Ensemble-Enhanced Graph Autoencoder with GAT and Transformer-Based Encoders for Robust Fault Diagnosis

TL;DR

The paper addresses robust fault diagnosis from vibration signals under varying operating conditions by proposing a graph-based framework that converts time-series data into a segment graph via entropy-based windowing and DTW-based edges. A Graph Autoencoder with a deep graph transformer encoder/decoder learns latent graph representations, augmented by an ensemble classifier to predict fault types. Across CWRU and HUST datasets, the model achieves near-perfect cross-domain performance (mean F1 around $0.99$ on CWRU) and demonstrates that richer, more diverse training data (3 HP) yields the strongest generalization. The approach highlights the importance of dataset diversity and provides a scalable, robust fault-diagnosis solution for industrial settings.

Abstract

Fault classification in industrial machinery is vital for enhancing reliability and reducing downtime, yet it remains challenging due to the variability of vibration patterns across diverse operating conditions. This study introduces a novel graph-based framework for fault classification, converting time-series vibration data from machinery operating at varying horsepower levels into a graph representation. We utilize Shannon's entropy to determine the optimal window size for data segmentation, ensuring each segment captures significant temporal patterns, and employ Dynamic Time Warping (DTW) to define graph edges based on segment similarity. A Graph Auto Encoder (GAE) with a deep graph transformer encoder, decoder, and ensemble classifier is developed to learn latent graph representations and classify faults across various categories. The GAE's performance is evaluated on the Case Western Reserve University (CWRU) dataset, with cross-dataset generalization assessed on the HUST dataset. Results show that GAE achieves a mean F1-score of 0.99 on the CWRU dataset, significantly outperforming baseline models-CNN, LSTM, RNN, GRU, and Bi-LSTM (F1-scores: 0.94-0.97, p < 0.05, Wilcoxon signed-rank test for Bi-LSTM: p < 0.05) -- particularly in challenging classes (e.g., Class 8: 0.99 vs. 0.71 for Bi-LSTM). Visualization of dataset characteristics reveals that datasets with amplified vibration patterns and diverse fault dynamics enhance generalization. This framework provides a robust solution for fault diagnosis under varying conditions, offering insights into dataset impacts on model performance.

Figures (7)

• Figure 1: Proposed Model: Pipeline Architecture Overview.
• Figure 2: Architecture of a Graph Neural Network (GNN) process for fault classification in bearings. The input graph is represented by node features $X$ and an adjacency matrix $A$, where nodes correspond to sensor data points, and edges capture the relationships between them. The Graph Attention Network (GAT) layer encodes the graph into a latent space that captures the attention weights of key neighbors. By contrast, the TransformerConv layer captures broader patterns and dependencies within the graph. The encoded latent representation is then passed to an ensemble classifier that predicts the fault class for the bearing based on the learned features. This architecture uses the power of GNNs to model complex relationships in sensor data, enabling accurate and robust fault classification.
• Figure 3: (a) Entropy analysis for identifying the optimal window size for segmentation. (b) Heatmap of pairwise DTW distances between segments, where darker shades indicate greater similarity.
• Figure 4: Comparison of reconstruction loss across models trained on the 1Hp, 2Hp, and 3Hp datasets.
• Figure 5: Cross-dataset heatmaps of Precision, Recall, and F1-score for models trained on 1Hp, 2Hp, and 3Hp datasets. Each heatmap displays performance across ten fault classes (0-9) when tested on 1Hp, 2Hp, and 3Hp datasets.

Theorems & Definitions (2)

• Theorem 1
• proof