An intelligent system for monitoring and diagnosis of the CO2 capture process

Expert Systems with Applications 38 (2011) 7935–7946

Contents lists available at ScienceDirect

Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa

An intelligent system for monitoring and diagnosis of the CO2 capture process Qing Zhou a, Christine W. Chan a,⇑, Paitoon Tontiwachwuthikul b a b

Energy Informatics Laboratory, Faculty of Engineering, University of Regina, Regina, Saskatchewan, Canada S4S 0A2 Process Systems Engineering Laboratory, International Test Centre for CO2 Capture (ITC), University of Regina, Regina, Saskatchewan, Canada S4S 0A2

a r t i c l e Keywords: CO2 capture DeltaV Simulate Intelligent system

i n f o

a b s t r a c t Amine-based carbon dioxide capture has been widely considered as a feasible ideal technology for reducing large-scale CO2 emissions and mitigating global warming. The operation of amine-based CO2 capture is a complicated task, which involves monitoring over 100 process parameters and careful manipulation of numerous valves and pumps. The current research in the field of CO2 capture has emphasized the need for improving CO2 capture efficiency and enhancing plant performance. In the present study, artificial intelligence techniques were applied for developing a knowledge-based expert system that aims at effectively monitoring and controlling the CO2 capture process and thereby enhancing CO2 capture efficiency. In developing the system, the inferential modeling technique (IMT) was applied to analyze the domain knowledge and problem-solving techniques, and a knowledge base was developed on DeltaV Simulate. The expert system helps to enhance CO2 capture system performance and efficiency by reducing the time required for diagnosis and problem solving if abnormal conditions occur. The expert system can be used as a decision-support tool that helps inexperienced operators control the plant; it can be used also for training novice operators. Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction The emission of large amounts of carbon dioxide (CO2) has caused increasing public concern regarding environmental pollution and global warming. To mitigate this serious environmental problem, the CO2 capture technology has become widely accepted as useful technology for reducing CO2 emissions from industrial sources. The goal of CO2 capture is to capture and remove CO2 from industrial gas streams before it is released into the atmosphere. The amine-based CO2 capture process has become a common method for CO2 removal because it is energy efficient (Sholeh, Svendsen, Karl, & Olav, 2007). In the amine-based CO2 capture process, an amine solvent is used to absorb CO2 from the flue gas, and CO2 is subsequently extracted from the amine solvent, which can then be regenerated and reused. Operation of an amine-based CO2 capture system is a complicated task because it involves monitoring and manipulation of 16 components and a number of valves/pumps. The 16 components are associated with over a 100 parameters, including temperatures, flow rates, pressures, and levels of reaction instruments. The monitoring and control of critical parameters is an important task in operation of the CO2 capture process because it directly impacts plant performance and capture efficiency of CO2. Since the monitoring and control task is complex, it is desirable to build a knowledge-based system that ⇑ Corresponding author. Tel.: +1 306 585 5225; fax: +1 306 585 4855. E-mail address: [email protected] (C.W. Chan). 0957-4174/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2010.12.010

can automatically monitor, control, and diagnose the CO2 capture process. In this paper, we present research conducted with the objective of building a knowledge-based expert system that can monitor, control, and diagnose the CO2 capture processes at the International Test Centre for CO2 Capture (ITC) located at the University of Regina in Saskatchewan, Canada. The system is called the Knowledge-Based System for Carbon Dioxide Capture (KBSCDC). The knowledge base consists of domain knowledge about: (1) the plant components and their attributes, and (2) the important process parameters and their desired operating ranges. The knowledge base also consists of the remedial actions that would address these abnormal situations. The KBSCDC system can help the operator monitor the operating conditions of the CO2 capture pilot plant by continuously comparing the measured values from sensors with normal or desired values. Plant components that have abnormal parameter values indicate that abnormal operating conditions have occurred. Deviations from the normal ranges would set off an alarm to advise the operator that a problem has occurred. The KBSCDC can conduct real-time monitoring and diagnosis, as well as suggest remedies for any abnormality detected, thereby improving the performance efficiency of the plant. An initial prototype of the system was developed on G2 (trademark of Gensym Corporation, USA), which is an object-oriented expert system development tool. However, the prototype can only monitor reaction instruments and diagnose their abnormalities. The system did not include the process control strategies applied

7936

Q. Zhou et al. / Expert Systems with Applications 38 (2011) 7935–7946

to the control devices. Details of the prototype system have been presented in Zhou, Chan, and Tontiwachiwuthikul (2009). This paper presents an improved version of the knowledge-based expert system that was implemented with DeltaV Simulate (trademark of Emerson Corp., USA). DeltaV Simulate provides control utilities which enables the configuration of control strategies in small modular components. These modules link algorithm, conditions, and provide control over the field devices such as pumps and valves. Modules can communicate directly with each other, and can be coordinated by other modules to perform higher-level control strategies. The modules deploy different algorithms such as sequential function chart (SFC) and function block diagram (FBD). SFC is made up of a series of steps, transitions, and actions and it is used for representing the sequence of controlling strategies which contain multiple states. FBD is made up of interconnected function blocks, which process the incoming signals and send the signal to the control devices. Each function block contains standard process control algorithm and parameters that customize the algorithm to perform a particular function in the process control. Therefore, the new version of KBSCDC has the following additional functions compared to the earlier G2 version: (1) modules for different control devices are configured based on their characteristics, and (2) the control strategies applied to the control devices are simulated. The paper discusses design and development of the improved version of the knowledge-based system, and demonstrates use of the system by using problems scenarios that occur due to abnormal conditions. The paper is organized as follows: Section 2 presents the background literature relevant to the area of CO2 capture and the knowledge acquisition approach of the inferential modeling technique adopted in this research. Section 3 describes the process of development of the knowledge base. Section 4 presents design and implementation of the system on DeltaV Simulate based on the developed knowledge base. Application of the system is demonstrated using a case study in Section 5. Section 6 gives a conclusion and includes some discussion about future work.

