Specification

Title: DETERMINATION APPARATUS, DETERMINATION METHOD
Technical field

[0001]
　The present invention relates to a determination device, a determination method, and a determination program.
BACKGROUND ART

[0002]
　Techniques for determining abnormality have been developed. For example, a technology for determining an abnormality by using a degree of divergence from a model as a criterion (see, for example, Patent Document 1) and a technique for determining an abnormality by using the integral of a difference with a model as a criterion (for example, 2) have been disclosed.
Prior Art Document

Patent literature

[0003]
Patent Document 1: International Publication No. 2010/082322
Patent Document 2: Japanese Examined Patent Publication No. 4-25565
Summary of the invention

Problem to be Solved by Invention

[0004]
　However, with the above technique, it is difficult to determine a sign of abnormality.
[0005]
　SUMMARY OF THE INVENTION The present invention has been made in view of the above problems, and it is an object of the present invention to provide a determination device, a determination method, and a determination program capable of determining a sign of abnormality.
Means for solving the problem

[0006]
　In one aspect, the determination device includes: a model creation unit that creates a reference model of a sensor detection value; a model generation unit that creates a reference model of a sensor detection value from a predetermined point in time; A judging section for judging whether or not the time is shorter than the time, and an output section for outputting a signal relating to the abnormality when judged to be short.
Effect of the Invention

[0007]
　It is possible to determine a sign of abnormality.
Brief Description of the Drawings

[0008]
[Fig. 1] (a) and (b) are examples of a temperature measurement method using an optical fiber. FIG. 2 is a diagram illustrating an example of object variables and explanatory variable groups. 3 (a) is an image diagram exemplifying a difference between an estimated value and an actually measured value in the case of using the fuel type A as a temperature difference, (b) is an image diagram in which an estimated value in the case of using the fuel type B and an actual value Value as a temperature difference. FIG. 4 is a diagram illustrating a relationship between a threshold value and an abnormality determination. FIG. 5 (a) is a schematic diagram of a determination device according to a first embodiment, and (b) is a block diagram for explaining a hardware configuration of a determination unit. [Fig. 6] An explanatory variable group is exemplified. FIG. 7 is a flowchart illustrating setting of estimation formulas of objective variables and setting of divergence degrees and threshold values ​​of estimated effective time. FIG. 8 is an example of a flowchart executed when abnormality determination is performed. FIG. 9 (a) is a schematic diagram of a determination device according to a second embodiment, and FIG. 9 (b) is an example of a temperature sensor. FIG. 10 is an example of a flowchart executed when abnormality determination is performed. [Fig. 11] (a) is the measured value of the objective variable, and (b) is the measured value of the explanatory variable. [Fig. 12] (a) is the instantaneous value of the estimation error, (b) is the integrated value of the estimation error, and (c) is the estimated effective time. FIG. 13 is an example of a temperature sensor. [Fig. 14] (a) is the Mahalanobis distance and (b) is the estimated effective time. FIG. 15 is a diagram illustrating a determination system according to a second modification.

[Fig. 16] Figs. 16 (a) and (b) are diagrams illustrating a sensor unit and a measuring instrument. [FIG. 17] (a) to (c) are diagrams illustrating sensor units. FIG. 18 is a diagram illustrating a flow chart representing a dimensionless procedure. [Fig. 19] Figs. 19 (a) and (b) are diagrams illustrating abnormality signs. FIG. 20 is an example of a flowchart executed when abnormality determination is performed. FIG. 21 is a flowchart illustrating a comparative example. FIG. 22 is a view exemplifying the results of a comparative example. FIG. 23 is a view exemplifying the results of the method described in FIG. 20. FIG. 24 is a diagram illustrating the results of a comparative example. FIG. 25 is a view exemplifying the results of the method described in FIG. 20. FIG. 26 is an example of a flowchart executed when abnormality determination is performed. FIG. 27 is a diagram illustrating a result of abnormality determination. FIG. 28 is a diagram exemplifying normalized values ​​of sensing data.

MODE FOR CARRYING OUT THE INVENTION

[0009]
　First, the outline of abnormality determination will be described.
[0010]
　In chemical plants, oil refineries, thermal power plants and the like, gas leakage or the like sometimes occurs due to corrosion or the like. Therefore, it is desired to be able to judge abnormality at an early stage. For example, it is conceivable that a temperature sensor is disposed on a piping system in which branch pipes are welded to a main pipe, and the leakage of gas or liquid is detected early as a temperature change. Alternatively, by monitoring the pipe temperature of the cooling water, it is possible to detect an abnormality in the temperature early before a fire occurs even if an accidental cooling failure or the like occurs. Note that "detection of symptoms" of anomaly as described below means "when monitoring a certain object, using some means to visualize a part of its operation, if the visualized state is different from the normal state It is assumed that a sign of abnormality is detected when it is judged that there is nothing wrong ".
[0011]
　For temperature abnormality detection, for example, a temperature measurement method using an optical fiber in which Raman scattering light is measured to obtain temperature information can be mentioned. For example, as shown in FIG. 1 (a), by laying optical fibers in branch pipes, it is possible to detect a leak early as a temperature change. As illustrated in FIG. 1 (b), by laying optical fiber in the cooling water pipe of the boiler, the cooling water temperature can be monitored. As a result, even if a cooling failure or the like occurs, an early abnormality in temperature can be detected before a fire occurs.
[0012]
　However, rather than "detecting an abnormality that minimizes accidents", it is more preferable for facility management to "prevent accidents beforehand by predicting abnormal symptoms". Therefore, regression analysis using parameters such as sensor data for various operation control as explanatory variables, posterior attachment of sensor information as a target variable, correlation analysis between posterior attachment information, A method of statistically judging whether or not a person is in need has been adopted. This is because network technology has improved to such an extent that parameters such as sensor data for operation control can be managed at once, and various regression analysis methods and correlation analysis methods can be applied in real time due to an increase in computer power It is due to that.
[0013]
　However, the above method focuses on how to minimize the "estimated error" (the difference between the estimated value and the measured value). Therefore, no practical discussion has been made as to the extent to which "estimation error" is judged to be abnormal. For example, in thermal power plants, oil types, coal types and the like are not pattern classified, and the characteristics change every time depending on the country of production and the amount of mixing thereof. Therefore, it is necessary to perform an initialization operation for estimation every time the characteristic changes. However, since the "estimation error" is influenced by the accuracy, it is easy to adopt only a simple setting method such as judging abnormality by how much the "estimation error" has changed in a fixed time or the like, for example. As a result, an ambiguous threshold value is set and it tends to be an impractical system.
[0014]
　FIG. 2 is a diagram illustrating an example of object variables and explanatory variables. In the example of FIG. 2, the objective variables 1 to 3 are the temperatures of respective portions of the outer wall metal of the boiler. The explanatory variable group is the output value of the sensor correlated with the objective variable. The measured values ​​of the object variables 1 to 3 can be acquired using a temperature sensor or the like installed on the outer wall metal. As illustrated in FIG. 2, the estimation formulas of objective variables 1 to 3 are obtained by setting coefficients and constants of each explanatory variable. Regarding these coefficients and constants, we can calculate the least squares regression (Ordinary Least Mean Square), the Principal Component Regression (Principal Components Regression), the partial least squares regression (Partial Least Squares) or the like. Every time explanatory variable information is collected, an explanatory variable is input to the estimation equation and an estimated value is derived.
[0015]
　It is possible to judge whether the system is normal or abnormal by comparing the actual value of the objective variable acquired at the same time with the estimated value of the estimation equation. In order to set the coefficients and constants of the estimation equation, the latest past data in a certain period is necessary. This immediate period is referred to as "modeling period". On the other hand, a period for actually comparing the estimated value with the measured value is referred to as a "scoring period". In the scoring period, if the "estimation error" exceeds a certain value, it is considered that a situation deviating from the estimation is occurring.
[0016]
　FIG. 3 (a) is an image diagram exemplifying the difference between the estimated value and the measured value of the objective variable 1 and the objective variable 2 in the scoring period in the case of using the fuel type A as a temperature difference. FIG. 3 (b) is an image diagram exemplifying the difference between the estimated value and the measured value of the objective variable 1 and the objective variable 2 in the scoring period using the fuel type B as a temperature difference.
[0017]
　In the case of fuel type A, it is assumed that the threshold value is ± 0.3 ° C. at 3σ. In the case of fuel type B, it is assumed that the threshold value is ± 0.9 ° C. at 3σ. However, it is difficult to make an objective judgment as to whether or not the condition exceeding 3σ is immediately abnormal. This is because there is a possibility of exceeding this value with a probability of 0.3%. Conversely, when 4 σ is set, there is a possibility of missing the abnormality.
[0018]
　FIG. 4 is a diagram illustrating the relationship between the threshold value and the abnormality determination. As exemplified in FIG. 4, when the threshold value is set to the relatively large threshold value 1, a delay occurs with respect to the case where the problem occurs truly, and the countermeasure is delayed. When the threshold value is set to the threshold value 2 which is smaller than the threshold value 1, it is judged to be abnormal in a situation different from the situation where the problem occurs truly, so that the function of the premonitory detection can not be fulfilled also. In other words, it is difficult to detect useful symptoms if it is impossible to make a sufficiently accurate estimation and to set a useful threshold for each objective variable and to make an abnormality judgment with that threshold value. It is not practical for business continuity to examine appropriate threshold for each update of content affecting the threshold such as periodic inspection and change of ratio of oil type.
[0019]
　In the following embodiments, a determination device, a determination method, and a determination program capable of determining a sign of abnormality will be described.
[0020]
　(First Embodiment)
　FIG. 5 (a) is a block diagram of a determination device 100 according to a first embodiment. In the present embodiment, the determination device 100 is installed, for example, in a thermal power generation facility utilizing the combustion cycle of coal. The determination device 100 includes an explanatory variable acquisition unit 10, a plurality of temperature sensors 20 a to 20 c, a determination unit 30, and the like. The determination unit 30 includes a model creation unit 31, a threshold setting unit 32, an abnormality determination unit 33, and an output unit 34.
[0021]
　FIG. 5 (b) is a block diagram for explaining the hardware configuration of the determination unit 30. As illustrated in FIG. 5 (b), the determination unit 30 includes a CPU 101, a RAM 102, a storage device 103, an interface 104, and the like. Each of these devices is connected by a bus or the like. A CPU (Central Processing Unit) 101 is a central processing unit. The CPU 101 includes one or more cores. A RAM (Random Access Memory) 102 is a volatile memory that temporarily stores programs executed by the CPU 101, data processed by the CPU 101, and the like. The storage device 103 is a nonvolatile storage device. As the storage device 103, for example, a solid state drive (SSD) such as ROM (Read Only Memory), flash memory or the like, a hard disk driven by a hard disk drive, or the like can be used. By executing the determination program stored in the storage device 103 by the CPU 101, the model creation unit 31, the threshold setting unit 32, the abnormality determination unit 33, and the output unit 34 are realized in the determination unit 30. The model creation unit 31, the threshold value setting unit 32, the abnormality determination unit 33, and the output unit 34 may be hardware such as dedicated circuits.
[0022]
　The explanatory variable acquiring unit 10 acquires each explanatory variable. FIG. 6 exemplifies explanatory variables. As exemplified in FIG. 6, the explanatory variables are electric energy, coal feed amount, internal temperature 1, internal temperature 2, air flow rate, pressure 1, pressure 2, pressure 3, ventilation port 1, ventilation port 2, operating speed , Operation rate, operation frequency, and the like. Each explanatory variable has a correlation with objective variables 1 to 3 (detected temperatures of the temperature sensors 20a to 20c). In addition, it is preferable that each explanatory variable can be regarded as independent of each other (low multiple co-linearity). The amount of electric power is generated electric power obtained by thermal power generation. The amount of coal supplied is the amount of coal supplied to the furnace. The internal temperatures 1 and 2 are, for example, temperatures at any point inside the furnace. The air flow rate is the flow rate of the air supplied to the furnace. Pressures 1 to 3 are, for example, pressures in a pipe connected to a furnace. The vents 1 and 2 are the temperature of the ventilation openings and the like. The operating speed, operating rate and operating frequency are the operating speed of the furnace, operating rate, operating frequency, and so on. These explanatory variables are the output values ​​of each sensor.
[0023]
　The plurality of temperature sensors 20a to 20c are installed, for example, at different places on the wall surface outside the furnace. In the present embodiment, temperature sensors are installed at three locations. As a method of measuring the temperature of the temperature sensors 20a to 20c, for example, a method using Raman scattered light in the optical fiber can be used. For example, in order to measure an accurate temperature, an optical fiber having a length of about 2 m is wound around a small region that can be regarded as substantially the same temperature. By doing so, each wound portion functions as one temperature sensor. In this embodiment, wall surface temperatures 1 to 3 detected by the temperature sensors 20a to 20c are used as measured values ​​of the object variables 1 to 3.
[0024]
　The model creation unit 31 creates estimation equations of the objective variables 1 to 3 using the explanatory variables acquired by the explanatory variable acquiring unit 10 and the detection values ​​of the temperature sensors 20 a to 20 c. This estimation formula is a reference model of object variables 1 to 3. The estimation equation is obtained by setting coefficients and constants of each explanatory variable similarly to the estimation formula illustrated in FIG. 2. These coefficients and constants can be set by regression analysis such as least squares regression, principal component regression, partial least squares regression, etc. based on past explanatory variables and measured values ​​of the temperature sensors 20a to 20c .
[0025]
　The threshold value setting unit 32 sets a threshold value for the deviation degree and the estimated effective time of the measured value of the objective variable with respect to the reference model. The degree of divergence of the measured value of the objective variable with respect to the reference model can be calculated by, for example, estimation error = (measured value of the objective variable) - (estimated value of the objective variable), integrated value for each data update of the estimation error, The ratio of the estimate of the objective variable, and so on. In the present embodiment, an estimation error and an integrated value of the estimation error are used as the deviation degrees. The estimated value of the objective variable is a numerical value obtained by inputting an explanatory variable in the estimation formula. The estimation error relating to objective variable 1 is (measured value of temperature sensor 20 a) - (estimated value of objective variable 1). The estimation error related to objective variable 2 is (measured value of temperature sensor 20 b) - (estimated value of objective variable 2). The estimation error related to objective variable 3 is (measured value of temperature sensor 20 c) - (estimated value of objective variable 3). The estimated effective time is the time from the start of the measurement of the estimation error in the scoring period using the reference model until the above-mentioned deviation degree exceeds the threshold value. The abnormality determination unit 33 determines whether or not the estimated effective time is less than the threshold value, thereby determining abnormality. The output unit 34 outputs a signal relating to the abnormality when the abnormality determination unit 33 determines the abnormality.
[0026]
　FIG. 7 is a flowchart exemplifying the setting of the estimation formula of the objective variable and the setting of the deviation degree and the threshold value of the estimated effective time. As exemplified in FIG. 7, the threshold setting unit 32 first detects an initial condition update flag (step S 1). The initial condition update flag is a flag that is a trigger for updating the estimation formula of the objective variable and the deviation degree and the threshold value of the estimated effective time. Next, the threshold value setting unit 32 sets appropriate allowable values ​​1 and 2 (step S2). Tolerance 1 is the threshold for the estimation error. Tolerance value 2 is a threshold value for the integrated value of the estimation error.
[0027]
　Next, the model creating unit 31 collects the data set of the modeling period (step S3). This data set includes explanatory variables at predetermined time intervals in the modeling period and detected values ​​(actually measured values) of the temperature sensors 20a to 20c. Next, the model generating unit 31 determines coefficients and constants of the estimation equations of the objective variables 1 to 3 using the data set collected in step S 3 (step S 4). By executing step S4, estimation formulas for object variables 1 to 3 are set.
[0028]
　Next, the threshold value setting unit 32 starts measuring the estimation error (scoring period), for example, the scoring period and the estimation error of the first 60 times (30 minutes in the case of the measurement of the 30-second period) of the modeling period And the standard deviation (step S5). Next, the threshold value setting unit 32 resets the average value + 1 σ value to 3 σ value as a tolerance value 1. In addition, the threshold value setting unit 32 resets the allowable value 2 so that the approximate estimated effective time is approximately 60 to 240 times the measurement cycle (30 minutes to 2 hours in measurement at 30-second intervals) (step S 6). Step S 6 corresponds to relaxation since it means that the tolerance values ​​1 and 2 are small if re-estimation occurs before the scoring period reaches 30 minutes in the measurement at 30-second intervals.
[0029]
　Next, after re-setting in step S 6, the threshold setting unit 32 sets a provisional threshold value of the estimated effective time and starts provisional measurement of the estimation error (step S 7). Next, the model creation unit 31 repeats re-creation of the reference model when the estimation error exceeds the tolerance value 1 or the integration value of the estimation error exceeds the tolerance value 2. The threshold value setting unit 32 determines whether or not data of such an extent that the re-creation is performed 30 times is accumulated (step S 8). If "No" is determined in step S 8, step S 8 is executed again. If "Yes" is determined in the step S 8, the threshold value setting unit 32 obtains the average value and the standard deviation of the estimated effective time obtained in the step S 8, and re-sets the threshold value of the estimated effective time using the 3σ value (Step S 9). In steps S 6 to S 9, if the provisional effective time also falls below the threshold value, the output unit 34 outputs a signal related to abnormality.
[0030]
　With respect to the tolerance values ​​1 and 2 and the threshold value of the estimated effective time, once the data is accumulated, it can be set as many times as it can go back in the past with a program incorporated in advance. Therefore, once resetting the allowable values ​​1 and 2, it is unnecessary to re-set the threshold value of the estimated effective time by again accumulating data from there and thereafter accumulating the data again and checking. Therefore, it is possible to construct a system that only needs to input information that "change has been made" to the system after periodic inspection, change of the ratio of oil type, or the like.
[0031]
　FIG. 8 is an example of a flowchart executed when the abnormality determination unit 33 performs abnormality determination after the threshold values ​​of the tolerance values ​​1, 2 and the estimated effective time are set by the threshold value setting unit 32. The abnormality determination unit 33 collects the data set after the tolerance values ​​1 and 2 and the threshold value of the estimated effective time are set by the threshold setting unit 32 (step S 11). Next, the abnormality determination unit 33 determines whether the estimation error exceeds the tolerance value 1 or whether the integrated value of the estimation error exceeds the tolerance value 2 (step S 12). If it is determined in step S 12 that none of them exceeded, step S 12 is executed again. If it is determined in step S 12 that any one has exceeded, the abnormality determination unit 33 determines whether the estimated effective time is shorter than a predetermined time (for example, 10 minutes in the measurement at the interval of 30 seconds) (step S 13) . If "Yes" is determined in the step S13, the output unit 34 outputs a signal relating to the abnormality (step S14). If "No" is determined in the step S13, the abnormality determination unit 33 performs the estimation again using the past data (for example, one hour in the measurement at the interval of 30 seconds) from that time, and sets the coefficients and constants of the estimation equation And updates it (step S15). Thereafter, the process is repeated from step S 11.
[0032]
　In the example of FIG. 8, the tolerance value 1 is the threshold value of the estimation error, and the tolerance value 2 is the integration value for each update of the estimation error data. If a certain estimate is made, the average value of the estimation error becomes almost 0 if it averages over a long period of time, but as the situation different from the estimation occurs, either positive or negative value starts to increase. The threshold value set for this change is the allowable value 2. Even if the average value of the estimation error is close to zero, in the case of an event that began to occur suddenly, since the estimation error increases, it can be regarded as an abnormality. In this case the tolerance value is 1.
[0033]
　Since tolerance values ​​1 and 2 contain ambiguity, the estimated error to be measured may exceed the tolerance value 1 or the integrated value of the estimation error may exceed the tolerance value 2. However, the model creation unit 31 performs estimation again at that time point, updates coefficients and constants of the estimation equation, and resumes counting with that time as the estimation start time. The method shown in FIG. 8 is to set the period until the allowable value 1 or the allowable value 2 is exceeded next time as the "estimated effective time" and set the threshold for the "estimated effective time". In the example of FIG. 8, two allowable values ​​are set, but one may be acceptable, or more numbers may be set. For example, it is also possible to set the "estimated effective time" at the stage where one or more tolerance values ​​are exceeded and then to estimate once again when one or more tolerance values ​​are exceeded I do not care.
[0034]
　In the present embodiment, frequent occurrence of a phenomenon that the threshold value is exceeded is allowed. Instead, importance is placed on the frequency at which the frequency occurs. The phenomenon gradually changes even in power plants, plants, and other supposed application destinations. Therefore, the estimation error tends to decrease after creation of the reference model. In that state, if the estimation error is large, it means that an event not occurring in the formulation is occurring. Therefore, the fact that the "estimated effective time" is shorter means that it can be judged that an abnormal situation is occurring even considering ambiguity of threshold setting and estimation. That is, according to the present embodiment, it is possible to determine a sign of abnormality.
[0035]
　Further, according to the present embodiment, the threshold value of the deviation degree of the temperature sensors 20a to 20c with respect to the reference model is determined on the basis of the deviation degree for a certain period after creation of the reference model. In this case, the setting accuracy of the deviation degree threshold is improved. Further, according to the present embodiment, the threshold value of the estimated effective time is determined based on the variation in time until the deviation degree exceeds the threshold value. In this case, the accuracy of setting the threshold value of the estimated effective time is improved.
[0036]
　Second Embodiment In
　the first embodiment, the reference model is created using the sensor detection value and the detection values ​​of a plurality of other sensors correlated with the sensor detection value, but the present invention is not limited to this. In the second embodiment, the reference model of the sensor detection value is created using the correlation between the detection values ​​of the plurality of sensors.
[0037]
　First, consider the correlation of temperature transition data detected by multiple temperature sensors and consider a sign detection method for finding signs of abnormality at an early stage. In this case, a method of finding the Mahalanobis squared distance to be calculated from the average and variance covariance matrix of the target data group, an MSD method of estimating the statistic corresponding to the Mahalanobis square distance by estimating the center and spread of data robustly Can be used.
[0038]
These are called "outlier detection". Procedures for detecting a symptom of finding an abnormality at an early stage using a procedure for finding the Mahalanobis squared distance are carried out by using n temperature data T1 (t), T2 (t), T3 (t), ... at time t, The following (1) to (3) are shown as Tn (t). (1) Set the modeling period (a period of time over which the data is accumulated older than the current time) and compare the average of the temperature data of the n temperature sensors of each period and the unbiased variance covariance matrix of the temperature data group Find a matrix. (2) For the temperature data T1 (m) to Tn (Tm) at each time Tm (m = 0, 1, 2, ...) within the modeling period, Temperature data group of temperature sensor ". A threshold that becomes abnormal is set from the standard deviation (3σ etc.) of those values. (3) Every time the data set T1 (T) to Tn (T) at the new time T is obtained, the Mahalanobis square distance between this "temperature data group of the n temperature sensors in the modeling period" is obtained, It is determined whether or not it is equal to or less than the threshold value.

