Structural Health Monitoring using deep learning with optimal finite element model generated data. (November 2020)

• Record Type: Journal Article
• Authors: Seventekidis, Panagiotis; Giagopoulos, Dimitrios; Arailopoulos, Alexandros; Markogiannaki, Olga
• Is Part Of: Mechanical systems and signal processing. Volume 145(2020)
• Publication Date: 2020-11
• Subjects: Structural Health Monitoring -- Deep learning -- Neural networks -- Damage identification -- FE model updating; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2020.106972
• Languages: English
• ISSNs: 0888-3270

Abstract

Highlights: Novel Structural Health Monitoring (SHM) method. Deep learning methods trained by numerically generated responses. Computational FE model updating. Binary and multiclass Damage Identification problems. Potential tool for Damage Identification tasks. Abstract: Identifying damage through Structural Health Monitoring (SHM) methods is increasingly attracting attention due to multiple maintenance and failure prevention applications. In order to create reliable SHM systems for structural damage identification (DI) tasks, access to large amounts of data containing measured structural responses is usually necessary. The data acquisition is mostly based on direct experimental responses up to now and requires time consuming measurements in various working and ambient conditions of the structure. In the present work, a novel SHM method is tested where all data is solely derived from FE calculated responses, after an initial experimental cost for FE model updating on the healthy structure state. The proposed method can be especially applied in cases where specific damage types are expected or anomalies are adequately defined so they can be effectively simulated by FE models. Origin of such models may then be the healthy experimental status. To test the proposed SHM system, the optimal FE model of an experimental benchmark linear beam structure is constructed, simulating an undamaged condition. In order to check the robustness of the proposed method the damage magnitudes imposed on the benchmark are kept small and combined with random excitations. Next, the optimal FE model is used for generating labeled SHM vibration data through a repetitive load case scheme which also includes uncertainties simulation. The data derived from the optimal FE model is finally used to train a Deep Learning (DL) Convolutional Neural Network (CNN) classifier which is after experimentally validated on the benchmark structure. The optimal FE generated data proves to be able to train an accurate CNN that can predict adequately the experimental benchmark states. A comparison is also given with a CNN trained by the corresponding nominal FE model data which is found not reliable on the experimental validations. The presented combination of optimal FE and DL is a potential solution for future SHM tools and further investigation is encouraged.