2. Background literature 2.1. Studies of amine-based CO2 capture The study of amine-based CO2 capture has been ongoing for the last decade. The general objective of the study is to improve effectiveness and efficiency of the CO2 capture process. The research has been primarily conducted in the following two areas: (1) Study of the behaviour of the conventional amine solvents and development of new or improved solvents with higher CO2 absorption capacities, faster CO2 reaction rates, higher degradation resistance, and lower heat consumption for regeneration. The studies of corrosion were conducted in CO2 absorption systems using different types of aqueous amine solvents including methyldiethanolamine (MDEA), diethanolamine (DEA), 2-amino-2-methyl1-propanol (AMP) and monoethanolamine (MEA), and it was found that the corrosiveness increased in the order of MDEA < DEA < AMP < MEA (Veawab, Tontiwachwuthikul, & Bhole, 1997; Veawab, Tontiwachwuthikul, & Chakma, 1999). It was suggested that 2-(2-aminoethyl-amino) ethanol (AEEA) is a potentially good absorbent for capturing CO2 because of its high absorption rate and high absorption capacity (Sholeh et al., 2007). Chakma (1997) proposed that utilization of mixed solvents could reduce energy consumption and solve a number of operational

problems. Idem et al. (2006) evaluated the benefits of using mixed MEA/MDEA solvent for CO2 capture and found that a very large heat-duty reduction could be achieved by using a mixed MEA/MDEA solvent instead of a single MEA solvent. (2) Selection of appropriate solvents for different applications to reduce the energy penalty. It was proposed that the crucial criteria of solvent selection include feed gas characteristics such as composition, pressure, temperature, and the treated gas specifications (Veawab, Aroonwilas, Chakma, & Tontiwachwuthikul, 2001). White, Strazisar, Granite, and Hoffman (2003) suggested that solvent selection is influenced by solvent characteristics such as CO2 absorption capability and rates and operational issues of the process such as corrosion potential and solvent stability. These factors influence the equipment size, solvent consumption and heat consumption. Tontiwachwuthikul (1996) proposed that the best solvents can be formulated by blending different amines to take full advantage of the desirable properties of each solvent. Some observations that can be derived from our survey of past research conducted in the field of CO2 capture include the following: (1) With respect to the objective of improving efficiency of the CO2 capture process, previous research studies rarely focused on using automation for supporting the process of monitoring and control of the CO2 capture system as a means for optimizing the plant performance and enhancing efficiency of the CO2 capture process. Operation of a CO2 capture system is a complicated task because it involves control of over 100 parameters. If these parameters are monitored and controlled effectively, the entire plant can work under desirable conditions and efficiency of the CO2 capture process can be greatly enhanced. (2) Application of artificial intelligence technologies has not been made to the CO2 capture domain. Since operation of a CO2 capture system is extremely complicated, the process operators have accumulated significant knowledge and problem-solving skills over time. This experience is exclusive and hard to develop, and it is desirable to capture and encode the human expertise into a knowledge-based system for documentation and training purposes. Therefore, the objective of this study is to develop a knowledgebased system for monitoring and control of the CO2 capture process. Such a research study would help fill the gap in research for the field of CO2 capture process. 2.2. Inferential modeling technique An important prerequisite for developing a knowledge-based system is to acquire expertise that can be encoded in the knowledge base. For acquiring knowledge on the CO2 capture process, we adopted the inferential modeling technique, which is derived from the inferential model. An inferential model is a generic categorization of knowledge types. It functions as a ‘‘conceptual map’’ to aid the knowledge engineer to identify and classify elements of the elicited expertise (Chan, Tontiwachwuthikul, & Cercone, 1995). Based on this map, the inferential modeling technique or IMT supports ‘‘an iterative-refinement of knowledge elements in a problem-domain that provides top-down guidance on the knowledge types required for problem solving’’ (Chan, Peng, & Chen, 2002). The resulting inferential model consists of the following four levels of knowledge:

