JP6632941B2 - Equipment monitoring device and equipment monitoring method - Google Patents

Description

  The present invention relates to a facility monitoring apparatus and a facility monitoring method, and more particularly to a facility monitoring apparatus and a facility monitoring method for monitoring the state of a facility based on information collected from the facility.

  Accumulates the detection data of the sensors provided in the facilities such as power plants, extracts the detection data when the equipment is normal from the accumulated detection data, and the learning model and the equipment that model the detection data when the equipment is normal Comparing with the detection data of the sensors detected during operation, if these differences exceed a predetermined threshold, the equipment is not in a normal state, that is, there is a possibility that an abnormality has occurred in the equipment An invariant analysis technique (SIAT) for determining is known (see Non-Patent Document 1).

  In this invariant analysis technique, in order to create a learning model when the equipment is normal, detection data when the equipment is normal is collected from accumulated sensor detection data. For this reason, it is important to improve the accuracy of collecting the detection data when the equipment is normal. In other words, it is important to exclude the detection data at the time of equipment abnormality from the accumulated data with high accuracy.

  Therefore, as a technique for removing detected data at the time of equipment abnormality from accumulated data, Patent Documents 1 and 2 disclose a method of recovering the accumulated data from a point in time when an alarm related to equipment abnormality, failure, or the like occurs. It is described that the detection data of the sensor is removed during the period until the operation is completed.

  Further, Patent Document 3 discloses that even if the generation of a plurality of warnings relating to equipment and the recovery of those warnings are mixed in an out-of-order state in chronological order, the process from the first warning to the recovery of the last warning is accurately performed. It is described that the detection data of the sensor during this period is not adopted.

Kei Fukushima and five others, "Plant failure sign monitoring system using invariant analysis technology (SIAT)", NEC Techniques, NEC Corporation, November 2014, Vol. 67 No. 1. Special Issue on Public Solutions to Support Social Safety and Security, p. 119-122

JP 2010-191556 A JP 2011-070635 A JP 08-297462 A

  The alarm generation (error occurrence) and alarm recovery (error recovery) of the equipment are closely related to the sensor detection data. When the correspondence between the alarm generation and the alarm recovery and the detection data of the sensor is clear, for example, the detection data of the sensor instantaneously changes in response to the alarm generation and the alarm recovery almost simultaneously with the alarm generation and the alarm recovery of the equipment. In such a case, it is possible to determine the abnormality of the equipment and its recovery from the change in the detection data of the sensor.

  However, when the correspondence between the alarm generation and the alarm recovery and the detection data of the sensor is not clear, for example, when the detection data of the sensor detects a warning sign (a warning sign) before the occurrence of the warning of the equipment, In the case where a slight abnormality occurs in the equipment and the abnormality proceeds to generate an alarm for the equipment, the detection data is gradually changing before the alarm is generated. Further, depending on the cause of the alarm of the equipment, the characteristics of the sensor, and the like, after the alarm is restored from the equipment, the detection data of the sensor may not be stable, and a predetermined period may be required until the detection data returns to normal.

  When the correspondence between the occurrence of an alarm or the like and the change in the detection data of the sensor is clear, the detection data of the sensor at the time of equipment abnormality is removed from the stored detection data of the sensor by the technology described in Patent Documents 1-3. be able to. However, if the correspondence between the occurrence of an alarm and the detection data of the sensor is not clear, detection data indicating a sign of abnormality or unstable detection data is included in the data for generating the learning model. As a result, the accuracy of the learning model decreases, and as a result, the accuracy of determining the equipment abnormality also decreases.

  In addition, depending on the cause of the equipment alarm and the characteristics of the sensor, there is a small change in the sensor detection data before and after the alarm is generated and the alarm is recovered, and the detection data of such a small change is used for the learning model. Sometimes it is possible to do so.

  Therefore, an object of the present invention is to improve the accuracy of a learning model by excluding abnormality detection data from learning data for generating a learning model or adding normal detection data to the learning data.

