JP2015018389A - Abnormality sign diagnosis device and abnormality sign diagnosis method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an abnormality sign diagnosis device and an abnormality sign diagnosis method for appropriately diagnosing the presence/absence of an abnormality sign even when the external environment of a mechanical apparatus largely varies.SOLUTION: An abnormality sign diagnosis device 1 includes: sensor data acquisition means 11 for acquiring multi-dimensional sensor data measured by a plurality of sensors installed in a mechanical apparatus 2; a data mining part 16 for, about diagnosis object data, diagnosing the presence/absence of an abnormality sign on the basis of the degree of abnormality indicating the degree of deviation from a normal model generated by learning using the sensor data when the mechanical apparatus 2 normally operates; and a remote monitoring part 17 for diagnosing the presence/absence of an abnormality sign on the basis of whether or not the value of individual sensor data falls within a predetermined range. The data mining part 16 reconstructs the normal model by executing learning in an initialization mode when the state of an external environment largely varies in accordance with an instruction by initialization instruction means 15.

Description

  The present invention relates to an abnormality sign diagnosis apparatus and an abnormality sign diagnosis method for diagnosing the presence / absence of an abnormality sign based on sensor data from mechanical equipment.

Mechanical equipment and system equipment composed of various devices are required to operate stably and normally. However, various troubles occur in mechanical equipment and system devices due to load fluctuations based on changes in operating conditions and other factors.
Normally, these devices are provided with various sensors for the purpose of grasping the operation state and detecting an abnormality, and monitoring the operation state. Furthermore, recent advances in IT equipment enable the storage of enormous operation record data and sensor data, and these data are stored in many equipments and systems.
In addition, for example, in a power generation system or the like, it is required to operate for 24 hours without stopping, and for that purpose, periodic inspection and maintenance for preventing the occurrence of a failure are indispensable. Furthermore, if a sign before the occurrence of an abnormal state can be detected, an early response can be made by analyzing sensor data before and after the sign.

Conventionally, two methods have been proposed as methods for detecting an anomaly sign. The first method is predictive abnormality diagnosis by a remote monitoring method using threshold setting. That is, a threshold value is set in advance for each sensor data, and when the sensor data value exceeds a predetermined threshold value, it is determined that there is a sign of abnormality.
The second method of detecting an abnormal sign is an abnormal sign diagnosis by a data mining method using a statistical classification method.

In the remote monitoring method described above, when detecting an abnormal sign, there are a small number (for example, one type) of sensors exceeding the threshold value. However, there are many complex causes of abnormalities that occur in complex machinery and equipment. For example, even if the failure items are the same, there are various processes leading to the failure, and it is not always possible to set an optimal threshold value. There is a problem.
Further, in the above-described data mining method, for example, when maintenance is performed on mechanical equipment, the system conditions may change greatly. Even in such a case, if a diagnosis based on data mining is performed using the accumulated past data as it is, there is a problem that a diagnosis is made as a sign of abnormality even though the mechanical equipment itself is normal.

Patent Document 1 discloses an abnormality sign diagnosis apparatus including a diagnosis unit using a remote monitoring method and a diagnosis unit using a data mining method. In this abnormality sign diagnosis device, when maintenance is performed, sensor data is accumulated for a certain period, and a normal model for diagnosis by a data mining method is generated using sensor data newly accumulated after maintenance is performed, Resume diagnosis using data mining techniques. Therefore, after the fixed period has elapsed, diagnosis using a data mining technique can be performed appropriately.
Further, while the sensor data is accumulated for a certain period, the diagnosis by the data mining technique is interrupted, but the abnormality sign diagnosis can be continued by the diagnosis by the remote monitoring technique that does not require data accumulation.

Japanese Patent No. 5081998

In the abnormality sign diagnosis apparatus described in Patent Document 1, the abnormality sign diagnosis can be appropriately continued even when maintenance is performed by combining diagnosis by the remote monitoring technique and diagnosis by the data mining technique. .
However, if the target machine / equipment for abnormal sign diagnosis is installed in an environment that is affected by large fluctuations such as temperature and humidity due to seasonal changes, it should be used for diagnosis by the data mining method. If the external environment changes drastically while a normal model that has been used has been used for a long time without being reconstructed from the time it was generated, it may not be possible to properly diagnose abnormal signs using the normal model. I understood.

Here, there is a method of generating a model including fluctuations in the external environment, but in this case, it is necessary to store a huge amount of data in advance, including when the external environment fluctuates. It will take a long time to operate. For example, when a new machine facility is introduced, an abnormal sign diagnosis cannot be performed until data for one year for building a model corresponding to a seasonal change is accumulated. Even if the same type of machinery and equipment is installed, it is built for machinery and equipment installed in other regions if the external environment is significantly different, such as differences in installation locations, for example, differences between Hokkaido and Okinawa. It may not be appropriate to divert the model.
It is also conceivable to reconstruct a normal model each time sensor data measured every day is acquired. However, regardless of the magnitude of the change in the state of the external environment, frequently restructuring the normal model unnecessarily increases the processing load for learning.

  Therefore, an object of the present invention is to provide an abnormality sign diagnosis apparatus and an abnormality sign diagnosis method that appropriately diagnose the presence or absence of an abnormality sign even when the external environment largely fluctuates.

  In order to solve the above problems, an abnormality sign diagnosis apparatus according to the present invention uses a sensor data detected by a plurality of sensors installed in a mechanical facility to diagnose the presence or absence of an abnormal sign in the mechanical facility. A sensor data acquisition means for acquiring sensor data from the plurality of sensors, a sensor data storage means for storing the sensor data acquired by the sensor data acquisition means, and a sensor data acquisition means. Specifying a period during which the sensor data was acquired, learning using a learning target data acquisition process for acquiring sensor data corresponding to the specified period from the sensor data storage unit, and learning using sensor data corresponding to the specified period Learning processing for generating a normal model indicating a normal range of the sensor data, and based on the normal model A first diagnosis means for diagnosing the presence or absence of an abnormality sign of the mechanical equipment, and an initial stage for restructuring the whole normal model in accordance with a change in the external environment that is an environment around the mechanical equipment Initialization instruction means for instructing the first diagnosis means to perform the learning process in the activation mode.

  According to the present invention, it is possible to provide an abnormality sign diagnosis apparatus and an abnormality sign diagnosis method that appropriately diagnose the presence or absence of an abnormality sign even when the external environment changes greatly.

1 is a block diagram showing a system configuration including an abnormality sign diagnosis apparatus according to an embodiment of the present invention. It is a block diagram which shows the structure of the data mining part in embodiment of this invention. It is explanatory drawing which shows the relationship between the learning period containing the learning object data acquired by the learning object data acquisition means in embodiment of this invention, and a diagnosis date. In embodiment of this invention, when a maintenance operation | work is performed, it is explanatory drawing which shows the relationship between the learning period containing the learning object data acquired by the learning object data acquisition means, and a diagnosis date. It is explanatory drawing of learning of the normal model by the data mining learning part in embodiment of this invention. It is a figure which shows the example of the learning result by the data mining learning part in embodiment of this invention. It is explanatory drawing for demonstrating the concept of the relationship between the change of the normal model learned by the data mining learning part in embodiment of this invention, and the change of an external environment. It is a figure which shows the example of the diagnostic result by the data mining diagnostic part in embodiment of this invention. It is a block diagram which shows the structure of the remote monitoring part in embodiment of this invention. It is explanatory drawing of the diagnosis of the abnormal sign by the separate determination means of a remote monitoring part. It is a flowchart which shows the learning process by the data mining learning part in embodiment of this invention. In embodiment of this invention, it is a flowchart which shows the learning target data acquisition process by the learning target data acquisition means after performing a maintenance operation | work. In the embodiment of the present invention, it is a flowchart showing learning object data acquisition processing by the learning object data acquisition means when an initialization instruction is given. It is a flowchart which shows the diagnostic process by the data mining diagnostic part in embodiment of this invention.

  Embodiments of the present invention will be described in detail with reference to the drawings as appropriate. In each figure, common portions are denoted by the same reference numerals, and redundant description is appropriately omitted.

<Embodiment>
FIG. 1 is a system configuration diagram including an abnormality sign diagnosis apparatus according to an embodiment of the present invention.
The abnormality sign diagnosis device 1 is a device that diagnoses whether there is an abnormality sign in the machine facility 2 based on measurement values acquired from a plurality of sensors installed in the machine facility 2. Here, the “abnormal sign” indicates that there is a high possibility that an abnormal state will occur in the mechanical equipment 2 in the near future (for example, within approximately two weeks) with reference to the diagnosis date. Further, “abnormal” indicates, for example, a situation in which a malfunction (failure) of the operation is clearly generated in the mechanical equipment 2 and the mechanical equipment 2 is not operating normally.
Note that the boundary for distinguishing “abnormality sign” from “abnormality” is not necessarily clear. For example, a scale indicating the degree of abnormality (an abnormality degree described later) is appropriately set, and the value of the scale is The two may be distinguished depending on whether or not a predetermined threshold value is exceeded.

As shown in FIG. 1, the abnormality sign diagnosis apparatus 1 according to the present embodiment is connected to at least one mechanical facility 2 via a communication network N. Usually, there are a plurality of machine facilities 2 (for example, 50), but even in such a case, in the following description, “machine facilities 2” will be collectively described. The mechanical equipment 2 indicates, for example, gas engines and gas turbines used in power plants, nuclear reactors in nuclear power plants, windmills in wind power plants, and various devices used in chemical plants. Moreover, the mechanical equipment 2 is not limited to the one where the installation location is fixed, and may be a moving body such as an automobile or a forklift.
In addition, in order to measure temperature, pressure, voltage, current, motor rotation speed, etc., a plurality of (not shown) sensors (for example, 30 sensors) are installed in the mechanical equipment 2, and a certain sample period ( For example, each measurement value is acquired every 30 seconds. In the transient state of the mechanical equipment 2, it is preferable to set the sample period shorter.

The machine facility 2 transmits the measured values acquired by the various sensors described above to the abnormality sign diagnosis apparatus 1 via the communication network N.
Further, each machine facility 2 is provided with a plurality of interlocks (not shown) as safety instrumentation in each machine facility 2 itself, in addition to being subjected to an abnormality sign diagnosis by the abnormality sign diagnostic apparatus 1. . When a predetermined state is reached (for example, when the fuel is low), alarm information or failure information from the interlock is transmitted to the abnormality sign diagnosis apparatus 1 via the communication network N. It is like that.

