JP4480019B2 - Refinery monitoring system - Google Patents

Description

  The present invention relates to a refinery monitoring system for diagnosing the operating state of an oil refinery plant, and in particular, using an Mahalanobis distance obtained from operation data including process data and acoustic and vibration data of the oil refinery plant, the oil refinery plant. The present invention relates to a refinery monitoring system for diagnosing the operating state of a plant.

  Generally, in an oil refining plant, petroleum refining is performed using a number of equipment such as a distillation column, a reactor, a heating furnace, a heat exchanger, a pump, and a compressor. These devices are equipped with thermometers, pressure gauges, flow meters, etc. Based on the process data obtained from these instruments, the temperature, pressure, and flow rate of fluids (petroleum and its refined products) in the oil refining process Etc. are automatically controlled.

  By the way, conventionally, as a method for automatically diagnosing the state of a device, a distance index representing how far the state of the device at a certain time is from a population configured in a steady state is used to check the soundness of the device. Methods for evaluation are known (see, for example, Patent Documents 1 and 2).

  The diagnosis method disclosed in Patent Document 1 includes a step of obtaining a Mahalanobis space from a steady state of a diagnosis object, a step of obtaining a Mahalanobis distance (MD) by extracting a feature amount from the object to be diagnosed, and a preset threshold value And MD, a step of specifying a distance element value when MD is larger than a threshold value, and a step of replacing a feature amount corresponding to the specified distance element value with an average value of feature amounts of reference data And a step of newly obtaining an MD using the replaced feature amount, and repeating the above steps until the newly obtained MD is equal to or less than a threshold value.

  Further, the diagnostic device disclosed in Patent Document 2 detects sound generated from a diagnosis target device using a microphone, converts the sound into a digital signal, calculates Mahalanobis distance, and diagnoses whether Mahalanobis distance is within a determination reference value. It diagnoses the soundness of the target device.

ここで、式（１）において、xはある時刻における診断対象の状態を表す各種測定値を要素としたｎ次元行ベクトル、μは定常状態で構成される母集団の測定値毎の平均値を要素としたベクトル、Sは定常状態で構成される母集団の共分散行列である。また、一般的な距離指標としては、マハラノビス距離の他にユークリッド距離が挙げられる。 The diagnostic method disclosed in Patent Document 1 and the diagnostic device disclosed in Patent Document 2 both use Mahalanobis distance as a distance index. The Mahalanobis distance is given by the following formula (1).

Here, in Equation (1), x is an n-dimensional row vector having various measurement values representing the state of the diagnosis target at a certain time as an element, and μ is an average value for each measurement value of the population configured in a steady state. An element vector, S, is a covariance matrix of a population composed of steady states. Further, as a general distance index, Euclidean distance can be cited in addition to Mahalanobis distance.

JP 2004-227279 A JP2004-340706

  Various measurement values used for the distance index usually vary (variance) for each measurement value, and the steady state constituting the population has a biased distribution. The Euclidean distance is just a geometrical distance in calculating the distance to the population that has a bias in the steady-state distribution, but the Mahalanobis distance is normalized by variance, that is, all distributions Correlation between measurements is taken into account. As described above, a large number of devices are used in an oil refinery plant, and there are various measured values representing the state of these devices. The inventors have found that the Mahalanobis distance, in which the correlation between all measured values is taken into account, is particularly useful in diagnosing the operating state of an oil refinery plant. In addition, there has been no confirmed case where the Mahalanobis distance was applied to the diagnosis of the operating state of an oil refinery plant.

  However, when diagnosing the state of a device using the Mahalanobis distance, it is necessary to fix a threshold for determining whether the state of the device is normal, as described in Patent Documents 1 and 2 above. There is. Here, in the oil refining plant, for example, the processing amount and the oil type may change, and since the oil refining process is various, for determining whether the operation state of the oil refining plant is normal or not. It is difficult to fix the threshold. In other words, if the Mahalanobis distance is simply applied to the diagnosis of the operating state of the oil refinery plant, for example, even if the value is larger than the threshold value, it may actually be normal and vice versa. There is a possibility that the reliability of diagnosis cannot be sufficiently secured.

