is an algorithm that creates a cluster from the correlation between data. The cluster in which the model building unit 6 shows the states of the measured values A and B in the learning mode is referred to as the cluster No. It is assumed that they are classified by 1 to 3. Cluster No. 1 is a state in which the measured value A is higher than the measured value B, and the cluster No. No. 2 is a state in which both the measured values A and B are low, and the cluster No. 3 represents a state in which the measured value B is higher than the measured value A.

After the cluster is generated, the model building unit 6 starts the diagnosis of the measured values A and B in the diagnosis mode from the start of the diagnosis. Then, it is assumed that a state has occurred in which both the measured values A and B show high values. This state is the cluster No. at the time of learning. It does not correspond to any of 1 to 3. Therefore, the model building unit 6 generates a cluster showing a state in which both the measured values A and B show high values, and the cluster No. 1 is generated in this cluster. Add 4. Then, when a new cluster is generated, the diagnostic device 100 will notify the user of the occurrence of an abnormality.

In this way, the model building unit 6 generates a new cluster when data that is considered not to belong to the cluster generated during the learning period appears in the diagnostic mode. Therefore, the model building unit 6 may output the presence / absence of the appearance of a new cluster as a flag. This flag is used when the presence or absence of a new cluster is used as the data judgment method of the diagnostic model.

Further, the model building unit 6 may output the degree of abnormality based on the clustering information.
FIG. 5 is a diagram in which measured

values

1 and 2 are clustered.
FIG. 5 shows a graph in which the value of the measured value 1 is taken on the horizontal axis, the value of the measured value 2 is taken on the vertical axis, and the points where the measured

values

1 and 2 measured at the same time intersect are plotted. The white circles in the figure represent learning data, and the black circles represent diagnostic data. When the diagnostic model is evaluated using the input data as the evaluation data as described above, the diagnostic data shown in FIG. 5 is replaced with the evaluation data.

The intersections of the measured

values

1 and 2 are clustered with circles of a predetermined size. This circle represents the cluster generated by the training data. The x mark in the center of the circle represents the center of gravity of the cluster generated by the training data, and is called the representative point of the cluster.

FIG. 5 shows an example of three clusters of clusters 1 to 3. The model building unit 6 calculates the degree of abnormality based on the points of each training data and diagnostic data and the distance on the vector space with respect to the representative points of the cluster. For example, for each learning data and diagnostic data, the distance from the center of gravity of the nearest cluster is defined as the degree of anomaly. The measured values 1 and 2 plotted farther than the circle of each cluster are outliers that do not belong to any of the clusters. As shown in this outlier, the greater the difference between the points of the diagnostic data and the data pattern formed by the training data, the longer the distance in the vector space and the higher the degree of anomaly. Therefore, the degree of abnormality is used as an index for measuring the degree of abnormality in the diagnostic data.

Although the model building unit 6 is supposed to build a diagnostic model by using a method based on a machine learning model, a diagnostic model may be built by using a method based on a physical model.

(Operation of abnormality determination unit 7)
Next, the operation of the abnormality determination unit 7 will be described with reference to FIG.
The abnormality determination unit 7 determines whether or not there is an abnormality in the time-series data input to the diagnostic model constructed by the model construction unit 6 and acquired by the data acquisition unit 4 during the evaluation period. At this time, the abnormality determination unit 7 determines whether or not the diagnostic data is regarded as an abnormality based on the determined rule. Here, the operation of the abnormality determination unit 7 will be described with reference to FIG.

FIG. 6 is a graph showing changes in the degree of abnormality. In this graph, the horizontal axis is time and the vertical axis is the degree of abnormality, and the time change of the degree of abnormality is shown. The abnormality determination unit 7 determines the presence or absence of an abnormality by comparing the degree of abnormality calculated by the model construction unit 6 in the diagnostic mode with a threshold value set in advance by the user. Here, the threshold value to be compared with the degree of abnormality is an example of the parameter adjusted by the parameter adjusting unit 9.

For example, as shown in FIG. 6, if the degree of abnormality is a value less than the threshold value indicated by the alternate long and short dash line in the figure, the abnormality determination unit 7 does not consider the degree of abnormality to be abnormal. However, when the degree of abnormality exceeds the threshold value, the abnormality determination unit 7 considers the degree of abnormality as an abnormality and notifies the occurrence of the abnormality. The abnormality determination unit 7 may use the presence / absence of the appearance of a new cluster as it is for the abnormality determination. For example, when the model construction unit 6 finds that a "new cluster" has appeared, the abnormality determination unit 7 determines that an abnormality has occurred.

(Operation of evaluation index calculation unit 8)
Next, the operation of the evaluation index calculation unit 8 will be described with reference to FIG.
The evaluation index calculation unit 8 calculates a specific evaluation index based on the determination result of the presence or absence of abnormality in the time series data. For example, the evaluation index calculation unit 8 calculates the evaluation index (for example, the correct answer rate) of the diagnostic model determined by the evaluation index determination unit 3 based on the abnormality determination result executed by the abnormality determination unit 7. In order for the evaluation index calculation unit 8 to calculate the evaluation index, a correct label corresponding to the abnormality determination result output by the diagnostic model is required.

This correct answer label gives a label that all the abnormality determination results are normal when the measurement value data acquired from the measurement value database 5 by the data acquisition unit 4 does not include the data at the time of abnormality. May be good. However, when the evaluation index calculation unit 8 gives a label that all are normal, the evaluation index calculation unit 8 divides the normal data into k sets by using, for example, a cross-validation method. Then, the evaluation index calculation unit 8 constructs a diagnostic model using the data of the (k-1) set of the k sets as training data, and performs a process of diagnosing using the remaining one set of data as diagnostic data. conduct. By this processing, the abnormality determination unit 7 makes an abnormality determination, and the evaluation index calculation unit 8 can calculate the evaluation index.

Further, when the abnormal data cannot be acquired among the measured value data acquired by the data acquisition unit 4 from the measured value database 5, the evaluation index calculation unit 8 simulates the abnormal data including the abnormal value artificially created by the user. It is also possible to create it as a target. The reason why the user artificially creates an abnormal value in this way is that, in general, the model building unit 6 can create a highly accurate diagnostic model if there is abnormal data. Specifically, the evaluation index calculation unit 8 adds ± nσ to the measurement values of one or more signals for the time-series data of the signals acquired by the data acquisition unit 4 from the measurement value database 5 during normal operation. Here, σ is a standard deviation and n is an arbitrary real number. Then, the evaluation index calculation unit 8 gives an abnormality label to the artificially created abnormality data.

The method of artificially creating anomalous data is not limited to this method, and in addition, a method of using a simulation, a mathematical model, or a physical model may be used. Also by using these methods, the model building unit 6 can build a more accurate diagnostic model.

