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

Newswise — In recent years, intelligent fault diagnosis methods have been applied to the condition monitoring of rotating machinery. Among them, graph neural networks, as emerging feature extraction tools, can mine the relationship features between samples. However, many existing graph construction methods suffer from structural redundancy or missing node relationships, thus limiting the diagnostic accuracy of models in practical applications.

A study led by researchers from Wuhan University of Technology and The Hong Kong University of Science and Technology published a research paper in Frontiers of Mechanical Engineering, 2025, Volume 20, Issue 1. The paper proposes an adaptive adjustment k-nearest neighbor graph-driven dynamic-weighted graph attention network (AAKNN-DWGAT) for fault diagnosis of rotating machinery under noisy signals.

The proposed fault diagnosis method can be summarized into two parts: the construction of AAKNN graph (AAKNNG) and fault diagnosis based on DWGAT. In AAKNNG construction, node features are first embedded from raw time-domain signals. Then, a dynamic frequency warping (DFW)-based method is used for preliminary edge connection construction, followed by adaptive edge adjustment using the 3σ criterion and second-order difference method to determine the optimal number of edge connections for each node and calculate edge weights. For the DWGAT-based fault diagnosis, the constructed AAKNNG is fed into DWGAT, which contains two graph attention layers for capturing global and local fault features, respectively. A dynamic weighting strategy adjusts edge weights periodically based on high-level output features to reduce noise interference, and fault diagnosis results are output through a fully connected layer and softmax classifier. The pseudocode of the algorithm is also provided.

Experiments using the axial flow pump dataset and XJTUGearbox dataset were conducted. All algorithms were implemented based on Python 3.7 and PyTorch 2.0.1, with five different types of faults created and the same training parameters adopted. In the axial flow pump fault diagnosis experiment, the data collection process, sample set construction, and model hyperparameter settings were introduced in detail. The performance of AAKNNG and KNNG was compared, and the impacts of the number of attention heads and similarity calculation methods were studied. Comparisons with other graph construction methods and fault diagnosis methods were also performed, verifying the proposed method’s advantages in diagnostic accuracy and noise robustness. The XJTUGearbox dataset experiment similarly introduced data conditions and sample set construction, and compared diagnostic results of different methods. The results showed that AAKNN-DWGAT achieved the highest diagnostic accuracy with low variance under different signal lengths, further validating its effectiveness and superiority.

The proposed AAKNN-DWGAT method for rotating machinery fault diagnosis demonstrates excellent fault recognition and noise robustness compared with other traditional and SOTA deep learning methods. However, in the process of constructing AAKNNG, the adaptive adjustment of edge connections inevitably increases computation time while improving diagnostic accuracy. In some real-time fault diagnosis scenarios, such adjustment may introduce extra problems, which should be further optimized in future research.

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