[0039]
　Also in the case of using the MSD method, the procedure of setting a modeling period, determining a threshold value from that period, and sequentially comparing the new data set and the threshold value is the same. In other words, also in predictive detection by "outlier detection" using correlation between measurement data, accuracy of predictive detection is determined by setting of threshold. In other words, it is difficult to detect useful symptoms if it is impossible to make a sufficiently high-precision estimation and to set a useful threshold and to make an abnormality judgment with that threshold. Therefore, also in the second embodiment, a determination device, a determination method, and a determination program capable of determining a sign of abnormality will be described.
[0040]
　FIG. 9 (a) is a schematic diagram of the determination device 100 a according to the second embodiment. The determination device 100a is different from the determination device 100 of the first embodiment in that the explanatory variable acquisition unit 10 is not provided and a temperature sensor 20 is provided instead of the temperature sensors 20a to 20c. The temperature sensor 20 detects temperatures at a plurality of places where the temperature values ​​are correlated with each other. For example, as illustrated in FIG. 9 (b), the temperature sensor 20 detects the temperature of each place based on the result obtained by the back scattered light of different length positions of the same optical fiber. In the example of FIG. 9 (b), each winding portion functions as an individual temperature sensor. The configuration of the determination unit 30 is the same as that of the first embodiment.
[0041]
　In the present embodiment, the model creation unit 31 creates a reference model by finding the Mahalanobis square distance to be calculated from the average and variance covariance matrix of the detection values ​​detected by the temperature sensor 20. In addition, the model generating unit 31 estimates the center and spread of the detection value in a robust manner, and creates a reference model by using the MSD method or the like for obtaining a statistical amount corresponding to the Mahalanobis square distance. The reference model here is the center of the variation that reflects the magnitude of the correlation between a plurality of detection values ​​(the attention focusing on the two sensors, the orientation).
[0042]
　The threshold value setting unit 32 sets a threshold value for the deviation degree and the estimated effective time of the actually measured value of the sensor detection value with respect to the reference model. The deviation degree of the actual measurement value of the objective variable with respect to the reference model is the Mahalanobis square distance, the integrated value every data update of the Mahalanobis squared distance, and the like. In the present embodiment, the integrated value of the Mahalanobis square distance and the Mahalanobis square distance is used as the deviation degree. The estimated effective time is the time from the start of the measurement of the Mahalanobis square distance in the scoring period after creation of the reference model until the above-mentioned deviation degree exceeds the threshold value. The abnormality determination unit 33 determines whether or not the estimated effective time is less than the threshold value, thereby determining abnormality. The output unit 34 outputs a signal relating to the abnormality when the abnormality determination unit 33 determines the abnormality.
[0043]
　The setting of the reference model and the setting of the deviation degree and the threshold value of the estimated effective time can be performed by the same processing as in FIG. 7. In addition to the allowable values ​​1 and 2, the threshold value setting unit 32 presets the allowable value 3 in advance. FIG. 10 is an example of a flowchart executed when the abnormality determination unit 33 performs abnormality determination after the threshold values ​​of the tolerance values ​​1, 2 and the estimated effective time are set by the threshold value setting unit 32. The abnormality determination unit 33 collects the data set after the tolerance values ​​1 and 2 and the threshold value of the estimated effective time are set by the threshold setting unit 32 (step S 21). This data set is collected for each detection value at each location of the temperature sensor 20.
[0044]
　Next, the abnormality judgment unit 33 judges whether the Mahalanobis square distance of any part exceeds the allowable value 1 or whether the integrated value of the Mahalanobis squared distance of the relevant part exceeds the allowable value 2 (step S22 ). If it is determined in step S22 that none of them exceeded, step S22 is executed again. When it is determined that any one of them exceeds in step S22, the abnormality determination unit 33 determines whether or not the Mahalanobis square distance exceeds the allowable value 3 (step S23).
[0045]
　If "Yes" is determined in step S23, the output unit 34 outputs a signal relating to the abnormality (step S24). If "No" is determined in the step S23, the abnormality determination section 33 determines whether or not the estimated effective time is less than a predetermined time (for example, 10 minutes in a measurement at an interval of 30 seconds) (step S25) . If "Yes" is determined in step S25, the output unit 34 outputs a signal relating to the abnormality (step S26).
[0046]
　If "No" is determined in step S25, the abnormality determination unit 33 again calculates the Mahalanobis square distance and the Mahalanobis square distance in the MSD method using past data (for example, 1 hour in the measurement at intervals of 30 seconds) from that time The parameter for deriving the statistic to be obtained is obtained (step S27). In this case, the average value, the unbiased variance covariance matrix, the inverse matrix, and the like in the new modeling period at each location of the temperature sensor 20 are included. Thereafter, the process is repeated from step S21.
[0047]
　In the example of FIG. 10, three allowable values ​​are set, and when the Mahalanobis squared distance exceeds the tolerance value 1 or the integrated value exceeds the tolerance value 2, estimation is performed again using the new data group, but before that , And when it exceeds the allowable value 3, it is judged as abnormal. For example, in the case of a method of deriving the Mahalanobis square distance, a method of comparing successively obtained Mahalanobis squared distances with an allowable value 1 and comparing the integrated value of the Mahalanobis square distances with the allowable value 2 can be mentioned. In addition, for example, when measuring the temperature at plural points, when the temperatures of the plural points coincide and change to dangerous temperature, etc., the Mahalanobis square distance is small but it can be said to be abnormal state. In order to avoid such a situation, allowable value 3 is set and compared with measurement data itself. The permissible value 3 may not be fixed, and a unique value may be determined for each measurement point. The set number of allowable values ​​and the setting method are not limited to the example of FIG. 10.
[0048]
　Also in the present embodiment, since the abnormality is determined when the estimated effective time becomes short, it is possible to determine the sign of abnormality. In addition, the threshold value of the deviation degree of each detected value of the temperature sensor 20 with respect to the reference model is determined based on the deviation degree for a certain period after creation of the reference model. In this case, the setting accuracy of the deviation degree threshold is improved. Further, the threshold value of the estimated effective time is determined based on the variation in time until the deviation degree exceeds the threshold value. In this case, the accuracy of setting the threshold value of the estimated effective time is improved.
Example 1
[0049]
　Specific examples will be described according to the above embodiment. In the first embodiment, the same object variables and explanatory variables as those in FIG. 6 were used according to the first embodiment. The system according to the first embodiment aims to predict from the explanatory variable group whether the wall surface temperatures 1 to 3 are kept within a proper range and to turn the operation cycle in the most efficient state. When the operation cycle becomes too high temperature and high pressure combustion before the boiler occurs, and if the operation temperature is too low, the combustion efficiency in the boiler decreases. It is required to control the optimum temperature and pressure by avoiding combustion.
[0050]
For these explanatory variables, threshold values ​​were set as follows according to the processing of FIG. In this embodiment, the same numerical values ​​are used for object variables 1 to 3.
Tolerance 1 ± 2 ° C
Tolerance 2 ± 10 ° C
Effective time threshold less than 20 minutes
[0051]
The modeling time for each estimation was 1 hour, and the principal component regression was used for the estimation. The specific method of principal component regression is shown below. (1) Generate a variance covariance matrix of 14 rows and 14 columns using the explanatory variable group and the value of objective variable 1 in the modeling period. Specifically, it is arranged at the last stage so that the variance of the objective variable 1 becomes 14th row and 14th column. (2) Inverse matrix is ​​generated for 13 rows and 13 columns excluding the row / column of the objective variable. (3) Obtain the product excluding the inverse matrix obtained in (2) and the element in the 14th row of the 14th column found in (1) to obtain 13 numerical values, and combine them with each coefficient of the explanatory variable . (4) Calculate the average value of each explanatory variable in the modeling period, multiply each average value by the coefficient obtained in (3), and take the sum of them. (5) Calculate the average value of the objective variables in the modeling period, subtract the value obtained in (4), and use this as the constant of the estimation equation.

