Applications of Self-Organizing Maps in Structural Health Monitoring

Abstract

The Structural Health Monitoring process includes several steps like feature extraction and probabilistic decision making, which need some form of data fusion and information condensation. These take place after data acquisition and before being able to decide, if a monitored structure has faced damage. Although feature selection is an important step, the processing and suitable preparation of these data are significant, influencing the potential of decision making in various ways. With Self-Organizing Maps (SOM) a multi-purpose instrument for these tasks of pattern recognition and data interpretation is presented here. Self-Organizing Maps belong to the group of artificial neural networks and by using the special map character provide the opportunity of additional visualization. Especially when monitoring a structure over a long period of time, environmental changes often occur, which can mask the effects of damage on the dynamic behavior of the structures. As one potential application of SOM, the possibility of distinguishing between environmental changes and damage of the structure is shown. In this application a self-organizing network is trained with data of the undamaged structure and via calculation of the distance to the map a damage indicator is developed. Moreover, the distinction between different damage modes of piezoelectric sensors is presented using SOM as a tool of pattern recognition and visualization. This application uses data recorded from different damage modes extracted from one specimen of a piezoelectric element. The trained network can be compared with other piezoelectric elements mounted in a similar way to be able to detect possible sensor damage. This helps avoiding false alarms even under changing environmental conditions. Both applications have been successfully used to analyze experimental data on coupon level showing the applicability of the presented concepts.