Q. Zhou et al. / Expert Systems with Applications 38 (2011) 7935–7946

(1) Domain knowledge consists of objects, attributes, values, and relations. The objects include a set of concrete domain objects. The attributes describe the properties of the objects, which can be defined as a set of functions that receive input values and return output values. The relations describe the relationships among the objects or the attributes. (2) Inference knowledge consists of abstract objects. These inference level objects can be described with inference relations and strength of inferences. The inference relations identify different types of relations among sets of abstract objects; the strength of the inference is associated with each inference relation and represents the relative inferential significance of the relation. (3) Task knowledge consists of a set of procedures or behaviours which are performed to complete a goal. A task is accomplished by means of a method that invokes the domain and inference objects or relations involved in this task. One task can be decomposed into a number of subtasks, and the objective of this task is accomplished by coordinating all the sub-goals. (4) Strategy knowledge is defined as the knowledge used during the diagnostic process to decide what is the most opportune choice to make or, alternatively, to judge if it is worth executing a certain action with respect to other possible actions (Mussi, 1993). The IMT was applied and the templates of domain knowledge and task knowledge were used in the process of knowledge base development for the domain of amine-based CO2 capture process. 2.3. Amine-based CO2 capture process The goal of CO2 capture is to separate CO2 from industrial gas streams before they are released into the atmosphere. The process of amine-based CO2 capture at the International Test Centre for CO2 capture (ITC) can be briefly described as follows: prior to CO2 capture, the flue gas is cooled down and particulates and other impurities such as SOx and NOx are removed as much as possible. The pre-treated flue gas is injected into the absorber column from the bottom, and it contacts solvent that is free of CO2 or lean amine solvent, which is injected from the top of the absorber column. The amine selectively absorbs CO2 from the flue gas. The amine solvent carrying CO2, which is called CO2-rich or rich amine, enters the stripper column, where the CO2 is extracted from the amine solvent and the lean amine solvent is regenerated. The lean amine solvent is returned to the absorber column and used in the CO2 removal process again. The CO2 stream produced is dried and post-treated, and it can be either developed to a food grade quality or pressurized and transported to a suitable site for geological storage. The CO2 capture process is depicted in Fig. 1. 3. Development of a knowledge base The knowledge base in this study was developed in three phases: knowledge acquisition, knowledge analysis, and knowledge representation. In the process of knowledge acquisition, the first author acted as the knowledge engineer and interacted with the domain expert, who is the chief engineer of ITC, to acquire knowledge about problem-solving in the domain. The process of knowledge acquisition lasted 1 year, from January to December 2005. During the phases of knowledge analysis and representation, the knowledge engineer analyzed the verbal information collected from the expert and configured them into a conceptual model. The IMT was applied in knowledge analysis, and the knowledge was

7937

formalized into an inferential model. The IMT decomposed knowledge into the two levels of domain knowledge and task knowledge. 3.1. Domain knowledge Domain knowledge includes three components: the objects, their attributes, and values related to the attributes. The objects in the plant can be classified into two categories: static and dynamic. The static objects include the constructive components of the plant, which can be divided into the three classes of reaction instruments, valves, and pumps. The dynamic objects include the substances that circulate and react in the plant, i.e., the water, amine solvent, and gases. The classification of objects is shown in Fig. 2, and the details are described in the following sections. 3.1.1. Static objects and their attributes The three kinds of static objects including reaction instruments, valves and pumps are discussed in details as follows. 3.1.1.1. Reaction instruments. There are 16 primary reaction instruments involved in the plant. They are grouped into three main classes based on their functions and listed as follows: (1) Pre-treatment section, which includes the steam boiler, micro turbine, inlet-gas scrubber. (2) Absorption-based CO2 section, which includes the absorber, off-gas scrubber, lean amine storage tank, lean amine cooler, rich amine vessel, lean/rich amine exchanger, stripper, reboiler, and reclaimer. (3) Post-conditioning section for product purification, which includes the reflux condenser, reflux accumulator, CO2 wash scrubber, and CO2 dryer unit. The attributes of the reaction instruments include the temperature, pressure, or level of the instruments and the attributes of their output dynamic objects. The details of the attributes of the reaction instruments will be discussed in Section 3.1.2. 3.1.1.2. Valves and pumps. The valves and pumps are manipulated to control the process parameters. Therefore, all the pumps and valves are associated with the attributes of the reaction instrument. In terms of representation in the knowledge hierarchy, the reaction instruments’ attributes are represented one level below the reaction instrument objects. Corresponding to this representation, the pumps and valves are defined as modules in the system design because the modules are one level lower than the plant area in the DeltaV representational hierarchy. Valves: The valves can be categorized into two types based on their control mechanism: PID (proportional-integral-derivative) control valves and solenoid valves. While all the solenoid valves are used for controlling water flow, the PID valves can be subdivided into four groups based on the substances they manipulate: (1) steam supply control valve, (2) amine control valve, (3) water control valve, and (4) gas control valve. All the PID control valves in the plant can be identified by five attributes: the three system attributes of: (1) tag number (the label for a valve/pump), (2) name (the brief description), (3) type (the mechanism of a valve/pump), and two design attributes of (4) location (where the valve/pump is installed in the plant), and (5) distribution flow (the dynamic object which a valve/pump controls). The solenoid valves can be identified by the additional attribute of status, which describes their ON/OFF state under normal conditions. Also, the attribute of distribution flow determines the process parameter controlled by a valve, and the attribute of location

7938

Q. Zhou et al. / Expert Systems with Applications 38 (2011) 7935–7946

Fig. 1. Amine-based CO2 capture process flow diagram.

CO2 Capture Plant

Static Objects

Valves

Dynamic Objects

Pumps

Water

Reaction Instruments

Flue Gas

Solvent Gases

Off Gas

CO2

Rich Amine

Lean Amine

Steam

Fig. 2. Objects in CO2 capture plant.

determines the plant area to which a valve belongs in the phase of system design. Pumps: Pumps include liquid distribution pumps and gas blower pumps. The liquid distribution pumps can be divided into three subclasses according to the types of flow they control: (1) water control pumps, (2) amine control pumps, and (3) chemical flow control pumps. Like solenoid valves, all the PID control valves in the plant can be identified by six attributes including tag number, name, type, location, distribution flow, and status. A sample valve and a sample pump are given in Table 1.

