Predicting disc cutter wear using two optimized machine learning techniques

The estimation of disc cutter wear (CW) remains a complex problem in mechanized tunneling using tunnel boring machines (TBM), despite the development of numerous TBM performance models. This research aimed to estimate the cutter life index (CLI) as an index to predict the CW by developing predictive models based on two machine learning algorithms, namely gradient boosting (GB) and random forest (RF), optimized by three optimization techniques: particle swarm optimization (PSO), differential evolution (DE), and simulated annealing (SA). To gain the aim, a dataset consisting of four rock parameters-density (ρ), uniaxial compressive strength, Brazilian tensile strength (BTS), and brittleness index-with 80 mechanized tunnel cases for each parameter has been utilized by obtaining the sample and then relevant tests on them were conducted in the laboratory. First, various parameter selection methods, such as mutual information, have been employed to reduce the dimensionality of the problem, and it has been revealed that ρ and BTS have been the most influential parameters to estimate the CLI. Then, by developing six optimized models, including GB-PSO, GB-DE, GB-SA, RF-PSO, RF-DE, and RF-SA, using the two mentioned parameters, their performance has been assessed via three performance evaluation indices of coefficient of determination (r2), root mean square error (RMSE), and mean absolute percentage error (MAPE). The results showed that among six predictive models, the two models of GB-SA (with r2, RMSE, and MAPE of 0.8274, 10.8329, and 0.3957, respectively) and RF-PSO (with r2, RMSE, and MAPE of 0.8213, 11.0249, and 0.4113, respectively) outperformed the other models, with 82.74% accuracy of GB-SA and with 82.13% accuracy of the RF-PSO, and the two can be utilized to estimate disc cutter via CLI for different type of rock in the range of established dataset.

---

Opracowanie rekordu ze środków MNiSW, umowa nr POPUL/SP/0154/2024/02 w ramach programu "Społeczna odpowiedzialność nauki II" - moduł: Popularyzacja nauki (2025).