A novel fault diagnosis scheme was developed to address the difficulty of feature extraction for planetary gearboxes using refined composite multiscale fluctuation dispersion entropy (RCMFDE) and supervised manifold mapping. The RCMFDE was first utilized in this scheme to fully mine fault features from planetary gearbox signals under multiple scales. Subsequently, as a supervised manifold mapping method, supervised isometric mapping (S-Iso) was applied to decrease the dimensions of the original features and remove redundant information. Lastly, the marine predator algorithm-based support vector machine (MPA-SVM) classifier was employed to achieve intelligent fault diagnosis of planetary gearboxes. The suggested RCMFDE combines the composite coarse-grained construction and refined computing technology, overcoming unstable and invalid entropy in the traditional multiscale fluctuation dispersion entropy. Simulation experiments and fault diagnosis experiments from a real planetary gearbox drive system show that the complexity measure capability and feature extraction effectiveness of the proposed RCMFDE outperform the multiscale fluctuation dispersion entropy (MFDE) and multi-scale permutation entropy (MPE). The S-Iso’s visualization results and dimensionality reduction performance are better than principal components analysis (PCA), linear discriminant analysis (LDA), and isometric mapping (Isomap). Moreover, the suggested fault diagnosis scheme has an accuracy rate of 100% in identifying bearing and gear defects in planetary gearboxes.

Fault diagnosis of rotating machinery is an attractive yet challenging task. This paper presents a novel intelligent fault diagnosis scheme for rotating machinery based on ensemble dilated convolutional neural networks. The novel fault diagnosis framework employs a model training strategy based on early stopping optimization to ensemble several one-dimensional dilated convolutional neural networks (1D-DCNNs). By varying the dilation rate of the 1D-DCNN, different receptive fields can be obtained to extract different vibration signal features. The early stopping strategy is used as a model update threshold to prevent overfitting and save computational resources. Ensemble learning uses a weighted mechanism to combine the outputs of multiple 1D-DCNN subclassifiers with different dilation rates to obtain the final fault diagnosis. The proposed method outperforms existing state-of-the-art classical machine learning and deep learning methods in simulation studies and diagnostic experiments, demonstrating that it can thoroughly mine fault features in vibration signals. The classification results further show that the EDCNN model can effectively and accurately identify multiple faults and outperform existing fault detection techniques.