In addition, maintenance work may be performed on the mechanical equipment 2. The maintenance work may be performed as regular maintenance, or alarm information or failure information from an interlock (not shown) installed in the mechanical equipment 2 itself, or diagnosis of an abnormal sign by the abnormal sign diagnosis device 1 It may be performed as maintenance corresponding to In addition, the maintenance work includes, for example, an operation for supplying cooling water from the outside when the mechanical equipment 2 is a gas engine.
In other words, “maintenance work” means an artificial operation including maintenance on the mechanical equipment 2. Below, the case where the maintenance with respect to the mechanical installation 2 is performed is demonstrated as an example of a maintenance operation | work.

  In addition, as shown in FIG. 1, the abnormality sign diagnosis apparatus 1 is connected to a computer 3 via a communication network N. The computer 3 includes input means, control means, storage means, display means, communication means, etc. (not shown). Then, the worker who has grasped the content of the maintenance (maintenance information) for each mechanical facility 2 can use the input means (keyboard or the like) of the computer 3 to perform identification information and maintenance work period of the mechanical facility 2 in which the maintenance operation is performed. Register maintenance information including

The maintenance information input to the computer 3 by the administrator when performing maintenance on the mechanical equipment 2 includes, for example, the following.
First, as “registration basic information”, a subject name, a correspondence type, an identification number for identifying the machine facility 2, an installation site name, and the like are input.

In addition, as the “defect information”, the date, state, error item (failure occurrence location), and the like when a defect such as a failure occurs in the mechanical equipment 2 are input.
In addition, as the “maintenance work period”, a period during which maintenance is to be performed on the machine facility 2 in which a failure has occurred is input. Here, the maintenance work period input to the computer 3 is information indicating a period during which maintenance is to be performed on the mechanical equipment 2 in the future (the time when the maintenance information is input is assumed to be present).

  Note that the maintenance work period input to the computer 3 may be information that the maintenance has been performed on the machine facility 2 in a predetermined period in the past (the current time point when the maintenance information is input). In the latter case, the abnormality sign diagnosis apparatus 1 again executes the diagnosis of the abnormality sign based on the information that the maintenance has been performed in the past predetermined period.

  In addition, as “replacement parts”, information for identifying the parts of the mechanical equipment 2 replaced during maintenance is input. Further, information indicating whether or not the maintenance work has been completed is input as information indicating whether or not the maintenance work has been completed. Further, as the “treatment content”, in addition to the maintenance content, the state of the mechanical equipment 2 before and after the maintenance is input. Maintenance information input to the computer 3 is stored in storage means (not shown) of the computer 3 by control means (not shown), and further transmitted to the abnormality sign diagnosis apparatus 1 via the communication network N.

<Configuration of abnormal sign diagnosis device>
As shown in FIG. 1, the abnormality sign diagnostic apparatus 1 includes a communication unit 10, a sensor data acquisition unit 11, a sensor data storage unit 12, a maintenance information acquisition unit 13, a maintenance information storage unit 14, and an initialization instruction. Means 15, data mining unit 16, remote monitoring unit 17, diagnostic result storage unit 18, display control unit 19, and display unit 20 are provided.
The communication means 10 receives data from the mechanical equipment 2 and the computer 3 via the communication network N, and is constituted by, for example, a router or various interfaces. The communication means 10 receives the data via the communication network N, for example, according to a TCP / IP communication protocol.

  The sensor data acquisition unit 11 acquires sensor data from various sensors (not shown) installed in the machine facility 2 among the data input from the communication unit 10 via the communication network N. As described above, the various sensors installed in the mechanical equipment 2 measure predetermined feature amounts (for example, temperature and pressure values) in the mechanical equipment 2. The mechanical equipment 2 acquires the measurement values obtained by the various sensors every predetermined time (for example, 30 seconds), and sequentially transmits them to the abnormality sign diagnosis apparatus 1 via the communication network N. In this case, the sensor data acquisition unit 11 acquires 2880 measurement values from the various sensors of the machine facility 2 per day.

In addition to the measurement value, the data input from the communication means 10 via the communication network N includes the data type of the measurement value detected by the sensor (the identification number of the mechanical equipment 2 and the sensor identification). Including the date) and the date and time when the measured value was acquired in the mechanical equipment 2. In addition, as data input from the communication unit 10, a predetermined command value (for example, a motor rotation speed command value) in the mechanical equipment 2 may be included.
Hereinafter, the measurement value and the data added to specify the measurement value are collectively referred to as “sensor data”. The sensor data acquisition unit 11 stores the sensor data acquired through the communication unit 10 in the sensor data storage unit 12.

  The sensor data storage unit 12 stores the sensor data acquired by the sensor data acquisition unit 11. That is, the sensor data storage unit 12 stores sensor data as a database for each site (not shown) that manages the machine facility 2 or at least one machine facility 2. The sensor data storage unit 12 is configured to be able to output the stored sensor data to a data mining unit 16, a remote monitoring unit 17, and a display control unit 19, which will be described later.

  The maintenance information acquisition unit 13 acquires maintenance information output from the computer 3 out of each data input from the communication unit 10 via the communication network N. The maintenance information acquired by the maintenance information acquisition unit 13 includes at least an identification number for specifying the machine equipment 2 and a period during which maintenance work is performed on the machine equipment 2 (maintenance work period). . The maintenance information acquisition unit 13 identifies whether the data acquired through the communication unit 10 is maintenance information, and stores the acquired maintenance information in the maintenance information storage unit 14.

  The maintenance information storage unit 14 stores the maintenance information acquired by the maintenance information acquisition unit 13. The maintenance information storage unit 14 stores, as a database, maintenance information including at least for which machine equipment 2 maintenance work has been performed for which period. The maintenance information storage unit 14 is configured to be able to output the stored maintenance information to a data mining unit 16 that will be described later and the display control unit 19.

  The initialization instruction means 15 outputs an initialization instruction signal for instructing initialization to the data mining unit 16 according to a change in the external environment. In this embodiment, as the contents of the initialization instruction by the initialization instruction signal, an instruction of “initialization learning” for executing learning processing in an initialization mode for reconstructing the entire normal model, and a normal model are determined in advance. There is an instruction of “initial value setting” for setting the initial value.

Here, the predetermined initial value is, for example, a set value determined as a default value of a normal model at the time of factory shipment. In addition, the initial value is not limited to one set value. For example, a plurality of set values are prepared in advance so as to cope with various climates and seasons in which the mechanical equipment 2 is installed, and the initial value setting is instructed. You may make it selectable when doing.
Note that the initialization instruction signal for instructing the initial value setting is mainly used for an artificial operation by an administrator when, for example, the machine facility 2 is newly installed or moved to a place where the external environment is greatly different. Based on the output.

  An initialization instruction signal for instructing execution of the learning process in the initialization mode is mainly output when the state of the external environment changes greatly. For this reason, the initialization instruction means 15 can be configured to output an initialization instruction signal manually or automatically in response to a change in the state of the external environment.

  Here, outputting the initialization instruction signal manually is based on an artificial input operation by an administrator. That is, for example, when an administrator performs an input operation intended for an initialization instruction using an input unit (not shown) of the computer 3, the input operation is performed via the communication unit 10. The signal shown is input to the initialization instruction means 15. When the initialization instruction unit 15 detects an input of a signal indicating that the input operation has been performed, the initialization instruction unit 15 outputs an initialization instruction signal to the data mining unit 16.

In addition, the automatic output of the initialization instruction signal means that the initialization instruction means 15 automatically outputs the initialization instruction signal when, for example, the temperature change or the elapsed time meets a predetermined condition. is there. Specifically, for example, the timing for outputting the initialization instruction signal can be determined as follows.
The initialization instruction unit 15 acquires, via the communication unit 10, a measured value of a sensor (not shown) that detects the state of the external environment such as the ambient temperature around the machine facility 2. Then, when the average daily temperature changes to a predetermined value or more compared to when the normal model is initialized last time, the output timing of the initial instruction signal can be set.
As another example, the initialization instruction unit 15 acquires calendar information via, for example, the communication unit 10 and initializes the normal model last time. When the time elapses, the output timing of the initial instruction signal can be set.

  The initialization instruction means 15 may be configured to perform both the manual initialization instruction signal output and the automatic initialization instruction signal output. Thereby, it is possible to cause the data mining unit 16 to execute the normal model learning process in the initialization mode at a more appropriate timing.

The data mining unit (first diagnostic means) 16 performs data mining by applying a statistical data classification method, and learns using normal data of the mechanical equipment 2 to generate a normal model for each mechanical equipment 2. Then, based on the normal model, the presence or absence of an abnormality sign of the mechanical equipment 2 is diagnosed. Further, the data mining unit 16 acquires sensor data from the sensor data storage unit 12 based on the maintenance information acquired from the maintenance information storage unit 14. The data mining unit 16 performs learning and diagnosis based on the acquired sensor data, and stores the diagnosis result in the diagnosis result storage unit 18.
In addition, when the initialization instruction signal is input from the initialization instruction unit 15, the data mining unit 16 initializes the normal model according to the contents of the initialization instruction.
Details of the data mining unit 16 will be described later.

The remote monitoring unit (second diagnostic means) 17 operates the sensor data value (or amount of change) acquired from the machine facility 2 based on the experience and knowledge of the administrator who performs various settings of the abnormality sign diagnostic apparatus 1. A threshold value is provided, and an abnormality sign is diagnosed when the sensor data value falls outside the range of the upper and lower threshold values. The remote monitoring unit 17 stores the diagnosis result in the diagnosis result storage unit 18.
Details of the remote monitoring unit 17 will be described later.
The abnormality sign diagnosis apparatus 1 independently performs diagnosis by the data mining unit 16 and diagnosis by the remote monitoring unit 17.

  The diagnosis result storage means 18 stores the diagnosis result by the data mining unit 16 and the diagnosis result by the remote monitoring unit 17 in separate storage areas (not shown). The diagnosis result storage means 18 stores each diagnosis result on the diagnosis date for each machine facility 2 (or a site that manages the machine facility 2), and is configured to be output to the display control means 19. .

  The display control unit 19 performs control for causing the display unit 20 to display the diagnosis result by the data mining unit 16 and the diagnosis result by the remote monitoring unit 17 stored in the diagnosis result storage unit 18. That is, the display control means 19 outputs a control signal necessary for displaying the diagnosis result by the abnormality sign diagnosis apparatus 1 on the display means 20.

