Enhanced Weakly Supervised Learning Method for Anomaly Detection in Aerospace Product Manufacturing Processes

Detecting anomalies in the manufacturing process of aerospace products is of great significance for ensuring product quality and reliability. Traditional anomaly detection methods suffer from limited availability of anomaly samples and uncertain labels of monitoring data from multiple manufacturing procedures. Therefore, an enhanced weakly supervised learning method is proposed. First, an anomaly sample generating model is designed to enhance its volume. Then, a multi-network model is designed, where a series of instances learning networks are used to estimate anomaly scores of data from some procedures separately and a stacking network is used to determine whether the whole manufacturing process contains anomalies. The proposed approach is verified on an experimental dataset of star sensor manufacturing, and the results substantiate that it outperforms existing methods.