[Paper Review] Benchmarking Domain Adaptation for Chemical Processes on the Tennessee Eastman Process

Eduardo Fernandes Montesuma, Michela Mulas+3 more|arXiv (Cornell University)|Aug 22, 2023

Fault Detection and Control SystemsEngineering3 citations

TL;DR

This paper proposes multi-source unsupervised domain adaptation (MSDA) for cross-domain fault diagnosis (CDFD) in chemical processes using the Tennessee-Eastman (TE) process benchmark. It demonstrates that leveraging multiple heterogeneous source domains during training significantly improves fault classification accuracy—achieving a 23.48% average improvement over single-source domain adaptation (SSDA) baselines—with optimal transport (OT)-based methods like DaDiL and JDOT outperforming others due to effective distributional shift minimization.

ABSTRACT

In system monitoring, automatic fault diagnosis seeks to infer the systems' state based on sensor readings, e.g., through machine learning models. In this context, it is of key importance that, based on historical data, these systems are able to generalize to incoming data. In parallel, many factors may induce changes in the data probability distribution, hindering the possibility of such models to generalize. In this sense, domain adaptation is an important framework for adapting models to different probability distributions. In this paper, we propose a new benchmark, based on the Tennessee Eastman Process of Downs and Vogel (1993), for benchmarking domain adaptation methods in the context of chemical processes. Besides describing the process, and its relevance for domain adaptation, we describe a series of data processing steps for reproducing our benchmark. We then test 11 domain adaptation strategies on this novel benchmark, showing that optimal transport-based techniques outperform other strategies.

Motivation & Objective

• To evaluate the effectiveness of multi-source domain adaptation (MSDA) in cross-domain fault diagnosis (CDFD) for chemical processes.
• To compare single-source (SSDA) and multi-source domain adaptation (MSDA) methods in handling distribution shifts across different modes of operation in the Tennessee-Eastman (TE) process.
• To investigate whether combining data from multiple heterogeneous source domains improves fault diagnosis performance compared to single-source or target-only baselines.
• To assess the performance of various domain adaptation techniques—particularly optimal transport (OT)-based methods—on real-world chemical process data with distributional shifts.
• To provide a comprehensive benchmark for future research in transfer learning for fault diagnosis in industrial chemical systems.

Proposed method

• The study uses the Tennessee-Eastman (TE) process as a benchmark for fault diagnosis, simulating multiple operational modes with distinct data distributions.
• Single-source domain adaptation (SSDA) methods are evaluated by training on labeled data from one source mode and unlabeled data from a target mode, with domain shift minimized via distribution alignment.
• Multi-source domain adaptation (MSDA) methods are applied by training on labeled data from multiple source modes and unlabeled data from a target mode, leveraging data diversity to improve generalization.
• Key techniques include optimal transport (OT) methods such as OTDA, JDOT, and DaDiL, which minimize Wasserstein distance between source and target domain distributions.
• The framework compares shallow methods (e.g., MMD, TCA, DANN) and deep neural network-based approaches, with feature extractors either fine-tuned or fixed during adaptation.
• Performance is evaluated using classification accuracy across all operational modes, with baselines including source-only (multi-mode labeled data) and target-only (unlabeled target data only).

Experimental results

Research questions

• RQ1 Does multi-source domain adaptation (MSDA) significantly improve fault diagnosis accuracy compared to single-source domain adaptation (SSDA) in chemical processes?
• RQ2 How do different domain adaptation techniques—especially optimal transport (OT)-based methods—perform in minimizing distributional shifts across heterogeneous operational modes?
• RQ3 To what extent does combining data from multiple source domains improve generalization in cross-domain fault diagnosis (CDFD), even without explicit adaptation?
• RQ4 Are shallow domain adaptation methods more effective than deep neural network-based approaches in the context of CDFD on the TE process?
• RQ5 How does the performance of domain adaptation methods vary across different operational modes, particularly when modes are closely related (e.g., modes 1 and 5) or dissimilar?

Key findings

• MSDA improves classification accuracy by an average of 23.48% over the SSDA baseline, demonstrating that data diversity from multiple sources enhances model generalization.
• The best-performing method, DaDiL (a Wasserstein barycenter-based OT method), achieves an average accuracy of 86.14% in the MSDA setting, outperforming the best SSDA method (JDOT at 83.13%).
• OT-based methods such as OTDA and JDOT consistently outperform other approaches like TCA, DANN, and MMD, confirming the effectiveness of optimal transport in distribution alignment.
• Shallow domain adaptation methods outperform deep learning-based methods, suggesting that effective feature extraction reduces the need for complex encoder training when distribution alignment is well-optimized.
• Only OT-based methods improve upon the source-only baseline, indicating that distributional shift minimization is essential for performance gains in CDFD.
• In mode 3, SSDA outperforms MSDA, likely due to the presence of a closely related mode (mode 6), highlighting that domain similarity influences adaptation effectiveness.