Sometimes an industrial component can be affected during its life by several external and internal conditions that can induce failure modes, or at least, contribute to the presence of one or more symptoms that can cause a certain amount of stress in the component and therefore facilitate the development of such failure modes. This paper presents a method to characterize the normal behaviour expected of an industrial component under typical working conditions. This is based on the use of neural networks techniques and training sets obtained from real data coming from the component dynamics. A normal behaviour model is created for each failure mode that is tried to be detected through its observable symptoms. These models are able to cover two objectives: to perform an incipient anomaly detection of which the cause will later be investigated and, especially relevant and new, to estimate the amount of stress of the component for the failure mode. A new method will be described in order to estimate a qualitative evaluation of the amount of stress in industrial components that can be used for rescheduling the maintenance planned for the component when this is really needed, contributing to prevent undesirable faults, saving unnecessary maintenance costs and making a better use of the asset over a longer period of time. The demonstration of the results of all the methods described: the creation of the normal behaviour models, the evaluation of the amount of stress evaluation and anomaly detection, will be applied to the case of a wind turbine.
Keywords: Diagnosis, multi-layer perceptron, normal behaviour models, anomaly detection, component stress
24th International Congress on Condition Monitoring and Diagnostics Engineering Management - COMADEM 2011. ISBN: 0-9541307-2-3. Stavanger, Norway. 30 Mayo - 1 Junio 2011
Published: May 2011.