3.1.2. Dynamic objects and their attributes The dynamic objects include amine solvent, water, and gas. The amine solvent can be classified into lean amine and rich amine based on the amount of CO2 it carries. The gases include flue gas (with CO2), off gas (free of CO2), CO2, and steam. All the dynamic objects can be specified by the three attributes of temperature, pressure, and flow rate. Since the dynamic objects circulate and react through the entire process, the values of their attributes are constantly changing. Therefore, a decision was made to identify the properties of dynamic objects at any particular

7939

Q. Zhou et al. / Expert Systems with Applications 38 (2011) 7935–7946 Table 1 Samples of valve/pump. Tag

Name

Type

Distribute flow

Location

Status

FCV-600 (valve) B-200 (pump)

Lean amine to absorber control valve Inlet flue gas blower

PID control valve Gas blower

Amine solvent Flue gas

Between absorber and lean amine storage tank Between flue gas scrubber and absorber

N/A ON

location with the tag of the sensor. Moreover, the knowledge engineer classified the attributes of the dynamic components with the attributes of the reaction instrument from which they flow. For example, the attributes of the lean amine storage tank include the level and temperature of the tank itself, as well as the attributes of its output lean amine. They include: (1) amine storage tank level (DPT-600), (2) lean amine storage tank temperature (TE-640), (3) lean amine to absorber flow rate (FT-600), and (4) lean amine to absorber temperature (TE-600). The attributes of dynamic objects are organized in this way due to three reasons. Firstly, the performance of the reaction instrument directly influences the attributes of its output dynamic objects; hence, it is logical to group the reaction instrument with the output dynamic objects. Secondly, this approach enables straightforward examination of the attributes of dynamic objects at different phases in the process. Thirdly, this organization simplifies the grouping of attributes and facilitates design and construction of the KBSCDC. In this way, over 100 process parameters in the plant can be grouped into 16 reaction instrument groupings. Therefore, the entire system can be viewed in terms of 16 reaction components, their relevant attributes or process parameters, and the relevant valves/pumps. The values of the process parameters are monitored by the system so that if any abnormal value is detected, the relevant pump/valves will be manipulated to remedy the abnormal conditions. 3.2. Task knowledge The objective of the KBSCDC is to maintain normal plant performance so that it produces CO2 at the desired rate. Therefore, the main task of the system is to monitor all the reaction instruments and ensure they operate under desirable conditions. The instruments that have abnormal parameter values indicate that abnormal operating conditions have occurred. The system can monitor the operating conditions of the CO2 capture plant by constantly comparing the measured values with desired parameter values. Deviation from the normal ranges of values triggers an alarm to advise the operator that a problem has occurred. The system then diagnoses the abnormal state and suggests the remedial control actions that would address the abnormal situation. The task of monitoring each reaction instrument includes the subtasks of monitoring its related attributes, i.e., process parameters. Therefore, it is important to obtain the desirable operating ranges of the process parameters. The knowledge engineer identified 25 critical process parameters and their normal operating ranges with the help of the domain expert; two sample parameters of the inlet-gas scrubber and their normal ranges of values are shown in Table 2. Therefore, monitoring of the inlet-gas scrubber consists of the four subtasks of: (1) controlling the flow rate of flue gas into absorber (FT-200), (2) controlling the temperature of flue gas into absorber (TE-201), (3) controlling the wash water flow rate of scrubber (FT-420), and (4) controlling the inlet-gas scrubber water level (LC-410). The two sample subtasks of controlling the water level (LC-410) and controlling the wash water flow rate (FT-420) are discussed here. They are controlled so that the parameter values fall within the values specified in Table 2. If the values should fall outside the normal ranges, the diagnosis and remedial control actions

Table 2 Sample parameters and their normal operating ranges. Tag

Parameter

Unit

Limit

Value

FT-420

Off gas scrubber wash water flow rate

kg/min

High Low

37.0 5.0

LC-410

Off gas scrubber water level control

%

High Low

65.0 5.0

are determined by various conditions. The details of diagnosis and control actions for the sample parameter of wash water flow rate (FT-420) are given in Table 3. If the wash water flow rate (FT420) of the inlet-gas scrubber is less than 5.0 kg/m, a warning is given to the operator. The diagnosis of the situation is that the over low flow rate of wash water could be caused by the closed water circulation pump P-420. Therefore, the remedial control action is to open pump P-420 to restart water circulation between the water tank and the inlet-gas scrubber. However, if P-420 is already open, then the PID valve FCV-420 should be opened to increase water flow.

4. System design and implementation 4.1. System design The intelligent system of KBSCDC was implemented on DeltaV Simulate (a trademark of Emerson Corp., USA). DeltaV Simulate logically decomposes the entire system into plant areas and control modules. It supports various algorithms for implementing process control logic, and it allows the simulation of dynamic processes and real-time monitoring. The implementation of KBSCDC on DeltaV involves a hierarchy of 5 levels: plant area (level 1), module (level 2), algorithm (level 3), function block (level 4), and parameter (level 5). The hierarchy is shown in Fig. 3. The plant areas are logical divisions of the process control system, which can be based on physical plant locations or main process functions. A plant area consists of modules, and each module is a logic control entity responsible for configuring the control strategies. It contains algorithms, alarms, and other characteristics that define the process control. Algorithms define the logic steps that describe how the module behaves and how the tasks are accomplished. In this intelligent system, the function block diagrams (FBD) were used to continuously execute control strategies. The basic component of a FBD is a function block, which contains the control algorithm and defines the behaviour of the module. Each function block contains parameters which are the userdefined data manipulated by the module’s algorithm in its calculations and logic. The five-level hierarchy of the KBSCDC system supports a top-down approach for encoding knowledge into DeltaV Simulate. System construction on DeltaV Simulate can be explained by describing sample components of each level as shown in Fig. 4, and the details are described below. Three sample plant areas include the stripper, inlet-gas scrubber, and absorber. In this discussion, the inlet-gas scrubber is used as an example to illustrate how a plant area is constructed. The

