Machine Learning-Based Tri-Sensor Fusion for Predictive Maintenance: Operational Reliability and Techno-Economic Validation of Rotating Machinery on Oil and Gas Platforms

Predictive maintenance (PdM) improves the reliability of rotating machinery on oil and gas platforms; however, few studies translate multi-sensor classification into validated reliability and economic outcomes under real industrial deployment. This study develops and operationally validates a deployment-oriented Tri-Sensor decision-support framework integrating vibration, ultrasound, and surface temperature measurements with supervised machine learning for semi-critical rotating machinery on oil and gas platforms. Using 11,136 inspection records from 464 assets across five platforms (2022–2025), four classifiers (Logistic Regression, Support Vector Machine, K-Nearest Neighbors, Random Forest) were evaluated under Bi-Sensor (vibration and temperature) and Tri-Sensor (vibration, ultrasound, and temperature) configurations. The Tri-Sensor configuration achieved a Macro-F1 score of approximately 0.97, representing an improvement of approximately 20 percentage points over the Bi-Sensor baseline, confirmed by paired McNemar’s tests (p < 0.001). The results demonstrate that ultrasound integration substantially improves diagnostic performance. Correlation analysis, permutation feature importance, and SHAP interpretation confirmed that all three sensing modalities provide complementary information and contribute to predictions. Beyond classification, model outputs were transformed into a calibrated Health Index and assessed using Cumulative Gains Analysis and Decision Curve Analysis for risk-based maintenance prioritization. Using four years of maintenance and financial records, the framework was associated with an increase in Mean Time Between Failures from 5.16 to 6.44 years. Under the evaluated 20% follow-up scenario, estimated benefits included a 60% reduction in follow-up workload and potential savings of 1108 maintenance labor hours. The five-year assessment yielded a net present value of USD 2.73 million and a return on investment of 226%, while the modeled HSE analysis estimated a 20% reduction in major-incident frequency. These findings support a deployment-oriented framework linking sensor-fusion-based condition monitoring to reliability, maintenance, economic, and safety indicators.