An equipment monitoring apparatus according to the present invention is an equipment monitoring apparatus that monitors the state of the equipment based on detection data of a sensor provided in the equipment, wherein the detection data, information on occurrence of an alarm due to equipment abnormality, and recovery of the alarm. A data accumulation unit for accumulating information in time series, and a correspondence between the detection data and the occurrence information and the restoration information in time series, and a first data period from an alarm occurrence to an alarm restoration in the detection data; A second data period from recovery of an alarm to generation of an alarm in the detection data, respectively, and determining a second data period from a boundary between the first data period and the second data period into the second data period. 3 and a fourth data period of a predetermined period from the boundary toward the first data period, and the third data period is set to the second data period. Exclude generates learning data from, or the detection data of the fourth data period to generate learning data by adding the fourth data period to the second data period when a normal data training data A generation unit, based on the learning data, a learning model generation unit that generates a learning model indicating the detection data of the sensor when the equipment is operating normally, the learning model, and when operating the equipment And an equipment state determining unit that determines the equipment state by comparing the detected data of the sensor with the detected data.

  Further, the learning data generation unit excludes the third data period from the second data period when the detection data in the third data period is abnormal data.

  Further, the equipment state determining unit compares the generated learning model with the detection data of the sensor detected during operation of the equipment, and determines that a difference between the learning model and the detection data is a predetermined value. A singularity extraction unit that extracts a singularity used for determining the equipment abnormality when the threshold value is exceeded.

Further, the equipment monitoring method of the present invention is a equipment monitoring method for monitoring the state of the equipment based on detection data of a sensor provided in the equipment, wherein the detection data, generation information of an alarm due to equipment abnormality, and the alarm Accumulate recovery information in time series, and associate the detection data in a time series with the occurrence information and the recovery information, and a first data period from an alarm occurrence to an alarm recovery in the detection data; The second data period from the alarm recovery to the generation of the alarm in the detection data is respectively determined, and the third data of the predetermined period from the boundary between the first data period and the second data period toward the second data period is determined. A data period and a fourth data period of a predetermined period from the boundary toward the first data period are set, and the third data period is excluded from the second data period. It generates learning data Te, or the fourth data interval and generates learning data in addition to the second data period when the detected data of the fourth data period is normal data, the training data A learning model indicating the detection data of the sensor when the facility is operating normally, and comparing the learning model with the detection data of the sensor detected during the operation of the facility. To judge the equipment state.

  According to the present invention, the accuracy of a learning model can be improved by excluding abnormality detection data from learning data for generating a learning model or adding normal detection data to learning data, and as a result, equipment abnormality Can be improved.

It is a block configuration diagram of a system including a facility monitoring device. It is a hardware block diagram of the computer of an equipment monitoring apparatus. 5 is a flowchart illustrating a process of setting a learning model according to the first embodiment. FIG. 3A is a characteristic diagram showing a relationship among history data, trend data, a learning model, and singularity extraction in the first embodiment. FIG. 3A is a characteristic diagram showing a relationship between history data, trend data, and a learning model. Shows a characteristic diagram of the generated learning model, and (C) shows a characteristic diagram for extracting a singular point. It is a schematic diagram of each margin when calculating the evaluation value of the learning model in the first embodiment, (A) is a schematic diagram in which the margin is not adjusted, and (B) is a diagram in which the initial margin is reduced. (C) is a schematic diagram in a state where the next margin is reduced, and (D) is a schematic diagram in a state where the next margin is further reduced. It is a flowchart which shows the singular point extraction process using the learning model in 1st Embodiment. FIG. 9 is a characteristic diagram of history data, trend data, and a learning model showing a modification of the first embodiment. FIG. 14 is a characteristic diagram of history data, trend data, and a learning model in the second embodiment. FIG. 13 is a characteristic diagram of history data, trend data, and a learning model showing a modification of the second embodiment.

  First, a first embodiment of the present invention will be described with reference to FIGS. FIG. 1 is a block diagram showing a schematic configuration of an equipment monitoring system 1 including an equipment monitoring device 10 according to the first embodiment. FIG. 4 is a characteristic diagram showing a relationship among history data, trend data, a learning model, and singular point extraction in the first embodiment.

