A fuzzy-neuro approach to predictive maintenance using vibration signature classification

Vibration-based predictive maintenance must remain reliable under variable speed/load and noise while delivering actionable, interpretable decisions. We propose a fuzzy-neuro framework that maps windowed vibration segments to class decisions and a calibrated risk score with uncertainty. Given a sampled signal , overlapping windows are formed and encoded into a learned signature . Physics-informed indicators (e.g., RMS, kurtosis, band-energy and short-horizon trend) are fused with to form . An end-to-end neuro-adaptive Takagi-Sugeno-Kang (TSK) layer produces a transparent maintenance risk enabling rule-level explanations via dominant firing strengths . To support safe decision-making near thresholds, we estimate an interval risk from predictive samples using quantiles , and width as confidence. Using leakage-safe, unit-aware temporal splitting, experiments on public rolling-bearing benchmarks achieve 0.956 accuracy, 0.952 macro-F1, and 0.941 MCC, while the full model improves calibration and yields sharper risk intervals (mean width 0.132), translating classifier evidence into auditable "monitor/schedule/urgent" actions. These results indicate that the proposed framework is accurate, interpretable, and decision-ready for predictive maintenance.