KLI

변속기 고장 진단 시스템의 진동 신호 분석을 위한 잡음 제거 기법

Metadata Downloads
Abstract
Gearbox fault diagnosis based on the vibration characteristic analysis has been widely developed and applied in research and industrial fields. Vibration characteristic of the gearbox systems accommodates the fault-related information of the rotating machines. In a simple gearbox, there are two engaged rigid blocks including a pinion wheel and a gear wheel for transferring the motion from the source to the load. The time variable gear-mesh stiffness causes the internal excitations during operation. If one or some of the teeth are faulty, abnormal movements will be appeared making the impulsive events in the vibration characteristic of a gearbox system. The vibration characteristic is normally sensed by vibration sensors (accelerometers) mounted on a gearbox sink. Thus the gearbox fault diagnosis model can be constructed based on the method of vibration signal analysis (VSA) to highlight the fault-related information. However, the gearbox vibration signals in real world are very complicated acting as non-stationary, nonlinear, and noisy overwhelming because of the inconsistent operation condition of a gearbox such as speed variation, fluctuation of the load, and the influence of mechanical resonances of other components. In addition, the several fault types create a similarly behavioral reflection on the gearbox vibration characteristic challenging the fault type discriminating process. Therefore, the enhanced denoising approaches and accurate fault type identifying methods are critically needed to construct the accurate and stable gearbox fault diagnosis model.
This study aims to propose novel adaptive denoising methodologies for filtering the original fault relative components from the raw noisy vibration signals regarding to variable operating speeds condition of a gearbox. Moreover, the AI based fault identification models are developed to process the fault-related information for accurately discriminating the fault types of a gearbox system. These advanced approaches of signal processing, feature engineering, and classification are incorporated to construct the sensitive and stable fault diagnosis frameworks for a multi-level gear fault gearbox under varying rotational speeds conditions. This thesis addresses four research keynotes.
For the first research topic, the adaptive noise reducer based Gaussian reference signal (ANR-GRS) approach is created for denoising the raw vibration signals. The ANR-GRS technique is established by following processes: first of all, the vibration characteristics of a gearbox system are structurally analyzed to classify noise types; then the speed-dependent Gaussian reference signals with adjustable parameters are generated, according to those noise types; finally, these generated noise-simulated reference signals were adaptively adjusted and accessed to the space between two consecutive fault-related frequency components and reduce the interference noise along with the whole frequency range of raw vibration signals. After denoising, the manual feature extraction method for extracting the optimized vibration subbands, outputted from the ANR-GRS, to many statistical features in time and frequency domains. Those fault representation features are used to input to a one-against-one support vector machine (OAOSVM) classifier for the fault type classification. The gearbox fault diagnosis scheme is validated for identifying the three fault types and one healthy state of the experimental testbed of a spur gearbox under varying speeds conditions. The result shows that the disturbance noises are significantly removed by using ANR-GRS method, thus feature extraction and OAOSVM based classifier provide excellent fault identification accuracy.
Secondly, the research topic focuses on the combined application of the adaptive noise control (ANC) method and genetic algorithm (GA) based feature selection to draw the sensitive fault diagnosis scheme. In this scheme, the applied adaptive noise control approach performs significantly removing noise elements and keeping original fault relative information from gearbox vibration signals. The outputs of ANC, optimized subbands, are then statistically extracted to many features configuring a feature pool. GA operates a heuristic searching process to select the most discriminative fault features, that represent samples of each fault type in clear separation allowing a simple machine learning model such as k-nearest neighbor (k-NN) for classifying defect categories into the respective types. This model is applied to classify six defective categories of a gearbox with multi-level gear defects. The accuracy result verifies the effectiveness of the combination fault identification model.
In the third research topic, the adaptive noise control (ANC) based Gaussian reference signal and stacked sparse autoencoder based deep neural network (SSA-DNN) are employed in combination for constructing a sensitive and speed invariant fault diagnosis model. The applied model is used for diagnosing seven health states of the multi-level tooth cut gear defects (MTCG) gearbox under variable speed conditions. The deep learning model is built up by stacking the sparse autoencoder layers as the hidden layers and using a Softmax layer as the output layer of the network. SSA-DNN is capable to extract the spectra of the optimal vibration subbands, significantly denoised by ANC, into the high dimension feature pool of latent representative fault features, then selecting the most fault discriminant features for identifying the MTCG fault types under various speeds conditions. The effective evaluation of the proposed fault diagnosis scheme is verified by the classification result of the experiment on the vibration signal dataset of an MTCG gearbox collected under four different rotational speeds. The experiment is arranged by four sub-experiments using the datasets corresponding to four rotational speeds. In each sub-experiment, the network model is trained using a one-speed dataset and tested by two other speed datasets. The highest accuracy results are achieved, which outperform the state-of-the-art methodologies, validating the sensitive and speed invariance capabilities of the proposed fault diagnosis model in this research topic.
For the fourth study, the new localized adaptive denoising technique (LADT) is developed based on the ANR-GRS approach for improving the efficiency of noise reduction. Thus, an accurate and stable gearbox fault diagnosis scheme, that combines LADT with wavelet-based vibration imaging approach and deep convolution neural network model, is established. The new localized adaptive denoising technique results in optimized vibration subbands with negligible background noise. To obtain fault-related information, the wavelet-based vibration imaging approach (WVI) is applied to the denoised vibration signal. The wavelet-based vibration imaging approach decomposes the vibration signal into different time-frequency scales, these scales are reflected by a two-dimensional image called a scalogram. The scalograms obtained from the wavelet-based vibration imaging approach are provided as an input to the deep convolutional neural network architecture (DCNA) for discriminant features extraction and classification of multi-degree tooth faults (MDTF) in a gearbox under variable speed conditions. The proposed scheme outperforms the already existing state-of-the-art gearbox fault diagnosis methods with the highest classification accuracy.
Author(s)
웬 꽁 다이
Issued Date
2022
Awarded Date
2022-02
Type
dissertation
URI
https://oak.ulsan.ac.kr/handle/2021.oak/10049
http://ulsan.dcollection.net/common/orgView/200000595401
Alternative Author(s)
Nguyen Cong Dai
Affiliation
울산대학교
Department
일반대학원 전기전자컴퓨터공학과
Advisor
김종면
Degree
Doctor
Publisher
울산대학교 일반대학원 전기전자컴퓨터공학과
Language
eng
Rights
울산대학교 논문은 저작권에 의해 보호 받습니다.
Appears in Collections:
Computer Engineering & Information Technology > 2. Theses (Ph.D)
Authorize & License
  • Authorize공개
Files in This Item:

Items in Repository are protected by copyright, with all rights reserved, unless otherwise indicated.