Quantitative Analysis of Multi-Source Wind Turbine Noise Using Adaptive Decomposition and SHAP

Abstract

Based on the Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN), a one-dimensional convolutional neural network (1D-CNN), and a synchrosqueezed wavelet transform, a new framework is proposed to identify noise sources of wind turbine generators. The contributions of multimodal noise from wind turbine generators are quantified for SHapley Additive exPlanations (SHAP) analysis. First, based on cross-channel correlation, the dual-channel vibration and acoustic signals are adaptively decomposed to realize accurate modal alignment. Then, a 1D-CNN network is used to select the Intrinsic Mode Function (IMF) and extract high-resolution time-frequency features through synchrosqueezing transform. Finally, the digital twin model describes the transmission of sound-vibration, and SHapley Additive exPlanations (SHAP) analysis quantifies the contribution of noise sources. The results show that aerodynamic and electromagnetic noise are dominant in the high-frequency IMF (>70% of the energy), and mechanical noise is dominant in low-frequency IMF (SHAP > 0.8). This model has a high accuracy and can accurately identify noise sources of wind turbine generators.