Abstract: To address the challenges of data scarcity and difficulty in extracting weak features in the fault diagnosis of flexible elliptical bearings (FEB), an integrated performance testing platform incorporating mechanical, electrical, and data acquisition systems was independently designed and developed. The platform is used to reveal the failure mechanisms of FEBs and validate the effectiveness of diagnostic models. Analysis of the fault characteristic frequencies shows that, unlike traditional rolling bearings, FEBs exhibit a triangular function-like variation in frequency within a certain range. Furthermore, a fault diagnosis model combining the hippopotamus optimization (HO) Algorithm with deep learning is proposed for FEB fault analysis. Experimental results demonstrate that frequency-domain signals are more effective in capturing subtle fault features of FEBs, and the proposed model outperforms other methods. This work provides both theoretical support and experimental foundation for the application of deep learning in FEB fault diagnosis.