Vibration signal transformation-based fault diagnosis for wind turbine gearboxes considering deep feature optimization

To address diagnostic challenges in feature extraction and parameter tuning for wind turbine gearboxes, this study proposes a diagnostic model that integrates the Gramian angular field, a convolutional neural network, and a least-squares support vector machine optimized by the dream optimization algorithm. First, Gramian angular field encodes one-dimensional vibration time-series signals into two-dimensional images to preserve complete temporal dependencies and global dynamic features, transforming the fault diagnosis problem into a more advantageous image recognition task. Second, a convolutional neural network is constructed to automatically extract highly discriminative deep abstract features from Gramian angular field images, which are then fed into least-squares support vector machine for classification. This achieves end-to-end feature learning and pattern recognition, overcoming the subjectivity and limitations of traditional feature extraction. Third, dream optimization algorithm is introduced for adaptive global optimization of least-squares support vector machine hyperparameters. Its simulated dream mechanism effectively enhances parameter tuning efficiency and accuracy while avoiding local optima. Finally, MATLAB simulation experiments based on the Southeast University gearbox dataset demonstrate that the proposed model achieves 99% accuracy. Its recall, F1 score, and probability calibration performance significantly outperform comparison models, providing a high-precision, highly reliable solution for intelligent operation and maintenance of wind turbine gearboxes.