KR102700495B1 - Apparatus and method for valve stiction diagnosing using machine learning - Google Patents

Abstract

The present invention relates to a valve stiction diagnosis device using machine learning and a method thereof. According to an embodiment of the present invention, the valve stiction diagnosis device using machine learning includes: a stiction index feature extraction unit (110) for extracting a valve stiction index for a control system in the field as a feature for valve stiction diagnosis; an odd harmonic feature extraction unit (120) for extracting odd harmonic variables for the control system as features for valve stiction diagnosis; a learning model generation unit (150) for generating and providing a learning model using a plurality of machine learning techniques based on supervised learning for learning data generated by performing a simulation for cases where a valve stiction model is applied and cases where it is not applied by simulating the control system; and a valve stiction diagnosis unit (160) for diagnosing valve stiction for the control system by applying the valve stiction index and the odd harmonic variables to the learning model.

Description

{APPARATUS AND METHOD FOR VALVE STICTION DIAGNOSING USING MACHINE LEARNING}

The present invention relates to a valve stiction diagnosis device and method using machine learning, and more specifically, to a valve stiction diagnosis device and method using machine learning, which performs valve stiction diagnosis by creating a learning model using machine learning techniques with features extracted through a valve stiction diagnosis algorithm using a stiction index and odd harmonic variables, thereby ensuring the accuracy of valve stiction diagnosis even in situations where it is difficult to comprehensively judge using existing valve stiction diagnosis algorithms.

In a plant, control is carried out by hundreds of control loops. As such, control loops must be maintained in an optimal state because they directly affect the stable operation of the plant.

These control loops can degrade in performance for a number of reasons, including improper tuning of controller parameters, fluctuating disturbances, equipment failures, interactions between control loops, control loop nonlinearities, and process design errors.

In general cases, the cause of performance degradation in the control loop is mostly considered to be improper tuning of controller parameters.

However, the cause of performance degradation in the control loop is often related to the nonlinearity of the control loop.

Nonlinearities in these control loops can appear in the control valve or in the process and instrumentation, and can manifest as stiction, deadband, hysteresis, and saturation.

In particular, the stiction phenomenon of the control valve is a phenomenon that has a negative effect on the performance of the control loop, and can be considered an important factor in valve failure as it is caused by excessive tightening of the valve packing, corrosion, foreign substances, deterioration of the valve seal, etc.

Stiction of these control valves can even cause the plant to be shut down while the control valves are being serviced.

Valve stiction can have serious effects on the control loop, requiring plant shutdown to resolve, but it is often overlooked and taken lightly.

Therefore, an environment that can diagnose and respond to the valve sticking phenomenon as a cause of performance degradation in the control loop is needed.

Yoshiyuki Yamashita, "An automatic method for detection of valve stiction in process control loops", Control Engineering Practice, Volume 14, Issue 5, May 2006, pp 503-510 Monir Ahmad, M.A.A. Shoukat Choudhury, "A Simple Harmonics Based Stiction Detection Method", DYCOPS 2010, pp 671-676

The purpose of the present invention is to provide a valve stiction diagnosis device and method using machine learning to secure the accuracy of valve stiction diagnosis even in situations where it is difficult to comprehensively judge existing valve stiction diagnosis algorithms by performing valve stiction diagnosis by creating a learning model using machine learning techniques with features extracted through a valve stiction diagnosis algorithm using a stiction index and odd harmonic variables.

A valve stiction diagnosis device using machine learning according to an embodiment of the present invention may include a stiction index feature extraction unit (110) for extracting a valve stiction index for a control system in the field as a feature for valve stiction diagnosis; an odd harmonic feature extraction unit (120) for extracting odd harmonic variables for the control system as features for valve stiction diagnosis; a learning model generation unit (150) for generating and providing a learning model using a plurality of machine learning techniques based on supervised learning for learning data generated by performing a simulation for cases where a valve stiction model is applied and cases where it is not applied by simulating the control system; and a valve stiction diagnosis unit (160) for diagnosing valve stiction for the control system by applying the valve stiction index and the odd harmonic variables to the learning model.

The above-mentioned sticktion index feature extraction unit may calculate and extract the valve sticktion index using the controller output (OP) and valve position (MV) input from the control system.

The above odd harmonic feature extraction unit may extract the odd harmonic variables by determining whether odd harmonics of the fundamental frequency exist using a controller error input from the control system.

According to an embodiment, the present invention may further include a learning data generation unit for generating the learning data by using simulation data generated through a simulation for cases where the valve stickition model is applied and not applied between a controller and a process model expressed in a block diagram.

The above simulation data may include controller output (OP), valve position (MV), setpoint (SP), and process value (PV).

The above controller may be a PI (Proportional Integral/PID (Proportional Integral Derivatve) controller, and a controller error, which is the difference between the setpoint (SP) and the process value (PV) included in the simulation data, may be input.

The above valve sticktion model may be one in which the controller output (OP) included in the simulation data is input, and the valve position (MV) included in the simulation data is output.

The above valve sticktion model may be implemented by any one of the Choudhury model, the Kano model, and the XCH model.

The above process model may be one in which the valve position (MV) included in the simulation data is input, and the process value (PV) included in the simulation data is output.

According to an embodiment, the method may further include a process model estimation unit for estimating the process model using the controller output (OP) and process value (PV) input from the control system and providing the estimated process model to the learning data generation unit.

The above learning data generation unit may generate the learning data by using a process model composed of process gain, time constant and delay time, which are parameters of the process model, and adjusting proportional gain and integral gain, which are parameters of the controller, and the sum of dead band and stick band and stick band, which are parameters of the valve stiction model.

The above learning data may include stiction authorization information indicating whether the valve stiction model is authorized, a stiction index extracted using the controller output (OP) and valve position (MV) included in the simulation data, and odd harmonic variables extracted using the setpoint (SP) and process value (PV) included in the simulation data.

