MDFF: Multi-Domain feature fusion for anomaly recognition

In the field of anomaly recognition in industrial processes, it has been observed that variations in the temporal frequency and spatial density of process data are a ubiquitous phenomenon, which lead to a reduction in the accuracy of anomaly recognition. In response to these challenges, a novel anomaly recognition approach, termed Multi-Domain Feature Fusion (MDFF), is proposed. In this method, temporal features are initially captured using a Gated Recurrent Unit (GRU) network. Subsequently, the characteristics of the frequency domain are extracted through a combination facilitated by the Fast Fourier Transform (FFT) algorithm and a convolutional network, enabling the detailed analysis of frequency components within anomalous signals. Spatial features are extracted from the input dataset through a one-dimensional Convolutional Neural Network (1D-DCNN), which is augmented by data segmentation and cascade connections. Additionally, for the purpose of augmenting the effective fusion of features across diverse domains, a feature fusion technique, grounded on a channel attention mechanism, is implemented. The efficacy of the proposed method is assessed through simulation experiments conducted on Tennessee Eastman process and blast furnace iron-making process. The results from these experiments demonstrate that our MDFF method is superior in the task of anomaly identification.