Polarized Direct Cross-Attention Message Passing in GNNs for Machinery Fault Diagnosis

Zongyu Shi; Laibin Zhang; Maoyin Chen

cs.LGMar 6, 2026

Polarized Direct Cross-Attention Message Passing in GNNs for Machinery Fault Diagnosis

Authors: Zongyu Shi, Laibin Zhang, Maoyin Chen

Organizations: Department of Automation, College of Artificial Intelligence, China University of Petroleum (Beijing), Beijing 102299, China · Department of Safety Engineering, College of Safety and Ocean Engineering, China University of Petroleum (Beijing), Beijing 102299, China

Abstract

The reliability of safety-critical industrial systems hinges on accurate and robust fault diagnosis in rotating machinery. Conventional graph neural networks (GNNs) for machinery fault diagnosis face limitations in modeling complex dynamic interactions due to their reliance on predefined static graph structures and homogeneous aggregation schemes. To overcome these challenges, this paper introduces polarized direct cross-attention (PolaDCA), a novel relational learning framework that enables adaptive message passing through data-driven graph construction. Our approach builds upon a direct cross-attention (DCA) mechanism that dynamically infers attention weights from three semantically distinct node features (such as individual characteristics, neighborhood consensus, and neighborhood diversity) without requiring fixed adjacency matrices. Theoretical analysis establishes PolaDCA's superior noise robustness over conventional GNNs. Extensive experiments on industrial datasets (i.e., XJTUSuprgear, CWRUBearing and Three-Phase Flow Facility datasets) demonstrate state-of-the-art diagnostic accuracy and enhanced generalization under varying noise conditions, outperforming seven competitive baseline methods. The proposed framework provides an effective solution for safety-critical industrial applications.

Figures & tables

Fig. 2: Framework of DCA-based message passing

Fig. 3: Framework of PolaDCA-based message passing

Method	Average (%)		Best (%)
Method	ACC	F1	ACC	F1
GCN	$94.97\pm 3.37$	$93.69\pm 2.97$	96.67	96.67
GAT	$78.97\pm 6.29$	$77.70\pm 7.07$	89.17	89.06
GraphSAGE	$87.57\pm 1.38$	$87.46\pm 1.48$	88.83	88.74
GTF(SCA)	$97.99\pm 1.00$	$98.00\pm 0.99$	99.00	99.00
GCL	$81.80\pm 2.95$	$81.48\pm 3.11$	85.00	84.82

TABLE I: Experimental Result on XJTUSuprgear Normal Data

Method	0 dB		-2 dB		-4 dB		-6 dB		-8 dB
Method	ACC(%)	F1(%)	ACC(%)	F1(%)	ACC(%)	F1(%)	ACC(%)	F1(%)	ACC(%)	F1(%)
GCN	75.17	74.57	71 83	71.15	71.77	71.06	70.50	69.92	62.17	61.13
GAT	62.17	55.94	60.17	54.39	61.33	57.89	60.67	57.67	46.67	37.37
GraphSAGE	69.33	69.51	67.83	67.70	61.89	61.85	59.17	58.06	60.33	60.70
GTF(SCA)	84.22	84.11	83.83	83.71	78.67	78.45	79.17	78.65	77.06	76.58
GCL	49.56	46.71	47.33	44.44	43.56	39.95	45.00	40.82	43.67	39.53

TABLE II: Average Experimental Result on XJTUSuprgear with Gaussian Noise

Method	Average (%)		Best (%)
Method	ACC	F1	ACC	F1
GNN	$95.46\pm 2.18$	$94.27\pm 2.06$	96.91	96.48
GAT	$83.33\pm 3.67$	$82.34\pm 3.07$	86.82	86.31
GraphSAGE	$97.17\pm 0.97$	$97.16\pm 0.87$	98.12	98.12
GTF(SCA)	$96.00\pm 0.57$	$96.00\pm 0.41$	96.85	96.76
GCL	$81.17\pm 1.19$	$79.07\pm 0.96$	82.08	80.15

TABLE III: Experimental Result on CWRUBearing Normal Data

Fig. 4: Accuracy comparison on CWRUBering data with gaussian noise

Method	Average (%)		Best (%)
Method	ACC	F1	ACC	F1
GCN	$67.67\pm 4.37$	$65.85\pm 3.97$	72.08	71.65
GAT	$77.97\pm 2.29$	$76.70\pm 2.07$	81.09	80.84
IAGNN	$91.19\pm 1.65$	$90.67\pm 1.58$	93.45	92.17
MA-STGNN	$93.80\pm 2.86$	$92.78\pm 2.57$	96.85	95.66
CDGNN	$97.58\pm 0.83$	$96.97\pm 0.71$	98.14	97.87

TABLE IV: Experimental Result on TFF Noraml Data

Fig. 5: Accuracy comparison on TFF data with gaussian noise

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Fig. H.1: XJTUSuprgear dataset detailed sensors position

