NEURAL NETWORKS FOR PROCESS FAULT DIAGNOSIS AND SAFETY

这项提案的主要目标是研究和展示一种新的 一种基于新兴神经网络技术的解决方案 人工智能(AI)学科走向设计 化学过程危险检测、预防和控制系统。 这样的系统对于改善职业发展极其重要 现代工艺复杂性带来的化工厂安全问题 植物。拟议的项目旨在预防和控制 这样频繁的、日复一日的、行业内的意外事件。过去时 故障诊断系统中的方法没有适当地包括 人类专家的推理策略和经验，因此没有 有足够的效率和正确的故障排除能力。在这个项目中，我们 提出了一种使用神经网络的新方法来解决 对以知识为基础的发展至关重要的重要问题 过程故障诊断和危险控制系统。我们的方法论 对于开发基于知识的系统来说，重要的是 神经网络的特点，即它们的能力 自动分类并学习输入和输出之间的关联 并且能够处理有噪声的数据。我们已经成功地 在一个基于真实工厂的原型案例研究上测试了我们的方法 年运行的催化裂化装置(FCCU)数据 埃克森美孚的一家炼油厂。我们计划通过以下方式进一步研究这一方法 用更大的化学加工厂原型进行试验 更现实、更复杂的各种单元的模型，如反应堆、 热交换器、蒸馏塔等。我们还计划探索 这类系统对不完整和不确定数据的稳健性。

The major goal of this proposal is to research and demonstrate a new approach based on the emerging technology of Neural Networks in the discipline of Artificial Intelligence (AI) towards the design of chemical process hazard detection, prevention, and control systems. Such systems are extremely important for improving the occupational safety of chemical plants owing to the complexity of modern process plants. The proposed project is aimed at the prevention and control of such frequent, day to day, accidental events in the industry. Past approaches in fault diagnostic systems did not properly include the human expert's reasoning strategies and experience and hence were not adequate in efficient and correct trouble-shooting. In this project, we propose a novel methodology using Neural Networks that address the important issues which are central to the development of knowledge-based systems for process fault diagnosis and hazard control. Our methodology for developing knowledge-based systems rests on the important characteristics of neural networks, namely, their ability to automatically classify and learn associations between input and output data and to be able to handle noisy data. We have already successfully tested our approach on a prototypical case study based on real-plant data, that of fluidized catalytic cracking unit (FCCU) in operation in an Exxon refinery. We plan to investigate this approach further by experimenting with larger prototypes of chemical process plants with more realistic, complex, models for the various units such as reactors, heat exchangers, distillation columns etc. We also plan to explore the robustness of such systems for incomplete and uncertain data.