For example, the display control unit 19 displays the diagnosis result (presence / absence of abnormality) by the data mining unit 16 in a matrix format with each machine facility 2 as a row and the diagnosis date as a column. In addition, the display control means 19 can display the date when maintenance was performed, the learning period thereafter, and the like (details will be described later) on the display means 20 in an identifiable manner. Moreover, the display control means 19 displays the presence or absence of alarm information or failure information from an interlock (not shown) installed in the mechanical equipment 2 itself on the display means 20 so as to be identifiable.
Further, when the remote monitoring unit 17 diagnoses that there is a sign of abnormality, the display control unit 19 can also display the diagnosis result on the display unit 20 in a pop-up manner, for example.
The display unit 20 is, for example, a monitor, and displays a diagnosis result or the like by the abnormality sign diagnosis device 1 according to a control signal input from the display control unit 19.

<Configuration of data mining unit>
Next, the configuration of the data mining unit 16 will be described with reference to FIG. 2 (refer to FIG. 1 as appropriate). As shown in FIG. 2, the data mining unit 16 includes a data mining learning unit 161 and a data mining diagnostic unit 162.

The data mining learning unit 161 performs data mining by applying clustering, which is a kind of statistical data classification method, and learns using normal data that is sensor data acquired when the machine facility 2 is operating normally. By doing so, a normal model is generated.
For this purpose, the data mining learning unit 161 includes a learning target data acquisition unit 161a, a learning target data storage unit 161b, a normal model generation unit 161c, and a normal model data storage unit 161d.
Details of the target indicated by the “normal model” will be described later.

  In the present embodiment, the learning process by the data mining learning unit 161 is performed once a day in advance, except for a certain period during maintenance work and after maintenance, and when initialization is instructed by the initialization instruction unit 15. It is executed at a set time (for example, midnight). A diagnosis process performed by the data mining diagnosis unit 162 described later is performed once a day with reference to a normal model generated by the learning process after the learning process is completed.

The learning target data acquisition unit 161 a acquires sensor data from the sensor data storage unit 12 based on the maintenance information stored in the maintenance information storage unit 14 and the initialization instruction signal input from the initialization instruction unit 15. The acquired sensor data is stored in the learning object data storage unit 161b. Hereinafter, the sensor data acquired by the learning target data acquisition unit 161a from the sensor data storage unit 12 is referred to as “learning target data”.
Details of the learning target data acquired by the learning target data acquiring unit 161a will be described later.

The learning target data storage unit 161b is a device for storing the learning target data acquired by the learning target data acquisition unit 161a. The learning target data storage unit 161b is configured to be able to output the stored learning target data to the normal model generation unit 161c.
Note that the learning target data acquired by the learning target data acquisition unit 161a may be directly output to the normal model generation unit 161c without providing the learning target data storage unit 161b.

The learning target data acquired by the learning target data acquiring unit 161a is provided with, for example, an operation mode dividing unit (not shown) subsequent to the learning target data acquiring unit 161a, and is divided for each “driving mode”. You may make it memorize | store in the data storage means 161b.
Here, the “operation mode” means an operation mode of the mechanical equipment 2. The operation mode is a mode in which the mechanical equipment 2 is stopped and in a steady state, a mode in which the mechanical equipment 2 is activated and in a transient state, a mode in which the mechanical equipment 2 is in operation and in a steady state, A mode in which the mechanical equipment 2 is stopped and is in a transient state is exemplified.
Then, the normal model is generated by learning using the learning target data corresponding to each operation mode by the normal model generation unit 161c. Thereby, the normal model which shows a normal range more appropriately can be produced | generated by the normal model production | generation means 161c.
In addition, when generating a normal model for each operation mode, the data mining diagnosis unit 162 described later also divides diagnosis target data described later for each operation mode, and performs diagnosis using the corresponding normal model. It is possible to diagnose abnormal signs with high accuracy.

  The normal model generation unit 161c generates a normal model of the machine facility 2 by learning using the learning target data acquired by the learning target data acquisition unit 161a and stored in the learning target data storage unit 161b. . The learning by the normal model generation unit 161c is performed by performing data mining by applying clustering, which is a kind of statistical data classification method.

For the learning performed by the normal model generation unit 161c, an initialization mode in which the entire normal model is newly constructed using the sensor data accumulated in a predetermined period and the sensor data accumulated every day are additionally applied to the normal model. By doing so, there are two learning modes: an update mode in which some parameters of the normal model are tuned daily.
Furthermore, the normal model generation unit 161c has a function of generating the normal model by setting an initial value that is predetermined as the normal model. The setting of the initial value is a function used when valid learning target data cannot be acquired, for example, when the machine equipment 2 is newly installed.
The normal model generation unit 161c stores the generated normal model in the normal model data storage unit 161d.
As shown in FIG. 2, the normal model generation unit 161c includes an initial value setting unit 161c1, an initialization learning unit 161c2, and an update learning unit 161c3.

The initial value setting means 161c1 is an initial value predetermined for the normal model (for example, a factory default value) when the instruction content of the initialization instruction signal input from the initialization instruction means 15 is "initial value setting". Value). The initial value data used for the setting includes, for example, storage means (not shown) such as a ROM (Read Only Memory), a Flash-RAM (Radom Access Memory), and a magnetic disk in the initial value setting means 161c1. It is assumed that the storage means stores the information in advance.
The initial value setting unit 161c1 stores the normal model generated by setting the initial value in the normal model data storage unit 161d.

The initialization learning unit 161c2 performs learning in the initialization mode when the instruction content of the initialization instruction signal input from the initialization instruction unit 15 is "initialization learning" and when initial learning after maintenance is performed. A normal model is generated. More specifically, the initialization learning unit 161c2 acquires the learning target data accumulated in a predetermined period from the learning target data storage unit 161b, and learns using the learning target data, so that the entire normal model can be obtained. To rebuild.
The initialization learning unit 161c2 stores the reconstructed normal model in the normal model data storage unit 161d.

  The update learning unit 161c3 performs learning in the update mode and updates the normal model every day after the normal model is initialized. More specifically, the update learning unit 161c3 updates the normal model by daily tuning (adjusting) some parameters of the normal model by additionally applying the sensor data accumulated every day to the normal model. Is. Learning in the update mode only adjusts some of the parameters of the cluster, which will be described later, so that it can exhibit a certain adaptive capability against changes in the state of the external environment without greatly increasing the processing burden. it can.

The normal model data storage unit 161d is a device for storing the normal model generated by the normal model generation unit 161c. The normal model data stored in the normal model data storage unit 161d is configured to be output to the abnormality degree calculation unit 162c of the data mining diagnosis unit 162, and is used for diagnosis of an abnormal sign by a data mining technique.
The configuration of normal model data stored in the normal model data storage unit 161d will be described later.

Here, the learning process performed by the normal model generation unit 161c will be described with reference to FIG.
(Learning process by initialization mode)
First, the learning process in the initialization mode will be described.
The learning process in the initialization mode is performed when the learning is instructed by the initialization instructing unit 15 or when the sensor data for a predetermined period is accumulated and learned first after the maintenance is performed. And executed.

The normal model generation unit 161c classifies the learning target data into several representative groups called clusters for each similar data. In the example shown in FIG. 5, the case of one cluster is shown. However, a plurality of clusters may exist for one operation mode of a specific machine facility 2.
In addition, the feature α, feature β, and feature γ shown in FIG. 5 are, for example, normalized feature amounts (temperature, pressure, rotational speed, etc.) measured by a sensor (not shown) installed in the mechanical equipment 2. It is a thing. FIG. 5 shows a three-dimensional case as an example, but it is assumed that there are as many dimensions as the number of sensors acquired by the sensor data acquisition unit 11 (see FIG. 1).

  As a clustering method, for example, a non-hierarchical clustering k-means method can be used. In the k-average method, first, clusters are randomly assigned to each data, and the center of each cluster (cluster center) is calculated based on the assigned data. As the cluster center, for example, the average of each element of the allocated data can be used. Next, the distance between the predetermined data and the cluster center is obtained, and the data is assigned to the nearest center cluster. If the cluster assignment of each data does not change for all data in such a process, the process ends. Otherwise, the cluster center is recalculated from the newly allocated cluster and the above process is repeated. Through the above processing, the normal model generation unit 161c can classify the learning target data into a plurality of clusters.

  Then, as shown in FIG. 5, a cluster center c and a cluster radius r are determined for each cluster. The cluster center c and the cluster radius r obtained for each cluster are normal models learned by the normal model generation unit 161c. The normal model generation unit 161c stores the learned normal model in the normal model data storage unit 161d (see FIG. 2).

  The normal model generation unit 161c is set to learn a normal model when there is a predetermined number or more of learning target data stored in the learning target data storage unit 161b. That is, the normal model generation unit 161c learns the normal model when it is determined that the learning target data acquisition unit 161a has not acquired more than the minimum number of pieces of learning target data necessary for learning the normal model. First, information indicating that the diagnosis has not been performed is stored in the diagnosis result storage unit 18 (see FIG. 2).

(Configuration of normal model data)
Here, a configuration example of normal model data, which is a learning result by the data mining learning unit, will be described with reference to FIG.
As shown in FIG. 6, the normal model data storage means 161d stores a normal model as a database for each operation mode. A cluster number 101 shown in FIG. 6 is an identification number assigned to identify each cluster. Sensors 1 to M (102) indicate sensors (not shown) installed in the machine facility 2.
The number of belonging data 103 shown in FIG. 6 indicates the number of learning target data belonging to each of the clusters 1 to N. The total error 104 indicates the total error (distance) between the cluster center of each cluster and each learning target data belonging to the cluster. The maximum error 105 is the maximum value of the distance between the cluster center of a specific cluster and a plurality of learning target data belonging to the cluster. The minimum error is a minimum value of the distance between the cluster center of a specific cluster and a plurality of learning target data belonging to the cluster.

  The maximum value 107 indicates the maximum value of the feature component (for example, the coordinate value of the feature α shown in FIG. 5) corresponding to each sensor in the plurality of learning target data included in the specific cluster. The minimum value 108 indicates the minimum value of the feature component corresponding to each sensor in the plurality of learning target data included in the specific cluster. The normal model generation unit 161c performs normalization processing using the maximum value 107 and the minimum value 108. The cluster center 109 indicates the values of the sensors 1 to M that specify the center of each cluster.

(Learning process by update mode)
Next, the learning process in the update mode will be described with reference to FIGS. 5 and 6 (see FIG. 2 as appropriate).
The learning process in the update mode is executed once a day after the day after the learning in the initialization mode.

As described above, the learning process in the update mode updates the normal model by daily tuning (adjusting) some parameters of the normal model by additionally applying the sensor data accumulated every day to the normal model. Is.
Specifically, for each diagnosis date, the cluster center c and the cluster radius r shown in FIG. 5 are adjusted using learning target data for a predetermined period relatively determined with the diagnosis date as a reference time. is there.

