The detection of structural damage using Convolutional Neural Networks on vibration signal

Vibration wave motion acting on a structure travels throughout the whole structure at a given instance. Instead of using many Convolutional Neural Networks (CNNs) to test every part of the structure in different testing sessions, one CNN is adopted to detect damage for the whole structure in one testing session. The fundamental goal of using CNN in structural damage detection is to answer the question of to what extent (severity of damage) the structure under observation is damaged. The proposed approach successfully uses the vibration signal from two measurement sessions to train only one dedicated CNN in order to know when the structure is damaged. The two measurement sessions are done by taking vibration signal data from a four floor steel structure when it is undamaged and when the structure is fully damaged. The CNN then predicts the degree of structural damage in respect to a given scale during testing. This approach is convenient for large structures since it involves using only two measurement sessions for training and also fuses all the sensor data into one CNN. The experiment results prove that the CNN accurately predicts the severity of damage of the structure as well as supporting the real-time vibration signal processing and the use of few resources since the method requires only one CNN for feature extraction and classification.