Integrated Deep Neural Network and Hypersphere Information Granule for Imbalance Fault Diagnosis

This paper proposes a integrated deep neural network and hypersphere information granule (DeepHIG) model, which specifically designed to enhance the fault detection and diagnosis (FDD) performance in the all-fault imbalanced (AFI) scenario. The deep learning-based feature representation learning (DL-FRL) module is used to manage complex nonlinear and temporal features. The hypersphere information granule-based classifier (HIGC) module employs an adaptive loss-based on the principle of justifiable granularity (AL-PJG) technology to granulate deep fuzzy features, and integrates the hypersphere information granules into the fuzzy rule antecedents to map features to labels. This novel granulation method accurately describes the feature structure, emphasizing fault categories with fewer samples. A successive learning technique is proposed during training to support the seamless transformation between original space, feature space, granule space, and label space, promoting the mutual adaptation of deep features and information granules, and achieving high-level FDD performance. Experiments are conducted on the Tennessee Eastman (TE) real benchmark process to demonstrate the effectiveness of the proposed DeepHIG model.