(Operation of parameter adjustment unit 9)
Next, the operation of the parameter adjusting unit 9 will be described.
The parameter adjustment unit 9 is a model including at least one of parameter items, time-series data usage range items, and evaluation index items based on the diagnostic purpose and data conditions input to the condition input unit 1. It is constructed according to the method and automatically adjusts the parameters of the diagnostic model evaluated by the evaluation index. The parameters adjusted by the parameter adjusting unit 9 are output to the data acquisition unit 4 and the abnormality determination unit 7.

For example, the parameter adjustment unit 9 updates the parameters to be adjusted selected in the parameter items shown in FIG. 2 in order to optimize the diagnostic model with respect to the evaluation index of the diagnostic model calculated by the evaluation index calculation unit 8. do. The parameters to be adjusted include, for example, Tag information, learning period, diagnosis period, normalization range of data, and the like. In this embodiment, Tag information is an example of data item parameters. The learning period and the diagnosis period are both examples of the model input period. The data normalization range is an example of a data item normalization range parameter.

The parameter adjustment unit 9 includes hyperparameters peculiar to the machine learning model as parameters for determining the parameters peculiar to the diagnostic model. For example, in the case of ART, resolution parameters and learning rates correspond to hyperparameters. The threshold value used by the abnormality determination unit 7 to execute the abnormality determination is also one of the parameters. The parameter update condition is not limited to one. For example, any optimization method such as grid search, steepest descent, stochastic gradient descent, conjugate gradient descent, Newton's method, quasi-Newton's method, genetic algorithm, Bayesian optimization, reinforcement learning, etc. can be applied.

The setting parameters set for constructing the diagnostic model in this way include at least one of the data item parameters, the model input period, the normalization range parameters of the data items, and the parameters that determine the model-specific parameters.

Further, the parameter adjusting unit 9 can aggregate the parameter search conditions into several patterns in advance. By this pattern aggregation, the processing amount required for the parameter adjustment unit 9 to search for parameters can be significantly reduced. Further, the parameter adjusting unit 9 may give priority to the parameter items. Since each parameter has a different role, it is possible to efficiently optimize the diagnostic model by setting the order of the parameter items to be searched preferentially according to the diagnostic purpose in advance.
Then, the model construction unit 6, the abnormality determination unit 7, the evaluation index calculation unit 8, and the parameter adjustment unit 9 repeat a series of processes a predetermined number of times.

(Operation of result output unit 10)
Next, the operation of the result output unit 10 will be described.
The result output unit 10 is a part that displays a diagnostic model optimized based on the evaluation index of the model. At this time, the result output unit 10 outputs a specific evaluation index calculated by the evaluation index calculation unit 8. Here, the result output unit 10 can output a specific evaluation index calculated repeatedly by the evaluation index calculation unit 8 a predetermined number of times. Therefore, a display example of the diagnostic model will be described with reference to FIG. 7.

FIG. 7 is a diagram showing a first display example of the diagnostic model. The diagnostic model is displayed as a diagnostic model display screen 21 having the form shown in FIG. 7. The diagnostic model display screen 21 displays elements that characterize the diagnostic model in a table format. The diagnostic model display screen 21 is composed of record No., parameter 1, parameter 2, number of data, learning abnormality degree average, diagnosis time abnormality degree average, precision rate, recall rate, F value, and file link items. Parameter 1 and parameter 2 represent parameter conditions searched during machine learning. Further, parameter 1, parameter 2, and the number of data are setting information required for machine learning.

The average degree of abnormality at the time of learning, the average degree of abnormality at the time of diagnosis, the precision rate, the recall rate, and the F value are values calculated as learning results of machine learning and are related to the evaluation index. Here, the graph No. shown in the item of the file link. 1, No. 2, No. 3 is used, for example, to link to the trend diagram file shown in FIG. 6 and display the trend diagram. The user can see the graph No. of the item of the file link. When the character 1 is clicked, the record No. 1 is displayed from the diagnostic model display screen 21. It is displayed by switching to the trend diagram of 1.

Each of the records (No. 1, No. 2, ...) Displayed on the diagnostic model display screen 21 represents the "diagnostic model" displayed on the result output unit 10. The display method of the diagnostic model is not limited to one, but the calculation result including the parameter conditions (values of parameter 1 and parameter 2) searched by the parameter adjustment unit 9 and information and evaluation indexes related to the construction model ( It is possible to output a list of the number of data to the F value).

The result output unit 10 selects the highest-level diagnostic model based on the evaluation index, and displays a trend diagram (see FIG. 6) in which the degree of abnormality and the threshold value are described in the time-series direction related to this diagnostic model. It is also possible to display. At this time, the result output unit 10 may display only the trend diagram having the best evaluation index, which is related to the diagnostic model obtained by the parameter adjustment unit 9 searching for the parameters. Further, the result output unit 10 may sort from the top of the evaluation index and display several top trend charts side by side so that the user can visually check the trend chart and select an arbitrary model. .. Whether the evaluation index is good (including the best) or bad is judged by the maximum value or the minimum value of the evaluation index.

FIG. 8 is a diagram showing a second display example of the diagnostic model.
The items constituting the diagnostic model display screen 22 shown in FIG. 8 are the same as those of the diagnostic model display screen 21 shown in FIG. 7, but the RANK item is added instead of the record No. on the left end of the screen, and the item on the right end of the screen is added. The difference is that instead of file links, graph items have been added. RANK represents the evaluation order of the diagnostic model, and the diagnostic model with the best specific evaluation index is represented by "1". Further, in FIG. 8, the user sets the graph No. of the graph item. When the cursor is placed on the character 2, the record No. 2 is superimposed on the diagnostic model display screen 22. It shows how the trend diagram of 2 is automatically displayed.

Note that there is a possibility that a plurality of optimum values of the evaluation index will appear on the diagnostic model display screen 22 shown in FIG. In that case, the result output unit 10 applies additional conditions to the higher evaluation index and displays the diagnostic model display screen 22. By such processing, the result output unit 10 can support the user to select a model through the diagnostic model display screen 22.

For example, the result output unit 10 can sort and present the index included in the above list as the second evaluation index from the top. Specifically, the result output unit 10 has an abnormality represented by the ratio of the characteristic amount of the diagnostic model (average degree of abnormality) in the normality determination period and the characteristic amount of the diagnostic model (average degree of abnormality) in the abnormality determination period. The average ratio can be used as the second evaluation index. For example, the graph No. 8 in FIG. As shown in 2, the period in which the degree of abnormality is less than the threshold value is called the normality determination period, and the period in which the degree of abnormality is greater than or equal to the threshold value is called the abnormality determination period.