  As shown in FIG. 1, the facility monitoring system 1 includes a building 2 and a facility monitoring device 10 connected to the building 2 via a network to monitor the building 2 remotely. The building 2 is provided with an air conditioner 3 as equipment. Further, the air conditioner 3 is provided with an air supply sensor SA for detecting the temperature of the air sent from the air conditioner 3 into the room and a return air sensor RA for detecting the temperature of the air drawn into the air conditioner 3 from the room. Although only one air conditioner 3 is shown in FIG. 1, a plurality of other air conditioners (not shown) are provided.

  The equipment monitoring device 10 includes a data storage unit 11 that stores various data relating to the building 2. The data storage unit 11 stores the air supply sensor SA, the return air sensor RA, and the ratio of these sensors SA, RA, that is, the detection data of the air supply sensor SA / return air sensor RA (hereinafter, these data are referred to as trend data TD). The trend data storage unit 11A that accumulates in time series, generation information indicating that an alarm has occurred due to an abnormality in the air conditioner 3, and recovery information (hereinafter, history data HD) indicating that the alarm has been restored are stored in time series. And a history data storage unit 11B for storing.

  Also, the equipment monitoring device 10 uses the trend data TD and the history data HD to generate learning data SD11, SD12, SD13,..., And the learning data SD11, SD12, SD13. And a facility state determination unit 14 that determines the state of the air conditioner 3 using the learning model SM. The generation of the detailed learning data SD11, SD12, SD13,... By the learning data generation unit 12 and the generation of the learning model SM by the learning model generation unit 13 will be described later.

  The equipment state determination unit 14 compares the detection data of the air supply sensor SA / return air sensor RA during operation of the air conditioner 3 with the learning model SM, and determines when the air conditioner 3 operates differently from the normal state. A singular point extraction unit 14A is provided for extracting as a singular point P (see FIG. 4). The singular point extraction unit 14A has a threshold L (see FIG. 4) for determining that the air conditioner 3 has operated differently from the normal state. The singular point P indicates a case where the air conditioner 3 is abnormal or a case where an abnormality may occur in the air conditioner 3. The singular point P indicates that the air conditioner 3 is operating differently from the normal state, and the extracted information of the singular point P is used as the maintenance and inspection information of the air conditioner 3.

  Although the building 2 is provided with a large number of sensors in addition to the air supply sensor SA and the return air sensor RA of the air conditioner 3, in the present embodiment, as an example, the air supply sensor SA / return air sensor RA Will be described.

  FIG. 2 is a hardware configuration diagram of a computer that forms the equipment monitoring device 10. The computer forming the equipment monitoring device 10 can be realized by a hardware configuration of a general-purpose personal computer (PC) that has existed before. That is, as shown in FIG. 2, the computer is provided with a CPU 21, a ROM 22, a RAM 23, an HDD controller 25 to which a hard disk drive (HDD) 24 is connected, a mouse 26 and a keyboard 27 provided as input means, and a display device. An input / output controller 29 for connecting the displays 28 and a network controller 30 provided as communication means are connected to the internal bus 31.

  The program for executing the facility monitoring method by the facility monitoring apparatus 10 can be provided not only by communication means but also stored in a computer-readable recording medium such as a CD-ROM or a DVD-ROM. The program provided from the communication means or the recording medium is installed in the computer, and various processes are realized by the CPU of the computer sequentially executing the program.

  Next, the control of the facility monitoring device 10, particularly the generation of the learning data SD11, SD12, SD13,... And the learning model SM will be described with reference to FIGS. The processing of the flowchart illustrated in FIG. 3 is performed in the learning data generation unit 12 and the learning model generation unit 13. In step S101 of the flowchart shown in FIG. 3, based on the alarm occurrence information and the recovery information stored in the history data storage unit 11B, a first data period from the alarm generation to the alarm recovery (hereinafter, an alarm generation period TA1). , TA2, TA3,...) And a second data period (hereinafter, referred to as a non-warning period TB1, TB2, TB3,...) From the recovery of the warning to the generation of the warning, and the process proceeds to step S102. In other words, in other words, an alarm is generated or an alarm is generated at the boundary between the first data period (alarm generation period TA1, TA2, TA3,...) And the second data period (non-alarm period TB1, TB2, TB3,...). There will be recovery.

