Lightweight Anomalous Detection of Hydro Turbine Operation Sound Using Fusion Network Enhanced by Load Information

Abstract

Degraded health of the hydro turbine generates anomalous sound during operations. Anomalous detection of operation sound is essential for health monitoring of hydro turbine operation reliability. However, hydro turbine load changes cause background noise variations, degrading anomalous detection accuracy. This article proposes a novel lightweight information enhancement fusion network (IEFNet) for accurate anomalous detection of hydro turbine operation sound. First, the filter bank (FBank) computes the sound tensor, serving as input to IEFNet’s feature extraction module. Next, sound features are extracted by the residual block (ResBlock) convolution of sound tensors. Then, an attention mechanism is adopted to fuse sound features and load information in the feature enhancement fusion module. Finally, based on the fusion results, anomalous detection of operation sound is achieved by fully connected (FC) layers. The IEFNet enables the detection of anomalous operational sound every 10 s, establishing a robust foundation for turbine health monitoring.