Session: SE-04-01 Machine Learning for Seismic Analysis of Industrial Facilities

Paper Number: 122434

122434 - Structural Health Monitoring Using Time-Augmented Response Spectrum and Deep Learning 

In recent years, the application of deep learning to structural health monitoring against seismic ground motions has been attempted. Chemical plants and nuclear power plants house multiple pieces of component crucial for safety, necessitating evaluations of their structural integrity. Numerous methods that do not employ deep learning have been suggested, centering on the frequency domain of structural responses. Wavelet transformation, considering the temporal change in frequency, lacks physical units, limiting its engineering applications. The utilization of deep learning for structural integrity assessments enables high-accuracy and swift evaluations. On the other hand, traditional deep learning models find it challenging to handle multiple input data.To address the above issues, this paper proposes two kinds of methods. The first is the Running Response Spectrum, which enhances the conventional response spectrum with temporal information. The Running Response Spectrum is articulated in a relationship of acceleration, natural frequency, and time, and since it is composed of physical quantities, it is applicable to seismic design and structural integrity assessments. The second is a deep learning model that links conventional Convolutional Neural Networks, using neuro-fuzzy logic to combine the outputs of each CNN. This enables making judgments based on the reliability of multiple input data. Additionally, a structural health monitoring system employing these two methods was developed. Traditional structural health monitoring systems only utilize the response data of the structures. However, this system estimates the probability of anomalies following a sigmoid function and the time of anomaly occurrence from the Running Response Spectrum of both seismic ground motions and structures.The effectiveness of this structural health monitoring system was validated using data from analyses and experiments. In the analyses, simulated earthquake motions and an elastoplastic analysis model were employed, defining anomalies at a certain plasticity rate. In the experiments, a suspended structure with elastoplastic behavior was used, defining anomalies by the presence or absence of plastic deformation. As a result, it was confirmed that this method can detect anomalies and estimate the time of anomaly occurrence. Moreover, in this system, the decision-making basis of deep learning, which is a black box, is visualized, facilitating its utilization in human decision-making.

Presenting Author: Takaki Aeba Tokyo Denki University

Presenting Author Biography: Takaki Aeba is a Ph.D. student at Tokyo Denki University. His work focuses structural health monitoring in response to seismic ground motions. He is dedicated to research focused on developing innovative methods to enhance the resilience and safety of infrastructure in response to seismic ground motions.

Authors:

Takaki Aeba Tokyo Denki University
Tsuyoshi Fukasawa Tokyo Denki University