Angle Encoding for Oriented Detection of Torsional Vibration of Rotating Parts

Visual measurement is widely used in the field of engineering machinery as a new type of sensor, but common visual methods are not applicable to the angular speed measurement of rotating parts. The strategy of rotating target detection is suitable for the vibration signal measurement of periodic rotating machinery, but the rotation angle, as a discontinuous function, produces great ambiguity in the network learning process. The prediction box of the rotating target predicted by the algorithm has low fit and rough position regression. This article proposes angle phase shift encoding (APSE) to convert the target rotation angle into a continuous function, and the angle value can be learned in a deep network. For the rough position regression of the prediction box, multiscale parallel prediction is proposed to improve the prediction accuracy of the bounding box. To constrain the offset of the prediction box and improve the center point positioning, the boundary is specified by the centerness loss term. In the experimental part, compared with visual algorithms, the proposed methods effectively measure periodic impact signals. In addition, the robustness and generalization of the proposed method are demonstrated in experiments on multiple working conditions and complex scenarios.