[Paper Review] Application of 2-D Convolutional Neural Networks for Damage Detection in Steel Frame Structures

Shahin Ghazvineh, Gholamreza Nouri+3 more|arXiv (Cornell University)|Oct 29, 2021

Structural Health Monitoring Techniques41 references4 citations

TL;DR

This paper proposes a lightweight 2-D convolutional neural network (2-D CNN) framework that directly processes raw acceleration signals from steel frame structures to detect damage with high accuracy and low latency. By optimizing individual CNNs for specific structural elements and reducing training data requirements, the method achieves superior classification performance and real-time feasibility on the Qatar University Grandstand Simulator benchmark dataset.

ABSTRACT

In this paper, we present an application of 2-D convolutional neural networks (2-D CNNs) designed to perform both feature extraction and classification stages as a single organism to solve the highlighted problems. The method uses a network of lighted CNNs instead of deep and takes raw acceleration signals as input. Using lighted CNNs, in which every one of them is optimized for a specific element, increases the accuracy and makes the network faster to perform. Also, a new framework is proposed for decreasing the data required in the training phase. We verified our method on Qatar University Grandstand Simulator (QUGS) benchmark data provided by Structural Dynamics Team. The results showed improved accuracy over other methods, and running time was adequate for real-time applications.

Motivation & Objective

• To develop a real-time, accurate damage detection system for steel frame structures using raw vibration data.
• To reduce the data requirements for training deep learning models in structural health monitoring.
• To enhance model efficiency and accuracy by using specialized, lightweight 2-D CNNs per structural element.
• To integrate feature extraction and classification into a single end-to-end learning framework.
• To validate the method on a real-world benchmark dataset (QUGS) for practical applicability.

Proposed method

• The method employs a set of lightweight 2-D CNNs, each optimized for a specific structural element, to process raw acceleration signals directly.
• Input signals are transformed into 2-D representations (e.g., spectrograms or time-frequency maps) to enable spatial convolution operations.
• Each CNN is trained end-to-end to perform both feature extraction and classification in a single model architecture.
• A novel framework is introduced to reduce the amount of training data required without compromising performance.
• The network design emphasizes computational efficiency to support real-time deployment on structural monitoring systems.
• The approach is evaluated on the Qatar University Grandstand Simulator (QUGS) dataset, a standard benchmark in structural dynamics.

Experimental results

Research questions

• RQ1 Can a lightweight 2-D CNN framework achieve high accuracy in damage detection using raw acceleration signals from steel frames?
• RQ2 How does the proposed method compare to existing deep learning approaches in terms of accuracy and inference speed?
• RQ3 To what extent can the training data requirement be reduced while maintaining model performance?
• RQ4 Can the use of element-specific CNNs improve detection accuracy and model efficiency?
• RQ5 Is the proposed method suitable for real-time structural health monitoring applications?

Key findings

• The proposed 2-D CNN framework achieved higher classification accuracy than existing methods on the QUGS benchmark dataset.
• The model demonstrated low inference latency, making it suitable for real-time damage detection in structural health monitoring.
• The use of specialized, lightweight CNNs per structural element improved both accuracy and computational efficiency.
• The framework significantly reduced the amount of training data required compared to conventional deep learning approaches.
• The end-to-end learning architecture successfully combined feature extraction and classification without manual engineering.
• The method outperformed baseline models in both detection accuracy and processing speed, validating its practical feasibility.