(339s) A Self-Labeled Deep Learning Method for Semi-Supervised Chemical Process Fault Diagnosis

The rapid development of modern chemical engineering industry brings benefits but meanwhile increases risk and accidents with more complex and large-scale process. Fortunately, progress has been made in process monitoring among which fault diagnosis is a vital step.

Fault diagnosis researches with supervised learning algorithms such as Support Vector Machine, Fisher Discriminant Analysis and Deep Belief Networks have reached high accuracy, nevertheless, the requirement of adequate labeled data for applying these algorithms can hardly be satisfied in real-world situations. When facing with limited amounts of labeled data and huge amounts of unlabeled data, semi-supervised methods show potentials in making use of both external and internal information of data.

In this work, a self-labeled technique integrated with deep learning algorithms is proposed for semi-supervised fault diagnosis of chemical process. Self-labeled techniques follow an iterative procedure aiming at enlarging labeled training dataset based on the hypothesis that their own predictions tend to be correct. The performance of self-labeled techniques is highly affected by the base classifier, which is a Deep Convolutional Neural Network in this paper. In addition, the data editing mechanism is introduced to remove erroneously predicted samples from the training dataset, and multiple classifiers are adopted to raise the confidence of predictions. These improvements contribute to better diagnosis result. A numerical instance and the benchmark Tennessee Eastman process are utilized to illustrate the effectiveness of the proposed semi-supervised method.