JP2006135412A - Remote supervisory system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a remote supervisory system with high detection sensitivity early and accurately detecting a fault of a supervisory object facility. <P>SOLUTION: The remote supervisory system includes: a sensor information acquisition section for acquiring a sensor value of a supervisory object; a prediction model buildup section that obtains a first correlation between sensor values in a normally operating state, builds up the first correlation as a basic prediction model for detecting a fault, obtains a second correlation between sensor values of parts of the sensors, and builds up the second correlation as a prescribed failure prediction model whose sensing sensitivity is higher in a particular fault of a facility than that of the basic prediction model; and a fault detection section for detecting the presence/absence of the fault of the facility on the basis of a difference between the sensor value to be acquired and the prediction sensor value during a supervisory period. When sensing the fault in the facility on the basis of the basic prediction model, the prediction model buildup section determines combinations of the sensors for maximizing the detection sensitivity with respect to the fault and builds up the prediction model for the prescribed fault with respect to the fault. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a system for remotely monitoring a plurality of facilities such as an air conditioner and a generator, and more particularly to a remote monitoring system that can detect a failure of a monitored facility at an early stage and correctly identify a failure site. .

  Conventionally, methods based on the concepts of post-maintenance and time-based maintenance have been used as maintenance methods for power equipment and the like. However, the post-maintenance that repairs after the occurrence of a failure has a problem that the operation of the facility must be stopped for a long time for the repair. Further, in the time-based maintenance in which the maintenance work is periodically performed, the maintenance work is performed even when the target equipment is in a healthy state, and thus there is a problem that time and cost are wasted.

  Therefore, there is a concept of state-based maintenance that performs appropriate maintenance based on the state of the target equipment. In such a method, the values of various sensors provided in the equipment and the status of each part of the equipment are inspected, and the result Maintenance work is determined based on the above. However, if there are multiple equipment to be maintained and they are scattered in remote locations, it is difficult for the workers to check the status of each equipment at the site. A system has been proposed to monitor the equipment from its history.

In such a remote monitoring system (for example, the device described in Patent Literature 1, 2, or 3 below), the value of a sensor provided in each facility is received by a monitoring system provided in a location away from the facility, and the system Determines whether the equipment is normal or abnormal based on the value. The determination method is generally based on whether the system notifies normality / abnormality for each sensor based on whether or not the value of each received sensor exceeds a preset threshold value. A person, such as a supervisor, determines the occurrence of a failure of a facility or a failure part. There is also a method in which the system determines that the equipment is faulty when some predetermined sensors have abnormal values.
JP 2001-41615 A JP 2003-4281 A JP 2001-21192 A Japanese Patent Application Laid-Open No. 2004-169989

  However, in the above conventional method, there is a problem that false detection occurs due to noise of each sensor that is closely related to, for example, temperature, load fluctuation, and other external factors. In order to avoid such false detection, a threshold value is avoided. However, there is a problem that it may be difficult to detect that a failure has occurred at an early stage.

  The technique disclosed in Patent Document 4 detects a failure using a statistically significant difference at the time of abnormality in univariate analysis and a correlation coefficient by two variables. Therefore, as the number of sensors increases, there is a problem that it is difficult to detect the abnormality of the entire monitoring target with high sensitivity.

  Accordingly, an object of the present invention is a system for remotely monitoring a plurality of equipment such as power equipment, which can detect a failure of the equipment to be monitored at an early stage, and has a relatively large number of sensors provided in the power equipment. Both are to provide a remote monitoring system that can detect a failure with high sensitivity.

  In order to achieve the above object, a first configuration of a remote monitoring system of the present invention is a remote monitoring system for remotely monitoring equipment having a plurality of sensors. Sensor values (measured sensor values) detected by the plurality of sensors are used. Obtaining a first correlation between sensor values of all of the plurality of sensors acquired when the equipment is operating normally, and a sensor information acquisition unit that acquires via a predetermined communication line, the first A correlation is constructed as a basic prediction model for detecting a failure of the equipment, a second correlation between sensor values of some sensors in the plurality of sensors is obtained, and the second correlation is A prediction model construction unit that constructs a specific failure prediction model having a higher detection sensitivity than the basic prediction model for a specific failure of the facility, and the sensor information acquisition unit acquires the facility during a monitoring period. A failure detection unit that detects the presence or absence of a failure of the equipment based on a difference between the sensor value to be detected and a prediction sensor value obtained from each of the basic prediction model and the prediction model for specific failure, and the prediction model construction A combination of sensors composed of a part of the plurality of sensors such that when the failure detection unit detects a failure of the facility based on the basic prediction model, the detection sensitivity to the failure is maximized. And determining the second correlation between the sensor values of the partial sensors to construct the specific failure prediction model for the failure.