When considering robustness, the result output unit 10 has a method of extracting the condition in which the first evaluation index is higher and selecting the median value of the parameter. For example, in the diagnostic model display screen 22 shown in FIG. 8, the first evaluation index is assumed to be an F value, and the second evaluation index is assumed to be an average degree of abnormality during learning. Then, on the diagnostic model display screen 22 shown in FIG. 8, a table in which the records are sorted in descending order by the F value of the first evaluation index is shown.

In this way, the diagnostic model display screen 22 displays a list of model candidates sorted by an arbitrary evaluation index (for example, F value) in a list format. When displaying the parameter search results on the list, the result output unit 10 sorts the diagnostic models in descending order of evaluation index. Therefore, it is possible to display the diagnostic model in a time-series trend according to various conditions.

Here, when the evaluation indexes of a plurality of model candidates have the same ratio, the median value or the average value of the parameters can be presented as a recommendation by the result output unit 10. That is, when the numerical values of the evaluation indexes are the same in the two candidate models selected based on the evaluation indexes, the result output unit 10 calculates, for example, the median value or the average value for the parameter values in each model. Can be presented.

Further, if there are a plurality of model candidates having the same first evaluation index among the model candidates displayed on the diagnostic model display screen 22, the model candidates are further sorted by the second evaluation index. When a plurality of high-ranking KPIs (Key Performance Indicators) appear in this way, they can be sorted by the second evaluation index.

As shown on the diagnostic model display screens 21 and 22, the user can select an appropriate diagnostic model by displaying two or more evaluation indexes of the diagnostic model together. In addition, the user can appropriately select an evaluation index for sorting model candidates. Therefore, when a certain evaluation index is emphasized, the user can select an appropriate model candidate while confirming how the other evaluation indexes change.

Further, the result output unit 10 can display the evaluation index on the screen in another form. For example, an example shown by an equivalence diagram of the evaluation index with each parameter as an axis will be described with reference to FIG. 9.
FIG. 9 is a contour diagram of the evaluation index.

On the left side of FIG. 9, a diagram is shown in which parameter 1 is taken on the horizontal axis and parameter 2 is taken on the vertical axis, and the search points searched by the parameter adjusting unit 9 are plotted. On the right side of FIG. 9, among the plotted search points, an equivalence diagram in which equivalence search points are connected by a line is shown. Since the evaluation index of the diagnostic model created at each search point is given, the result output unit 10 can draw an equivalence diagram. This contour diagram has an advantage that the distribution status of the evaluation index is visually easy for the user to understand. Further, the distribution 23 in which the evaluation index is closest to the center and the evaluation index is high indicates that the combination of the

parameters

1 and 2 is appropriate.

Note that parameter 1 may be read as a first evaluation index and parameter 2 may be read as a second evaluation index, and when a plurality of high-ranking KPIs occur, the first and second evaluation indexes may be displayed in an equivalence diagram. Further, the result output unit 10 can display not only the two-dimensional diagram but also the three-dimensional diagram with another parameter as the axis as the contour diagram. Further, the result output unit 10 may aggregate conditions with similar trends by pattern recognition and display only representative figures as candidates. In this case, among the many search points shown on the left side of FIG. 9, the search points having similar values are aggregated, and the number of search points is reduced and displayed.

Further, as another form of the screen displayed by the result output unit 10, it is also possible to collectively display a representative trend diagram from the high-ranking candidate model group obtained as the search result. This is because, for example, a diagnostic model in which the parameter search points are close to each other is expected to have almost the same appearance of the trend diagram. Since it is difficult for a person to visually discriminate these similar trend charts, it is effective for the result output unit 10 to aggregate a plurality of trend charts into patterns. As a specific pattern aggregation method, a pattern recognition method based on machine learning can be mentioned, but there is no particular limitation.

As described above, in the diagnostic apparatus 100 according to the present embodiment, at least the diagnostic purpose and the data condition are input, and the information is used to determine the evaluation index and the search condition of the diagnostic model. Therefore, in the diagnostic apparatus 100, the evaluation index determination unit 3 determines the evaluation index according to not only the conditions of the time-series data acquired by the data acquisition unit 4 but also the diagnostic purpose of the diagnostic model constructed by the model construction unit 6. Therefore, the parameter adjusting unit 9 can quickly and automatically adjust the parameters related to the construction of the diagnostic model. At this time, the adjustment parameters related to the evaluation index and the search condition are automatically searched, so that the user can be shown the information for constructing the optimum diagnostic model. Therefore, it is possible to reduce the man-hours for designing a diagnostic model and support the development of a highly accurate model.

Further, the diagnostic device 100 automatically extracts candidates for diagnostic purposes based on the past maintenance history of the device and the like, and supports the construction of a diagnostic model capable of assuming a failure that may occur for each device. Can be done. Therefore, the diagnostic apparatus 100 can create a diagnostic model related to the failure diagnosis of the device by setting the minimum necessary items such as the setting for the purpose of diagnosis by the user.

Next, an example in which the diagnostic apparatus according to this embodiment is applied to each of the following embodiments will be described in order.

[First Embodiment]
In the first embodiment, an example of ART, which is one of the machine learning algorithms, is taken up. Then, as shown in FIG. 3, when the diagnostic purpose is set to detect a failure and the data condition is set to include data at the time of abnormality in the time series data, and the exact time of occurrence of the abnormality cannot be specified. , The operation of the diagnostic apparatus 100 that automatically adjusts the parameters will be described.

<Example of processing of diagnostic device 100>
FIG. 10 is a flowchart showing an example of processing of the diagnostic apparatus 100 according to the first embodiment. Here, a method of adjusting parameters of a diagnostic model performed in cooperation with each part of the diagnostic apparatus 100 will be described.

First, the diagnostic device 100 operates the condition input unit 1 (S1). The user selects a diagnostic purpose and data conditions through the condition setting screen 20 (see FIG. 3) displayed on the display device. When the candidate of the model evaluation index registered in advance is presented according to the information of the selected diagnostic purpose and the data condition, the user determines one evaluation index from the presented candidates. In the present embodiment, as shown in FIG. 3, it is assumed that the user selects "failure detection" as the diagnostic purpose and "abnormal data exists" as the data condition. After that, the diagnostic apparatus 100 determines the diagnostic purpose and the data condition selected by the condition input unit 1 by operating the condition determination unit 2.

Next, the diagnostic device 100 operates the evaluation index determination unit 3 (S2). Here, it is assumed that the evaluation index determination unit 3 selects the F value from the candidates of the model evaluation index registered in advance based on the conditions determined by the condition determination unit 2. Then, the F value selected by the evaluation index determination unit 3 is displayed on the condition setting screen 20 shown in FIG.

