Vibration frequency measurement based on machine learning and stereo vision

Traditional vibration frequency measurement methods using stereo vision systems (SVS) often require explicit extraction of vibration signal time histories, rely on complex image processing algorithms, and depend on optical cues (e.g., markers or speckling) or techniques like edge and feature detection to track small movements on the target surface. These limitations increase implementation complexity and reduce adaptability to diverse scenarios. This paper introduces the SVS/ML method, a straightforward approach combining stereo vision techniques with machine learning (ML) for accurate and robust vibration frequency measurement. Unlike conventional methods, SVS/ML eliminates the need for explicit time history extraction and simplifies the tracking process. Experimental results comparing SVS/ML with reference methods employing industrial-grade sensors and known excitation sources demonstrate that the proposed method directly generates pixel-level vibration frequency maps with minimal error, achieving comparable accuracy to industrial-grade sensors. Moreover, SVS/ML exhibits strong robustness in both laboratory and field conditions, producing results that are ready-to-use without additional post-processing. These advantages make the method highly suitable for practical engineering applications, including structural health monitoring and machinery diagnostics.