To address the issue of rolling bearing fault diagnosis, this paper proposes a novel model combining the Improved Gorilla Troop Optimization (IGTO) algorithm, Variational Mode Decomposition (VMD), Permutation Entropy (PE), and Long Short-Term Memory (LSTM) networks. The IGTO algorithm is used to optimize the parameters of VMD and LSTM, enhancing signal decomposition and feature extraction. The proposed model achieves fault classification accuracies of 96.67% and 98.96% in the testing and training phases, respectively, on the Case Western Reserve University dataset, with minimal accuracy fluctuations. Furthermore, on the Jiangnan University dataset, the model reaches an average testing accuracy of 98.85%, with the highest accuracy reaching 99.48%. The results also demonstrate high stability, as indicated by low standard deviations (1.2148 and 1.3217) and narrow 95% confidence intervals ([95.75%, 97.58%] and [96.73%, 97.49%]). Despite a longer average runtime of 13.88 s per sample, the model’s superior accuracy justifies the computational cost. These results demonstrate the model’s excellent diagnostic performance, adaptability to different datasets, and practical applicability for rolling bearing fault diagnosis. This approach provides a valuable reference for predictive maintenance and fault detection systems in industrial applications.

The accuracy of hydro-turbine fault diagnosis directly impacts the safety and operational efficiency of hydroelectric power generation systems. This paper addresses the challenge of low diagnostic accuracy in traditional methods under complex environments. This is achieved by proposing a signal preprocessing method that combines complete ensemble empirical mode decomposition with adaptive noise and multiscale permutation entropy (CEEMDAN-MPE) and that is optimized with the crested porcupine optimizer algorithm for the bidirectional long- and short-term memory network (CPO-BILSTM) model for hydro-turbine fault diagnosis. The method performs signal denoising using CEEMDAN, while MPE extracts key features. Furthermore, the hyperparameters of the CPO-optimized BILSTM model are innovatively introduced. The extracted signal features are fed into the CPO-BILSTM model for fault diagnosis. A total of 150 sets of acoustic vibrational signals are collected for validation using the hydro-turbine test bench under different operating conditions. The experimental results demonstrate that the diagnostic accuracy of the method is 96.67%, representing improvements of 23.34%, 16.67%, and 6.67% over traditional models such as LSTM (73.33%), CNN (80%), and BILSTM (90%), respectively. In order to verify the effectiveness of the signal preprocessing method, in this paper, the original signal, the signal processed by CEEMDAN, CEEMDAN-PE, and CEEMDAN-MPE are input into the CPO-BILSTM model for controlled experiments. The results demonstrate that CEEMDAN-MPE effectively denoises hydro-turbine acoustic vibrational signals while preserving key features. The method in this paper integrates signal preprocessing and deep learning models and, with the help of intelligent optimization algorithms, significantly enhances the model’s adaptive ability, improves the model’s applicability under complex operating conditions, and provides a valuable supplement for hydro-turbine fault diagnosis.