(509b) Unsupervised Incremental Learning Framework for Online Fault Diagnosis

The importance of safe operation in industrial processes cannot be overstated, because accidents in chemical processes cause casualties, environmental problems, and economic losses. Therefore, it is critical to detect and analyze faults accurately in a timely manner. As modern industrial processes become larger and more complicated, the need to deal with huge data has arisen. Data-driven methods have been investigated and applied to the field of fault diagnosis to capture the underlying complexity of processes [1]. Data-based fault diagnosis is a classification task that predicts the corresponding fault from available data, and various methodologies based on machine learning have been introduced [2-4].

Incremental learning is a machine learning technique that updates an existing model when new input data becomes available over time [5]. It is used to incorporate new fault data into the model when the data belongs to a class that has not been previously learned and is receiving attention in the field of fault diagnosis [6-8]. However, previous studies assume that the labels of new online data obtainable over time are known, which is not always the case in real-world situations. In actual chemical processes, the data can either belong to the class of trained data or to the fault of a new class, making it necessary to predict its anomality and update the model without knowing the labels. Therefore, existing studies that utilize incremental learning with a supervised approach suffer from a major drawback, which is using the label information about the online data that is unobtainable in real chemical processes.

In this study, we present a novel methodology for fault diagnosis in chemical processes using an unsupervised incremental learning approach. This approach does not require any prior knowledge about the new data and is performed in two steps. In the first step, we use out-of-distribution methods to assign labels to the data and determine whether the data belongs to the existing training data or new fault data, utilizing the statistics of the pretrained model, softmax function, and temperature scaling. If new fault data is detected, the pretrained model is updated using replay and regularization-based methods to incorporate both existing and new data. To address the imbalance between existing and new fault data, we utilize a generative model to estimate the distribution of the existing data and find a few exemplars that represent each class and are used for the model training. An attention-based structure is used for the pretrained classification model as it takes into account the time series of the data, is easy to analyze, and has high classification performance. We apply the proposed approach to the Tennessee Eastman Process in various scenarios and compared it to existing incremental learning methodologies. The results show that it is possible to update a model with accurate prediction performance for new fault classes with only a few iterations even when prior information about the new fault data is unavailable. As such, we proposed an online fault diagnosis methodology suitable for chemical processes.

References

[1] Narasingam, Abhinav, and Joseph Sang-Il Kwon. "Data-driven identification of interpretable reduced-order models using sparse regression." Computers & Chemical Engineering 119 (2018): 101-111.

[2] Islam, MM Manjurul, and Jong-Myon Kim. "Automated bearing fault diagnosis scheme using 2D representation of wavelet packet transform and deep convolutional neural network." Computers in Industry 106 (2019): 142-153.

[3] Wang, Yalin, et al. "Deep learning for fault-relevant feature extraction and fault classification with stacked supervised auto-encoder." Journal of Process Control 92 (2020): 79-89.

[4] He, Xiao-hui, et al. "A novel bearing fault diagnosis method based on gaussian restricted boltzmann machine." Mathematical Problems in Engineering 2016 (2016).

[5] van de Ven, Gido M., Tinne Tuytelaars, and Andreas S. Tolias. "Three types of incremental learning." Nature Machine Intelligence (2022): 1-13.

[6] Yu, Wanke, and Chunhui Zhao. "Broad convolutional neural network based industrial process fault diagnosis with incremental learning capability." IEEE Transactions on Industrial Electronics 67.6 (2019): 5081-5091.

[7] Zhou, Han, et al. "Incremental Learning and Conditional Drift Adaptation for Non-Stationary Industrial Process Fault Diagnosis." IEEE Transactions on Industrial Informatics (2022).

[8] Gu, Xiaohua, et al. "An imbalance modified convolutional neural network with incremental learning for chemical fault diagnosis." IEEE Transactions on Industrial Informatics 18.6 (2021): 3630-3639.