Surveillance continue des transformateurs de puissance par réseaux de neurones auto-organisés

An outage of a power transformer, generally, has heavy financial consequences for electric power systems utilities. In order to prevent any failure and to optimize their maintenance, a growing number of operating parameters are measured on-line. The important amount of data together with the ever increasing number of units to be monitored calls for automatic alarm devices. Due to the complexity of power transformers, the evolution of the operating parameters is hardly known and, in fact, difficult to model. That is why the alarms, generally, start on thresholds defined independently for each recorded variable. With such an approach, neither variations in relation to operating conditions, nor existing relations between different monitored quantities are taken into consideration. This thesis is devoted to the study of an "intelligent" on-line monitoring system, able to adapt itself depending on those variations. The developed solution uses self-organizing maps (SOM). It consists of unsupervised neural networks inspired by the organization of cortical zones in the brains of mammals. Their characteristics enable them to represent a complex set of data in a generally dimensional space preserving at the same time, the topology of the original set as truthfully as possible. This work shows, more specifically, that these systems allow to model the different states of a transformer preserving not only the general structure of data, but also their local correlations. The topological preservation feature of the maps implies that close operating states will also be close on the map. Based on this observation, a powerful graphic interface is proposed to exploit qualitatively the measurement results. The monitoring is considered in the real context for which, in case of a defect, the transformer behaviour is generally unknown. This constraint resulted in the study of the self-organizing map response for states located outside the original model space. A confidence indicator quantifying the accuracy of the model response was proposed. It is shown that this indicator allows to define an adaptive alarm threshold that complements the developed tool for qualitative interpretation. The developed approach is finally validated using data collected on a pilot-transformer, operating in the Swiss electric network. The obtained results show the pertinence of the proposed technique and its ability to reveal any sudden change of one or several variables with a relatively sharp sensitivity.