Data based models are frequently and successfully used to monitor the operation state and detect faults in industrial plants. Deriving these models from data, however, can be quite difficult and unreliable, especially if data from actual operation have to be used, as the resulting optimization problem often becomes ill conditioned: available data may not contain sufficient information, in particular for continuous production systems which are often run for longer times under almost constant operating conditions. This paper presents a local steady state approach for such problems based on complexity reduction: data are pre-processed prior to the modeling phase to allow the build-up of reduced complexity models, thus replacing the original ill conditioned problem with a better conditioned one. A method is presented to extract the relevant information about the process from measurement data in such a way to guarantee that the resulting simplified models will retain the essential characteristics of the original system required to perform fault diagnosis successfully. This approach has been used in different industrial applications and has proved reliable and efficient.