Attitude deviation prediction of shield tunneling machine using Time-Aware LSTM networks

Abstract

In shield tunneling projects, the precise prediction and control of the shield machine's attitude is critical for quality assurance. Existing prediction methods utilize historical data to construct machine learning frameworks for forecasting future attitude deviations. However, these methods typically require data collected at uniform time intervals. In engineering practice, data collection intervals can sometimes be irregular. These irregularities can undermine the accuracy of predictions when still assumed to be uniform. To address this challenge, this study presents an attitude deviation prediction model based on the principles of the Time-Aware Long-Short-Term Memory (Time-Aware LSTM) method. This model tackles the irregular time intervals by employing a subspace decomposition process for cell memory. It incorporates the concept of time decay, allowing for a gradual discounting of memory content for different historical data in a manner that accurately reflects their respective elapsed times. Furthermore, this study discusses the intrinsic significance of the time decay function, drawing insights from the characteristics inherent in shield tunneling behaviors. Validated in a practical field application, the presented model demonstrated superior accuracy in predicting attitude deviations compared to existing methods. The presented model and findings may provide a reference for advancing adaptive control technologies in shield tunneling.