Data Mining for Fault Diagnosis in Dynamic Processes: An approach based on SVM

The major inconvenient for the fault detection and isolation (FDI) in technical processes based on analytical redundancy is the requirement of a very accurate model of the system. By contrary, in the methods based on data handling it is not required of a precise model, since they are based on the manipulation of the information by means of the measured data. Thus, in this work a set of statistical indices allowing quantifying the amount of information contained in the collected data is presented, in order to realize the FDI. These indices are used for the reconstruction of the fault patterns, and next for its classification is used the machine learning technique, in particular the support vector machines (SVM). For verification of the results, generated data by two nonlinear models are used; one of discreet time that simulates the Logistic Application, which is used under different types of dynamic states behavior that represent the occurrence of faults. The other model is a continuous time that represents the control of a magnetic levitation system. For the first model, the results show that by means of the obtained statistical indices the reconstruction of fault patterns is obtained, which allow separating the different dynamic behaviors (faults) and using a SVM considering different kernel, a classification between the faults is obtained (classes). For the second model, the classification of the faults by means of a SVM is realized, obtaining one diagnosis index. Key-words: Fault diagnosis, Data mining, Operational classification, Machine learning, Support Vector Machines.