Graph neural networks (GNNs) have made significant achievements in the field of artificial intelligence, providing innovative approaches for numerous classification and prediction problems. However, early graph neural network models struggled to capture different types of relationships between nodes. Inspired by the attention mechanism in the Transformer architecture, which can dynamically allocate weights based on the relationships between nodes, graph attention networks has been proposed. Nevertheless, these node relationships are primarily based on feature relationships and find it challenging to address cases where the correlations rely more on human experience than on features. This limitation hampers their effectiveness in industrial production scenarios. To address this issue, we proposes a graph node aggregation method based on fuzzy logic, called FGAT. On top of the attention mechanism, a fuzzy membership function is added to introduce a strong relational inductive bias, allowing for more flexible weight adjustments. Tests on three different public datasets demonstrate that FGAT performs better than traditional graph aggregation methods.