(314c) Exploring Generative Adversarial Networks for Fault Detection and Diagnosis in Chemical Processes: A Distillation Column Study

The growing complexity of modern chemical process systems has heightened the importance of fault detection and diagnosis (FDD) in ensuring process safety, reliability, and product quality. With large-scale operations and numerous variables at play, faults such as catalyst deactivation, valve blockage, or compressor failure have become inevitable occurrences. Early detection and diagnosis of these faults are crucial for enhancing safety, and productivity, and reducing machinery defectiveness, potentially saving billions of dollars for the process industries [1]. However, implementing effective FDD systems in the chemical industry faces challenges. Fault occurrences in chemical processes are not frequent, and even when data is available, it may not be labeled as normal or faulty. Also, the available data may suffer from class imbalance, complicating the training of accurate fault detection models and potentially leading to biased detection performance.

One potential solution to the imbalance issue in fault detection and diagnosis (FDD) is the generation of synthetic data using Generative Adversarial Networks (GANs) [2]. GANs have been claimed to be able to create synthetic data that can help balance datasets and improve the performance of FDD algorithms [3, 4]. However, challenges arise when employing GANs for FDD. Firstly, GANs typically demand a significant amount of labeled data to effectively learn the underlying data distribution, which can be impractical or costly to obtain. Secondly, through the generation of synthetic data, GANs effectively model the system's distribution, resembling a form of system modeling. This raises questions about the necessity and practicality of integrating GANs into FDD processes, as it may overlap with existing efforts in system modeling, potentially duplicating resources and complicating the FDD framework. We hypothesize that the synthetic data created by GANs provides no additional value or real information about the process.

The presentation will test the hypothesis using a distillation column, an ethylene splitter (C2 splitter), simulated using ASPEN dynamic software. The plant incorporates a Proportional-Integral (PI) controller for purity control, where changes in upstream conditions can lead to flooding. In our study, we aim to first manipulate inputs to simulate normal and flooding conditions in the distillation column. Then utilize a portion of the generated data for GAN training apply existing GAN methods to the data and generate synthetic samples. After that, an FDD framework will be developed using the synthetic data generated by GANs. A performance comparison between the GAN-based FDD framework and existing fault detection methods (only based on the `real' data available) will be conducted. The performance of the framework will also be compared to previous work [5] that has been applied to the C2 splitter. By this approach, the practicality and efficacy of the GAN-based approach in comparison to conventional FDD methods will be assessed, particularly in the context of distillation columns, which are critical components in the chemical industry. In this approach, two scenarios will be explored. First, unsupervised FDD using a single GAN trained on both normal and faulty data, and second, supervised or semi-supervised FDD using two GANs—one generating normal data and the other generating faulty samples. By conducting this study, we aim to evaluate the feasibility and practicality of GAN-based FDD frameworks in addressing the challenges of imbalanced datasets in the chemical industry.

References

[1] Md Nor, N., Che Hassan, C.R. and Hussain, M.A., 2020. A review of data-driven fault detection and diagnosis methods: Applications in chemical process systems. Reviews in Chemical Engineering, 36(4), pp.513-553.

[2] Xu, R. and Yan, W., 2020, July. A comparison of gans-based approaches for combustor system fault detection. In 2020 International Joint Conference on Neural Networks (IJCNN) (pp. 1-8). IEEE.

[3] Yan, K., Chong, A. and Mo, Y., 2020. Generative adversarial network for fault detection diagnosis of chillers. Building and Environment, 172, p.106698.

[4] Liu, J., Qu, F., Hong, X. and Zhang, H., 2018. A small-sample wind turbine fault detection method with synthetic fault data using generative adversarial nets. IEEE Transactions on Industrial Informatics, 15(7), pp.3877-3888.

[5] Jalanko, M., Sanchez, Y., Mhaskar, P. and Mahalec, V., 2021. Flooding and offset-free nonlinear model predictive control of a high-purity industrial ethylene splitter using a hybrid model. Computers \& Chemical Engineering, 155, p.107514.