Adversarial Regularized Graph Autoencoder for Intelligent Anomaly Detection With Multisensor Signal Fusion

Gears play an indispensable role in the transmission systems of advanced industrial equipment. Timely detection of anomalies in gear components is crucial for ensuring operational efficiency and safety. This article introduces an Adversarial Regularized Graph Autoencoder (ARGAE) enhanced by a multiscale Chebyshev Convolutional (ChebConv) encoder (MSCE) (ARGAE-MSCE), designed for unsupervised anomaly detection (UAD) in gear multisensor signals. The methodology constructs graph-structured datasets from multisensor data, treating each sensor as a node with features extracted via fast Fourier transform (FFT) and connected using the RadiusGraph approach. The proposed MSCE captures data characteristics at various scales, enhancing feature extraction capabilities. Integrating adversarial training into the graph autoencoder (GAE) framework imposes regularization on the latent space, improving the distinction between normal and abnormal data representations. The anomaly score is defined by the discrepancy between the node feature matrices before and after reconstruction. Experimental validation on two gear multisensor datasets demonstrates that ARGAE-MSCE achieves an outstanding accuracy of 99.7% and 99.8%, highlighting its effectiveness and robustness in multisensor signal anomaly detection, offering a powerful solution for predictive maintenance and anomaly detection in industrial applications.