Detecting Helical Gearbox Defects from Raw Vibration Signal Using Convolutional Neural Networks

Area of Science:

• Mechanical Engineering
• Machine Learning
• Vibration Analysis

Background:

• Gearbox defects impact machinery reliability and performance.
• Early detection of gear faults is crucial for preventing catastrophic failures.
• Vibration analysis is a common technique for condition monitoring.

Purpose of the Study:

• To develop and evaluate convolutional neural network (CNN) models for gearbox defect detection.
• To assess the effectiveness of 1D-CNN and 2D-CNN architectures using raw vibration signals.
• To identify optimal sensor axes for defect classification.

Main Methods:

• Raw vibration signals from a triaxial accelerometer were analyzed.
• 1D-CNN and 2D-CNN models were employed for feature extraction and classification.
• Models were trained and tested on data from various rotating velocities and load levels.

Main Results:

• The best 1D-CNN model (Y-axis data) achieved 98.91% testing accuracy.
• Data from X and Z axes yielded slightly lower accuracies (97.15% and 97%).
• A 2D-CNN model using all three axes reached a high accuracy of 99.63%.

Conclusions:

• CNN-based models are highly effective for gearbox defect detection.
• Multi-axis data fusion in a 2D-CNN architecture enhances detection accuracy.
• Vibration analysis combined with deep learning offers a robust solution for gearbox condition monitoring.