With the rapid development of large language models (LLMs) in the field of natural language processing, their application potential in mental health role simulation, early screening, and auxiliary diagnosis has attracted significant attention. Existing work still lacks an implementation framework for simulating individuals with different depression severity levels based on LLMs, as well as solutions for localized deployment. This paper proposes a framework for simulating communication with depressed patients and conducting psychological cognitive diagnosis using LLMs based on chain-of-thought (CoT) prompts. By integrating Chinese diagnostic dialogue data on depression and the Self-Rating Depression Scale (SDS), this study systematically explores the capabilities of different open-source LLMs in simulating the roles of depressed patients with varying severity, generating diverse linguistic behaviors, and evaluating self-perceived psychological cognitive distortions. The experimental results show that mainstream open-source LLMs perform well in terms of role simulation accuracy, consistency in self-assessment using psychological scales, and quality of language generation. Furthermore, the integration of patient background information can effectively enhance the contextual relevance and semantic expression accuracy of the language generated by the models. The framework proposed in this paper provides effective insights for improving the practicality of LLMs in the field of mental health and enhancing their reliability in multi-turn psychological counseling and auxiliary diagnosis, verifying the application value of a multi-level comprehensive evaluation method that combines professional domain data and psychological scales in patient role simulation.