Résumé:
This work investigates the application of deep learning techniques, specifically 2D Convolutional Neural Networks (CNNs), for the diagnosis and detection of faults in electrical rotating machines, with a focus on bearing fault detection. The study leverages a comprehensive database of vibration signals acquired from a test rig at Case Western Reserve University, comprising various bearing fault conditions and operating conditions.
The vibration signals are preprocessed and converted into image representations suitable for input to the CNN models. Two different input image sizes, 32x32 and 64x64 pixels, are explored to evaluate the impact of spatial resolution on the model's performance. Additionally, the study compares the efficacy of two widely-used optimization algorithms, Stochastic Gradient Descent (SGD) and Adam, in training the CNN models.
The proposed 2D CNN architectures are designed to automatically learn discriminative features from the input vibration signal images, enabling accurate classification of different bearing fault types, including inner race faults, outer race faults, and ball faults, as well as normal bearing conditions. Data augmentation techniques are employed to mitigate overfitting and enhance the model's generalization capabilities.
Extensive experiments are conducted, and the results of using the CNN models, achieve classification accuracies of up to 99.58% for the 32x32 input size and 98.41% for the 64x64 input size when trained with the Adam optimizer. Detailed analysis of the confusion matrices and classification metrics provides insights into the strengths and weaknesses of each configuration.
This work highlights the potential of deep learning-based approaches, particularly 2D CNNs, for bearing fault diagnosis and remote detection of faults in electrical machines. The study contributes to the advancement of predictive maintenance strategies in industrial applications and paves the way for further research in this domain
