Unsupervised and auto-adaptive neural architecture for on-line monitoring. Application to a hydraulic process

Generally, the functioning of a monitored system is characterized by some parameters. Each evolution of these parameters may be assimilated with a drift or a fault. Therefore it may be considered to estimate the functioning state of the system by analyzing the drifts of these critical parameters evolving with time. So, this paper presents a new architecture, based on a new neural network classifier, for on-line monitoring of complex industrial systems. The first stage consists in the choice and the extraction of the significant parameters. The second stage extracts some parameter characteristics to be used in order to discriminate the different kinds of evolution. Then, in the third stage, a specialized neural network is used to classify, on-line, the evolutions for each significant parameter. The neural network is initialized with a set of synthetic and experimental data and a copy of this classifier is used to achieve the trend detection for each monitored parameter. Due to unsupervised rules and auto-adaptive abilities, each copy is adapted according to the specific evolutions of its monitored parameter in order to take account of new kinds of evolutions in situ and on-line. In the fourth stage, a decision tree is used to establish the diagnostic by analyzing the results of the trend detection stage for all the monitored parameters. The application to the monitoring of an industrial hydraulic process shows the ability of the classifier to learn and detect the different drifts.