JP2006120049A - Plant failure prediction method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform a failure prediction based on a logic of a failure prediction system that can perform a simulation including a control system, has few setting items, and has high versatility.
SOLUTION: After compressing air with a compressor 11, carbon dioxide and moisture are removed by an MS adsorber 12. The compressed air is cooled and liquefied by the main heat exchanger 13, separated into nitrogen, oxygen, argon, and the like by the rectifying tower 14, and high-purity nitrogen is sent to the product recovery system 3. The control system is not only a control system connected by a control loop for each system but also a system identification by combining control systems that influence each other as a disturbance to form one ARX model. Even when the liquid nitrogen injection valve LV3501 of the distillation separation system 2 is clogged and liquid nitrogen cannot normally be injected into the rectification column 14, failure prediction can be easily performed.
[Selection] Figure 3

Description

  The present invention relates to a plant failure prediction method for predicting a plant failure based on a predetermined logic, and in particular, a chemical / chemical process for controlling a process by implementing a control system such as a valve of a piping system, a compression / expansion turbine, or a heater. The present invention relates to a failure prediction method for a plant for properly controlling steel and oil plants.

  In recent years, the complexity and specialization of chemical, steel, and petroleum plants are increasing. On the other hand, however, the number of operators is declining in the operating field due to reasons such as reduced operating costs. In such an environment, it is very difficult to monitor a complex and non-generalized plant with a small number of people, and plant abnormalities will surface for the first time after a plant stop alarm is issued when an abnormality occurs. There are many cases of doing. Therefore, a failure prediction system that detects an abnormality in advance by performing simulation of plant failure prediction in real time is very useful.

  Various plant failure prediction systems have been reported in the past. For example, for large-scale refrigeration equipment, a system that predicts individual control variables using a process simulator, or chemical parameters that are expected to fluctuate due to plant abnormalities are predicted from plant management data collected in advance. A method of diagnosing an abnormal event of a plant by comparison with fluctuation is known (for example, see Non-Patent Document 1). A technique for diagnosing an abnormal event in a plan by collecting data related to chemical management of a power plant and comparing it with a predetermined evaluation standard is also disclosed (for example, see Patent Document 1). However, in the system that predicts these control variables, fine adjustment of control constants such as investigation of all parameters such as piping and valve CV values, which are the prediction target devices, and collection of plant management data is performed. There is a need to. Therefore, such a failure prediction system cannot be applied universally to plants having different characteristics.

Therefore, as a method applied to a general-purpose failure prediction system, model identification (or system identification) is performed based on normal input / output data, and the characteristic model parameters in the reference plant and the characteristic model parameters in the operational plant are determined. In comparison, a method for diagnosing plant abnormality is disclosed (for example, see Patent Document 2). However, since the method disclosed in Patent Document 2 is intended only for abnormality diagnosis, detailed setting is necessary for each model, and there is still a problem in terms of versatility. Therefore, as a general-purpose failure prediction method that can be applied to individual models, process control response tendencies are stored, and anomalies are detected by comparing the relationship between response time and response tendencies. Is disclosed (for example, see Patent Document 3).
Maekawa R., et.al., Simulation system for a large scale cryogenic facility at NIFS, Proceedings of the 19th International Cryogenic Engineering Conference, Narosa Pubilishing House, 691-694 (2002) JP 2003-122429 A JP 2003-303020 A Japanese Patent Laid-Open No. 5-40517

  However, although the failure prediction system disclosed in Patent Document 3 has the advantage of high versatility, the failure diagnosis logic is simple and the characteristics of each control system in a complex plant are not fully considered. It can only be applied to some failures in the plant process. That is, in the above-described conventional failure prediction system, if the control system is considered sufficiently, the trend characteristics differ from plant to plant. Therefore, it cannot be applied to dissimilar plants such as chemical / steel / petroleum plants. On the other hand, if a simple logic is sought for versatility, there is a problem that it can be applied only to a part of the plant failure.

  The present invention has been made in view of the problems as described above, and can be simulated including a control system, has few setting items, and is based on the logic of a highly versatile failure prediction system. An object of the present invention is to provide a plant failure prediction method for predicting a failure of a plant.

  A plant failure prediction method according to the present invention was devised in order to achieve the above-described object, and is a plant failure prediction method for automatically predicting a failure of a plant that performs process control, and has a mutual influence. The input / output system between control systems that affect each other is a multi-input / multi-output type, and based on the mathematical model that identifies the system identification procedure to obtain the transfer function using the least squares method using the measured plant data The procedure for simulating the two actual measurement data of the control variable of minutes before and the manipulated variable from several minutes ago to the present time to calculate the current estimated value, and the difference between the actual measured data and the estimated value over several hours And a procedure for issuing a failure prediction alarm when the integrated value of the difference exceeds a preset threshold value.

