The transition of industrial maintenance toward intelligent and autonomous operation requires non-destructive evaluation frameworks that remain reliable under variable operating conditions and limited labeled data. In practical inspection scenarios, data driven diagnostic models often suffer from performance degradation due to changes in operating speed and insufficient fault samples, which restrict their applicability in real machinery monitoring. To address this challenge, this study presents a multistage transfer learning framework designed to enhance vibration-based fault evaluation of rotating machinery under varying rotational speeds. The proposed framework progressively adapts a pretrained convolutional network through generic feature learning, vibration specific representation refinement, and task focused fine tuning, enabling stable feature reuse while preserving inspection relevant signal characteristics. A layer selective adaptation strategy is employed to retain general low-level representations while refining higher level features that capture speed dependent fault signatures. Experimental investigations conducted on vibration signals acquired from a milling machine operating at five spindle speeds demonstrate that the proposed approach achieves consistent and reliable fault discrimination across operating conditions, with classification accuracy ranging from 97.70% to 100%. The results indicate improved separation of fault induced time frequency patterns compared with conventional single stage transfer strategies. By enhancing robustness to operating variability and reducing dependence on extensive labeled datasets, the proposed framework contributes to reliable nondestructive evaluation and supports practical deployment of vibration-based condition monitoring systems in industrial environments.
Siddique et al. (Sat,) studied this question.