  In step S102, first, the characteristics of the detection data of the air supply sensor SA / return air sensor RA are determined based on the trend data TD. The characteristics of the detection data of the air supply sensor SA / return air sensor RA are roughly determined in advance depending on the sensor type and the like.

  According to the characteristics of the detection data of the air supply sensor SA / return air sensor RA in the present embodiment, the trend data TD of the air supply sensor SA / return air sensor RA indicates a warning sign (sign of abnormality) before the alarm of the air conditioner 3 is generated. Has the property of detecting For example, when a slight abnormality occurs in the air conditioner 3 and the abnormality proceeds to generate an alarm of the air conditioner 3, the trend data TD gradually changes before the alarm is generated (from the time of occurrence of the abnormality). . That is, the trend data TD includes data indicating a sign of an abnormality before an alarm occurs. Since the characteristics of the detection data can be grasped in advance, the following control is performed based on the characteristics. As described above, in the present embodiment, a case will be described in which the trend data TD includes data indicating a sign of abnormality before the occurrence of an alarm.

  In step S102, as shown in FIG. 4A, the boundary between the alarm generation periods TA1, TA2, TA3... And the non-warning periods TB1, TB2, TB3. The margins σ11, σ12, σ13,... As the third data period are set in the period. The margins σ11, σ12, σ13,... Correspond to a period including detection data indicating a sign of abnormality before the occurrence of an alarm. Next, learning data SD11, SD12, SD13,... Are generated from the non-warning periods TB1, TB2, TB3,. The learning data SD11, SD12, SD13,... Exclude data indicating a sign of abnormality before an alarm occurs.

  In step S103, a learning model SM is generated based on the learning data SD11, SD12, SD13,... Generated in step S102. Since a method for generating the learning model SM based on the learning data SD11, SD12, SD13,... Is known, detailed description thereof is omitted. By modeling the learning data SD11, SD12, SD13,..., A model for predicting detection data of the air supply sensor SA / return air sensor RA detected when the air conditioner 3 is in a normal state is generated. Then, the process proceeds to step S104.

  In step S104, the evaluation value of the generated learning model SM is calculated, and the process proceeds to step S105. In step S105, the margins σ11, σ12, σ13,. That is, if the periods of the margins σ11, σ12, σ13,... Are lengthened, the detection data indicating the sign of the abnormality before the alarm is generated can be surely excluded, but the normal detection data is also excluded. Conversely, if the periods of the margins σ11, σ12, σ13,... Are shortened, it is not possible to exclude the detection data indicating the sign of abnormality before the alarm is generated, and the detection data indicating the sign of abnormality is included. there is a possibility. Therefore, the margins σ11, σ12, σ13,. In order to optimize the margins σ11, σ12, σ13,..., An optimization algorithm called metaheuristics is used. In particular, in the present embodiment, the margins σ11, σ12, σ13,... Are optimized using the optimization algorithm of the hill-climbing method of the neighborhood search method.

  The optimization of the margins σ11, σ12, σ13... Will be described with reference to FIG. The optimization of the margins σ11, σ12, σ13,... Gradually reduces the values of the margins σ11, σ12, σ13,. In FIG. 5A, initially, the optimal values of the margins σ11, σ12, σ13,... Are not set, so that the margins σ11, σ12, σ13,. The learning model SM is generated based on this setting. The generated learning model SM is compared with evaluation data (normal data) for evaluating the learning model SM, and an evaluation value for the learning model SM is calculated.

  The evaluation data is detection data of the air supply sensor SA / return air sensor RA determined by the operator of the equipment monitoring device 10 to be normal data, and serves as an index of the evaluation value. The evaluation value indicates a scale by which the learning model SM approximates the evaluation data. The higher the evaluation value, the closer the learning model SM and the evaluation data are. That is, the learning model SM having a high evaluation value is close to normal detection data.

  After calculating the evaluation value, in step S105, it is determined whether the learning model SM is optimal. As a method of determining the optimal learning model SM, in the present embodiment, steps S102 to S104 are repeated according to the number of margins, and a learning model SM that provides the highest evaluation value is selected.

