Data-driven disciplines such as biostatistics and chemometrics are undergoing a period of transformation propelled by powerful advances in computational hardware, parallel processing and algorithmic efficiency. Process systems engineering is positioned for concurrent advances in data-driven sub-disciplines such as modeling, optimization, control, fault detection and diagnosis. This work embodies this transformation as it addresses a novel industrial fault detection problem from both traditional and contemporary approaches to process analytics. Traditional approaches such as partial least squares are compared with powerful new techniques inspired by deep representation learning such as convolutional neural networks. Novel contributions include the formulation and introduction of a novel industrial predictive classification problem, the design and implementation of a comprehensive machine learning workflow that converts raw industrial data into critical operational insights, and the presentation of a robust comparative analysis between traditional and contemporary approaches to representation learning and binary classification. Specifically, this work addresses the unexpected loss of plasma arc in the electric arc furnace of a large-scale metallurgical process. The objective is to learn an efficient and informative representation from the raw industrial data that enables the prediction of an arc loss event such that operators can take corrective actions. A comprehensive representation learning and predictive classification framework is presented for development of the inferential sensor from large quantities of historical industrial process data.

诸如生物统计学和化学计量学之类的数据驱动学科正经历着一段变革时期，这是由计算硬件，并行处理和算法效率的强大进步推动的。过程系统工程的定位是在数据驱动子学科（例如建模，优化，控制，故障检测和诊断）中同时取得进步。这项工作体现了这种转变，因为它解决了传统的和现代的过程分析方法中的新型工业故障检测问题。传统方法（例如偏最小二乘）与强大的新技术进行了比较，这些新技术的灵感来自深度表示学习，例如卷积神经网络。新颖的贡献包括提出和引入了新颖的工业预测分类问题，全面的机器学习工作流程的设计和实现，该工作流程将原始的工业数据转换为关键的操作见解，并展示了传统和现代的表示学习和二进制分类方法之间的强大比较分析。具体而言，这项工作解决了大规模冶金工艺的电弧炉中等离子体电弧的意外损失。目的是从原始的工业数据中学习有效且信息丰富的表示形式，从而能够预测电弧损失事件，从而使操作员可以采取纠正措施。提出了一个全面的表示学习和预测分类框架，用于从大量历史工业过程数据中开发推理传感器。