Abstract
Disc cutters are essential for shield tunnel construction, and monitoring their wear is vital for safety and efficiency. Due to their position in the soil silo, it is more challenging to observe the wear of disc cutters directly, making accurate and efficient detection a technical challenge. However, existing methods that treat the problem as a classification task often overlook the issue of data imbalance. To solve these problems, this paper proposes an end-to-end detection method for disc cutter wear state called the Multivariate Selective Attention Prototype Network (MVSAPNet). The method introduces an attention prototype network for variable selection, which selects important features from many input parameters using a specialized variable selection network. To address the problem of imbalance in the wear data, a prototype network is used to learn the centers of the normal and wear state classes, and the detection of the wear state is achieved by detecting high-dimensional features and comparing their distances to the class centers. The method performs better on the data collected from the Ma Wan Cross-Sea Tunnel project in Shenzhen, China, with an accuracy of 0.9187 and an F1 score of 0.8978, yielding higher values than the experimental results of other classification models.