Real systems are usually nonlinear and their modeling and monitoring remains adifficult task. However, with advances in technology and the availability of big amounts of data,we have a facility to operate these systems.This work presents a methodology for fault diagnosis and modeling which is in large part basedon the method of Support Vector Machines (SVM) which data-based. The proposedmethodology is applied to various nonlinear multivariable systems including: wastewatertreatment processes, wind turbines and stirred tank reactors.The objective of this PhD is to examine the possibility of extracting the maximum of informationfrom data to effectively monitor the behavior of real systems and rapidly detect any faults whichmay impair their proper functioning. The same method is used for modeling the differentsystems. Several challenges were identified and surmounted such as the complexity of thesystem behavior, large amount of data varying at different time scales, the presence of noise anddisturbances. A generic method of fault diagnosis is proposed for the generation of the faultcharacteristics followed by an evaluation of these characteristics as well as an improved transferof knowledge in modeling.In this thesis the usefulness of the tool Support Vector Machines in Classification has beendemonstrated by the construction of decision models dedicated to evaluating the characteristicsof faults, and also its usefulness for modeling/ or as estimator for the nonlinear systems usingsupport vector machines dedicated for regression (SVR).The combination of SVM and a method based on models “observer” was also considered andwas found to be interesting in some cases to ensure proper fault diagnosis.
