State estimation in structural dynamics through RNN transfer learning

Abstract

Model construction for state estimation is a pivotal concern in structural dynamics, driven by the need for effective control and health monitoring of structures. In general, state estimation models are derived from physics-based finite element (FE) models. However, the limited capability of FE models to simulate actual structures and the complexity of the environments in which these structures operate pose a significant challenge to the accuracy of model-based state estimation. This paper proposes a novel approach that leverages recurrent neural network (RNN)-based transfer learning to construct state estimation models, aiming to enhance the accuracy of state estimation for actual structures even if the FE model is not accurate enough. A calibrated FE model is used to generate extensive response data under synthetic excitations. The data is then processed and integrated to train an RNN model specifically designed for state estimation. Considering the diverse sensors involved in real-world structure monitoring, this study innovatively utilizes the collected data in a dual-purpose manner. A portion of the data serves as input for the RNN model, while the complete dataset facilitates the transfer learning process for the RNN model. This strategy enables the RNN model to adapt to real-structure state prediction. To ensure effective convergence in transfer learning, a method is proposed where parameters within the RNN cells at the network's front end are fine-tuned, while those close to the output layers are frozen. This approach deviates from conventional transfer learning methods used for other neural networks, while it is particularly useful for RNN models designed for state estimation. Numerical and experimental studies demonstrate that the proposed RNN transfer learning approach could effectively integrate both model-generated and actual measurement data. Under the same data acquisition condition, the transfer learning-RNN models could achieve significantly higher accuracy than state estimation models that rely solely on FE models.