An Unsupervised Vibration Noise Reduction Approach and Its Application in Lubrication Condition Monitoring

Abstract

Accelerometers are sensitive devices that capture vibrational fault signatures from industrial machines. However, noise often contaminates these fault signatures and must be eliminated before analysis. A data-driven (DD) denoising algorithm capable of filtering useful vibrational fault signatures from background noises was derived in this study. The algorithm was first validated by comparing its denoised result with a numerically generated ideal signal with a known exact solution. The DD denoising approach reduced the Mean Squared Error (MSE) from 0.459, when no denoising was performed, to 0.068, indicating an 85.2% decrease in noise. This novel approach outperformed the Discrete Wavelet (DW) denoising approach, which had an MSE of 0.115. The proposed DD denoising algorithm was also applied to preprocess vibration data used for the real-time lubrication condition monitoring of the plastic injection molding machine’s toggle clamping system, thereby reducing false positive relubrication alarms. The false positive rates, when analysis was performed on the raw vibration and the DW denoised vibration, were 10.7% and 7.6%, respectively, whereas the DD denoised vibration yielded the lowest false positive rate at 1%. This low false positive rate of the DD denoised vibration indicates that it is a more reliable condition monitoring system, thereby making this technique suitable for the smart manufacturing industry.