PSO-optimised autoencoder for fault prediction in wind turbine planet carrier bearing

• A thresholding framework combining Particle Swarm Optimization, autoencoder, and discrete wavelet transform enhances planet carrier bearing fault diagnostics. • Particle Swarm Optimization reduces autoencoder's mean squared error, improving reconstruction accuracy. • Sequential threshold exploration effectively determines the optimal fault detection threshold for vibration data. • The proposed framework serves as a robust tool for condition monitoring of wind turbine bearings, enhancing operational reliability and profitability. This study introduced a novel thresholding framework based on a hybrid of Particle Swarm Optimization (PSO), autoencoder and discrete wavelet transform for planet carrier bearing (PLCB) fault diagnostics. Vibration signals from the PLCB are decomposed using discrete wavelet transform, with the resulting approximation coefficients serving as input to a PSO-optimized autoencoder model. The autoencoder model is first trained on the normal dataset to establish a baseline representing typical behaviour. The latter is evaluated on a validation set with reconstruction errors computed to identify a threshold for fault identification. This research determines the most effective threshold for the fault diagnostic model through an innovative sequential threshold exploration approach. The study results identified the autoencoder model's optimal hyperparameters as a latent space dimension of six (6) and a leaky ReLU activation function for the hidden layer. Following optimization, the model's mean squared error was reduced by 13.7 %, demonstrating a significant improvement in reconstruction capacity. Using the proposed thresholding framework, the optimal threshold was identified as 17.89. At this threshold, the model achieved exceptional diagnostic performance, with 98.4 % accuracy, a 98.4 % F1-score, and a 96.8 % Matthews correlation coefficient. These results highlight the model's viability as a robust tool for wind turbine condition monitoring, offering increased turbine uptime, reduced LCOE, and improved profitability of wind power investments.