7940

Q. Zhou et al. / Expert Systems with Applications 38 (2011) 7935–7946

Table 3 Diagnosis and control strategies for sample parameter of inlet-gas scrubber. Object

Task

Conditions

Diagnosis and controlling actions

Inlet-gas scrubber

Control wash water flow rate (FT-420)

5.0 kg/min < = FT420 < = 37.0 kg/min FT-420 < 5.0 kg/min

Normal operation No action taken Warning is given Diagnosis: The wash water flow to water tank is low or stopped by the closed water circulation pump-420 Remedial control action: Open P-420 to restart water circulation between water tank and scrubber; otherwise, open up FCV-420 to increase volume of water returning to water tank if P-420 is open Warning is given Diagnosis: The volume of wash water between the off gas scrubber and water storage tank is too high Remedial control action: Close FCV-420 to reduce wash water flow from the scrubber wash water storage tank

FT-420 > 37.0 kg/min

Level 1

Plant Area

Level 2

Module

Level 3

Function Block Diagram

Level 4

Function Block

Level 5

Parameter Fig. 3. Hierarchy of KBSCDC on DeltaV.

plant area of inlet-gas scrubber contains eight modules. Four of these are PID valve control modules for the process parameters of: (1) flue gas flow rate into absorber (FC-200), (2) inlet-gas scrubber water level control (LC-410), (3) temperature of flue gas to absorber (TC-201), and (4) wash water flow rate of inlet-gas scrubber (FC-420). The other four are 2-state control modules for the pumps and solenoid valves of: (1) flue gas blower (B-200), (2) make up water control valve (EV-300), (3) wash water control valve (EV420), and (4) wash water pump (P-420). The algorithm used in the module of wash water flow rate into absorber (FC-420) is represented in a function block diagram, which consists of the function blocks of data input simulation, data output simulation, and primarily a PID control function. Since the KBSCDC system is not connected to the CO2 capture plant at the current stage, the data input and output to the system are simulated by using function blocks. The PID control function block contains the most important parameters, which includes the set-point (SP) of the PID control and alarm activation limits, whose variable names in the system are HIGH_LIMIT and LOW_LIMIT. The details of system construction and implementation are given in Section 4.2. More implementation details about the knowledge hierarchy of the KBSCDC in DeltaV Simulate are explained as follows. Fig. 5 shows how the system knowledge base was developed in the DeltaV system. In Fig. 5, the white boxes on the left side contain the components of the knowledge base. As observed vertically from the top to bottom, the components of the domain knowledge consist of objects, the attributes of the objects, and values of the attributes. The blue or shaded boxes on the right side contain the

components of the DeltaV Simulate. As observed vertically, the components of the DeltaV Simulate are also displayed from the higher to lower hierarchical level from the top to bottom (refer to Fig. 3). More details on how the knowledge components are represented as the components of the DeltaV Simulate at different levels are given as follows: Level 1: The objects of the plant include the reaction instruments, pumps, and valves. The reaction instruments are defined as the plant area in DeltaV Simulate. Since the CO2 capture plant contains 16 reaction instruments, there are together 16 plant areas defined in the system. Level 2: As mentioned in Section 3.1.1.2, the attribute of location of a valve or pump determines the plant area to which a valve or pump belongs in the system design. Each plant area can consist of a number of valves and pumps which manipulate multiple attributes of this plant area. Therefore, the objects of pumps and valves are defined as modules under the level of plant area in the DeltaV Simulate, although they are at the same level of objects as the reaction instruments in the knowledge base. As mentioned in Section 3.1.1.2, the pumps and valves have two different control mechanisms of PID control and two-state control. Therefore, the pumps and solenoid valves based on a two-state control mechanism are defined as two-state control modules; the PID control valves are defined as proportionalintegral-derivative or PID control modules. Level 3: The function block diagram (FBD) is a diagram that contains multiple interconnected function blocks. However, since FBD represents a type of algorithm and is not directly related to the knowledge base, it is not shown in Fig. 5. Level 4: A function block is a logic processing unit that defines the behaviour of an algorithm for a particular module. Two types of function blocks are available: the two-state function block and the PID control function block. At this level, the attributes of the objects in the knowledge base are analyzed. As mentioned in Section 3.1.1.1, the attributes of a reaction instrument include the attributes of the reaction instrument and the attributes of its output dynamic objects, such as the pressure and temperature. Since the PID control valves are used to control the attributes of the reaction instruments and the dynamic objects, the attributes of reaction instruments and dynamic objects and their relative PID control valves are combined into the PID control function blocks, which enable the present values of the attributes to approach their desired values by controlling the PID valves. As mentioned in Section 3.1.1.2, the pumps and solenoid valves have another important attribute of status, which manipulate the attributes of the reaction instruments by switching between the ON/OFF states. Therefore, the pumps

7941

Q. Zhou et al. / Expert Systems with Applications 38 (2011) 7935–7946

DeltaV Hierarchy

Plant Stripper

Inlet-Gas Scrubber

Absorber

Plant Area Module

FC-200

LC-410

TC-201

FC-420

B-200

EV-300

Function Block Diagram

Data Input Simulation

PID Control

HIGH_ LIMIT

LOW_ LIMIT

EV-420

P-420

Algorithm

Data Output Simulation

….

