A Multisensor Feature Representation and Fusion Method for Data-Driven Industrial Process Fault Diagnosis

Data-driven fault diagnosis is a critical component of industrial process monitoring. To more effectively capture the temporal relationships within multisensor data and the spatial relationships among sensors, this article proposed a feature representation and fusion method for fault diagnosis in complex industrial processes. First, a Gramian angular field method with temporal weighting is developed to convert process time series signals into image sequences, where each subimage is assigned corresponding temporal weights. Subsequently, features are represented from each sensor's image sequence, and a spatial-weighted multichannel convolutional neural network is employed to fuse these features and generate new feature embeddings. Finally, a fault diagnosis model is jointly trained to optimize and balance the weight loss across multiple channels. The effectiveness of the proposed method is validated by utilizing the Tennessee–Eastman process and vinyl acetate monomer process, showing significant enhancements in fault diagnosis accuracy and reliability compared to other similar methods.