Next, the diagnostic device 100 operates the data acquisition unit 4 (S3). Here, the learning data and the diagnostic data to be used by the user for constructing the diagnostic model are selected. In that case, signal information and data period are required.

The signal information may be specified in advance by the user in a list or a text file, or may be selected by the user through a screen (not shown). Alternatively, a method may be used in which signal information candidates are presented on the condition setting screen 20 and the user checks and selects the signal information. The measured value data of the type specified as the signal information is acquired from the measured value database 5 by the data acquisition unit 4.

Similarly, for the data period, it is possible to apply a method in which the user specifies the period in advance using a list or a text file, and a method in which the user inputs and specifies the period from the screen (not shown). Alternatively, the trend of the measured value data (also referred to as “target data”) required for model construction may be displayed, and the period may be specified by a method of dragging and dropping by the user.

Here, a display example of a screen on which the user specifies a data period will be described with reference to FIG.
FIG. 11 is a diagram showing a display example of a screen on which the user specifies a data period. In FIG. 11, the time is taken on the horizontal axis and the measured value is taken on the vertical axis, and the state of the time change of the measured values A and B is shown in a graph (trend diagram).

When the case where there is abnormal data in the measured value data is selected as the data condition, the data acquisition unit 4 attaches information indicating that the time series data acquired during the period when the correct answer is known is normal. For example, as shown in the graph of FIG. 11, the data acquisition unit 4 attaches a correct answer label to a region known to be the correct answer in advance, and designates the data in this region as normal data. The measured value data acquired by the data acquisition unit 4 during the period designated as normal data is labeled with "normal" on the correct answer label. Further, the data acquisition unit 4 attaches information indicating that the time series data acquired during the period in which the abnormality is known is abnormal. For example, the data acquisition unit 4 attaches a correct answer label to an area that is clearly known to be abnormal, and designates the data in this area as abnormal time data. The measured value data acquired by the data acquisition unit 4 during the period designated as the abnormal time data is labeled with "abnormal" on the correct answer label.

Then, the data acquisition unit 4 excludes unknown points from the modeled data when selecting the range to be labeled as normal or abnormal. Therefore, the area (period) other than the area designated as normal data and the area (period) designated as abnormal data by the data acquisition unit 4 is excluded from the evaluation target. It is often difficult to determine whether the measured values A and B included in the period not subject to evaluation are normal or abnormal. Therefore, the measured value data of the measured values A and B included in the period not subject to evaluation is not used when calculating the evaluation index of the model.

In this way, the data acquisition unit 4 acquires the target data from the measured value database 5 in which the time series data is stored, based on the signal information and the data period specified through the screen (not shown).

Next, the diagnostic device 100 operates the model building unit 6 (S4). The model building unit 6 constructs a diagnostic model using time-series data with information indicating that it is normal and time-series data with information indicating that it is abnormal. do. At this time, the model building unit 6 operates the learning mode of ART by combining the learning data, the normalization range, and the resolution parameters of the data prepared by the data acquisition unit 4. For details on the ART clustering algorithm, refer to the description in Patent Document 1 (Japanese Unexamined Patent Publication No. 2017-117034). The outline of ART processing is as follows.
Process 1: The model building unit 6 normalizes the input vector.
Process 2: The model building unit 6 selects a suitable cluster candidate by comparing the input data with the weighting coefficient.
Process 3: The model building unit 6 evaluates the validity of the selected cluster by comparison with the resolution parameter.

If the model building unit 6 determines that the selected cluster is appropriate, it classifies the input data into that cluster and proceeds to the next process 4. If the model building unit 6 does not determine that the selected cluster is valid, it resets the cluster and selects a suitable cluster candidate from other clusters (that is, repeats process 2). If the value of the resolution parameter is increased, the classification of clusters becomes finer and the number of clusters generated increases. On the other hand, if the value of the resolution parameter is made small, the classification of clusters becomes coarse and the number of clusters generated decreases.

Process 4: When the model building unit 6 resets all the existing clusters in the process 2, it determines that the input data belongs to the new cluster and generates a new weighting coefficient representing the new cluster.
Process 5: When the input data is classified into clusters, the model building unit 6 updates the weighting factors corresponding to the clusters using the past weighting factors and the input data.

In this way, when the model building unit 6 executes the learning mode, the input data is classified into clusters by ART, and each weighting factor is updated, so that the model building unit 6 can obtain model information. Therefore, when new input data is obtained in the trained ART, the model building unit 6 can determine which cluster generated in the past can be classified by the above algorithm. Further, the model building unit 6 can generate a new cluster if it cannot be classified into any cluster. Further, as shown in FIG. 5, the model building unit 6 generates an abnormality degree for the learning data and the diagnostic data based on the distance from the representative point of the cluster. Then, the model building unit 6 can measure the degree of abnormality in the diagnostic data by using the degree of abnormality.

Next, the diagnostic device 100 operates the abnormality determination unit 7 (S5). The abnormality determination unit 7 performs abnormality determination by comparing the abnormality degree output by the model construction unit 6 with the threshold value which is one of the parameters.

Next, the diagnostic device 100 operates the evaluation index calculation unit 8 (S6). Since the correct answer label is given to the diagnostic data in advance, the evaluation index calculation unit 8 can calculate the F value by comparing with the determination result output from the abnormality determination unit 7.

Next, the diagnostic apparatus 100 stores each parameter condition and a series of calculation results in a temporary memory or a database (not shown) (S7).

Next, the diagnostic device 100 determines the number of parameter searches (S8). The diagnostic apparatus 100 according to the present embodiment specifies the number of searches according to the initially set start value, end value, and step width of each parameter based on the grid search method. This number of searches is called a "specified number of times". Then, the diagnostic apparatus 100 branches the subsequent processing depending on whether or not the number of parameter searches indicated by the value of the search count counter satisfies the specified number of times.

When the number of parameter searches is less than the specified number of times (NO in S8), the diagnostic device 100 operates the parameter adjustment unit 9 (S9). The parameter adjusting unit 9 according to the present embodiment updates the parameters one by one to the next search area according to the step size of each parameter set in advance. After that, the processes after step S4 are repeated. Therefore, the series of processes of steps S4 to S9 is repeated a designated number of times.

When the number of parameter searches satisfies the specified number of times (YES in S8), the diagnostic device 100 operates the result output unit 10 (S10). The result output unit 10 presents on the screen the conditions in which the model evaluation index is the best from the stored results.

In this way, the diagnostic device 100 was able to construct a failure detection model (an example of a diagnostic model) assuming that a failure phenomenon to be diagnosed can be detected. In the present embodiment, the case where ART is used as the classification means has been described, but the classification means is not limited to ART. For example, various classification means such as k-means clustering and vector quantization can be adopted.