  Further, in the second configuration of the remote monitoring system of the present invention, in the first configuration, the first correlation and the second correlation obtained by the prediction model construction unit are obtained by principal component analysis. It is characterized by that.

  The third configuration of the remote monitoring system according to the present invention is the difference between the predicted sensor value and the sensor value acquired by the sensor information acquiring unit in the first or second configuration. The combination of the sensors is determined based on the S / N ratio.

  Further, in a fourth configuration of the remote monitoring system of the present invention, in any one of the first to third configurations, when the failure detection unit detects a failure of the facility based on the basic prediction model, It further comprises a failure type determination unit that determines the type of failure based on the tendency of the difference between the sensor value acquired by the sensor information acquisition unit and the predicted sensor value.

  Furthermore, a remote monitoring program according to the present invention for achieving the above object is a remote monitoring program for causing a remote monitoring system to execute a process of remotely monitoring a facility including a plurality of sensors. The first correlation between the sensor values of all of the plurality of sensors acquired when the first process is acquired via a predetermined communication line and the facility is operating normally , Constructing the first correlation as a basic prediction model for detecting a failure of the equipment, obtaining a second correlation between sensor values of some sensors in the plurality of sensors, A second process of constructing the correlation as a prediction model for specific failure with a higher detection sensitivity than the basic prediction model for the specific failure of the facility, and monitoring the facility And detecting the presence or absence of a failure of the equipment based on the difference between the sensor value acquired by the first process and the predicted sensor value obtained from each of the basic prediction model and the specific failure prediction model. In the third process, the failure of the equipment is detected based on the basic prediction model in the third process. A combination of sensors composed of a part of the plurality of sensors that maximizes the detection sensitivity to the failure is determined, and the second correlation between sensor values of the part of the sensors is determined. It is a process for constructing the specific failure prediction model for the failure by obtaining the relationship.

  Preferably, in the second processing of the remote monitoring program, the first correlation and the second correlation are obtained by principal component analysis.

  Preferably, in the second process of the remote monitoring program, the combination of the sensors is determined based on the SN ratio of the difference between the sensor value acquired by the sensor information acquisition unit and the predicted sensor value.

  In the third process, preferably, the remote monitoring program may detect the sensor value obtained by the first process and the prediction when a failure of the facility is detected based on the basic prediction model. Based on the tendency of the difference from the sensor value, the remote system is further caused to execute a fourth process for determining the type of failure.

  According to the remote monitoring system and the remote monitoring program of the present invention, it becomes possible to detect a failure of a power facility with high sensitivity for each type of failure, and an early and accurate failure detection is desired.

  Embodiments of the present invention will be described below with reference to the drawings. However, such an embodiment does not limit the technical scope of the present invention.

  FIG. 1 is a diagram illustrating a configuration example of a remote monitoring system according to an embodiment of the present invention. The remote monitoring system 1 remotely monitors a plurality of power equipments 2 installed at a plurality of locations. Such a remote monitoring system 1 can detect failures in the power equipment 2 early and with high sensitivity based on the correlation between sensor values of a plurality of sensors 21 (21a, 21b, 21c,...) Attached to each power equipment 2. It is something to be detected.

  As described above, the power facility 2 shown in FIG. 1 is a facility to be monitored by the remote monitoring system 1, and each power facility 2 includes a plurality of sensors 21. The power equipment 2 is, for example, a generator or a gas engine heat pump air conditioner (GHP), and the sensor 21 is, for example, a thermometer or a pressure gauge attached to each part of the equipment. Further, the communication line 3 shown in FIG. 1 is a line for transmitting values (for example, temperature, pressure, etc.) detected by the sensors 21 to the remote monitoring system 1, and each power equipment 2 and the remote monitoring system. Connecting 1

  The remote monitoring system 1 includes a sensor information acquisition unit 11, a prediction model construction unit 12, a prediction model storage unit 13, a failure detection unit 14, a failure type determination unit 15, and a failure model storage unit 16 as illustrated. It can be configured by a so-called computer system including a CPU, a memory, an OS, an application program, and the like.

  The sensor information acquisition unit 11 is a part that forms an interface with the power equipment 2 and receives the values of the sensors 21 periodically detected and transmitted by the plurality of sensors 21 via the communication line 3. Store.

  The prediction model construction unit 12 constructs a correlation between the sensor values of the plurality of sensors 21 from the values of the plurality of sensors 21 acquired when the power equipment 2 is operating normally. The constructed correlation is called a prediction model. A specific method for constructing the prediction model will be described later. For example, a correlation between a plurality of predetermined sensors 21 is constructed as a mathematical expression by multivariate analysis such as principal component analysis. For example, the correlation between the values detected by the three sensors 21a, 21b, and 21c of the power facility 2 shown in FIG. 1 during normal operation is modeled by a mathematical formula, and the values of the three sensors 21a, 21b, and 21c are calculated by this prediction model. Can be predicted.

