TOWARDS AN EFFICIENT HEALTH CONDITION MONITORING STRATEGY APPLIED FOR GEARING MECHANICAL SYSTEMS

Abstract

Cracked teeth are a common phenomenon for gears. The dynamic behavior of gears in the presence of cracks and poor lubrication conditions that lead to rubbing between teeth are not well understood, which limits the development of precision and reliability of gear transmission. In this research, a comprehensive model is proposed to study the dynamic behavior of a one-stage spur gearbox with cracked teeth in a condition of poor lubrication. The inter-teeth rubbing generated from the sliding of one surface relative to another is caused by the variation of curvature radii of both mating teeth when the contact point is moving along the line of contact. Under the transmitted loading between mating teeth, the common elastic deformed area, computed by Hertz theory, keeps varying, generating a combination of modulated and constant amplitudes noises. Adding this friction-induced vibration to the impulsive periodic response is found to realistically mimic the actual behavior experimentally measured on a test rig specially developed for this investigation. Based on time-domain statistical indicators, the study concluded that the combination of both components of friction-induced noise with the primary impacting response was found to accurately and realistically simulate the dynamic behavior of the gearbox. Afterward, the experimental setup utilized in this investigation was modeled numerically in the initial part of the research to get a variety of operating circumstances. Those results are included in the dataset used in this work.Getting an adequate generalization capability by training traditional, deep 2-dimensional (2D) CNNs is a time-consuming operation, and this often requires massive datasets, exponentially increasing the processing challenges. Conversely, 1D CNNs have been suggested for use in various 1D signal processing tasks, such as gear condition monitoring. Recently developed methods aim to eliminate these drawbacks by focusing just on 1D signals, where they perform much better. This research addresses the challenge of failure identification in gears. It provides a solution based on combining raw and residual vibration data with a convolutional neural network (CNN) and recurrent neural network (LSTM). The described technique begins by training a one-dimensional convolutional neural network (CNN) using the raw vibration signals and then a long short-term memory (LSTM) network with the remaining vibration signals. Next, a deep learning structure for gear teeth fault identification is built by combining the two networks' findings on gear teeth fault characteristics. The provided approach is put through its paces by being applied to 41 unique gear fracture circumstances. The gear crack fault diagnostic result demonstrates that the provided technique achieves an accuracy of over 93% with little training data. The suggested approach provides more precise diagnostic outcomes than either the CNN or LSTM network alone. The reliability of the provided technique for gear fracture defect identification is shown by contrasting the results obtained using different sample sizes and methods. Further, a novel method for gear defect classification that integrates time-frequency analysis with image processing is proposed. This method may identify and categorize gear defects using the vibrating signals induced by cracked gear teeth. Empirical Mode Decomposition (EMD) and Principal Components Analysis (PCA) are used to deconstruct the signals into their principal components. To visualize the time- frequency connection of the primary components of the studied signal, the Short-Time Fourier Transform (STFT) is used to create the spectrogram for each element. Further, spectrogram pictures of primary components are converted into an array of features for each signal by extracting Image Moments. Then, a deep machine learning approach called a 2-dimensional convolutional neural network is used to achieve the classification (2-D CNN) utilizing image processing. As shown by the findings, the established method provides reliable classification, and the given deep structure can be readily expanded to include more sensor input signals for future gear crack failure diagnostics.