Robust System Modeling, Process Monitoring and Fault Diagnosis in the Era of Big Data

多变量统计方法从采集的数据中提取有价值的信息，用于过程监控和故障诊断，对保证工业过程的正常运行起着重要作用。近年来，随着技术发展的爆炸式加速，收集的数据量呈指数级增长，传统的统计方法因其计算能力的限制而无法实时处理数十亿个数据样本，因此变得不那么可取。此外，大量数据带来了新的挑战，例如杂乱的数据集（具有异常值，缺失值和大噪声的异常数据），异构数据源以及复杂的动态和非线性。为了确保现代大规模工业过程中的能源效率、产品质量和工厂安全，解决这些问题并尽早发现过程异常和设备故障非常重要。因此，在大数据时代，开发大规模鲁棒系统建模、过程监控和故障诊断框架具有很大的动力。长期目标是设计一个系统的框架，从大数据集中提取知识，用于智能决策，包括研究可扩展性和鲁棒性增强，动力学和非线性处理，模型结构重新设计（深度学习和强化学习），以及开发大规模统计分析平台。该计划将同时关注两个短期目标：(1)设计大规模数据预处理技术和直接鲁棒模型，以及(2)设计大规模鲁棒时间序列模型。具体来说，O1将专注于开发先进的并行数据清洗技术和鲁棒融合来融合混合数据源，而O2将结合时间信息并开始设计用于异常检测的鲁棒时间序列聚类。O1和O2将集成可伸缩性和健壮性，以处理上述数据问题。将制定监测和诊断框架，并通过模拟过程和工业过程加以验证。该项目将为过程系统工程（PSE）领域的大数据分析研究提供重要的理论基础，同时也有利于制药行业的药品质量检测等其他领域。早期发现和诊断潜在危害将提高生产率和操作安全性，并减少对环境的影响，这将为加拿大带来显著的经济效益。例如，通过监测电弧炉的漏水，可以有效地避免炉膛爆炸。该计划的研究人员将获得强大的建模和分析技能，这将使他们能够在PSE和其他领域（如计算机科学）从事学术和工业职位，为加拿大劳动力提供高质量的专业人员。

Multivariate statistical methods have been widely studied to extract valuable information from collected data for process monitoring and fault diagnosis, which play an important role to ensure normal operation of industrial processes. Recently, with the explosive acceleration in technologies development, the amount of data collected has grown exponentially, and the traditional statistical methods become less preferable, since they cannot process billions of data samples in real time due to their computational limitations. Moreover, new challenges arise with voluminous data, such as messy datasets (abnormal data with outliers, missing values and large noise), heterogeneous data sources, and complex dynamics and nonlinearity. To ensure energy efficiency, product quality and plant safety in modern large-scale industrial processes, it is important to address these issues and detect process anomalies and equipment malfunctions as early as possible. Therefore, there is a great incentive to develop large-scale robust system modeling, process monitoring and fault diagnosis frameworks in the era of big data. The long-term goal is to design a systematic framework to extract knowledge from big datasets for smart decision making, including research on scalability and robustness enhancement, dynamics and nonlinearity handling, model structure re-design (deep learning and reinforcement learning), and development of large-scale statistical analytics platforms. The proposed program will focus on two shorter-term objectives simultaneously: (O1) design of large-scale data preprocessing techniques and direct robust models, and (O2) design of large-scale robust time series models. Specifically, O1 will focus on developing advanced parallel data cleaning techniques and robust fusion to fuse mixed data sources, while O2 will incorporate temporal information and initiate the design of robust time series clustering for anomaly detection. Together, O1 and O2 will integrate scalability and robustness to handle the aforementioned data issues. The monitoring and diagnosis frameworks will be developed, which will be verified through both simulated and industrial processes. The program will provide a crucial theoretical foundation for research on big data analytics in process systems engineering (PSE) area, while also benefitting other areas such as drug quality detection in the pharmaceutical industry. Early detection and diagnosis of potential hazards will improve productivity and operation safety, and reduce environmental impact, which will bring significant economic benefits to Canada. For instance, furnace explosion can be effectively avoided by monitoring water leakage in electric arc furnaces. The research personnel in the program will obtain strong modeling and analytical skills, which will equip them to pursue academic and industrial positions in PSE and other areas such as computer science, contributing high quality professionals to the Canadian workforce.