Abstract: In order to extract machinery fault characteristics that are submerged in strong background noise, a general singular value decomposition (SVD) based subspace noise reduction algorithm is applied to signal processing, i.e., μ-SVD based denoising method. It can be proved that the traditional SVD based denoising method is a special case of the μ-SVD based one where μ=0. μ-SVD based denoising method contains a filter factor that plays a role in restraining information contributions of the noise-domain singular values to the denoised signal. μ-SVD based denoising method involves five parameters, including delay time, embedding dimension, noise reduction order, noise power and Lagrange multiplier. The selection methods for these parameters are discussed. In particular, the effects of noise reduction order and Lagrange multiplier on denoising performance are also studied. The experimental results of simulation signal with local fault and vibration signal with early crack fault in gear demonstrate that the μ-SVD based denoising method is superior to the traditional one in denoising performance, and can more effectively extract the gear fault characteristics at the presence of strong background noise.

Key words: fault diagnosis, filter factor, gear, μ-singular value decomposition, noise reduction