  There is also a demand for a system in which an abnormal event that has not occurred conventionally is fed back to future diagnosis.

  The present invention has been made to solve the above-described problems, and an object of the present invention is to provide a refinery monitoring system capable of automatically and appropriately diagnosing the operating state of an oil refinery plant.

The object is achieved by the refinery monitoring system according to the present invention described in the following (1) to ( 4 ).

(1) A refinery monitoring system comprising the following steps (a) to (d) and diagnosing the operating state of an oil refinery plant.
(A) a data collection step for collecting operation data including fluid process data in the oil refinery process of the oil refinery plant and acoustic and vibration data of the equipment of the oil refinery plant; (b) obtained in the data collection step Mahalanobis calculation step for obtaining a Mahalanobis distance in consideration of the correlation between the process data of the operation data and the acoustic and vibration data. (C) From the normal operation database in which the operation data in the normal operation state of the oil refinery plant is stored, A data mining step for extracting operation data similar to the operation data obtained in the data collection step and obtaining a Mahalanobis distance in consideration of the correlation between the process data of the extracted operation data and the acoustic and vibration data. (D) The Mahalanobis calculation Mahalanobis requested in steps Operating state determination that calculates the difference between the separation and the Mahalanobis distance obtained in the data mining step, and compares the difference with a preset threshold value to determine whether the operating state of the oil refinery plant is normal Step

  (2) Collecting the data from the abnormal operation database in which the operation data in the abnormal operation state of the oil refinery plant and abnormal events related to the operation data are accumulated when it is determined as abnormal in the operation state determination step The refinery monitoring according to (1), further comprising an alarm issuing step for extracting operation data similar to the operation data obtained in the step and notifying an abnormal event related to the extracted operation data system.

(3) A refinery monitoring system comprising the following steps (a) to (e) and diagnosing the operating state of an oil refinery plant.
(A) a data collection step for collecting operational data including fluid process data in the oil refinery process of the oil refinery plant and acoustic and vibration data of the equipment of the oil refinery plant;
(B) Mahalanobis calculation step for obtaining the Mahalanobis distance in consideration of the correlation between the process data of the operation data obtained in the data collection step and the acoustic and vibration data.
(C) From the normal operation database in which the operation data in the normal operation state of the oil refinery plant is accumulated, operation data similar to the operation data obtained in the data collection step is extracted, and the extracted operation data Data mining step for obtaining Mahalanobis distance considering the correlation between process data and acoustic and vibration data
(D) calculating a difference between the Mahalanobis distance obtained in the Mahalanobis calculation step and the Mahalanobis distance obtained in the data mining step, and comparing the difference with a preset threshold value to operate the oil refinery plant. Operation state determination step for determining whether the state is normal
(E) Collecting the data from the abnormal operation database in which the operation data in the abnormal operation state of the oil refinery plant and abnormal events related to the operation data are accumulated when it is determined as abnormal in the operation state determination step Alarm issuing step that extracts operation data similar to the operation data obtained in the step and notifies the abnormal event related to the extracted operation data and the probability of occurrence of the abnormal event represented by the following formula
Accuracy [%] = 100 × (similarity threshold−similarity) / similarity threshold ( 4 ) Driving data obtained in the data collection step when it is determined to be abnormal in the driving state determination step, and The refined monitoring system according to (2) or (3) above, wherein the abnormal events that have occurred are stored in the abnormal operation database.

According to the present invention, the Mahalanobis distance is obtained from the operation data including the collected oil refining plant process data and acoustic and vibration data, and the operation state of the plant is diagnosed using the Mahalanobis distance. Thereby, it is possible to accurately diagnose the operating states of various plants.
The Mahalanobis distance takes into account the correlation between the acoustic and vibration data and the process data. Thereby, in the diagnosis of the operation state of an oil refinery plant that needs to process a wide variety of data, the reliability of the diagnosis can be sufficiently ensured.
Furthermore, the difference between the Mahalanobis distance obtained from the collected operation data and the Mahalanobis distance obtained from similar operation data stored in the normal operation database is obtained, and the difference is compared with the threshold value. Even if the threshold value used for determining whether the operating state is normal or not is fixed, there is no possibility that the reliability of the diagnosis is lowered, and a diagnosis reflecting the actual situation of various oil refining processes can be performed.
Thus, according to the present invention, it is possible to automatically and appropriately diagnose the operating state of the oil refinery plant.