  Further, in the present embodiment, as described above, the prediction model construction unit 12 is a basic prediction model (hereinafter referred to as basic prediction) from the sensor values of all of the plurality of sensors 21 acquired during normal operation. In addition, a prediction model for specific faults with increased detection sensitivity only for specific faults is also built. When there are a plurality of types of failures, a plurality of specific failure prediction models are constructed for each type of failure.

  As will be described in detail later, this specific failure prediction model performs principal component analysis when a certain failure occurs, excluding the values (variables) of the sensors, and performs SPE (square prediction error) for the failure. The combination of sensors that maximizes the S / N ratio is determined, and a prediction model based on the sensor values (data variables) of some of the sensors that are excluded from a certain number from all the sensors is constructed. That is, before and after the occurrence of a specific failure, by removing sensor values whose value changes are relatively small from the target of the prediction model construction, it is possible to construct a prediction model with high detection sensitivity for the failure.

  Note that the prediction model construction unit 12 can be configured by a program describing the procedure of construction processing of a prediction model, an arithmetic device that executes processing according to the program, and the like.

  The prediction model storage unit 13 stores the basic prediction model generated by the prediction model construction unit 12 and a plurality of specific failure prediction models. The prediction model storage unit 13 can be configured by a storage device such as a hard disk device.

  The failure detection unit 14 detects a failure based on the sensor value acquired by the sensor information acquisition unit 11 during operation of the power equipment 2 and the prediction model stored in the prediction model storage unit 13. Although details of the failure detection method will be described later, a difference between the predicted sensor value based on the prediction model and the actual sensor value is obtained, and it is determined whether the power equipment 2 is in an abnormal state based on the difference. . In the comparison with the prediction model, a comparison is made between the basic prediction model stored in the prediction model storage unit 13 and all the prediction models for specific failure, and if any abnormality is detected by any prediction model, the abnormality is detected. It is judged. The specific failure prediction model has higher detection sensitivity than the basic prediction model with respect to a specific failure. Therefore, in the case of an abnormality due to the specific failure, the abnormality is usually detected by the specific failure prediction model. In the case of a failure (new type of failure) other than the failure in the prediction model for specific failure stored in the prediction model storage unit 13, an abnormality is detected by the basic prediction model, and the failure type determination unit 15 described below detects a failure. And a specific failure prediction model for the failure is constructed, stored, and used from the next time.

  The failure type determination unit 15 specifies the type of failure when the failure detection unit 14 detects a failure based on the basic prediction model. Specifically, the type of failure is determined by comparing the difference between a failure model stored in a failure model storage unit 16 to be described later and a predicted sensor value obtained by the failure detection unit 14 and an actual sensor value. Although the details of the determination method will be described later, the failure type is information for identifying the content of the failure such as a failure part and a failure phenomenon. When the failure type is determined by the failure type determination unit 15, the prediction model construction unit 12 constructs a specific failure prediction model corresponding to the failure.

  The failure model storage unit 16 stores a failure model used by the failure type determination unit 15. The failure model means a failure pattern in which a difference pattern (trend) between a predicted sensor value and an actual sensor value at the time of the failure, such as a failure experienced in the past, is known and stored in advance. . Although details of such a failure model will be described later, this failure model is generally determined for each failure type of the power equipment 2.

  A prediction model construction method in the prediction model construction unit 12 will be described. The prediction model construction unit 12 first constructs a basic prediction model. Specifically, first, after the power facility 2 to be monitored is newly installed or after the repair of the power facility 2 is completed, the power facility 2 is started and a test operation is performed to confirm that it is normal operation. Check. Thereafter, detection is performed periodically (for example, every hour) by a plurality of sensors 21 attached to the power facility 2, and the detected sensor values are sequentially transmitted to the remote monitoring system 1 via the communication line 3. In the remote monitoring system 1, the sensor information acquisition unit 11 sequentially receives and stores transmitted sensor values.

  The above equation (1) exemplifies a time series of data acquired at a certain time interval by the sensor information acquisition unit 11. Sensor values (data variables) acquired from each sensor at a certain time are arranged in a column and are overlapped in the row direction in time series.

  The sensor information acquisition unit 11 continuously receives and stores sensor values, but at the time of acquiring a predetermined number of sensor values (data variables), the prediction model construction unit 12 starts the construction process of the prediction model. To do. Note that the predetermined number of times is a number necessary for constructing the prediction model, for example, a number sufficient to perform so-called learning by performing statistical processing to be described later. Determined.

  As described above, the construction of the prediction model is to determine the correlation between the sensor values in the plurality of sensors 21, and the prediction model construction unit 12 uses, for example, the sensor values for the determined number of times, for example, the main component Using an analysis method, a mathematical formula that defines the correlation between sensor values is obtained. The principal component analysis method will be described below.