No.	Fault type	Training data	Testing data
1	15Hz_Normal	140	60
2	15Hz_Crack02	140	60
3	15Hz_Crack06	140	60
4	15Hz_Crack10	140	60
5	15Hz_Crack14	140	60
6	20Hz_Normal	140	60

Appendix

TABLE H.1: Fault Types in the XJTUSuprGear Dataset

No.	Fault type	Training data	Testing data
1	Normal	140	60
2	Rolling Element Fault (0.007 in)	140	60
3	Rolling Element Fault (0.014 in)	140	60
4	Rolling Element Fault (0.021 in)	140	60
5	Inner Race Fault (0.007 in)	140	60
6	Inner Race Fault (0.014 in)	140	60

Appendix

TABLE H.2: Fault Types in the CWRUBearing Dataset

Variables	Measured magnitude	Unit
1	Air delivery pressure	MPa
2	Pressure in the bottom of the riser	MPa
3	Pressure in top of the riser	MPa
4	Pressure in top separator	MPa
5	Pressure in 3 phase separator	MPa
6	Diff. pressure (PT401-PT408)	MPa

Appendix

TABLE H.3: Variables in Three-Phase Flow Facility Dataset

No.	Fault type	Training data	Testing data
1	Air line blockage	50	21
2	Water line blockage	60	26
3	Top separator input blockage	100	43
4	Open direct bypass	73	32
5	Slugging conditions	35	15
6	Pressurization of the 2” line	52	22

Appendix

TABLE H.4: Fault Types in Three-Phase Flow Facility Dataset

Fig. I.1: Confusion matrix of XJTUSuprgear normal data. (a) GCN, (b) GAT, (c) GraphSAGE, (d) GTF, (e) GCL, (f) MRF-GCN. (g) IAGNN, (h) DCA-GNN, (i) PolaDCA-GNN

Fig. I.2: Evolution of adaptive attention weights of XJTUSuprgear dataset

Fig. I.3: t-SNE visualization comparison of XJTUSuprgear dataset

Fig. I.4: Visual impact of different conditions on PolaDCA weights

Fig. I.5: Convergence on PolaDCA-GNN on TTF data

Fig. I.6: Confusion matrix of TTF data fault diagnosis

Explore similar work

Multi-Level Temporal Graph Networks with Local-Global Fusion for Industrial Fault Diagnosis
Apr 20, 2026Bibek Aryal, Gift Modekwe, Qiugang LuTemporal GNNs Multi-Scale Feature Fusion
Department of Chemical Engineering, Texas Tech University, Lubbock, TX 79409, USA
Parametric Generalized Adaptive Moment Features (PG-AMF) for Bearing Fault Diagnosis and Machine Health Monitoring
Jun 24, 2026Rajeev KumarRepresentation Learning Predictive Maintenance
Department of Mechanical Engineering University of Alberta Edmonton, Canada
An Explainable GNN Framework for Component-Level Anomaly Diagnosis
Aug 10, 2026Sena Ozgunay, Louise Trav{é}-Massuy{è}s, Jean-Michel Loubes +1Industrial Anomaly Detection Root Cause Analysis
Comue de Toulouse, LAAS-DISCO, IMT · Comue de Toulouse, LAAS-DISCO · REGALIA, IMT +1

Apr 20, 2026J/K move · Enter open · S save

Multi-Level Temporal Graph Networks with Local-Global Fusion for Industrial Fault Diagnosis

Bibek Aryal, Gift Modekwe, Qiugang Lu

Department of Chemical Engineering, Texas Tech University, Lubbock, TX 79409, USA

Temporal GNNs Multi-Scale Feature Fusion Anomaly Detection

Open paper AskPDF arXiv

Fault detection and diagnosis are critical for the optimal and safe operation of industrial processes. The correlations among sensors often display non-Euclidean structures where graph neural networks (GNNs) are widely used therein. However, for large-scale systems, local, global, and dynamic relations extensively exist among sensors, and traditional GNNs often overlook such complex and multi-level structures for various problems including the fault diagnosis. To address this issue, we propose a structure-aware multi-level temporal graph network with local-global feature fusion for industrial fault diagnosis. First, a correlation graph is dynamically constructed using Pearson correlation coefficients to capture relationships among process variables. Then, temporal features are extracted through long short-term memory (LSTM)-based encoder, whereas the spatial dependencies among sensors are learned by graph convolution layers. A multi-level pooling mechanism is used to gradually coarsen and learn meaningful graph structures, to capture higher-level patterns while keeping important fault related details. Finally, a fusion step is applied to combine both detailed local features and overall global patterns before the final prediction. Experimental evaluations on the Tennessee Eastman process (TEP) demonstrate that the proposed model achieves superior fault diagnosis performance, particularly for complex fault scenarios, outperforming various baseline methods.

Abstract

Explore similar work

Multi-Level Temporal Graph Networks with Local-Global Fusion for Industrial Fault Diagnosis

Parametric Generalized Adaptive Moment Features (PG-AMF) for Bearing Fault Diagnosis and Machine Health Monitoring

An Explainable GNN Framework for Component-Level Anomaly Diagnosis