(432g) Break the Trade-Off Relationship between Detection and Diagnosis Performance through Explainable Deep Learning

Deep learning has recently emerged as a promising method for nonlinear process monitoring. However, ensuring operational reliability based on the decisions from the black-box models remains a challenge [1]. In this study, the process monitoring system that integrates the deep learning model and its explanation models is proposed. Breaking the traditional trade-off between accuracy and interpretability of deep learning models, the proposed method can explain why the model detects the fault both in local and global scopes. First, the state-of-the-art fault detection model is developed using an adversarial autoencoder (AAE) [2] to bridge the gap between data and process system insights. AAE model makes the latent space follow a specific distribution by adding the discriminator network and thus is advantageous for extracting enhanced representative latent space while capturing the nonlinear relationships of the process data. To explain the faults detected by this complex model, the fault diagnosis method using Shapley additive explanation (SHAP) [3] is investigated. The diagnosis method aims to provide a comprehensive explanation of the influence of input variables on process states. The combined monitoring index [4] is used for calculating the Shapley value to consider the latent space and the reconstruction space simultaneously. The whole proposed method is validated by two case studies, including the CSTR process [5] and the Tennessee Eastman process [6]. The fault detection rate and false alarm rate of the proposed method are calculated for all types of faults and compared with various methods such as principal component analysis [7], autoencoder [8], and variational autoencoder [9]. In addition, the normality test is conducted to verify the feature extraction capability of the model discussed above. To examine the fault diagnosis performance, the proposed model was tested on single and multiple faults and compared with the traditional reconstruction-based diagnosis method [10]. The contribution map based on SHAP values isolates the faulty variables well reducing the smearing effect. Fault maps are also created by clustering the SHAP values to provide insights into fault patterns. Overall, this study proposes the process monitoring system using an explainable deep learning model where the process data flow seamlessly from the detection to diagnosis to help decide on the operational strategies.

Reference.

[1] S.J. Qin, L.H. Chiang, Advances and opportunities in machine learning for process data analytics, Comput. Chem. Eng. 126 (2019) 465–473.

[2] A. Makhzani, J. Shlens, N. Jaitly, I. Goodfellow, B. Frey, Adversarial Autoencoders, ArXiv:1511.05644. (2015). http://arxiv.org/abs/1511.05644.

[3] S.M. Lundberg, S.-I. Lee, A Unified Approach to Interpreting Model Predictions, 31st Conf. Neural Inf. Process. Syst. (NIPS 2017). (2017) 552–564. https://doi.org/10.1016/j.ophtha.2018.11.016.

[4] H.H. Yue, S.J. Qin, Reconstruction-based fault identification using a combined index, Ind. Eng. Chem. Res. 40 (2001) 4403–4414.

[5] K.E.S. Pilario, Y. Cao, Canonical variate dissimilarity analysis for process incipient fault detection, IEEE Trans. Ind. Informatics. 14 (2018) 5308–5315.

[6] J.J. Downs, E.F. Vogel, A Plant-wide Industrial Problem Process, Comput. Chem. Eng. 17 (1993) 245–255. https://doi.org/10.1016/0098-1354(93)80018-I.

[7] M.E. Tipping, C.M. Bishop, Probabilistic principal component analysis, J. R. Stat. Soc. Ser. B Stat. Methodol. 61 (1999) 611–622. https://doi.org/10.1111/1467-9868.00196.

[8] A.J. Holden, D.J. Robbins, W.J. Stewart, D.R. Smith, S. Schultz, M. Wegener, S. Linden, C. Hormann, C. Enkrich, C.M. Soukoulis, D. Schurig, A.J. Taylor, C. Highstrete, M. Lee, R.D. Averitt, P. Markos, D. Mcpeake, S.A. Ramakrishna, J.B. Pendry, V.M. Shalaev, M. Maksimchuk, D. Umstadter, W. Chen, Y.R. Shen, J. V Moloney, Reducing the Dimensionality of Data with Neural Networks, Science (80-. ). 313 (2006) 504–507.

[9] D.P. Kingma, M. Welling, Auto-Encoding Variational Bayes, ArXiv:1312.6114. (2013). http://arxiv.org/abs/1312.6114.

[10] B. Mnassri, M. Ouladsine, Reconstruction-based contribution approaches for improved fault diagnosis using principal component analysis, J. Process Control. 33 (2015) 60–76.