(584d) Transfer Learning Method for Chemical Plant Fault Diagnosis

Modern chemical and petrochemical engineering plants are working within complex and large variety of operating and performance conditions. In particular, many flammable and explosive chemicals are involved in these complex reactions and/or are stored in plant areas and harsh operational/working conditions might be required, such as high temperature, high pressure, acid/alkaline environment and etc [1]. Hence, any component singularity or failure may lead to severe catastrophe, property damage, and environmental contamination. On the other hand, multi-reactions and multi-equipment cascading systems often cause the uncertainty, nonlinearity and multi-working modes (changeable working steady states) of chemical plants, which poses a great challenge of plant condition monitoring and fault detection, alarming and diagnosis.

Hence, considering the complex and varying chemical process, well-trained traditional fault detection and diagnosis models for one working condition may tend to lapse under another working condition. Given that, more effective and efficient online monitoring and fault diagnosis techniques need to be proposed and applied. This work introduces state-of-the-art transfer learning (TL) methods into the area of chemical plant condition monitoring and fault intelligent diagnosis.

Compared to traditional machine learning (ML) methods, transfer learning theories specialize in cross-domain learning, when the testing data are following different data distributions [2-3], which provides the natural benefits for varying condition fault diagnosis. In particular, stemming from well-formed transfer component analysis algorithm [3], more representative and adaptive cross-domain features will be extracted, selected and fused for singularity detection and fault recognition by combining classical classifiers such as support vector machine (SVM), extreme learning machine (ELM) application, etc [4-5].

In order to eliminate differences brought by changeability of working conditions, several similarity evaluation criteria such as Euclidean distance and Kullback-Leibler divergence (KL) are utilized for addressing the nonlinearities and uncertainties masked in the monitoring data and bridging the data distribution gaps of difference working conditions. Finally, Tennessee Eastman process (TEP) dataset is utilized for fault detection and diagnosis to test the performance of the proposed method. In particular, reactor cooling water inlet temperature change and valve sticking faults are detected and diagnosed by the proposed enhanced transfer component analysis technique, and other conventional feature selection and fusion methods such as principle component analysis (PCA) and kernel principle component analysis (KPCA) are used as baseline methods to verify the superior performance of the presented approach.

[1] Frank P. M. 1990. Fault diagnosis in dynamic systems using analytical and knowledge-based redundancy: A survey and some new results. Automatica, 26(3), 459-474.

[2] Pan S. J., Tsang I. W., Kwok J. T. and et al. 2011. Domain adaptation via transfer component analysis, Neural Networks IEEE Transactions, 22(2):199-210.

[3] Pan S. J., Yang Q. 2010. A survey on transfer learning, IEEE Transactions on Knowledge and Data Engineering, 22(10): 1345-1359.

[4] Wang J., Xie J., Zhao R. and et al. 2016. A New Probabilistic Kernel Factor Analysis for Multisensory Data Fusion: Application to Tool Condition Monitoring. 65(1): 2527-2537.

[5] Xie J., Zhang L., Duan L. and et al. 2016. On cross-domain feature fusion in gearbox fault diagnosis under various operating conditions based on Transfer Component Analysis. 2016 IEEE International Conference on Prognostics and Health Management (ICPHM). 1-6.