A Robust Noise Reduction Approach with Multi-Channel LSTM-GRU Integration for Fault Diagnosis in Chemical Industry Processes.

In the domain of fault diagnosis for chemical production processes, traditional methods predominantly rely on manual feature engineering, which introduces several limitations. These include a high susceptibility to signal noise, restricted capacity for feature extraction, and inefficient data utilization. While deep learning approaches, especially those focused on temporal sequence modeling, show promise in automatically generating and utilizing temporal features, they are still hindered by insufficient architectural flexibility and limited integration of multiscale information. To overcome these challenges, this study proposes a new computational framework that combines multichannel long short-term memory networks with multichannel gated recurrent units. This framework uses a joint preprocessing module that incorporates wavelet transform and fast Fourier transform, which helps preserve important signal details while suppressing high-frequency noise. The framework also dynamically adjusts the depth of the network through a variable-order multichannel structure, enhancing its ability to capture the complexity of faults. Additionally, it combines bidirectional temporal modeling with the efficient information transfer of GRU, enabling the hierarchical fusion of multiscale features. Experiments conducted on the Eastman Tennessee process data set demonstrate that this approach achieves an average accuracy of 96.10% across 21 fault classification tasks. This performance is 14.75% and 5.65% higher than traditional LSTM and GRU, respectively. These results underline the method's effectiveness in high-dimensional noise environments, providing a more reliable solution for real-time fault diagnosis in industrial settings.