Identification of abnormal conditions in induction motors from current spectrum images using a two-stage approach with progressive learning
Background and objectives: This study presents a fault diagnosis system for induction machines based on a two-stage architecture using Convolutional Neural Networks (CNN). The aim is to improve fault identification by simplifying the classification process through sequential modeling. Methods: A dat...
| Autores: | , , , |
|---|---|
| Tipo de recurso: | artículo |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/449157 |
| Acceso en línea: | https://hdl.handle.net/2117/449157 https://dx.doi.org/10.1016/j.measurement.2025.119892 |
| Access Level: | acceso abierto |
| Palabra clave: | Convolutional Neural Network Deep learning Induction machine Fault diagnosis Visual geometry group Predictive maintenance Àrees temàtiques de la UPC::Enginyeria mecànica::Processos de fabricació mecànica::Màquines i mecanismes |
| Sumario: | Background and objectives: This study presents a fault diagnosis system for induction machines based on a two-stage architecture using Convolutional Neural Networks (CNN). The aim is to improve fault identification by simplifying the classification process through sequential modeling. Methods: A dataset of 5404 images was generated from the Fourier spectra of current signals acquired from a test bench under five conditions: healthy machine, rotor asymmetry fault, broken rotor bar, race bearing fault, and ball bearing fault. A CNN based on the Visual Geometry Group (VGG) architecture was trained from scratch and then adapted using transfer learning. The classification strategy follows two steps: first, distinguishing healthy from faulty machines; then, identifying the specific fault type. Results: The system reached 96.5% accuracy in the first stage and 98.5% in the second. All main performance metrics (sensitivity, specificity, precision, F1-score) remained above 95%. The behavior of the models was examined using Uniform Manifold Approximation and Projection (UMAP), which showed clearer separation between conditions in the latent space when using the sequential approach. In addition, Gradient-Weighted Class Activation Mapping (Grad-CAM) visualizations provided insights into the image regions influencing each decision, showing consistent focus on spectral areas related to each condition. Conclusions: The combination of image-based preprocessing, sequential classification, and model interpretation techniques leads to accurate predictions and helps to understand how the models behave. These features support its use in predictive maintenance tasks for industrial applications. |
|---|