INDUSTRIAL SAFETY

Software for safety

Interview with Dr. Venkat Venkatasubramanian.

When a safety alarm goes off in an oil refinery, control-room personnel must make crucial decisions in a matter of minutes. Should the process be shut down? Which unit should be stopped? Should the other units continue to run? For how long? Every minute counts, for some poisonous, carcinogenic or inflammable substance that can kill people and destroy property could be escaping into the air.

A case in point is the 1984 Bhopal gas tragedy, in which more than 3,000 people died and thousands were affected, when a pesticide plant run by the Union Carbide Corporation accidentally released a toxic gas. Had the leak been detected before the safety alarm went off, the disaster could probably have been contained, if not averted. But is such detection possible?

Yes, asserts Dr. Venkat Venkatasubramanian, Professor of Chemical Engineering, who leads a research group in the Intelligent Process Systems Laboratory at Purdue University, United States. After working on the problem for 16 years, he and his team have come up with an online, real-time computer program that could quickly determine the cause of process abnormality and recommend appropriate action. Called DKit, or Diagnostic ToolKit, it can serve a variety of industries, including chemical plants.

While DKit works backwards from the symptom to the cause, another program, HAZOPExpert, also developed by the team, works forward, from the causes to the hazard and helps design safer plants.

Troubled by the Bhopal tragedy, Venkatasubramanian, who had just completed his Ph.D in chemical engineering from Cornell University, U.S., decided to explore methods to prevent such industrial accidents. After a stint as Research Associate, Artificial Intelligence, in the Department of Computer Science at Carnegie-Mellon and Columbia universities, Venkatasubramanian joined Purdue University' s Chemical Engineering Department.

Venkatasubramanian has published more than 100 research papers, which cover areas of process fault diagnosis and supervisory control, hazard and safety analysis, operating procedures synthesis and product design using knowledge-based systems, neural networks, genetic algorithms, artificial intelligence, mathematical programming and statistical approaches.

DKit and HAZOPExpert systems combine Venkatasubramanian' s expertise in various specialities. His experience as a consultant to global corporations and institutions such as Exxon, Honeywell, DowElanco, Lubrizol, the United Nations Industrial Development Organisation, the United Nations Development Programme, the Indian Oil Corporation, Imperial Chemical Industries, American Cynamid, Amoco, Arthur D.Little and G.D. Searle, has helped him understand the needs of various industries.

S. THANTHONI

Recipient of several awards and honours, Venkatasubramanian serves on the editorial board of the Process Safety Progress journal published by the American Institute of Chemical Engineers. He was recently selected by Industry Week magazine as Ňone of the 50 stars in the U.S. whose achievements are shaping the future of our industrial culture and America' s technology''.

Recently in Chennai to establish links with academic institutions and industries in order to test DKit and HAZOPExpert, Venkatasubramanian spoke to Asha Krishnakumar on his research work. Excerpts from the interview:

What motivated you to work on the problem of industrial accidents?

The 1984 Bhopal incident. Several things went wrong there. Incidents like that do happen when you design complex systems. There are risks involved and we try to minimise them to absolutely low levels. We cannot, however, eliminate them. The role of the plant manager is crucial. He is concerned about how to run the plant not only safely but optimally and economically. In this case I was interested in designing systems which could monitor chemical plants in real time, 24 hours a day and 365 days a year, and recognise abnormal situations. Typically, the alarm system in a plant goes off when the problem has progressed to a considerable extent, but just short of the danger level.

The way alarms are set, it may not be too late before they go off. But when several things conspire to happen at the same time and when people do not know how to react, it may be too late. Most often there is a panic reaction to abnormal situations as it is difficult to reason through all possibilities within a few minutes.

I was interested in designing a system that could detect the problem and warn people of the abnormal situation much before the alarms went off and tell operators what had gone wrong and why, and suggest solutions. An alarm does not tell you why something has gone wrong.

The control systems in chemical plants do not do this. For example, if a system has to control a plant at a certain temperature, it takes periodic temperature readings and compares them with what is required. If the reading is higher or lower than normal, it will open or close the valve appropriately. The problem is that the sensors that take the readings may fail, or the valve that has to open or close may fail or something else might fail. This is when you need human help. But, then, it may be difficult for humans to diagnose what went wrong and how to rectify it quickly. So, I was interested in designing the next generation control systems.

What represents next generation control systems?

Systems that are capable of sophisticated reasoning. So, in some sense I was, and still am, interested in designing computers that can reason like humans but much faster, covering a lot more ground. This is where artificial intelligence comes in - designing systems to think and reason like people. With this knowledge-base we developed DKit, the first part of our research work, and HAZOPExpert systems, the second part.

What is DKit?

It is an intelligent control system that can help figure out what is going wrong in a plant and what should be done. That is, abnormal situation management (ASM) wherein you reason from symptoms to causes.

