Journal
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS
Volume 18, Issue 12, Pages 8618-8628Publisher
IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
DOI: 10.1109/TII.2022.3150805
Keywords
Principal component analysis; Detectors; Feature extraction; Transformers; Sensitivity; Informatics; Covariance matrices; Deep learning; ensemble learning; fault detection; incipient faults; sliding window patch
Categories
Funding
- National Key Research and Development Program of China [2020YFF0304904]
- National Natural Science Foundation of China [61873143]
Ask authors/readers for more resources
In this article, a novel feature ensemble net (FENet) is proposed for detecting difficult-to-detect faults in the Tennessee Eastman process (TEP). FENet integrates features extracted by basic detectors in the input feature layer, transforms the previous feature matrix using sliding-window patches and principal component analysis (PCA) in the hidden feature transformer layers, and performs sliding technique and uses normalized singular values in the decision layer for detection. Results show that FENet effectively detects Faults 3, 9, and 15 in TEP compared to other methods.
How to detect incipient faults has been an important problem in the field of fault detection. Although many types of machine and deep learning methods have been proposed, their performance is not as good as expected. In this article, a novel feature ensemble net (FENet) was developed, particularly for faults 3, 9, and 15 in the Tennessee Eastman process (TEP), which are notoriously difficult to detect. For the input feature layer, features extracted by the basic detectors are integrated to expand the detection ability of FENet. For the hidden feature transformer layers, with sliding-window patches and principal component analysis (PCA), the previous feature matrix is transformed. The sliding-window patches can be used to generate singular values, whereas the patches in the well-known convolution technique can only be vectorized, primarily for performing PCA in PCA-based networks. This enhances the sensitivity of the FENet to incipient faults. For the output feature layer, all feature matrices in the last hidden layer are completely stacked into a large feature matrix. The sliding technique is performed at the decision layer, and a detection index is designed with normalized singular values. The superiority of FENet can be completely verified by a continuous stirred tank heater and TEP. As compared with deep PCA, PCA-based monitoring network, and typical ensemble strategies, such as averaging, voting, stacking, and Bayesian inference, FENet can effectively detect Faults 3, 9, and 15 in TEP.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available