이 연구에 대하여
This research presents innovative solutions by applying artificial intelligence, specifically deep learning techniques, to the data analysis of architectural structures. By accurately predicting the complex and nonlinear behavior of structures and effectively addressing data corruption and noise issues inherent in measurement processes, it has opened new horizons in structural health monitoring and maintenance. This will make a significant contribution to the future development of smart construction technologies.
주요 발견
- • ANFIS demonstrated significantly lower error (RMSE) than traditional RNNs, exhibiting superior accuracy in dynamic response simulations.
- • This study demonstrates the successful restoration of high-frequency information in dynamic data acquired at low or lossy sampling rates using ANFIS.
- • ANFIS effectively removed outliers (noise) from the repeated load test data of the RC column, however, the observed peak value smoothing tendency requires further refinement.
핵심 개념
심층 분석
ANFIS: Insightful Prediction of Nonlinear Systems
This study employs Adaptive Neuro-Fuzzy Inference Systems (ANFIS) as its core methodology. ANFIS is a deep learning algorithm combining the learning capabilities of artificial neural networks with the human-like reasoning capabilities of fuzzy logic. It is particularly effective in modeling and predicting the behavior of complex, non-linear systems. In simulations of forced vibration response in a real five-story steel structure, ANFIS demonstrated significantly higher accuracy than conventional Recurrent Neural Networks (RNNs), successfully reproducing even minute frequency components, thereby proving its superiority. This establishes a foundation for detecting and predicting even subtle changes in the structure.
Data Revitalization: Value Creation through Restoration and Refinement
ANFIS has been employed beyond simple prediction to revitalize damaged or incomplete measurement data. It successfully restored information lost in high-frequency bands due to low sampling rates, thereby enhancing the data's usability. Furthermore, it effectively filtered electrical noise contaminating data from repeated load tests on RC columns, improving the reliability of the analysis. While further improvement in peak value processing is needed, this data cleansing and restoration capability will play a crucial role in utilizing unstable data collected in real-world applications.
그림 및 표
Figure 7: Five-story Test Building and Sensor Installation
This figure shows the five-story steel structure used in the study and the experimental equipment installed on the building's top floor. Accelerometers for measuring building vibration data and a shaker for artificially generating vibrations are the core equipment. Data on the actual structure's response, collected using this equipment, were used for deep learning model training and validation.
Figure 8: Simulation Results of Top-Layer Acceleration Response Based on RNN
This figure compares the predicted top-floor acceleration response of a building, generated by a recurrent neural network (RNN) model, with measured data. The graphs illustrate the agreement between simulation results and actual data in both the time domain (left) and frequency domain (right). The root mean square error (RMSE) shown in each graph indicates the prediction accuracy of the model, highlighting limitations of the RNN model.