The above learning model generation unit may divide the learning data into learning data and verification data, and generate learning models for each of a plurality of machine learning techniques based on supervised learning using the learning data, and perform verification on each learning model using the verification data to provide the learning model with the highest accuracy to the valve sticktion diagnosis unit.

The above multiple machine learning techniques may include a Naive Bayes technique, a Neural Network technique, and a Support Vector Machine technique.

The above learning model generation unit may divide the learning data and the verification data into a predetermined ratio, randomly extract the learning data, and determine the verification data as the remainder after excluding the learning data.

In addition, a method for diagnosing valve stiction using machine learning according to an embodiment of the present invention may include a step of extracting a valve stiction index and odd harmonic variables for a control system in the field as features for diagnosing valve stiction; a step of generating a learning model using a plurality of machine learning techniques based on supervised learning on learning data generated by performing a simulation for cases where a valve stiction model is applied and cases where it is not applied by simulating the control system; and a step of diagnosing valve stiction for the control system by applying the valve stiction index and the odd harmonic variables to the learning model.

The step of generating and providing the above learning model may include a step of dividing the learning data into learning data and verification data; a step of generating learning models for each of a plurality of machine learning techniques based on supervised learning using the learning data; and a step of performing verification on each learning model using the verification data to provide a learning model with the highest accuracy.

The step of dividing the data for learning and the data for verification may be to divide the data for learning and the data for verification according to a predetermined ratio, randomly extract the data for learning, and determine the data for verification as the remainder after excluding the data for learning.

In addition, a valve stiction diagnosis device according to an embodiment of the present invention comprises: at least one processor; and a memory for storing computer-readable instructions; wherein the instructions, when executed by the at least one processor, cause the valve stiction diagnosis device to extract a valve stiction index and odd harmonic variables for a control system in the field as features for valve stiction diagnosis, and generate a learning model using a plurality of machine learning techniques based on supervised learning for learning data generated by performing a simulation for cases where a valve stiction model is applied and cases where it is not applied by simulating the control system, and to diagnose valve stiction for the control system by applying the valve stiction index and the odd harmonic variables to the learning model.

The present invention performs valve stiction diagnosis by creating a learning model using machine learning techniques with features extracted through a valve stiction diagnosis algorithm using a stiction index and odd harmonic variables, thereby ensuring the accuracy of valve stiction diagnosis even in situations where it is difficult to comprehensively judge using existing valve stiction diagnosis algorithms.

In addition, the present invention estimates the process using the controller output and process value of the control loop to be diagnosed, generates a learning model using the simulation results using the estimated process model and the valve stiction model and a machine learning technique, and diagnoses valve stiction using features for valve stiction diagnosis extracted using the generated learning model and control data from the control loop.

In addition, since the present invention has various control loops to be diagnosed and each control loop has different process characteristics, the accuracy of valve stiction diagnosis can be improved by using process identification of the corresponding control loop and simulation data using a valve stiction model.

In addition, the present invention can identify the influence of valve sticktion on a plant through diagnosis of the valve sticktion phenomenon of a control loop, thereby contributing to the stable operation of a plant.

In addition, the present invention can extract the valve stiction index and the presence or absence of odd harmonics by using, in addition to the method of using the relationship between the controller output and the valve position and the controller error, the relationship between the controller output and the process value, the controller output or the valve position.

In addition, the present invention can diagnose by optimizing the parameter values of the algorithm using a valve stiction diagnosis algorithm and a process model estimated from the control data of the control loop and a simulation using the valve stiction model, and then reflecting the results in an algorithm that diagnoses using the control data.

Figure 1 is a drawing showing a control valve structure according to an embodiment of the present invention;
Figure 2 is a drawing showing the input/output operation characteristics of the valve stickition phenomenon using the input/output data of the control valve of Figure 1.
Figures 3 to 6 are diagrams showing control loop trends when valve sticking occurs.
Figure 7 is a drawing explaining the relationship between the controller output and the valve position in the stickition state.
Figure 8 is a drawing explaining the determination of the synthetic direction of the controller output and the valve position.
FIG. 9 is a drawing showing a valve stickion diagnostic device according to an embodiment of the present invention;
Figure 10 is a drawing showing a system block diagram for valve stiction simulation.
Figure 11 is a diagram showing the input/output flow in the XCH model.
Figure 12 is a diagram showing an example of learning data generated through valve stiction simulation according to changes in process model parameters, controller parameters, and valve stiction model parameters.
Figure 13 is a diagram showing the simulation trend when valve sticktion is applied.
Figure 14 is a diagram showing the simulation trend when valve stickion is not applied.
Figure 15 is a diagram showing learning data randomly extracted from the learning data of Figure 12.
Figure 16 is a diagram showing the verification data from which the learning data is excluded from the learning data of Figure 12 and the verification result thereof.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the attached drawings. However, detailed descriptions of well-known functions or configurations that may obscure the gist of the present invention in the following description and the attached drawings will be omitted. In addition, it should be noted that identical components are indicated with the same drawing reference numerals as much as possible throughout the drawings.

The terms or words used in this specification and claims described below should not be interpreted as limited to their usual or dictionary meanings, but should be interpreted as meanings and concepts that conform to the technical idea of the present invention, based on the principle that the inventor can appropriately define the terms to best describe his or her invention.

Therefore, the embodiments described in this specification and the configurations illustrated in the drawings are only the most preferred embodiments of the present invention and do not represent all of the technical ideas of the present invention. Therefore, it should be understood that there may be various equivalents and modified examples that can replace them at the time of filing this application.

In the attached drawings, some components are exaggerated, omitted, or schematically depicted, and the size of each component does not entirely reflect the actual size. The present invention is not limited by the relative sizes or spacings drawn in the attached drawings.

When a part of the specification is said to "include" a component, this does not mean that it excludes other components, but rather that it may include other components, unless otherwise specifically stated. Also, when a part is said to be "connected" to another part, this includes not only the case where it is "directly connected," but also the case where it is "electrically connected" with other elements in between.

