With the rapid development of Industrial Internet of Things technology, the scale of its software is expanding, the complexity of the system continues to rise, and the problem of software defects has become increasingly prominent. Software defect prediction technology can locate potential defects in advance and improve software reliability. However, most of the traditional software defect prediction methods rely on a single code metric feature, which makes it difficult to fully characterize the complex characteristics of Industrial Internet of Things software in the semantic information, program structure, and software evolution process, resulting in limited prediction performance. In view of the above problems, this paper focuses on the research of multi-feature fusion in software defect prediction in Industrial Internet of Things scenarios, focusing on the analysis of the role of different types of features in defect prediction and their fusion mechanism. Firstly, the features of code metrics, semantic features, and structure features involved in Industrial Internet of Things software defect prediction are analyzed. Secondly, the influence of different feature fusion methods on prediction performance is studied, including feature concatenation, weighted feature concatenation, attention fusion, and gating fusion. The research results have a certain reference value for Industrial Internet of Things software quality assurance and intelligent defect analysis.