Access Type

Open Access Dissertation

Date of Award

January 2025

Degree Type

Dissertation

Degree Name

Ph.D.

Department

Chemical Engineering and Materials Science

First Advisor

Helen Durand

Abstract

In the digitalization era and Smart Manufacturing, companies are harnessing the power of artificial intelligence (AI) and machine learning (ML) across multiple sectors, including process engineering optimization, process control and fault detection, to enhance efficiency and engineering decision making. Although AI and ML are widely used in anomaly detection and handling, there are still areas where it has been less explored. One of the major areas where AI’s potential in manufacturing needs to be characterized is with respect to the applications of large language models (LLMs) in manufacturing troubleshooting for fault/attack handling. A second major area where the potential of AI and ML in manufacturing is not fully characterized is with respect to how the use of AI/ML impacts the security of process operation. In this thesis, we address these gaps.

With respect to leveraging LLMs in anomaly handling, we provide investigations into how an industry could move toward enhanced troubleshooting for fault/attack diagnosis and handling where an engineer can provide queries on an encountered issue/anomaly and receives a reply from an LLM model that includes the reasons behind a fault or attack and how to fix it. However, a major challenge for the “enhanced troubleshooting” workflow is that it would require that the LLM always returns verifiably correct outputs without being fine-tuned. Improvements can be performed by grounding the LLM’s responses on proprietary data sourced from vectorized sources through a process called retrieval augmented generation (RAG). Through this process documents that are similar to the user’s query will be retrieved from a vector database to assist the LLM’s model in generating a response that aligns with the user’s query. However, the enhanced troubleshooting strategy has two major challenges related to the relevancy and the accuracy of the retrieved information and the response generated by the LLM model, respectively. In this thesis, we focus on one challenge related to the relevancy problem of the information retrieved from a centralized vector store. Although there are various approaches for addressing the relevancy problem, a company may want to try to assess which strategy would perform best in a constantly growing literature. A common approach for validating the performance of different techniques for utilizing LLMs is to benchmark their performance with a substantial dataset. However, because LLMs are less explored for critical applications like engineering troubleshooting, there is no current database available to be used in the evaluation strategy. Therefore, in this thesis we aim to provide guidance for an industry on how to generate their own dataset to test various enhanced troubleshooting workflows and to examine the effect of the questions generated in various ways on the success of these workflows using an identified success assessment metric.

Next, we seek in this thesis to provide initial steps toward formulating the components of the enhanced troubleshooting workflow from a control-theoretic perspective which might provide steps toward generating guarantees for the enhancement of the relevancy of the information retrieval as well as the accuracy of the LLM’s output. This thesis focuses on providing initial concepts toward this in the domains of fault and cyberattack handling.

In addition to the empirical and control theoretic LLM-based strategies for anomaly diagnosis and handling, this thesis also develops control-theoretic detection and handling strategies for anomalies and more specifically for cyberattacks on sensors. In this anomaly detection and handling direction, we first introduce the concept of cyberattack handling for classical and quantum systems using model-based control. We do this through discussing the extension of an active detection strategy elaborated in our group prior work for chemical processes under an advanced control formulation known as Lyapunov-based economic model predictive control (LEMPC). This strategy probed for cyberattacks by modifying thesteady-state around which the LEMPC was designed, as well as the associated Lyapunov function and Lyapunov-based stability constraints, at random times to check whether the modified Lyapunov functions decreased over the subsequent sampling period. In this thesis, we investigate whether this type of active detection policy can be generalized to other economic model predictive control (EMPC) formulations. In another work in our group, a ‘directed randomization’ strategy based on randomly selecting between two predefined control actions at every sampling time was developed to attempt to force an attacker to reveal himself when continuously attacking the sensor of a process. This thesis provides initial steps toward the extension of such detection mechanism to other cyberphysical systems such as quantum systems. Next, we considered the theme of this thesis, regarding the integration of AI/ML in the design of anomaly detection and we focus on the issue of having the advanced controller fully replaced by learned models such as neural networks (NN) and how that affect the security of the process in the presence of an undetected attack. More specifically, we extend a passive detection strategy elaborated in our group’s prior work to the case of an NN that approximates an LEMPC. We then elaborate a cyberattack handling strategy integrated with reachability analysis and inspired by the NN repair for attempting to ensure safety of controllers for one sampling period after undetected attacks. We then examine the potential conservatism differences between the LEMPC-based safety strategy and reachability-based method, and consider an extension of the repaired input concept to fighting back against attackers.

Download

Share

COinS