Neural network-based prediction of topside mass of an in-service jacket platform

The determination of the topside mass of a jacket platform is imperative in structural health monitoring (SHM). This study therefore proposed a novel neural network with denoising autoencoder (DAE) and multitask learning strategy to predict the topside mass of an in-service jacket platform based on available SHM measurements. The DAE was introduced to learn a denoised representation against the noise encountered by the measurements on the platform. Initially, a traditional multilayer perceptron network was established with regularization techniques to predict the topside mass. However, the result indicated that the network encountered overfitting problems or relatively large discrepancies with observed data. To overcome this problem, a multitask learning strategy was introduced to learn the vibration features of an idealized model, and the division-based layer was applied to theoretically compute the topside mass. Ultimately, the applied multitask strategy improved the generalization performance and training efficiency compared to traditional deep learning methods. The practical applicability of this method under random wave excitation was then verified and discussed.