Monitoring, fault detection and estimation in processes using multivariate statistical

Abstract

Multivariate statistical techniques are one of the most widely used approaches in data driven monitoring and fault detection schemes in industrial processes. Concretely, principal component analysis (PCA) has been applied to many complex systems with good results. Nevertheless, the PCA-based fault detection and isolation approaches present some problems in the monitoring of processes with different operating modes and in the identification of the fault root in the fault isolation phase. PCA uses historical databases to build empirical models. The models obtained are able to describe the system¿s trend. PCA models extract useful information from the historical data. This extraction is based on the calculation of the relationship between the measured variables. When a fault appears, it can change the covariance structure captured, and this situation can be detected using different control charts. Another widely used multivariate statistical technique is partial least squares regression (PLS). PLS has also been applied as a data driven fault detection and isolation method. Moreover, this type of methods has been used as estimation techniques in soft sensor design. PLS is a regression method based on principal components. The main goal of this Thesis deals with the monitoring, fault detection and isolation and estimation methods in processes based on multivariate statistical techniques such as principal component analysis and partial least squares. The main contributions of this work can be arranged in the three following topics: ¿ The first topic is related with the monitoring of continuous processes. When a process operates in several operating modes, the classical PCA approach is not the most suitable method. In this work, an approach for monitoring the whole behaviour of a process, taking into account the different operating modes and transient states, is presented. The monitoring of transient states and start-ups is studied in detail. Also, the continuous processes which do not operate in a strict steady state are monitored in a similar way to the transient states. ¿ The second topic is related with the combination of model-based structural model decomposition techniques and principal component analysis. Concretely, the possible conflicts (PCs) approach is applied. PCs compute subsystems within a system model as minimal subsets of equations with an analytical redundancy property to detect and isolate faults. The residuals obtained with this method can be useful to perform a complete fault isolation procedure. These residuals are monitored using a PCA model in order to simplify and improve the fault detection task. ¿ The third topic addresses the estimation task in soft sensor design. In this case, the soft sensors of a real process are studied and improved using neural networks and multivariate statistical techniques. In this case, a dry substance (DS) content sensor based on indirect measurements is replaced by a neural network-based sensor. This type of sensors take into account more variables of the process and obtain more robust and accurate estimations. Moreover, this sensor can be improved using a PCA layer at the network input in order to reduce the number of inputs in the network. Also, a PLS-based sensor is designed in this topic. It also improves the sensor based on indirect measurements. Finally, the different approaches developed in this work have been applied to several process plants. Concretely, a two-communicated tanks system, the evaporation section of a sugar factory and a reverse osmosis desalination plant are the systems used in this dissertation