As the complexity of software systems continues to increase, traditional software testing methods face problems such as low test efficiency, insufficient path coverage, and high labor costs, and intelligent solutions are urgently needed. This study aims to build a software engineering intelligent testing system based on deep learning algorithms to improve test quality and efficiency by improving test case generation and defect detection. At the method level, a bidirectional long short-term memory network (Bi-LSTM) is first used to extract features from historical test data and construct a code semantic vector representation; secondly, a dual-channel model that integrates convolutional neural networks (CNNs) and graph attention networks (GATs) is designed to process code text features and program dependency graph structure features, respectively; then, the adversarial generative network (GAN) framework is introduced to dynamically generate high-coverage test cases, and the test strategy selection is optimized through a reinforcement learning module. Experimental results show that on standard data sets such as JIRA and SIR, the system achieves 95.7% path coverage and 93.2% defect detection accuracy, while reducing the average test time to 2.9 h. The system effectively solves the problem of test coverage blind spots in complex scenarios and provides a verifiable technical path for automated testing of software engineering.