Realtime Knowledge Based Systems RTKBS

Chemical process industries (CPI) require a high level of supervision in realtime. Supervision tasks may include scheduling processing stages, supervising data acquisition, distributed control systems, and alarm management. This means low level process operations such as adjustment of PID control settings and high level qualitative decisions such as implementing different operational policies and fault handling are to be dealt with together. All these activities are realized with accumulated expertise over the years. Experience of process operators and engineers is an invaluable asset and should be incorporated in an automated supervisory system. Real-time knowledge-based systems (RTKBS) provide such an environment where a high level automated process supervision can be achieved.

KBSs have been one of the rapidly growing applications of Artificial Intelligence (AI) in the scientific and engineering arena during the last two decades. A KBS is a computer program that emulates the decision-making ability of a human expert. The terms KBS and Expert Systems are often used synonymously. In this book we will use the term KBS.

Figure 8.11 illustrates the general framework and the common components of a KBS [191]. The knowledge-base contains facts, rules and heuristics that are used by the inference engine to draw conclusions. The user interfaces with the KBS by using an interface to input information or learn the conclusion reached by the KBS. Many algorithms have been proposed for inferencing by AI researchers [502, 428]. In the context of KBSs for

Figure 8.11. Basic structure of a KBS.

supervising process operations, forward-chaining is used to update all measurement information and derived variables when new process information is made available to the KBS. Then, backward-chaining is used to infer the status of process operation (normal or faulty), diagnose the source cause of abnormal operation, and formulate the proper intervention.

The process of building a KBS, developing and implementing a problemsolving strategy is called knowledge engineering. The knowledge of an expert or a group of experts is transferred into KBS by knowledge engineer. The customary way of performing this task is to repeat a cycle of interviewing the expert(s), constructing a prototype, testing and re-interviewing that is a very time consuming and laborious task. This task is called knowledge acquisition and elicitation [221], Recently, more effective systematic techniques have been proposed for knowledge acquisition and elicitation [193] and techniques for automatic knowledge acquisition have been suggested [46, 513]. In the early days of the technology, knowledge engineers had to develop the entire system from scratch by using one of the available AI programming languages such as LISP (LISt Processing language), Prolog (Programming in Logic), Smalltalk, OP5 (its current version is OP83) and NASA's CLIPS software. Today's KBS development software such as Gensym's G2, make the development easier. One of the major differences between conventional programming languages such as FORTRAN and C and AI programming languages is that the former rely on the numbers and algorithms while the latter are designed over symbols, lists and searches. KBSs use inferences to achieve a reasonable solution that is the best that can be expected based on data, facts and rules currently available, in contrast to a numerical optimization approach based on an objective function, process model, constraints to equations, and numerical optimization algorithm.

Several types of knowledge are used in a KBS. Most of the early KBSs for CPI are developed using shallow knowledge which is based on empirical and heuristic knowledge [303, 314, 492, 617]. Heuristics are rules of thumb or empirical knowledge gained by experience which may provide a quick solution to a specific problem by relating the symptoms with causes without using a system model. Deep knowledge is based on the basic structure, function and behavior of a process such as underlying physiological phenomena about microbial activities in fermentations (a process model in a mathematical form). Shallow and compiled knowledge provide the basis for various kinds of knowledge. Rules formed from information derived from deep knowledge are called compiled knowledge and rules derived from shallow and compiled knowledge are called rule-based knowledge. Several KBSs have been proposed for FDD based on these knowledge abstractions in CPI [467, 619],

Rule-based, knowledge representation became dominant in early KBSs developed in many fields. The initial number of rules to be retained in the KBS depends on the complexity of the process and the amount of knowledge acquired from experts. One of the advantages of a KBS is that rule library can be expanded by adding new rules as they become available. As the number of rules increases, the use of information becomes challenging. Some commercial KBSs had tens of thousands of rules for knowledge representation and inferencing, putting on an enormous burden on the execution of the software. Recent KBS shells such as G2 of Gensym Inc., and Nexpert Object of Neuron Data have adopted a hybrid structure based on object based systems. The object framework is used to develop classes, objects, and instances to represent knowledge, and rules are used for inferencing. This results in significant reduction in the number of rules and increase in computation speed. Rules are conditionally true and can be cast into IF-THEN statements such as

If the substrate consumption rate is lower than that expected and the fermentation is in the fed-batch operation mode Then the flow of substrate feed rate is high.

