Expert systems are intended to solve problems in difficult, unstructured domains, where knowledge guides the expert reasoning.
The paper presents an expert based approach to the problem of risk assessment at various work places, underlying the advantages in using expert system methods and techniques to deal with risk identification and evaluation.
Classic risk assessment
At the work place we should consider in the risk assessment process, the man-machine system with its four components, the human operator, the task, the machine and the working environment. An efficient risk assessment must study all these components and also their interaction that results in the activity being done at the work place.
However, the majority of risk assessment systems are centered on the machine. There are some reasons for that approach, considering the great subjectivity of human risk assessment. The programs that are implementing these methods could be called classical programs, having a pre-defined algorithm to perform the risk assessment. This classic, machine-centered approach is not very suitable in many cases, especially where the man generated accidents are prevailing.
Expert systems and their applications to risk assessment
What is an expert system?
Expert systems are programs that offer advice or solve problems by reasoning with volumes of knowledge highly specific to a particular domain. The activity of designing and implementing expert systems is known as knowledge engineering.
A general structure of an expert system is presented in the following figure.
Figure 1 General schema of an expert system
Why an expert approach?
Risk assessment is a very important process, regarding the safety and health of many workers. A well done risk assessment process could lead to significant improvements not only just to the work conditions but also to the efficiency of the work.
Considering the core of the risk assessment process at the work place, it is mainly an expert job. The human expert identifies specific risks, is giving them severity and probability weight and is proposing measures to reduce risks at ALARP levels or simply eliminate them. Risk assessment is not a repetitive process; the work place conditions are unique. So, indifferent of how performing are classic assessment programs, indifferent of the complexity of the implemented assessment method (Hazop, FMEA,FMECA, or other such like) they could not perform expert tasks regarding risk assessment but just the repetitive and monotonous part of the process.
What could an expert system do regarding risk assessment?
Having established that risk assessment is mainly an expert approach we could see the expert system as an integrator and also as an assistant to the human assessor .The expert system could:
-analyze and logically validate the results obtained by running machine centered assessment methods;
-interpret the obtained results considering the global picture of the work place being assessed;
-perform an objective assessment for the human component at the work place;
-generate risk models and use these models to predict risk development for short and medium prognosis;
-integrate all these pieces in a single view and use this view to develop prevention plans so to reduce risks at ALARP levels or just eliminate them ; the next figure presents a general schemata of a risk assessment expert system.
Figure 2 General schema of an expert risk assessment system
Expert elements in risk assessment
Machine risk assessment
There are many programs that perform more or less efficiently machine risk assessment, using a variety of assessment methods centered on the machine component
The classic assessment methods are centered mainly on the machine, giving little importance to the external risk sources. But, in many situations, the external risk sources (such as environment conditions, task design, and task execution) are triggers to incidents and accidents. So, for the correct image of risks at the work place, there must be analyzed all the chained components of the man-machine system. Classical programs are not doing this-an expert system may perform this analysis.
Also, the expert system may quantify the obtained results into significant cases that would simplify further similar assessments:
-by using the obtained results as a case oriented knowledge source and simply cross-check these results against the real situation, instead of performing again all the assessment process which can be very time consuming;
-by working as an example for similar assessments;
The expert system could work assisted by the human specialist to develop such cases and put those cases into a specific case base. For example, results for similar work places must be somehow similar; giving a similarity check-list for a specific machine, such a check-list could be quickly cross-checked against the real situation and established if the case and the real situation are perfectly matching, partially matching or if there are singularity points that must be analyzed separately.
Also, having the risk assessment results, the expert system could develop and run scenarios that will predict risk trends over a period of time, considering various degree of intervention towards preventing risk effects. These scenarios could be very efficient in developing and prioritizing prevention measures.
Using its learning mechanism the expert system could also improve the machine risk assessment process, mainly in his primary parts, establishing the data that will be collected and processed.
To sumarize, in the machine risk assessment process, the expert system can action as an assistant to the human specialist and also as an improver of the assessment results and as a predictor of risk trends.
Human operator assessment
The problem of human operator assessment from the risk point of view is a crucial one. Most of the occupational accidents have as their principal cause human operator failures and mistakes..The principal problem with the risk assessment methods regarding human operator is their subjectivity degree combined also with the environment and task influence. A checklist results will be influenced by the operator’s self-perception, stress and so on. In establishing a decent and objective human risk assessment a very important role can be played by corporate specific memories regarding the human operator (mode of behaviour, incidents, accidents, capacity to work alone/to work inside a team, etc.).For a specific operator, these memories could be invaluable-generally, an operator that performs badly a task and is not corrected will be performing in the same way, till an incident is happening. Defining best work procedures and checking up the real activity against these procedures could be another objective assessment factor..
Based on expert system techniques, it is possible to define, inside expert assessment system behaviour models (based on best work procedures and also on normal work condition and physiologic paramethers) and follow the real situation by cross-checking the actual behaviour of human operators against these models. The over-passing of model defined thresholds could be alarm signals and could lead to the stop of machines and of the whole process, so that the human operator will not be hurt.
