Hybrid and machine learning based fault detection and control

The operation of industrial processes to achieve economic and safety objectives significantly depends on the use of advanced fault diagnosis and process control strategies. To design model-based fault diagnosis and control frameworks, developing sufficiently accurate models is an essential task. With the recent advance in computational power and data storage technologies, there have been growing interests in the use of machine learning techniques for process modeling and control. The performance of these techniques, however, depends on the use of a sufficient amount of high-quality data. The presence of uninformative and redundant data can hinder the model's ability to accurately predict dynamic behaviors. In this situation, the over-fitting also remains a challenging problem. As a result, it is necessary to use appropriate pre-processing tools, data mining techniques, and first-principles knowledge (if applicable) to achieve a reliable model. Motivated by the above considerations, this thesis focuses on the problem of hybrid and machine learning based modeling, fault detection, and control when collected data are not sufficient or there exist correlations between data samples. The first part of this thesis addresses the problem of system identification for heating, ventilation, and air conditioning (HVAC) systems when an insufficient amount of data is available. To this end, a hybrid machine learning based approach is developed where a pre-trained recurrent neural network (RNN) model (trained on a large amount of data from a representative zone) is leveraged to build a model for the zone in question. In the next phase, first-principles knowledge is integrated with data to develop a fault detection mechanism for HVAC systems using principal component analysis (PCA). The superior performance of the proposed approach, over the individual first-principles and data-driven based methods is shown. Finally, the problem of handling correlated data for the RNN-based model predictive control (MPC) implementations is addressed. To this end, PCA and autoencoder (AE)-based strategies are used to recognize the correlations that exist between data samples in both the input and output spaces. The constrained RNN-based MPC is then formulated by adding a PCA-based squared prediction error (SPE) constraint. PCA and AE-based optimization problems are also defined to calculate the achievable set-points. The efficacy of the proposed approaches over the nominal (standard) RNN-based MPC is demonstrated using different set-point tracking scenarios for a chemical reactor example. Thesis Doctor of Philosophy (PhD)