A Novel Residual Graph Representation Learning Method Towards Multi-Source Data Fusion and Fault Diagnosis of Machinery

Abstract As an important part of an industrial mechanical system, harmonic reducer undertakes the task of high-precision motion transmission. Once a fault occurs and is not handled in time, it may cause huge safety accidents, making comprehensive monitoring and accurate fault diagnosis of the harmonic reducer an important problem. However, the current methods fail to fully exploit the rich information embedded in various data sources, resulting in suboptimal fault diagnosis performance. Moreover, the interactions and dependencies between different data channels are not adequately captured, which hampers the accurate identification of faults. To address these challenges, this paper proposes a multi-source data fusion and fault diagnosis model based on residual graph representation learning. The running state of the harmonic reducer is monitored through a sensor network composed of multi-sensor. To reflect the interaction and dependence between multi-channel data, a signal preprocessing module based on FFT and RadiusGraph is proposed. This module transforms the multi-sensor data into a multi-sensor graph network composed of nodes and weighted edges. The feature representation of the graph is then learned using the bilayer ChebyNet with residual connection (BiCNR) to mine and update the interactions and dependencies between multiple sensors. Finally, based on the learned graph representation, the fault type of the harmonic reducer is diagnosed by the proposed model, and potential faults or anomalies in the device are identified. To verify the method, we designed harmonic reducer fault experiments under different laboratory conditions to diagnose the obtained multi-sensor data using the model. The experimental results show that the proposed model has excellent fault diagnosis performance and outperforms the current methods commonly used for harmonic reducer fault diagnosis.