Singular expressions include plural expressions unless the context clearly indicates otherwise. It should be understood that terms such as "comprises" or "have" are intended to specify the presence of a feature, number, step, operation, component, part, or combination thereof described in the specification, but do not preclude the possibility of the presence or addition of one or more other features, numbers, steps, operations, components, parts, or combinations thereof.

Also, the term "part" used in the specification means a software, hardware component such as an FPGA or an ASIC, and the "part" performs certain functions. However, the "part" is not limited to software or hardware. The "part" may be configured to be on an addressable storage medium and may be configured to execute one or more processors. Thus, by way of example, the "part" includes components such as software components, object-oriented software components, class components, and task components, processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuitry, data, databases, data structures, tables, arrays, and variables. The functionality provided in the components and "parts" may be combined into a smaller number of components and "parts" or further separated into additional components and "parts."

Hereinafter, with reference to the attached drawings, embodiments of the present invention will be described in detail so that those skilled in the art can easily practice the present invention. However, the present invention may be implemented in various different forms and is not limited to the embodiments described herein. In addition, in order to clearly describe the present invention in the drawings, parts that are not related to the description are omitted, and similar parts are assigned similar drawing reference numerals throughout the specification.

Hereinafter, preferred embodiments of the present invention will be described with reference to the attached drawings.

FIG. 1 is a drawing showing a control valve structure according to an embodiment of the present invention.

Referring to FIG. 1, a control valve according to an embodiment of the present invention controls the opening and closing of the valve by the operation of an actuator (10) according to the intake or exhaust of air.

At this time, the control valve determines the direction of movement of the valve stem (20) as air is introduced or exhausted. Here, one end of the valve stem (20) is connected to a diaphragm (12), and the other end is connected to a plug (30).

That is, when air is introduced into the control valve, the diaphragm (12) moves upward due to air pressure, thereby opening the plug (30) connected to the valve stem (20), and when the air is exhausted, the diaphragm (12) moves downward due to the elastic force of the spring (11), thereby closing the plug (30) connected to the valve stem (20).

Valve stiction is defined as the resistance to initiation of motion required to overcome static friction and measured as the difference between the up and down driving forces.

The valve stem (20) moves against static or dynamic friction. Accordingly, the valve stem (20) does not move with small changes in air pressure due to valve stiction, and moves when the valve stiction is overcome.

Fig. 2 is a diagram showing the input/output operation characteristics of the valve sticktion phenomenon using the input/output data of the control valve of Fig. 1, and Figs. 3 to 6 are diagrams showing the control loop trend when valve sticktion occurs.

Referring to Figure 2, the controller output is related to air pressure and the valve position is related to elastic force.

The input-output operation characteristics of a valve, which exhibit valve stiction, consist of four components: dead band, stick band, slip jump, and moving phase.

When the valve stops at point N or tries to change direction, the valve sticks.

After the controller output overcomes the dead band (NA') and stick band (A'C), the valve jumps to a new position (point C) and continues moving.

While moving in the same direction, the valve may get stuck again at point D due to zero velocity or very slow velocity. At this time, the dead band is zero and only the stick band (DE) exists. In this case, the valve can overcome the valve sticktion if the controller output is greater than the stick band.

The valve operation characteristics for cases where the valve does not move despite changes in valve input (controller output) can be expressed by dead band and stick band.

The magnitude of the slip jump plays a significant role in determining the oscillation dynamics induced by valve sticktion. That is, once a slip jump occurs, the valve continues to move until it sticks again.

From this perspective, we can redefine valve stiction to mean the characteristic of a valve in which a slip jump occurs, a sudden jump before the valve moves smoothly in response to a changing input. The slip jump is expressed as a percentage of the output span.

Referring to FIGS. 3 to 6, OP represents controller output, PV represents process value, and MV represents valve position feedback.

The XY graph of the controller output (OP) and the valve position feedback (MV) represents a quadrilateral as shown in Fig. 3, and the XY graph of the controller output (OP) and the process value (PV) represents an ellipse as shown in Fig. 4.

The control loop causes oscillation, and the process value (PV) exhibits a sine wave shape as shown in Fig. 5, and the valve position feedback (MV) exhibits a square wave shape as shown in Fig. 6.

Hereinafter, a valve stiction diagnosis technique (e.g., shape-based stiction detection method) based on time domain data reflecting the input/output operation characteristics of Fig. 2 and the control loop trends of Figs. 3 to 6 will be examined with reference to Figs. 7 and 8.

This valve stiction diagnostic technique was proposed in 'Yoshiyuki Yamashita, An automatic method for detection of valve stiction in process control loops, Control Engineering Practice, Volume 14, Issue 5, May 2006, pp 503-510'.

Figure 7 is a drawing explaining the relationship between the controller output and the valve position in the stickition state, and Figure 8 is a drawing explaining the determination of the synthetic direction of the controller output and the valve position.

This valve sticktion diagnosis technique recognizes a sticktion state when the X-Y relationship between the controller output (OP) and the valve position (MV) is a quadrilateral. In other words, this recognizes valve sticktion by utilizing the characteristic that even if the controller output, which is the value of the X-axis, changes, the valve position, which is the Y-axis, does not change.

Fig. 7 (a) shows the operating characteristics of the dead band, Fig. 7 (b) shows the operating characteristics of the slip jump, and Fig. 7 (c) shows the operating characteristics of the dead band of (a) and the slip jump of (b) simultaneously.

In this valve sticktion diagnosis technique, the characteristic that the valve position (MV) does not change even if the controller output (OP) changes in the sticktion state is utilized to define the operating directions of OP and MV as shown in Fig. 8. Here, D is Decreasing, S is Steady, and I is Increasing.

Three stiction indices, namely SI_C1, SI_C2, and SI_C3, are calculated as follows.

