ML Intern @ IIT Roorkee
Recently, I worked as a machine learning research intern at IIT Roorkee during the summer, collaborating with Professor R. S Anand on a project focused on Health Condition Monitoring of Ball Bearings using Vibration Signals. My responsibilities included signal processing, vibrational data analysis, implementing machine learning pipelines, and feature extraction.
Recently, I worked as a machine learning research intern at IIT Roorkee during the summer, collaborating with Professor R. S Anand on a project focused on Health Condition Monitoring of Ball Bearings using Vibration Signals. My responsibilities included signal processing, vibrational data analysis, implementing machine learning pipelines, and feature extraction.
The primary issue addressed in our research was the impact of faults in ball bearings on the induction motor. Identifying and differentiating these faults, such as cracks in the outer ring, was crucial to preventing long-term damage to the motor. We discovered that different faults produced unique vibration signals, making it possible to differentiate them using machine learning algorithms.
Our approach involved three main phases. The initial phase focused on data collection and preprocessing, with an emphasis on normalizing signals and dividing them into frames. In the second phase, I conducted feature extraction in both time and frequency domains, resulting in around 14 statistical features. The final phase included the analysis of the produced work, featuring visualizations of statistical features, ROC curves, and feature graphs.
We achieved three key objectives. First, we successfully interpreted different faults’ profiles using time and frequency features. Second, we identified the best feature selection and training techniques, with Genetic Algorithm standing out as the optimal feature selection method, and Random Forest achieving the highest accuracy at 97%. Lastly, we prioritized proper documentation to ensure the reproducibility of our work.
In terms of results, our efforts demonstrated the effectiveness of our approach. We interpreted fault profiles, identified optimal techniques, and achieved a high accuracy rate of 97% with the fastest algorithm being Random Forest. Additionally, we developed a web-based application for signal analysis and machine learning model training, enhancing accessibility and usability.
In conclusion, our research successfully addressed the challenges of health condition monitoring in ball bearings, providing valuable insights and effective solutions. The combination of advanced signal processing, feature extraction, and machine learning proved to be a powerful approach in identifying and mitigating faults, with practical applications demonstrated through our web-based tool.