JP2001356818A - On-line maintenance device for in-plant detector - Google Patents

Abstract

PROBLEM TO BE SOLVED: To evaluate the integrity of an in-plant detector through an on-line. SOLUTION: An on-line maintenance device 10 of an in-plant detector is provided with an input means 19 for receiving detector signals from plural detectors correlated with each other in a plant 60, at least two estimating means 11 and 13 for estimating the true values of the detector signals received from the input means 19, an integrity evaluating means 15 for evaluating the integrity of the detector from the deviation of the detector signals and the true values outputted from the estimating means 11 and 13, a switching means 17 for receiving the output of the integrity evaluating means 15, and for switching the estimating means 11 and 13 for receiving the detection signals from the input means, and an output means 21 for outputting the estimated results of the true values of the detector signals or the integrity evaluated results.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus for online evaluation of the soundness of various detectors arranged in a plant or a mechanical device.

[0002]

2. Description of the Related Art To check the soundness of a detector used in a plant or the like, conventionally, calibration and adjustment work has been performed when the operation of the plant or the machine is stopped, such as during a periodic inspection. Specifically, a reference signal generator is attached to the target detector in the plant to provide a reference input signal, and it is checked whether or not the signal output from the detector matches the reference output value. If not, change the setting based on the adjustment procedure of the detector. For example, in the case of a semiconductor type differential pressure detector, a dedicated tool is connected at the time of calibration and adjustment, and the set value inside the semiconductor is changed from the dedicated tool.

[0003] In the detector soundness confirmation method as described above,
It cannot be carried out unless the operation of the plant is stopped, which is not preferable in that the state of the detector is always kept in a suitable state, and the plant operation is frequently stopped to check the soundness of the detector. In this case, the operation cost increases, which is not preferable. From such a viewpoint, attempts have recently been made to check the soundness of the detector online during the operation of the plant. FIG. 6 is a block diagram of an example of what is actually proposed and tested. In general, a plurality of correlated detector signals from a plant or a mechanical device 60 are input to the maintenance device 50 by the input means 54. Capture and send to the switching means 53. Switching means 5
Numeral 3 designates an input means 5 for a detector which has not been determined by the soundness evaluation means 52 to be described later that a drift has occurred.
4 is sent to the estimating means 51.
For the detectors that have been determined by the soundness evaluation means 52 to have drifted, the estimation values previously estimated by the estimation means 51 described later are sent to the estimation means 51. Estimating means 51
Has a neural network (described later) therein.
In this neural network, internal parameters are learned in advance by a detector signal at the time of soundness, and the estimating means 51 inputs a plurality of detector signals to the learned network, and estimates an estimated value of each detector signal. And output. The deviation of the estimated value signal from the detector signal fetched from the input means 54 is obtained and sent to the soundness evaluation means 52. In the soundness evaluation means 52, the deviation in each detector is evaluated by a general test means called a sequential probability ratio test, and when it can be stochastically determined that the deviation exceeds a predetermined reference value, the detector is drifted. Is determined to have occurred, and it is determined that the detector is not sound.

Here, the outline of the sequential probability ratio test will be described. This is a method in which data is extracted at any time and a test is performed. FIG. 7 shows a conceptual diagram. In the detector soundness evaluation, the integrated value S of the deviation between the actually measured value and the estimated value of the detector signal is shown.
_{If n} exceeds the upper reference line U _n , it is determined to be abnormal, and if _n is lower than the lower reference value L _n , it is determined to be normal. It is assumed that the data to be tested has a normal distribution, and the logarithm of the sequential probability ratio when the i-th deviation is obtained is expressed by the following equation.

(Equation 1) Here, σ ² indicates the variance of the data.
In the equation (1), μ ₁ has no drift,
It is the average of the deviations when the detector is sound, and is substantially μ ₁ = 0. On the other hand, μ ₂ is the average value of the deviation after the drift has occurred. For example, if an abnormality is detected during a 1% drift, μ ₂ = 0.01. When n deviations are obtained, the statistic T _n is obtained by the following equation.

(Equation 2) The sequential probability ratio test adopts an anomaly when the statistic T _n has the relationship of the equation (3),

(Equation 3) When the statistic T _n has the relationship of equation (4), it cannot be said that the statistic is abnormal or normal

(Equation 4) Further, when the statistic T _n has the relationship of the equation (5), normal is adopted.