  Hereinafter, an embodiment of a refinery monitoring system according to the present invention will be described with reference to the drawings.

  As shown in FIG. 1, the refinery monitoring system of the present embodiment is obtained in a data collection step S1 for collecting operation data including process data and acoustic and vibration data of an oil refinery plant, and the data collection step S1. Operation data similar to the operation data obtained in the data collection step from the Mahalanobis calculation step S2 for obtaining the Mahalanobis distance of the operation data and the normal operation database in which the operation data in the normal operation state of the oil refinery plant is accumulated. Data mining step S3 for obtaining the Mahalanobis distance of the extracted driving data, and calculating the difference between the Mahalanobis distance obtained in the Mahalanobis calculation step and the Mahalanobis distance obtained in the data mining step, And the preset threshold value Comprises a driving state determination step S4 determines whether the normal the operating condition of the oil refining plant.

  When the refinery monitoring system of this embodiment is determined to be abnormal in the operation state determination step S4, the operation data in the abnormal operation state of the oil refinery plant and the abnormal event related to the operation data are accumulated. The system further includes an alarm issuing step S5 for extracting operation data similar to the operation data obtained in the data collection step S1 from the abnormal operation database 4 and notifying an abnormal event related to the extracted operation data.

<Data collection step S1>
The process data is obtained from a DCS (Distribution Control System) 1 already installed in the oil refinery plant.

  DCS does not control the entire plant with a single control device, but distributes the control devices in several control loops, thereby causing the risk of simultaneous shutdown of all plants due to control device failure. Process control system that collects and manages process data obtained from thermometers, pressure gauges, flow meters, power meters, etc. installed in plant equipment. Various process data can be used as the process data in the present invention, and among them, temperature, pressure, flow rate, electric power and the like are preferably used.

  Sound and vibration data are obtained from sensors 2 installed in plant equipment.

  Diagnosis by sound is state-based maintenance, and enables comprehensive, three-dimensional, and comprehensive diagnosis from the characteristic that sound propagates not only in solid but also in the air. This feature enables stable operation of the plant in that a wide range / site can be diagnosed with a small number of sensors, and the need for precise diagnosis can be objectively determined before making a precise diagnosis locally. Contributes to optimizing the balance between maintenance costs. In plant equipment, for example, bearings of rotating equipment such as motors and pumps, various dynamic sounds from the dynamic machinery and static machinery to the audible range (20 Hz to 20 kHz) and the ultrasonic range (20 kHz to) as abnormalities occur. Will occur. Further, it is known that an AE signal exceeding 100 kHz is generated upon microscopic destruction. For example, a parabolic sound collector can be used as a sensor for collecting such sound. When collecting sound, it is necessary to collect the sound of the target device and not to pick up the sound of other devices as much as possible, and the parabolic sound collector has directivity, so it is suitable as a sensor for collecting sound. .

  Diagnosis by vibration is state-based maintenance, and precise diagnosis is possible as compared with acoustic diagnosis by local measurement of the sensor installation position. There are two types of sensors that collect vibrations: contact type using a piezoelectric element and non-contact type such as laser Doppler type. In consideration of installation, 1) the sensor itself can output signals, 2) only dynamic phenomena can be measured, 3) wide dynamic range, 4) easy to make robust structures, 5) frequency range It is preferable to use a contact-type sensor using a piezoelectric element that has the advantage that it can be operated with a wide 6) two-wire system.

  In the data collection step S1, the period for taking operation data including process data and sound and vibration data into the system can be arbitrarily selected from 1 minute to 60 minutes.