[0052]
　In this embodiment, the system malfunction is substituted with the system operation stop. 11 (a) and 11 (b), the operating state of the system is changed from around 17:20, but no abnormality accompanying it has occurred, but at 18: 18, the explanatory variable Suddenly fluctuate. This is because the operation of the system was stopped. It is effective as a system if it can estimate system shutdown at 18:18 as soon as several minutes.
[0053]
　FIG. 12 (a) shows the instantaneous value of the estimation error. FIG. 12 (b) shows the integrated value of the estimation errors. FIG. 12 (c) shows the estimated effective time. The position where the integrated value is reset to zero and the data position of FIG. 12 (c) are the same in FIG. 12 (b). This is because the instantaneous value in FIG. 12 (a) exceeds the tolerance value 1 or the integrated value in FIG. 12 (b) exceeds the tolerance value 2 at that time, and the re-estimation calculation is performed. As can be seen from FIG. 12 (a), as described above, the estimation error immediately after the start of re-estimation is small. Also, when comparing Fig. 11 (a), Fig. 11 (b) and Fig. 12 (c), the estimated effective time gradually decreases toward 18:18, and from the time of 17:55 at the 18:12 time It meets the condition that it can be judged as abnormal, but you can see that the automatic judgment of this abnormality is reasonable. It means that the abnormality confirmation was made 6 minutes before 18:18, and the system can issue a stop command promptly.
Example 2
[0054]
　The second embodiment is an example according to the second embodiment. As exemplified in FIG. 13, four sets of winding parts were prepared and adhered closely to the wall surface of the furnace. Each of these four sets was used as a temperature sensor. In FIG. 13, a portion drawn with a circle is a wound portion, and each wound portion is connected by the same optical fiber. In addition, temperature distribution is drawn using mesh pattern. The temperature is low in the coarse mesh part, and the temperature is high in the fine mesh part. If the excessive heat accumulation starts inside each device, the temperature rises partly, so that abnormality can be detected by this.
[0055]
In the present embodiment, three wound portions are extracted from each of the four winding portion sets, and correlation analysis is performed on a total of twelve winding portions, thereby detecting a predictor of abnormality. Specifically, let us assume that the lower left corner of each winding part set is the reference origin (X, Z) of the local coordinate = (0, 0) and the coordinates that each winding part is included at the lower left and the upper right two points Set the area with.
In other words,
the winding unit set 1 region 1 (X1a1, Z1a1), ( X1a2, Z1a2)
　　　　　　　region 2 (X1b1, Z1b1), ( X1b2, Z1b2)
　　　　　　　region 3 (X1c1, Z1c1), ( X1c2, Z1c2)
winding part set 2 areas 1 (X 2 a 1, Z 2 a 1), (X
　　　　　　　2 b 2, Z 2 a 2 ) area 2 (X 2 b 1, Z 2 b 1), (X 2 b 2, Z 2 b 2), and the
like.
[0056]
　Furthermore, the average value, the maximum value, the minimum value and the like thereof are obtained from the temperature of each position of the optical fiber included in each region and used as the temperature data of each region. In this system, a threshold value is set for each of these 12 temperatures, and this is set to a tolerance value 3. In the present embodiment, since the temperature corresponding to the allowable value 3 was never exceeded, only the allowable values ​​1 and 2 and the valid time threshold were set.
[0057]
　Tolerance value 1 is set for the value of Mahalanobis square distance for the data of the modeling period used at that time in the new data set. Tolerance value 2 is set for the average value of the Mahalanobis square distance of one sample ago and the Mahalanobis square distance of the new data set.
[0058]
　The reason for not using the integral value is as follows. That is, the Mahalanobis squared distance indicates how far the new data set is with respect to the center of gravity of the dataset in the modeling period. The Mahalanobis squared distance can be updated in such a way that the data set is updated so as to rotate at a constant distance around the center of gravity when adding in consideration of the vectorial component in any direction and the case where the data set transits to different quadrants across the center of gravity Such cases can be indicated by different numerical values. However, in scalar addition, they are considered to be the same.
[0059]
Therefore, based on the assumption that the nearby data goes in the same direction at least as the data set comes off gradually, the average value is adopted as the allowable value 2.
Tolerance 1 60
Tolerance value 2 50
Effective time threshold less than 20 minutes
[0060]
　The method of obtaining the Mahalanobis square distance is as described above. Specifically, the results of predictive detection by finding the Mahalanobis square distance are exemplified in FIGS. 14 (a) and 14 (b). As exemplified in FIG. 14 (a), in the scoring period, the Mahalanobis square distance is the smallest immediately after the modeling period. The Mahalanobis distance increases as time elapses, but its increasing trend differs in each time zone. In the re-estimation at 17:04, the estimated effective time at which it reached 50 minutes became shorter after that, the estimated effective time of re-estimation at 17: 54 was 14 minutes, the estimated effective time at 18: 08 was 8 minutes . Compared to the threshold value of 20 minutes, abnormality confirmation was made at 18:08 after 14 minutes at 17: 54. This is 10 minutes earlier than 18:18 when the system was stopped, and in this example it was described about the system shutdown, but even in the case where signs of some kind of accident are actually seen, the time during which initial response is possible It is possible to find the possibility to secure.
[0061]
　Modified Example 1 In
　the first and second embodiments, the temperature measuring method of the plurality of temperature sensors 20a to 20c or the temperature sensor 20 is a method using Raman scattered light in the optical fiber, but this But it is not limited to this. For example, as the temperature sensors 20a to 20c or the temperature sensor 20, a thermocouple, a temperature measuring resistor, a camera imaging type infrared thermography, or the like is used.
[0062]
　However, when using a thermocouple or resistance temperature detector, two conductors with insulation maintained between the conductors and between the conductors and the wall are required for each measurement point. Also, when infrared thermography is used, it is necessary to inspect that the surface to be measured is not insulated or the like, there is no shielding object, it is possible to take an image from the outside, a plurality of temperature data in the vicinity of the position corresponding to the measurement point To be integrated into a single measurement point by averaging or the like and to obtain the emissivity of the outer wall surface beforehand for accurate conversion from luminance to temperature.
[0063]
　In the first embodiment and the first embodiment, the temperature is set as the objective variable and the other sensing data is set as the explanatory variable. However, other sensing data may be used as the objective variable, and the temperature to be measured may be other sensing It may be an explanatory variable together with data. Since it is only what kind of value to pay attention to, for example, it is possible to use a usage method in which the efficiency of the power generation system is deteriorated if an abnormality prediction is detected using the amount of electric power as a target variable . This is also true in the second embodiment and the second embodiment. For example, instead of the temperature sensor 20, the above embodiment may be applied to other sensing data. This will be described later in Example 4.
[0064]
　(Modification 2)
　FIG. 15 is a diagram illustrating a determination system according to Modification 2. In the second embodiment, the determination unit 30 directly acquires data from the temperature sensor 20. On the other hand, in the determination system of Modification 2, the server having the function of the determination unit acquires data from the temperature sensor through the electric communication line.
[0065]
　The determination system of Modification 2 includes a temperature sensor 20, a server 202, and a monitoring server 203. The temperature sensor 20 has a sensor unit 21 for acquiring temperature data of the measurement object, and a measuring device 22 for acquiring measurement data from the sensor unit 21 and generating temperature data.
[0066]