  According to such a procedure, system identification is performed by using the measured data of the plant as a multi-input multi-output type of control systems that influence each other. Then, based on the mathematical model identified by the system, a current simulation value is calculated from two actually measured data of a control variable several minutes before the current operation and an operation variable from several minutes before to the present. Further, the difference between the actually measured data and the calculated simulation value is integrated over several hours, and an alarm is issued when the integrated value of the difference exceeds a preset threshold value. As a result, it is not necessary to perform control setting for each process or to perform individual control, so that failure prediction can be easily performed even for complicated plant processes.

  Further, in the plant failure prediction method according to the present invention, the control systems are a combination control system in which a control system connected by a control loop and a control system that causes disturbance to each other are combined. It is characterized by creating a multi-input multi-output mathematical model by a combination control system. As a result, it is possible to predict the failure of a plant having processes that cause disturbances to each other, and therefore it is possible to perform a more complicated failure prediction of a plant.

  In the plant failure prediction method according to the present invention, the system-identified mathematical model is an ARX (Auto-Regressive eXogenous) model. That is, since the ARX model is the simplest mathematical model in system identification, it can be easily modeled even with complex plant control variables and operation variables.

  According to the plant failure prediction method according to the present invention, by making the logic into an automatic system, the user only sets a control system that affects each other as a disturbance with the control system connected by the control loop when the system is introduced. Thus, an accurate failure prediction simulation can be performed. Then, by comparing the integral value of the difference between the actual measurement value and the simulation value for several hours with a predetermined threshold value, it is possible to easily perform a prior prediction of a failure in consideration of the process control system. As a result, it is not necessary to perform control settings for each process or to perform individual control, so that it is possible to easily predict failure of a complex plant such as a chemical / steel / petroleum plant.

<Outline of the invention>
According to the plant failure prediction method of the present invention, in a plant that performs process control, control systems that influence each other are set as a multi-input multi-output type, and system identification is performed using measured plant data. Then, based on the mathematical model identified by the system, a current simulation value is calculated from two actually measured values of “control variable several minutes ago” and “manipulated variable from several minutes ago to the present time”. In addition, it integrates the difference between the measured value for several hours and the current simulation value, and predicts a plant failure in advance by issuing an alarm when the integrated value of the difference exceeds a predetermined threshold. Features. In the system identification, failure prediction is performed in advance using a multi-input multi-output mathematical model that combines a control system connected by a control loop and a control system that causes disturbance.

  The system identification is to create a mathematical model of an object from input / output time series data in a system to be controlled, and is widely used in the field of process control. Since the simplest mathematical model in this system identification is the ARX model, the present invention realizes a plant failure prediction method using this ARX model.

<Embodiment>
Hereinafter, embodiments of a plant failure prediction method according to the present invention will be described in detail with reference to the drawings. Usually, in the control and monitoring system of an air cryogenic separator that separates and purifies nitrogen and the like from air, the measured values of the plant during the past operation of the air cryogenic separator, for example, valve opening, temperature, pressure, gas flow rate, etc. And numerical values such as the liquid level are input to the mathematical model for system identification as operation variables. That is, system identification is performed by the least square method based on these actually measured values, and a coefficient of a polynomial transfer function is obtained. This identification model uses an ARX model as shown in equation (1) based on the equation error model.
A (q) y (k) = B (q) u (k) + w (k) (1)

  An ARX model like the above equation (1) is obtained as follows. That is, when equation (2) is defined as a parameter vector and equation (4) is defined as a parameter vector in equation (2) where the disturbance term is white noise w (k) in the difference equation, output y (k) is expressed by equation (2). It can be expressed as (5).

_{y (k) + a 1 y} (k-1) + ... + a na y (k-n a)
= B ₁ u (k−1) +... + B _nb u (k−n _b ) + w (k) (2)

θ = [a ₁ ,..., a _na , b ₁ ,..., b _nb ] ^T (3)
φ (k) = [− y (k−1),..., −y (k−n _a ), u (k−1),... u (k−n _b )] ^T (4)
y (k) = θ ^T φ (k) + w (k) (5)

Here, the following two expressions that are irreducible polynomials of the shift operator q are introduced into the above expression (4).
A (q) = 1 + a ₁ q ⁻¹ +... + A _na q ^−na
B (q) = b ₁ q ⁻¹ +... + B _nb q ^−nb

Then, equation (2) can be rewritten as equation (1). This equation (1) is an ARX model, which is a mathematical model often used in system identification. As in the conceptual diagram of the multi-input / multi-output mathematical model shown in FIG. 1, the ARX model as shown in Equation (1) is inputted with the manipulated variable u (k) and added with the disturbance w (k) to control variable y. It is a model of a control system that outputs (k). Then, past output data of a finite number current output y a (k) - the failure prediction by associating the data u _{(k-n b)} - data y _{(k-n a)} and input data.