  First, the data group of Formula (1) is plotted on the m-dimensional variable coordinate axis. The dimension represents the number of data variables (sensor values) acquired at one time. Here, m = 3 for visual explanation. FIG. 2 is a diagram showing data groups plotted in a three-dimensional space.

  A new axis is taken in the direction of the largest variance in the coordinate system in which the data is plotted, and this is called the main component. In particular, the first main component is expressed as a first main component (PC1). The second principal component (PC2) is taken in the direction orthogonal to the first principal component and secondly in the largest dispersion. In this way, the principal components that can be taken are equal to the dimension of the data, but if there is a correlation between the data variables to be modeled, only some of those principal components will have their properties approximately Can be represented. This is called dimension reduction, and means modeling for representing multivariate information in a smaller dimension (the number of data variables remains m). Each main component can be expressed by the following equation (2).

Here, the vector t is called a score vector and represents the distance from the origin when each data point is projected onto the principal component axis. p is a loading vector and represents a cosine formed by a principal component axis and a variable axis. In formula (2), t ₁ p ₁ represents the first principal component, and t _k p _k represents the k-th principal component. FIG. 3 is a diagram showing a score vector t and a loading vector p. As an example, score vectors t ₁ and t ₂ and a loading vector p ₁ when two data points x ₁ and x ₂ are projected onto the axis of the first principal component PC ₁ are shown.

  In FIG. 3, the distance projected from the data point onto the principal component axis indicates an error. Therefore, the above equation (1) can be expressed as the following equation (3) by the error from the above equation (2).

In this way, a normal model is constructed. That is, the construction of the prediction model is a process for determining the loading vector p that defines the principal component axis. The basic prediction model is a prediction model constructed using sensor values (data variables) from all m sensors. In the prediction model, all the data points x that are sensor values are predicted to exist on the principal component axis, and by monitoring whether the magnitude of the error from the principal component axis exceeds a predetermined allowable range, a failure ( Determine whether there is any abnormality. A parameter called Q statistic is used as an abnormality determination parameter. The Q statistic is also called a squared prediction error (SPE) and is defined as the following equation (4).

here,

Is the coordinate of the data point (actual sensor value),

Is

The coordinates (predicted sensor value) projected on the principal component coordinates from the above are shown. A large SPE means that there is a data point on the m-a residual subspace that deviates from the a-dimensional normal model, suggesting the possibility of an abnormal situation.

  Next, the construction method of the specific failure prediction model will be described. The specific failure prediction model is newly constructed every time an abnormality is detected by the basic prediction model. If an abnormality is detected by the basic prediction model without being detected by the stored prediction model for specific failure, which has higher detection sensitivity than the basic prediction model for a specific failure, Therefore, a specific failure prediction model for the failure is newly constructed. The specific failure prediction model is basically constructed using the same principal component analysis method as the basic prediction model described above, but at this time, data variables that are relatively unrelated to the specific failure are selected. Delete and construct by principal component analysis using remaining data variables (number less than m). As a method for selecting a data variable to be deleted, in the embodiment of the present invention, a principal component analysis is performed when data variables when a failure occurs in the power unit 2 one by one, and the square at that time is analyzed. Data variables that maximize the SNR of the prediction error (SPE) (maximum sensitivity) are removed. The process of deleting data variables one by one is repeated until the SPE SN ratio no longer increases (sensitivity is high), and the combination of data variables that maximizes the SPE SN ratio is determined.

  FIG. 4 is a flowchart of data variable selection processing. The initial value in step S100 is set, and in step S101, the i-th variable is deleted. The first is the first variable. In step S102, it is confirmed that i is not larger than m, and principal component analysis is performed when the i-th variable is deleted (S103). The dimension which can be taken at this time is m-1. Then, the SN ratio (SNRQi) of the square prediction error SPE is obtained and stored (S104). The i-th variable is returned (S105), i = i + 1 is set (S106), the process returns to step S101, and the next variable is deleted. When each SNRQi is obtained by deleting up to m-th variables, the process proceeds to step S107. In step S107, m variables are sequentially deleted, and the obtained SNRQi values are compared. If the SNRQi when the j-th variable is deleted is maximum, the SNRQi is set to the maximum value Max [SNRQi. (m-1)] j.

  The maximum value Max [SNRQi (m-1)] j is compared with the maximum value Max so far (S108), and if it is larger than the maximum value Max, the maximum value Max [SNRQi (m-1)] j is converted to the maximum value Max. (S109). Then, the jth variable is deleted (S110), m = m-1 is set (S111), the process returns to step S100, and the above processing is repeated. Since the deleted variable exists from the second round, if i = i + 1 and the number of the deleted variable is reached, the process proceeds to the next number not deleted. In step S108, when the maximum value Max [SNRQi (m-1)] j does not exceed the maximum value Max, the process ends. At the end of the process, the remaining variables are combinations of variables that maximize the SN ratio of the square prediction error SPE.

