Data-based process monitoring, fault detection and diagnostics

The objective of this work is to develop a framework, along with the tools required, for the development of process monitoring solutions. In most cases, it is impractical to develop precise models from first principles for monitoring purposes as it requires consideration of not only the complex physics involved in the process but also the interactions between different components constituting the process. In these cases, soft computational methods, which can make use of process data to capture its trends and dynamics, provide an attractive alternative for the quick development and deployment of process monitoring solutions. Firstly, signal based methods based on feature-level sensor fusion are considered for monitoring purposes. The problem of optimal sensor selection is formulated as the problem of selecting optimal groups of inputs during linear or non-linear model training from data. Novel penalty terms called as Group Selection Terms (GST) are derived based on hierarchical Bayesian modeling. A generalized algorithm based on the “Bound Optimization” approach is derived for simultaneously selecting the optimal sensors, sensor-features and model parameters from data. Three specific algorithms called Linear Embedded Sensor Selection (L-ESS), Nonlinear Embedded Sensor Selection (NL-ESS) and Sparse Multiple Kernel Learning (SMKL) are derived for training models which allow the user to tradeoff the number of points they can handle versus the degree of non-linearity allowed by the model. The ability of these algorithms to learn models which use fewer groups of features while achieving a high degree of prediction accuracy is tested using real data sets. Finally, NL-ESS is tested on experimental data obtained for the purpose of monitoring burn and chatter conditions during cylindrical plunge grinding. While the performance of the regularization based methods was found to be satisfactory in terms of prediction accuracy as well as the sparsity of the obtained solution, one of its main drawbacks is the fact that it requires the tuning of one or more tradeoff parameters. In an attempt to further automate the process of sensor and sensor-feature selection, the hierarchical Bayesian formulation is extended further and two algorithms, Variational Relevant Group Selector (VRGS) and the Relevant Group Selector (RGS), are derived to estimate the parameter probability distributions directly from the data. Experimental results using data from plunge grinding monitoring and diesel engine diagnostics verify the excellent performance of the algorithms without the need for any manual parameter tuning. As a complementary approach, some of the insights gained from classical model based fault detection and isolation (FDI) methods are exploited by making use of data based dynamic process models. In this work Recurrent Neural Networks (RNN) are used to capture the dynamics of any system that can be represented in the state space form. A novel constructive procedure for training RNNs from data is proposed. A systematic method for incorporating partial state measurements into the structure of a RNN is proposed where the measured states are augmented with hidden node activations to get a complete dynamic model of the system. A robust stochastic nonlinear discrete observer called the Adaptive Divided Difference Filter (ADDF) is developed for simultaneous state and parameter estimation in uncertain systems with the goal of combining it with the RNN system model for data based FDI. All the modules of this framework are validated using simulation examples