Structure Crack Identification Based on Surface-mounted Active Sensor Network with Time-Domain Feature Extraction and Neural Network

In this work the condition of metallic structures are classified based on the acquired sensor data from a surface-mounted piezoelectric sensor/actuator network. The structures are aluminum plates with riveted holes and possible crack damage at these holes. A 400 kHz sine wave burst is used as diagnostic signals. The combination of time-domain S0 waves from received sensor signals is directly used as features and preprocessing is not needed for the dam age detection. Since the time sequence of the extracted S0 has a high dimension, principal component estimation is applied to reduce its dimension before entering NN (neural network) training for classification. An LVQ (learning vector quantization) NN is used to classify the conditions as healthy or damaged. A number of FEM (finite element modeling) results are taken as inputs to the NN for training, since the simulated S0 waves agree well with the experimental results on real plates. The performance of the classification is then validated by using these testing results.