(Equation 5) It is a technique to judge. Here, α is a first type error rate,
β is a value called a second error rate, for example, α = β = 0.
05.

[0005] In addition, equation (3) showing the above-mentioned judgment criteria,
Expressions (4) and (5) are expressed as Expressions (7), (8) and (9), respectively, by using the following Expression (6).

(Equation 6)

(Equation 7)

(Equation 8)

(Equation 9) However, in the above relational expressions, there are the following expressions (10) and (11).

(Equation 10)

[Equation 11]

[0006]

SUMMARY OF THE INVENTION The above-mentioned conventional apparatus includes:
There are roughly two problems. The first is the simultaneousness of the input signals to the estimating means 51 when it is determined that the detector is not sound. The second relates to the sequential probability ratio testing means employed in the soundness evaluating means 52. . First, the problem of the synchronism of the input signals will be described. As described above, for the detector determined to be sound by the soundness evaluation means 52, the switching means 53 makes the estimation means 51 The estimated detector signal value is adopted and sent to the estimating means 51. The estimating means 51 performs an estimation process of the time t using the signal at the time t of the sound detector (actual measurement value) and the estimated value of the time t-1 for the unhealthy detector. There is no synchronization in time.
Therefore, the accuracy of the estimated value obtained by the estimating unit 51 may be reduced due to the noise signal included in the detector signal or the calculation cycle of the conventional device. In addition, in the sequential probability ratio test, as described above, the sensitivity of the soundness determination is greatly influenced by the behavior of the detector signal, that is, the variance of the signal, as is apparent from the definition equations of U _n and L _n. May be delayed or a false alarm may occur. Therefore, the present invention detects the drift of the detector early regardless of the noise or behavior of the detector signal and regardless of the operation cycle of the device,
It is an object of the present invention to provide a maintenance device for accurately estimating a true value of a detector signal.

[0007]

According to the present invention, there is provided, in accordance with the present invention, an on-line maintenance system for an in-plant detector which receives detector signals from a plurality of correlated detectors in the plant. Input means,
At least two estimating means for internally incorporating a neural network for estimating a true value of the detector signal received from the input means; and detecting the detector based on a deviation between the detector signal and the true value output from the estimating means. Soundness evaluating means for evaluating the soundness of the signal, switching means for switching the estimating means for receiving a detector signal from the input means in response to the output of the soundness evaluating means, and estimation results and soundness evaluation of the true value of the detector signal It has an output means for outputting the result. Then, the soundness evaluation means calculates a moving average value of the deviation for each detector, compares the moving average value with a predetermined threshold, and evaluates the soundness for each detector, or calculates the deviation of the deviation for each detector. The increase rate of the integrated value is calculated, and the soundness is evaluated for each detector from the increase rate. Further, a plurality of true value estimating units each including two estimating units, a soundness evaluating unit, and the switching unit may be arranged in parallel between the input unit and the output unit.

[0008]

Embodiments of the present invention will be described below with reference to the accompanying drawings. It is to be noted that the same portions are denoted by the same reference numerals throughout the drawings including the drawings related to the above-mentioned conventional technology. First, referring to FIG. 1, a detector online maintenance device (hereinafter abbreviated as a maintenance device) 10 includes two estimating units 11 and 13 for estimating a true value of a detector signal, and evaluates the soundness of the detector. Switching means 1 for switching the input of soundness evaluation means 15 and estimation means 13
7. It has an input means 19 which is connected to a plurality of detectors (not shown) of the plant 60 and receives the detector signals, and an output means 21 which outputs an estimation result of a true value of the detector signal and a soundness evaluation result. Note that the deviation calculating circuit 23 that calculates the deviation between the output of the estimating unit 13 and the output of the input unit 19 may be incorporated in the soundness evaluating unit 15.

