Indirect Condition Monitoring of the Transmission Belts in a Desalination Plant by Using Deep Learning

Condition monitoring is a basic technique in contemporary maintenance, since it can be used to identify problems in equipment and machinery before catastrophic failures occur. In the present work, an indirect monitoring system of the state of deterioration of the transmission belt of a water desalination plant is proposed. To achieve this goal, the mechanical vibrations in the three axes, measured at the bearing of the drive pulley, are taken as input signals. They are preprocessed by applying a fast Fourier transform and combining the respective outcomes into an image, where each basic channel corresponds to an axis. These images are used as the inputs of a two-block convolutional neural network, which is trained by using the Adam algorithm. The trained convolutional network allows the belts to be classified into three categories: new, medium used, and worn out. The proposed system was more than 90% effective for both the training and validation sets.