Real-time and energy-efficient bearing fault diagnosis using discriminative wavelet-based fault features on a multi-core system

Abstract

This paper presents a comprehensive bearing fault diagnosis methodology to prevent unscheduled interruptions in machinery. This method consists of fault signature extraction, discriminative fault feature selection (not only to improve diagnosis but also to reduce computational overhead), and decision making using a k-nearest neighbor classifier. Although the presented method yields very accurate classification, its computational complexity limits its use in real-time applications. To address that issue, this paper introduces a single instruction, multiple data-based multi-core system including 64 low-cost processing elements to support the high computational fault diagnosis method, where the multi-core system is implemented in a Xilinx Virtex-7 field programmable gate array device, operating at 200 MHz clock frequency. In addition, this study compares the performance of the multi-core system with that of high-performance digital signal processors (DSPs) to demonstrate improvement in execution time and energy efficiency. Experimental results using one-second acoustic emission signals sampled at 1 MHz indicate that the multi-core approach achieves 82.5x and 7.9x speedups over Texas Instruments (TI) TMS320C6748 and TMS320C6671, respectively, when executing the same bearing fault diagnosis algorithm. In addition, the presented multi-core approach reduces 21.5x and 4.2x less energy than TMS320C6748 and TMS320C6671, respectively. Keywords-Acoustic emission; discrete wavelet packet transform; discriminative fault feature selection; energy efficiency; multi-core system; online bearing fault diagnosis I. * Corresponding author 978-1-4799-1894-2/15/$31.00 ©2015 IEEE

Power Equipment Fault Diagnosis Method Based on Energy Spectrogram and Deep Learning

“…Time-frequency domain analysis methods mainly make use of the joint distribution of time domain information and frequency domain information to conduct correlation analysis of signals, among which wavelet transform [ 2 ], singular value decomposition [ 3 ], short-time Fourier transform [ 4 ], wavelet packet transform [ 5 ], ensemble empirical mode decomposition [ 6 ], and other time-frequency analysis methods are widely used in the field of fault detection. However, although the abovementioned methods claim certain achievements, there are still some common limitations: the first is the feature extraction conducted mainly through technical personnel artificial extraction, which relies on the expert’s experience and lack of generalization, especially when the power equipment is more complex or in operation mode, etc., making the traditional feature extraction method difficult to effectively extract the fault characteristic information; Secondly, most of the current feature extraction methods are equipped with shallow classification models [ 7 , 8 , 9 ], and the simple architecture of these models limits the nonlinear processing of fault feature information. Therefore, with the rapid development of industrial big data, it is necessary to study feature extraction methods with more adaptability and generalization and classification models with better classification performance.…”