  Thus, when a combination of data variables (sensors) is determined, the above-described principal component analysis is performed based on the data variables of the combination, and a specific failure prediction model having a high sensitivity to a generated failure can be constructed. it can. In the embodiment described later, an embodiment of the data variable selection process will be described.

  FIG. 5 is a flowchart illustrating the processing contents in the embodiment of the present invention. After the power equipment 2 to be monitored is newly installed or after repair of the power equipment 2 is completed, the power equipment 2 is started and a test operation is performed to confirm that it is in normal operation. In this state, the sensor information acquisition unit 11 of the remote monitoring system 1 sequentially receives and stores sensor values transmitted from the plurality of sensors 21 (S11). First, since it is necessary to construct a basic prediction model, the number of data necessary to construct it is acquired (S12), and the prediction model construction unit 12 uses the above-described principal component analysis method, A basic prediction model is constructed and stored in the prediction model storage unit 13 (S13).

  When the basic prediction model is constructed and stored, the actual monitoring process starts. Each time the sensor information acquisition unit 11 receives a sensor value transmitted from the power equipment 2 (S21), the failure detection unit 14 determines whether or not the received sensor value is an abnormal value. Specifically, the predicted value of the sensor value in each model is obtained from the basic prediction model constructed by the prediction model construction unit 12 or the specific failure prediction model described later and the acquired sensor value, and the predicted value and the received sensor value ( A difference from the actual measurement value) is obtained, and whether or not there is an abnormality is analyzed based on the magnitude of the difference (S22), and whether or not there is an abnormality (failure) is determined (S23). A difference is calculated | required by the square prediction error (SPE) by the said (4) Formula. In equation (4), the predicted sensor value is

The measured sensor value is

It is. When the obtained square prediction error (SPE) exceeds the 95% confidence limit value, it is determined that the sensor value (actually measured value) is abnormal. The 95% confidence limit value is a value normally used in statistical analysis, and means a value that does not exceed this value with a probability of 95%. The 95% confidence limit is predetermined.

  The failure detection unit 14 preferably determines that there is a failure in the power equipment when it is determined that the acquired sensor value is abnormal continuously for a predetermined number of times. This is because there may be a case where the sensor value shows an abnormal value on a one-off basis even during normal operation without any failure in the power equipment 2. For example, when the above square prediction error SPE exceeds the 95% confidence limit value 10 times, it is determined that there is a failure. Of course, it may be determined that there is a failure by detecting the abnormality of the sensor value once.

  If no failure is detected as a result of the failure detection in the failure detection unit 14 (No in step S23), monitoring by the remote monitoring system 1 such as stopping the power equipment 2 to be monitored is performed. As long as there is no reason to end (No in step S24), the processing from step S21 of FIG. 6 described above is repeatedly executed. If there is a reason for terminating the monitoring (Yes in step S24), the monitoring is terminated.

  On the other hand, when the failure detection unit 14 determines that there is a failure, it is further determined whether the determination is based on the basic prediction model or the specific failure prediction model (S25). If it is based on the comparison with the specific failure prediction model, the failure type is specified as a failure corresponding to the specific failure prediction model, and failure information is issued (S33). In the case of comparison with the basic prediction model, since the content of the failure is unknown, it is necessary to determine the content (type) of the failure.

  FIG. 6 is a diagram for explaining failure type determination processing. First, the failure type determination unit 15 of the remote monitoring system 1 analyzes the difference between the predicted sensor value obtained by the failure detection unit 14 and the actual measurement value (step S31 in FIG. 5). Specifically, as shown in FIG. 6A, the difference (residual vector) between each data point x1, x2, x3 and the principal component axis t when three sensor values a, b, c are acquired. E1), e2, and e3. The square sum of the residual vector e is a square prediction error SPE. The failure type determination unit 15 extracts the components (ea, eb, ec) of each residual vector e and stores them in the failure model storage unit 16 as the “actual value pattern” shown in FIG. Compare with multiple failure models.

  FIG. 6B is a diagram illustrating the determination of the failure type more specifically. What is shown on the left side of the figure is an example of a failure model stored in the failure model storage unit 15, and corresponding residual vectors for a plurality of failure types (type A, type B, type C, etc.). The component (pattern) of is stored. The failure type includes, for example, “heat exchanger clogged”, “refrigerant leak”, and a failure part in the electric heat pump. Further, for these failure models, residual vectors obtained from sensor values actually obtained in the case of failures experienced in the past are used.

  As described above, the “measured value pattern” shown on the right side of the figure indicates the residual vector component obtained this time, and this pattern is collated with the pattern of each failure model. In the example shown in the figure, since the pattern of the actual measurement values is similar to the failure model of type B, in this case, the type of failure is determined to be type B. In the determination of similarity, for example, the inner product of the unit vector of the residual vector related to the actual measurement value and the unit vector of the residual vector related to each failure model is calculated, and if the result is close to 1, it is determined to be similar . That is, as described above, if the directions of the residual vectors are almost the same, it is determined that the failure types are the same.

