Comparison of wavelet based denoising schemes for gear condition monitoring: An Artificial Neural Network based Approach

International Conference on Advances in Materials and Manufacturing Applications 2017

Ahmed, Rounaq^1 ; Srinivasa Pai, P.^2 ; Sriram, N.S.^3 ; Bhat, Vasudeva^4

Department of Mechanical Engineering, VTU, PACE, Mangalore, Karnataka, India^1

Department of Mechanical Engineering, NMAMIT, Nitte, VTU, Karnataka, India^2

Department of Mechanical Engineering, VVIET, VTU, Mysore, Karnataka, India^3

Department of Mechanical Engineering, SIT, VTU, Mangalore, Karnataka, India^4

关键词: Adaptive thresholds;    Artificial neural network models;    Classification accuracy;    De-noised signals;    Empirical Mode Decomposition;    Fault identifications;    Gear fault diagnosis;    Root mean square errors;   

Others  :  https://iopscience.iop.org/article/10.1088/1757-899X/310/1/012010/pdf
DOI  :  10.1088/1757-899X/310/1/012010

来源: IOP

【 摘 要 】

Vibration Analysis has been extensively used in recent past for gear fault diagnosis. The vibration signals extracted is usually contaminated with noise and may lead to wrong interpretation of results. The denoising of extracted vibration signals helps the fault diagnosis by giving meaningful results. Wavelet Transform (WT) increases signal to noise ratio (SNR), reduces root mean square error (RMSE) and is effective to denoise the gear vibration signals. The extracted signals have to be denoised by selecting a proper denoising scheme in order to prevent the loss of signal information along with noise. An approach has been made in this work to show the effectiveness of Principal Component Analysis (PCA) to denoise gear vibration signal. In this regard three selected wavelet based denoising schemes namely PCA, Empirical Mode Decomposition (EMD), Neighcoeff Coefficient (NC), has been compared with Adaptive Threshold (AT) an extensively used wavelet based denoising scheme for gear vibration signal. The vibration signals acquired from a customized gear test rig were denoised by above mentioned four denoising schemes. The fault identification capability as well as SNR, Kurtosis and RMSE for the four denoising schemes have been compared. Features extracted from the denoised signals have been used to train and test artificial neural network (ANN) models. The performances of the four denoising schemes have been evaluated based on the performance of the ANN models. The best denoising scheme has been identified, based on the classification accuracy results. PCA is effective in all the regards as a best denoising scheme.