Much of the earlier work presented in the area of computerised on line fault diagnosis, has been based on heuristic principles, where knowledge has been represented in some kind of rule structure. This paper presents an overview of the possibilities and pitfalls of different modelling techniques in fault diagnosis. Experiences from a large industrial project, KSM, are presented, where results from risk identification are used in on line fault diagnosis in two industrial plants. The novelties of this system is the introduction of characterisations in order to evaluate the most probable cause among several possible causes. Another feature is the support system for detecting possible operator errors after a deviation has occurred. This facilitates the decision of which information should be given to the operator from the support system. Future development, where qualitative mathematical modelling is used in order to find the most probable cause among several possible, is also discussed.