软件导刊2026，Vol.25Issue(3)：63-68,6.DOI:10.11907/rjdk.251065
基于Transformer与BiLSTM的Docker构建预测
Prediction of Docker Construction Based on Transformer and BiLSTM
摘要
Abstract
In the process of building Docker,establishing an early warning mechanism can accurately predict the occurrence of build fail-ures,thereby effectively saving time costs incurred due to waiting for the completion of the build.A deep neural network model DeepPDBR based on improved source code representation and multi head attention mechanism is proposed to address the problems of difficulty in effective-ly extracting semantic features,capturing complex feature patterns,and deep semantic relationships in existing Docker prediction models based on configuration semantic features and traditional machine learning algorithms.Firstly,the source code is characterized using AST staged parsing to extract deep semantic features,and the AST nodes are serialized and vectorized using Word2Vec for feature vectorization;Secondly,using Transformer encoder to capture global connections and further extract input data features;Again,using BiLSTM network for long-distance dependency feature extraction to preserve the serialized features of the data;Finally,the classification results are obtained through the fully connected layer and Softmax layer.The experimental results show that compared to existing models,DeepPDBR has im-proved accuracy,F1 score,and area under the curve by 5.63％～12.28％,1.47％～50.49％,and 61.20％～93.03％,respectively,confirming its effectiveness in Docker based prediction tasks.关键词
Docker/构建预测/抽象语法树/Transformer/BiLSTMKey words
Docker/building prediction/abstract syntax tree/Transformer/BiLSTM分类
信息技术与安全科学引用本文复制引用
蔡美玲,杨佶,张锦..基于Transformer与BiLSTM的Docker构建预测[J].软件导刊,2026,25(3):63-68,6.基金项目
全军共用信息系统装备预研项目(31511010402) （31511010402）
湖南省重点研发计划项目(2023GK2042) （2023GK2042）
