Edge-Aware EfficientViT-FPN Model for Vibration Source Separation in Hydropower Units

Source abstract

Identifying the vibration source of a hydropower unit is difficult because hydraulic excitation, bearing looseness, runner disturbance, generator-frame response and mixed-mode structural transmission may occur within the same monitoring window. To support edge-side diagnosis, this paper proposes an edge-aware EfficientViT-FPN model for multi-point vibration source separation. Synchronous vibration signals are acquired from guide-bearing, generator-frame and draft- tube-related positions and converted into reliability-aware spectrogram windows. EfficientViT extracts compact spectral features, while a feature pyramid branch fuses narrow harmonic ridges, local impulse traces and broad low-frequency response zones. A point-quality mask distinguishes weak source evidence from unstable acquisition, and a confidence branch provides review priority for engineering inspection. A simulated monitoring dataset of 6,640 sequence windows is generated under stable-load, partial-load, transition and noisy acquisition conditions. The proposed model achieves 92.8% overall accuracy, a macro-F1 score of 0.914 and an average edge inference latency of 24.7 ms. Compared with the compact convolutional baseline, mixed-source misclassification is reduced by 18.6%. These results show that the model balances computational efficiency, source separation accuracy and evidence traceability. The output reports source label, confidence level, dominant point group and supporting spectral evidence, helping maintenance staff distinguish immediate inspection, remeasurement, trend tracking and low-risk recording cases.