Object-based knowledge representation is another technique in which field of knowledge representation, the object is the central notion. The design of G2 knowledge representation relies on this technique. Knowledge is here expressed by means of two kinds of objects: (1) classes (which describe families of individuals), and (2) instances (which describe the individuals). Classes are organized in hierarchies by a specialization relation upon which an inheritance mechanism is settled. This mechanism allows a more specific sub-class to inherit from all the properties of it super-class it does not redefine. Inference mechanisms are also proposed in order to complete knowledge; default value, classification, and procedural attachment. Classification is a central mechanism which determines for an instance the set of sub-classes of its current class to which it also could be linked. Procedural attachment consists in specifying a peace of code to be executed in order to obtain the value of a property in a class, if needed [183, 184].

A number of KBS are proposed in early nineties for knowledge-based (KB) process control and control systems design in CPI [43, 56, 274, 601]. KB control technologies are also proposed for bioprocesses and fermentation industries [1, 27, 236]. A review of knowledge-based control systems for fermentations is given by Konstantinov and Yoshida [288] where they summarize the functions of a supervisory KBS for fermentation control as

1. Input data validation. KB is structured such that contradictory measurements with respect to previous fermentation can be identified.

2. Identification of the state of the cell culture. On-line detection and evaluation of physiological phase of the cell population is one of most challenging tasks to be achieved. Phase specific control activities can then be performed.

3. Detection and diagnostics of instrumentation faults. Instrument measurements should be closely monitored and failures detected/diagnosed by the KBS.

4. Supervision of conventional control. Phase detection type of high level decisions are used to change low level control parameters.

5. Communication with user. KBS should be able to inform the user (operator) about the process, explain its activities and give advice.

6. Plantwide supervision and scheduling. KBS should be extended to perform supervision and scheduling activities for upstream and downstream processes.

On-line estimation of infrequently measured variables and prediction of product quality variables can also be added to the list above. Achieving all of these tasks in a KBS environment can be realized with the combined use of different techniques. For example, ANNs (Section 4.6), are integrated with fermentation KBS for estimating state variables such as biomass concentration [28, 490].

A variety of RTKBS applications can be found for bioprocesses including novel interface design [580], extended Kalman filter integration for on-line state estimation [399], use of qualitative physics for behavior monitoring [574] and simple rule-based intensive designs [205]. Successful development and implementation of RTKBS using Gensym's G2 for supervision of industrial fermentation plants are reported [13, 14]. This application which contained approximately 300 rules and integrated with large plant databases is credited for increasing the plant yield by 4%, reducing process variability by more than 10% and saving more than 10 production fermentation batches from total loss over a period of two years [14].

Fuzzy set theory (widely known as fuzzy logic (FL)) has also received attention during the last decade in control applications and integrated with RTKBSs [466, 543]. A brief introduction of FL is presented next in conjunction with its use in fermentation technologies and RTKBS integration. Techniques and applications for using integrated ANN-FL controllers have also been reported [280, 293, 342],

Fuzzy logic and its integration with KBS for supervision of fermentation processes

FL is inspired by the way human thinking deals with inexact uncertain information and can be interpreted as the generalization of classical set theory. In classical set theory, a set is comprised of a finite or infinite number of elements belonging to some specified set called universe of discourse. An element x of the universe of discourse (U) may belong to a set A which is included in U so that

where the membership function (or characteristic function) is defined as (Figure 8.12(a))

This is a crisp or Boolean description. FL is based on Fuzzy Set Theory which was introduced by Zadeh in 1965 [680]. A fuzzy set is a generalization of a classic set so that it allows the degree of membership for each element in a range over, say closed unit interval [0, 1]. A fuzzy set A (also called as support set) in the universe of discourse U can be defined as a set of ordered pairs,

where ¡i(x) is called the membership function of set A and it maps each element of the universe of discourse to its range space that is the unit interval in most cases [342]. A variety of membership functions illustrated in Figures 8.12(b) and 8.12(c) can be used.

Consider the classical Boolean description of the level of temperature in a fermenter: the temperature is high or low based on a reference point. In contrast, FL defines vague qualifiers such as quite high, very high, rather high, rather low, very low, quite low on temperature. Figure 8.12(d) illustrates how the linguistic (fuzzy) variable 'temperature' is mapped for a few of its values onto the universe of discourse (temperature scale in this example) for a range of [0, 100 °C] through linguistic descriptors and their assigned values. For the fuzzy value VeryLow for instance, the mapping is described in terms of a set of positive integers in the range [0, 100 °C}. The support set (A) expresses the degree to which the temperature is considered VeryL0W over the range of all possible temperatures in discrete values specified in degrees Centigrade using Ha(x) suc^ that ma(0) = Ma(5) = 1, Mi(10) - ^(15) = 0.8 /^(20) = ^(25) = 0.6, ^(30) = Ma(35) = .. • ,^(100) = 0. (8.116)

(a) Function of a Boolean set (b) Function of a triangular fuzzy set

(a) Function of a Boolean set (b) Function of a triangular fuzzy set

(c) Function of trapezoidal (d) The linguistic variable temperature and some fuzzy sets of its values

Figure 8.12. Examples of membership functions and mapping of a linguistic variable (d) [280].

This relation is usually represented in the following more compact form n

A = nA(xi)/xi + nA{x2)/x2 + ... + fJ.A(xn)/xn = HA(Xi)/Xi (8.117)

where '+' denotes the union of elements (/ does not indicate division) and )iA{xi) is the grade of membership of .x't for n membership values. For the temperature example, Eq. 8.116 becomes

A = 1/0 + 1/5 + 0.8/10 + 0.8/15 + 0.6/20 + 0.6/25 + ... + 0/100. (8.118)

Theory and applications of FL, and integration of KBS and FL for control of fermentation processes are discussed in the literature [285]-[422].

Control of Fermentation Processes using an Integrated KBS-FL System. Konstantinov and Yoshida proposed a methodology (Figure 8.13) for detection of the physiological state of fermentation and control of bio-processes based on expert identification of the physiological state of a cell population [285, 286, 287] using FL, RTKBS and temporal reasoning [465], The physiological state (PS) vector (x) is defined quantitatively by a set of on-line measured variables (u, y) such as ammonia flow rate and substrate feed rate that are used to calculate variables such as specific oxygen to substrate consumption rate, forming the physiological state-space of the culture. Based on the practical experience on the process, a finite number of physiological situations (PSN) are defined where the physiological characteristics of the cell population and its reactions to different control actions are well known. The description of PSNs is mostly in qualitative terms such as "situation of optimal productivity." Hence, PSNs can be interpreted as fuzzy variables. When the process passes from one state to the next, it often exhibits variation in structure behavior and therefore proper alteration in control strategies is required for each state. Adaptive weighting of the membership functions is also introduced to give more importance to certain physiological states. The synthesis of physiological recognition algorithm consists of the development of a decision procedure in which qualitatively defined PSNs are related quantitatively to PS vector x. State recognition is performed by means of expert decision rules using fuzzy sets defined over PSNs. An example of a decision rule relating the current PS vector x and PSNs is:

If xi is high and x2 is low

Then the current x belongs to PSNi with the possibility Mi = 1. The fuzzy values used in the rules are described by fuzzy sets in the general form given in Eq. 8.117 leading to the following system of nonlinear decision functions w/x(x) = M (8.119)

where w denotes the matrix of weights, /x(x) matrix of fuzzy membership functions and M vector of possibilities for the recognition of the current PS as an element of PSNi. Mi is a real number in the range [0, 1] and equal to WjHijixj) = Mi [285]. Once the physiological state is determined, RTKBS uses another rule-base to decide on switching to appropriate control algorithm:

If PSN is PSNi

Then activate control algorithm aj where the control action is defined as Uj = aj(y, x).

They have also proposed variants of this methodology by developing temporal shape libraries for real-time detection of physiological phenomena in a KBS framework [289].

Figure 8.13. The structure of physiological state control system [285].

8.4.2 Real-time Supervisory KBS for Process Monitoring and FDD

KBS applications discussed in Section 8.4.1 for FDD and supervision of fermentation processes lacked multivariate statistical inference. MV statistical techniques are found to be very suitable for on-line SPM and FDD of fermentations processes as discussed in detail in Chapter 6 and Sections 8.1 and 8.2. There is a growing interest in the use of MV techniques in fermentation process modeling, monitoring and FDD [199, 248, 333, 608]. The synergistic integration KBS and MSPM tools offers advantage. Integrating MSPM and RTKBS enables the automated interpretation of MV charts during the abnormal situations and relate this information with process knowledge. The basic structure of the overall integrated framework based on Gensym's G2 KBS development environment is given in Figure 8.14 [184],

Research on developing integrated KBS-SPM for process supervision and FDD progressed during the last decade. Norvilas et al. proposed an intelligent SPM framework by interfacing KBS and MV techniques [337, 436] and demonstrated its performance with simulation studies. Integrated use of MSPM techniques and RTKBS for real-time on-line monitoring and FDD of fermentation processes is proposed by Undey et al. [607] and Glassey et al. [193]. Applications of the integrated RTKBS and MSPM techniques are also reported by industrial researchers [11]. Most of the recent applications are developed using Gensym's G2 software. G2 offers a graphical, object-oriented environment for creating intelligent applications that monitor, diagnose, and control dynamic events in on-line and simulated environments. It features a structured natural language for creating rules, models, and procedures. G2 includes concurrent execution of rules and procedures and the ability to reason about behavior over time. Communication between

Fermentation Process Rule-Base

Multivariate Statistics Rule-Base

G2 Knowledge-Base

Figure 8.14. The structure of integrated KBS [607],

G2 and external programs such as executable C routines and real-time data systems including relational databases and distributed control systems is realized by using G2 Bridge Products provided by Gensym [185].

