Abstract: The production of hot rolled strip is faced with many problems such as multiple working conditions,complex mechanism and various process parameters,which makes it difficult for experts to analysis the causes of quality defects in time and effectively.A method based on knowledge graph reasoning was proposed to analyze the causes of quality defects.The prediction results of random forest model were explained by the Shapley Additive exPlanations (SHAP) method,and the data mining results from SHAP were integrated with domain knowledge such as process mechanism and expert experience by using the knowledge graph.Further,the subgraphs representing the dependence of process parameters and quality parameters in the knowledge graph were extracted and mapped to the Bayesian network in order to infer the posterior probability of product quality defects caused by different process parameters.Actual production data validation was performed,and the results showed that the proposed method could effectively identify the process parameters that caused quality defects in each batch for different working conditions and get good recognition rate.

Key words: hot rolled strip, quality defect, knowledge graph, explainable artificial intelligence, Bayesian networ

摘要： 热轧带钢生产面临多工况、机理复杂、工艺参数繁多等问题,造成专家很难及时有效给出生产中导致质量缺陷的原因。由此提出一种基于知识图谱推理的质量缺陷追溯方法。首先通过可解释方法SHAP实现对随机森林模型预测结果的解释,并通过知识图谱将数据挖掘结果与工艺机理、专家经验等知识进行融合,进一步将图谱中表示工艺参数与质量参数依赖关系的子图映射到贝叶斯网络,推断不同工艺参数导致产品质量缺陷的后验概率。实际生产数据验证表明,针对不同生产工况,该方法能有效识别各个批次中导致质量缺陷的工艺参数,表现出良好识别率。

关键词: 热轧带钢, 质量缺陷, 知识图谱, 可解释人工智能, 贝叶斯网络