Project Details
Projekt
Data Generation and Knowledge-based Augmentation: Continuous OME Production
Applicant
Professor Dr.-Ing. Jakob Burger
Subject Area
Technical Thermodynamics
Chemical and Thermal Process Engineering
Chemical and Thermal Process Engineering
Term
since 2022
Project identifier
Deutsche Forschungsgemeinschaft (DFG) - Project number 459419731
Recent advances in machine learning (ML) gave rise to novel methods for detecting anomalies and faults in chemical processes. Due to a lack of actual process data with public access, they are typically developed and bench-marked with synthetic data from dynamic process simulations. This procedure has considerable limitations since the data is idealized, and many plant anomalies are hardly predictable in simulations without experiments. The primary objective of Project B2 is to overcome these limitations and provide large amounts of experimental process data of an actual continuous chemical plant in non-anomalous and anomalous operation points. The plant of study is an existing mini-plant for the production of synthetic diesel fuels at the applicant's technology lab. It consists of a reactor, a distillation train, and recycles. The plant is equipped with industry-typical sensors (temperature, pressure, flowrate, levels, offline analyses) and advanced sensors (cameras for detecting precipitation and changes in product color). The produced experimental data is the essential base for developing the ML methods for anomaly detection in Research Area A of the Research Unit (RU). However, the generated experimental data is still too sparse for training the developed deep learning methods. Therefore, another major objective of Project B2 is to provide additional, pseudo-authentic synthesized data based on the experimental plant data and physical knowledge in mechanistic model equations. The equations consist of conservation laws for material and energy and equations describing the mixture's chemical reactions and thermodynamic properties. They will be implemented into a steady-state simulator of the plant. Parts of the equations will be selected and modified, yielding guaranteed relationships among process variables for Projects A2, A3, and A4. In collaboration with projects A4 and B1, methods to generate synthetic data sets are developed. Thereby, the results of the mechanistic process simulation of the plant are modified and augmented by noise, non-measure process variables, and dynamic interpolations using ML methods that are trained by comparing synthetic and actual experimental data. For supporting the ML methods, dynamic interpolations are produced using mechanistic Hammerstein models. The generated experimental and synthetic data will be collected, stored, and disseminated with open access in collaboration with Project B1 for the RU and the community beyond. The combined data serves as training and evaluation data for the advanced ML methods of anomaly detection (Project A1), exploration, explanation, and visualization (Project A3). In turn, Project B2 will test the methods developed in Projects A1 and A3 in plant operations and provide precious feedback.
DFG Programme
Research Units
