A generalization-oriented fusion model for real-time TBM load prediction

The accurate real-time prediction of load parameters is crucial for the design and operation of Tunnel Boring Machines (TBMs). However, the limited quantity and diversity of training data often result in predictive models with poor generalization. Poor generalization severely restricts the practical deployment of deep learning models in engineering construction. To address this issue, this paper proposes a fusion model for real-time load prediction that is compatible with various types of TBMs, thereby expanding the quantity and diversity of the training data and ultimately improving model generalization. The fusion model integrates two sub-models: a low-frequency trend prediction model and a high-frequency wave prediction model, ensuring compatibility for projects with diverse TBM types, geological conditions, and TBM data frequencies. Moreover, it is trained using a hybrid dataset collected from eight distinct TBM projects. Analyses show that the fusion model achieves superior predictive performance on the testing dataset compared with other time-series prediction techniques. The accuracy of the trend prediction model increases with a larger window size, although the rate of improvement gradually declines. Moreover, the fusion model trained on data from multiple projects using different TBM types exhibits significantly stronger generalization than the model trained on a single-project dataset when evaluated on a completely unseen project. This finding confirms that greater diversity and richness in the training dataset effectively enhance the model’s generalization capability.