The human operator assessment may be modeled by belief networks that will define, regarding the human operator, believed relations (uncertain, stochastic, and imprecise) between sets of risk relevant variables to build patterns toward the birth and development of human operator related risks.
Belief updating may be done for each new assessment, by:
-mistake-the learning from trial and errors;
The expert system may infere new beliefs from the already existing knowledge and also may check the beliefs against the real situation.
An expert approach to risk assessment
An integrated expert approach to the risk assessment problem will be developed around the expert system methods for solving specific problems. It will use a goal-oriented strategy, choosing a goal (to exactly and correctly identify the risks at work place and to establish the optimum prevention plan for the risks being identified) and attempting to achieve it.
It is difficult to think of a universal expert system that will be able to assess every risk in every industrial domain, regardless of the specific activities, machinery and so on.
A better approach would be, in this aspect, to imagine a dual expert system, with an external knowledge base, containing risk assessment methods and also significant cases and an internal, corporate specific knowledge base, which will contain the internal knowledge required to perform an efficient risk assessment.
This internal knowledge base (IKB) will be based on corporate knowledge and will act partly like a corporate safety memory . IKB will be a guide for the external risk auditing system, having captured and updated inside the significant aspects that deserve study. It will also preserve safety events inside the corporation and will be a source of lessons for safety specialists in learning from the past mistakes and improving safety at the work place.
A possible design for such a base is shown in the next figure.
Figure 3-Possible design for Corporate Internal Knowledge Base
Generally, the risk assessment problem could be interpreted as a search problem in which a solution must be found in a state space pool.
This search may be optimized using heuristic evaluation functions H(x).How does a heuristic evaluation function work? By asigning to each significant risk state x an estimate of the minimum cost to reach the goal state-in our case the definition of the subsequent risk. With such a function available, all we have to do in order to obtain efficient risk identification will be to sort the found risk list according to H(x).H(x) is only an estimator. To make it a predictor it must be done to search exhaustively the state space. The risk identifying performance varies directly with the accuracy of the heuristic evaluation function. So, if
H(x) =d 
where d is the true remaining distance to the nearest risk identification for all the risk states x, then, the search is optimal. If
H(x) =1/d 
the search will always find the goal node last, after exploring the entire state space. The perfect heuristic function, which will lead to perfect balance between the assessment costs (g (x) and assessment results H (x), could be defined using the following equation
H(x) =H’(x) for all risk state spaces x 
We could also define a cost-results function
f(x) =g(x) +H(x) 
The optimal path to the risk identification, called P, must be perfect regarding the balance assessment costs/assessment results. So, we could write
H’ (P) =0 
f (P) = g (P) +0 
and the estimate of P costs turns out to be perfect-the real cost
As presented before, expert approach could improve the risk assessment for the machine and lead to an objective and efficient assessment regarding the human operator, giving also a global risk view for the work place,.It takes into account also the task and the working environment.
In risk assessment, the expert approach could give:
-a reality approach evaluation, considering the corporate internal memories of past events and also the expert internal capability of the corporation being assessed;
-optimum prevention solutions, arranged in order of their efficiency;
-a efficient and optimized assessment;
-reduced assessment costs, time and resources;
There are also some problems. Expert based assessment using cases as an assessment base may lead to misjudgements in case of singularities existence. Also, expert based assessment needs expert knowledge which, in some situations may be very expensive.
We have realized some progress towards the development of efficient expert tools for risk assessment, some of our results being presented at various international scientific congresses. One primary research direction was the development of tools for human operator assessment from the risk/safety point of view. .We hope that in the future, with a consistent international co-operation, to be able to launch an expert system for risk assessment as soon as 2015.
 Siu N.-Risk assessment for dynamic systems-an owerview-in Reliability Engineering and System Safety 60, pg.153-164,1994
 Rasmussen J.-Major accident prevention:what is the basic research issue ? in Safety and Reliability (eds:Lydersen and others),Balkema,Rotterdam, pg.739-740,1998
 Reason J.-Human error,Cambridge University Press, 1990
 „A new duty to investigate accidents”,HSE Consultative document CD 169
 T.A.Kletz-Accident investigation-missed opportunities-keynote paper at the Hazards XVI-Analysing the past, planning the future symposium,
 Dr.J.Bond-A Janus approach to safety--keynote paper at the Hazards XVI-Analysing the past, planning the future symposium, Manchester, 2001,pg.13-23
 K.Pearl-Heuristics:Intelligence search strategies for computer problem solving, Bantam, 1997
 J.Nilsson Problem solving methods in AI, McMillan Books, 1996
 Benjamin R.jr.,Dinamics of job evaluation,in Management Review nr.42,pg.190-194,April 1993
 Ph.D.Stefan Kovacs and others-Human operator assessment-basis for a safe work place in the process industry, Hazards XVI Proceedings, Symposium series no.148,ISBN 85295-441-7, pg.819-832.
 Ph.D.Stefan Kovacs-Expert systems in risk assessment, SafeCon Proceedings, Athens, 2000