ここで、式（２）において、xはある時刻における運転データであり、温度、圧力、流量、電力、音響、振動、等の測定値を要素としたｎ次元行ベクトルである。また、μは、定常状態で構成される母集団の複数の運転データについて測定値毎の平均値を要素としたベクトルである。また、Sは、定常状態で構成される母集団の複数の運転データの共分散行列である。尚、母集団を構成する定常状態での複数の運転データは、このマハラノビス演算ステップＳ２に先立って収集されている。 <Mahalanobis calculation step S2>
Next, the Mahalanobis distance of the operation data collected in the data collection step S1 is obtained. In the Mahalanobis calculation step S2, the Mahalanobis distance is calculated by the following equation (2).

Here, in Expression (2), x is operation data at a certain time, and is an n-dimensional row vector having measured values of temperature, pressure, flow rate, power, sound, vibration, and the like as elements. Further, μ is a vector having an average value for each measurement value as an element for a plurality of operation data of a population configured in a steady state. S is a covariance matrix of a plurality of operation data of a population configured in a steady state. A plurality of operation data in a steady state constituting the population is collected prior to this Mahalanobis calculation step S2.

An example of when arranged in time series the Mahalanobis distance D ² calculated using the above equation (2) shown in FIG. As shown in FIG. 2, the Mahalanobis distance D ^2, in the case of a state close to the population showed a value close to 0, in the case of distant state from the population a larger value. For this reason, it is possible to express the change of a state sensitively.

<Data mining step S3>
In parallel with the Mahalanobis calculation step S2, the operation data similar to the operation data obtained in the data collection step S1 is extracted from the normal operation database 3 in which the operation data in the normal operation state of the plant is accumulated. Find the Mahalanobis distance of driving data.

  In the normal operation database 3, for example, operation data when the plant is operating normally is stored as average data for one hour or more. The more operation data in the normal operation state stored in the normal operation database 3, the more appropriate diagnosis can be made for various operation states.

  In the data mining step S3, the Mahalanobis distance of the extracted driving data may be calculated each time using the above equation (2), but by modeling, the driving data as input and the Mahalanobis distance as output are calculated. It is also possible to create an inter-function and estimate the output from any input. Among such modeling methods, neural networks, GMDH (Group Method of Data Handling) are particularly applicable to non-linear inputs and outputs such as driving data and Mahalanobis distance. PLS (Partial least squares regression), TCBM (Topological Case Base Modeling), and the like.

<Operation state determination step S4>
Next, a difference between the Mahalanobis distance obtained in the Mahalanobis calculation step S2 and the Mahalanobis distance obtained in the data mining step S3 is calculated, and the difference is compared with a preset threshold value to determine whether the operation state of the oil refinery plant is normal. Determine whether or not.

  If the operation state of the plant is normal when the operation data is collected in the data collection step S1, there is no difference between the Mahalanobis distance obtained in the Mahalanobis calculation step S2 and the Mahalanobis distance obtained in the data mining step S3. If not normal or slight, but not normal, a deviation occurs between the two. Therefore, by setting an appropriate threshold value (fixed value) for the difference between the two, it can be determined whether or not the operation state of the plant is normal. For example, a value of about 50 is preferably set as the threshold value.

<Alarm announcement step S5>
When it is determined that there is an abnormality in the operation state determination step S4, it is obtained in the data collection step S1 from the abnormal operation database 4 in which the operation data in the abnormal operation state of the oil refinery plant and the abnormal events related to the operation data are accumulated. The operation data similar to the operation data is extracted, and the abnormal event related to the extracted operation data is notified.

  The abnormal operation database 4 stores operation data when an abnormality is confirmed in the plant and abnormal events at that time. As the operation data and abnormal events in the abnormal operation state stored in the abnormal operation database 4 are increased, the abnormal event can be appropriately predicted for various operation states.

ここで、抽出された運転データの類似度は、図３に示すように、量子化された空間で以下のように定義する。即ち、抽出された運転データがデータ収集ステップＳ１で収集した運転データ（図３において符号Ａで示す）と同一のメッシュに入る場合に当該運転データの類似度は０、隣接するメッシュに入る場合に当該運転データの類似度は１、という具合に定義する。 In the refinery monitoring system of the present embodiment, the abnormal event related to the extracted operation data is notified together with the accuracy representing the possibility that the abnormal event will occur. The accuracy is calculated from the following equation (3).

