(62c) Genetic Process Visualization Using Parametric t-Sne

Distributed control systems (DCS) are essential for modern chemical plants. Plant operators use the DCS to monitor the operation in real time and keep the process within safe and productive domains of operation. Commonly, the monitoring interface in the DCS is designed for single variable analysis, which is expressed in two ways: tags with current values on the process diagram or a line chart on a time axis. Based on the thresholds of each variable, alarms notify personnel if any single variable drifts out of its safe limitations. These strategies are clear to operators and are also simple to implement, but they can only provide limited information of the process due to the handicap of single variable statistics. In such a system, the DCS is blind to any correlation information among multiple variables.

In our work, we proposed a generic process visualization method using dimensionality reduction techniques based on deep learning, which is the state-of-the-art technique that uses deep neural networks (DNN) to extract features hidden within large data sets [1]. It has already achieved notable success in many areas, including image analysis [2], natural language processing [3] and even human-level intelligence for control tasks [4]. For the objective of process monitoring and data visualization, we adopted a deep learning model, namely parametric t-SNE [5]. As a dimensionality reduction method derived from t-SNE [6], it uses a DNN to optimize the projection of the data into the latent space by minimizing the Kullback-Leibler divergence from original space. In this method, different measurements in each data record are considered as a whole, allowing correlation features to be learned parametrically by the DNN. The objective of utilizing the parametric t-SNE in process visualization is to project the high-dimensional measurement onto a 2D map, where different normal operating regions and faults can be separated. Thus, a notable change can be observed in the latent space which can provide qualitative information regarding the process operation to the plant operator.

In practice, we observed that the nonuniformed variance in each variable can create bias, causing the DNN to greedily learn the features from only the most contributing variables in the training set. In other words, variables that show less variance in the training set are muted in the model, which is misleading for visualizing unknown conditions of the process in the future. Therefore, beyond following the two-stage training procedure of the parametric t-SNE, we used combinatorial variation creation method to introduce variance equally to each variable. Process data from a pyrolysis reactor was used as a case study for the proposed method. Data from one operating cycle was used to training the DNN and the trained model was tested on another cycle. The visualization model is especially effective in collaboration with data-driven fault detection methods (e.g. PCA-T2), where it can provide more in-depth information about process operation on the 2D map. This can relieve the operators from staring at the numerical values and line charts of each variable in interest and move their work to a more intuitive interface.

Reference

[1] LeCun, Yann, Yoshua Bengio, and Geoffrey Hinton. "Deep learning." Nature 521.7553 (2015): 436-444.

[2] Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton. "Imagenet classification with deep convolutional neural networks." Advances in neural information processing systems. 2012.

[3] Graves, Alex, Abdel-rahman Mohamed, and Geoffrey Hinton. "Speech recognition with deep recurrent neural networks." Acoustics, speech and signal processing (icassp), 2013 ieee international conference on. IEEE, 2013.

[4] Mnih, Volodymyr, et al. "Human-level control through deep reinforcement learning." Nature 518.7540 (2015): 529-533.

[5] van der Maaten, Laurens. "Learning a parametric embedding by preserving local structure." RBM 500.500 (2009): 26.

[6] Maaten, Laurens van der, and Geoffrey Hinton. "Visualizing data using t-SNE." Journal of Machine Learning Research 9.Nov (2008): 2579-2605.