Reliable fault detection is an essential requirement for safe and efficient operation of mechanical systems in various industrial applications. As machine complexity increases, the number of sensors required to measure over time to infer abnormal and normal behavior increases dramatically, while despite the abundance of existing approaches and the maturity of fault detection research field, the interdependencies between the multivariate data have often been overlooked. However, current graph autoencoders for multivariate data mining have fixed receptive fields and ignores the distribution of latent space, limiting their ability to extract multiscale features and model performance. To overcome these limitations, a novel graph autoencoder, namely adversarially regularized graph wavelet autoencoder (ARGWAE), is proposed in this work. ARGWAE consists mainly of the spectral graph wavelet convolutional encoder for multivariate data multiscale feature extraction, a feature decoder for data reconstruction, and an adversarial regularizer to force the latent space match the Gaussian prior. To verify the effectiveness of the proposed method, two case studies are carried out and the experimental results show that the proposed method can obtain performance improvements by around 4–9% compared to the comparison methods. The code is available at: https://github.com/HazeDT/ARGWAE.