IndustFlow: MLLMs-Driven Industrial Processes Flowchart Synthesis with Cross-Domain Knowledge Alignment

The generation of industrial process flow charts is crucial in modern production, but traditional deep learning models struggle with inconsistent formats, terminology variations, and noisy data. To address these limitations, we propose a novel framework for industrial flowchart generation based on Multi-Modal Large Language Models (MLLMs), named IndustFlow. IndustFlow leverages MLLMs to convert industrial graphics into detailed, interpretable textual descriptions, improving the utility and comprehensibility of industrial data. Additionally, we introduce a Graph-Based Retrieval-Augmented Generation (GraphRAG) model, which integrates knowledge graph structures to generate precise and contextually enriched outputs. The framework also enables cross-domain knowledge alignment by bridging domain-specific terminologies and leveraging vast external knowledge sources. Finally, we enhance the logical reasoning and accuracy of flowchart generation through Group Relative Policy Optimization (GRPO), enabling reliable handling of complex industrial workflows. We collected thousands of noisy industrial documents to evaluate our fine-tuned model, achieving exceptional experimental results that outperform many advanced methods.