Modeling, Analysis, and Fault Prediction of Planetary Gearboxes Using Vibration Analysis, Signal Processing, and Machine Learning

نوع: Type: Thesis

مقطع: Segment: PHD

عنوان: Title: Modeling, Analysis, and Fault Prediction of Planetary Gearboxes Using Vibration Analysis, Signal Processing, and Machine Learning

ارائه دهنده: Provider: Ali Hemati

اساتید راهنما: Supervisors: Dr.Alireza Shooshtari

اساتید مشاور: Advisory Professors:

اساتید ممتحن یا داور: Examining professors or referees: Dr. Mehdi Karimi, Dr. Seyed Ali Asghar Hosseini, Dr. Kourosh Khorshidi

زمان و تاریخ ارائه: Time and date of presentation: 2026

مکان ارائه: Place of presentation: 54

چکیده: Abstract: Planetary gear systems, due to their high-power transmission capacity, reliable performance, and ability to achieve large reduction ratios within a compact structure, have become essential components in advanced power transmission mechanisms across various industries. The critical role of these gear sets is such that even minor faults can lead to unexpected shutdowns, reduced efficiency, increased maintenance costs, and, in severe cases, structural damage to industrial equipment. Therefore, early detection of faults in these mechanisms plays a pivotal role in maintenance management and improving system reliability. In the present study, common faults in planetary gear assemblies are identified and analyzed, and diagnostic strategies based on vibration analysis, signal processing, and machine learning are proposed. To achieve this, a detailed 18-degree-of-freedom dynamic model of a planetary gear system is developed, incorporating extremely small backlash levels (below 12 micrometers), time-varying mesh stiffness, and dynamic interactions between planetary components. Several fault scenarios—including increased gear backlash, rolling-element bearing defects in the sun gear, planetary gear tooth wear, and pitting of the contact surfaces—are investigated under different excitation conditions. For model validation, two dynamic configurations—one with constant mesh stiffness and another with time-varying mesh stiffness—are employed. Furthermore, vibration data obtained from field experiments are used to extract fault-specific frequency formulations and diagnostic indicators. These characteristic frequencies are then applied to the dynamic model, and the simulation results exhibit strong agreement with the measured vibration responses, confirming the model’s accuracy in predicting fault-related signatures. Finally, to develop an intelligent condition-monitoring framework, various machine learning and deep learning algorithms are employed for automatic fault detection. Due to the limited availability of data, only two frequency-domain signatures—representing healthy and faulty conditions—are used for training. The performance of three classification approaches, namely Support Vector Machine (SVM), Random Forest (RF), and Long Short-Term Memory (LSTM) neural networks, is evaluated and compared based on these limited datasets. The results demonstrate that each method exhibits different capabilities in extracting discriminative features and separating healthy from faulty patterns. This comparison enables the identification of the most suitable algorithm for condition-monitoring applications under data-scarce environment.