Looseness diagnosis method for connecting bolt of fan foundation based on sensitive mixed-domain features of excitation-response and manifold learning

Abstract

Looseness diagnosis for connecting bolt of fan foundation is an important task for ensuring proper operation of fan and safe traffic. Aiming at solving the key problems of bolt looseness diagnosis including looseness feature extraction, looseness feature set construction, non-sensitive or poor sensitive feature interference and feature set nonlinear dimension reduction, a looseness diagnosis method for connecting bolt of fan foundation based on sensitive mixed-domain features of excitation response and manifold learning is proposed. Firstly, the response signal is collected by applying a pulse excitation signal to the fan, and the frequency response function is calculated. The looseness of fan foundation is characterized by response signal and the frequency response function. Then, the looseness mixed-domain feature set is constructed through fusion time-domain feature and frequency-domain feature of response signal and frequency response function. Secondly, the looseness sensitivity index is calculated based on scatter matrix for sensitive feature selection to avoid the interference of non-sensitive and poor sensitive feature, thus the looseness sensitive feature set is constructed. Moreover, orthogonal neighborhood preserving embedding (ONPE), an effective manifold learning algorithm with non-linear dimensionality reduction capability, is applied to compress the high-dimensional looseness sensitive feature set into the low-dimensional one. Finally, the low-dimensional looseness sensitive feature set is imported into weight K nearest neighbor classifier (WKNNC) as input to recognize different loosening of connecting bolt, and the stability of recognition accuracy rate is ensured. The feasibility and performance of the proposed method were proved by successful looseness diagnosis application on a tunnel fan foundation's connecting bolt.