Health diagnosis of the rotating machinery can identify potential failure at its early stage and reduce severe machine damage and costly machine downtime. In recent years, the adaptive decomposition methods have attracted many researchers’ attention, due to less influences of human operators in the practical application. This paper compares two adaptive methods: local mean decomposition (LMD) and empirical mode decomposition (EMD) from four aspects, i.e., local mean, decomposed components, instantaneous frequency, and the waveletlike filtering characteristic through numerical simulation. The comparative results manifest that more accurate instantaneous frequency and more meaningful interpretation of the signals can be acquired by LMD than by EMD. Then LMD and EMD are both exploited in the health diagnosis of two actual industrial rotating machines with rub-impact and steam-excited vibration faults, respectively. The results reveal that LMD seems to be more suitable and have better performance than EMD for the incipient fault detection. LMD is thus proved to have potential to become a powerful tool for the surveillance and diagnosis of rotating machinery.

Traditional machine fault diagnosis techniques are labor-intensive and hard for non-experts to use. In this paper, a novel three-stage intelligent fault diagnosis approach is proposed for practical industrial process monitoring. A new feature processing technique is developed to enhance the identification accuracy and reduce computation burden, which incorporates variational mode decomposition (VMD)-based trend detection and self-weight algorithm. Furthermore, an adaptive density peaks search (ADPS) algorithm has been primarily proposed for adaptive clustering, whose effectiveness is verified in comparison with the original DPS, affinity propagation clustering and K-medoids. The three-stage intelligent fault diagnosis approach is subsequently applied to three specific industrial cases. Results of bearing and gear fault diagnosis have well demonstrated that the proposed method is able to reliably and accurately identify different faults with less prior knowledge and diagnostic expertise. Moreover, the proposed technique can be adopted to adaptively monitor different conditions using unlabeled bearing run-to-failure testing data, which also shows it is well suitable for industrial online applications.

Feature extraction plays an essential role in bearing fault detection. However, the measured vibration signals are complex and non-stationary in nature, and meanwhile impulsive signatures of rolling bearing are usually immersed in stochastic noise. Hence, a novel hybrid fault diagnosis approach is developed for the denoising and non-stationary feature extraction in this work, which combines well with the variational mode decomposition (VMD) and majoriation–minization based total variation denoising (TV-MM). The TV-MM approach is utilized to remove stochastic noise in the raw signal and to enhance the corresponding characteristics. Since the parameter λ is very important in TV-MM, the weighted kurtosis index is also proposed in this work to determine an appropriate λ used in TV-MM. The performance of the proposed hybrid approach is conducted through the analysis of the simulated and practical bearing vibration signals. Results demonstrate that the proposed approach has superior capability to detect roller bearing faults from vibration signals.

Rolling element bearings are one of the main elements in rotating machines, whose failure may lead to a fatal breakdown and significant economic losses. Conventional vibration-based diagnostic methods are based on the stationary assumption, thus they are not applicable to the diagnosis of bearings working under varying speeds. This constraint limits the bearing diagnosis to the industrial application significantly. A hybrid approach to fault diagnosis of roller bearings under variable speed conditions is proposed in this work, based on computed order tracking (COT) and variational mode decomposition (VMD)-based time frequency representation (VTFR). COT is utilized to resample the non-stationary vibration signal in the angular domain, while VMD is used to decompose the resampled signal into a number of band-limited intrinsic mode functions (BLIMFs). A VTFR is then constructed based on the estimated instantaneous frequency and instantaneous amplitude of each BLIMF. Moreover, the Gini index and time-frequency kurtosis are both proposed to quantitatively measure the sparsity and concentration measurement of time-frequency representation, respectively. The effectiveness of the VTFR for extracting nonlinear components has been verified by a bat signal. Results of this numerical simulation also show the sparsity and concentration of the VTFR are better than those of short-time Fourier transform, continuous wavelet transform, Hilbert–Huang transform and Wigner–Ville distribution techniques. Several experimental results have further demonstrated that the proposed method can well detect bearing faults under variable speed conditions.