　The temperature sensor 20 has a configuration connected to the server 202 via an electric communication line 201 such as the Internet. The monitoring server 203 that monitors the measurement target on which the sensor unit 21 is installed is connected to the electric communication line 201. The server 202 includes the CPU 101, the RAM 102, the storage device 103, the interface 104, and the like in FIG. 5 (b), and realizes the function as the judgment unit 30.
[0067]
　In such a judgment system, for example, the server 202 installed in Japan receives measurement data measured by a coal bunker of a power plant in another country, and detects a sign of abnormal heat generation in the coal bunker. The result output from the server 202 is transmitted to the monitoring server 203.
[0068]
　This modified example can also be applied to the first embodiment. For example, instead of the temperature sensor 20, a plurality of temperature sensors 20a to 20c may be used.
[0069]
　(Modification 3)
　FIGS. 16A and 16B are diagrams illustrating the sensor unit 21 and the measuring device 22. As exemplified in FIG. 16 (a), the sensor part 21 is attached to the outer wall of the mill intermediate housing part 40 of a pulverized coal machine for crushing coal and producing fine powder, for example. 16 (a), the mill middle housing part 40 includes a storage part 42 in which the coal 41 falls and is temporarily stored, and a crushing ring 42 for crushing the coal 41 stored in the storage part 42 43 and a roller 44. The pulverized coal 45 obtained by pulverization rises due to the air in the primary air chamber 46.
[0070]
　As illustrated in FIG. 16 (b), the measuring device 22 includes a laser 11, a beam splitter 12, an optical switch 13, a filter 14, a plurality of detectors 15 a and 15 b, a calculation unit 16, and the like. The laser 11 is a light source such as a semiconductor laser and emits a laser beam having a predetermined wavelength range. For example, the laser 11 emits light pulses (laser pulses) at predetermined time intervals. The beam splitter 12 makes the optical pulse emitted from the laser 11 incident on the optical switch 13. The optical switch 13 is a switch for switching the emission destination (channel) of the incident optical pulse. In the double-end system, the optical switch 13 alternately applies light pulses to the first end and the second end of the optical fiber 23 of the sensor section 21 at a constant period. In the single-end system, the optical switch 13 causes an optical pulse to be incident on either the first end or the second end of the optical fiber 23. The optical fiber 23 is arranged along a predetermined path of the temperature measurement target.
[0071]
　The optical pulse incident on the optical fiber 23 propagates through the optical fiber 23. The optical pulse gradually attenuates and propagates in the optical fiber 23 while generating forward scattered light traveling in the propagation direction and backward scattered light (returning light) proceeding in the feedback direction. The backscattered light passes through the optical switch 13 and reenters the beam splitter 12. The backscattered light incident on the beam splitter 12 is emitted to the filter 14. The filter 14 is a WDM coupler or the like, and extracts a long wavelength component (Stokes component) and a short wavelength component (anti Stokes component) of the back scattered light. The detectors 15 a and 15 b are light receiving elements. The detector 15 a converts the received light intensity of the short wavelength component of the backscattered light into an electric signal and transmits it to the calculation unit 16. The detector 15 b converts the received light intensity of the long wavelength component of the backscattered light into an electric signal and transmits it to the calculation unit 16. The calculation unit 16 measures the temperature distribution in the extending direction of the optical fiber 23 using the Stokes component and the anti-Stokes component.
[0072]
　FIG. 17 (a) is a transmission view of the sensor unit 21, and is a view passing through the seat 24 b of FIG. 17 (b). 17 (b) is a cross-sectional view taken along line A - A in FIG. 17 (a). The sensor section 21 is a fiber sheet in which the optical fiber 23 is arranged at a predetermined position. As illustrated in FIGS. 17 (a) and 17 (b), the sensor section 21 includes a pair of sheets 24 a and 24 b for holding the optical fiber 23 therebetween and a pair of sheets 24 a and 24 b for holding the gap between the sheets 24 a and 24 b, 25, and a slit metal pipe 27 roughly determining the position of the wound portion 26 of the optical fiber 23.
[0073]
　The optical fiber 23 has wound portions 26a to 26h (hereinafter collectively referred to as wound portions 26) held on the sheets 24a and 24b in a state of being wound plural times. FIG. 17 (c) shows an example in which the optical fiber 23 is wound in a single fashion. The wound portions 26a to 26h are constituted by one optical fiber 23 or one each in the upper and lower stage, and are composed of a total of two optical fibers 23. In the latter case, for example, fusion splicing is performed at the upper-stage lower connecting portion in FIG. 17 (a). The sheet 24 a is in contact with the object to be measured. An adhesive tape 28 is provided on the sheet 24 a. Thereby, the sheet 24 a can be pasted on the temperature measurement target.
[0074]
　Each of the wound portions 26a to 26h is wound, for example, 2 to 8 times. The diameter of the optical fiber 23 varies depending on the heat resistant temperature, but since it is 0.16 to 0.4 mm, the inner diameter of the metal tube 27 is smaller than the diameter of the optical fiber 23 About 1 to 2 mm which is twice the diameter or more is required. Since the thickness of the metal pipe 27 is about 0.5 mm, it has a thickness of about 2 to 3 mm from the seat 24 a to the seat 24 b.
[0075]
　In the second embodiment, the average value, the maximum value, the minimum value and the like thereof are obtained from the temperature of each position of the optical fiber included in each region, and these are taken as the temperature data of each region, but the winding The portion 26 has a thickness as shown in FIG. 17 (b), and the temperature of the portion of the wound portion 26 that is distant from the measurement target may deviate considerably from the actual temperature of the measurement target in some cases.
[0076]
　Therefore, from among temperature values ​​indicated by a plurality of measurement points included in the winding portions 26a to 26h, five predetermined points, for example, five points are selected in descending order of the highest temperature to obtain an average value, The temperature of the region corresponding to each winding part is set as the temperature of each turning part. As a result, the accuracy of the temperature to be measured is increased, and the predictor of the abnormal state of the temperature of the measurement object can be detected more accurately.
[0077]
(Third Embodiment) A
　third embodiment will be described as an embodiment that holds a different purpose from the second embodiment. In the second embodiment illustrated in FIG. 10, although it is assumed that various sensing data are used, a concrete method has not been shown. Therefore, in this embodiment, an example of the method will be described first. This method itself is similar to that disclosed as conversion to a random variable also in Patent No. 5308501, and it is a general method in ordinary multivariate analysis.
[0078]
　FIG. 18 is a diagram illustrating a flow chart illustrating a dimensionless procedure for modeling using N pieces of sensing data S1 (t) to SN (t) at time t. This dimensionlessization procedure is performed when the model creation unit 31 executes steps S3 and S4 in FIG. 7, when the abnormality determination unit 33 executes S15 in FIG. 8 or step S27 in FIG. 10. The N pieces of sensing data S1 (t) to SN (t) are each object variable and each explanatory variable. Here, a description will be given of a case where the model creation unit 31 is mainly used.
[0079]
The model creation unit 31 obtains an average value and a standard deviation of each sensing data in a predetermined past time (modeling period) from the current reference time t (step S 31). Each average value S1_ave to SN_ave between a predetermined past time ΔT used for modeling from the time t0 of each of N pieces of the sensing data S1 (t) to SN (t) at the time t is expressed by the following equation. ..., SN_ave = Average (SN (t 0), ...,
S 2 _ ave = Average (S 2 (t 0), ..., S
2 (t 0 - ΔT) SN (t 0 - ΔT)) The standard deviations S 1 _sigma to SN _ sigma are expressed by the following equations. S1_sigma = Standard Deviation (S1 (t0), ..., S1 (t0-[Delta] T)), S2_sigma = Standard Deviation (S2 (t0), ..., S2 (t0-[Delta] T)), ... SN_sigma = Standard Deviation (SN (t0) , ..., SN (t 0 - ΔT))

