An unsupervised context-free forecasting method for structural health monitoring by generative adversarial networks with progressive growing and self-attention

Abstract

Ensuring the robust operation of bridges demands swift and precise forecasting of structural performance within the health monitoring system. However, challenges arise in the realm of long-time series forecasting context-free data. These challenges encompass scenarios where there is a lack of reference data pre- and postforecasting, instances of missing data before forecasting (near-forecasting), or predictions of the distant future (far-forecasting). Addressing these issues, a current imperative is the development of a framework adept at efficiently and directly forecasting context-free long-time series data. This article introduces a framework, the convolutional generative adversarial network with progressive growing and self-attention (PSA-CGAN) mechanisms, tailored for forecasting context-free data. The approach employs generative adversarial networks in tasks related to long-time series. Additionally, progressive growing and self-attention mechanisms are harnessed to capture both long- and short-term features in the time series, notably enhancing the efficiency and accuracy of the forecasting method. The proposed method undergoes validation through application to two distinct bridge cases, confirming its generality and real-time forecasting prowess. On two bridges, PSA-CGAN can effectively predict acceleration data in various context-free scenarios and is capable of forecasting progressively changing damage data. It provides a valuable reference for predicting damage data. Additionally, the results indicate that PSA-CGAN is a promising and practical solution for the prediction of context-free data. It represents an efficient and rapid tool for damage prediction.