In addition, the purpose of diagnosis includes detecting an abnormality in equipment or equipment, or detecting deterioration in the performance of equipment or equipment. Further, as the data condition, the user can select the presence / absence of abnormal data or the type of deterioration data (maintenance information or numerical data). Therefore, the diagnostic device 100 can realize CBM (Condition Based Maintenance) that predicts the state of the equipment and performs maintenance in response to not only abnormality detection of the equipment but also deterioration detection. ..

[Second Embodiment]
Next, an operation example of the diagnostic apparatus according to the second embodiment of the present invention will be described. As the diagnostic device according to the second embodiment, the diagnostic device 100 according to the first embodiment is used. Here, the operation that enables the diagnostic device 100 according to the second embodiment to automatically adjust the parameters when there is no abnormality data will be described.

Similar to the first embodiment, the diagnostic device 100 according to the second embodiment operates the condition determination unit 2 in step S1 shown in FIG. At this time, the user selects the diagnostic purpose as "failure detection" and the data condition as "no abnormal data" through the condition setting screen 20. However, at this point, abnormal data cannot be given to the data acquisition unit 4 as a correct answer label. Therefore, in step S2, as shown in the first embodiment, the user cannot select the F value as the evaluation index of the diagnostic model.

Therefore, the condition setting screen 20 outputs a false alarm rate that can be evaluated only with normal data as a candidate. For example, in the item of "model evaluation index" on the condition setting screen 20 of FIG. 3, "F value" is grayed out and "false alarm rate" is displayed instead. The erroneous report rate represents the ratio of the number of data erroneously determined by the abnormality determination unit 7 to all the data given the normal correct answer label. In this definition, as described in the first embodiment, when the abnormality determination unit 7 erroneously determines certain data as an abnormality by using the threshold value as one of the parameters and comparing the threshold value with the degree of abnormality, this definition is used. To prevent the data from being judged as abnormal, the threshold value may be increased. Therefore, when the case where there is no abnormal data in the time series data is selected as the data condition, the data acquisition unit 4 provides information indicating that the time series data acquired in the period when the correct answer is known is normal. Attach.

The parameter adjustment unit 9 adjusts the parameters for the time-series data to which the information indicating that it is normal is attached, based on the false alarm rate that the abnormality determination unit 7 erroneously determines as an abnormality. However, it is expected that a number of high-ranking model candidates will appear in the process of the parameter adjustment unit 9 performing the parameter search, or a problem of falling into a local optimum solution will occur. Therefore, these problems can be avoided by including a certain tolerance (margin of error) in the false alarm rate. As a method of avoiding such a problem, for example, the abnormality determination unit 7 does not determine that the abnormality is abnormal at the same time as the abnormality exceeds the threshold value, but when the abnormality degree exceeds the threshold value continuously for 10 unit hours. Examples include a method of counting abnormalities. Further, a method in which the user sets an allowable range of the false alarm rate and a method in which the diagnostic apparatus 100 statistically sets the allowable range based on the learning data can be mentioned.

Then, the diagnostic apparatus 100 according to the second embodiment executes the processes of steps S3 to S10 as the same processes as those of the first embodiment. By this process, the diagnostic apparatus 100 can construct an optimum failure diagnosis model even under the condition that there is no abnormality data.

Further, the construction of the diagnostic model may be carried out using a machine learning algorithm other than clustering. For example, regardless of the type of machine learning algorithm, a diagnostic model can be introduced by estimating the false alarm rate assumed during operation using cross validation and introducing it into the model evaluation index together with the margin of error. Can be optimized. The processing when the cross-validation method is used will be described later (see FIG. 13).

Note that the types of parameter search algorithms include, for example, grid search, Bayesian optimization, and genetic algorithms. By using such a parameter search algorithm, the diagnostic apparatus 100 can improve the speed of constructing a diagnostic model.

[Third Embodiment]
Next, an operation example of the diagnostic apparatus according to the third embodiment of the present invention will be described. As the diagnostic device according to the third embodiment, the diagnostic device 100 according to the first embodiment is used. Here, the operation of the diagnostic device 100 according to the third embodiment to automatically adjust the parameters when deterioration detection is selected for the purpose of diagnosis will be described.

Similar to the first embodiment, the diagnostic device 100 according to the third embodiment operates the condition determination unit 2 in step S1 shown in FIG. At this time, the user checks the "deterioration detection" of the diagnostic object through the condition setting screen 20. As described above, the data condition is displayed on the condition setting screen 20 so that the user can select either qualitative data or quantitative data. That is, in the item of "data condition" on the condition setting screen 20, "qualitative data" and "quantitative data" are selectively displayed instead of "with abnormal data" and "without abnormal data" shown in FIG. To. Here, it is assumed that the user selects "qualitative data" from the "data condition" item.

After that, the diagnostic device 100 operates the evaluation index determination unit 3 in step S2. The evaluation index determination unit 3 can select the recall rate as the evaluation index in order to reliably detect the deterioration state of the device based on the diagnostic purpose and the data conditions set through the condition setting screen 20.

Then, the diagnostic apparatus 100 according to the third embodiment executes the processes of steps S3 to S10 as the same processes as those of the first embodiment. Then, the parameter adjusting unit 9 adjusts the parameters based on the reproducibility selected as the evaluation index when the diagnostic purpose is to detect deterioration of the diagnosis target. By this process, the diagnostic apparatus 100 can construct an optimum diagnostic model even when the diagnostic purpose is deterioration detection.

[Fourth Embodiment]
Next, an operation example of the diagnostic apparatus according to the fourth embodiment of the present invention will be described. As the diagnostic device according to the fourth embodiment, the diagnostic device 100 according to the first embodiment is used. The diagnostic apparatus 100 according to the fourth embodiment operates in consideration of operability in addition to the first embodiment.

The difference between the fourth embodiment and the first embodiment is that the number of clusters is taken into consideration on the input screen. As described in Patent Document 1 (Japanese Unexamined Patent Publication No. 2017-117034), the cluster No. and the event are registered in association with each other in the cluster of the classification result. Therefore, at the stage of operating the diagnostic apparatus 100, it is possible to monitor the operating state of the equipment according to the appearance of the cluster.

As mentioned earlier, the frequency of new cluster generation during diagnosis and operation differs depending on the adjustment of each parameter including the resolution parameter. Therefore, when the operator monitors the operating status of the equipment, when a large number of new clusters appear in a day, it is necessary to register the event in the cluster each time, which increases the load on the operator.

