A two-layer approach to predictive maintenance using unsupervised anomaly detection and rule-based fault interpretation : industrial case study of a thermo-oil pump

Abstract

This thesis presents an Industry 4.0 method for assessing the health of industrial machinery using vibration data, emphasising the shift from schedule-based to data-driven maintenance. Developed in co-operation with SIA Latvijas Finieris at a veneer thermo-oil plant, the study describes an early-stage predictive maintenance strategy for the plywood industry. The case study examines a thermo-oil centrifugal pump used in veneer drying to detect mechanical degradation amidst high thermal and mechanical stress. The method uses a two-layer approach, combining unsupervised anomaly detection with Isolation Forest and a rule-based interpretation layer to identify and explain irregularities in machine behavior. Results are endorsed through physical inspection. This approach enables early fault detection and assists informed maintenance decisions in production facilities.