基于多维差异感知的生产工况异常检测与故障预测方法研究

It is very necessary that anomaly of operating condition is perceived timely and accurately so as to ensure production quality and safety, eliminate potential accidents, avoid accidents for complex process industry. Considering the problems that anomaly is caused by complex and varied factors, features reflecting the conditions is correlate with production targets, numerous operational variables and parameters, and it is very difficult to condition changes are perceived and then fault or failure foretastes and potential accidents are perceived through these changes, approaches to anomaly detection and fault prediction based on multi-way dissimilarity perception are researched systematically in this project for process operating condition. A novel method based on multi-way day for describing the modes of features vectors, features distributions and features trending sequence is proposed. Multi-way dissimilarity measurement methods based on heterogeneous and uneven property of features are presented. Dynamic detection for complex condition is achieved by multi-way dissimilarity perception and pattern matching of multi-way features vectors, multi-way features distributions and multi-way features trending sequences. Theories and methods of anomaly detection and fault prediction based on multiple kernel learning are formed for complex process. A new solutions schema is offered for process condition of anomaly detection, fault prediction, decision on optimization control and maintenance for the production of complex process industrial. It has also great theoretical and significant practical values to improve product quality, ensure safety, reduce maintenance cost.

对流程工业等复杂过程，及时准确感知工况异常变化并采取相应措施，是确保生产质量及安全、及时消除事故隐患、避免生产事故的必然要求。本项目针对引起流程工业生产工况异常的因素复杂多变，反映工况的特征和生产指标、操作变量及工艺参数众多且相互关联，感知工况变化及通过该变化感知故障或故障先兆和事故隐患困难等特点，系统研究基于差异感知的工况异常检测与故障预测方法，主要研究：基于多维数据体结构描述的特征向量、特征分布及特征趋势序列等工况特征模式表征方法；基于异构不平坦属性的多维差异性测度方法；通过多维差异感知和多维特征向量、特征分布及特征趋势序列模式匹配实现复杂工况动态检测，形成基于多核框架的复杂工况异常检测与故障预测理论与方法。项目的研究成果将为流程工业生产过程的工况异常检测、故障预测与优化控制及维修决策提供新的解决模式，对提高产品质量、确保生产安全、降低维护成本具有十分重要的理论意义和显著的实用价值。

针对浮选生产中普遍存在的工况复杂多变、影响因素众多，反映工况的特征类型繁多属性各异且与生产指标（精矿品位）、入矿品位、操作变量（加药量等）、工艺参数之间相互关联，感知工况变化以及通过这些变化感知异常或故障工况困难等问题，首次将Kinect传感器引入到锑粗选泡沫浮选工业现场，研发了基于Kinect传感器的泡沫图像获取装置；结合浮选过程不同工况下泡沫图像特征的差异分析，提出了辅以深度信息的浮选泡沫图像视觉特征提取方法、基于RGB-D数据的浮选泡沫特征提取方法、基于差异分析的改进的泡沫特征提取方法，有效提高了所提取泡沫特征对工况差异的感知能力；提出了反映不同浮选生产工况泡沫图像视觉特征向量、特征分布及特征序列等视觉特征模式多维表征方法以及基于异构不平坦属性的多维差异性测度方法，提升了多维异构泡沫特征对工况差异的理解能力。提出了基于特征差异分析的工况识别方法、异常/病态/故障工况检测/预测方法，有效提高了多维异构泡沫特征对工况差异的识别能力。构建了基于差异分析的工况识别框架，研发了在线工况识别系统，为锑粗选生产现场提供了浮选生产管理和操作指导信息，为浮选生产的安全、高效、稳定运行创造了条件，为实现浮选过程不同工况下的有效控制和优化奠定了基础。项目的研究成果可为流程工业生产过程的工况异常检测、故障预测与优化控制及维修决策提供新的解决模式，对提高产品质量、确保生产安全、降低维护成本具有十分重要的理论意义和显著的实用价值。