  As shown in FIG. 6B, the failure type determination unit 15 collates the difference between the predicted value and the actual measurement value (residual vector component) and the failure model stored in the failure model storage unit 15. The type of the detected failure is determined (step S32 in FIG. 5). Specifically, from the failure models stored for each failure type, a failure model in which the difference between the predicted value obtained this time and the actual measurement value and the pattern (trend) match or is very similar is selected, The failure type of the selected failure model is determined as the failure type detected this time. This means that if the failure type is the same, the difference pattern (trend) between the predicted value and the actual measurement value is almost the same, and if the failure type is different, the pattern (trend) is different. This is because it is known experimentally.

  FIG. 6A illustrates three residual vectors. In the figure, x1, x2, and x3 are three actually measured values of sensor values, and e1, e2, and e3 are residual vectors corresponding to them. As described above, the pattern (trend) of the difference between the predicted value and the actually measured value, in this case, the pattern of each component of the residual vector, that is, the direction of the residual vector is substantially the same for each failure type. Become. Therefore, in FIG. 6A, since the directions of the residual vectors e1 and e3 are substantially the same, it can be determined that the failure types relating to the actual measurement values x1 and x3 are the same, and the actual measurement values having different residual vector directions. It can be determined that the failure type related to x2 is different from these.

  When the determination of the failure type is completed, the failure type determination unit 15 notifies the failure information to the user of the remote monitoring system 1 (step S33 in FIG. 5). The content of the failure information is, for example, the power equipment 2 that has detected the failure and its failure type. Further, the notification means includes display on a monitor and printing on paper by a printer. Then, while notifying the failure, the prediction model construction unit 12 constructs the specific failure prediction model for the failure by the above-described method and stores it in the prediction model storage unit 13 (S34).

  Then, the remote monitoring system 1 ends the monitoring if there is a reason to end the monitoring process such as when the power equipment 2 should be stopped due to the detected failure (Yes in step S24 in FIG. 5). And the inspection, repair, etc. of the power equipment 2 in which the failure is detected are performed as necessary. On the other hand, if there is no reason to end the monitoring process (No in step S24 in FIG. 5), the monitoring process is continued, and the process from step S21 in FIG. 5 described above is repeatedly executed.

  As described above, in the remote monitoring system 1, the failure type can be identified based on the tendency of the difference between the predicted value of the sensor value and the actual measurement value, which can accurately identify the failure. Accurate fault type identification and fault diagnosis are possible. In addition, according to the remote monitoring system 1 according to the present embodiment, failure detection and failure type focusing on the correlation between the plurality of sensors 21 by multivariate analysis on values detected by the plurality of sensors 21. Identification is made, and early detection of a failure and accurate failure diagnosis are possible. In addition, since a plurality of monitoring objects (power equipment 2) scattered at different locations can be monitored at one place, labor saving in equipment monitoring work and maintenance work can be achieved. Further, since the failure diagnosis is automatically performed by the system, no skill is required as in the case where a person makes a judgment, and there is no individual difference.

  In addition to the basic prediction model constructed using all sensor values (all data variables) for the predicted value to be compared with the actual measurement value, a predetermined sensor is used to increase the detection sensitivity for a specific failure. By constructing the prediction model for specific failure constructed by the combination of the deleted data variables for each type of failure, it becomes possible to detect the failure early and accurately.

  In the present embodiment, the power facility 2 is the monitoring target of the remote monitoring system 1, but the monitoring target is not limited to the power facility, and any other device may be used as long as it can acquire a sensor value.

  The protection scope of the present invention is not limited to the above-described embodiment, but covers the invention described in the claims and equivalents thereof.

  In principal component analysis, methods for reducing variables while maintaining the analysis capability are being studied. This is used in cases where it is necessary to reduce the number of data points per time while maintaining the appearance of the principal component model in terms of cost and practical use. Reduce the variable.

  As one of variable selection methods for principal component analysis, Jolliffe's variable selection method based on principal component contribution criteria is known. This compares the loading vector size of each variable in each principal component, deletes the least important variable, drops it to the appropriate number of variables (variable reduction method), or importance It is a method of acquiring from a variable with a high (variable increasing method). For Jolliffe's method, see the following references (1) and (2).

(1) Jolliffe, IT (1972) Discarding variables in a principal component analysis. I. Artifical data. Appl. Atatist., 21, 160-173.
(2) Jolliffe, IT (1973) Discarding variables in a principal component anal ysis. II. Real data. Appl. Atatist., 22, 21-31.
On the other hand, in the above-described embodiment of the present invention, a method of reducing variables based on the SN ratio of the square prediction error SPE is adopted. In the variable selection by Jolliffe's method, the variable is not reduced based on the SN ratio of the square prediction error that is directly related to the sensitivity of abnormality detection. Therefore, the detection sensitivity is not necessarily improved by the variable selection.

