Abstract
Accurate condition monitoring of industrial cyber-physical systems/components demands the use of reliable fault detection and isolation (FD&I) methodologies. Meta-heuristic algorithms for feature selection have good exploration capability for optimal discriminative feature selection for fault isolation/classification of which the Binary Particle swarm optimization (BPSO) is superior to its counterparts. This study presents a robust approach for vibration-based failure diagnostics of electromagnetic/solenoid pumps which employ a multi-domain feature extraction procedure (statistical time-domain and frequency-domain features, Mel frequency cepstral coefficients, and continuous wavelet coefficients) for capturing linear and nonlinear properties from the signals. Compared with other filter and wrapper methods for supervised feature selection, a hybrid filter-wrapper (Pearson’s correlation-BPSO ( $\rho $ -BPSO)) feature selection procedure is proposed for global search of optimal discriminative (uncorrelated) features for fault diagnosis with an RBF-kernel support vector machine (SVM*). Subsequently, a practical case study involving five VSC63A5 solenoid pumps at various operating/fault conditions is presented for validating the performance of the proposed approach. Results show the superior performance of the proposed hybrid filter-wrapper approach against filter-based and wrapper-based techniques for discriminative feature selection. Also, the proposed $\rho $ -BPSO-SVM* diagnostics model performance was compared with other standard fault isolation/classification methods.
Highlights
Pump failure/malfunction is one of the high-ranking concerns in the safety and productivity of hydraulic systems
The results showed the relatively better diagnostics capability of the random forest especially in the presence of additive Gaussian noise, a general conclusion cannot be drawn considering that (1) the statistical features used for comparison may be insufficient for a more in-depth feature extraction, (2) each classifier depend on parameter optimization for optimum performance, and (3) feature selection could have improved the classifiers’ performance for a more reliable comparative analysis
In each case, 70% of the selected features were randomly selected for training while the remaining 30% were used as testing samples and following a grid search on the on 9 support vector machine (SVM)* architectures consisting of C, γ ∈ {100, 101, 102} combinations, Fig. 8 shows the comparison in test accuracy between the architectures after a 10-fold cross-validation
Summary
Pump failure/malfunction is one of the high-ranking concerns in the safety and productivity of hydraulic systems. These methods are quite reliable, extensive studies suggest that by taking a comprehensive feature extraction approach across a wide range of domains/methods, a more reliable feature assessment can be achieved for a more accurate diagnostic result Inspired by their robustness in extracting linear and nonlinear characteristics from non-stationary signals [23], [24], the Mel Frequency Cepstral Coefficients (MFCCs) and their n-order differentials have strong capabilities for vibration-based diagnostics [15], [16], [25]. A. MULTI-DOMAIN FEATURE EXTRACTION The effectiveness of vibration monitoring has recently motivated hydraulic component manufacturers to integrate on-board sensors in their products; processing the non-stationary signals from these sensors for accurate condition monitoring has been one of the major issues facing vibration monitoring and other sensor-based real-time monitoring of ICPSs in general [51]. As a result of the wavelet basis function of the WT, wavelet decomposition of a non-stationary signal into linear forms of time-scale units is possible, thereby reconstructing the signal into several components according to the wavelet function translation [29], [30]; as fault features, wavelet coefficients are quite efficient [32], [33]
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