[0080]
　Next, at the time of deriving the parameters, the model generating unit 31 first subtracts the average value from each sensing data value to obtain a zero reference value, and further divides the average value by a standard deviation value or a value several times larger than that (step S 32 ). As a result, each sensing data is made dimensionless. When the abnormality determination unit 33 performs the dimensionless procedure, when executing the step S 11 of FIG. 8 or the step S 21 of FIG. 10, the dimensionless processing is performed on the new data group.
[0081]
　According to the dimensionless processing, in each modeling period, the average value of each sensing data becomes zero and the variation of values ​​becomes uniform. Since the variation of the value can be corrected by the method of deriving the parameter later, even if the standard deviation value is not used and each sensing data value is divided by each calculated average value and the parameter is derived by the value obtained as a result Good. In that case, the average value in the modeling period is 1. By this dimensionless procedure, it is possible to handle sensing data with different wall dimensions such as the wall surface temperature 1 to 3, the amount of electric power, the amount of coal feed, the pressure 1 to 3, etc. illustrated in FIG. 6 by the method of FIG. 8 and FIG. 10, for example It becomes possible.
[0082]
　Incidentally, in most cases, an abnormality predictor in a plant does not converge immediately after only one event of some kind occurs, and it can be said that a natural ignition phenomenon of coal exemplified in FIG. 19 (a) As in the case of bearings degradation faults that are slowly deteriorating, they start to occur frequently at an accelerated pace, which leads to a continuous phenomenon leading to an accident, and it is said that there is a sudden occurrence in actual power plants, plants, etc. It is necessary to think differently from the case that some kind of abnormality occurs only for a certain period of time, but it will fit normally shortly afterwards. It is to be noted that in FIG. 19 (a), it is indicated that the spontaneous ignition phenomenon occurs frequently in any of coal types A to H at an accelerated rate.
[0083]
　Therefore, the procedure of FIG. 20 is disclosed as one method of this embodiment. The device configuration is the same as that of the second embodiment. The abnormality determination unit 33 collects the data set after the tolerance values ​​1 and 2 and the threshold value of the estimated effective time are set by the threshold setting unit 32 (step S41). This data set is collected for each detection value at each location of the temperature sensor 20.
[0084]
　Next, the abnormality determination unit 33 determines whether or not the Mahalanobis square distance of any part exceeds the tolerance value 1 and the integrated value of the Mahalanobis square distance of the place exceeds the allowable value 2 (step S 42) . If "No" is determined in step S 42, the process is executed again from step S 41. If "Yes" is determined in step S 42, the abnormality determination unit 33 determines whether or not the Mahalanobis square distance exceeds the allowable value 3 (step S 43).
[0085]
　If "Yes" is determined in the step S43, the output unit 34 outputs a signal relating to the abnormality (step S44). If "No" is determined in the step S43, the abnormality determination unit 33 again calculates the Mahalanobis square distance and the Mahalanobis square distance in the MSD method using past data (for example, 1 hour in measurement at 30-second intervals) from that time The parameter for deriving the statistic to be obtained is obtained (step S 45). In this case, the average value, the unbiased variance covariance matrix, the inverse matrix, and the like in the new modeling period at each location of the temperature sensor 20 are included.
[0086]
　Next, using the current data set collected in step S41, the abnormality determination unit 33 starts the outlier test of the following steps S47 and S48 (step S46). First, the abnormality judgment unit 33 judges whether or not the Mahalanobis square distance of any part exceeds the allowable value 1 and the integrated value of the Mahalanobis squared distance of the place exceeds the allowable value 2 (step S47). If "No" is determined in the step S47, the abnormality determination unit 33 determines whether or not the estimated effective time is less than a predetermined time (for example, 10 minutes in the measurement at the interval of 30 seconds) (step S48) . If "Yes" is determined in the step S48, the abnormality determination section 33 outputs a signal relating to the abnormality (step S49).
[0087]
　If "No" is determined in the step S47, the abnormality determination unit 33 sets the estimated effective time to a fixed value larger than the predetermined time (step S50). Thereafter, it is executed again from step S41. If "No" is determined in step S48, the abnormality determination unit 33 resets the estimated effective time to zero (step S51). Thereafter, it is executed again from step S41.
[0088]
　In the example of FIG. 20, unlike FIG. 10, instead of performing outlier test using a new data set after modeling, an outlier test is performed again with the current data set and the estimated effective time is reset according to the result It is deciding whether or not. This is because if the new data set exceeds the tolerance value 1 or the tolerance value 2 at a point in time when a certain period of time elapses without modeling, it is actually caused by the occurrence of an unexpected event Or whether it was simply due to change in coal type or output command change. This technical idea can also be applied to the first embodiment. If the permissible value 1 or the allowable value 2 is exceeded again after modeling, it may be considered that it corresponds to the former. By setting the state in which the estimated effective time can be evaluated, it is possible to determine whether or not it is in an acceleration situation It becomes possible to judge. When using this method, regardless of the dimensions of the data set, using the method disclosed in FIG. 18, or as described above, instead of dividing by the standard deviation, division by the average value is used to make dimensionless It is more preferable to perform. In step S 42 and step S 47 of FIG. 20, "and" determination may be "or" determination.
Example 3
[0089]
　Comparison between the third embodiment illustrated in FIG. 20 and the comparative example in FIG. 21 which always performs modeling in successive outlier tests, which is a standard moving window also used in Patent No. 5308501 and the like Then, the effect of the third embodiment is clarified. In the comparative example shown in FIG. 21, when the set data group is used (step S61) and the Mahalanobis distance of any part exceeds the allowable value 1 and the allowable value 2 (step S62), a signal relating to the abnormality is output (Step S 63). If the Mahalanobis distance exceeds the tolerance value 3 in the case of "No" in step S 62 (step S 64), a signal relating to the abnormality is also output (step S 65). In the case of "No" in step S 64, from the point in time that past data (for example, one hour in the measurement at 30-second intervals) is used again to derive the statistical amount corresponding to the Mahalanobis square distance and the Mahalanobis squared distance in the MSD method Parameters are obtained (step S 66).
[0090]
　In each of the examples of FIGS. 20 and 21, the dimensionless processing exemplified in FIG. 18 was performed. In this case, however, instead of dividing by the standard deviation, which is the other method described above, a method of dividing by the average value is used. As an example, a thermal power generation facility utilizing the coal combustion cycle shown in the first embodiment will be taken as an example. Instead of the wall surface temperatures 1 to 3 in FIG. 6, those processed into a sheet shape having a large number of wound portions as shown in FIG. 9 (b), FIG. 13, and FIG. 17 are installed on the wall surface of the facility as shown in FIG. 16 Then, the measurement data of 64 points are extracted and treated as the data set for each measurement.
[0091]