Each element shown in FIG. 1 is defined as follows.
A (q): polynomial (1 + a ₁ q ⁻¹ + a ₂ q ⁻² +...)
B (q): polynomial (b ₁ q ⁻¹ + b ₂ q ⁻² +...)
y (k): Control variable (for example, pressure, temperature, etc.)
u (k): manipulated variable (for example, valve opening, etc.)
w (k): Noise (for example, noise)

In the simulation using the ARX model as shown in FIG. 1, as shown below, the current control variable y (k) is used to predict the current value using data from the previous m stages.
The control variable before m-1 stage is
y (k−m + 1) = B (q) u (k−m + 1) + [1−A (q)] y (k−m + 1)
The control variable before m-2 stage is
y (k−m + 2) = B (q) u (k−m + 2) + [1−A (q)] y (k−m + 2)
And so on,
The current control variable is
y (k) = B (q) u (k) + [1-A (q)] y (k)
Asking.

  The input / output system of the ARX model for system identification is not only a control system connected by a control loop but also a multi-input multi-output type by combining control systems that influence each other as a disturbance. Further, the order (power number) of the transfer function in the ARX model for system identification is determined so that the loss function is minimized. Using the ARX model identified in this way, the current simulation value is calculated from two actually measured values of “control variable several minutes ago” and “current operation variable several minutes ago”. This method employs a “control variable several minutes ago” as the initial value of the simulation, and sequentially calculates the control variable for the next stage time from the operation variable at each time and the ARX model. Such failure prediction is performed by integrating the difference between the actual measurement value and the simulation value over several hours and setting so as to issue an alarm when the integration value exceeds a predetermined threshold value.

<Example>
FIG. 2 shows a conceptual diagram in the case of performing plant failure prediction using the ARX model as a mathematical model for system identification.
The database 21 stores past data of the plant. The system identification unit 22 reads the data in the database 21, determines the transfer function 23, and outputs it. The transfer function 23 once determined is held in a memory or a transfer function database or the like until a new system identification is performed.
Actual plant data 26 (operation variables and control variables) is collected from the plant 25 to be predicted for failure.
The simulation unit 24 simulates the current control variable estimated value using the transfer function 23 and the actual plant data 26. The value simulated in the simulation unit 24 is held as the current control variable estimated value 27. The stored value is stored in the memory or the transfer function database until the next simulation.
An integral value of a difference between the current control variable estimated value 27 and the actual plant data 26 is sequentially calculated in the abnormality determination unit 28, and an alarm is issued when the value exceeds a certain value.
Next, an example in the case of performing failure prediction of the chilled air separation device will be described. FIG. 3 is a system diagram of a cryogenic air separation apparatus to which a plant abnormality diagnosis system is applied as an embodiment of the present invention. The control system of the cryogenic air separation apparatus is divided into a raw material air purification system 1, a distillation separation system 2, and a product recovery system 3. The raw material air purification system 1 is mainly composed of a compressor 11 that compresses air and an MS adsorber 12 that removes carbon dioxide and moisture contained in the compressed air. The distillation separation system 2 is composed of a main heat exchanger 13 that cools and liquefies air and a rectifying column 14 that separates nitrogen, oxygen, argon, and the like from the cooling air. The product recovery system 3 is composed only of piping and valves.

  The chilled air separation apparatus shown in FIG. 3 removes carbon dioxide and moisture solidified at a low temperature by the MS adsorber 12 after compressing air as a raw material by the compressor 11. Further, after the compressed air is cooled and liquefied by the main heat exchanger 13, it is separated into nitrogen, oxygen, argon and the like by the rectifying column 14. In this way, high-purity nitrogen is purified from the raw air as a product by the rectification column 14 and taken out to the product recovery system 3. At this time, the control system forms not only a control system connected by a control loop for each system but also a combination of control systems that influence each other as a disturbance to form one ARX model.

