In order to address the challenge of signal denoising and fault diagnosis for rolling bearings in complex environments, a fault diagnosis model based on an improved pelican optimization algorithm (IPOA) combined with variational mode decomposition (VMD) and neural networks was proposed. To enhance the algorithm efficiency, adaptive inertia weights and nonlinear convergence factors were first introduced to improve and optimize the pelican optimization algorithm. The improved algorithm was then used to optimize and determine the critical parameters K and α of VMD. Subsequently, the optimized parameters were applied for VMD decomposition of the original signal, and time-frequency domain features were extracted from the resulting K intrinsic mode functions (IMFs). Finally, a CNN-BiLSTM network was employed to classify the different states of rolling bearings. Experimental results on a rolling bearing dataset demonstrate that the feature data extracted using the IPOA-VMD method achieves higher fault diagnosis accuracy compared to feature data extracted using POA-VMD, PSO-VMD and VMD methods. On the test dataset, the proposed method achieves an average diagnostic accuracy of 99.867% with a large sample size.