Deep Denoising Convolutional Autoencoder-Based Fault Diagnosis of A Centrifugal Pump

This study addresses the challenge of assessing the health condition of centrifugal pumps in industrial settings without the need for traditional vibrational sensors, which can reduce system reliability. Instead, the focus is on utilizing electrical signals for fault detection and diagnosis (FDD). The proposed approach involves a deep denoising convolutional autoencoder-based classifier that is trained to be robust to noise. A autoencoder is initially trained to extract noise-robust features, and these features are then transferred to a deep classification neural network for fine-tuning using a noisy dataset. The method is tested on an experimental dataset, demonstrating its robustness to various noise levels and its ability to accurately identify both mechanical and operational faults in industrial scenarios with diverse noise amplitudes.