  FIG. 4 is a diagram showing an ARX model creation table in the raw material air purification system 1 of FIG. That is, in the raw material air purification system 1, one ARX model is created by a combination of the operation variable u (k) and the control variable y (k) as shown in the table of FIG. For example, when the raw material air pressure adjustment valve PV1791 of the raw material air purification system 1 is set as the operation variable u (k), the raw material air pressure PT1791 is combined as the control variable y (k), and the precooler TV2401 is set as the operation variable u (k). When the input temperature TE2401 of the MS adsorber 12 is combined as a control variable y (k), and when the aftercooler TV 2403 is an operation variable u (k), the outlet temperature TE2403 of the MS adsorber 12 is controlled by the control variable y (k). ).

  FIG. 5 is a diagram showing an ARX model creation table in the distillation separation system 2 of FIG. In the distillation separation system 2, one ARX model is created by a combination of the operation variable u (k) and the control variable y (k) as shown in the table of FIG. For example, when the liquefied nitrogen injection valve LV3501 of the distillation separation system 2 is the operation variable u (k), the liquid level LT3501 of the rectification column 14 is combined as the control variable y (k), and the precooler FV2441 is the operation variable u (k). In this case, the liquid air flow rate FT2441 is combined as the control variable y (k).

  FIG. 6 is a diagram showing an ARX model creation table in the product collection system 3 of FIG. In the product collection system 3, one ARX model is created by combining the operation variable u (k) and the control variable y (k) as shown in the table of FIG. For example, when the product nitrogen flow rate adjustment valve FV3521 of the product recovery system 3 is set as the operation variable u (k), the product nitrogen flow rate FT3521 and the product nitrogen pressure PT3521 are combined as the control variable y (k). In this way, as many ARX models as combinations of multiple inputs and multiple outputs are created for the number of combinations of control variables to be monitored. At this time, the number of combinations of ARX models is not limited.

Hereinafter, a simulation operation as a result of system identification of the cryogenic air separation device shown in FIG. 1 using the ARX model created as described above will be described. FIG. 7 is a diagram showing an example of simulation based on the ARX model in which the system is identified based on the data collected in the distillation separation system 2 of the cryogenic air separation device shown in FIG. 3, and (a) is an operation variable u (k). Shows the relationship between the characteristic when the liquefied nitrogen injection valve LV3501 is operated (upper figure) and the characteristic when the liquid level LT3501 changes as the control variable y (k) (lower figure), and (b) shows the operational variable u (k ) Shows the relationship between the characteristic when the precooler FV 2441 is operated (upper figure) and the characteristic when the liquid air flow rate FT2441 changes as the control variable y (k) (lower figure). In all the characteristic diagrams, the horizontal axis represents time (h), the vertical axis represents the change rate (%) of the valve opening of the liquefied nitrogen injection valve LV3501 and the precooler FV2441, and the liquid level LT3501 represents the rectification column 14 It represents the liquid level position (mm), and the liquid air flow rate FT2441 represents the flow rate (Nm ³ / h).

  The data in FIG. 7 is the ARX model of the distillation separation system 2 shown in FIG. 5, and was obtained by system identification of the liquid level control and the air flow rate control of the rectification column 14 in the distillation separation system 2 with two inputs and two outputs. The simulation result is compared with the actual measurement value. In each characteristic, the actual measurement value is indicated by a solid line and the simulation result is indicated by a broken line.

The operation variable u (k) at this time is the opening degree of the liquefied nitrogen injection valve LV3501 and the opening degree (%) of the liquid air flow rate control valve FV2441 of the precooler in the distillation separation system 2, and the corresponding control variable y (k ) Are the height (mm) of the liquid level LT 3501 at the bottom of the rectifying column 14 and the flow rate value (Nm ³ / h) of the liquid air flow rate FT2441 obtained as a result of the respective opening degrees.

  That is, the upper diagram of FIG. 7A shows the change over time of the valve opening degree of the liquid nitrogen injection valve LV3501, and the valve opening is performed by feedback control based on the measured value of the liquid level of the rectifying column 14. The degree has changed by about 82-2%. Moreover, the lower figure of Fig.7 (a) shows the measured value (solid line) and simulation value (broken line) of the height of the liquid level LT3501 of the tower bottom part of the rectification column 14. FIG. As is clear from FIG. 7A, the simulation value (broken line) well represents the change in the liquid level corresponding to the change in the valve opening degree of the liquid nitrogen injection valve LV3501.