SP (Setpoint)

Function

Parameter

Fig. 4. Implemented hierarchy of the KBSCDC in DeltaV Simulate.

and solenoid valves are constructed by two-state control function blocks, which enable the pumps and solenoid valves to switch between their ON/OFF states to reach their desired state. Level 5: At this level, the values of the attributes of the objects are analyzed. For the object of reaction instruments, the important attributes include their related operation properties, such as temperature, pressure, etc. Therefore, the values of these attributes are the high/low limits that specify the limits beyond which the present value of the parameter is outside the normal (i.e. desired) operating range and an alarm is activated. For the object of the PID control valves, the value of the attribute of setpoint is critical as it defines the desired value for the PID control. Since the attributes of the reaction instruments and the dynamic objects and their relative PID control valves are combined into the PID control function blocks, their individual values including the high/low limits of the attributes of the reaction instruments and the dynamic objects, and the setpoint values of the PID valves, are specified as critical parameters in the PID control function block. Since the solenoid valves and pumps are used to control the process parameters by switching between their ON/OFF states, the ON or OFF are the values of the attribute of status for the pumps and solenoid valves. Since the pumps and solenoid valves are constructed by two-state control function blocks, their ON/OFF status are defined as an important parameter in the two-status control function blocks. The details of system implementation are given in Section 4.2. 4.2. System implementation The DeltaV Simulate consists of three applications of DeltaV Explorer, Control Studio, and DeltaV Operate. Each of the applications conducts a specific function in the construction of the intelligent system. The DeltaV Explorer is a navigation tool for the user to construct and view the overall system, i.e., the hierarchy of plant areas and modules. The Control Studio is used to design and graphically create the individual modules and templates that make up the control strategy. The DeltaV Operate functions in two modes

of configure and run. In the configure mode, it is used to build process graphics and develop the user interface. In the run mode, it allows the operator to monitor and control the process by interacting with the system through the user interface. 4.2.1. Plant areas The division of plant areas and specification of modules were completed in DeltaV Explorer, as shown in Fig. 6. The overall structure of the system, containing 16 reaction instruments or plant areas can be viewed in the left panel. One sample plant area of ‘‘inlet-gas-scrubber’’ is selected and the eight modules in this plant area are shown in the right panel. These include: (1) flue gas flow rate into absorber (FC-200), (2) inlet-gas scrubber water level control (LC-410), (3) temperature of flue gas to absorber (TC-201), (4) wash water flow rate (FC-420), (5) flue gas blower (B-200), (6) make up water control valve (EV-300), (7) wash water control valve (EV-420), and (8) wash water pump (P-420). These modules correspond to the modules shown at the module level in Fig. 4. 4.2.2. Control modules A control module is constructed by combining various function blocks. The function blocks are connected with graphic wires, and data values pass through the wires from one block to another. The function blocks are executed in a logical order, which determines the behaviour of the module. The module of FC-420 in the plant area of ‘‘inlet-gas-scrubber’’ is used to illustrate the construction of the PID valve control modules. In Fig. 7, the FC-420 module is selected, which opens the list of function blocks listed in the top left panel and illustrated graphically in the right panel. The function block diagram of module FC-420 shown in the right panel of Fig. 7 is expanded for clearer illustration in Fig. 8. It can be seen that the function blocks are connected to perform the control strategy. The details are described as follows. The control module of FC-420 contains the following function blocks: math multiply (MLTY), addition (ADD), analog control signal generator (SGGN), filter (FLTR), rate limit (RTLM), PID function (PID), and analog output (AO) blocks. The numbers at the bottom right of each function block show the execution order of the

7942

Q. Zhou et al. / Expert Systems with Applications 38 (2011) 7935–7946

Level 1

Reaction Instruments

Plant Areas

Level 2 Objects

Pumps

Two-state Control Modules Solenoid

Valves PID control

PID Control Modules

Reaction Instruments Level Pressure

Level 4

Domain Knowledge Attributes

P ID Control Function Blocks

Dynamic Objects Temperature Pressure Flow Rate

Pumps Status Two-state Control Function Blocks Solenoid Valves Status

High/LowLimits, SP

Level 5 Values

Parameters ON/OFF Status

Fig. 5. Design of knowledge-based system on DeltaV Simulate.

function blocks. If the execution involves the data being passed from the function block with a lower number to the one with a higher number, this is called ‘‘forward execution’’, and the connection wire is represented by a solid line between the two function blocks. If the execution is backward, i.e., the data passes from the function block with a higher number to the one with a lower number, the connection wire is represented as a dotted line. The input generated for module FC-420 is passed to function block FT-420, which sends the input to PID1 for processing. The control signal generated by PID1 is sent to the water control valve FCV-420. This is shown in Fig. 8. However, as the KBSCDC system is not connected to the field bus devices, the analog input is simulated using the function blocks of SGGN, ADD and FLRT. The SGGN function block uses a specified combination of a sine wave, a square wave, and a bias value to generate an output signal that simulates a process signal (wire 1 in Fig. 8). The simulated input signal is summed with the feedback signal from the AO function block (wire 2) through an ADD block and generates an actuating signal, which is sent to a FLTR block (wire 3). The FLTR block filters large changes in this signal and generates a smooth output signal (wire 4). The output signal is sent to the RTLM block, which limits the change rate of the signal to specific limits and maintains stability of the signal. The output signal of the RTLM block is input to the PID control block (wire 5), and the PID function block performs the functions of analog input/output processing, proportional-integral-derivative (PID) control, and alarms. The PID function block sends a signal to the AO block and updates

