机器学习和统计分析相结合的多模态工业过程故障诊断

Industrial manufacturing process generally contains many stable operating conditions. Fault diagnosis of different operating conditions and their switches is very important. With the advent of the artificial intelligence age, the rapid development of machine learning methods have brought new ideas to multimodel process monitoring. In this project, multimode industrial process is dealt with, the new machine learning and statistical analysis methods are adopted, the difference and relationship between the modes are in-depth investigated, dynamic characteristics of stable mode and transition mode are analyzed, the mode features are extracted, and the stable/transition mode are identified accurately; the fault samples are collected, the fault features are extracted, the fault/normal pattern classifier is designed, the fault detection model is set up, and the alarm threshold is set and optimized; the causes of the fault are traced based on the fault characteristics,and the fault diagnosis methods are developed. Finally, the fault detection and diagnosis scheme of multimode industrial process is established, and a series of original results of mode/fault feature extraction, mode division, fault detection/recognition are obtained; simulation and software platform are constructed to verify the theoretical research results, which will lay foundation for the large-scale application of these results. This project is in the subject frontier, which comes from the actual production and returns to the engineering application, and has great theoretical significance and practical value.

工业制造过程一般具有多个稳定工况，不同工况及其切换过程中的故障诊断至关重要。随着智能时代的来临，迅速发展的机器学习方法给多模态过程监测带来了新的思路。本项目针对多模态工业过程，以最新的机器学习方法和统计分析方法为手段，深入挖掘不同模态之间的区别和关系，分析稳定模态和过渡模态的动态过程，提取模态特征，精确区分稳定/过渡模态；采集故障样本，提取故障特征，设计故障/正常模式分类器，建立故障检测模型，设置和优化报警阈值；从故障特征出发追溯故障原因，发展故障诊断方法。最终建立多模态工业过程的故障检测和诊断体系，取得一系列模态/故障特征提取、模态划分、故障检测/识别方法的原创成果，并在仿真平台和软件平台上验证，为未来该方法在制造工业中的推广应用做好准备。本项目处于学科前沿，既来源于生产实际，又回归于工程应用，具有重要的理论意义和实用价值。

工业制造过程一般具有多个稳定工况，不同工况及其切换过程中的故障诊断至关重要。本项目针对多模态工业过程，取得一系列模态/故障特征提取、模态划分、故障检测/识别方法的原创成果。其中重要成果包括：采用慢特征分析的方法辨识稳定模态和过渡模态，提出了基于轨迹的过渡模态操作故障的监控方法；提出了基于时间加权核稀疏表示的实时非线性多模态过程监控方法，同时解决了实时性、非线性、可解释性问题；针对带有大量离群点的多模态过程，提出了基于矩阵差分分解的鲁棒过程监控方法，并进一步提出了基于核函数分解的鲁棒非线性过程监控方法；提出基于k近邻（KNN）的密度峰值聚类（DPC）方法来识别不同的稳定/过渡模态；采用统计模式分析（SPA）全面挖掘多模态数据集潜在的统计特征，提出了基于改进加权支持向量数据描述（SVDD）的非线性多模态过程监控方法；提出了多操作阶段的全流程多模态工业过程广义监测方法，解决了多操作下平稳和非平稳状态交替出现的问题；在多维重构贡献（RBC）的基础上，提出了基于贝叶斯决策或类间差别分析的多变量故障分离方法；发展基于一维卷积神经网络的多通道多头故障分类方法；设计了强覆盖脉冲幅度差（ADSCI）特征来获取滚动轴承的退化信息，提出了一种通用的基于似然概率比较（LPC）的健康阶段划分方法。在本项目的支持下，课题组出版专著1部在IEEE Trans、JPC等顶级期刊上发表论文35篇，其中SCI论文19篇，EI论文13篇；授权国家发明专利20项。并且，相关成果在半导体制造业、船舶系统、核电工业和薄膜打印业中得到了应用，解决了实际问题。本项目处于学术前沿，既来源于生产实际，又回归于工程应用，具有重大的理论意义和实用价值。