Vibration-Based Gear Fault Diagnosis and Real-Time Condition Monitoring Using Machine Learning Techniques

Abstract

Modern industrial systems operating under varying speed and load conditions widely employ gear assemblies for smooth power transmission. Unexpected gear failures can lead to costly downtime and safety hazards. With the advancement of Industry 4.0, sensor-based predictive maintenance has become essential for early fault detection. Traditional condition monitoring methods rely heavily on human expertise and specialized instruments, which may fail to identify early-stage faults. This study presents a low-cost intelligent predictive maintenance framework for vibration-based gear fault diagnosis and real-time condition monitoring using machine learning techniques. Vibration signals were acquired from a laboratory-scale gearbox test rig using an ADXL345 accelerometer under three gear conditions: healthy, root crack, and broken tooth, over a speed range of 600–2400 RPM. After preprocessing, approximately 930 vibration samples were used to train and evaluate Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN) models. Among the implemented models, the ANN achieved the highest classification accuracy of 99.46% on the held-out test set, with precision, recall, and F1-scores of approximately 0.99 for all three classes and convergence within roughly 53 training epochs. A 5-fold stratified cross-validation of the ANN yielded a mean accuracy of 99.89 ± 0.22%, with the worst-fold accuracy matching the single-split result, confirming that the reported performance is stable and reproducible. In addition, a Python-based graphical user interface (GUI) using the trained ANN model, was developed for real-time condition monitoring, enabling continuous sensor-data acquisition, live fault prediction, vibration visualization, and automatic alarm activation during abnormal operating conditions. The proposed framework demonstrates high diagnostic reliability, low implementation cost, and practical applicability for Industry 4.0–based predictive maintenance systems.

Conclusion

This study proposes a low-cost and reliable AI-based condition monitoring system for gear fault diagnosis using vibration data from an ADXL345 accelerometer. Some key conclusions from this study are: • A cost-effective vibration-based condition monitoring framework was developed using the ADXL345 accelerometer and Arduino Uno, collecting data along three axes (X, Y, Z) under Healthy, Root Crack, and Broken Tooth gear conditions. • From over 1400 samples, 930 processed vibration samples (310 per class) were used to train and evaluate three AI models—Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN). • The ANN model achieved the highest performance, with an accuracy of 99.46%, outperforming RF and SVM. • A Tkinter-based real-time GUI was implemented to continuously classify incoming sensor data and visualize vibration signals. • An alarm system with distinct alerts for Root Crack and Broken Tooth conditions was integrated to enhance operational safety. This study has several limitations. Data came from a single laboratory-scale rig under controlled conditions, which may limit generalisability to industrial systems in variable, noisy environments. Only three gear conditions were considered, the consumer-grade ADXL345 has sensitivity constraints relative to industrial transducers, and the framework relies on a manually set speed input. Future work could explore deeper architectures such as CNN and LSTM networks for more complex faults, and transfer learning for cross-machine generalisation. Expanding the dataset across more fault types, loads, and gearbox configurations would improve robustness, while deployment on edge platforms such as Raspberry Pi or NVIDIA Jetson would enable autonomous, low-latency monitoring in industrial settings.

References