Here, as for the sensor data newly added to the learning target data, the belonging cluster is determined according to the normal model adjusted on the previous day. When the belonging cluster of all the added sensor data is determined, the average of the belonging sensor data is calculated for each cluster and set as a new cluster center c. In other words, a kind of moving average is used as the cluster center c. Further, for each cluster, for example, the distance (maximum error 105 shown in FIG. 6) from the new cluster center c to the farthest sensor data belonging to the cluster is set as a new cluster radius r.
As described above, the normal model stored in the normal model data storage unit 161d is updated for each diagnosis date based on the learning target data acquired by the learning target data acquisition unit 161a corresponding to the diagnosis date.

  Here, with reference to FIG. 7 (refer to FIGS. 2 and 5 as appropriate), the relationship between the change in the normal model generated by the update mode and the initialization mode and the change in the state of the external environment will be described. In FIG. 7, in the upper and lower graphs, the horizontal axis indicates the time of the scale common to both graphs, the vertical axis of the upper graph indicates the state quantity in the multidimensional space spanned by the multidimensional sensor data, The vertical axis of the lower graph shows temperature as an example of the state quantity of the external environment. For convenience of explanation, FIG. 7 shows a one-dimensional space as an example of the multidimensional space.

In this example, the upper graph shows the change of the cluster center (indicated by black circles in each of the clusters C _{1 to} C ₁₃ in FIG. 7) for one cluster indicating a normal model. For convenience of explanation, description of changes in the cluster radius is omitted. Clusters C _{1 to} C ₁₃ are extracted and shown at intervals of one month for clusters that are learned in the update mode _every day. The lower graph shows the change in temperature for about half a year.
FIG. 7 is a diagram for conceptually explaining the relationship between the two, and does not show a specific relationship.

First, it is assumed that learning in the initialization mode is executed at time t1. Cluster C ₁ generated at that time is learned to better conform to the temperature in the state of the external environment at that time. Take over time from time t1 to temperatures rise, as shown in the cluster C ₁ -C _4, cluster centers are gradually changed. This indicates that the cluster center is adjusted to reflect the change in the output of the sensor data in the normal range corresponding to the temperature change. In this way, in the present invention, by performing learning in the update mode, the cluster that is a normal model can be adjusted within a certain range with respect to a change in the external environment state.

Then, with the rapid increase in temperature, as shown in the cluster C ₅ -C _8, the cluster centers are greatly changed. When the state change of the external environment is large as described above, the difference between the normal model and the newly acquired sensor data becomes too large and exceeds the adjustment range of learning in the update mode. Although not shown in the figure, the cluster radius indicating the normal range becomes large at this time, so that the sensitivity for detecting an abnormal sign is lowered.
Therefore, at time t2 when a large change in the state of the external environment is detected, for example, sensor data acquired in a predetermined period C immediately before time t2 is used as learning target data, and learning by the initialization mode is performed by the initialization learning unit 161c2. Is executed to reconstruct the cluster which is a normal model, so that the cluster C ₈ is replaced with the cluster C ₁₁ . That is, in the state of the external environment at the time t2, the cluster C ₁₁ is indicative of the appropriate normal range. Therefore, when the cluster C ₈ before learning in the initialization mode is used, it may not be possible to properly diagnose the abnormality sign.

Clusters C _{9 to} C ₁₀ show the state of the cluster when learning in the update mode is continued without executing learning in the initialization mode. On the other hand, the clusters C _{11 to} C ₁₃ show the state of the cluster when learning in the initialization mode is executed at time t2 and then learning in the update mode is continued.
As described above, when the state of the external environment changes to some extent, the normal model can be appropriately adjusted by learning in the update mode by performing learning in the initialization mode.

(Learning target data acquisition process)
Next, a period for acquiring learning target data used in learning in the initialization mode and the update mode will be described.
In the present embodiment, learning in the initialization mode is performed by the initialization instruction means 15 (see FIG. 2) when the learning target data is acquired after the maintenance work, at a normal time after a certain period of time has elapsed since the maintenance work. Sometimes it is different.
Each case will be described in turn.

(A. Learning target data acquisition process in a normal case)
First, with reference to FIG. 3 (refer to FIG. 2 as appropriate), the acquisition target period of the learning target data acquired by the learning target data acquisition unit 161a at a normal time after a lapse of a fixed period from the maintenance work will be described.
As shown in FIG. 3, when the maintenance of the mechanical equipment 2 is not performed between the first day (1 day) and the diagnosis date (22 days) of the learning period B, which is an acquisition target period for normal learning. Shows about. Further, the dates shown in FIG. 3 are given with the first day of the learning period B as a reference (one day).
As shown in FIG. 3, the learning target data acquisition unit 161a goes back from the diagnosis date (22nd) to the past for a predetermined period A (for example, one week) and learns for a predetermined learning period B (for example, two weeks). Target data is acquired from the sensor data storage means 12. Here, the predetermined period A and the learning period B are preset periods.
The predetermined period A can cope with the detection of an event in which the equipment state slowly changes to an abnormal state over a long period of time, and can also cope with a change in the operating state of the mechanical equipment 2 itself. These may be set as appropriate. The learning period B may be a period in which normal model generation means 161c, which will be described later, can sufficiently acquire learning target data necessary for learning a normal model, and may be set as appropriate.

In addition, the learning target data acquisition unit 161a is set as the learning period B so as to exclude days when an alarm or a failure is detected by an interlock (not shown) installed in the machine facility 2. That is, since there was an alarm at the machine facility 2 on the 2nd shown in FIG. 3, the learning target data acquisition unit 161a acquires sensor data for two weeks excluding 2 days from the sensor data storage unit 12 as learning target data. .
The learning process by the data mining learning unit 161 is performed once a day at a preset time (for example, midnight). A diagnosis process by the data mining diagnosis unit 162 described later is diagnosed once a day with reference to a normal model learned by the learning process after the learning process is completed.
The predetermined period A and the learning period B shown in FIG. 3 are shifted one day at a time as the diagnosis date elapses.
The learning process and the diagnosis process described above are not limited to once a day, and may be performed every predetermined period.

In the above description, the day on which an alarm or failure is detected by an interlock (not shown) installed in the mechanical equipment 2 itself is excluded from the learning target day. It is preferable to exclude the day diagnosed as having a sign from the learning target. This is because the data mining unit 16 needs to learn normal data in order to perform highly accurate diagnosis using clustering. In this way, the sensor data of the mechanical equipment 2 on the day when the alarm or failure is detected and on the day excluding the day diagnosed as the abnormal sign is referred to as “normal data”.
That is, the learning target data acquisition unit 161a acquires normal data for a predetermined number of days in the past from the sensor data storage unit 12, and stores it in the learning target data storage unit 161b.

(B. Learning target data acquisition process when there is maintenance)
Next, the acquisition target period of the learning target data acquired by the learning target data acquisition unit 161a when maintenance (maintenance work) is performed will be described with reference to FIG. 4 (refer to FIG. 2 as appropriate).
As shown in FIG. 4, for example, it is assumed that maintenance is performed on a specific machine facility 2 during a period of 3 to 5 days. Further, the maintenance period (3 to 5 days) shown in FIG. 4 includes a period until the operation of the mechanical equipment 2 is stabilized after the maintenance work is completed.

When such maintenance is performed, there is a high possibility that the sensor data value will fluctuate significantly before and after the maintenance. If the learning target data acquisition unit 161a acquires the learning target data by the method shown in FIG. 3 that is applied in a normal case when the diagnosis date is immediately after maintenance, the following situation occurs. Occur. That is, the learning target data acquisition unit 161a acquires sensor data before maintenance and sensor data under maintenance as learning target data. Therefore, in this case, there is a possibility that the abnormal sign diagnostic apparatus 1 may continue to diagnose that there is an abnormal sign, even though the mechanical equipment 2 itself that has undergone maintenance is normal.
Therefore, when there is maintenance, the learning target data acquisition unit 161a acquires the sensor data of the learning period shown in FIG. 4 as learning target data.

As illustrated in FIG. 4, the learning target data acquisition unit 161 a is configured to acquire sensor data before maintenance (before two days) and sensor data for three to five days during maintenance from learning target data to be acquired. exclude. That is, the learning target data acquisition unit 161a acquires sensor data of a predetermined learning period B1 (for example, one week) as learning target data from 6 days after the maintenance period (3 days to 5 days) has elapsed.
Note that the learning period B1 shown in FIG. 4 is a minimum period necessary for the normal model generating unit 161c to create a new normal model using the learning target data acquired from the machine equipment 2 after the maintenance. Therefore, as shown in FIG. 4, the diagnostic means 162d described later does not diagnose abnormal signs during the maintenance period (3 to 5 days) and the learning period B1 after maintenance (6 to 12 days). . Then, the diagnosis unit 162d resumes the diagnosis of the abnormal sign from the 13th day after the learning period B1 has elapsed.

As shown in FIG. 4A, until the learning period B2 (for example, two weeks) that is a sufficient period as the learning period can be secured, the learning target data acquisition unit 161a starts from the sixth day after the maintenance to the diagnosis date. Is acquired from the sensor data storage unit 12 as learning target data.
Further, as shown in FIG. 4B, until the number of days from the last day (19th day) of the learning period B2 to the diagnosis date reaches a predetermined period A2 (for example, one week), the learning target data acquisition unit 161a Then, sensor data in the learning period B2 is acquired. This is because, as described above, it is possible to cope with detection of an event in which the equipment state slowly changes to an abnormal state over a long period of time.
And after 27th shown in FIG. 4, the learning object data acquisition means 161a performs the process similar to the above-mentioned "normal learning data acquisition process" (refer FIG. 3, FIG.4 (b)). Further, as shown in FIG. 4C, from the diagnosis date (28 days) when the number of days from the last day (19 days) of the learning period B2 exceeds the predetermined period A2, the learning object data acquisition unit 161a performs diagnosis. The learning period B3 and the predetermined period A3 are sequentially shifted with the passage of days.

In addition, the day when an alarm or a failure is detected by an interlock (not shown) installed in the mechanical equipment 2 itself, or the day when the abnormality sign diagnosis apparatus 1 diagnoses that there is an abnormality sign is excluded from the learning target days. This is the same as in the case of the “learning target data acquisition process in the normal case” described above.
Further, the settings of the learning periods B1 and B2 and the predetermined period A2 can be changed as appropriate.