Here, the contents of the condition input unit 1 and an example of the operation of the condition determination unit 2 according to the fourth embodiment will be described with reference to FIG.
FIG. 12 is a diagram showing the relationship between the precondition input and the modeling method determination.

It is shown that the precondition input shown in the upper part of FIG. 12 has an operability item added in addition to the diagnostic purpose and the data condition included in the condition input unit 1. The items included in the diagnostic purpose and the data conditions are the same as the items shown in FIG. On the other hand, the item of operability includes at least one item of the target number of clusters and the margin of error.

The target number of clusters is the target value of the number of clusters generated at the time of learning or diagnosis of the model building unit 6. By giving the target number of clusters in advance by the user, the number of clusters generated at the time of learning or diagnosis is within the target number of clusters. This simplifies the registration of events in the cluster and facilitates the selection of diagnostic models.

The number of clusters generated is the number of clusters generated at the time of learning or diagnosis of the model building unit 6.
Items of purpose selection, presence / absence of abnormal data, target number of clusters, and tolerance are set by the user through the condition setting screen 20 shown in FIG.

As shown in the lower part of FIG. 12, the condition determination unit 2 includes an adjustment parameter item, a usage data range, and a model evaluation index based on the diagnostic purpose, data condition, and operability items input by the condition input unit 1. The modeling method is automatically determined.

Specifically, the condition determination unit 2 sets the conditions to be presented to the user based on the pre-registered rules for each item of the modeling method to be set by the user according to the diagnostic purpose, the data condition and the operability. decide. The adjustment parameter items and the items included in the usage data range are the same as the items shown in FIG. On the other hand, the model evaluation index shown in FIG. 12 includes an item of cluster classification number error in addition to the detection performance shown in FIG. A database (not shown) is a file or list that defines the diagnostic model evaluation index including the usage data range and detection performance by patterning the item priority and search condition of the adjustment parameters in advance based on the diagnostic purpose, data conditions, and operability. ) Can be saved.

The item of the cluster classification number error included in the model evaluation index is given from the target number of clusters included in the operability shown in the upper part of FIG. 12 and the number of clusters generated described above.
The cluster number classification error is a value obtained by converting the number of clusters generated with respect to the target number of clusters as an error. At the same time, as in the first embodiment, the user can also select the F value as an index for evaluating the model accuracy.

Similar to the first embodiment, the diagnostic device 100 according to the fourth embodiment operates the condition determination unit 2 in step S1 shown in FIG. Next, the diagnostic device 100 operates the evaluation index determination unit 3 in step S2 shown in FIG. At this time, the F value is valid only within the margin of error from the target number of clusters shown in the upper part of FIG. For example, the evaluation index outside the permissible range is 0, and the evaluation index within the range is the F value. The specific formula can be applied arbitrarily.

Then, the diagnostic apparatus 100 according to the fourth embodiment executes the processes of steps S3 to S10 as the same processes as those of the first embodiment. By this process, the diagnostic apparatus 100 can construct a diagnostic model with the highest accuracy within the range of the target number of clusters and the tolerance.

In the diagnostic apparatus 100 according to the fourth embodiment described above, the preconditions including the diagnostic purpose, data conditions, target cluster number / cluster number tolerance are input from the condition input unit 1. Then, based on this precondition, the item priority / search condition of the adjustment parameter is patterned, and the diagnostic modeling method including the usage data range, the model evaluation index including the detection performance and the cluster classification number error is determined. Therefore, the diagnostic device 100 can construct an appropriate diagnostic model based on the diagnostic purpose, data conditions, and operability.

Further, the diagnostic apparatus 100 according to the fourth embodiment uses information regarding at least one of the diagnostic purpose, data conditions, and operability. That is, unlike the above-described embodiment, it is possible to construct a diagnostic model using only operability, diagnostic purpose and operability, or data conditions and operability information.

[Fifth Embodiment]
Next, an operation example of the diagnostic apparatus according to the fifth embodiment of the present invention will be described. The diagnostic apparatus according to the fifth embodiment represents another embodiment of the diagnostic apparatus according to the fourth embodiment. Also in the diagnostic apparatus according to the fifth embodiment, the user inputs the number of clusters and the allowance for the number of clusters through the condition setting screen 20. However, the difference between the diagnostic device according to the fifth embodiment and the diagnostic device according to the fourth embodiment is that the model is modeled before the model building unit 6 operates in the process of step S4 shown in FIG. The construction unit 6 performs cross-validation to estimate the number of clusters generated during normal operation.

Here, a configuration example of the diagnostic apparatus according to the fifth embodiment will be described with reference to FIG.
FIG. 13 is a block diagram showing a configuration example of the diagnostic apparatus 100A according to the fifth embodiment.

The diagnostic device 100A has the same configuration as the diagnostic device 100 shown in FIG. 1, except that a data division unit 12 is provided between the data acquisition unit 4 and the model construction unit 6.
The data division unit 12 divides the time-series data acquired by the data acquisition unit 4 into predetermined numbers. For example, the data division unit 12 divides the data acquired by the data acquisition unit 4 into k sets. Of the k sets of data, the (k-1) set of data is used as learning data, and the remaining one set of data is used as diagnostic data. The data division unit 12 can prepare a combination of such learning data and diagnostic data in advance for k types of learning and diagnosis.

In the process of step S4, the model building unit 6 performs cross-validation before building the diagnostic model based on the divided time-series data for each predetermined number, and the number of clusters generated during normal operation of the diagnosis target. To estimate. Then, the model building unit 6 constructs k different diagnostic models by using the prepared learning data and diagnostic data. Then, in the process of step S6, the evaluation index calculation unit 8 calculates the average value of the evaluation indexes obtained from k different diagnostic models. By such processing, the model building unit 6 of the diagnostic apparatus 100A can pseudo-predict how many clusters will be generated per unit time at the start of operation of the diagnostic model.

After that, the parameter adjusting unit 9 compares the estimated number of clusters with the target cluster number input from the condition input unit 1, and the difference between the estimated number of clusters and the target cluster number is the margin of error. If it is within the range, select the parameter of the condition for which the evaluation index is optimal. At this time, the parameter adjustment unit 9 compares the number of clusters input in advance from the condition input unit 1 with the permissible error with the estimated cluster generation number estimated by the evaluation index calculation unit 8, and sets the conditions within the permissible range. The optimum parameter conditions for the model evaluation index can be selected. The user confirms the predicted number output to the user interface by the result output unit 10, and the user selects the parameter condition or sets the optimum allowable range of the parameter condition for selecting the highly accurate diagnostic model. can do.

As a method of displaying the result displayed on the result output unit 10, only the optimum conditions may be presented. However, other display forms are also conceivable. Here, another form of the result display method displayed on the result output unit 10 will be described with reference to FIG.
FIG. 14 is a diagram showing the evaluation index of the diagnostic model and the number of clusters for each search point.

