Research on ChatGLM-Based Tower Machine Intelligent Q&A System with Cross-Layer Shared Dynamic Attention

With the increasing demand for automation and intelligence in the construction industry, intelligent Q&A systems in the field of tower crane operation have gained significant attention. Although existing large language models such as GPT and ChatGLM perform well in general natural language processing tasks, they still face significant challenges in the field of tower crane operations due to insufficient ability to focus on key terms caused by scattered attention, as well as information loss and double calculation problems caused by inter layer independence. To address these issues, this paper has improved ChatGLM by combining a dynamically composable multi-head attention mechanism and a cross-layer sharing module. By dynamically adjusting attention allocation, the model can focus on core terms in the field of tower cranes and reduce interference from non core words; Meanwhile, the cross layer sharing module reduces redundant calculations and improves inference efficiency by sharing KV key value pairs and calculation results across layers, thereby performing better in long-distance context dependency, complex task processing, and multitasking. The experimental results show that the improved ChatGLM architecture ChatGLM-DCLA performs well in the intelligent Q&A task within the tower machine domain, with ROUGE-1, ROUGE-2, ROUGE-L and BLEU-4 scores reaching 59.72, 41.01, 57.98, and 31.46, respectively, which significantly enhances the accuracy and effectiveness of the intelligent Q&A systems in the tower crane domain, providing powerful intelligent support for efficient and safe tower crane operation.