(C. Learning target data acquisition process when there is an initialization instruction)
Next, referring to FIG. 7 (see FIG. 2 as appropriate), the learning target acquired by the learning target data acquiring unit 161a when the initialization instruction signal instructing the execution of the learning process in the initialization mode is input. The data acquisition target period will be described.
Learning in the initialization mode based on the initialization instruction signal is executed when the state of the external environment has changed greatly, and is for reconstructing a normal model. Therefore, it is preferable to use sensor data acquired during a minimum period necessary for reconstructing a normal model and a period in which the state of the external environment does not change significantly.

In the present embodiment, the target period for acquiring sensor data as learning target data is a predetermined period C closest to the date (time t2) on which the learning process is performed, as shown in FIG. The length of the predetermined period C can be set to, for example, three months in order to respond to changes in the four seasons. Depending on the degree of influence, it can be appropriately determined between several days to several months.
Note that it is preferable to perform learning in the initialization mode at a timing that does not exceed the adaptive capability of the normal model by learning in the update mode.

In addition, it is preferable to remove the sensor data of the day on which the occurrence of abnormality or the presence of abnormality sign is diagnosed in the predetermined period C from the learning target data.
In addition, when maintenance is performed in the predetermined period C, the sensor output may change greatly before and after the maintenance as described above, so the sensor data acquired after the maintenance work is used as learning target data. Is preferred.

Prior to the “normal period” after a fixed period after maintenance (“learning period B + predetermined period A” in FIG. 3), the learning process in the initialization mode executed after maintenance is prioritized and the initial The learning process based on the digitized signal is not executed. As a result, it is possible to avoid redundant execution of learning in the initialization mode within a short period of time.
Note that it may be configured to be able to select which of the learning process in the initialization mode executed after the maintenance and the learning process based on the initialization instruction signal has priority.
In addition, after maintenance, when a certain period does not elapse and sensor data sufficient for reconstructing a normal model is not accumulated, an instruction for learning processing based on an initialization signal may be canceled. The fixed period can be, for example, “learning period B1”, “learning period B2” or “learning period B2 + predetermined period A2” shown in FIG.

Further, as the predetermined period C, sensor data acquired at the same time (month and day) may be used in the previous year or a year before it, not in the latest period. In addition, when using sensor data before the previous year, sensor data acquired in a period before and after the date on which the learning process in the initialization mode is executed may be used. For example, when the predetermined period C is 3 months long and the learning process is executed on July 15, sensor data acquired from June 1 to August 31 of the previous year may be used. . Moreover, you may make it use both the sensor data acquired in the latest period, and the sensor data acquired during the same period before the previous year.
Furthermore, when outputting the initialization instruction signal, the predetermined period C may be set to an arbitrary length, or the various periods described above may be selected.

  When the instruction content of the initialization instruction signal is an initial value setting, the normal model generating unit 161c sets an initial value predetermined as normal model data by the initial value setting unit 161c1. Learning using the learning target data is not performed.

Returning to FIG. 2, the description of the configuration of the data mining unit 16 will be continued.
The data mining diagnosis unit 162 diagnoses the presence or absence of an abnormality sign from the sensor data of the machine facility 2 based on the normal model generated by the data mining learning unit 161.
As shown in FIG. 2, the data mining diagnosis unit 162 includes a diagnosis target data acquisition unit 162a, a diagnosis target data storage unit 162b, an abnormality degree calculation unit 162c, a diagnosis unit 162d, and a contribution calculation unit 162e. I have.

  The diagnosis target data acquisition unit 162a acquires sensor data to be diagnosed (hereinafter referred to as diagnosis target data) from the sensor data storage unit 12. In the present embodiment, the diagnosis target data acquisition unit 162a uses the sensor data for the previous day as diagnosis target data once a day at a predesignated time (for example, midnight) as the diagnosis target data. 12 from. The diagnosis target data acquisition unit 162a stores the acquired diagnosis target data in the diagnosis target data storage unit 162b.

  Further, as described above, before the data mining diagnosis unit 162 performs the diagnosis process, the data mining learning unit 161 acquires the learning target data corresponding to the diagnosis date in advance (see FIGS. 3 and 4), and the learning is performed. Normal model data generated by learning using target data is stored in the normal model data storage means 161d.

  The diagnosis target data storage unit 162b is a device for storing the diagnosis target data acquired by the diagnosis target data acquisition unit 162a. The diagnosis target data stored in the diagnosis target data storage unit 162b can be output to the abnormality degree calculation unit 162c and the contribution degree calculation unit 162e.

  In the above-described data mining learning unit 161, when the learning target data is divided for each operation mode and learning is performed for each operation mode to generate a normal model, the same procedure is performed in the data mining diagnosis unit 162. Suppose that the data to be diagnosed is divided for each operation mode. In this case, the diagnosis object data divided for each operation mode is stored in the diagnosis object data storage unit 162b for each operation mode.

The degree-of-abnormality calculation unit 162c reads out the diagnosis target data from the diagnosis target data storage unit 162b and calculates the degree of abnormality of the diagnosis target data.
For example, it is assumed that the diagnosis target data exists at the position shown in FIG. 5 and the cluster center closest to the diagnosis target data is a cluster center c shown in FIG. In this case, the degree of abnormality u is expressed by the following equation (1) using the cluster radius r and the distance d from the cluster center c to the diagnosis target data.
u = d / r Formula (1)

  The degree of abnormality calculation unit 162c outputs the degree of abnormality u calculated using Equation (1) to the diagnosis unit 162d. Further, the abnormality degree calculation means 162c associates the diagnosis target data with the abnormality degree u and stores the data in the diagnosis result storage means 18. Based on the abnormality degree u input from the abnormality degree calculating means 162c, the diagnosis means 162d diagnoses the presence or absence of an abnormality sign for the mechanical equipment 2 (see FIG. 1) that is the object from which the diagnosis object data has been acquired.

When the degree of abnormality u ≦ 1, it can be said that the diagnosis target data exists in the cluster shown in FIG. That is, in this case, since the diagnosis target data exists within the range of the specific normal model, the diagnosis unit 162d determines that the corresponding mechanical equipment 2 is normal (no abnormality sign).
On the other hand, when the degree of abnormality u> 1, it can be said that the diagnosis target data exists outside the cluster shown in FIG. That is, in this case, since the diagnosis target data does not belong to any normal model, the diagnosis unit 162d determines that there is a sign of abnormality in the corresponding mechanical equipment 2.

Then, the diagnosis unit 162d stores the presence / absence of an abnormality sign using the diagnosis target data in the diagnosis result storage unit 18 in association with the diagnosis target data.
In the above description, the diagnosis unit 162d determines whether or not there is an abnormality sign depending on whether or not the value of the degree of abnormality u is 1 or more. The value to be is not limited to 1. That is, the diagnosis unit 162d considers the distribution (such as variance) of the learning target data in the cluster having the cluster center closest to the diagnosis target data, the temporal change rate of the diagnosis target data, and the like. Can be set as appropriate.

The contribution degree calculation means 162e reads the diagnosis object data from the diagnosis object data storage means 162b, and calculates the contribution degree of each sensor (not shown) to the diagnosis object data. Each of the sensors indicates a plurality of sensors installed in the mechanical equipment 2 shown in FIG.
Although detailed description is omitted, for example, when the characteristic β shown in FIG. 5 is a normalized value of a temperature detected by a specific sensor P (not shown) provided at a predetermined location of the mechanical equipment 2, The contribution degree i is expressed as a ratio of the distance f which is the characteristic β component of the distance d shown in FIG. That is, the contribution degree i is expressed by the following formula (2).
i = f / d Formula (2)

In short, the contribution degree i is a value representing how much the sensor data detected by each sensor contributes to the abnormality degree u. That is, by comparing the value of the contribution degree i for each sensor, when it is detected that there is a sign of abnormality in the specific mechanical equipment 2, which sensor among the plurality of sensors installed in the mechanical equipment 2 is detected. It is possible to specify whether or not contributes most to the detection of an abnormal sign.
The contribution degree calculation unit 162e associates the diagnosis target data with the contribution degree i and stores the data in the diagnosis result storage unit 18.

(Example of diagnosis result by data mining)
Next, an example of a diagnosis result by the data mining diagnosis unit will be described with reference to FIG. As shown in FIG. 8, the diagnosis result storage means 18 (see FIG. 2) stores the diagnosis result as a database.
A time 201 shown in FIG. 8 indicates the date and time when the sensor data acquisition unit 11 (see FIG. 1) acquired the measurement value detected by a sensor (not shown) installed in the mechanical facility 2. The measurement value is a value obtained by the sensor of the mechanical equipment 2 for each preset period. Therefore, the time 201 is a value obtained by sequentially adding the period (for example, 30 seconds) to a predetermined time (for example, midnight).

  The learning target data acquisition date 202 indicates that a cluster that is a normal model has been learned using the sensor data of this day. The cluster number 203 is an identification number of the cluster having the shortest distance from the cluster center c (see FIG. 5) to the diagnosis target data among the plurality of clusters as the normal model described above. In the abnormality flag 204, a value of 1 is stored when there is a sign of abnormality, and a value of 0 is stored when there is no sign of abnormality, according to the diagnosis result by the diagnosis unit 162d (see FIG. 2). The operation state number 205 is an identification number assigned corresponding to the operation mode of the mechanical equipment 2. As described above, the abnormality degree 206 is an abnormality degree calculated by the abnormality degree calculation unit 162c (see FIG. 2). In the area 207 shown in FIG. 8, the measured value which is a value detected by each sensor installed in the machine facility 2 and the value of the component corresponding to the sensor 1 among the coordinates representing the cluster center c described above are shown. The learning value is stored. The contribution 208 is a contribution calculated for each sensor described above.

<Configuration of remote monitoring unit>
Next, the configuration of the remote monitoring unit 17 of the abnormality sign diagnosis apparatus 1 will be described with reference to FIG.
As described above, the remote monitoring unit 17 sets the upper and lower thresholds for the value (or change amount) of the sensor data acquired from the mechanical equipment 2 based on the experience and knowledge of the administrator, When a value or the like falls outside a predetermined range defined by the upper and lower thresholds, it is diagnosed that there is an abnormal sign.
The remote monitoring unit 17 may be referred to as individual determination means 171 ₁₁ , 171 ₁₂ ,..., 171 _1N , 171 ₂₁ , 171 _{22 ,.} And diagnostic means 172a, 172b,... (Hereinafter, simply referred to as diagnostic means 172). The individual determination means 171 ₁₁ , 171 ₂₁ ,... Acquire sensor data 1 (for example, temperature) from the sensor data storage means 12, and each outputs the determination result to the diagnosis means 172a, 172b,. To do. Similarly, the individual determination means 171 _1N , 171 _2N ,... Acquire the sensor data N from the sensor data storage means 12, and each outputs the determination result to the diagnosis means 172a, 172b,.
Each diagnosis means 172 has a predetermined logic circuit corresponding to the diagnosis contents (for example, diagnosis of whether or not the cooling water pressure of the mechanical equipment 2 is lowered).

