MATLAB Machine Learning for a two-stage gearbox using multi-sensor data from gears for predictive maintenance

Abstract

Feed pellets, which hold a significant position in various industries, utilize gearboxes for the proper gear ratio required for pellet making. Moreover, these gearboxes are not only vital for pellet machines but for other machines as well in different industries. In this regard, their timely maintenance is a significant challenge that can be catered to by the use of modern techniques like CM and SHM. This study aims to use ML models for predictive maintenance. A multi-sensor technique comprising an acoustic and vibration sensor was used for a two-stage gearbox. There were three scenarios used in the current study, which were healthy, edge broken, and broken half tooth mounted on an intermediate shaft.  A gearbox with a combined ratio of 5.122 was used. The ADXL335 accelerometer and a MAX9814 microphone were used for extracting data. This data acted as input to MATLAB, which was then used for ML. All the built-in machine learning models in MATLAB were trained. The ML model named Fine Tree gave an initial accuracy of 48.8%. Then, by training the ML models by features instead of all the compiled data and modifying the code to detect data leakage, leak-free accuracy came out to be 63.98% and Fine Tree as 91.8%. The findings support that successful machine learning for predictive maintenance was done on the data extracted from multiple sensors for a 2-stage gearbox, and some future recommendations were also discussed.