Here, the similarity of the extracted operation data is defined as follows in a quantized space as shown in FIG. That is, when the extracted operation data enters the same mesh as the operation data (indicated by symbol A in FIG. 3) collected in the data collection step S1, the similarity of the operation data is 0, and when it enters an adjacent mesh The similarity of the operation data is defined as 1, for example.

  The notification means is not particularly limited, and for example, notification is displayed on a monitor.

  In the refinery monitoring system of the present embodiment, the operation data obtained in the data collection step S1 and the abnormal event that has occurred are accumulated in the abnormal operation database 4 when it is determined that there is an abnormality in the operation state determination step S4. . As a result, various events are accumulated, and abnormal events can be predicted with high accuracy.

As described above, according to the refinery monitoring system of the present embodiment, the Mahalanobis distance is obtained from the operation data including the collected process data and the acoustic and vibration data of the oil refinery plant, and the Mahalanobis distance is used. Diagnose plant operating conditions. Thereby, it is possible to accurately diagnose the operating states of various plants.
The Mahalanobis distance takes into account the correlation between the acoustic and vibration data and the process data. Thereby, in the diagnosis of the operation state of an oil refinery plant that needs to process a wide variety of data, the reliability of the diagnosis can be sufficiently ensured.
Further, since the difference between the Mahalanobis distance obtained from the collected operation data and the Mahalanobis distance obtained from similar operation data stored in the normal operation database 3 is obtained and this difference is compared with a threshold value, the plant Even if the threshold value used for determining whether or not the operation state is normal is fixed, there is no possibility that the reliability of the diagnosis is lowered, and a diagnosis reflecting the actual situation of various oil refining processes can be performed.
Furthermore, when the operation state of the plant is determined to be abnormal, the collected operation data and the abnormal events that have occurred are stored in the abnormal operation database 4, so that the events can be fed back to future diagnosis. It is possible to properly predict an abnormal event even in a difficult operating state.

  As an embodiment of the present invention, an example applied directly to a desulfurization apparatus in an oil refinery plant will be described.

  As shown in FIG. 4, a plurality of sensors 2 are installed in the make-up compressor 6, the recycle compressor 7, the high-pressure pump 8, and the hot oil pump 9 of the direct desulfurization apparatus 5.

  In the present embodiment, the DC desulfurization apparatus 5 is monitored by the following procedures 1 to 5.

Procedure 1. Operation data is collected from the DCS 1 and the plurality of sensors 2 at 1-minute intervals.
Procedure 2. The Mahalanobis distance M1 of the operation data is calculated from the operation data collected in the procedure 1 using the above equation (2).
Procedure 3. Using commercial data mining software, the operation data similar to the collected operation data is extracted from the normal operation database 3 in which the operation data in the normal operation state of the plant is accumulated, and the Mahalanobis distance M2 of the extracted operation data is calculated. Ask.
Procedure 4. A difference between the Mahalanobis distance M1 obtained in the procedure 2 and the Mahalanobis distance M2 obtained in the procedure 3 is calculated, and the difference is compared with a threshold value.
Procedure 5. When the difference is larger than the threshold value in the procedure 4, the operation data in the abnormal operation state of the plant and the abnormal operation database 4 in which abnormal events related to the operation data are accumulated are similar to the operation data collected in the procedure 1. The operation data is extracted, an abnormal event related to the extracted operation data is displayed on the monitor, and an alarm is issued.

  Then, as a model at the time of occurrence of abnormality, abnormal data is input to the acoustic data after a predetermined time has elapsed. The results are shown in FIG. As shown in FIG. 5, the two Mahalanobis distances M1 and M2 obtained in steps 2 and 3 are almost the same before abnormal data is input to the acoustic data (normal), but the acoustic data is abnormal. It can be seen that there is a large difference between the two after the correct data is input.