　FIG. 22 illustrates the results of the comparative example. The instantaneous value is the outlier distance calculated using the new data set, and the two-point average is the two-point average value of the outlier distance calculated last time and the outlier distance calculated this time. Allowable value 1 is the average value of the instantaneous value of the model update period plus 3 times (3σ) the standard deviation of the instantaneous value and the tolerance 2 is the average value of the 2 point average value plus 3 times the standard deviation of the instantaneous value (3σ ) Value. Tolerance value 3 was set to 8 times the average value of instantaneous value plus standard deviation of instantaneous value (8σ). In the abnormality determination, 0 is outputted when it is not abnormal in FIG. 21, and 1 is outputted in case of abnormality determination. In the example of FIG. 22, abnormality judgment frequently occurs at 18:14 which is 4 minutes faster than 18:18 when the system was stopped. However, in FIG. 22, since erroneous detection occurs three times from midnight, it is understood that reliability as a system is impaired.
[0092]
　Figure 23 illustrates the results of the method described in Figure 20. The definition of instantaneous value, two-point average, abnormality judgment, allowable value 1 to 3 is the same as in FIG. The threshold of the estimated effective time was 4 minutes. Although it is 2 minutes earlier than 18:18 when the system was stopped, abnormality judgment occurred frequently at 18:16 which is two minutes later than the method in FIG. 21. However, unlike FIG. 22, there is no erroneous detection, and it can be seen that the reliability of abnormality determination is kept.
[0093]
　22 and 23, the allowable value 1 is set to be the average value of the instantaneous value + 2 times the standard deviation of the instantaneous value (2σ) value, and the allowable value 2 is set to the average value of the two point average + the standard deviation of the instantaneous value Fig. 24 and Fig. 25 show the change to 2 times (2σ) value. In both cases, the time when abnormality judgment starts to occur frequently does not change from FIG. 22 and FIG. 23. In the method described in FIG. 21, further erroneous detection occurs more frequently, but in the method of FIG. 20, erroneous detection does not occur and the reliability of abnormality determination is kept.
[0094]
　That is, by using this embodiment which is different from the outlier test etc. of the moving window type as shown in Japanese Patent No. 5308501, it is possible to calculate, for the allowable values ​​1 and 2, from the average value + twice the standard deviation (2σ) It can be said that it is possible to detect a predictor even if it is set with an ambiguous width of 3 times the deviation (3σ).
Example 4
[0095]
　In the fourth embodiment, in addition to the temperature, detection of an abnormality prediction exemplified in FIG. 26 is executed for various sensing data of a thermal power generation facility utilizing the combustion cycle of coal exemplified in the first embodiment. Specifically, the processing exemplified in FIG. 26 is performed on the time series data of 16 pieces of sensing data including the object variables 1 to 3 and the explanatory variable group of the coal thermal power generation facility shown in FIG. 6. First, non-dimensionalization is performed using the method illustrated in FIG. However, here, as in the third embodiment, a method of dividing by the average value instead of dividing by the standard deviation is used.
[0096]
　The difference between the processing in FIG. 26 and the processing in FIG. 20 lies in forcibly re-modeling if a certain long-term time is exceeded while sequential processing is being performed. Specifically, the abnormality determination unit 33 differs in that it determines whether or not the current time is within the forced update time (step S52) after execution of step S41 and before execution of step S42. If "Yes" is determined in step S52, step S42 is executed. If "No" is determined in step 52, step S 45 is executed. The reason why step S 52 is executed is that since the data contributing to outliers clearly changes after a certain period of time passes even if the outliers are transitioning with no problem, it is more preferable to correct them . This was taken as the forced update time. The forced update time is set to be several times longer than the estimated effective time.
[0097]
　In the present embodiment, as in Example 3, the tolerance value 1 is an average value of instantaneous values ​​between model updates + triple (3σ) value of the standard deviation of instantaneous values, and the allowable value 2 is an average value of 2 point average values + 3 times the standard deviation of the instantaneous value (3σ) value. Tolerance value 3 was set to 8 times the average value of instantaneous value plus standard deviation of instantaneous value (8σ). The estimated effective time was 5 minutes, and the forced update time was 40 minutes. Note that each sensing data is collected every 2 minutes.
[0098]
　The results are illustrated in FIG. 27. In FIG. 27, abnormality is judged at 17:24 which is one hour earlier than 18:18 when the system was stopped. After that, once it returned to normal once again, abnormality frequently occurs again at 18:10. Therefore, FIG. 28 exemplifies each of the sensing data normalized by using the average value from 0 o'clock to 2 o'clock on the same day. The reason why we standardized with fixed values ​​is because we did not do modeling etc. and wanted to compare with the same index.
[0099]
　Referring to FIG. 28, it can be seen that the amount of coal is reduced from 17: 22, and accordingly the plurality of sensing data starts showing different trends. Since the current estimated effective time is 5 minutes, there is a high possibility that the abnormality will be determined after 17: 27, but it became abnormal at the time of 17:24 of the next measurement that started reducing the coal feed . This is because several data including the coal feed amount has further decreased, but the other data exceeded the value of 8 times the standard deviation of the mean value + instantaneous value (8σ) because there was no clear change is there. Although the decrease in the amount of coal was relieved and it settled once, many parameters including the amount of coal supply started changing again from 18:10, but since the change is steep, at that point the judgment result will be the average + instant Since it exceeds the value of 8 times the standard deviation of the value (8σ), an abnormality determination is made.
[0100]
　As described above, according to the premonitory detection using the present invention, not only the temperature but also the correlation between the sensing data having various dimensions, even if a threshold including a certain degree of ambiguity is set, it is highly responsive and accurate It can be seen that it is possible to detect a predictor of abnormality. Incidentally, in steps S 42 and S 47 of FIG. 26, "and" determination may be "or" determination.
[0101]
　Although the embodiments of the present invention have been described in detail above, the present invention is not limited to the specific embodiments, and various modifications and changes may be made within the scope of the present invention described in the claims. It is possible to change. For example, in each of the above examples, the threshold value for the estimated effective time is one, but a second threshold value larger than the first threshold value may be further set. In this case, when the estimated effective time is longer than the first threshold value and less than the second threshold value, even if the warning information alarm is outputted as being one that is higher in safety level than the abnormality warning but calls attention Good.
Explanation of sign

　A process of generating a reference model 　of a sensor detection value
　from a predetermined time after creation of the reference model and a time until a degree of deviation between the reference model and the sensor detection value exceeds a threshold value for a predetermined time And
　a process of outputting a signal relating to abnormality in a case where it is determined to be sho