Ensemble Learning based Model for Multi-Sensor Vibration Data Fusion in Gearbox Diagnosis

Document Type : Research Article

Authors

1 Department of Mechanical Engineering, University of Zanjan, Zanjan, Iran

2 Master of Mechanical Engineering, Department of Mechanical Engineering, University of Zanjan, Zanjan, Iran

Abstract

This study investigates recorded vibration signals from a laboratory gearbox to assess health condition and identify fault types, using a proposed ensemble-based machine learning algorithm. A single-stage gearbox was designed and tested in laboratory under four health states: no faults, root tooth crack, tooth breakage, and pitting on the tooth, across varying loads and speeds. Vibration measurements were recorded at six points (vertical, horizontal, and axial directions at input and output). A total of 792 signals (6 signals from 132 tests) were collected. For the data from each sensor, a Support Vector Machine (SVM) classifier with a linear kernel was trained, and fault detection accuracy was assessed and compared for each transducer individually. A new data fusion algorithm, inspired by Random Forest (RF), was developed to combine data from the six sensors. The results showed that the proposed ensemble algorithm provides higher detection accuracy than the individual classifiers for each sensor. In addition, a method is introduced to calculate the confidence level of the diagnoses from the proposed algorithm. It demonstrated that the ensemble algorithm can effectively diagnose faults with incomplete data (regardless of how many sensors are used from the total of six). As expected, using data from fewer sensors resulted in reduced accuracy and confidence levels for the detected fault.

Keywords

Main Subjects