Multi-Node Feature Learning Network Based on Maximum Spectral Harmonics-to-Noise Ratio Deconvolution for Machine Condition Monitoring

Since the cyclostationarity in vibration signals is the key to judge the rotating machine health state, spectral harmonics-to-interference ratio (SHIR) has been used to construct single node feature learning network (SHIR-based blind deconvolution, BD-SHIR) to realize condition monitoring of machine. However, BD-SHIR still has several obvious limitations, including the need for prior fault information and the tendency to fall into local optimal solutions which will affect its feature learning and state monitoring performance. Therefore, this paper proposes a multi-node feature learning network based on BDSHIR, which does not rely on prior information, multi-node adaptive BDSHIR (MABD-SHIR). It uses each filter in the 1/3 binary tree filter bank to design each initial node (filter) in the learning network correspondingly, and introduces an adaptive fault characteristic frequency (FCF) detection process. Since this initialization can give some filters close to the fault resonant band, iterations starting from these nodes are more likely to detect the precise FCF and converge to the optimal solution. Finally, MABD-SHIR can obtain multiple local optimal solutions, among which the one with the largest SHIR is generally closer to the global optimal solution than the one obtained by BD-SHIR. Simulation and experimental data of defective rotating machinery verify the effectiveness of MABD-SHIR in machinery condition monitoring and fault feature learning. Note to Practitioners—This paper is motivated by the problems that the rotating machine condition monitoring based on BDSHIR is easy to prematurely fall into local optimal solution and requires prior fault information. These problems can be effectively solved by extending BDSHIR from single node structure to multi-node structure and introducing adaptive period detection technique (A-DPT). The core idea is to start iterations from multiple initial filters and to detect the FCF adaptively during each iteration. Multiple features conforming to filtered signals’ SHIR maximization can be finally learned. Therefore, in this paper, MABD-SHIR is proposed, which is equivalent to a multi-node feature learning network. Since the objective function of BD is usually non-convex, MABD-SHIR can obtain multiple local optimal solutions without requiring prior FCFs. This is more likely to achieve a global optimal solution than BDSHIR, which starts iteration from only one initial filter and ends up with one local optimal solution. In addition, the objective function of MABD-SHIR has lower mathematical complexity, which reduces the difficulty of optimization.

I. Introduction

Impulsive signature extraction has been widely studied in the field of condition monitoring for rotating machinery. The presence of impulsive signature usually indicates the existing defects or potential faults of rotating machine. Currently, there have been many representative researches on recovering fault features excited by machine defects from collected vibration signals, mainly including classical decomposition [1], morphological filtering [2], health index-guided filtering [3], sparse representation [4], [5], [6], [7], spectral analysis [8], [9], [10], [11], improved decomposition [12], [13], [14], [15], [16], machine learning [17], [18], [19], [20], [21], [22] and blind deconvolution (BD) [23], [24], [25]. For example, Miao et al [26] designed a systematic diagnostic framework for bearing condition monitoring based on Gini index. Tao et al. separately designed two unsupervised bearing fault diagnosis networks for rolling bearings by combining STFT with generated neural networks [27], and based on time-frequency information fusion [28]. Hou et al. [29] designed a differential mode decomposition for adaptive bearing fault characteristic (FC) extraction. Among the above methods, due to the adaptive design of finite impulse response (FIR) filter that can maximize or minimize health indices of rotating machinery, BD performs very well in FC learning.