Integrating Knowledge Graphs (KGs) into recommendation systems as supplementary information has become a prevalent strategy. By leveraging the semantic relationships between entities in KGs, recommendation systems can better comprehend user preferences. Due to the unique structure of KGs, methods based on Graph Neural Networks (GNNs) have emerged as the current technical trend. However, existing GNN-based methods struggle to (1) filter out noisy information in real-world KGs, and (2) differentiate the item representations obtained from the knowledge graph and bipartite graph. In this paper, we introduce a novel model called Attention-enhanced and Knowledge-fused Dual item representations Network for recommendation (namely AKDN) that employs attention and gated mechanisms to guide aggregation on both knowledge graphs and bipartite graphs. In particular, we firstly design an attention mechanism to determine the weight of each edge in the information aggregation on KGs, which reduces the influence of noisy information on the items and enables us to obtain more accurate and robust representations of the items. Furthermore, we exploit a gated aggregation mechanism to differentiate collaborative signals and knowledge information, and leverage dual item representations to fuse them together for better capturing user behavior patterns. We conduct extensive experiments on two public datasets which demonstrate the superior performance of our AKDN over state-of-the-art methods, like Knowledge Graph Attention Network (KGAT) and Knowledge Graph-based Intent Network (KGIN).