Investigation of the applicability of recurrent neural networks for structural health monitoring in the frequency domain

Structural health monitoring (SHM) aims to detect or predict state changes or damages in engineering structures. In order to discriminate between various damage characteristics and locations, the SHM system requires relevant information about the structure as well as a suitable method to evaluate these. This paper explores a data-driven SHM approach that models damage processes using recurrent neural networks. As model input data, the machine learning algorithm uses sequential frequency domain data at consecutive steps during the advancing damage process. From the change in transfer functions that occurs during structural changes, the model derives information about the state of the monitored object. In order to discuss the potentials and limitations of this modeling approach, a simple bolted structure with non-trivial changes in the frequency response is employed. A gradual damage process is simulated by incrementally loosening one of the joints. The resulting sequences of transfer functions are used as input to the recurrent neural network model and related to the respective preload force. By varying the data sequences used for model training and application, the functioning of the modeling process is investigated. The possibility of inversely learning from the model about damage indicators by analyzing effective input values is discussed.