Denoising and Recognition Method for Weak Acoustic Abnormal Signals in Hot-Wall Hydrogenation Reactors Using DnCNN-CNN

Accurate identification of damage in hot-wall hydrogenation reactors is a challenging and costly task. Acoustic Emission (AE) technology is considered a promising approach owing to its high sensitivity to micro-damages. However, during the operation of the hot-wall hydrogenation reactor, high-intensity operational noise often masks weak damage signals, presenting a significant challenge for damage monitoring under strong noise conditions. To address this issue, this study proposes a deep double convolutional neural network that combines denoising convolutional neural networks (DnCNN) and convolutional neural networks (CNN) for denoising and recognition of weak abnormal AE signals under strong noise conditions. Short-Time Fourier Transform (STFT) is employed to convert the AE time-domain signals into time-frequency domain joint representations, constructing a compressed spectrogram as the network input. The DnCNN module removes noise from the mixed spectrogram, while the CNN module performs signal classification. The model was validated through three representative case studies. Experimental results demonstrate that the proposed method achieves excellent signal denoising performance with limited prior knowledge of signals and noise. Furthermore, the model demonstrates strong generalization and robustness in recognition tasks, significantly improving the accuracy of abnormal signal recognition. This study provides technical support for the application of AE monitoring technology in the safety operation of hot-wall hydrogenation reactors.