Adaptive Adjustment Graph Representation Learning Method for Rotating Machinery Fault Diagnosis Under Noisy Signals

Newswise — In the field of rotating machinery fault diagnosis, intelligent fault diagnosis methods based on graph neural networks (GNNs) have become a research focus due to their ability to mine relational features between samples. However, existing graph construction methods, such as fixed k-nearest neighbor graphs (KNNGs), face significant limitations: fixed k values lead to structural redundancy or missing node relationships, introducing uncertain neighbor information and increasing training time. Particularly under noisy signal conditions common in industrial environments, redundant edges may bring more noise interference, reducing the model’s diagnostic accuracy. Additionally, traditional similarity metrics like Euclidean distance fail to fully capture frequency-domain fault feature relationships, and conventional graph attention networks (GATs) cannot dynamically adjust edge weights based on high-level features, further limiting the reliability of fault diagnosis results.

Therefore, a research team consisting of researchers from the School of Mechanical and Electronic Engineering at Wuhan University of Technology and the Department of Mechanical and Aerospace Engineering at The Hong Kong University of Science and Technology has conducted a study entitled “Adaptive Adjustment Graph Representation Learning Method for Rotating Machinery Fault Diagnosis Under Noisy Signals”.

This study proposes an adaptive adjustment k-nearest neighbor graph-driven dynamic-weighted graph attention network (AAKNN-DWGAT) to address the aforementioned challenges. The method consists of two core parts: adaptive adjustment k-nearest neighbor graph (AAKNNG) construction and dynamic-weighted graph attention network (DWGAT)-based fault diagnosis.

Extensive validations were conducted using two datasets: the axial flow pump dataset and the XJTUGearbox dataset. For the axial flow pump dataset (including normal state and five fault types), compared with fixed KNNGs, AAKNNG reduced structural redundancy, achieving an average diagnostic accuracy of 94.26%–95.42% (vs. lower accuracy for KNNGs) with negligible additional graph construction time. Under different signal-to-noise ratios (SNRs), AAKNNG outperformed other graph construction methods (KNNG, dynamic-weighted graph, affinity graph, SuperGraph) at low SNRs (e.g., -10 dB to 0 dB), verifying its noise robustness. Compared with state-of-the-art methods like AG+MRF-GCN and KNNG+MHGAT, AAKNNG-DWGAT achieved over 95.0% test accuracy with lower variance. For the XJTUGearbox dataset (including normal state and eight fault types), with sample lengths of 512, 1024, and 2048, AAKNNG-DWGAT consistently achieved the highest diagnostic accuracy (average 85%, 94%, and 99% respectively), outperforming classical methods like 1DCNN, MLP, and single-head GAT.

The paper “Adaptive adjustment graph representation learning method for rotating machinery fault diagnosis under noisy signals” is authored by Lei WANG, Peijie YOU, Xin ZHANG, Li JIANG, and Yibing LI. Full text of the open access paper: