Decision Support System for Process Analysis and Supervision

by Galia Weidl

The increasing complexity of industrial processes and the struggle for cost reduction and higher profitability means automated systems for processes diagnosis in plant operation and maintenance are required. The system proposed here is designed to address this issue.

As system complexity increases, condition monitoring and fault diagnosis become demanding tasks for process operators, who face an overflow of data and may have to undertake extensive searches for unexpected faults. This can lead to work overload and high stress levels. Simultaneously, technical, economical and time constraints on production efficiency and quality require that condition monitoring and adaptation of production and service processes be undertaken on a continuous basis.

In this context the ABB group, along with its customers from process industries and the Danish company Hugin Expert, have collaborated on a project (which commenced in 2000) targeting the development of a methodology for root cause analysis (RCA) and decision support on process operation. The author, who was involved in technology evaluation and methodology development while associated with ABB, is now associated with IFF - Stuttgart University and is working on an extension of the methodology to automotive and manufacturing industries and services. It is expected that this extension will be implemented within a European project known as 'The Butler Concept'.

Motivation and the Techniques Employed
In finding the origin of a process disturbance or fault, there exists a need for a quick and flexible guidance tool for decision support at higher automation levels, including analysis of process conditions and advice on cost-efficient actions.

The technology of probabilistic graphical models such as Bayesian networks and influence diagrams have turned out to be the best of a number of alternatives, in the case where high diagnostics capabilities, explanation of conclusions for transparency in reasoning, and trustworthy decision support are expected by the users (process engineers, operators and maintenance crew).

Due to the existence of a number of first-level diagnostic tools, the aim has been to provide decision support in process operation. The framework of Bayesian networks has been found to be an efficient and flexible tool in overall-level process operation analysis, since not all conditions are measurable or computable in real time, and the combinatorial reasoning procedure is subject to uncertainties.

The development of the methodology incorporated the following system requirements and modelling issues:

root cause analysis of industrial processes with adaptation to process operation/grade changes, aging and wear
reusable system design for various process applications
reusable modelling of repetitive structures (eg sensors, control loops) and assets (pumps, valves)
risk assessment of disturbances by analysis of signals' level-trend, while adapting to changes in process operation mode
ease of communication and explanations of conclusions at different process levels, eg process overview, (sub)sections, units and instrumentation.

These modelling and system requirements have been met in the methodology. This is supported by the integration of the methodology and the Hugin-tool into the ABB Industrial IT platform. This RCA-integration allows efficient data exchange with all available IT-applications, eg distributed control systems, diagnostics of sensors and control loops, and physical-model computations. The infrastructure for applying this methodology in different domains is therefore ready for immediate use. The method and the analyser for producing information have been the subject of five separate patent applications.

Figure 1: Digester Fiber-line. Case-study: Monitoring of the digester operating conditions.

The first prototype of the root cause analysis system was tested by ABB with real process data during 2002. The monitoring and root cause analysis of the digester operating conditions in a pulp plant have been chosen as a real-world application (see Figure 1). The structure of one of the developed Bayesian Networks is shown in Figure 2.

Figure 2: An example of a Bayesian Network for root cause analysis of process operation.

The real application development has been closely related with its integration into the Industrial IT platform. The extension includes the use of object-oriented Bayesian networks (OOBN). OOBN facilitate the modelling of large and complex domains and allow reusability.

Benefits, Results and Consequences
The resulting RCA-system provides process operators with information on condition overview and advice on the most efficient sequence of corrective actions. In addition, the advantages for process industries include the seamless integration of all relevant information sources on the Industrial IT platform, resulting in automated and early assessment of abnormal conditions with flexible diagnosis and advice. Future enhancements will involve integration of the functionalities providing advice on a suitable time for maintenance activities under technical constraints and order deadlines in process scheduling, as well as simulation of the impact of intended corrective actions on process efficiency. This allows pro-active (instead of reactive) troubleshooting to be undertaken, which increases the process performance, availability and output, avoids potential process breakdowns and cuts both operation downtime and maintenance cost.

Expected Function - RCA as a Powerful Complement to System Control
We believe that this methodology will provide a powerful complement to system control in process industries. Our expectations and the proof of the system's potential are based on successful tests at ABB with real plant data. Future challenges are primarily related to the application of hybrid learning systems for abnormality detection, diagnosis and advice in large-scale industrial settings.

Links:
http://www.abb.com
http://www.hugin.com
http://www.iff.uni-stutgart.de
Press-release Web Site: http://www.hugin.com/cases/Industry/ABB/ABB18122003.pdf

Please contact:
Galia Weidl, University Suttgart
E-mail: giwiff.uni-stuttgart.de, galiamathematik.uni-stuttgart.de