  As shown in FIG. 5A, after calculating the first evaluation value, steps S102 to S104 are repeated, and then, as shown in FIG. 5B, only the margin σ11 is set to a smaller value. Then, learning data SD11, SD12, SD13,... Are generated to generate a learning model SM. The evaluation value of the learning model SM is calculated by comparing the learning model SM with the evaluation data. Further, steps S102 to S104 are repeated to set only the margin σ12 to a small value as shown in FIG. 5C, generate learning data SD11, SD12, SD13,. Generate. The evaluation value of the learning model SM is calculated by comparing the learning model SM with the evaluation data. Similarly, as shown in FIG. 5D, the learning model SM is created by setting only the margin σ13 to a small value, and the evaluation value is calculated.

  As described above, steps S102 to S104 are repeated according to the number of margins. As a result, among the obtained evaluation values, the learning model SM having the highest evaluation value is set as the optimal learning model SM. That is, the learning model SM closest to the evaluation data is set. In FIG. 5, when the margin σ12 is set to a small value, the evaluation value becomes the evaluation value 80, which indicates that this learning model SM is the learning model SM closest to the evaluation data as compared with the other learning models SM. ing. As a result, in step S106, the learning model SM when the margin σ12 is set to a small value, here, the learning model SM with the evaluation value 80 is set as the learning model SM used when the air conditioner 3 is operated, and the learning is performed. The setting of the model SM ends. FIG. 4B shows the set learning model SM.

  Note that the target evaluation value of the learning model SM may be set in advance, and steps S102 to S105 may be repeated until the calculated evaluation value approaches the target evaluation value. In addition, as an optimization algorithm, in addition to the hill-climbing method used in the present embodiment, annealing method, neighborhood search method such as taboo search, genetic algorithm, evolution strategy, evolutionary programming, genetic programming, ant colony optimization, Evolutionary computation methods such as particle swarm optimization can also be applied. Also, algorithms such as simulated, evolution, artificial immune system, and neural network can be applied.

  Next, control for judging the state of the detection data of the air supply sensor SA / return air sensor RA detected during the operation of the air conditioner 3 using the set learning model SM will be described. The processing of the flowchart shown in FIG. 6 is performed in the singular point extraction unit 14A of the equipment state determination unit 14. In step S110 of the flowchart shown in FIG. 6, the difference between the learning model SM set in step S106 of the flowchart of FIG. 3 and the detection data of the air supply sensor SA / return air sensor RA detected during the operation of the air conditioner 3 is calculated. After the calculation, the process proceeds to step S111. FIG. 4C shows the characteristic TM of the difference between the two.

  In step S111, it is determined whether or not the difference between the two exceeds a predetermined threshold L. Here, the threshold value L is for determining that the learning model SM and the detection data of the air supply sensor SA / return air sensor RA have greatly deviated. In the characteristic TM, a point where the difference between the two exceeds the threshold value L is extracted as a singular point P.

  As shown in FIG. 4C, when the difference between the two is small, the difference does not exceed the threshold L. However, when the detection data of the air supply sensor SA / return air sensor RA greatly deviates from the learning model SM, the difference between the two becomes large, and exceeds the threshold value L as indicated by the symbol P in FIG. This code P is extracted as a singular point P. The singular point P is detection data that is significantly different from the learning model SM, and indicates that the air conditioner 3 may have an abnormality or that the air conditioner 3 has an abnormality. The singular point P is a point where the difference between the learning model SM and the detection data of the air supply sensor SA / return air sensor RA exceeds the threshold value L, and is not necessarily directly related to the occurrence of an abnormality in the air conditioner 3. Absent. That is, the singular point P may be extracted even if the alarm of the air conditioner 3 does not occur. Therefore, when the singular point P is extracted, the operator is notified of the fact that the singular point P is extracted.

  If the singular point P is not extracted in step S111 (No), the process returns to step S110, acquires the detection data of the air supply sensor SA / return air sensor RA again, and repeats steps S110 and 111. When the singular point P is extracted in step S111 (Yes), the process proceeds to step S112, and information on the extracted singular point P is used as maintenance and inspection information of the air conditioner 3. Thereafter, the detection data of the air supply sensor SA / return air sensor RA is acquired again, and steps S110 and S111 are repeated.

