This paper proposes software architecture reconstruction technology based on layered architecture pattern recognition. The input of the recognition is the source code and the unnecessary source code will be filtered out at first. Then the approach relies on lexical information from the source code to mine the semantic relation between system entities and using a topic model to extract the responsibility of entities

which is then used to cluster these entities into cohesion components. Later

the approach supplements the structural information between components to generate the component graph and use the ILD property to recognize the actual software layers. Based on the results of pattern recognition and the principle of layered pattern

position the nonstandard existing in the system as the reconstruction point

and relevant reconstruction suggestions to assist the designers and developers in the reconstruction implementation. Finally

this paper selects 10 open source software systems in Github and SourceForge as experimental objects to verify the effectiveness of the technology in this paper. This technology can greatly improve the effectiveness of pattern recognition

the recognition effect of illegal refactoring points and the implementation effect of refactoring suggestions. This technology can also assist developers in the implementation of architecture reconstruction to a certain extent

improve the situation of the system violations

and improve the quality of the software.