  After the abnormal data is input to the acoustic data, as shown in FIG. 6, the risk of the current state from the past case on the monitor by the procedure 5 (that is, the probability that the abnormal event that occurred in the past will occur) on the monitor. Is displayed as a numerical value, and the operation state of the plant can be grasped.

It is a flowchart which shows the structure of one Embodiment of the refinery monitoring system which concerns on this invention. It is a graph which shows an example at the time of arranging the Mahalanobis distance of driving data in time series. It is a conceptual diagram for demonstrating the similarity of driving data and the extracted similar driving data. It is a block diagram of the direct desulfurization apparatus to which the refinery monitoring system according to the present invention is applied. FIG. 5 is a graph in which Mahalanobis distances of operation data collected in the direct desulfurization apparatus shown in FIG. 4 and Mahalanobis distances of similar extracted operation data are arranged in time series. It is a figure which shows an example of warning issuing.

Explanation of symbols

1 DCS
2 Sensor 3 Normal operation database 4 Abnormal operation database S1 Data collection step S2 Mahalanobis calculation step S3 Data mining step S4 Operation state determination step S5 Alarm issuing step

Claims (4)

A refinery monitoring system comprising the following steps (a) to (d) and diagnosing the operating state of an oil refinery plant.
(A) a data collection step for collecting operation data including fluid process data in the oil refinery process of the oil refinery plant and acoustic and vibration data of the equipment of the oil refinery plant; (b) obtained in the data collection step Mahalanobis calculation step for obtaining a Mahalanobis distance in consideration of the correlation between the process data of the operation data and the acoustic and vibration data. (C) From the normal operation database in which the operation data in the normal operation state of the oil refinery plant is stored, A data mining step for extracting operation data similar to the operation data obtained in the data collection step and obtaining a Mahalanobis distance in consideration of the correlation between the process data of the extracted operation data and the acoustic and vibration data. (D) The Mahalanobis calculation Mahalanobis requested in steps Operating state determination that calculates the difference between the separation and the Mahalanobis distance obtained in the data mining step, and compares the difference with a preset threshold value to determine whether the operating state of the oil refinery plant is normal Step   When it is determined that the operation state is determined to be abnormal, the data collection step obtains from the abnormal operation database in which the operation data in the abnormal operation state of the oil refinery plant and abnormal events related to the operation data are accumulated. The refinery monitoring system according to claim 1, further comprising an alarm issuing step of extracting operation data similar to the extracted operation data and notifying an abnormal event related to the extracted operation data.   A refinery monitoring system comprising the following steps (a) to (e) and diagnosing the operating state of an oil refinery plant.
  (A) a data collection step for collecting operational data including fluid process data in the oil refinery process of the oil refinery plant and acoustic and vibration data of the equipment of the oil refinery plant;
  (B) Mahalanobis calculation step for obtaining the Mahalanobis distance in consideration of the correlation between the process data of the operation data obtained in the data collection step and the acoustic and vibration data.
  (C) From the normal operation database in which the operation data in the normal operation state of the oil refinery plant is accumulated, operation data similar to the operation data obtained in the data collection step is extracted, and the extracted operation data Data mining step to obtain Mahalanobis distance considering the correlation between process data and acoustic and vibration data
  (D) calculating the difference between the Mahalanobis distance determined in the Mahalanobis calculation step and the Mahalanobis distance determined in the data mining step, and comparing the difference with a preset threshold value to operate the oil refinery plant Operation state determination step for determining whether the state is normal
  (E) Collecting the data from the abnormal operation database in which the operation data in the abnormal operation state of the oil refinery plant and abnormal events related to the operation data are accumulated when it is determined as abnormal in the operation state determination step Alarm issuing step that extracts operation data similar to the operation data obtained in the step and notifies the abnormal event related to the extracted operation data and the probability of occurrence of the abnormal event represented by the following formula
  Accuracy [%] = 100 × (similarity threshold−similarity) / similarity threshold If it is determined that the abnormality in the operation state determination step, the operation data obtained by the data collection step, and to claim 2 or 3 generated abnormal event is characterized in that it is stored in the abnormal operation database The refinery monitoring system described.

Images