the present value (wire 6). The AO block generates a control signal that indicates the valve position of FCV-420, which enables the present value of FT-420 to approximate the set-point value as much as possible. The feedback signal of the AO block is wired back to the PID block to provide a bumpless transfer when the mode changes in the PID control block (wire 7). As well, the feedback bias signal from the AO block is sent through the MLTY block (wire 8) to the SGGN block (wire 9). Since FT-420 is not only controlled by the valve FCV-420, but also controlled by the pump P-420, the status of P-420 is involved in this module as an external parameter (the two-state module for P-420 is separately developed). The status of P-420 is connected to the SUB and MLTY blocks (wire 10, 11, 12), sent to the RTLM block (wire 13, 14) and then sent to the PID control block (wire 5). In this way, once the status of P-420 is changed (0–1 or 1–0), the present value of FT-420 will be changed accordingly and the impact on FT-420 due to the status change of P-420 can be simulated. The details of the PID function block are explained in the next section. 4.2.3. Function blocks and parameters One set of parameters contained in the PID control block is execution parameters: (1) PV: present value, (2) SP: set point of PID control, (3) Gain: the normalized proportional gain value, (4) Reset: the integral action time constant and (5) Rate: the derivative action time constant. The PID loop control is implemented by taking the error difference between SP and PV and calculating a control output signal using the PID control execution.

Q. Zhou et al. / Expert Systems with Applications 38 (2011) 7935–7946

7943

Fig. 6. Overall structure and modules in plant area ‘‘inlet-gas-scrubber’’.

Since the PID block supports alarm limit detection, the other set of parameters in the PID block are for alarm detection, which is defined by the desirable operating ranges of the process parameters. There are three classes of parameters for alarm detection: ALARM_TYPE, ALARM_LIMIT, and ALARM_ACT. They are shown in Fig. 9 and explained as follows. ALARM_TYPE specifies two types of alarms for warning conditions: HIGH_ALARM and LOW_ALARM. ALARM_LIMIT specifies the alarm limits or the values at which the alarms are activated. Based on the two types of alarms, there are two alarm limits: HIGH_LIMIT and LOW_LIMIT. The system compares these limit values to the present value to determine if an alarm condition exists. ALARM_ACT is in a true/false (1/0) state to specify the alarm condition state. Corresponding to ALARM_LIMIT that includes HIGH_LIMIT and LOW_LIMIT, ALARM_ACT includes HIGH_ACT and LOW_ACT. When any alarm condition exists or when the present value exceeds any values of ALARM_LIMIT, ALARM_ACT is set to true to indicate that the alarm is activated. For example, the desirable operating range of FT-420 is between 5.0 and 37.0 kg/min. Therefore, the high limit (HIGH_LIMIT) of FT-420 is 37 and the low limit (LOW_LIMIT) is 5. If the present value (PV) of FT-420 is higher than 37.0 kg/min (HIGH_LIMIT), HIGH_ ACT becomes true and the alarm for HIGH condition is activated. If the present value (PV) of FT-420 is lower than 5.0 kg/min (LOW_LIMIT), LOW_ACT becomes true and the alarm for LOW condition is activated. If the present value is neither higher than 37.0 kg/min nor lower than 5.0 kg/min, neither HIGH_ACT nor LOW_ACT becomes true and no alarm is activated. The alarm is displayed on the user interface to remind the operator of the abnormal condition. The desirable operating ranges of the process parameters are not only used for alarm activation, but also trigger the diagnosis displays, which is the message box on the left top corner of screen shown in Fig. 10.

5. Case study Fig. 10 describes a scenario in which the inlet-gas scrubber wash water flow rate (FT-420) is in the abnormal condition. The current value of FT-420 shown on the panel of FC-420 is 3.2 kg/ min, which is lower than its low limit of 5 kg/min. The alarm is activated and displayed on the interface, so that the panel for FC420 is turned to the blue color. The diagnosis and control suggestions are sent to a message board on the user interface: ‘‘If P-420 is on, open up FCV-420 to increase water returning from the scrubber to water tank; if P-420 is off, turn off P-420’’. The green colour of pump P-420 indicates its open status. Therefore, the remedial controlling action should be applied on the PID control valve FCV-420. By pushing the control button of module FC-420, the faceplate of FC-420 is displayed on the user interface, as shown in Fig. 11. The faceplate mainly contains two bar sliders. On the right is the PV (present value) slider, which indicates the present value of FT-420. The two yellow arrowheads vertically positioned relative to the PV slider indicate the high and low alarm limits of the present value. The large, white arrowhead is the SP (set-point) indicator. To reset the set-point, the indicator is moved up to a desired set-point value, in this case at 25.0 kg/min. This reset value brings the present value of FT-420 to the normal operating range of between 5.0 and 37.0 kg/min, in this case 23.8 kg/min, which is displayed on top of the PV slider as the new present value of FT-420. The output bar slider on the left indicates the output value of the PID loop. On top of the left slider is displayed the output value of FC-420. In this case, it shows the valve position of FCV-420 is 26.2% open. Fig. 12 shows the interface after the normal operation of FT-420 is restored. The background panel for FC-420 has returned to its normal black colour, and the current value of FT-420 shown on

7944

Q. Zhou et al. / Expert Systems with Applications 38 (2011) 7935–7946

Fig. 7. Construction of PID control module FC-420.

