Application of Pattern Recognition Techniques to non Invasive Insulation Monitoring

This article resumes results reached at the R&D project named "Study for ultra-sound sensor application as predictive technique at the management of substations and distribution and transmission lines". The samples were constituted from insulators acquired in trade, which had been previously submitted to X ray tests. This samples, after classified as "defective" or "non defective", were submitted to electrical tests, during which a bank of ultra-sound signs was constituted. This bank enabled the building of a training set for diagnostic system elaboration. When at the setting up of this system, many pattern recognition techniques were tested: Fisher, Karhunen-Loeve, linear and non-linear artificial neural network, as well as the application of many acquired signal features. This article presents the results and describes the prototype built according to the technique that presented the best performance. The paper even presents some considerations over refinements that may be added to future versions of the implemented diagnostic process.