Here,

SI_C1 = (T _IS + T _DS ) / (T _total -T _SS )

SI_C2 = (T _ISII ＋T _ISSI ＋T _DSDD ＋T _DSSD )／(T _total －T _SS )

SI_C3 = SI_C1 - (T _ISDD ＋T _ISDI ＋T _ISSD ＋T _ISID ＋T _ISDS ＋T _DSDI ＋T _DSS ＋T _DSID ＋T _DSII ＋T _DSIS )／(T _total －T _SS )

Here, T _IS is the sum of times for the IS section, T _SS is the sum of times for the SS section, T _DS is the sum of times for the DS section, T _ISII is the sum of times for the IS section in the section that is IS II, T _ISSI is the sum of times for the IS section in the section that is IS SI, T _DSDD is the sum of times for the DS section in the section that is DS DD, T _DSSD is the sum of times for the DS section in the section that is DS SD, T _ISDD is the sum of times for the IS section in the section that is IS DD, T _ISDI is the sum of times for the IS section in the section that is IS DI, T _ISSD is the sum of times for the IS section in the section that is IS SD, T _ISID is the sum of times for the IS section in the section that is IS ID, T _ISDS is the sum of times for the IS section in the section that is IS DS, T _DSDI is the sum of times for the DS section in the section that is DS DI, T _DSSI is the sum of times for the DS section in the section that is DS SI, T _DSID is the sum of times for the DS section in the section that is DS ID, T _DSII is the time for the DS section in the section that is DS II. Total, T _DSIS is the time sum of the DS interval in the DS IS interval, and T _total represents the time window for diagnosis.

This valve stiction diagnosis technique is an algorithm that uses a stiction index, and thus requires setting a threshold value for the stiction index to determine whether stiction has occurred. Depending on the threshold value setting, there may be cases where it is difficult to determine whether stiction has occurred.

Meanwhile, other valve stiction diagnostic techniques focus on the fact that odd harmonic components appear in the power spectrum of square waves, while odd and even harmonic components appear in triangle waves/sawtooth waves.

This valve stiction diagnostic technique has been proposed in 'Monir Ahammad, M.A.A. Shoukat Choudhury, A Simple Harmonics Based Stiction Detection Method, DYCOPS 2010, pp 671-676'.

In this case, considering that the controller error and the controller output take the form of square waves and sawtooth/triangular waves when valve sticktion occurs, the presence of odd harmonics is determined by Fourier transforming the signal for the controller error to determine whether valve sticktion occurs.

This valve stiction diagnosis technique is an algorithm that utilizes the presence of odd harmonic components. Since the presence of odd harmonic components varies depending on the process and valve status, a standard for determining whether valve stiction occurs for each odd harmonic component is required.

The valve stiction diagnosis technique using the two algorithms described above requires threshold setting of the stiction index and judgment criteria based on the presence of each odd harmonic component, and if the diagnosis results by each algorithm are different, valve stiction diagnosis may be difficult.

Hereinafter, a valve stickion diagnostic device according to an embodiment of the present invention will be described with reference to FIG. 9 described later.

FIG. 9 is a drawing showing a valve stickition diagnosis device using machine learning according to an embodiment of the present invention.

As illustrated in FIG. 9, the valve stiction diagnosis device (100) according to an embodiment of the present invention performs valve stiction diagnosis by creating a learning model using a machine learning technique with features extracted through a valve stiction diagnosis algorithm using a stiction index and odd harmonic variables, thereby ensuring the accuracy of valve stiction diagnosis even in situations where it is difficult to comprehensively judge using an existing valve stiction diagnosis algorithm.

Meanwhile, the valve sticktion diagnosis device (100) can be configured to include a sticktion index feature extraction unit (110), an odd harmonic feature extraction unit (120), a process model estimation unit (130), a learning data generation unit (140), a learning model generation unit (150), and a valve sticktion diagnosis unit (160).

This valve sticktion diagnostic device (100) receives data on the controller output (OP), valve position (MV), controller setpoint (SP), and process value (PV) of the control loop from the control system (50) located at the site where valve sticktion is to be diagnosed.

First, the stiction index feature extraction unit (110) extracts three valve stiction indices (SI_C1, SI_C2, SI_C3) calculated using the controller output (OP) and valve position (MV) input from the control system (50) as features for valve stiction diagnosis.

Calculation process of valve stiction index in the stiction index feature extraction unit (110)

① Set the initial values of the dead band (db_op) for determining the D/I/S direction of △OP(t) and the dead band (db_mv) for determining the D/I/S direction of △MV(t).

② Enter the time window (T _total ) to diagnose valve sticktion.

③ Read the controller output (OP) and valve position (MV) during the input time window (T _total ).

④ The controller output difference value (△OP) and valve position difference value (△MV) are calculated using the controller output (OP) and valve position (MV) at each step during the input time window (T _total ).

⑤ The individual directions (D/I/S) of the controller output [OP(t)] based on the time domain and the valve position [MV(t)] based on the time domain are determined using the dead band (db_op) for determining the D/I/S direction of △OP(t) set as the initial value and the dead band (db_mv) for determining the D/I/S direction of △MV(t).

⑥ Using the directions of OP(t) and MV(t) obtained in ⑤, the synthetic direction (IS, II, DD, DS, IS, SI, SD, DS, etc.) of the controller output (OP) and valve position (MV) as in Fig. 8 is determined.

⑥ Calculate T _IS , T _SS , T _DS , T _SS , T _ISII , T _ISSI , T DSDD, T _DSSD , T _ISDD , T _ISDI , T _ISSD , T _ISID , T _ISDS , T _DSDI , T _DSSI , T _DSID , T _DSII , T _DSIS during the entered time _window .

⑦ Calculate the three valve stickition indices as follows.

SI_C1 = (T _IS + T _DS ) / (T _total -T _SS )

SI_C2=(T _ISII ＋T _ISSI ＋T _DSDD ＋T _DSSD )／(T _total －T _SS )

SI_C3 = SI_C1 - (T _ISDD ＋T _ISDI ＋T _ISSD ＋T _ISID ＋T _ISDS ＋T _DSDI ＋T _DSS ＋T _DSID ＋T _DSII ＋T _DSIS )／(T _total －T _SS )

Next, the odd harmonic feature extraction unit (120) extracts odd harmonic variables (H3E, H5E, H7E) indicating the presence of odd harmonics (i.e., 3rd harmonic, 5th harmonic, 7th harmonic) of the fundamental frequency based on frequency analysis of the controller error using the controller setting value (SP) and process value (PV) input from the control system (50) as features for valve sticktion diagnosis.

