Dynamic Structural Health Monitoring With Filter Net De-noising and SSDBN Model Using Vibrational Data

Abstract

Measurement noise is always part of the vibration data in vibration-based structural health monitoring (SHM). However, it might be challenging to regulate the state in which civil constructions are tested in the field. Moreover, strong noise from a variety of sources, make damage detection inaccurate. Additionally, the precision of the current studies will eventually begin to saturate and possibly deteriorate. To overcome the mentioned limitations, this research proposed a deep learning framework for monitoring the structural health. First, Filter Net is suggested, which integrates neural network techniques for de-noising observed vibration signals with skip connection, dropout and shuffling. The next step was to propose a smooth sparse deep boltzmann network to detect structural degradation. A sparse penalty component built on the inverse function norm was added to improve performance. In addition, a greedy algorithm is used to perform unsupervised learning, which trains the first Restricted Boltzmann Machines (RBM) using the sampling data before using the first RBM's parameters to initialize the Deep belief networks (DBNs) first layer's parameters. Then, a BP network is used in a fine-tuning method to get the final systematic parameters. As a result, the RBM provides the Smooth Sparse Deep Boltzmann Network (SSDBN) with a decent starting value and therefore ensures higher performance.