A critical challenge in structural control is the design of controllers capable of rapid adaptation to uncertainties, including changes in local dynamics and unknown parameters. A solution is the use of data-based controllers, but these typically require pre-training and may be computationally expensive to utilize. In this paper, the authors proposed a fast neuro-controller capable of pure online learning. The controller is built using an ensemble of recurrent neural networks (ERNNs) based on long short-term memory (LSTM) cells running in parallel. A multi-rate sampler is used to sequentially construct the individual RNN input vectors using local measurements at sampling rates pre-determined from the structure's natural frequencies. Each RNN performs an individual estimation with the associated sampling rate, and these estimations are assembled into an attention and a dense layer to produce an estimated control force. Critical advantages of the proposed architecture are (1) rapid online adaptation and processing; (2) capability to adapt to local dynamics and non-stationarities; and (3) pure online learning architecture minimizing reliance on pre-training. First, the performance of the ERNNs control algorithm is validated on an actively controlled single-degree-of-freedom system, and stability of the controller is demonstrated. Second, the ERNNs is simulated on a full-scale 20-story building equipped with a semi-active tuned liquid wall damper (TLWD) system that consisted of distributed semi-active tuned liquid multiple column dampers. The ERNNs controller effectively improved the passive performance of the control strategy and performed similarly to that of a full state-feedback sliding mode controller under multi-hazard scenario. The evolution of attention weights under both the wind and seismic hazards demonstrated that the importance of each RNN follows the importance of the structural frequency responses, showing that the controller was capable of capturing and adapting fast-changing dynamics.
Ensemble of long short-term memory recurrent neural network for semi-active control of tuned liquid wall damper
Ubertini F.;
2022
Abstract
A critical challenge in structural control is the design of controllers capable of rapid adaptation to uncertainties, including changes in local dynamics and unknown parameters. A solution is the use of data-based controllers, but these typically require pre-training and may be computationally expensive to utilize. In this paper, the authors proposed a fast neuro-controller capable of pure online learning. The controller is built using an ensemble of recurrent neural networks (ERNNs) based on long short-term memory (LSTM) cells running in parallel. A multi-rate sampler is used to sequentially construct the individual RNN input vectors using local measurements at sampling rates pre-determined from the structure's natural frequencies. Each RNN performs an individual estimation with the associated sampling rate, and these estimations are assembled into an attention and a dense layer to produce an estimated control force. Critical advantages of the proposed architecture are (1) rapid online adaptation and processing; (2) capability to adapt to local dynamics and non-stationarities; and (3) pure online learning architecture minimizing reliance on pre-training. First, the performance of the ERNNs control algorithm is validated on an actively controlled single-degree-of-freedom system, and stability of the controller is demonstrated. Second, the ERNNs is simulated on a full-scale 20-story building equipped with a semi-active tuned liquid wall damper (TLWD) system that consisted of distributed semi-active tuned liquid multiple column dampers. The ERNNs controller effectively improved the passive performance of the control strategy and performed similarly to that of a full state-feedback sliding mode controller under multi-hazard scenario. The evolution of attention weights under both the wind and seismic hazards demonstrated that the importance of each RNN follows the importance of the structural frequency responses, showing that the controller was capable of capturing and adapting fast-changing dynamics.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.