  The upper diagram of FIG. 7B shows the change over time of the valve opening of the liquid air flow rate adjustment valve FV2441 of the precooler. The valve opening changes by about 48 to 56% by feedback control based on the measured value of the liquid air flow rate. Further, the lower diagram of FIG. 7B shows an actual measurement value (solid line) and a simulation value (broken line) of the liquid air flow rate FT2441. As is clear from FIG. 6 (b9), the simulation value (broken line) well represents the change in the liquid air flow rate corresponding to the change in the valve opening degree of the liquid air flow rate adjustment valve FV2441.

  Next, an example in which the ARX model of the present invention is applied to a failure example in which liquid nitrogen injection valve LV3501 of distillation separation system 2 is clogged and liquid nitrogen cannot be normally injected into rectification column 14 will be described. FIG. 8 is a diagram illustrating an example of a simulation when a failure occurs in the cryogenic air separation device of FIG. 1. That is, FIG. 8A shows the time change of the opening degree of the liquid nitrogen injection valve LV3501 of the distillation separation system 2. FIG. FIG. 8B shows an actual measurement value (solid line) and a simulation value (broken line) of the time change of the liquid level LT 3501 at the bottom of the rectifying column 14. As shown in FIG. 8A, when the time is around 2 h, the liquid nitrogen injection valve LV3501 is clogged (that is, the opening degree of the liquid nitrogen injection valve LV3501 is 0%). If liquid nitrogen is not injected into the column 14, as shown in FIG. 8B, the measured value of the liquid level LT3501 gradually increases to the bottom of the rectifying column 14 as shown by the solid line. Going down. Therefore, the liquid nitrogen injection valve LV3501 holds the valve opening again at 82% in order to maintain the liquid level as shown in FIG.

  On the other hand, the simulation value of the liquid level LT3501 indicated by the broken line in FIG. 8B gradually deviates from the actual measurement value of the liquid level in the rectifying column 14 after the occurrence of a failure at a time near 2h. By setting an alarm for the integrated value of this difference, a plant abnormality can be detected at an early stage. For example, if there is a model capable of simulating the time change of the liquid nitrogen injection valve LV3501 or the liquid level LT3501 at the bottom of the rectifying column 14, it is possible to detect an abnormality by comparing these measured values with the simulation results. However, as described above, it is difficult to create a model applied to each actually measured value. On the other hand, if the fault prediction method according to the present invention is applied, it is not necessary to create a model specific to each actual measurement value by using the ARX model. If an ARX model is created by combining variables, abnormality detection of the entire plant can be easily performed.

  According to the plant failure prediction method of the present invention, an abnormal state of a plant can be easily predicted without performing individual setting and control by creating an ARX model by combining operation variables and control variables. Therefore, it can be used for failure prediction of heterogeneous plants and complex plants such as chemical, steel, and petroleum plants.

It is a conceptual diagram of the identification model of multiple input multiple output applied to this invention. It is a conceptual diagram of this invention in the case of performing failure prediction of a plant using an ARX model as a mathematical model of system identification. 1 is a system diagram of a cryogenic air separation device to which a plant abnormality diagnosis system is applied as an embodiment of the present invention. It is a figure which shows the creation table of the ARX model in the raw material air purification system | strain 1 of FIG. It is a figure which shows the creation table of the ARX model in the distillation separation system 2 of FIG. It is a figure which shows the creation table of the ARX model in the product collection system | strain 3 of FIG. It is a figure which shows an example of the simulation by the ARX model identified by the system identification based on the data collected by the distillation separation system 2 of the cryogenic air separation apparatus shown in FIG. It is a figure which shows an example of the simulation when a failure arises in the cryogenic air separation apparatus of FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Raw material air purification system 2 Distillation separation system 3 Product recovery system 11 Compressor 12 MS adsorber 13 Main heat exchanger 14 Rectification tower

Claims (3)

A plant failure prediction method for automatically predicting a failure of a plant performing process control,
The input / output system between the control systems that affect each other is a multi-input multi-output type, and the system identification to obtain the transfer function by the least square method using the measured data of the plant,
Based on the mathematical model identified by the system, a procedure for simulating two actual measurement data of a control variable several minutes ago and an operation variable several minutes before to the present time, and calculating a current estimated value;
A procedure for integrating the difference between the current measured data and the estimated value over several hours;
A failure prediction method for a plant, comprising: a step of issuing a failure prediction alarm when the integral value of the difference exceeds a preset threshold value. The control systems are a combination control system that combines a control system connected by a control loop and a control system that causes disturbance to each other,
2. The plant failure prediction method according to claim 1, wherein the system identification creates a multi-input multi-output mathematical model by the combination control system. The plant failure prediction method according to claim 1, wherein the mathematical model identified by the system is an ARX model.

Images