TY - JOUR
T1 - Manifold learning-empowered lightweight unsupervised deep representation for edge-side fault detection
AU - Mo, Hangfeng
AU - Chu, Shilong
AU - Liu, Fengchun
AU - Liu, Jiameng
AU - Zhang, Ming
AU - Mao, Zhiwei
PY - 2024/12/1
Y1 - 2024/12/1
N2 - This paper addresses the problem of fault diagnosis in rotating systems by monitoring and analysis of their vibration signals. To tackle the prevalent challenges posed by the scarcity of high-quality labelled data and the complexity of models in the context of intelligent fault diagnosis for edge computing deployments, this research introduces a lightweight unsupervised deep representation method for edge-side fault detection. By leveraging manifold learning, this approach places a strong emphasis on unsupervised learning and model compactness. Firstly, the approach utilises the non-linear fitting capabilities of deep learning to implement manifold learning's dimensionality reduction, which can efficiently extract structured relationships from high-dimensional data, enabling rapid mapping of unlabelled samples to a lower-dimensional space. Then, the proposed method includes a clustering algorithm to refine the model's accuracy by correcting pseudo-labels near category centres, reducing intra-class distances and enhancing low-dimensional clustering. Finally, the model employs the isolation forest (iForest) algorithm for the rapid identification of both known and novel fault types at the edge, an innovative step in fault detection technology. Experiments on an open-source dataset and a real-world fault dataset collected by a customised test-rig indicate that the model proves effective in complex scenarios involving unlabelled data, sample imbalances and emerging fault categories. Its success underscores its significant potential for real-time online fault diagnosis in industrial environments, marking a substantial advancement in the field of intelligent fault identification.
AB - This paper addresses the problem of fault diagnosis in rotating systems by monitoring and analysis of their vibration signals. To tackle the prevalent challenges posed by the scarcity of high-quality labelled data and the complexity of models in the context of intelligent fault diagnosis for edge computing deployments, this research introduces a lightweight unsupervised deep representation method for edge-side fault detection. By leveraging manifold learning, this approach places a strong emphasis on unsupervised learning and model compactness. Firstly, the approach utilises the non-linear fitting capabilities of deep learning to implement manifold learning's dimensionality reduction, which can efficiently extract structured relationships from high-dimensional data, enabling rapid mapping of unlabelled samples to a lower-dimensional space. Then, the proposed method includes a clustering algorithm to refine the model's accuracy by correcting pseudo-labels near category centres, reducing intra-class distances and enhancing low-dimensional clustering. Finally, the model employs the isolation forest (iForest) algorithm for the rapid identification of both known and novel fault types at the edge, an innovative step in fault detection technology. Experiments on an open-source dataset and a real-world fault dataset collected by a customised test-rig indicate that the model proves effective in complex scenarios involving unlabelled data, sample imbalances and emerging fault categories. Its success underscores its significant potential for real-time online fault diagnosis in industrial environments, marking a substantial advancement in the field of intelligent fault identification.
KW - edge devices
KW - fault recognition
KW - lightweight models
KW - manifold learning
KW - unsupervised learning
UR - https://www.ingentaconnect.com/content/bindt/insight/2024/00000066/00000012/art00006?authentication=failed
UR - http://www.scopus.com/inward/record.url?scp=85212856976&partnerID=8YFLogxK
U2 - 10.1784/insi.2024.66.12.747
DO - 10.1784/insi.2024.66.12.747
M3 - Article
SN - 1354-2575
VL - 66
SP - 747
EP - 757
JO - Insight Journal
JF - Insight Journal
IS - 12
ER -