Recently, Undey et al. [609] have extended the integrated RTKBS framework by including process landmark detection and time alignment for a more refined SPM and FDD in their work for fed-batch fermentation processes. G2 knowledge-base is comprised of two kinds of rule-bases: (1) Fermentation process rule-base, where fermentation specific rules are stored such as physiological phase related heuristics, (2) Multivariate statistics rule-base, where interpretation about the MV charts are stored. The first rule-base is process-specific, hence the level of knowledge depends on the knowledge acquired from process experts, while the second rule-base is process-independent so that it can be used for different types of batch processes. Examples of statistical and process related rules are, respectively

IF the T2 chart is out-of-control

Then start checking T2 contribution plots and identify faulty variables whose contributions exceed contribution limits.

If the glucose feed rate is diagnosed as out-of-control and fermentation is in the fed-batch operation mode then Check the condition of the glucose feed pump.

These two rule-bases are also connected to an external database where process related data such as historical data sets and reference batches as well as statistical limits and parameters are stored. There is a continuous flow of information between the G2 KBS and external databases. MV statistical algorithms are first developed in Mathworks' MATLAB software environment because it allows faster prototyping. A number of software bridges are created by using G2 Standard Interface (GSI) to provide communication between the KBS and external statistical modules. Since GSI bridge development requires C code, the Matlab functions developed are compiled into C functions. Performing statistical calculations outside the G2 KBS environment allows faster execution and better computational performance.

Detection of abnormal process operation can be performed on-line in real-time by implementing one of the on-line SPM techniques discussed in Section 6.5. In this application, AHPCA technique (Section 6.5.2) is preferred due to its superior computational performance and elimination of the need to estimate future values of variable trajectories. Fault diagnosis is performed by means of contribution plots and inferencing. Statistical limits for variable contributions to T2 and SPE are calculated as discussed in Section 8.1. When an out-of-control signal is observed on either T2 or SPE charts, the corresponding contribution plots are investigated automatically and variable(s) exceeding control limits are diagnosed as major contributor(s) to the abnormal situation. The RTKBS helps operators on this interpretation. Once the MSPM rule-base detects and diagnoses the abnormal situation, the process specific rule-base is activated to further investigate the problem by emulating the reasoning of the human expert.

The penicillin fermentation simulator developed based on the unstructured mathematical model discussed in Section 2.7.1 is integrated with the RTKBS as a test-bed. It is run as an external C executable to provide fermentation data in real-time through another GSI bridge. G2 also allows a nice representation of the process flow chart. Each processing unit can be interpreted as objects that can inherit a general class information. Changes in the process operation can also be animated by means of changing color schemes of objects such as active pumps or showing attributes on each object such as temperature value in the fermentor next to fermenter object. A typical fed-batch fermentation for penicillin production flow chart is developed in a G2 workspace as a part of integrated G2 RTKBS (Figure 8.15). Figure 8.16 shows a case where a small downward drift is introduced in glucose flow rate. First, the RTKBS uses its statistical inference rule-base to detect the out-of-control situation and reports it along with the time of its occurrence. The RTKBS continues to use the statistical rule-base to find out the responsible variable(s) by analyzing contribution plots. Based on the contribution limit violations, a conclusion is reached on the responsible variable(s). At this point process expertise is required. Hence, the RTKBS turns to process specific rule-base to further investigate the situation and generate some advice to isolate the problem. In the example shown in Figure 8.16, the process rule-base is used by the RTKBS to infer that the problem is with the glucose feed. The RTKBS also checks certain variables that are highly correlated with glucose feed such as glucose and biomass concentrations in the fermenter to verify its conclusion. Since these variables are also affected (determined by analyzing their contribution values), the certainty about a potential glucose feed failure is high and these findings are reported to the operator by displaying them on the monitor. The messages include the time of detection of the deviation, the input variable(s) responsible, and the process variable(s) affected by the

Data Monitoring Messages Displays Help Batch Age: 129.0 hr Exponentiel Growth Phase

Was this article helpful?

0 0

Post a comment