Research on intelligent diagnosis method of bolt loosening for drum screen in nuclear power plant

Addressing challenges posed by the complex structure, weak signals, and significant noise interference in nuclear power drum screen systems, this study proposes a fault diagnosis method for bolt loosening. Building upon Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) combined with wavelet packet analysis for denoising, the method incorporates a Convolutional Neural Network (CNN) and a Bi-directional Long Short-Term Memory network with an attention mechanism (BILSTM-attention). Initially, CEEMDAN decomposition extracts multiple intrinsic modal components, with noise-containing components identified using correlation coefficients and variance contribution ratios and processed using wavelet packet thresholding noise reduction. Features are extracted from Vibration signals of bearing seats using CNN, while the BILSTM layer preserves global and local signal features in the time dimension. The integration of an attention mechanism (AM module) enhances feature encoding and fault diagnosis accuracy of BILSTM. Experimental validation on a custom drum screen bolt dataset demonstrates a 99% accuracy in diagnosing bolt loosening faults using the proposed method. Compared to CNN, CNN-LSTM, and CNN-BILSTM-AM models without denoising, the proposed method achieves accuracy improvements of 16.4%, 5.31%, and 3.1%, respectively.