Odd harmonic feature extraction unit (120) odd harmonic judgment process

① Set the initial values of the size threshold (odd_F_th) and frequency band threshold (odd_F_bw) to determine whether odd harmonics of the fundamental frequency exist.

② Initialize the variable indicating the presence of the 3rd harmonic (H3E), the variable indicating the presence of the 5th harmonic (H5E), and the variable indicating the presence of the 7th harmonic (H7E) to ‘0’.

③ Enter the time window (T _total ) to diagnose valve sticktion.

④ Read the controller setting value (SP) and process value (PV) during the input time window (T _total ).

⑤ The controller error is calculated using the controller setting value (SP) and process value (PV) at each step during the input time window (T _total ). Here, the controller error is the difference between the controller setting value (SP) and the process value (PV) (i.e., |SP－PV|).

⑥ The fast Fourier transform (FFT) is performed using the controller error signal, and then the fundamental frequency is extracted. Here, the frequency response size of the fundamental frequency is referred to as A.

⑦ If (frequency response size/A) in the range of (fundamental frequency × 3±odd_F_bw) falls within the range of odd_F_th, the 3rd harmonic is determined to exist. At this time, the 3rd harmonic variable (H3E) indicating the existence of the 3rd harmonic becomes '1'.

⑧ If the (frequency response size/A) in the range of (fundamental frequency × 5±odd_F_bw) falls within the range of odd_F_th, the 5th harmonic is determined to exist. At this time, the 5th harmonic variable (H5E) indicating the existence of the 5th harmonic becomes '1'.

⑨ If the (frequency response size/A) in the range of (fundamental frequency × 7±odd_F_bw) falls within the range of odd_F_th, the 7th harmonic is determined to exist. At this time, the 7th harmonic variable (H7E) indicating the existence of the 7th harmonic becomes '1'.

Next, the process model estimation unit (130) estimates the first process model of the control loop using the controller output (OP) and process value (PV) input from the control system (50).

At this time, the first-order process model consists of a time constant, process gain, and delay time, and is called the first-order delay model (First Order Plus Dead Time, FOPDT). This first-order process model can be expressed as the following mathematical expression 1.

Here, U(s) is the Laplace transform of the input u(t), Y(s) is the Laplace transform of the output y(t), T is the time constant, K is the process gain, and Td is the delay time.

In addition, the process model estimation unit (130) can estimate not only the first-order process model as in the mathematical expression 1 above, but also a higher-order process model.

Next, the learning data generation unit (140) generates learning data for generating a learning model for valve stiction diagnosis by performing valve stiction simulation for cases where the valve stiction model is applied and cases where it is not applied by simulating the control system (50).

This learning data generation unit (140) can configure a valve stiction simulation process for generating learning data for a control loop as shown in the block diagram in Fig. 10. Fig. 10 is a drawing showing a system block diagram for valve stiction simulation.

Referring to Fig. 10, the learning data generation unit (140) performs a valve stickition simulation process for cases where a valve stickition model (142) is applied between the controller (141) and the process model (143) and cases where it is not applied.

Here, the process model (143) assumes a first-order process model that can be provided from the process model estimation unit (130). The process model (143) is composed of process model parameters such as process gain (K), time constant (T), and delay time (Td).

At this time, the learning data generation unit (140) generates simulation data for cases where valve stiction is applied and cases where valve stiction is not applied, i.e., controller output (OP), valve position (MV), set value (SP), and process value (PV), as the valve stiction simulation process proceeds.

Specifically, the controller (141) applies a PI (Proportional Integral/PID (Proportional Integral Derivatve) controller of the following mathematical expression 2.

The above mathematical expression 2 represents the transfer function of the controller (141).

Here, the controller input is the controller error e(t), which is expressed as the difference between the setpoint (SP) and the process value (PV). That is, e(t) = SP―PV. The Laplace transform of the controller input e(t) is E(s).

And, in the above mathematical expression 2, Kp is the proportional gain and Ki is the integral gain.

In addition, the valve stiction model (142) exhibits the same characteristics as a physical model of the stiction phenomenon of the valve. For example, the valve stiction model (142) can be implemented as any one of the Choudhury model, the Kano model, and the XCH model.

Here, we will examine the XCH model with reference to Fig. 11. Fig. 11 is a diagram showing the input/output flow in the XCH model.

In Fig. 11, x(k) denotes the valve input (controller output) signal at the kth time step (or the current time point), and x(k-1) denotes the valve input (controller output) signal at the previous step.

Also, y(k) represents the valve position at the kth time step (or the current point in time), and y(k-1) represents the valve position at the previous step.

And, S represents the dead band and the stick band and the sum, and J represents the stick band.

Also, xss stands for stuck location, and I stands for whether a stuck state occurs or not.

And, Δt means the scan period (time between steps) for the valve input (controller output) signal, and sign(v) means the sign of v. Here, when v is greater than 0, it has the value 1, when it is less than 0, it has the value -1, and when it is 0, it has the value 0.

Also, fs stands for static friction, which is (S+J)/2.

In this way, the learning data generation unit (140) uses the process model (143) estimated by the process model estimation unit (130), but changes the characteristics of the control loop by adjusting the proportional gain (Kp) and the integral gain (Ki) of the controller (141), and changes the sum of the dead band and the stick band (S) and the stack band (J) representing the degree of stiction in the valve stiction model (142), thereby obtaining simulation data for cases where valve stiction is applied and cases where it is not applied.

Then, the learning data generation unit (140) can generate learning data using simulation data.

To this end, the learning data generation unit (140) further includes a stickiness index feature extraction simulation unit (144) that implements the same function as the stickiness index feature extraction unit (110) described above, and an odd harmonic feature extraction simulation unit (145) that implements the same function as the odd harmonic feature extraction unit (120) described above.

