基于大规模数据驱动的轧钢过程故障诊断及自愈控制

Though most of the rolling production lines have been equipped with the perfect 3-level control systems in China, the massive accumulated datum of production and maintenance as well as product application have not been utilized effectively. There are little theoretical researches and practical applications in this field at home and abroad concern about how to further decrease failure rate, improve productivity and product quality, and let the system have the functions of integrated fault diagnosis and health management in highly modern rolling lines. In this project, aiming at big data in rolling process, based on the feature extraction and ensemble learning methods of multiple Taylor and D-S theory, the best learning machine of character subset is confirmed, also multicollinearity problems are avoided and suppressed effectively. Then a high-dimensional data processing model with CPU+GPU heterogeneous hardware platform in rolling process is established to research the optimization design for fast real-time parallel computation of multiple character subsets. Relevant data information in different components of rolling process are excavated to realize the numerical quantify in every physical part and achieve robust fault detection and diagnosis of the equipment using state observer, then based on uncertain union resolution the prediction and diagnosis of exceptional quality are came true. The models are decomposed with general Bender’s decomposition method. Self-healing control and intelligent maintenance of the faults and potential quality problems in the rolling systems are realized with the help of multi-level and multi-scale coordinating and optimizing technological parameters. Finally, the above theoretical methods and system will be evaluated and verified in field application.

我国轧钢产线大多配置有完备的三级控制系统，积累了大量的生产、维护和产品应用数据，但尚未有效开发利用。如何在高度现代化的轧线上，进一步降低故障率、提高产品质量，使系统具有综合故障诊断和自愈控制功能，国内外此领域理论研究和应用较少。本项目针对轧钢过程的大规模数据，基于多元泰勒和D-S证据理论的特征提取与集成学习方法，确定特征子集的最佳基学习机，规避和抑制多重共线性问题，建立轧钢过程高维数据处理模型，并构建CPU+GPU异构硬件平台，研究多个特征子集实时快速并行运算的优化设计。挖掘轧钢过程各环节关联数据信息，实现各物理环节内在关系的数值量化，使用状态观测器完成设备的鲁棒故障检测与诊断，基于不确定度联合解析实现质量异常预测与诊断。采用广义Bender分解等对模型进行分解，通过多层次、多尺度的工艺参数协调优化，实现轧钢系统的故障和质量异常自愈控制和智能化维护。通过现场应用验证其方法的有效性和实用性。

我国轧钢产线大多配置有完备的三级控制系统，积累了大量的生产、维护和产品应用数据，但尚未有效开发利用。本项目依托高度现代化的轧线，研究设备和产品质量故障诊断方法，成功开发出了面向轧钢生产过程智能故障诊断和自愈控制的工业软件，相关研究成果在沙钢、鞍钢等钢铁企业进行离线和在线应用，进一步降低故障率、提高产品质量，实现了轧制全流程的安全稳定运行。.（1）研究了生产过程高维度、大规模数据的信息特征提取和集成学习方法，开发了高维数据降维技术和集成建模技术，实现了生产过程各物理环节内在关系的数值化和模型化表述，建立了轧钢过程的质量预测模型。.（2）实现了基于模糊状态观测器的设备故障智能检测与诊断；从输入性故障、输出型故障和模型参数偏移故障的特征入手，研究基于状态观测器的设备故障诊断机制；采用现场数据进行仿真实验，完善并测试设备故障检测与诊断方法的性能。.（3）以热轧板凸度、冷轧板形等关键质量指标为例，分析并确定出各工序数据的输入输出关系，实现了多质量指标的集成建模；利用生产过程数据建立产品质量的在线预测模型，基于多元产品质量信息联合解析进行动态调整控制和诊断分析。.（4）建立了以生产成本和多元质量参数稳定度为目标的生产过程多目标优化模型，研究故障干扰下的轧制运行参数协调优化策略和动态调整运行机制；提出了轧制过程运行的多目标动态协同优化方法，实现了轧制生产过程的自愈控制。.（5）构建了CPU+GPU异构硬件平台，通过GPU多个特征子集并行运算需求算法优化设计，满足了大规模数据处理速度需求；开发了基于傅里叶神经算子的GPU加速与自适应重采样随机搜索算法，达到了提高协同优化控制实时性能的目的。.发表相关学术论文88篇，其中SCI收录63篇；申请发明专利35项，其中美国专利1项，授权18项；培养博士研究生毕业8人，硕士研究生毕业14人；获得省科技进步二等奖1项。