Nonparametric Vibration Based Damage Detection Technique for Structural Health Monitoring Using 1D CNN.

Structural Health Monitoring (SHM) aims to give a diagnosis of the “state” of the constituent materials, different parts, and full assembly of the structure as a whole. The continuous real-time diagnosis allows optimal use of the structure, a minimized downtime, and the avoidance of catastrophic failures. A 1-Dimensional Convolutional Neural Network (CNN) approach for SHM is proposed in this paper which can minimize human involvement, and thus improving safety and reliability. An experiment pipeline is designed for analysis of Phase I-AISC benchmark structure. The pipeline comprises of four phases based analysis where the first phase involves multiclass classification of vibration data based on the type of damage in the structure. Average accuracy of 96.11% was obtained for all the nodes of the structure. The second phase uses the maximum damage and undamaged data to train the model and compute the level of overall damage in the structure by testing the trained model for different types of damage patterns. The obtained result accurately detects the pattern for increasing instances of damage using the probability of damage (PoD) parameter. The third phase is used to figure out damage localization for different nodes. The first three phases use a distributed SHM approach where each node has an independently trained model whereas in the fourth approach a unanimously trained single model for the merged dataset, as per hypothesis, was prepared to predict damage or undamaged state for data obtained from any node of the structure. The hypothesis was rightly proven with around 97.89% probability of classification (PoC) values obtained.