At this time, the stiction index feature extraction simulation unit (144) extracts three stiction indices (SI_C1, SI_C2, SI_C3) using simulation data generated as the valve stiction simulation process proceeds, i.e., the controller output (OP) and valve position (MV).

In addition, the odd harmonic feature extraction simulation unit (145) extracts odd harmonic variables (H3E, H5E, H7E) indicating the existence of three 3rd/5th/7th harmonics using simulation data generated as the valve stickition simulation process proceeds, i.e., the set value (SP) and the process value (PV).

Then, the learning data generation unit (140) can generate learning data characterized by sticktion authorization information, sticktion index, and odd harmonic variables using simulation data.

In this way, the learning data generation unit (140) performs a valve stiction simulation while changing the process model parameters, namely, process gain (K), time constant (T), and delay time (Td), the controller parameters, namely, proportional gain (Kp) and integral gain (Ki), and the valve stiction parameters, namely, the sum of dead band and stick band (S) and the stack band (J).

Figure 12 is a diagram showing an example of learning data generated through valve stiction simulation according to changes in process model parameters, controller parameters, and valve stiction model parameters.

The learning data in Fig. 12 represents a case where 50 cases were generated. In this case, the valve stiction simulation was performed with the process gain (K), which is a process model parameter, fixed to 1, and the stiction indices (SI_C1, SI_C2, SI_C3) and odd harmonic variables (H3E, H5E, H7E) were extracted through the simulation data.

As shown in Fig. 12, the learning data may include process model parameters, controller parameters, valve stiction model parameters, stiction authorization information indicating whether the valve stiction model is authorized, stiction indices (SI_C1, SI_C2, SI_C3), and odd harmonic variables (H3E, H5E, H7E) indicating whether odd harmonics exist.

However, the learning model generation unit (150) utilizes the sticktion authorization information, sticktion index, and odd harmonic variables included in the learning data for learning.

However, the simulation trends for cases where the valve stickion model is authorized and not authorized are as shown in Figures 13 and 14.

Fig. 13 is a drawing showing a simulation trend when valve sticktion is applied, which is the case when case 29 of Fig. 12 is applied. Fig. 14 is a drawing showing a simulation trend when valve sticktion is not applied, which is the case when case 46 of Fig. 12 is applied.

Referring to FIGS. 13 and 14, it can be seen that when the valve stiction model is applied, oscillation occurs in the control loop, and the XY graph between the controller output (CO) and the process value (PV) is close to a quadrilateral or ellipse.

Looking at each case in Fig. 12, it can be seen that even when the valve stiction model is approved, there are cases where odd harmonics do not exist, and thus the valve stiction diagnosis results may appear differently, and it can also be seen that valve stiction diagnosis is difficult in cases where the threshold value of the stiction index is not set.

Referring again to FIG. 9, the learning model generation unit (150) generates a learning model for valve stickition diagnosis using learning data transmitted from the learning data generation unit (140).

That is, the learning model generation unit (150) generates a learning model by applying multiple machine learning techniques based on supervised learning using the sticktion authorization information included in the learning data as a target variable for whether or not sticktion occurs.

Here, the multiple machine learning techniques may be the Naive Bayes technique, the Neural Network technique, and the Support Vector Machine (hereinafter referred to as 'SVM') technique.

In this way, the learning model generation unit (150) applies multiple machine learning techniques to each learning data, determines the learning model with the highest accuracy as the learning model for valve stiction diagnosis, and provides it to the valve stiction diagnosis unit (160).

Then, the valve sticktion diagnosis unit (160) performs valve sticktion diagnosis by applying the valve sticktion diagnosis learning model provided from the learning model generation unit (150) to the sticktion index and odd harmonic variable characteristics extracted using the data of the field control system (50).

Specifically, the learning model generation unit (150) divides the learning data transmitted from the learning data generation unit (140) into learning data for generating learning models according to each of a plurality of machine learning techniques based on supervised learning, and verification data for use in verifying each of the generated learning models.

Here, the training data and the verification data are divided according to a predetermined ratio. For example, 70% (35) of the 50 training data are randomly extracted as training data (see Figure 15), and the remaining 30% (15) after excluding the training data are determined as verification data (see Figure 16).

Figure 15 is a diagram showing training data randomly extracted from the training data of Figure 12, and Figure 16 is a diagram showing verification data with the training data excluded from the training data of Figure 12 and the verification results thereof.

Then, the learning model generation unit (150) generates a learning model for machine learning techniques based on supervised learning, such as the naive Bayes technique, neural network technique, and SVM technique, using the learning data of Fig. 15.

And, the learning model generation unit (150) performs verification for each learning model using the verification data of Fig. 16.

At this time, the learning model generation unit (150) verifies whether the stickition authorization information and the diagnostic results of each learning model match each other and selects the learning model with the best performance (accuracy).

Referring to Figure 16, the naive Bayes technique shows an accuracy of 73.3%, the neural network technique shows an accuracy of 80%, and the SVM technique shows an accuracy of 93.3%.

The SVM technique shows that the stickiness authorization information and the diagnosis results are consistent with each other in 14 out of 15 validation data.

Accordingly, the learning model generation unit (150) transfers the valve stiction diagnosis learning model generated through the SVM technique to the valve stiction diagnosis unit (160).

Next, the valve sticktion diagnosis unit (160) diagnoses valve sticktion for the field control system (50) using the learning model for valve sticktion diagnosis generated by the learning model generation unit (150). That is, the valve sticktion diagnosis unit (160) uses control loop data for the time (T _total ) for diagnosing valve sticktion to generate sticktion indices (SI_C1, SI_C2, SI_C3) extracted by the sticktion index feature extraction unit (110) and odd harmonic variables (H3E, H5E, H7E) extracted by the odd harmonic feature extraction unit (120), and then applies these to the valve sticktion diagnosis learning model transmitted from the learning model generation unit (150) to diagnose whether valve sticktion occurs and calculate the result.