What kind of research work is involved in DKit?

We began laying the foundation for DKit in 1984-85, soon after the Bhopal incident. We looked at the different approaches to ASM. Reasoning backward was not easy. To understand ASM we worked on neural networks, trend monitoring, causal and analytical modelling, artificial intelligence and statistical techniques. Each had its own strengths and weaknesses. No single approach addressed all the complexities of industrial diagnostic problems. So, we combined all the five approaches in DKit.

DKit is equipped with a moderator or a scheduler, which pools recommendations from different approaches, assesses the strengths and weaknesses of each, applies conflict-resolution algorithms and comes to a conclusion. If DKit does not offer one answer, it narrows the possible causes of an abnormal situation for the plant operator' s consideration.

What is HAZOPExpert?

In order to have a safe plant, it is best to design one to begin with. This is called process hazards analysis (PHA). When you design a new plant or take an existing plant and make changes to it, retro-fitting, you analyse all the implied hazards. That is, you build into the system all that can possibly go wrong with the various parts of the plant, their causes, consequences and ways to protect the system. All these things have to be worked out by a detailed, systematic analysis, involving an enormous number of man-years.

Since we had already developed some techniques for going back, I figured that I must be able to develop techniques to go forward as well using a computer. So, we designed a whole series of systems, called the HAZOPExpert. In PHA, the well-known approach is hazard and operability analysis, or Hazop. So we developed two systems, HAZOPExpert for continuous process plants such as refineries and Batch HazopExpert for batch processing plants such as those used by pharmaceutical and speciality chemicals companies.

For this we developed algorithms using which, sitting in front of the computer screen and looking at every single unit of the process drawing, you can identify thousands of things that can go wrong in a plant, identify the causes and consequences, and provide the results in a summarised form, as a table.

When people do PHA, they look at two scenarios. First, the routine functioning that is common to all plants. You ask the same kind of questions and go through the same kind of logic. These are the routine hazards. Some situations are unique to particular plants. We have focussed on routine situations, as 80 per cent of the effort is expended on solving routine problems.

So Hazop models hazards that may arise routinely in a plant are common to all plants. So, the Hazop system can take care of 80 per cent of the problems quickly.

For a large chemical plant it is almost impossible to analyse and build into the system all possible system failures. A typical chemical plant has hundreds of units, each responsible for 10 different activities. And there may be thousands of such combinations. It is humanly impossible even to identify all possible hazards. But that is how it is done now. This takes enormous amounts of time, effort and money. In the U.S., industries spend $ 2-3 billion a year on PHA.

For long, people did it because it was the right thing to do. But after the Bhopal incident, U.S. companies became concerned; they carried out a number of investigations. PHA was always being done by good chemical companies because that was the right thing to do. But in 1992, when Hazop was made mandatory, all U.S. plants became interested in our work at Purdue.

We are in the process of commercialising all our software. Batch HAZOPExpert is being tested in G.D. Searle and Company in Chicago.

Ideally, we would like to have one comprehensive system that is both diagnostic, like DKit, and predictive, like HAZOPExpert. As they are both related and part of the same problem, our team is now working on combining the two programs.

Are you thinking of marketing your product?

One of the techniques we developed was recently licensed to Honeywell Inc. to make it into a product. It is the first of its kind - a technique licensed from a university for control purposes - anywhere in the world. It is called the qualitative trend analysis (QTA) technique.

Have you had a trial run of the process safety software in industries?

I have been working on this technique for over 10 years, and in the last year and a half it has been on real-time trial in a major chemical plant, Exxon, in the U.S. and the results have been spectacular. It detected problems in the plant long before the alarms went off and advised the operator about what went wrong and what to do.

After the Bhopal tragedy there has been a lot of research work to understand the problem. How is your work different from the rest?

There is nothing like this (HAZOPExpert). We are acknowledged as the leaders in process safety, both on-line and off-line design. It is something of a pioneering effort.

How long did it take to complete the program?

Since we had to start from scratch, we took longer than would have been the case otherwise. QTA and the diagnosis work took us 15 years, roughly about 50 man-years of work. HAZOPExpert and Batch HAZOPExpert took nearly half of that.

How relevant are the products for India?

They are especially relevant for developing countries, such as India. After all, the worst accident in the history of chemical plants happened in India. Also, a number of chemical plants are being built in India. Many plants are also shifting base from developed countries to countries like India because of the advantages of cheap labour, environmental considerations and so on.

Have you patented your work?

We have licensed part of our work to Honeywell. The intellectual property rights are with Purdue. When we set up the company and market the products, we will acquire the rights. Purdue encourages the faculty to start companies to disseminate knowledge as it sees its role not only as one which creates knowledge but also one that transfers the knowledge to society. The two ways of transfering knowledge are through licensing and by setting up companies. Purdue does both.