Here, referring to FIG. 10 (refer to FIG. 9 as appropriate), an abnormality sign diagnosis process by the individual determination means 171 of the remote monitoring unit 17 will be described.
As described above, each individual determination unit 171 is associated with each sensor installed in the machine facility 2 on a one-to-one basis. As shown in FIG. 10, the individual determination unit 171 determines that the measured value in the sensor data is smaller than a preset threshold value A (see the one-dot chain line in FIG. 10) or from the threshold value B. When it becomes larger (see the broken line in FIG. 10), a predetermined ON signal is output to the diagnostic means 172 connected to itself.
In the individual determination means (for example, 171 ₁₁ and 171 ₂₁ ) for acquiring the same type of sensor data, the threshold values (see A and B in FIG. 10) may be set to be the same, and the diagnosis Different threshold values may be set corresponding to.

The individual determination unit 171 performs the determination on the sensor data stored in the sensor data storage unit 12 every preset sampling period (for example, 30 seconds), and outputs the result to the diagnosis unit 172. The diagnosis unit 172 has a predetermined logic circuit as described above, and outputs the presence / absence of an abnormality sign corresponding to each diagnosis content corresponding to the presence / absence of a signal from each individual determination unit 171. It is supposed to be.
Then, the diagnosis unit 172 stores each diagnosis result in the diagnosis result storage unit 18 (see FIG. 2).

Note that the threshold setting in the individual determination unit 171 of the remote monitoring unit 17 shown in FIG. 9 and the logic circuit setting in the diagnosis unit 172 can be changed as appropriate. Therefore, for example, the administrator of the abnormality sign diagnosis apparatus 1 (see FIG. 1) can refer to the maintenance period and the content of the maintenance and appropriately change the settings such as the threshold value.
As shown in FIG. 1, the display control unit 19 stores the sensor data stored in the sensor data storage unit 12 and the maintenance information storage in addition to the diagnosis result by the data mining unit 16 and the diagnosis result by the remote monitoring unit 17. Maintenance information stored in the means 14 can also be displayed on the display means 20.

<Processing flow by the data mining unit>
Next, the flow of learning processing and diagnosis processing by the data mining unit will be described with reference to the drawings as appropriate.

(Learning process flow by the data mining learning unit)
First, the flow of learning processing by the data mining learning unit 161 will be described with reference to FIG. 11 (see FIGS. 2, 4 and 7 as appropriate).
The data mining learning unit 161 uses the normal model generation unit 161c to check whether or not the initialization instruction signal is input from the initialization instruction unit 15 (step S11), and when there is no initialization instruction signal input (step S11). No), the process proceeds to step S12. In step S12, the data mining learning unit 161 uses the learning target data acquisition unit 161a to execute the learning target data acquisition process 1 when there is maintenance including the normal period, and performs a predetermined period (the learning period B or FIG. 3). The learning target data in any of the learning periods B1 to B3 in FIG. 4 is acquired from the sensor data storage unit 12. In addition, the data mining learning unit 161 causes the learning target data acquisition unit 161a to store the acquired learning target data in the learning target data storage unit 161b.

Next, the data mining learning unit 161 uses the learning target data acquisition unit 161a to
In step S12, it is confirmed whether or not the learning target data has been acquired (step S13). If the learning target data has not been acquired (No in step S13), the learning is not performed and the process ends.

  On the other hand, when learning target data has been acquired (Yes in step S13), the data mining learning unit 161 confirms whether or not the first learning after maintenance is performed by the normal model generation unit 161c (step S14). Whether it is the first learning after the maintenance can be confirmed by whether it is the first day when the minimum “learning period B1” necessary for the reconstruction of the normal model shown in FIG. 4 has elapsed.

  If it is the first learning after the maintenance (Yes in step S14), the data mining learning unit 161 is initialized by using the learning target data acquired in step S12 by the initialization learning unit 161c2 of the normal model generation unit 161c. A normal model is reconstructed by executing learning by mode (step S20). Further, the data mining learning unit 161 causes the initialization learning unit 161c2 to store the reconstructed normal model, which is the learning result of step S20, in the normal model data storage unit 161d (step S16), and ends the process.

  On the other hand, when it is not the first learning after the maintenance (No in step S14), the data mining learning unit 161 uses the learning target data acquired in step S12 by the update learning unit 161c3 of the normal model generation unit 161c. The normal model is adjusted by executing learning by (step S15). In addition, the data mining learning unit 161 causes the update learning unit 161c3 to store the adjusted normal model, which is the learning result of step S15, in the normal model data storage unit 161d (step S16), and ends the process.

  If an initialization instruction signal is input in step S11 (Yes in step S11), the data mining learning unit 161 determines whether the content of the initialization instruction is “initial value setting” by the normal model generation unit 161c. Confirm (step S17). When the content of the initialization instruction is initial value setting (Yes in step S17), the data mining learning unit 161 sets a predetermined initial value by the initial value setting unit 161c1 of the normal model generation unit 161c. A normal model is generated (step S21). In addition, the data mining learning unit 161 stores the normal model generated in step S20 in the normal model data storage unit 161d by the initial value setting unit 161c1, and ends the process.

  On the other hand, when the content of the initialization instruction is not the initial value setting (No in step S17), the data mining learning unit 161 performs learning processing in the initialization mode based on the initialization instruction by the learning target data acquisition unit 161a. The learning target data acquisition process 2 is executed to acquire learning target data for a predetermined period (predetermined period C in FIG. 7) from the sensor data storage unit 12 (step S18). In addition, the data mining learning unit 161 causes the learning target data acquisition unit 161a to store the acquired learning target data in the learning target data storage unit 161b.

Next, the data mining learning unit 161 uses the learning target data acquisition unit 161a to
In step S18, it is confirmed whether or not the learning target data has been acquired (step S19). If the learning target data has not been acquired (No in step S19), the learning is not performed and the process ends.

  On the other hand, if the learning target data can be acquired (Yes in step S19), the data mining learning unit 161 uses the learning target data acquired in step S18 by the initialization learning unit 161c2 of the normal model generation unit 161c to perform initial processing. The normal model is reconstructed by executing learning in the conversion mode (step S20). Further, the data mining learning unit 161 causes the initialization learning unit 161c2 to store the reconstructed normal model, which is the learning result of step S20, in the normal model data storage unit 161d (step S16), and ends the process.

(Flow of learning target data acquisition process 1 when there is maintenance)
Next, with reference to FIG. 12 (refer to FIGS. 2 to 4 and 11 as appropriate), the learning target data acquisition process 1 by the learning target data acquisition means 161a when there is maintenance will be described. The learning target data acquisition process 1 shown in FIG. 12 corresponds to the details of the learning target data acquisition process 1 in step S12 in FIG.

First, the learning target data acquisition unit 161a determines whether or not the diagnosis date is within the normal period (step S31). Here, the “normal period” refers to the period when the learning period B and the predetermined period A shown in FIG. 3 are secured with reference to the date when the maintenance performed at the time closest to the diagnosis date is completed. Show.
When the diagnosis date is within the normal period (Yes in step S31), the learning target data acquisition unit 161a acquires the learning target data from the sensor data storage unit 12 in the normal mode (step S32). Here, the “normal mode” is a mode in which the learning target data acquisition unit 161a acquires, as learning target data, sensor data in the learning period B that is traced back by a predetermined period A from the diagnosis date as shown in FIG. Show.

  If the diagnosis date is not within the normal period (No in step S31), the process of the learning target data acquisition unit 161a proceeds to step S33. In step S33, the learning target data acquisition unit 161a determines whether or not the diagnosis date is within the maintenance period. If the diagnosis date is within the maintenance period (Yes in step S33), the processing of the learning target data acquisition unit 161a proceeds to step S34. In step S <b> 34, the learning target data acquisition unit 161 a causes the diagnosis result storage unit 18 to store information indicating that the diagnosis date is within the maintenance period. In this case, since there is no sensor data that can be used as learning target data, the learning target data acquisition unit 161 a does not acquire learning target data from the sensor data storage unit 12.

  On the other hand, if the diagnosis date is not within the maintenance period (No in step S33), the process of the learning target data acquisition unit 161a proceeds to step S35. In step S35, the learning target data acquisition unit 161a determines whether or not the diagnosis date is included in the learning period B1 shown in FIG. That is, the learning target data acquisition unit 161a determines whether or not a minimum learning period B1 that is necessary as a learning period after maintenance is secured.

When the diagnosis date is included in the learning period B1 (Yes in step S35), the process of the learning target data acquisition unit 161a proceeds to step S36. In step S36, the learning target data acquisition unit 161a causes the diagnosis result storage unit 18 to store information indicating that the diagnosis date is within the learning period after the maintenance. In this case, since the sensor data that can be used as the learning target data is insufficient, the learning target data acquisition unit 161 a does not acquire the learning target data from the sensor data storage unit 12.
On the other hand, when the diagnosis date is not included in the learning period B1 (No in step S35), the processing of the learning target data acquisition unit 161a proceeds to step S37.

In step S37, the learning target data acquisition unit 161a determines whether or not the learning period B2 and the predetermined period A2 shown in FIG. When the learning period B2 and the predetermined period A2 are secured (Yes in step S37), the processing of the learning target data acquisition unit 161a proceeds to step S32. Note that this case corresponds to the case of (b) or (c) in FIG. That is, in this case, it can be said that the diagnosis date corresponds to the “normal period” described above.
On the other hand, when the learning period B2 is not secured, or when the learning period B2 is secured but the predetermined period A2 is not secured (No in step S37), the learning target data acquisition unit 161a is in the maintenance mode. Thus, learning target data is acquired (step S38). Here, the “maintenance mode” is a transitional learning period selection mode until returning to the “normal mode” as described above with reference to FIG.

(Flow of learning target data acquisition process 2 when learning process is performed in the initialization mode based on the initialization instruction)
Next, referring to FIG. 13 (refer to FIGS. 2 to 4, 7, and 11 as appropriate), learning in the case of performing learning processing in the initialization mode based on the initialization instruction from the initialization instruction means 15. The learning target data acquisition process 2 by the target data acquisition unit 161a will be described. The learning target data acquisition process 2 shown in FIG. 13 corresponds to the details of the learning target data acquisition process 2 in step S18 in FIG.

