A novel residual graph representation learning method towards multi-source data fusion and fault diagnosis of machinery
Zhuojun Dai, Weidong Xu, Zhuyun Chen, Kairu Wen, Bin Zhang, Yi He, Weihua Li · Measurement Science and Technology · 2025
Abstract The harmonic reducer, a critical component in industrial mechanical systems, is responsible for high-precision motion transmission. Timely fault detection and diagnosis are essential to prevent catastrophic safety incidents and ensure system reliability. However, existing fault diagnosis methods often fail to fully utilize the rich information embedded in multi-source data, leading to suboptimal performance. Additionally, the interactions and dependencies between different data channels remain inadequately captured, hindering accurate fault identification. To address these limitations, this paper introduces a novel fault diagnosis framework based on residual graph representation learning and multi-source data fusion. The proposed model monitors the harmonic reducer’s operational state using a multi-sensor network. A signal preprocessing module, leveraging fast fourier transform and RadiusGraph, transforms raw multi-sensor data into a graph structure, where nodes represent sensor channels and weighted edges capture interdependencies. This graph structure effectively reflects the interaction and dependency among multi-channel data. For feature extraction, we propose a bi-layer ChebyNet with residual connections to mine and update the relationships between sensor data while addressing challenges such as gradient vanishing or explosion in deep graph neural networks. Finally, the learned graph representations are used to classify fault types, enabling precise identification of potential faults or anomalies in the harmonic reducer. To validate the proposed method, we conducted fault experiments under various laboratory conditions, simulating six types of faults commonly encountered in real-world scenarios, including gear tooth breakage, bearing faults, and flexspline issues. Experimental results demonstrate that our model achieves superior fault diagnosis performance compared to existing methods, as evidenced by higher classification accuracy, F1 scores, and improved confusion matrix metrics. This work highlights the effectiveness of integrating multi-source data fusion with advanced graph-based learning techniques for enhancing fault diagnosis in harmonic reducers, offering a robust solution for industrial applications.