  As described above, when the learning model SM is generated, the learning data SD11, SD12, SD13,... Are generated excluding the detection data indicating the sign of the abnormality of the air conditioner 3 before the alarm is generated. The learning model SM can be generated based on normal detection data. As a result, the accuracy of the learning model SM can be improved, and the accuracy of the state determination of the air conditioner 3 using the learning model SM can also be improved.

  Further, when the detection data of the air supply sensor SA / return air sensor RA has a characteristic that requires a predetermined period to return to the normal detection state after the alarm is restored, that is, the detection data of the air supply sensor SA / return air sensor RA Is the characteristic including the detection data in the unstable state after the alarm is restored, as shown in FIG. 7, the margins σ21, σ22, σ23,.・ ・ Set

  In FIG. 7, the boundaries between the alarm generation periods TA1, TA2, TA3,... And the non-warning periods TB1, TB2, TB3,. Set. The margins σ21, σ22, σ23... Correspond to a period including the detection data in an unstable state after the alarm recovery. Next, learning data SD21, SD22, SD23,... Are generated from the non-warning periods TB1, TB2, TB3,. Then, a learning model SM is generated based on the learning data SD21, SD22, SD23. The optimization of the margins σ21, σ22, σ23... And the setting of the optimal learning model SM are the same as in the above-described embodiment.

  As described above, by excluding the detection data in the unstable state after the alarm recovery according to the sensor characteristics, the accuracy of the learning model SM can be improved. In addition, both the detection data indicating the sign of abnormality before the occurrence of the alarm described above and the detection data in an unstable state after the recovery from the alarm may be excluded, or either one of the detection data may be excluded. May be excluded.

  Next, a second embodiment will be described with reference to FIG. In the first embodiment, the margins σ11, σ12, σ13,... Are set before the alarm is generated. However, in the second embodiment, the detection data of the air supply sensor SA / return air sensor RA indicates that the alarm is generated. When normal detection data is detected for a predetermined time later, for example, when the detection data of the air supply sensor SA / return air sensor RA changes after a predetermined time from the generation of the alarm, that is, when the detection data changes slightly after the generation of the alarm , The margins σ31, σ32, σ33,... Are set as the fourth data period after the alarm is generated.

  In FIG. 8, similarly to the first embodiment, based on the alarm generation information and the recovery information stored in the history data storage unit 11B, the alarm generation periods TA1, TA2, TA3,. .. And the non-warning periods TB1, TB2, TB3,...

  Next, margins σ31, σ32, σ33,. Set. The margins σ31, σ32, σ33,... Correspond to the period of the detection data that has hardly changed from the detection data before the occurrence of the alarm. In other words, it can be considered that the detection data is substantially the same as the normal detection data before the alarm is generated.

  .. Are added to the non-warning periods TB1, TB2, TB3,... To generate learning data SD31, SD32, SD33,. After that, learning models SM corresponding to the learning data SD31, SD32, SD33,... Are generated. The evaluation and optimization of each learning model SM are the same as those in the first embodiment, and a description thereof will be omitted. The learning model SM with the highest evaluation value is set as the learning model SM used when the air conditioner 3 is operated, and the air supply sensor SA // detected when the air conditioner 3 is operated using the set learning model SM. The state of the detection data of the return air sensor RA is determined. The determination of the state of the detection data is performed by extracting the singular point P.

  As described above, when the learning model SM is generated, the detection data that hardly changes from the normal state after the alarm is generated is added to the learning data SD31, SD32, SD33, and so on, so that the learning data SD31, SD32, SD33,. .. Can increase the amount of data. By increasing the data amount of the learning data SD31, SD32, SD33,..., An accurate learning model SM can be generated, and as a result, the accuracy of the state determination of the air conditioner 3 can be improved. In particular, when the amount of learning data is small, the amount of learning data can be increased, which is effective.