  Hereinafter, the variable selection method of Jolliffe and the variable selection method according to the present invention will be compared by examples. In the examples, Tennessee-Eastman plant simulation, which is often used as a benchmark for process control, was used (for details, see “Benchmark process for industrial-scale plant-wide process control” JJ Downs and EF Vogel A Plant-Wide Industrial Process Control Problem, Computer & Chemical Engineering, Vol. 17, No. 3, pp.245-255, 1993). In this plant simulation, a disturbance occurs when a product is manufactured in a chemical plant, and it is used as a process control benchmark for avoiding an abnormality caused by the disturbance. In this analysis, the disturbance is regarded as an abnormal phenomenon. The detection sensitivity was examined. The disturbance (IDV02) shown in this example is a disturbance of the inert gas concentration and is an analysis of abnormal fluctuations that occur in the entire plant by principal component analysis.

  Of the measurement variables and operation variables in Tennessee-Eastman plant simulation, 33 data were used to detect process abnormalities due to disturbances by principal component analysis.

An image of the data set to be analyzed is shown in FIG. Among these, XMEAS01 to XMEAS22 are operation variables such as flow rate and temperature pressure, and XMV01 to XMV11 are operation variables for operating the control amount.
(1) Variable selection based on S / N ratio criterion of square prediction error SPE FIG. 8 is a graph showing the principal component analysis using 32 variables and normal data 1 to 960 by sequentially deleting each of the 33 variables. Modeling is performed and the state at each time is represented. Here, the horizontal axis represents the time axis, and the vertical axis represents the SN ratio (hereinafter referred to as the SPE-SN ratio) obtained by normalizing the SPE with the 95% confidence limit value at the normal time. It can be seen that the SPE signal-to-noise ratio increases in the vicinity of 980, where the disturbance occurred, until 980 during normal operation.

  FIG. 9 is an enlarged view of the change in the SPE-SN ratio after the occurrence of the disturbance shown in FIG. 8, and compares the change in the SPE-SN ratio of the 33 analysis results analyzed by removing one of the 33 variables. FIG. Among these, the analysis result (the arrow line shown in FIG. 9) that maximizes the SPE-SN ratio was analyzed excluding the variable XMEAS01. This is because the XMEAS01 variable represents a change in the state of the plant system due to disturbance. It indicates that it contains the most non-contributing noise.

In this way, a variable that maximizes the SPE-SN ratio is obtained, an analysis data set from which the variable is deleted is newly created, the same analysis is performed, and the number of variables is reduced. 10 and 11 show changes in the SPE-SN ratio when variables are dropped and variables to be deleted. The variable selection process ends when the SPE-SN ratio decreases. In this example, when the variable is 24, the SPE-SN ratio showed the maximum, so modeling of 24 variables is optimal for the failure of IDV02.
(2) Variable selection by Jolliffe's variable reduction method Fig. 12 shows the order of the absolute value of the coefficient of each loading vector (corresponding to the principal component) p when SPE analysis is performed using all 33 variables using IDV02 data. The ranking is shown, and the ranking of the contribution ratio of each sensor to each loading vector is also shown. That is, p1 represents the loading vector of the first principal component having the largest absolute value, and p33 represents the loading vector of the thirty-third principal component having the smallest absolute value. At this time, in the variable selection by Jolliffe's variable reduction method, paying attention to the contribution rate of each sensor of the loading vector (that is, the unimportant principal component) from the smaller absolute value, the contribution rate to the unimportant principal component is large ( Delete from the variable with the highest ranking. In FIG. 12, the shaded variables are deleted. Here, the parameters to be deleted are selected until 24 variables are the same as the variable selection criteria based on the SPE SN ratio.

  FIG. 13 is a diagram showing a comparison of detection sensitivity between the variable selection criterion of the present invention based on the S / N ratio of the square prediction error SPE and the variable selection criterion using the Jolliffe method. The detection sensitivity is about 524 for the SPE signal-to-noise ratio, and about 265 for the Jolliffe method. This indicates that the method based on the S / N ratio of the square prediction error SPE reduces the influence of noise, and a model suitable for abnormalities is created. FIG. 14 is an enlarged view of the first to twentieth points of FIG. 13. This figure also shows that the rise of the SPE-SN ratio is early and that an abnormality can be found early.