Next, the configuration and operation of each means constituting the maintenance device 10 will be described. First, the estimating means 11,
13 has the same configuration, and has a neural network in which internal parameters are learned in advance using a detector signal obtained from a detector of the plant 60.
FIG. 2 shows the concept of the neural network. The left side is an input layer, the right side is an output layer, and an intermediate layer is between them. Then, the detector signal is input to the input layer, an operation is performed by a weight, a bias, and a conversion function inside the network, and an estimated value of each detector signal is output to the output layer. The number of input layers and the number of output layers match, and the estimated value for the i-th input signal is output to the i-th output layer. In FIG. 2, the number of nodes in each of the input layer, the intermediate layer, and the output layer is changed according to the number of input signals to the network, and is not limited to the example.

The soundness evaluation means 15 receives a deviation between the measured value of the detector signal input from the input means 19 and the estimated value of the detector signal calculated by the estimating means 13, and evaluates the soundness as described later. Then, the output is sent to the switching means 17 and the output means 21. When drift occurs in the detector, the deviation increases. The soundness evaluation means 15 stores the deviation inputted every moment for a certain period of time, and calculates an average value (hereinafter, referred to as a moving average value). For example, if it is stored for 100 times, the moving average value μ (t) of the deviation at time t at this time is expressed by the following equation.

(Equation 12) Here, e (t) is soundness evaluation means 1 at time t.
5 is the difference between the measured and estimated values of the detector signal. Regarding the soundness evaluation, when the moving average value μ (t) for each detector exceeds a preset threshold δ, it is determined that the detector is not sound. FIG. 3 is a conceptual diagram of soundness determination.

The switching means 17 selects a signal to be input to the means based on the determination result of the soundness evaluation means 15 and outputs it to the estimating means 13. (A) For a detector signal that has not been determined to be unhealthy, the measured value of the detector signal input from the input means 19 is selected. The detector signal (b) detector has been determined not to be healthy, at time t _1, which is determined to be not healthy, selects the detector signal measured value inputted from the input means 19. Then, the next time t ₁ determined to be unhealthy
At +1, an estimated value of the detector signal estimated by the estimating means 11 is selected. Further, at time t ₁ +2, which is the next time after the time determined to be unhealthy, and thereafter, at time t ₁ +1 until it is determined to be healthy, the estimating means 1
3 is used and fixed to this value.

The input means 19 receives the measured value of the detector signal from the plant 60 and sends it to the estimating means 11, the switching means 17 and the deviation calculating circuit 23. Output means 21
Sends the estimated value of the detector signal received from the estimation unit 13 and the soundness evaluation result received from the soundness evaluation unit 15 to, for example, a display unit such as a CRT or a printer of a monitoring system (not shown) or a control device. In the above embodiment, two estimating means are used. However, even if the number of estimating means is further increased, similar effects can be obtained by modifying the related switching means.

Next, another modified embodiment of the present invention will be described. The arrangement relationship of each means of this embodiment is the same as that of the above embodiment, but the evaluation procedure of the soundness evaluation means is different. In conventional health evaluation unit described above, in the manner of the sequential probability ratio test, the integrated value S _n of upper reference line U _n and deviation at a point crossing, were determined to be abnormal. As described above, in this determination method, the variance of the signal used for the test has a large effect on the determination, and thus there is a risk of erroneous detection. Therefore, in this modified embodiment, according to the concept shown in FIG.
Evaluate soundness. That is, it is determined the integrity of the detector increase of the deviation integrated value S _n at a certain time interval. Referring to FIG. 4, if the determination is made by dividing the time interval into 100 time steps (calculation cycles), the increments of the upper-limit reference line and the deviation integrated value are calculated by triangles A and B (B is exactly It is not a triangle, but there is no problem if it is roughly a triangle.) If the areas of triangles A and B are A <B, then the bases of the triangles are equal in length, so the slope of each is also upper reference line U _n <deviation accumulated value S _n, and the unpredictable and upper reference line intersects the deviation accumulated value curve in the future, it is determined that the detector is not healthy. Since this determination procedure does not use the dispersion of the signal, the abnormality of the detector can be stably detected without being affected by the properties of the signal. The time division is not divided into time intervals. For example, the latest value such as the integrated value of the deviation of the past 100 steps (meaning from time t-99 to current time t) is used. May be used as a value obtained by integrating a predetermined number of deviations.

