Comparison of machine learning methods for limited predictive maintenance

Authors : Timur Ozkul, Ayça Topallı

Pages : 183-191

Doi:10.28948/ngumuh.1465282

View : 48 | Download : 50

Publication Date : 2025-01-15

Article Type : Research Paper

Abstract :Predictive maintenance has gained increasing attention recently with the availability of sensors and connectivity of equipment. Yet, it would be difficult to obtain a wide range of data, especially with legacy devices. This paper describes an intelligent method for predicting a near future condition using the past information for an environment in which data are limited to the alarm logs from industrial machinery. Since machine learning methods are proven to be efficient in classification tasks using time series data, three of them are selected to predict an alarm two hours in advance using the past occurrences. These methods are neural networks, random forests, and extreme gradient boosting. The performances of these three methods are compared, and it is aimed to find the optimal configuration among hyper-parameter values. According to the obtained results, extreme gradient boosting gives the highest F1-score of 0.767 with number of trees equal to 500, maximum depth of 128, and an input window of alarm occurrences from the last day. This work consists of a comparative study aiming to identify the best machine learning method for alarm predictions, which potentially provides important insights into the operation and maintenance of machinery, bringing the possibility of considerable cost reductions.
Keywords : Makine Öğrenmesi, Sinir Ağları, Kestirimci Bakım, Rassal Orman, Aşırı Eğim Arttırma