INET researchers developed a novel multisensory fusion technology which provides new research ideas for health monitoring and fault diagnosis of large-scale equipment in scenarios where high-quality fault samples are scarce

Yan Xunshi | INET news

November 11, 2024

Research Associate Prof. Yan Xunshi and Prof. Shi Zhengang from the Institute of Nuclear and New Energy Technology (INET) of Tsinghua University have developed a novel multisensor fusion technology for health monitoring and fault diagnosis of large-scale equipment. This method establishes for the first time high-order corrections among multiple sensors in a complex system in the form of hypergraphs, exploring the relationships among multiple sensor signals in the sensor space. It breaks the constraints imposed by the independent and identically distributed (i.i.d.) assumption on intelligent diagnostic methods, addressing the challenge of insufficient relevant information in fault analysis under small sample conditions, and providing new research ideas for fault diagnosis in scenarios where high-quality fault samples are scarce. Recently, this research was published in the August 2024 issue of IEEE Transactions on Industrial Informatics, titled “Multisensor Fusion on Hypergraph for Fault Diagnosis”. This journal is a top-tier journal in the field of industrial informatics, classified in the top tier (Zone 1) by the Chinese Academy of Sciences, with an impact factor of 11.7.

Abstract of the paper

Multisensor information fusion techniques based on deep learning are crucial for machinery fault diagnosis. However, there are two major issues in previous research. First, the relationship between multisensor samples is disregarded, which is important to enhance the diagnostic performance. Second, the structure of the fusion algorithm becomes extremely complex with prolonged training when dealing with machinery equipped with a large number of sensors. To address the aforementioned two issues, our study proposes a new multisensor fusion mechanism that fuses multisensor information on hypergraphs, by building a single-sensor fusion hypergraph and a multisensor fusion hypergraph in the sensor space to embed the fault samples as nodes. In addition, a dual-branch hypergraph neural network is designed to compute the two hypergraphs to obtain the feature representation of the samples and diagnose faults. The algorithm is validated on two datasets for its performance.

Link to access the full paper

https://ieeexplore.ieee.org/document/10517296