A Small Sample Fault Detection Method Based on an SE-SAGFD

Modern industry poses the challenge of limited data volume and the difficulty in extracting effective spatial relational features. To tackle this issue, an enhanced self-adaptive graph convolutional network (E-SAGCN) is proposed. Different from traditional structural learning methods with predefined fixed edge, the proposed E-SAGCN model enhances the adaptability to different node information by adjusting the learnable edges. On this basis, a fault detection model (E-SAGCFD) is developed, which further mines out spatial relational features and applies it to fault detection effectively. To address the small-sample challenge, the E-SAGCFD model is improved to a small sample E-SAGCFD (SE-SAGFD) model by combining the Siamese Neural Networks (SNN) and E-SAGCFD. This model leverages historical production data from similar products to successfully extract spatial relational features under small sample conditions, enhancing the accuracy of fault detection. The performance of the proposed fault detection algorithm is validated on a real thermal power generation where experimental results show the superiority in the fault detection of small sample industrial process.