A vibro-acoustic signals hybrid fusion model for blade crack detection

Blade crack detection is the key to ensuring the smooth and safe operation of centrifugal fans. However, a single vibration signal is difficult to fully reflect the health state of the blade and is susceptible to noise interference in the industrial field, which makes it difficult to detect blade cracks. Therefore, a vibro-acoustic signals hybrid fusion model (VAS-HFM) is proposed for blade crack detection of centrifugal fans. Firstly, the kernel cosine similarity is designed to exploit the nonlinear relationship between multi-source vibro-acoustic signals and complete the data-level fusion of homogeneous vibro-acoustic signals. Then, a one-dimensional feature pyramid network (1D-FPN) is proposed to extract the vibro-acoustic features and generate initial decisions. Finally, a basic probability assignment (BPA) acquisition method based on the precision of the initial classifiers and a weighted average method based on the Jousselme evidence distance (JED) are proposed to minimize the conflict between multi-source evidence, and the Dempster-Shafer theory of Evidence (DSE) method is used to obtain the blade crack detection results. The effectiveness of the proposed method is verified by two centrifugal fan blade crack detection experiments. Compared with other related methods, the proposed method has better detection performance and stronger robustness.