No abstract

No abstract

This paper presents a novel approach to detect structural damage based on combining independent component analysis (ICA) extraction of time domain data and artificial neural networks (ANN). The advantage of using time history measurements is that the original vibration information is used directly. However, the volume of data, measurement noise and the lack of reliable feature extraction tools are the major obstacles. To circumvent them, the independent component analysis technique is applied to represent the measured data with a linear combination of dominant statistical independent components and the mixing matrix [ A]. Such a representation captures the essential structure of the measured vibration data. The vibration features represented by the mixing matrix provide the relationship between the measured vibration response and the independent components and are then employed to build the simplified neural network model for damage detection. Two examples are included to demonstrate the effectiveness of the method. First, a truss structure with simulated displacement data was used, and the results show that healthy and damage states located in the nine elements may be classified. Second, a bookshelf structure together with measured time history data from 24 piezoelectric single axis accelerometers was used to demonstrate the approach on a physical structure. The results show the successful detection of the undamaged and damaged states with very good accuracy and repeatability.

This paper deals with structural damage detection using measured frequency response functions (FRF) as input data to arti®cial neural networks (ANN). A major obstacle, the impracticality of using full-size FRF data with ANNs, was circumvented by applying a datareduction technique based on principal component analysis (PCA). The compressed FRFs, represented by their projection onto the most signi®cant principal components, were used as the ANN input variables instead of the raw FRF data. The output is a prediction of the actual state of the specimen, i.e. healthy or damaged. A further advantage of this particular approach is its ability to deal with relatively high measurement noise, which is a common occurrence when dealing with industrial structures. The methodology was applied to detect three different states of a space antenna: reference, slight mass damage and slight stiffness damage. About 600 FRF measurements, each with 1024 spectral points, were included in the analysis. Six 2-hidden layer networks, each with an individually-optimised architecture for a speci®c FRF reduction level, were used for damage detection. The results showed that it was possible to distinguish between the three states of the antenna with good accuracy, subject to using an adequate number of principal components together with a suitable neural network con®guration. It was also found that the quality of the raw FRF data remained a major consideration, though the method was able to ®lter out some of the measurement noise. The convergence and detection properties of the networks were improved signi®cantly by removing those FRFs associated with measurement errors.

No abstract