  When the detection data of the air supply sensor SA / return air sensor RA detects normal detection data before the alarm is restored, for example, the detection data is changed to a substantially normal state at the same time as the equipment abnormality is resolved. After that, when the alarm is restored, margins σ41, σ42, σ43... As the fourth data period are set before the alarm is restored.

  As shown in FIG. 9, by setting margins σ41, σ42, σ43... Before the alarm recovery, the detection data of the air supply sensor SA / return air sensor RA before the alarm recovery is substantially normal. Can be taken in, and the data amount of the learning data SD41, SD42, SD43,... Can be increased.

  In this manner, the learning data SD41, SD42, SD43,... Fetching the detection data indicating substantially the same state as the normal state after the alarm is generated or before the alarm is restored into the learning data SD41, SD42,. , SD43... Can be increased, and an accurate learning model SM can be obtained. Note that, depending on the sensor characteristics, both detection data after the alarm is generated and before the alarm is restored may be captured, or any one of the detection data may be captured.

  REFERENCE SIGNS LIST 1 equipment monitoring system, 2 building, 3 air conditioner, 10 equipment monitoring device, 11 data storage unit, 11A trend data storage unit, 11B history data storage unit, 12 learning data generation unit, 13 learning model generation unit, 14 equipment state judgment Unit, 14A Singularity point extraction unit, RA return air sensor, SA air supply sensor, SD11, SD12, SD13, SD21, SD22, SD23, SD31, SD32, SD33, SD41, SD42, SD43 Learning data, SM learning model, TA alarm Occurrence period, TB non-warning period, σ11, σ12, σ13, σ21, σ22, σ23, σ31, σ32, σ33, σ41, σ42, σ43 margin.

Claims (4)

A facility monitoring device that monitors a state of the facility based on detection data of a sensor provided in the facility,
A data accumulation unit that accumulates the detection data, alarm occurrence information due to equipment abnormality, and alarm recovery information in time series,
Correlating the detection data in a time series with the occurrence information and the recovery information, a first data period from the alarm generation to the alarm recovery in the detection data, and a first data period from the alarm recovery to the alarm generation in the detection data. 2 data periods, a third data period of a predetermined period from the boundary between the first data period and the second data period to the second data period, and the first data period from the boundary. A fourth data period of a predetermined period going into the period is set, learning data is generated by excluding the third data period from the second data period, or detection data of the fourth data period is generated. A learning data generating unit configured to add the fourth data period to the second data period to generate learning data when the data is normal ;
Based on the learning data, a learning model generation unit that generates a learning model indicating the detection data of the sensor when the equipment is operating normally,
The learning model, an equipment state determination unit that determines the equipment state by comparing the detection data of the sensor detected during operation of the equipment,
An equipment monitoring device comprising:   The equipment monitoring according to claim 1, wherein the learning data generator excludes the third data period from the second data period when the detection data in the third data period is abnormal data. apparatus. The equipment state determining unit compares the generated learning model with the detection data of the sensor detected during operation of the equipment, and determines a difference between the learning model and the detection data to be a predetermined threshold. when it exceeds, equipment monitoring device according to claim 1 or 2, characterized in that it comprises a singular point extraction unit that extracts as a singular point to be used for the equipment abnormality judgment. A facility monitoring method for monitoring a state of the facility based on detection data of a sensor provided in the facility,
The detection data, information on the occurrence of alarms due to equipment abnormalities and information on the recovery of the alarms are stored in chronological order,
Correlating the detection data in a time series with the occurrence information and the recovery information, a first data period from the alarm generation to the alarm recovery in the detection data, and a first data period from the alarm recovery to the alarm generation in the detection data. And 2 data periods respectively,
A third data period of a predetermined period from the boundary between the first data period and the period of the second data into the second data period, and a fourth data period of a predetermined period from the boundary to the inside of the first data period. Data period and
The learning data is generated by excluding the third data period from the second data period, or when the detection data in the fourth data period is normal data, the fourth data period is changed to the second data period. To generate learning data,
Based on the learning data, to generate a learning model indicating the detection data of the sensor when the equipment is operating normally,
An equipment monitoring method, comprising: comparing the learning model with the detection data of the sensor detected during operation of the equipment to determine an equipment state.

Images