Fault Feature Extraction Based on Unsupervised Graph Embedding for Harmonic Reducers Diagnosis

In recent years, there has been a growing interest in data-driven intelligent fault diagnosis, focusing on extracting fault features, which has emerged as a critical and challenging aspect of fault diagnosis. However, most existing intelligent diagnostic methods are supervised, relying on many labeled samples and treating them as independent entities, disregarding their correlations. Moreover, the failure of harmonic reducers, as crucial components in robots, can significantly impact the performance and practical applications of the robots. To address these issues, we proposed an unsupervised feature extraction method based on graph embedding, utilizing vibration signal data from harmonic reducers in this study. Specifically, a graph data structure was constructed based on the specific context, and a graph convolutional autoencoder approach was employed to learn representations of subgraphs that captured the sample data, resulting in low-dimensional vector representations of the vibration signals. In the final stage, the extracted features were treated as labeled samples. A subset comprising only 6% of the available samples was used to train several simple models. Subsequently, the remaining sample data was employed for testing purposes. The experimental results demonstrated notably higher accuracy in fault diagnosis, thereby providing strong empirical evidence to validate the effectiveness of the proposed approach.