Sound-to-Vibration Transformation for Sensorless Motor Health Monitoring

Abstract

Automatic detection of motor failures, such as bearing faults, is essential for predictive maintenance across various industries, as bearing faults alone account for up to 51% of motor failures. While vibration-based diagnostics remain the de-facto standard, acquiring reliable vibration data is costly and sensitive to variations of the sensor model and quality, location, mounting and many other factors. To address this, we propose a novel sound-to-vibration transformation method that eliminates the need for onboard vibration sensors. Using any audio recorder (e.g., a mobile phone) and the proposed machine learning model, realistic vibration signals can be synthesized directly from the acquired sound under diverse operating conditions and fault scenarios. Experimental results on the Qatar University Dual-Machine Bearing Fault Benchmark (QU-DMBF) dataset show that the classification accuracy achieved with synthesized signals differs from real vibration data by less than 0.5%, demonstrating negligible loss in performance. This approach offers a low-cost, practical, and scalable alternative for accurate fault detection. The QU-DMBF benchmark dataset, results, and the optimized PyTorch implementation of the proposed sound-to-vibration transformer are publicly available for further research.