Technical contradictions identification, a foundational step in TRIZ-based systematic innovation, faces significant challenges when automated from patent documents due to implicit expressions and logical complexity. This paper introduces an integrated large language model (LLM)-based framework enhanced by a novel Two-Stage Knowledge Fusion mechanism and Dialectical Validation for precise contradiction extraction and validation. The proposed method consists of five sequential stages: (1) contradiction context identification; (2) template-guided contradiction summarization; (3) dialectical validation to ensure logical consistency; (4) precise technical parameter extraction; and (5) systematic TRIZ parameter mapping. Our Two-Stage Knowledge Fusion approach dynamically selects and integrates relevant TRIZ domain knowledge, significantly improving model precision and interpretability. Experimental results on patent datasets demonstrate substantial enhancements in contradiction extraction accuracy and validation reliability compared to existing approaches. Our framework not only bridges unstructured patent texts with structured TRIZ methodologies effectively but also provides interpretable outputs directly usable for subsequent inventive solution generation processes.