Damage Diagnostic Technique Combining Machine Learning Approach With a Sensors Swarm

A Model-free approach is particularly valuable for Structural Health Monitoring because real structures are often too complex to be modelled accurately, requiring anyhow a large quantity of sensor data to be processed. In this context, this paper presents a machine learning technique that analyses data acquired by swarm of a sensor. The proposed algorithm uses unsupervised learning and is based on the use principal component analysis and symbolic data analysis: PCA extracts features from the acquired data and use them as a template for clustering. The algorithm is tested with numerical experiments. A truss bridge is modelled by a finite element model, and structural response is produced in healthy and several damaged scenarios. The present research shows also the importance of considering a sufficient number of measurements points along the structure, i.e. the swarm of sensors. This technology, which nowadays is easily attainable with the application of optical Fiber Bragg Grating strain sensors. The difficulties related to the early stage damage detection in complex structures can be skipped, especially when ambient, narrow band, moving loads are considered, enhancing the prediction capabilities of the proposed algorithm.