Dual heterogeneous attention-based deep learning model for multi-output prediction of TBM operations

Abstract

Predicting tunnel boring machine (TBM) performance in real-time is challenging due to the complex, dynamic, and multi-output nature of TBM operations. To address the challenges, this paper proposes a deep-learning method to provide an effective and efficient solution for predicting multi-output TBM performance in real-time, while also guiding TBM operations. This method integrates various essential components, including two parallel bi-directional long short-term memory (BiLSTM), a dual heterogeneous attention module (DHAM), a loss function, and evaluation metrics to ensure precise predictions while maintaining computational efficiency for real-time deployment. Experiments on real-world TBM operation data showcase the model's enhanced capabilities, achieved through the model featuring the learning rate of 0.00001, the batch size of 4, the full training set, the 2-step time window, the utilizations of the Nadam optimizer and the DHAM, and the ensemble of multiple modules. A comparative analysis reveals that the proposed method outperforms existing state-of-the-art models. This paper not only demonstrates the capabilities of the proposed method but also opens up opportunities for further advancements in utilizing deep learning to enhance decision-making processes and operational efficiency within the infrastructure construction fields.