A NOVEL APPROACH TO FORECASTING SHORT UNIVARIATE TUNNEL SETTLEMENT DATA THROUGH MULTI-TASK MULTI-CHANNEL TRAINING OF ADJACENT TUNNEL POINTS

Abstract

With the increased density of urban environments and increased connectivity globally, tunnel constructions have increased tremendously, and settlement monitoring and forecasting is crucial to reduce loss of life and resources. There are three methods of tunnel settlement forecasting, empirical, numerical and intelligent methods. Empirical and numerical methods are restricted by the need for assumptions for the model to work and do not perform well when used for different soil types. Intelligent methods were proved to have great potential in predicting tunnel settlement as they do not require complex understanding of the mathematical or physical mechanisms of the whole environment. Tunnel settlement data collection is difficult and it requires significant resources and time to collect sufficient datasets for intelligent methods. Hence a significant number of datasets are univariate time series data collected from measuring the vertical displacement along the longitudinal axis of the tunnel, making it difficult for deep learning models to be implemented. This study looked into the relationship between short univariate datasets collected from this scenario and found that the data from adjacent points are linearly correlated. Hence, a multi-task multi-channel method based on nested long short term memory was proposed to exploit the inter-point relationship and overcome the shortness of the data. However, the experiment was unable to prove the superiority of the proposed method over the baseline models and hence a recommendation for future work into using different machine learning methods, wavelet types for transformation and number of adjacent points to batch for multi-task prediction is proposed.