Structural health monitoring through deep learning: A study on scalogram-based vibration signal classification

Structural health monitoring (SHM) is of critical importance in a variety of fields. Oftentimes the studied system also possesses complex behavior, turning the assessment of such structure difficult. One of the solutions for high complexity systems is the use of data-driven approaches, such as machine learning and deep learning techniques. However, training traditional deep learning architectures can require large amounts of data, which may not be available for both health and damaged conditions in operating structures. This study proposes the adaptation of GoogleNet allied with transfer learning for damage detection in vibrational structures. By converting vibrational signals into scalograms, using the Continuous Wavelet Transform (CWT) both the time and frequency information of the measurement can be kept and the Convolutional Neural Network (CNN) can classify the time-series dataset into different damage scenarios. This methodology initially employs full-field measurements from the structure, allowing fine-tuning of the GoogleNet, which also requires less amount of data. Later, traditional accelerometer measurements are employed for testing and classification of different damages. This approach poses good accuracy for anomaly detection, using diverse frequency excitation as input, demonstrating a great capability to perform image processing with vibration signal analysis, allowing local testing for damage detection and real-time signal classification.