An Intelligent Approach to Detecting Novel Fault Classes for Centrifugal Pumps Based on Deep CNNs and Unsupervised Methods

TL;DR

This paper presents a two-stage deep learning and unsupervised clustering approach for detecting and updating models with novel faults in centrifugal pumps, enhancing fault diagnosis capabilities.

Contribution

It introduces a novel combination of CNNs with t-SNE and clustering to identify previously unknown faults and adapt the model accordingly.

Findings

High accuracy in detecting novel faults

Effective model augmentation with new data

Validated on centrifugal pump setup

Abstract

Despite the recent success in data-driven fault diagnosis of rotating machines, there are still remaining challenges in this field. Among the issues to be addressed, is the lack of information about variety of faults the system may encounter in the field. In this paper, we assume a partial knowledge of the system faults and use the corresponding data to train a convolutional neural network. A combination of t-SNE method and clustering techniques is then employed to detect novel faults. Upon detection, the network is augmented using the new data. Finally, a test setup is used to validate this two-stage methodology on a centrifugal pump and experimental results show high accuracy in detecting novel faults.