Browsing by Author "Bastias Ramirez, Cristian Alejandro"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Diagnóstico de fallas en rodamientos mediante imágenes tiempo-frecuencia utilizando aprendizaje profundo.
    (Universidad de Concepción, 2023) Bastias Ramirez, Cristian Alejandro; Leaman Weiffenbach, Félix Alberto
    Bearings are crucial for the operation of rotating machines, so in this thesis is developed a methodology that uses Convolutional Neural Networks (CNN) to automatically classify time-frequency images, which are generated by the Short-Time Fourier Transform (STFT) and are associated with different bearing failure modes. This methodology is developed in order to deliver a tool that allows improving the reliability of these components, through the application of artificial intelligence. The application of this methodology contemplates the evaluation of the influence associated with the use of different configurations, such as the duration of the signal, normalization of colors and window width of the STFT. The signal duration directly affects the number of images available for training. Where a greater number of images improves the performance of CNN, while low availability negatively affects its performance. In respect to color normalization, three types of normalization are evaluated: local, global and by signal. Local normalization achieves the best results, since it effectively highlights representative characteristics associated with bearing failure modes. Subsequently, for the final configuration, a small window width presents the best results. Being one of the most important configurations in order to increase the robustness and generalization of the resulting model. Among other considerations, the correct selection of CNN hyperparameters contributes to mitigate the influence of random initialization of weights, and therefore contributes to reduce the variability of the CNN results. Finally, the best configurations correspond to a duration of 10 revolutions of the shaft, local normalization and window width of 4096 signal points, for which a model is generated that reaches 100% accuracy in the classification of these time-frequency images. Among other contributions of this memory, there is the evaluation of the feasibility of the use of TL between different datasets of bearings. Which is done in order to address the problems associated with the scarcity of available failure data, a common situation in the industry. These datasets consist of different operating conditions and failure modes in common. The TL is carried out using the knowledge previously acquired during the training of the CNN in the dataset of the University of Ottawa, considering the best configurations mentioned. Subsequently, this knowledge is transferred to a new dataset of the American Society for Machinery Fault Prevention Technology using TL techniques, such as Fine-Tuning. For which, as a result, a greater training speed and convergence of the model is obtained, with an average precision of 100% ± 0.00 and presenting itself as a solution to a low amount of available failure data. Which turns the use of TL as a feasible alternative for its application in the industry.
Síguenos en...