Another embodiment of the present invention will be described with reference to FIG. Also in the present embodiment, the estimating units 11 and 13, the soundness evaluating unit 15, the switching unit 17, and the deviation calculating circuit 23 in the embodiment of FIG. In other words, these means are put together as one of the detector signal true value estimation blocks 35a, 35b, 35c, and are arranged in parallel between the input means 31 and the output means 33. Then, the input means 31 receives detector signals from a plurality of detectors in the plant 60, for example, al to l, and outputs these signals to each of the blocks 35a, 35b, 35c.
Distribute to Table 1 shows an example of distribution.

[Table 1] As shown in Table 1, in this embodiment, the input means 31 detects the signals of the detectors a to i in the block 35a, detects the signals of the detectors d to L in the block 35b, and detects the signals of the detectors d to L in the block 35c. The signals of the devices a to c and g to L are transmitted.

Each of the blocks 35a, 35b, 35
In c, the same processing as in the above-described embodiment is performed on the distributed detector signal. The output unit 33 receives the estimated values of the detector signals and the results of the soundness determination described in Table 1 from the blocks 35a, 35b, and 35c. Further, here, for example, the following comprehensive evaluation is performed, and a final estimated value of the detector signal is obtained and output. (I) A block having the smallest number of detectors determined to be unhealthy is given top priority, and the estimated value transmitted from the block is adopted. (Ii) If the number of detectors determined to be unhealthy is the same, the average of the estimated values transmitted from each detector is used as the final estimated value. For example, in each of the blocks 35a, 35b, and 35c, the number of detectors determined to be unhealthy is 0, 1
In the case of one detector, the detectors a to i use the estimation result of the block 35a, and the detectors j to L use the average value of the estimation values of the blocks 35b and 35c. The output destination of the output unit 33 is a display device such as a CRT or a printer of the monitoring system or a control device.

According to the above-described embodiment, the soundness of a plurality of detectors can be monitored and evaluated simultaneously, regardless of the noise or behavior of the detector signals, and the operation of the online maintenance device can be performed. Irrelevant detectors can be detected early, regardless of the period, and the true value of the detector signal can be accurately estimated.

[0017]

As described above, according to the present invention,
In the soundness evaluation means for evaluating the soundness of the detector, since the successive probability ratio test method is not used, the true value of the detector signal with high accuracy can be estimated with high reliability.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention.

FIG. 2 is a conceptual diagram illustrating the operation of one component of the embodiment.

FIG. 3 is a graph for explaining the operation of the embodiment.

FIG. 4 is a conceptual diagram illustrating the operation of one component of another embodiment.

FIG. 5 is a block diagram showing still another embodiment.

FIG. 6 is a block diagram of a conventional device.

FIG. 7 is a conceptual diagram for explaining the operation of the conventional device.

[Explanation of symbols]

 Reference Signs List 10 detector online maintenance device 11, 13 estimation means 15 soundness evaluation means 17 switching means 19 input means 21 output means 31 input means 33 output means 35a, 35b, 35c block

Continuation of the front page (72) Inventor Daisuke Asada 1-1-1 Wadazakicho, Hyogo-ku, Kobe-shi, Hyogo F-term in Kobe Shipyard, Mitsubishi Heavy Industries, Ltd. (reference) 5H004 GA28 JB07 JB10 JB17 JB20 JB29 JB30 KD33 KD42 5H223 AA01 EE02 FF06

Claims (3)

[Claims] 1. An input means for receiving a detector signal from a plurality of mutually correlated detectors in a plant, and at least two built-in neural networks for estimating a true value of the detector signal received from the input means. Estimating means, a soundness evaluating means for evaluating the soundness of the detector from a deviation between the detector signal and the true value output from the estimating means, the input means receiving an output of the soundness evaluating means Switching means for switching the estimating means for receiving the detector signal from the output means for outputting the estimation result of the true value of the detector signal and the soundness evaluation result, the estimating means, the soundness evaluating means and the switching The means forms a true value estimator, and the soundness evaluation means calculates a moving average value of the deviation and evaluates the soundness. Maintenance equipment. 2. The on-line maintenance device for an in-plant detector according to claim 1, wherein said soundness evaluating means evaluates said soundness from an increasing rate of an integrated value of said deviation instead of a moving average value of said deviation. . 3. The on-line maintenance device for an in-plant detector according to claim 1, wherein a plurality of the true value estimating units are arranged in parallel between the input means and the output means.