Online Learning Deep Neural Network Fuzzy Control of Structures Under Earthquake Motions: Numerical and Experimental Tests
Abstract
The conventional control methods typically assume that the controlled structure behaves as a deterministic system, overlooking variations in structural dynamic properties and uncertainties in earthquake motions. To overcome this constraint, this study introduces an innovative approach: an online learning deep neural network fuzzy control method (DFNN). In the proposed method, Offline training was conducted using training samples generated through the Linear–quadratic regulator (LQR) to determine the initial parameters of DFNN. An ON–OFF system was introduced for real-time control signal adjustment. The input and corrected control signals were utilized as training samples to train and modify the parameters of the DFNN system, enabling online learning capabilities. Numerical and experimental investigations were performed to evaluate the effectiveness of passive control (OFF), fuzzy logic control, deep neural network fuzzy control, and online learning deep neural network fuzzy control using a numerical three-story steel frame structure and an experimental two-story steel frame structure with one magneto-rheological (MR) damper. The simulation and shaking table test results demonstrate that DFNN can adaptively adjust the parameters of the neural network, leading to significantly higher control efficiency compared to fuzzy logic control and deep neural network fuzzy control, particularly when the structural properties and earthquake motions differ from the training samples.
Remember to check out the Most Cited Articles! |
---|
Remember to check out the structures |