(125f) Comparison of Decision Fusion Strategies for Combining Heterogeneous Diagnostic Fault Classifiers

Introduction:

Diagnosis of process faults in chemical processes has been an
active area of research for several decades. Successful identification of
process faults at an early stage can increase the rate of fault recovery during
operations and prevent unnecessary shutdowns. Also, automatic detection and diagnosis
of faults are necessary to prevent costly accidents by providing time critical diagnostic
information to plant operators. In the literature, several fault diagnosis
methodologies have been proposed for fault detection and identification (FDI) [1-3].
Each FDI method has its strengths and shortcomings that are process dependant.
A method that works well under one circumstance might not work well under
another when different features of the process come to the fore. For instance,
the multifaceted operations of a process shown under different operating modes
and transitions often complicate the process of fault diagnosis, as different
types of data analysis methods might be required under different mode of
operations. In previous work, we proposed a multi-agent framework that combines
heterogeneous types of FDI methods and allows collaboration among these methods
[4]. Each FDI method is represented as an agent in a multi-agent environment. FDI
results from the different agents are integrated by a knowledge-based consolidator
agent. However, this approach for evidence aggregation is difficult to maintain
when a large number of FDI methods are used. A more efficient means of decision
fusion is necessary to systematically resolve conflicts among the agents. In
this paper, we perform a detailed comparison of various approaches for
combining decisions from heterogeneous FDI methods.

Decision fusion methods:

We compare strategies for decision fusion based on (1) voting
theory, (2) Bayesian-combination theory, and (3) Dempster-Shafer theory when
implemented in a multi-agent environment.

Voting strategy: Voting-based technique is the most
commonly used method to combine decisions. When predictions are obtained from
fault classifiers, the predictions from all classifiers can be counted as votes
with a majority or plurality decision rule adopted for fusion.

Bayesian strategy: Bayesian-based technique is a
popular technique in evidence gathering and uncertainty reasoning. The method
proposed by Xu et. al, (1992) [5] is considered here to generate a prior
probability for each fault classifier. FDI results from each fault classifier
to different classes of faults are first studied offline and its results collated
into a confusion matrix. Posteriori probabilities of faults are then computed
by the consolidator agent  for online decision support based on the Bayesian
rule.  

Dempster-Shafer strategy: Dempster-Shafer theory [6]
is a mathematical theory of evidence based on belief functions. It uses degrees
of belief collected from previous predictions to merge two pieces of information.
The belief function is usually represented as a basic probability assignment (bpa).
The Dempster-rule for evidence combination [6] can be used recursively to merge
the prediction results from different fault classifiers.       

Implementation & Case Studies:

The three evidence fusion strategies are applied to diagnose
faults in two different case studies, namely the Tennessee Eastman Challenge problem
[7] and the startup of a lab-scale distillation unit [8]. In each case, online process
measurements are analyzed using different FDI methods ? neural-networks [1],
principal components analysis [2], and self-organizing maps [3]. Each FDI
method is capable of diagnosing faults by extracting different features from
the process measurements and the classification results obtained from each
method also varies considerably. The performance of each individual method are
studied and compared to the performance achieved using different decision fusion
scheme.  A comparison between the proposed multi-agent approach [4] with hybrid
methods such as blackboard architecture [9] for integration of FDI methods is
also presented. 

Results & Conclusions:

Results obtained through decision fusion suggest that combining
diagnostic classifiers outperform approaches based on any single approach. The decision
fusion methods that utilize evidence gathering, i.e. Bayesian and Dempster-Shafer
techniques are found to perform better than averaging techniques such as voting.
Misclassification rates in terms of both false positives and false negatives are
greatly reduced when decisions are fused.

In summary, fusion of heterogeneous FDI methods provide an
effective way to combine the strengths of various FDI methods, as results from the
method which shows high rate of successful diagnosis over certain fault classes
will dominate during the process of fusion, thus resolving possible conflicts
and disagreements among the FDI methods (agents).

References

[1] Srinivasan, R.,
Wang, C., Ho, W.K., Lim, K.W., (2005). Context-based recognition of process
states using neural networks, Chemical Engineering Science 60, 935-949.

[2]
Qin, S. J., (2003). Statistical process
monitoring: basics and beyond, Journal of Chemometrics 17, 480-502.

[3]
Ng, Y.S., and Srinivasan, R., (2004). Monitoring
of Distillation Column Operation through Self-organizing Map, 7^th
International Symposium on Dynamics and Control of Process System (DYCOPS), Massachusetts,
USA,
July 5 ? 7.

[4]
Ng, Y.S., and Srinivasan, R., (2004).
Collaborative Decision Support during Process Operations using Heterogeneous
Software Agents, AIChE annual meeting, Paper 425q, Computers in Operations and
Information Processing, Texas, USA, Nov 7 ? 12.

[5]
Xu, L., Krzyżak, A., Suen,
C.Y., (1992). Methods of combining multiple classifiers and their applications
to handwriting recognition, IEEE Transactions on Systems, Man, and
Cybernatics 22(3), 418-435.

[6] Shafer, G., (1976). A Mathematical Theory of Evidence, Princeton University Press.

[7] Downs, J.J., and Vogel.,
E.F., (1993). A plant-wide industrial process control problem, Computers and
Chemical Engineering 17(3), 245-255.

[8]
Ng, Y.S., and Srinivasan, R., (2004). Distillation
unit case study homepage, iACE-Laboratory, Singapore, http://www.iace.eng.nus.edu.sg/research/Distillation_column/index.htm,
National University of Singapore.

[9]
Mylaraswamy, D., and Venkatasubramanian, V., (1997). A hybrid framework for
large-scale process fault diagnosis, Computers and Chemical Engineering 21,
935-940.