  First, the learning target data acquisition unit 161a checks whether or not the predetermined period C shown in FIG. 7 has been secured after the last maintenance was performed (step S41). If the predetermined period C is secured (Yes in step S41), the learning target data acquisition unit 161a acquires the sensor data acquired during the predetermined period C from the sensor data storage unit 12 as learning target data. (Step S42), it is stored in the learning object data storage means 161b.

  On the other hand, when the predetermined period C is not secured (No in step S41), the learning target data acquisition unit 161a further performs a certain period (for example, “learning shown in FIG. 4” after the last maintenance is performed). (Period B1 "," Learning period B2 "or" Learning period B2 + predetermined period A2 ") is confirmed (step S43). If the certain period is secured (Yes in step S43), the sensor data storage unit 12 obtains the sensor data of all the latest periods within the predetermined period C and after the maintenance (step S44). ) And stored in the learning object data storage means 161b.

(Diagnosis process flow by the data mining diagnosis unit)
Next, the flow of diagnosis processing by the data mining diagnosis unit 162 will be described with reference to FIG. 14 (refer to FIG. 2 as appropriate).
First, the data mining diagnosis unit 162 acquires diagnosis target data from the sensor data storage unit 12 by the diagnosis target data acquisition unit 162a (step S51) and stores it in the diagnosis target data storage unit 162b. For example, the data mining diagnosis unit 162 uses the sensor data storage unit 12 from the sensor data storage unit 12 by using the diagnosis target data acquisition unit 162a as a diagnosis target data of sensor data for the previous day at a time designated in advance (for example, midnight). get.

  Next, the data mining diagnosis unit 162 calculates the degree of abnormality of the diagnosis target data by the degree of abnormality calculation unit 162c (step S52). That is, the data mining diagnosis unit 162 acquires a normal model from the normal model data storage unit 161d by the abnormality level calculation unit 162c, and calculates the abnormality level based on the normal model.

Next, the data mining diagnosis unit 162 executes diagnosis by the diagnosis unit 162d (step S53). That is, the data mining diagnosis unit 162 uses the diagnosis unit 162d to diagnose the presence or absence of an abnormality sign in the mechanical equipment 2 (see FIG. 1) based on whether or not the abnormality degree value calculated in step S52 is greater than or equal to a predetermined value. .
Then, the data mining diagnosis unit 162 causes the diagnosis unit 162d to store the diagnosis result executed in step S53 in the diagnosis result storage unit 18.

As described above, the normal model generation unit 161c cannot generate an appropriate normal model during the maintenance period shown in FIG. 4 and the minimum learning period B1 required after the maintenance. The unit 16 does not output the diagnosis result.
However, in the remote monitoring unit 17 (see FIG. 1), diagnosis is performed based on sensor data input every sampling period (for example, 30 seconds), and diagnosis is performed on the diagnosis result storage unit 18 (see FIG. 1). The result is output. That is, even during the maintenance period and the minimum necessary learning period B1 after maintenance, the abnormality sign diagnosis apparatus 1 outputs the diagnosis result by the remote monitoring unit 17 (see FIG. 1), and the display means 20 (see FIG. 1). Can be displayed.

<Effect of the present invention>
As described above, the abnormality sign diagnosis apparatus 1 according to the present embodiment includes the sensor data acquisition unit 11 that acquires sensor data detected by a plurality of sensors installed in the machine facility 2, and the sensor data acquisition unit 11. The sensor data storage means 12 for storing the sensor data acquired by the above and the period when the sensor data is acquired by the sensor data acquisition means 11 are specified, and the specified period (learning period B1, learning period) is specified from the sensor data storage means 12. B2, a learning period B3 or a predetermined period C), a learning target data acquisition process for acquiring sensor data, and a normal model indicating a normal range of the sensor data by learning using sensor data corresponding to a specified period A learning process to be generated, a diagnostic process for diagnosing the presence or absence of an abnormality sign of the mechanical equipment 2 based on a normal model, The first diagnosis is performed by the first diagnosis (data mining unit 16) and the learning process in the initialization mode in which the entire normal model is reconstructed according to the change of the external environment that is the environment around the mechanical equipment 2. Initialization instruction means 15 for instructing the means.
For this reason, the abnormality sign diagnostic apparatus 1 can initialize the normal model in response to a change in the state of the external environment and appropriately diagnose the abnormality sign. In addition, when performing the learning process in the initialization mode in which the entire normal model is reconstructed, parameters such as the outside temperature indicating the state of the external environment are not directly used, and therefore the process for the learning process of the data mining unit 16 The load can be reduced.

  Further, the learning process has an initialization mode and an update mode in which the normal model is partially adjusted with a first period (learning period B) defined as a reference time when the learning process is performed as a designated period. The first diagnosis means (data mining unit 16) performs learning in the update mode periodically at a predetermined cycle (for example, once a day) or at a predetermined time (for example, midnight). It is preferable to perform processing. As a result, normally, by learning in the update mode, the normal model can be made to respond to a change in the state of the external environment with a certain adaptive capability without increasing the processing load. For this reason, even when the mechanical equipment 2 is installed in an environment in which the state of the external environment changes greatly, the interval for performing learning in the initialization mode can be increased.

  Further, it is preferable that the initialization instruction means 15 has a function of instructing the first diagnosis means to set an initial value with a predetermined initial value as a normal model. As a result, when setting the mechanical equipment 2, it is possible to start diagnosis of an abnormal sign by a data mining technique without acquiring sensor data by a trial run.

  The initialization instructing unit 15 acquires a measurement value of a sensor that detects the state of the external environment, and when the change in the state of the external environment becomes a predetermined threshold value or more, the first diagnosis unit (data mining) It is preferable that the learning process in the initialization mode is instructed to the unit 16). As a result, the normal model can be automatically initialized and reconstructed according to changes in the external environment without any artificial operation. Thereby, it is possible to automatically maintain the accuracy of the abnormality sign diagnosis by the data mining technique.

  In addition, when the initialization instruction unit 15 detects an input based on an artificial operation instructing the learning process in the initialization mode, the initialization instruction unit 15 performs the learning process in the initialization mode with respect to the first diagnosis unit (data mining unit 16). Is preferably indicated. As a result, the normal model can be initialized and reconstructed at an arbitrary timing. As a result, it is possible to more appropriately maintain the diagnostic accuracy of the abnormal sign by the data mining technique.

  In addition, when the learning instruction in the initialization mode is instructed by the initialization instruction unit 15, the first diagnosis unit (data mining unit 16) performs the learning process in the initialization mode in the learning target data acquisition process. It is preferable that the learning process in the initialization mode is performed for the normal model using the sensor data of the second period (predetermined period C) determined with the initialization instruction time when the instruction is given as a reference time. . Thereby, a normal model adapted to the external environment at the time of execution of the learning process, that is, the diagnostic process can be reconstructed.

  Further, it is preferable that the second period (predetermined period C) is the most recent period including the time of acquisition of the sensor data acquired last when the initialization instruction is given. As a result, a normal model more adapted to the external environment at the time of execution of the learning process, that is, the diagnostic process can be constructed.

  Also, maintenance information acquisition means 13 for acquiring maintenance information including at least information for specifying the mechanical equipment on which the maintenance work is performed and the maintenance work period, and maintenance information in which the maintenance information acquired by the maintenance information acquisition means 13 is stored. Storage means 14, and the first diagnosis means (data mining unit 16) performs maintenance work on the mechanical equipment 2 and at least in advance between the end of the maintenance work and the time of diagnosis. When it is determined that the third period (learning period B1) of a predetermined length is secured, sensor data for a predetermined period after the maintenance work is completed is acquired from the sensor data storage unit 12, and the sensor data It is preferable to perform learning processing in the update mode using. Thereby, a normal model can be appropriately generated after maintenance work.

  The first diagnosis means (data mining unit 16) is instructed by the initialization instruction means 15 to perform learning processing in the initialization mode, and the maintenance work is performed on the mechanical equipment 2 and the maintenance is performed. If it is determined that at least the third period (learning period B1) has not been secured between the end of the work and the time of diagnosis, the learning process in the initialization mode according to the instruction of the initialization instruction unit 15 is performed. It is preferable not to do so. As a result, it is possible to prevent execution of learning processing in which an appropriate normal model cannot be generated.

  Further, the first diagnosis means (data mining unit 16) performs maintenance work on the mechanical equipment 2, and at least the third period (learning period B1) between the end of the maintenance work and the time of the diagnosis. ) Is ensured, and the learning process performed first after the maintenance work is preferably performed as the learning process in the initialization mode. Thereby, a normal model can be appropriately constructed after maintenance work.

  The first diagnosis means (data mining unit 16) is represented by the center of at least one cluster indicating the normal range of the sensor data and the cluster radius that is the distance from the cluster center to the boundary indicating the normal range. A learning process is performed by learning a normal model, and the diagnosis means (data mining unit 16) calculates the ratio of the distance between the cluster center closest to the diagnosis target data and the diagnosis target data among the cluster centers to the cluster radius. It is preferable to calculate the degree of abnormality shown and determine the presence or absence of an abnormality sign of the mechanical equipment 2 based on the value of the degree of abnormality. As a result, an abnormality sign can be detected with high sensitivity.

  Moreover, it is preferable that a 1st diagnostic means (data mining part 16) adjusts the center and radius of the said cluster for every cluster in the learning process in update mode. As a result, the normal model can be adjusted without increasing the processing load.

Further, the abnormality sign diagnosis apparatus 1 acquires sensor data from the sensor data storage unit 12 and diagnoses that there is an abnormality sign when each sensor data exceeds a predetermined threshold value set in advance. It is preferable to further include a second diagnostic means (remote monitoring unit 17).
Thereby, before the learning in the initialization mode for the first means (data mining unit 16) is executed, the state of the external environment changes greatly, and the data mining technique using the normal model before the initialization is performed. Even if the detection sensitivity of the abnormality sign is lowered, by combining the diagnosis by the remote monitoring method of the second diagnosis means (remote monitoring unit 17), at least by the remote monitoring method without significant reduction in sensitivity. Abnormal signs can be diagnosed with sensitivity.

<Modification>
In the above-described embodiment, the sensor data and maintenance information of the machine facility 2 are transmitted to the abnormality sign diagnosis apparatus 1 via the communication network N, but the present invention is not limited to this. For example, the abnormality sign diagnosis apparatus 1 can also acquire each data from a storage medium storing sensor data or maintenance information.

