БАГАТОЗАДАЧНІ АРХІТЕКТУРИ РЕКУРЕНТНИХ НЕЙРОННИХ МЕРЕЖ ДЛЯ ІНФОРМАЦІЙНОЇ ТЕХНОЛОГІЇ МОНІТОРИНГУ БЕЗПЕКИ МОНОРЕЙКОВОГО КРАНА

Read full text Summarise with VIRAT
Authors: Oleksandr Terentyev, Bohdan Solovei
Cited by: 0
References: 14

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Ensuring the safety of monorail cranes is a critical task in modern construction. A contemporary solution involves the application of recurrent neural networks (GRU) for real-time data analysis. The aim of this study is to develop and comparatively analyze architectures of multitask recurrent neural networks for the simultaneous prediction of monorail crane stability with respect to overturning and sliding criteria. For this purpose, two architectures were developed and investigated: a multitask GRU with a self-attention mechanism and a multitask GRU with a Bayesian attention mechanism. The study includes the evaluation of regression accuracy (MAE, RMSE, R²), the effectiveness of binary classification of hazardous states (Precision, Recall, Specificity, F1), and the characteristics of predictive uncertainty (width of the 95% confidence interval, correlation between prediction error and uncertainty estimate). The Wilcoxon signed-rank test was applied to statistically assess differences between the models. The results demonstrate that both models provide high predictive performance (R² > 0.85). For the task of overturning stability prediction, the self-attention model showed statistically significantly higher accuracy (p < 0.001) and a 25.8% reduction in mean absolute error compared to the Bayesian model. For the sliding stability task, no statistically significant difference in regression accuracy was found (p = 0.54); however, the Bayesian model achieved higher recall in detecting hazardous states. Uncertainty analysis revealed a more conservative behavior of the Bayesian model, which generates wider confidence intervals, whereas the self-attention model demonstrated better uncertainty calibration for the sliding task (ρ = 0.93). Statistical tests confirmed significant differences in the models’ behavior regarding uncertainty estimation (p < 0.001). The findings indicate the feasibility of combining deterministic and Bayesian architectures in IT-based crane stability monitoring systems. Future research will focus on developing a decision support method that integrates simultaneous predictions from both models to form an adaptive risk indicator that accounts for both prediction accuracy and uncertainty.

Fields of Study & Subjects

Construction Engineering and Safety Structural Health Monitoring Techniques Neural Networks and Applications

Publication details

Journal / source
Herald of Khmelnytskyi National University Technical sciences
Volume
363
Issue
2
Pages
576–586
Publication date
26 March 2026
Language
EN