FIG. 14 shows a diagram in which parameter 1 is taken on the horizontal axis and parameter 2 is taken on the vertical axis, and the evaluation index of the diagnostic model and the predicted value of cluster generation are written together for each search point of the parameter. The predicted value of cluster generation can be, for example, the predicted number of cluster generations per day. For example, at the search points surrounded by the broken line in the figure, the evaluation index is 0.92, which is closer to 1.00 than the other search points. Therefore, it can be determined that the

parameters

1 and 2 indicated by the search points surrounded by the broken line are useful for constructing the diagnostic model.

The user can select a diagnostic model based on arbitrary parameter conditions by checking the screen on which the figure shown in FIG. 14 is displayed. For example, the diagnostic model can be selected under the condition that the evaluation index is the maximum within the range where the amount of work for registering the event is judged to be operationally acceptable for the number of newly generated clusters per day.

Further, in the diagnostic apparatus 100A according to the fifth embodiment, the number of clusters generated by the diagnostic model can be used as one of the evaluation indexes. In this case, the number of generated clusters is evaluated as good as long as it is equal to or less than the upper limit of the number of clusters generated by the user. Therefore, it is possible to use the number of clusters generated as a diagnostic model at the time of diagnosis as one of the evaluation indexes. In addition, when operating the diagnostic model, the number of generated clusters predicted by the diagnostic model during operation can be used as one of the evaluation indexes. Therefore, the evaluation index calculation unit 8 may calculate the number of generated clusters predicted at the time of operation by using the cross validation method. By this process, the number of generated clusters can be predicted accurately.

[Sixth Embodiment]
Next, an operation example of the diagnostic apparatus according to the sixth embodiment of the present invention will be described. Here, a method of registering an evaluation index of a model will be described with reference to FIG. 15, together with a configuration example and an operation example of the diagnostic apparatus according to the sixth embodiment.
FIG. 15 is a block diagram showing a configuration example of the diagnostic apparatus 100B according to the sixth embodiment.

The diagnostic device 100B has the same configuration as the diagnostic device 100 shown in FIG. 1, except that the evaluation index determination unit 3 provides an evaluation index storage database 13 capable of reading the evaluation index. In the evaluation index storage database 13, for example, model evaluation indexes recommended according to diagnostic purposes and data conditions, item priorities and patterns of parameters are organized and stored in the form of a list shown in FIGS. 7 and 8. To.

The evaluation index determination unit 3 reads the evaluation index from the evaluation index storage database 13 based on the diagnostic purpose and data conditions determined by the condition determination unit 2, and displays any item from the evaluation index in a selectable manner. Therefore, when the diagnostic purpose and the data condition are input from the condition determination unit 2 to the evaluation index determination unit 3, the evaluation index determination unit 3 selects the diagnostic purpose and the diagnostic purpose from the evaluation indexes stored in the evaluation index storage database 13. The evaluation index information 14 corresponding to the combination of data conditions is read out. The evaluation index extracted from the evaluation index information 14 is presented to the user on the condition setting screen 20 shown in FIG.

As described above, the information registered in the evaluation index storage database 13 includes not only the evaluation index but also other parameter items that should be set by the user according to a certain diagnostic purpose and data condition. For example, it is assumed that "failure detection" is selected for the diagnostic purpose and "abnormal data is present" is selected for the data condition through the condition setting screen 20. At this time, the condition setting screen 20 is displayed so that the user can select the "F value" as the evaluation index based on the rules registered in the evaluation index storage database 13. At the same time, the condition setting screen 20 is displayed so that the user can select the input data range of "learning data" and the input data range of "correct answer label: normal" and "correct answer label: abnormal".

Next, assume that "Failure detection" is selected for the diagnostic purpose and "No abnormal data" is selected for the data condition. At this time, the condition setting screen 20 is displayed so that the user can select the "false alarm rate" as the evaluation index based on the rules registered in the evaluation index storage database 13, and the input data range of the "learning data". Is displayed so that the user can select.

The operation of the evaluation index determination unit 3 may be selected according to similar condition input results by referring to past model construction examples. Therefore, when similar diagnostic objectives, data conditions and operating conditions are input from past diagnostic models based on utilization rates and evaluations, recommended item priorities and search conditions are patterned, used data range, and detection performance. The result output unit 10 may be able to present a diagnostic model evaluation index including. For example, for each diagnostic model constructed by the model construction unit 6 in the past, the condition determination method is scored based on the number of applications, model accuracy, and artificial evaluation. Then, when the user gives a similar condition input, a screen showing a candidate condition determination method from the top of the scoring is displayed, so that the user can be prompted to select the condition determination method. be.

<Hardware configuration of computer>
Next, the hardware configuration of the computer 30 constituting the diagnostic apparatus according to each embodiment will be described.
FIG. 16 is a block diagram showing a hardware configuration example of the computer 30. The computer 30 is an example of hardware used as a computer that can operate as a diagnostic device according to the present embodiment.

The computer 30 includes a CPU (Central Processing Unit) 31, a ROM (Read Only Memory) 32, and a RAM (Random Access Memory) 33, which are connected to the bus 34, respectively. Further, the computer 30 includes a display device 35, an input device 36, a non-volatile storage 37, and a network interface 38.

The CPU 31 reads the program code of the software that realizes each function according to the present embodiment from the ROM 32, loads it into the RAM 33, and executes it. Variables and parameters generated during the arithmetic processing of the CPU 31 are temporarily written in the RAM 33, and these variables and parameters are appropriately read out by the CPU 31. However, an MPU (Micro Processing Unit) may be used instead of the CPU 31. The functions of the functional units in the

diagnostic devices

100, 100A, and 100B are realized by a program executed by the CPU 31.

The display device 35 is, for example, a liquid crystal display monitor, and displays the result of processing performed by the computer 30 to the user. The result output unit 10 can output the result to the display device 35. Therefore, the condition setting screen 20 and the like shown in FIG. 3 are displayed on the display device 35. For example, a keyboard, a mouse, or the like is used as the input device 36, and the user can perform predetermined operation inputs and instructions. The condition input unit 1 can take in the information input through the input device 36.

As the non-volatile storage 37, for example, an HDD (Hard Disk Drive), SSD (Solid State Drive), flexible disk, optical disk, optical magnetic disk, CD-ROM, CD-R, magnetic tape, non-volatile memory, or the like is used. Be done. In the non-volatile storage 37, in addition to the OS (Operating System) and various parameters, a program for operating the computer 30 is recorded. The ROM 32 and the non-volatile storage 37 record programs, data, and the like necessary for the CPU 31 to operate, and as an example of a computer-readable non-transient storage medium that stores a program executed by the computer 30. Used. The measured value database 5 shown in FIGS. 1, 13, and 15 and the evaluation index storage database 13 shown in FIG. 15 are configured in the non-volatile storage 37.

