Abstract: Traditional metric-based software defect prediction models suffer from performance limitations due to their inability to capture syntactic and structural information of source code. While deep learning techniques have demonstrated effectiveness in extracting deep-seated code features to enhance prediction accuracy, existing approaches predominantly rely on singular feature types, resulting in insufficient feature representation and suboptimal model performance. To address this issue, this paper proposes a software Defect Prediction method based on Combines Multiple Features(DP-CMF). The proposed method uses attention-based Transformer and Tree-Based Convolutional Neural Network（TBCNN） to extract and merge semantic and structural features of the source code respectively for feature fusion, then the fused features are combined with the traditional features for defect prediction. This paper is validated in 7 public datasets of PROMISE repository, and experimental results show that compared with existing software defect prediction methods, DP-CMF obtains higher F1-score.