# LSTM-based prediction method for shape error of steel truss during incremental launching construction

## TL;DR

This paper introduces an LSTM-based method to predict and control shape errors in steel truss bridges during construction, improving accuracy and reducing manual effort.

## Contribution

A novel LSTM-based real-time prediction method for shape errors during steel truss incremental launching is proposed.

## Key findings

- The LSTM model achieved an RMSE of 0.03 on the test set for shape error prediction.
- Field experiments showed predicted values aligned with measured values within 3 mm for short-term predictions.
- Model updates with measured data significantly reduced accumulated prediction errors.

## Abstract

Accurate shape control during steel truss incremental launching remains a persistent challenge in bridge engineering, primarily due to dynamic geometric variations induced by continuous spatial translation. Conventional measurement-based approaches often lead to inaccurate error determination and insufficient control criteria due to continuous geometric variations during structural movement. This study presents a novel Long Short-Term Memory (LSTM)-based methodology for real-time prediction and adaptive control of shape errors in launching processes. First, an error matrix is established based on the actual pushing measurement plan, and numerous error splines are generated using virtual assembly technology. These splines are output as an error matrix and encoded into a machine-readable format, leading to the establishment of a sliding window method for recursive prediction and updates of the predictions (measured values). Then, the LSTM model is trained using this sliding window approach, achieving a root mean square error(RMSE) of 0.03 on the test set. Field experiments demonstrate that the predicted values from the LSTM model closely align with the measured values, maintaining short-term shape error prediction accuracy within 3 mm. However, prediction accuracy diminishes for longer time steps as the step length increases. Following model updates with measured data, the accumulated prediction error rapidly decreases. The proposed prediction method for shape errors during pushing exhibits high accuracy and versatility in similar projects, significantly reducing time spent on manual error handling and minimizing computational inaccuracies.