Unsupervised Anomaly Detection of Hydro-Turbine Generator Acoustics by Integrating Pre-Trained Audio Large Model and Density Estimation

ObjectiveHydro-Turbine Generator Units (HTGUs) require reliable early fault detection to maintain operational safety and reduce maintenance cost. Acoustic signals provide a non-intrusive and sensitive monitoring approach, but their use is limited by complex structural acoustics, strong background noise, and the scarcity of abnormal data. An unsupervised acoustic anomaly detection framework is presented, in which a large-scale pretrained audio model is integrated with density-based k-nearest neighbors estimation. This framework is designed to detect anomalies using only normal data and to maintain robustness and strong generalization across different operational conditions of HTGUs.MethodsThe framework performs unsupervised acoustic anomaly detection for HTGUs using only normal data. Time-domain signals are preprocessed with Z-score normalization and Fbank features, and random masking is applied to enhance robustness and generalization. A large-scale pretrained BEATs model is used as the feature encoder, and an Attentive Statistical Pooling module aggregates frame-level representations into discriminative segment-level embeddings by emphasizing informative frames. To improve class separability, an ArcFace loss replaces the conventional classification layer during training, and a warm-up learning rate strategy is adopted to ensure stable convergence. During inference, density-based k-nearest neighbors estimation is applied to the learned embeddings to detect acoustic anomalies.Results and DiscussionsThe effectiveness of the proposed unsupervised acoustic anomaly detection framework for HTGUs is examined using data collected from eight real-world machines. As shown in Fig. 7 and Table 2, large-scale pretrained audio representations show superior capability compared with traditional features in distinguishing abnormal sounds. With the FED-KE algorithm, the framework attains high accuracy across six metrics, with Hmean reaching 98.7% in the wind tunnel and exceeding 99.9% in the slip-ring environment, indicating strong robustness under complex industrial conditions. As shown in Table 4, ablation studies confirm the complementary effects of feature enhancement, ASP-based representation refinement, and density-based k-NN inference. The framework requires only normal data for training, reducing dependence on scarce fault labels and enhancing practical applicability. Remaining challenges include computational cost introduced by the pretrained model and the absence of multimodal fusion, which will be addressed in future work.ConclusionsAn unsupervised acoustic anomaly detection framework is proposed for HTGUs, addressing the scarcity of fault samples and the complexity of industrial acoustic environments. A pretrained large-scale audio foundation model is adopted and fine-tuned with turbine-specific strategies to improve the modeling of normal operational acoustics. During inference, a density-estimation-based k-NN mechanism is applied to detect abnormal patterns using only normal data. Experiments conducted on real-world hydropower station recordings show high detection accuracy and strong generalization across different operating conditions, exceeding conventional supervised approaches. The framework introduces foundation-model-based audio representation learning into the hydro-turbine domain, provides an efficient adaptation strategy tailored to turbine acoustics, and integrates a robust density-based anomaly scoring mechanism. These components jointly reduce dependence on labeled anomalies and support practical deployment for intelligent condition monitoring. Future work will examine model compression, such as knowledge distillation, to enable on-device deployment, and explore semi-/self-supervised learning and multimodal fusion to enhance robustness, scalability, and cross-station adaptability.