
Smart Science, Год журнала: 2024, Номер unknown, С. 1 - 21
Опубликована: Июнь 11, 2024
Vibration-based fault diagnosis from rotary machinery requires prior feature extraction, selection, or dimensionality reduction. Feature extraction is tedious, and computationally expensive. selection presents unique challenges intrinsic to the method adopted. Nonlinear reduction may be achieved through kernel transformations, however there often a trade-off in information achieve this. Given above, this study proposes novel autoencoder (AE) pre-processing framework for vibration-based wind turbine (WT) gearboxes. In study, AEs are used learn features of WT gearbox vibration data while simultaneously compressing data, obviating need costly engineering The effectiveness proposed was evaluated by training genetically optimized linear discriminant analysis (LDA), multilayer perceptron (MLP), random forest (RF) models, with AE's latent space features. models were using known classification metrics. results showed that performance depends on size space. As increased, quality extracted improved until plateau observed at dimension 10. AE pre-processed RF, MLP, LDA designated AE-Pre-GO-RF, AE-Pre-GO-MLP, AE-Pre-GO-LDA, accuracy, sensitivity, specificity seven (7) conditions. AE-Pre-GO-RF model outperformed its counterparts, scoring 100% all metrics, though longest time (239.50 sec). Comparable found comparing similar investigations involving traditional processing techniques. More so, it established effective can manifold learning without expensive engineering.
Язык: Английский