As another embodiment, the valve stickition diagnostic device (100) can perform the valve stickition diagnostic method according to an embodiment of the present invention when executing computer-readable instructions stored in a memory by at least one processor.

Here, the processor may be at least one processor, and may also be called a controller, a microcontroller, a microprocessor, a microcomputer, etc. And, the processor may be implemented by hardware, firmware, software, or a combination thereof.

Also, memory may be a single storage device, or may be a collective term for multiple storage elements. Computer-readable instructions stored in the memory may be executable program code, parameters, data, etc. And, memory may include random access memory (RAM), or may include non-volatile memory (NVRAM), such as magnetic disk storage or flash memory.

FIG. 17 is a diagram illustrating a valve stickition diagnosis method using machine learning according to an embodiment of the present invention, and FIG. 18 is a diagram explaining a process of generating a learning model in FIG. 17.

Referring to Fig. 17, the valve sticktion diagnosis device (100) extracts a valve sticktion index and odd harmonic variables for the control system (50) as features for valve sticktion diagnosis (S210).

At this time, the valve sticktion diagnostic device (100) calculates and extracts a valve sticktion index using the controller output (OP) and valve position (MV) input from the control system (50), and extracts odd harmonic variables by determining whether odd harmonics of the fundamental frequency exist using the controller error (i.e., the difference between the controller setpoint (SP) and the process value (PV)) input from the control system (50).

And, before diagnosing valve stickition for the control system (50), the valve stickition diagnosis device (100) simulates the control system (50) and performs a simulation for cases where the valve stickition model is applied and cases where it is not applied, thereby generating a learning model using multiple machine learning techniques based on supervised learning for the generated learning data (S220).

At this time, the valve stickition diagnostic device (100) generates learning data using simulation data generated through simulation for cases where the valve stickition model (142) is applied and not applied between the controller (141) and the process model (143) expressed in the block diagram.

Here, the simulation data includes controller output (OP), valve position (MV), setpoint (SP), and process value (PV).

And, the controller (141) is a PI (Proportional Integral/PID (Proportional Integral Derivatve) controller, and a controller error, which is the difference between the setpoint (SP) and the process value (PV) included in the simulation data, is input.

In addition, the valve stiction model (143) is input with the controller output (OP) included in the simulation data, and the valve position (MV) included in the simulation data is output. Here, the valve stiction model (143) is implemented as one of the Choudhury model, the Kano model, and the XCH model.

Additionally, the process model (143) inputs the valve position (MV) included in the simulation data and outputs the process value (PV) included in the simulation data.

Referring to Fig. 18, the valve stickion diagnostic device (100) generates a learning model as follows.

First, the valve stickion diagnosis device (100) divides the learning data into learning data and verification data (S221). At this time, the valve stickion diagnosis device (100) divides the learning data and verification data according to a predetermined ratio, randomly extracts the learning data, and determines the verification data as the remainder after excluding the learning data.

Thereafter, the valve stiction diagnosis device (100) generates a learning model for each of multiple machine learning techniques based on supervised learning using the learning data. At this time, the valve stiction diagnosis device (100) can generate a learning model using the naive Bayes technique (S222), generate a learning model using the neural network technique (S223), and generate a learning model using the SVM technique (S224).

Thereafter, the valve stickion diagnosis device (100) performs verification for each learning model using the verification data. At this time, the valve stickion diagnosis device (100) performs a diagnosis and calculates the accuracy for the naive Bayes learning model using the verification data (S225), performs a diagnosis and calculates the accuracy for the neural network learning model using the verification data (S226), and performs a diagnosis and calculates the accuracy for the naive Bayes learning model using the verification data (S227).

Then, the valve sticktion diagnosis device (100) selects the learning model with the highest accuracy as the learning model for valve sticktion diagnosis (S228).

Thereafter, the valve sticktion diagnosis device (100) diagnoses valve sticktion for the control system (50) by applying the valve sticktion index and odd harmonic variables to the learning model (S230).

The method according to some embodiments may be implemented in the form of program instructions that can be executed by various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., alone or in combination. The program instructions recorded on the medium may be those specially designed and configured for the present invention or may be those known to and available to those skilled in the art of computer software. Examples of the computer-readable recording medium include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CDROMs and DVDs, magneto-optical media such as floptical disks, and hardware devices specially configured to store and execute program instructions such as ROMs, RAMs, flash memories, etc. Examples of the program instructions include not only machine language codes generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter, etc.

Although the above description has focused on the novel features of the present invention as applied to various embodiments, it will be understood by those skilled in the art that various deletions, substitutions, and changes in the form and details of the devices and methods described above may be made without departing from the scope of the present invention. Accordingly, the scope of the present invention is defined by the appended claims rather than by the above description. All changes which come within the scope of equivalency of the claims are intended to be embraced therein.

10 ; Actuator
11 ; Spring
12; diaphragm
20 ; valve stem
30 ; plug
50 ; Control System
100 ; Valve sticktion diagnostic device
110 ; Stiction index feature extraction section
120 ; Odd harmonic feature extraction unit
130; Process model estimation section
140 ; Learning data generation section
141 ; Controller
142 ; Valve sticktion model
143 ; Process Model
144; Stiction Index Feature Extraction Simulation Section
145 ; Odd harmonic feature extraction simulation
150 ; Learning model generation section
160 ; Valve Stiction Diagnostic Section
OP ; Controller output
MV; valve position
PV; process value
SP ; setting value

Claims (20)

