Deep learning for fault diagnosis of monoblock centrifugal pumps: a Hilbert–Huang transform approach

Abstract

Fault detection in monoblock centrifugal pumps plays an important role in ensuring the safe and efficient use of mechanical equipment. This study proposes a deep learning-based method using transfer learning for fault detection in monoblock centrifugal pumps. A MEMS sensor was used to acquire vibration signals from the experimental setup and these signals were subsequently processed and stored as Hilbert-Huang transform images. By leveraging 15 pretrained networks such as InceptionResNetV2, DenseNet-201, GoogLeNet, ResNet-50, VGG-19, Xception, VGG-16, EfficientNetb0, ShuffleNet, InceptionV3, ResNet101, MobileNet-v2, AlexNet, NasNetmobile and ResNet-18, fault diagnosis was performed on the acquired data. To achieve high classification accuracy, various hyperparameters including, batch size, learning rate, train-test split ratio and optimizer were systematically varied and optimized. The aim was to identify the most suitable configuration for the deep learning model. By leveraging transfer learning and preprocessing the acquired vibration signals into Hilbert–Huang transform images, the classification accuracy was significantly improved. Optimizing hyperparameters through extensive experimentation proved instrumental in elevating the models performance. Following thorough trials and meticulous tuning, the GoogleNet architecture emerged as the optimal setup, attaining a peak classification accuracy of 100.00%, all while upholding computational efficiency at 80 s.