It is a figure which shows the structural example of the remote monitoring system in embodiment of this invention. It is a figure which shows the data group plotted in three-dimensional space. It is a figure which shows the score vector t and the loading vector p. It is a flowchart of the selection process of a data variable. It is the flowchart which illustrated the processing content in embodiment of this invention. It is a figure for demonstrating failure classification determination processing. It is a figure which shows the image of the data set of analysis object. It is the figure which deleted the variable one by one among 33 variables in order, modeled by principal component analysis using the data at the time of 32 variables, and represented the state of each time. It is the figure which compared the change of SPE-SN ratio of the 33 analysis results analyzed except 1 variable among 33 variables. It is a figure which shows the change of SPE-SN ratio when a variable is deleted. It is a table | surface which shows the variable to delete when deleting a variable. It is the figure which showed the ranking in large order about the absolute value of each loading vector p when SPE analysis is performed using all 33 variables. It is a figure which shows the comparison of the detection sensitivity of the variable selection reference | standard of this invention based on SN ratio of the square prediction error SPE, and the variable selection reference | standard using the Jolliffe method. FIG. 14 is an enlarged view of points 1 to 20 in FIG. 13.

Explanation of symbols

  1: remote monitoring system, 2: power equipment, 3: communication line, 11: sensor information acquisition unit, 12: prediction model construction unit, 13: prediction model storage unit, 14: failure detection unit, 15: failure type determination unit, 16: Failure model storage unit, 21: Sensor

Claims (8)

In a remote monitoring system for remotely monitoring equipment having a plurality of sensors,
A sensor information acquisition unit that acquires sensor values detected by the plurality of sensors via a predetermined communication line;
Obtaining a first correlation between the sensor values of all of the plurality of sensors acquired when the facility is operating normally, the basic correlation is used to detect the failure of the facility. A second correlation between sensor values of some sensors in the plurality of sensors is obtained, and the second correlation is determined for the specific failure of the facility. A prediction model building unit that builds a prediction model for specific faults with higher detection sensitivity;
Based on the difference between the sensor value acquired by the sensor information acquisition unit during the monitoring period of the facility and the predicted sensor value obtained from each of the basic prediction model and the prediction model for specific failure, the presence or absence of failure of the facility A failure detection unit that detects
When the failure detection unit detects a failure of the facility based on the basic prediction model, the prediction model construction unit starts from a part of the plurality of sensors such that the detection sensitivity for the failure is maximized. The remote monitoring system is characterized in that the specific failure prediction model for the failure is constructed by determining a combination of sensors and obtaining the second correlation between sensor values of the partial sensors.   The remote monitoring system according to claim 1, wherein the first correlation and the second correlation obtained by the prediction model construction unit are obtained by principal component analysis.   The said prediction model construction part determines the combination of the said sensor on the basis of the SN ratio of the difference of the sensor value which the said sensor information acquisition part acquires, and the said prediction sensor value, The Claim 1 or 2 characterized by the above-mentioned. The remote monitoring system described.   When the failure detection unit detects a failure of the facility based on the basic prediction model instead of the specific failure prediction model, the sensor value acquired by the sensor information acquisition unit and the prediction sensor value The remote monitoring system according to claim 1, further comprising a failure type determination unit that determines a type of failure based on a difference tendency. In a remote monitoring program for causing a remote monitoring system to execute a process for remotely monitoring equipment having a plurality of sensors,
A first process of acquiring sensor values detected by the plurality of sensors via a predetermined communication line;
First correlation between sensor values of all of the plurality of sensors acquired when the facility is operating normally is obtained, and the correlation is constructed as a basic prediction model for detecting a failure of the facility Then, a second correlation between sensor values of some of the plurality of sensors is obtained, and the second correlation has higher detection sensitivity than the basic prediction model for a specific failure of the equipment. A second process to be constructed as a prediction model for a specific failure;
During the monitoring period of the equipment, based on the difference between the sensor value acquired by the first process and the predicted sensor value obtained from each of the basic prediction model and the prediction model for specific failure, the failure of the equipment Causing the remote monitoring system to execute a third process for detecting presence or absence;
The process for constructing the specific failure prediction model in the second process is such that, in the third process, when a failure of the facility is detected based on the basic prediction model, the detection sensitivity for the failure is maximized. Determining a combination of sensors composed of a part of the plurality of sensors and obtaining the second correlation between sensor values of the part of the sensors, thereby predicting the specific fault for the fault A remote monitoring program characterized by being a model building process.   6. The remote monitoring system according to claim 5, wherein in the second processing, the first correlation and the second correlation are obtained by principal component analysis.   The combination of the sensors is determined based on the SN ratio of the difference between the sensor value acquired by the sensor information acquisition unit and the predicted sensor value in the second process. The remote monitoring program described in 1.   In the third process, when a failure of the facility is detected based on the basic prediction model instead of the specific failure prediction model, the sensor value and the prediction acquired by the first process are detected. 8. The remote monitoring program according to claim 5, further causing the remote system to execute a fourth process for determining a failure type based on a tendency of a difference from a sensor value. 9.

Images