Sensor-Based Air-Gap Monitoring of Elevator Brakes via RMS-Envelope-Guided Transient Impact Extraction and RBF-SVM

Air-gap variation in elevator brakes affects armature actuation, brake-shoe release/contact processes, and transient vibration responses, making it an important indicator for brake condition monitoring. However, long-duration vibration recordings acquired from brake-mounted sensors contain mixed operating stages, which makes it difficult to isolate short action-related impacts. This study develops an engineering-oriented sensor-based framework for elevator brake air-gap monitoring by combining RMS-envelope-guided transient impact extraction with an RBF-SVM classifier. A triaxial accelerometer was mounted on the right-side brake armature, and a fixed y-axis vibration channel was used as the baseline input for feature construction. The short-time RMS envelope was first used to identify operating-state transition boundaries. Brake-release and brake-engagement impact peaks were then localized within the neighborhoods of these boundaries, and fixed-length transient impact samples were extracted. Time-domain, frequency-domain statistical, and band-energy features were constructed to characterize impact intensity, spectral structure, and energy redistribution under different air-gap conditions. Experiments were conducted on an elevator traction-machine brake under six controlled air-gap states from 0.30 mm to 0.80 mm. The results show that brake-release impact features are more sensitive to air-gap variation than brake-engagement or combined impact features. Using brake-release features, the proposed RMS-IE-SVM method achieved an accuracy of 89.77% and a Macro F1 of 89.85% under the last-file split setting, and a mean accuracy of 83.62% and a mean Macro F1 of 79.07% under file-grouped cross-validation. In the common-sample multi-axis comparison, X + Y + Z feature-level fusion achieved a file-grouped cross-validation accuracy of 89.63% and a Macro F1 of 85.94%. Feature-group ablation shows that time-domain features provide the dominant information, while band-energy features offer complementary information. These results indicate that RMS-envelope-guided transient impact extraction provides an interpretable framework for controlled-condition elevator brake air-gap identification and that multi-axis vibration information can further improve cross-file generalization.