Comparative study of neural network damage detection from a statistical set of electro-mechanical impedance spectra

The detection of structural damage from the high-frequency local impedance spectra is addressed with a spectral classification approach consisting of features extraction followed by probabilistic neural network pattern recognition. The paper starts with a review of the neural network principles, followed by a presentation of the state of the art in the use of pattern recognition methods for damage detection. The construction and experimentation of a controlled experiment for determining benchmark spectral data with know amounts of damage and inherent statistical variation is presented. Spectra were collected in the 10-40 kHz, 10-150 kHz, and 300-450 kHz for 5 damage situations, each situation containing 5 members, "identical", but slightly different. A features extraction algorithm was used to determine the resonance frequencies and amplitudes contained in these high-frequency spectra. The feature vectors were used as input to a probabilistic neural network. The training was attained using one randomly selected member from each of the 5 damage classes, while the validation was performed on all the remaining members. When features vector had a small size, some misclassifications were observed. Upon increasing the size of the features vector, excellent classification was attained in all cases. Directions for further studies include the study of other frequency bands and different neural network algorithms.