Moreover, although FIG. 1 demonstrated the case where the mechanical installation 2 and the computer 3 existed in different positions and the maintenance information of each mechanical installation 2 was registered into one computer 3 collectively, it is not restricted to this. Absent. That is, there are a plurality of computers 3, each computer 3 manages maintenance information of one or a plurality of (for example, five) machine facilities 2, and the maintenance information is transmitted to the abnormality sign diagnosis apparatus 1 via the communication network N. It is good also as transmitting.
In the above-described embodiment, the data mining diagnosis unit 162 (see FIG. 2) diagnoses sensor data for one day of the previous day once a day at a preset time (for example, midnight). The target data is acquired from the sensor data storage unit 12, but the present invention is not limited to this. For example, the data mining diagnosis unit 162 may diagnose an abnormal sign in real time.

In the above-described embodiment, the contribution calculated by the contribution calculation unit 162e of the data mining diagnosis unit 162 is simply stored in the diagnosis result storage unit 18 and further displayed on the display unit 20. Not limited to. As described above, it is possible to grasp how much the measurement value by the sensor contributes to the degree of abnormality based on the contribution calculated for each sensor (not shown) installed in the mechanical equipment 2.
Therefore, the contribution calculated by the contribution calculation unit 162e may be reflected on the logic circuit of the diagnosis unit 172 of the remote monitoring unit 17 (see FIG. 9). For example, let us consider a case where the data mining unit 16 finds that the contribution degree of the sensor A is the largest for the content of a certain abnormality sign. In this case, if the logic circuit of the diagnostic means 172 is rearranged so that the change in the value of the sensor A is most sensitively reflected in the diagnostic result in the remote monitoring unit 17, the remote monitoring unit 17 can more appropriately display the diagnostic result. It becomes possible to output.

In addition, as shown in FIG. 2, the normal model data storage unit 161d, the learning target data storage unit 161b, and the diagnosis target data storage unit 162b are included in the data mining unit 16, but the present invention is not limited thereto. That is, each of these storage means may be provided outside the abnormality sign diagnosis apparatus 1.
In the above-described embodiment, the case where the normal model generation unit 161c performs clustering using the k-average method as non-hierarchical clustering and learns a normal model has been described, but the present invention is not limited thereto. That is, for example, the normal model generation unit 161c may perform learning processing using fuzzy clustering, a mixed density distribution method, or the like as non-hierarchical clustering.

In addition, this invention is not limited to above-described embodiment, Various modifications are included. For example, the above-described embodiment has been described in detail for easy understanding of the present invention, and is not necessarily limited to one having all the configurations described. Further, a part of the configuration of an embodiment can be replaced with the configuration of another embodiment, and the configuration of another embodiment can be added to the configuration of an embodiment. In addition, it is possible to add, delete, and replace other configurations for a part of the configuration of each embodiment.
Each of the above-described configurations, functions, processing units, processing means, and the like may be realized by hardware by designing a part or all of them with, for example, an integrated circuit. In addition, each of the above-described configurations, functions, and the like may be realized by software by the processor interpreting and executing a program that realizes each function. Information such as programs, tapes, and files that realize each function can be stored in a recording device such as a memory, a hard disk, or an SSD (Solid State Drive), or a recording medium such as an IC card, an SD card, or a DVD.
Further, the control lines and information lines indicate what is considered necessary for the explanation, and not all the control lines and information lines on the product are necessarily shown. Actually, it may be considered that almost all the components are connected to each other.

DESCRIPTION OF SYMBOLS 1 Abnormal sign diagnostic apparatus 10 Communication means 11 Sensor data acquisition means 12 Sensor data storage means 13 Maintenance information acquisition means 14 Maintenance information storage means 15 Initialization instruction means 16 Data mining part (1st diagnosis means)
161 Data mining learning unit 161a Learning target data acquisition unit 161b Learning target data storage unit 161c Normal model generation unit 161c1 Initial value setting unit 161c2 Initialization learning unit 161c3 Update learning unit 161c Normal model data storage unit 162 Data mining diagnosis unit 162a Diagnosis target Data acquisition means 162b Diagnosis object data storage means 162c Abnormality calculation means 162d Diagnosis means 162e Contribution calculation means 17 Remote monitoring unit (second diagnosis means)
171 _{11 to} 171 _1N , 171 _{21 to} 171 _2N individual determination means 172a, 172b diagnosis means 18 diagnosis result storage means 19 display control means 20 display means 2 machine equipment 3 computer N communication network

Claims (16)

Using the sensor data detected by a plurality of sensors installed in the mechanical equipment, an abnormality sign diagnosis device for diagnosing the presence or absence of an abnormal sign of the mechanical equipment,
Sensor data acquisition means for acquiring sensor data from the plurality of sensors;
Sensor data storage means for storing the sensor data acquired by the sensor data acquisition means;
A learning target data acquisition process for specifying a period during which the sensor data is acquired by the sensor data acquisition unit and acquiring sensor data corresponding to the specified period from the sensor data storage unit, and sensor data corresponding to the specified period A first diagnosis for performing a learning process for generating a normal model indicating a normal range of the sensor data by learning using a diagnostic process and a diagnostic process for diagnosing the presence or absence of an abnormality sign of the mechanical equipment based on the normal model Means,
Initialization instruction means for instructing the first diagnosis means in the learning process in an initialization mode for reconstructing the whole normal model in response to a change in the external environment that is an environment around the mechanical facility;
An abnormality sign diagnosis apparatus comprising: The learning process includes the initialization mode, and an update mode in which the normal model is partially adjusted with the first period defined as a reference time when the learning process is performed as the designated period,
2. The abnormality sign diagnosis apparatus according to claim 1, wherein the first diagnosis unit performs a learning process in the update mode at a predetermined cycle or at a predetermined time.   The said initialization instruction | indication means further has a function which instruct | indicates the initial value setting which makes a predetermined initial value the said normal model to the said 1st diagnostic means, The Claim 1 or Claim 2 characterized by the above-mentioned. Abnormal sign diagnosis device.   The initialization instruction means acquires a measurement value of a sensor that detects the state of the external environment, and when the change in the state of the external environment becomes a predetermined threshold value or more, the first instruction means The abnormality sign diagnosis apparatus according to any one of claims 1 to 3, wherein a learning process in the initialization mode is instructed.   The initialization instruction means instructs the first diagnosis means to perform the learning process in the initialization mode when detecting an input based on an artificial operation that instructs the learning process in the initialization mode. The abnormality sign diagnostic apparatus according to any one of claims 1 to 4, characterized in that: The first diagnosis means includes
When the learning instruction in the initialization mode is instructed by the initialization instructing means, the initialization instruction time when the learning process in the initialization mode is instructed in the learning target data acquisition process is a reference 6. The abnormality sign according to claim 1, wherein a learning process in the initialization mode is performed for the normal model, with the second period defined as time being the specified period. Diagnostic device.   The abnormality sign diagnosis apparatus according to claim 6, wherein the second period is the most recent period including the time of acquisition of the last acquired sensor data at the time of the initialization instruction. Maintenance information acquisition means for acquiring maintenance information including at least information for specifying the mechanical equipment on which maintenance work is performed and the maintenance work period; and
Maintenance information storage means for storing maintenance information acquired by the maintenance information acquisition means;
Further comprising
The first diagnosing means performs a maintenance operation on the mechanical equipment, and at least a third period of a predetermined length is secured between the end of the maintenance operation and the time of the diagnosis. The sensor data is acquired from the sensor data storage means for a predetermined period after the maintenance work is completed, and learning processing in the update mode is performed using the sensor data. The abnormality sign diagnosis apparatus according to any one of claims 3 to 7, which refers to claim 2 or claim 2. The first diagnosis means includes
A learning process in the initialization mode is instructed by the initialization instruction means;
When it is determined that at least the third period is not secured between the completion of the maintenance work and the time of diagnosis after the maintenance work is performed on the mechanical equipment, the initialization instruction unit The abnormality predictor diagnosis apparatus according to claim 8, wherein the learning process in the initialization mode according to the instruction is not performed. The first diagnosis means includes
A maintenance work is performed on the mechanical equipment, and it is determined that at least the third period is secured between the end of the maintenance work and the time of diagnosis, and the maintenance work is finished. The abnormality sign diagnosis apparatus according to claim 8 or 9, wherein a learning process that is first performed later is a learning process in the initialization mode. The first diagnosis means learns the normal model represented by a center of at least one cluster indicating a normal range of the sensor data and a cluster radius that is a distance from the cluster center to a boundary indicating a normal range. To perform the learning process,
The first diagnosing means calculates a degree of abnormality indicated by a ratio of a distance between the cluster center closest to the diagnosis target data and the diagnosis target data to the cluster radius among the cluster centers, 11. The abnormality sign diagnosis apparatus according to claim 1, wherein the diagnosis process is performed by determining presence / absence of an abnormality sign of the mechanical equipment based on the information.   The abnormality sign according to claim 11, wherein the first diagnosis unit adjusts the center and radius of the cluster for each of the clusters in the learning process in the update mode. Diagnostic device.   The apparatus further comprises second diagnosis means for acquiring the sensor data from the sensor data storage means and diagnosing that there is a sign of abnormality when each sensor data exceeds a predetermined threshold value set in advance. The abnormality sign diagnosis apparatus according to any one of claims 1 to 12, characterized by: Using the sensor data detected by a plurality of sensors installed in a mechanical facility, an abnormality sign diagnosis method used in an abnormality sign diagnosis device for diagnosing the presence or absence of an abnormality sign of the mechanical facility,
The abnormal sign diagnostic apparatus is
Comprising sensor data storage means for storing the sensor data measured by the plurality of sensors;
Specifying a period during which the sensor data was acquired, learning using a learning target data acquisition process for acquiring sensor data corresponding to the specified period from the sensor data storage unit, and learning using sensor data corresponding to the specified period A first diagnosis step for performing a learning process for generating a normal model indicating a normal range of the sensor data and a diagnosis process for diagnosing the presence or absence of an abnormality sign of the mechanical equipment based on the normal model;
An initialization instruction step for instructing the learning process in an initialization mode for reconstructing the whole normal model in response to a change in the external environment that is an environment around the mechanical facility;
A method for diagnosing abnormal signs, comprising: The learning process includes the initialization mode,
An update mode for partially adjusting the normal model, with the first period defined as a reference time when performing the learning process as the designated period;
15. The abnormality sign diagnosis method according to claim 14, wherein in the first diagnosis step, the learning process in the update mode is periodically performed at a predetermined cycle or at a predetermined time.   The method further includes a second diagnosis step of acquiring the sensor data from the sensor data storage means and diagnosing that there is an abnormality sign when each sensor data exceeds a predetermined threshold value set in advance. 16. The abnormality sign diagnosis method according to claim 14 or claim 15.

Images