For the network interface 38, for example, a NIC (Network Interface Card) or the like is used, and various data can be transmitted and received between the devices via a LAN (Local Area Network) connected to the terminal of the NIC, a dedicated line, or the like. It is possible. For example, the measured value data (time series data) output from the measuring instrument 2 is taken into the measured value database 5 through the network interface 38.

[Modification example]
The present invention is not limited to each of the above embodiments as it is, and each embodiment can be modified or omitted within the scope of the invention.

For example, each embodiment of the present invention includes, for example, a power generation plant (boiler, condenser, turbine), a chemical plant (reaction tank, heat exchanger), an industrial plant (filtration filter), and a water treatment facility (aggregator). It can be used to diagnose the operation of industrial equipment such as injection tanks) and small boilers.

In addition, each of the above-described embodiments is a detailed and specific description of the configurations of the apparatus and the system in order to explain the present invention in an easy-to-understand manner, and is not necessarily limited to those including all the described configurations. Further, it is possible to replace a part of the configuration of the embodiment described here with the configuration of another embodiment, and further, it is possible to add the configuration of another embodiment to the configuration of one embodiment. It is possible. Further, it is also possible to add, delete, or replace a part of the configuration of each embodiment with another configuration.
In addition, the control lines and information lines indicate those that are considered necessary for explanation, and do not necessarily indicate all the control lines and information lines in the product. In practice, it can be considered that almost all configurations are interconnected.

1 ... Condition input unit, 2 ... Condition determination unit, 3 ... Evaluation index determination unit, 4 ... Data acquisition unit, 5 ... Measurement value database, 6 ... Model construction unit, 7 ... Abnormality judgment unit, 8 ... Evaluation index calculation unit, 9 ... Parameter adjustment unit, 10 ... Result output unit, 11 ... Time series data, 100 ... Diagnostic device

Claims

A condition input unit for inputting data conditions of time-series data representing the state of the diagnosis target, which is used for the diagnostic purpose of the diagnosis model for diagnosing the state of the diagnosis target and for constructing the diagnosis model.
Parameter items in which parameters for adjusting the diagnostic model are defined based on the input diagnostic purpose and data conditions, items in the range of use of the time-series data, and evaluation indexes for evaluating the diagnostic model. A diagnostic apparatus comprising: a parameter adjusting unit constructed according to a modeling method including at least one of the items and automatically adjusting the parameters of the diagnostic model evaluated by the evaluation index.
The parameter item includes an item priority for determining the priority of the parameter item searched by the parameter adjustment unit, and a patterning item for aggregating the search range in the parameter item into a predetermined number of patterns.
The items of the usage range of the time-series data include a learning period required for constructing the diagnostic model, an evaluation period in which a normal period and an abnormal period are separately set, and an abnormal period in which an abnormality or deterioration occurs in the diagnosis target.
The diagnostic device according to claim 1, wherein the item of the evaluation index includes a detection performance indicating the performance of the diagnostic model to detect an abnormality of the diagnosis target.
A condition determination unit that determines the contents of the parameter item and the item of the usage range of the time series data based on the diagnostic purpose and the data condition input from the condition input unit.
An evaluation index determination unit that determines a specific evaluation index defined by the item of the evaluation index based on the diagnostic purpose and the data condition input from the condition input unit.
A data acquisition unit that acquires the time-series data according to the usage range of the time-series data,
A model building unit that builds the diagnostic model from the time series data according to the modeling method.
An abnormality determination unit for determining the presence or absence of an abnormality in the time-series data acquired during the evaluation period, which is input to the diagnostic model constructed by the model construction unit, and
An evaluation index calculation unit that calculates the specific evaluation index based on the determination result of the presence or absence of abnormality in the time series data.
The diagnostic apparatus according to claim 2, further comprising a result output unit that outputs the calculated evaluation index.
When the case where the time-series data has abnormal data is selected as the data condition, the data acquisition unit attaches information indicating that the time-series data acquired during the period when the correct answer is known is normal. However, information indicating that the abnormality is added to the time-series data acquired during the period when the abnormality is known is added.
A claim that the model building unit builds the diagnostic model using the time-series data with information indicating that it is normal and the time-series data with information indicating that it is abnormal. 3. The diagnostic device according to 3.
When the case where there is no abnormal data in the time-series data is selected as the data condition, the data acquisition unit attaches information indicating that the time-series data acquired during the period when the correct answer is known is normal. death,
According to claim 4, the parameter adjusting unit adjusts the parameter based on the false alarm rate that the abnormality determining unit erroneously determines as abnormal with respect to the time-series data to which the information indicating that the condition is normal is attached. The diagnostic device described.
The diagnostic device according to claim 4, wherein the parameter adjusting unit adjusts the parameters based on the reproducibility selected as the evaluation index when the diagnostic purpose is to detect deterioration of the diagnostic object.
An operability item including at least one item among the target number of clusters and the tolerance is added to the condition input unit.
The diagnostic device according to claim 4, wherein the condition determination unit determines the modeling method based on the diagnostic purpose, the data condition, and the operability item.
A data division unit that divides the time-series data acquired by the data acquisition unit into predetermined numbers is provided.
The model building unit performs cross-validation before building the diagnostic model based on the time-series data for each divided predetermined number, and estimates the number of clusters generated during normal operation of the diagnostic target. ,
The parameter adjusting unit compares the estimated number of clusters with the target cluster number input from the condition input unit, and the difference between the estimated number of clusters and the target cluster number is the difference. The diagnostic device according to claim 7, wherein the parameter is selected under the condition that the evaluation index is optimal when the tolerance is within the range.
Equipped with an evaluation index storage database that stores evaluation indexes
The evaluation index determination unit reads the evaluation index from the evaluation index storage database based on the diagnostic purpose and the data condition determined by the condition determination unit, and displays an arbitrary item selectable from the evaluation index. The diagnostic device according to claim 4.
The purpose of diagnosing the diagnostic model for diagnosing the state of the diagnostic target, and the process of inputting the data conditions of the time-series data used to construct the diagnostic model.
Parameter items in which parameters for adjusting the diagnostic model are defined based on the input diagnostic purpose and data conditions, items in the range of use of the time-series data, and evaluation indexes for evaluating the diagnostic model. A parameter adjustment method including a process of automatically adjusting the parameters of the diagnostic model constructed according to a modeling method including at least one of the items and evaluated by the evaluation index.