Abstract: To address the scarcity of labeled data, weak early fault signals, and severe background noise interference in the acoustic diagnosis of hydraulic turbine flow channels, An unsupervised acoustic fault diagnosis method based on the Convolutional Block Attention Module (CBAM) and Generative Adversarial Networks (GAN) was proposed. This method constructs a deep diagnostic model (CBAM-WGAN-AE) that integrates a Wasserstein Generative Adversarial Network (WGAN) with an Autoencoder (AE). It is trained using exclusively acoustic signal data from normal operating conditions to learn the deep feature distribution, and leverages the CBAM to enhance sensitivity to key fault features while suppressing irrelevant noises. Additionally, an anomaly detection mechanism based on the K-Nearest Neighbors (KNN) algorithm was introduced and the turbine’s abnormal states were detected by calculating the nearest neighbor distance deviations of test samples from the corresponding normal samples in the feature space. Through validating against the fault experimental data from a model hydraulic turbine, The new model improves the Area Under Curve (AUC) up to 91.71%, outperforming previous diagnostic models reported in both overall accuracy and the ability to identify weak fault features.

Key words: hydraulic turbine, unsupervised learning, fault diagnosis, generative adversarial networks, noise