A valve stickiness index feature extraction unit (110) for extracting a valve stickiness index for a field control system as a feature for valve stickiness diagnosis;
An odd harmonic feature extraction unit (120) for extracting odd harmonic variables for the above control system as features for valve sticking diagnosis;
A learning model generation unit (150) for generating a learning model using multiple machine learning techniques based on supervised learning on learning data generated by performing a simulation for cases where the valve stiction model is approved and not approved by simulating the above control system;
A valve stickion diagnosis unit (160) for diagnosing valve stickion for the control system by applying the valve stickion index and the odd harmonic variable to the learning model; and
It further includes a learning data generation unit for generating the learning data by using simulation data generated through simulation for cases where the valve stickition model is applied and not applied between the controller and the process model expressed in the block diagram;
The above learning data generation unit uses a process model composed of process gain, time constant, and delay time, which are parameters of the process model, and generates the learning data while adjusting the proportional gain and integral gain, which are parameters of the controller, and the sum of the dead band and the stick band, and the stick band, which are parameters of the valve stiction model, in a machine learning-based valve stiction diagnosis device.
In the first paragraph,
The above-mentioned stickition index feature extraction unit is,
A valve sticktion diagnosis device using machine learning that calculates and extracts the valve sticktion index using the controller output (OP) and valve position (MV) input from the above control system.
In paragraph 1,
The above odd harmonic feature extraction unit is,
A valve stickition diagnosis device using machine learning, which extracts odd harmonic variables by determining whether odd harmonics of a fundamental frequency exist using a controller error input from the above control system.
delete In the first paragraph,
The above simulation data is,
A valve sticktion diagnostic device using machine learning, which includes a controller output (OP), a valve position (MV), a setpoint (SP), and a process value (PV).
In paragraph 5,
The above controller is a PI (Proportional Integral/PID (Proportional Integral Derivatve) controller,
A valve stickition diagnosis device using machine learning, in which a controller error, which is the difference between the setpoint (SP) and the process value (PV) included in the above simulation data, is input.
In paragraph 6,
The above valve sticktion model is,
A valve sticktion diagnosis device using machine learning, wherein a controller output (OP) included in the above simulation data is input and a valve position (MV) included in the above simulation data is output.
In paragraph 7,
The above valve sticktion model is,
A valve stiction diagnostic device using machine learning, wherein the device is implemented with one of the Choudhury model, Kano model, and XCH model.
In Article 8,
The above process model is,
A valve stickition diagnosis device using machine learning, wherein a valve position (MV) included in the above simulation data is input and a process value (PV) included in the above simulation data is output.
In Article 9,
A process model estimation unit for estimating the process model using the controller output (OP) and process value (PV) input from the above control system and providing the same to the learning data generation unit;
A valve stickition diagnostic device using machine learning including more.
delete In the first paragraph,
The above learning data is,
A valve stiction diagnosis device using machine learning, comprising: stiction authorization information indicating whether the valve stiction model is authorized; a stiction index extracted using a controller output (OP) and a valve position (MV) included in the simulation data; and odd harmonic variables extracted using a setpoint (SP) and a process value (PV) included in the simulation data.
In paragraph 1,
The above learning model generation unit is,
A valve stiction diagnosis device using machine learning, wherein the learning data is divided into learning data and verification data, learning models for each of multiple machine learning techniques based on supervised learning are created using the learning data, and verification is performed on each learning model using the verification data to provide the learning model with the highest accuracy to the valve stiction diagnosis unit.
In Article 13,
The above multiple machine learning techniques are:
A valve stiction diagnosis device using machine learning, including the Naive Bayes technique, the Neural Network technique, and the Support Vector Machine technique.
In Article 13,
The above learning model generation unit is,
A valve stickition diagnosis device using machine learning, wherein the learning data and the verification data are divided at a predetermined ratio, the learning data is randomly extracted, and the verification data is determined as the remainder after excluding the learning data.
A step of extracting valve stickiness indices and odd harmonic variables for the control system in the field as features for valve stickiness diagnosis;
A step of generating a learning model using multiple machine learning techniques based on supervised learning on learning data generated by performing a simulation for cases where the valve stiction model is applied and cases where it is not applied by simulating the above control system; and
A step of diagnosing valve stickiness for the control system by applying the valve stickiness index and the odd harmonic variable to the learning model;
The learning data generation unit for generating the above learning data is:
Simulation data generated through simulation for cases where the valve stickition model is applied and not applied between the controller and the process model expressed in the block diagram is used,
A valve stiction diagnosis method using machine learning, which generates the learning data by using a process model composed of process gain, time constant and delay time, which are parameters of the above process model, and adjusting the proportional gain and integral gain, which are parameters of the controller, and the sum of dead band and stick band and the stick band, which are parameters of the valve stiction model.
In Article 16,
The step of creating and providing the above learning model is:
A step of dividing the above learning data into learning data and verification data;
A step of creating a learning model for each of multiple machine learning techniques based on supervised learning using the above learning data; and
A step of performing verification on each learning model using the above verification data to provide the learning model with the highest accuracy;
A method for diagnosing valve stickiness using machine learning including:
In Article 17,
The above multiple machine learning techniques are:
A method for diagnosing valve stiction using machine learning, including the Naive Bayes technique, the Neural Network technique, and the Support Vector Machine technique.
In Article 17,
The step of dividing the above data into learning data and verification data is as follows:
A method for diagnosing valve stickiness using machine learning, wherein the learning data and the verification data are divided according to a predetermined ratio, the learning data is randomly extracted, and the verification data is determined as the remainder after excluding the learning data.
As a valve sticktion diagnostic device,
at least one processor; and
A memory for storing computer-readable instructions;
The above commands, when executed by the at least one processor, cause the valve stickition diagnostic device to:
Extract valve stickiness index and odd harmonic variables for the field control system as features for valve stickiness diagnosis.
By performing simulations for cases where the valve stiction model is authorized and not authorized by simulating the above control system, a learning model is created using multiple machine learning techniques based on supervised learning on the generated learning data.
Applying the valve sticktion index and the odd harmonic variables to the learning model to diagnose valve sticktion for the control system;
The learning data generation unit for generating the above learning data is:
Simulation data generated through simulation for cases where the valve stickition model is applied and not applied between the controller and the process model expressed in the block diagram is used,
A valve stiction diagnosis device that generates the learning data by using a process model composed of process gain, time constant and delay time, which are parameters of the above process model, and adjusting proportional gain and integral gain, which are parameters of the controller, and the sum of dead band and stick band and stick band, which are parameters of the above valve stiction model.

Images