Intelligent fault diagnosis of bearings based on unsupervised domain adaptive adversarial graph neural network under variable operating conditions

In practical industrial diagnostic scenarios, there is often a significant lack of labeled information for equipment fault samples. Additionally, dynamic changes in operating parameters cause shifts in the spatial distribution of sample features, resulting in a dual challenge of label scarcity and time-varying data distribution. The existing unsupervised domain adaptation framework relies on a single supervised signal mechanism, making it difficult to effectively address the challenges in dynamic migration scenarios. Consequently, the paper introduces an unsupervised domain adaptive adversarial graph neural network, known as UDGAT, which fuses three key information sources into the model: specific class labels, domain labels, and data structure de tails. The class and domain label information are respectively handled by classifiers and domain discriminators. The proposed framework extracts input signal features from the data structure information using a variational autoencoder (VAE). Subsequently, the extracted VAE characteristics are fed into the suggested graph attention module for constructing graph-based data. Attention mechanisms and speed information are incorporated into the graph data to produce instance graphs. To assess structural variations across these graphs from different domains, the maximum average difference metric is u sed. Ultimately, this study employed bearing experimental data and real electric vehicle (EV) transmission system data to validate the proposed method and conducted a comparative analysis with cutting-edge approaches. This bearing fault diagnosis approach demonstrates strong adaptability to varying operational conditions, offering practical significance for intelligent maintenance in EV.