This paper proposes an innovative bearing fault diagnosis method aimed at enhancing the accuracy and effectiveness of transfer learning. The innovation lies in the signal preprocessing stage, where a Noise Eliminated Intrinsic Time-Scale Decomposition (NEITD) algorithm is introduced. This algorithm adaptively decomposes unified-phase sine wave signals to effectively extract the geometric mean of the intrinsic rotational component, and selects the optimal decomposition result based on the orthogonality index, significantly improving the quality and reliability of the signals. In addition, fault diagnosis parameters are adaptively optimized using an improved adaptive deep transfer learning (ADTL) network combined with the Jellyfish Search (JS) algorithm, further enhancing diagnostic performance. By innovatively combining signal noise reduction, feature extraction, and deep learning optimization techniques, this method significantly improves fault diagnosis accuracy and robustness. Comparative simulations and experimental analyses show that the NEITD algorithm outperforms traditional methods in both signal decomposition performance and diagnostic accuracy. Furthermore, the NEITD-ADTL-JS method demonstrates stronger sensitivity and recognition capabilities across various fault types, achieving a 5.29% improvement in accuracy.