Planetary gear train microcrack detection using vibration, grayscale images and convolutional neural networks
Planetary Gear Trains (PGT) are widely used in many industrial applications from wind turbines to automobile transmissions due to its high power to weight ratio. Since PGT can be subjected to faults and prone to failure over time, a non-intrusive method of monitoring the condition of PGT is required. Vibration-based machine learning algorithms are mostly used in fault diagnosis and classification in PGT, but due to the epicyclic motion of gears; vibration signals from one gear can get neutralized or amplified by signals from another gear. Identification of smaller cracks in the sun gear can be challenging when using vibration-based fault diagnosis methods due to cancellation effects from the planet pinions. In this paper, the application chosen was based on the epicyclic drivetrain of a Chevy Volt hybrid car. The main research goal of this paper is to solve two research gaps; first is determining if extremely small elliptical cracks the size of 0.02 mm in length can be detected in vibration analysis and second is determining if sun gear faults neutralized by planetary gear faults can be detected. For that purpose 4 different health conditions of a planetary gear system were simulated using MSC ADAMS software. The conditions are a healthy PGT, PGT with a sun crack, PGT with a planet crack and a PGT with a sun and a planet crack. The joint forces at the exterior ring gear were extracted and a Blackman function was applied to convert to frequency domain values. Each of the frequency domain amplitude values was converted to pixel values in a grayscale image and the generated images were fed into a Convolutional Neural Network to train, validate and test the datasets. The results indicated that the proposed Grayscale – 2D CNN algorithm has an accuracy of 100% for the train and validation sets and 92% accuracy for the test set.