To address the challenges of difficult feature extraction for rolling bearing vibration signals, low efficiency in optimizing diagnostic model parameters, and the tendency to get trapped in local optima, this paper proposes an improved ICEEMDAN feature extraction method based on Bayesian optimization and adaptive noise signal ratio enhancement (BAYES-ICEEMDAN-SNR) and combines it with the improved Coriolis force optimization algorithm (ITOC) to optimize the least squares support vector machine (LSSVM) fault diagnosis model. Firstly, Bayesian optimization is used to adaptively determine the noise parameters and introduce a dynamic signal-to-noise ratio adjustment mechanism to enhance the robustness of feature extraction; secondly, Chebyshev chaotic mapping, Cauchy mutation, and dynamic reverse learning strategies are applied to enhance the global search and local escape capabilities of ITOC, thereby optimizing the hyperparameters of LSSVM; and finally, the Keesey-Chestnut University bearing dataset and Huazhong University of Science and Technology gear dataset are used for verification. The experimental results show that the average fault identification accuracy of the proposed method reaches over 97%, which is superior to that of the comparison models, and the effectiveness of each core improvement module of the proposed model is verified through ablation experiments, providing an effective solution for intelligent fault diagnosis of rolling bearings and gears.

Non-stationary, multi-component vibration signals in rotating machinery are easily contaminated by strong background noise, which masks weak fault features and degrades diagnostic reliability. This paper proposes a joint denoising method that combines an improved cordyceps fungus optimization algorithm (ICFO), successive variational mode decomposition (SVMD), and an improved wavelet thresholding scheme. ICFO, enhanced by Chebyshev chaotic initialization, a longitudinal–transverse crossover fusion mutation operator, and a thinking innovation strategy, is used to adaptively optimize the SVMD penalty factor and number of modes. The optimized SVMD decomposes the noisy signal into intrinsic mode functions, which are classified into effective and noise-dominated components via the Pearson correlation coefficient. An improved wavelet threshold function, whose threshold is modulated by the sub-band signal-to-noise ratio, is then applied to the effective components, and the denoised signal is reconstructed. Simulation experiments on nonlinear, non-stationary signals with different noise levels (SNR = 1–20 dB) show that the proposed method consistently achieves the highest SNR and lowest RMSE compared to VMD, SVMD, VMD–WTD, CFO–SVMD, and WTD. Tests on CWRU bearing data and gearbox vibration signals with added −2 dB Gaussian white noise further confirm that the method yields the lowest residual variance ratio and highest signal energy ratio while preserving key fault characteristic frequencies.