Uso de técnicas de inteligência artificial para identificação da necessidade de limpeza de cadeias de isoladores

Abstract

String insulators are components present in high voltage towers responsible for preventing energy dissipation through the tower structure, that is, they are responsible for the isolation of the high voltage present in the cables of the electrical network. These string insulators need to be clean for best performance and to avoid malfunctions. The verification of the need for cleaning/washing is most often performed by human visual observation, which can lead to interpretation errors, in addition to bringing risks to the physical in- tegrity of humans who are in the vicinity of these electrical systems. Thus, this research project aims to develop a support algorithm for these professionals, who observe and classify these insulators. The support algorithm uses artiĄcial intelligence techniques and analyzes the image, inferring a second opinion on the state of cleanliness of the analyzed insulator. For the development of this algorithm, it was necessary to create a simulated database, from the use of software such as Inventor and Unity-3D, due to image limitations available from dirty insulator strings. Thus, in this research, two distinct neural networks are built using supervised learning techniques, where the Ąrst one is for the detection of the chain of insulators, and the second for detecting the type of impurity on the disk surface. In the Ąrst stage, techniques that use supervised learning are studied, more speciĄcally aimed at semantic segmentation networks, and in the second stage, are classiĄcation networks were checked to detect the type of impurities. In the process of detecting insulator strings, an average dice coefficient of 0.95 was achieved for simulated images and 0.92 for real images, based on a database with only simulated images. In the impurity type identiĄcation stage, the average accuracy obtained was 0.98, with a completely different dataset from the Ąrst, but also simulated.