Comparative analysis of Machine Learning and Deep Learning Algorithms using Fault diagnosis in Bevel gearbox

S. Ravikumar, Sharik N, C. Chandraprakash, V. Muralidharan, Syed Shaul Hameed

Abstract: The bevel gearbox is one of the most common types of gearbox or gearhead used in automation and energy transmission applications. The term refers to the gearbox’s tools, such as bevel gears, which are a single-stage unit that interlocks the beveled edges of gears and transfers rotation, much like interlocking fingers. The use of vibration measurement analysis for diagnosing gearbox failure has been proven to be efficient. This study compares several time-frequency signal processing methods for extracting diagnostic information from transient vibration signals. Accelerometer vibration measurements were used in experiments conducted on a bevel gearbox test rig. The discrete wavelet transform was originally used in vibration signal scrutiny to extract the frequency content of a signal from the appropriate frequency region. Then, to extract statistical features from vibration signals, several time-frequency signal processing methods were used, and their characteristic performance was compared. Due to the difficulties in selecting a suitable window length to capture the impulse signal, it was found that the Short Time Fourier Transform (STFT) could not give a good time resolution to detect the periodicity of the broken gear tooth. As a result, statistical features were used as inputs for a variety of deep and machine learning algorithms, yielding results such as a confusion matrix and a high-accuracy comparison. The algorithms used in this study were Random Forest, Decision Tree, K-Star, ANN, Naïve Bayes, and SVM. Of all the algorithms used in this study, LSTM yielded the highest accuracy of 98.65 percent.