Fig. 8. The connections of function blocks in module FC-420.

Q. Zhou et al. / Expert Systems with Applications 38 (2011) 7935–7946

HIGH_ ALARM

HIGH_ LIMIT

HIGH _ACT

LOW_ ALARM

LOW_ LIMIT

LOW_ ACT

ALARM TYPE

7945

KBSCDC system helps to enhance performance and efficiency of the CO2 capture system by dramatically reducing the time for problem diagnosis and resolution when abnormal operating conditions occur. The system also can be used as a decision-support tool for inexperienced operators and aids them in controlling the plant; it can be used for training novice operators. The knowledge base can be extended or refined with future additions of knowledge.

Fig. 9. Parameters for alarm activation.

the panel of FC-420 indicates that the water circulation rate has increased to 25.0 kg/min, which is within the normal operating range. As a remedial action has been applied to address the problem which triggered the alarm, the alarm and the diagnosis message have disappeared from the top left corner of the interface.

6. Discussion and conclusion In the present study, a knowledge-based expert system for monitoring and control of the CO2 capture process was developed. To develop the system, the IMT was applied to analyze the domain knowledge and problem-solving techniques and a knowledge base was established. The KBSCDC system was implemented on DeltaV Simulate, which provides features useful for development, such as graphic displays, alarms, and pop-up menus. Compared to G2, DeltaV Simulate provides function blocks, which supported the configuration of control mechanisms and definition of the specific parameters for different valves and pumps, such as set-point value of a PID control valve and status of a two-state pump. Therefore, the tool of DeltaV Simulate enabled us to simulate the control strategies applied to the vales and pumps and the influence of the abnormal conditions caused by these control actions. The

Fig. 10. System interface for case study.

Fig. 11. Faceplate for control FT-420.

7946

Q. Zhou et al. / Expert Systems with Applications 38 (2011) 7935–7946

Fig. 12. Case study back to normal operation.

The current version of the KBSCDC system has two disadvantages. Since there are altogether 16 components involved in the process, an abnormal condition can be caused by incorrect performance of more than one component or parameter. However, the system can only deal with abnormal operation of one component at a time. Moreover, the knowledge captured for system diagnosis and control represents the problem-solving expertise of only one expert operator. It is likely that other senior operators would conduct problem solving in a different manner. To address these weaknesses, future development of the system will focus on enabling monitoring, control, and diagnosis of multiple malfunctioning components or parameters and integrating expertise from additional operators to improve the capability and reliability of the system. Acknowledgements We are grateful for the generous support of Natural Science and Engineering Research Council (NSERC) and the Canada Research Chair Program, and the help of Dr. Raphael Idem and Don Gelowitz. References Chakma, A. (1997). CO2 capture processes-opportunities for improved energy efficiency. Energy Conversion and Management, 38, 51–56. Chan, C. W., Peng, Y., & Chen, L. L. (2002). Knowledge acquisition and ontology modeling for construction of a control and monitoring expert system. International Journal of Systems Science, 33, 485–503.

Chan, C. W., Tontiwachwuthikul, P., & Cercone, N. (1995). Knowledge engineering for a process design domain. International Journal of Expert Systems, 8, 47–76. Idem, R., Wilson, M., Tontiwachwuthikul, P., Chakma, A., Veawab, A., Aroonwilas, A., et al. (2006). Pilot plant studies of the CO2 capture performance of aqueous MEA and mixed MEA/MDEA solvents at the University of Regina CO2 capture technology development plant and the boundary dam CO2 capture demonstration plant. Industrial & Engineering Chemistry Research, 45, 2414–2420. Mussi, S. (1993). A method for putting a strategic common sense into expert system. IEEE Transactions on Knowledge and Data Engineering, 5, 369–385. Sholeh, M., Svendsen, H. F., Karl, A. H., & Olav, J. (2007). Selection of new absorbents for carbon dioxide capture. Energy Conversion and Management, 48, 251–258. Tontiwachwuthikul, P. (1996). Research and development activities on high efficiency separation process technologies for carbon dioxide removal from industrial sources at University of Regina, Canada. Energy Conversion and Management, 37, 935–940. Veawab, A., Aroonwilas, A., Chakma, A., & Tontiwachwuthikul, P. (2001). Solvent formulation for CO2 separation from flue gas streams. In Proceedings of the first national conference on carbon sequestration, Washington DC, USA. Veawab, A., Tontiwachwuthikul, P., & Bhole, S. D. (1997). Studies of corrosion and corrosion control in a CO2-2-amino-2-methyl-1-propanol (AMP) environment. Industrial & Engineering Chemistry Research, 36, 264–269. Veawab, A., Tontiwachwuthikul, P., & Chakma, A. (1999). Corrosion behavior of carbon steel in CO2 absorption process using aqueous amine solutions. Industrial & Engineering Chemistry Research, 38, 3917–3924. White, M., Strazisar, R., Granite, J., & Hoffman, S. (2003). Separation and capture of CO2 from large stationary sources and sequestration in geological formationscoalbeds and deep saline aquifers. Journal of the Air & Waste Management Association, 53, 645–715. Zhou, Q., Chan, C. W., & Tontiwachiwuthikul, P. (2009). A monitoring and diagnostic expert system for carbon dioxide capture. Expert System with Applications, 36, 1621–1631.