Perancangan Modul Predictive Maintenance Menggunakan Model Machine Learning untuk Prediksi Risiko Kerusakan MesinPada Sistem Maintenance Management

Abstract

Kegiatan pemeliharaan mesin di Divisi Engineering Plant II PT XYZ masih bersifat reaktif, dengan downtime tahun 2025 mencapai 283 jam atau 7,15% dari total 3.958 jam operasi. Penelitian ini bertujuan merancang Modul Predictive Maintenance berbasis web untuk menghasilkan peringatan kerusakan, tingkat keparahan kerusakan, dan estimasi interval downtime. Penelitian menggunakan pendekatan desain keteknikan melalui analisis kebutuhan, pengembangan model machine learning, perancangan sistem, serta validasi fungsional dan pengguna. Pengembangan menggunakan 267 data aktual dan 1.000 observasi sintetis pada setiap target. Model terpilih meliputi Logistic Regression untuk breakdown warning, Random Forest untuk breakdown severity dan estimasi downtime kategori low serta high severity, dan Extra Trees untuk kategori medium severity. Rata-rata akurasi kelima target 83,70%. Seluruh fungsi modul dinyatakan valid melalui black-box testing, sedangkan User Acceptance Test menghasilkan tingkat penerimaan 83,20% dalam kategori baik. Modul yang dirancang digunakan sebagai prototipe pendukung keputusan maintenance, tetapi masih memerlukan validasi menggunakan data operasional aktual.

Machine maintenance in the Engineering Division of Plant II, PT XYZ, remains reactive, with 2025 downtime reaching 283 hours or 7.15% of 3,958 total operating hours. This study aimed to design a web-based Predictive Maintenance Module that provides failure warnings, failure severity levels, and downtime intervals. An Engineering design approach was applied through requirement analysis, machine learning model development, system design, and functional and user validation. Model development used 267 actual maintenance records and 1,000 synthetic observations for each prediction target. The selected models were Logistic Regression for breakdown warning, Random Forest for breakdown severity and downtime estimation in the low- and high-severity categories, and Extra Trees for the medium-severity category. The mean accuracy across the five prediction targets was 83.70%. All module functions were declared valid through black-box testing, while the User Acceptance Test achieved an acceptance rate of 83.20%, categorized as good. The developed module can be used as a maintenance decision-support prototype but still requires further validation using actual operational data.