Under complex working conditions, the vibration signals of rotating machinery usually exhibit significant non-stationarity and are accompanied by strong noise interference, which leads to greater challenges for traditional signal processing methods in extracting fault features. In response to the above problems, this paper proposes a fault diagnosis method for rotating machinery based on Bayesian optimization and signal reconstruction. Firstly, the Bayesian optimization algorithm is utilized to adaptively adjust the core parameters of successive variational mode decomposition. Secondly, a vertical distance discrimination method based on the average instantaneous frequency is proposed to distinguish the high and low frequency components of the inherent mode function obtained by successive variational mode decomposition. Then, a correlation coefficient weighting strategy is designed to conduct weighted reconstruction of different inherent mode functions. Finally, multi-dimensional features are extracted from the time domain, frequency domain and time-frequency domain to construct the feature set and conduct fault classification. Through experimental comparisons and analyses on multiple datasets and comparisons with existing methods, the effectiveness and superiority of the proposed method are verified.