Memory unit-based autoencoder method for health monitoring of offshore substation platforms

A novel health monitoring method for offshore steel structure platforms, based on a memory unit autoencoder, is introduced. De-noising and angle domain re-sampling of high-frequency vibration data from these platforms amplify fault-related features, enabling a deep-convolutional neural network-based memory autoencoder to learn historical data patterns. The establishment of a hidden state memory library allows sample encoding features to sparsely match with the memory library, effectively capturing the complex working conditions of the platform. Subsequently, a Gaussian mixture probability model is employed to model the association scores of hidden states encoded by the deep learning model, providing an assessment of the steel structure's health status. The application of this method to an offshore boosting station in Rudong County, Jiangsu, verifies its effectiveness. Overall, this research presents a new technical approach and theoretical foundation for offshore steel structu re platform health monitoring.