Traditional process monitoring methods of off-line analysis involve removing a sample from the process and taking it to a centralised analytical laboratory. It takes time for the analytical result to be achieved and the result is used retrospectively to determine the yield or quality of a batch, and not to control the process. This leads to batches being produced that do not meet specifications, so may require re-working, wasting time and money. The process should be monitored to allow control of the batch to ensure it meets specifications first time, and every time. The use of at-line or on-line analysis, such as near infrared spectroscopy, provides quicker process analysis and allows the results to be used to monitor and control the process. These techniques are usually nondestructive so less waste is produced, and are safer as they can be located away from the process environment. Within the analysis of processes, sampling is a key issue. The sample must be representative of the process to ensure the analysis gives a true indication of the batch. This is a problem when the process is heterogeneous as a sample taken from one region of the process may give a different analytical result from a sample taken from another region. Guided microwave spectroscopy (GMS) has been investigated for its use as an on-line process analyser. The GMS has a sample chamber in which a process can be carried out and this whole chamber is analysed. This removes the sampling issue. This method is not well understood or used in process analysis due to the complicated MW spectra. Near infrared (NIR) spectroscopy is a tried and tested method of process analysis and many examples of applications exist of its use in industry. The spectra are easy to interpret and relate to the process. The main problem with NIR is that a probe must be used for on-line analysis. This produces sampling issues, and any process variation, such as a process upset, must be in the vicinity of the probe to be detected. In this work, a new process analysis technique, GMS, has been compared to an established technique, NIR, to determine their effectiveness within process analysis. NIR is used as a reference method for the GMS to aid interpretation of the spectra, and relate it to the process. Various processes have been investigated to determine the effectiveness of NIR and GMS to monitor them. A drying process has been monitored which has a problem of sampling due to huge cakes of several tonnes of material that are dried. The drying process was first simulated by adding solvent to a material to determine if the process can be monitored and the limits of solvent that can be detected. NIR data was collected using a diffuse reflectance probe. The spectra were found to be unrepresentative of the process as it was reliant on the solvent added being in the vicinity of the probe. GMS was used to monitor the process as it provides a representative measurement. Three different systems were analysed: the addition of water to sand, propanol to ascorbic acid and ethanol to salicylic acid. Simple partial least squares (PLS) models were built to predict the amount of solvent present in the solid sample from MW spectra. Various pre-processing techniques were examined to produce the best model. The models were built using auto-scaled followed by Box-Cox logarithmically transformed data, and allow prediction of the amount of water in sand, and the amount of propanol in ascorbic acid down to 1 % w/w with relative errors below 5%. The calibration models can predict up to 30% solvent, so the technique was shown to be very useful for monitoring the drying of a solid. The model for the addition of ethanol to salicylic acid gave relative errors of 32% so seems to be an unsuitable method. However, models built using above 2% ethanol gave relative errors of only 20/0, suggesting the MW spectra are not sensitive to levels of ethanol below this. Propanol was then removed from ascorbic acid by drying to prove that the actual drying method can be monitored. The use of principal component analysis (PCA) scores plotted against time and the residuals (process spectra minus the reference dry spectra) show that the drying process has the possibility of being monitored in a representative way using MW spectroscopy. An esterification reaction has been monitored and various aspects of this process have been investigated. Traditionally calibration models are built using reference concentration spectra. Ideally process samples should be used to build the model which means a reference method such as GC must be used to give concentration data. These methods take time to develop and within this work it was found difficult to get reproducible results. Calibration free techniques have been used to extract the concentration profiles of the reaction to allow the rate constants of the reaction to be determined. A calibration free technique has also been used to determine the endpoint of the process, and also detect process upsets. During these processes, it is desirable to be able to predict the endpoint of a reaction, instead of waiting for it to be reached, which may waste time. It is also advantageous to be able to detect process upsets to allow the batch to be corrected. Multivariate curve resolution (MCR) was used to extract the concentration profiles from the MW and NIR spectra, and these profiles used to calculate the rate constants, k of the reaction. The MW and NIR calculated k values do not agree, suggesting the two techniques do not capture the same process variation. The rate constants have also been calculated using GC measurements as a comparison. These values also do not agree with the spectroscopic methods, but it is unknown which method provides the correct determination of the rate constant. However, it has been found that the use of MW and NIR spectroscopy provides a much more reproducible method to monitor esterification reactions than GC. An adaptive algorithm called caterpillar has been used to determine the endpoint of an esterification reaction, and also to detect a variety of process upsets. This allows the reaction to be monitored to ensure it proceeds as expected without the need for building a calibration model. The endpoint was detected reproducibly for MW spectra taken for repeat reactions showing the spectra are suitable for monitoring the reaction. The same endpoint was not detected for corresponding NIR spectra, so this does not appear to be as reproducible a method. MW spectroscopy was found to detect process upsets of addition of incorrect catalyst, addition of water, addition of an interferant and incorrect changing of reactants. The NIR was found to only pick up the addition of water and incorrect charging of reactants. It has been found that the MW spectra are more sensitive to small disturbances in the process variation and it is a better technique for endpoint determination and process upset detection. The NIR spectra does not appear to be as representative of the process, possibly due to the limitations of sampling with the probe used.

This thesis focuses on the development of data-driven automated techniques to enhance performance assessment methods. These techniques include process control loop status monitoring, fault localisation in a number of interacting control loops and the detection and isolation of multiple oscillations in a multi-loop situation. Not only do they make use of controlled variables, but they also make use of controller outputs, indicator readings, set-points and controller settings. The idea behind loop status is that knowledge of the current behaviour of a loop is important when assessing MVC-based performance, because of the assumptions that are made in the assessment. Current behaviour is defined in terms of the kind of deterministic trend that is present in the loop at the time of assessment. When the status is other than steady, MVC-based approaches are inappropriate. Either the assessment must be delayed until steady conditions are attained or other methods must be applied. When the status is other than steady, knowledge of current behaviour can help identify the possible cause. One way of doing this is to derive another statistic, the overall loop performance index (OLPI), from loop status. The thesis describes a novel fault localisation technique, which analyses this statistic to find the source of a plant-wide disturbance, when a number of interacting control loops are perturbed by a single dominant disturbance/fault. Although the technique can isolate a single dominant oscillation, it is not able to isolate the sources of multiple, dominant oscillations. To do this, a novel technique is proposed that is based on the application of spectral independent component analysis (ICA). Spectral independent component analysis (spectral ICA) is based on the analysis of spectra derived via a discrete Fourier transform from time domain process data. The analysis is able to extract dominant spectrum-like independent components each of which has a narrow-bank peak that captures the behaviour of one of the oscillation sources. It is shown that the extraction of independent components with single spectral peaks can be guaranteed by an ICA algorithm that maximises the kurtosis of the independent components (ICs). This is a significant advantage over spectral principle component analysis (PCA), because multiple spectral peaks could be present in the extracted principle components (PCs), and the interpretation of detection and isolation of oscillation disturbances based on spectral PCs is not straightforward. The novel spectral ICA method is applied to a simulated data set and to real plant data obtained from an industrial chemical plant. Results demonstrate its ability to detect and isolate multiple dominant oscillations in different frequency ranges.

Partial Least Squares (PLS) has been shown to be a versatile regression technique with an increasing number of applications in the areas of process control, process monitoring and process analysis. This Thesis considers the area of nonlinear PLS; a nonlinear projection based regression technique. The nonlinearity is introduced as a univariate nonlinear function between projections, or to be more specific, linear combinations of the predictor and the response variables. As for the linear case, the method should handle multicollinearity, underdetermined and noisy systems. Although linear PLS is accepted as an empirical regression method, none of the published nonlinear PLS algorithms have achieved widespread acceptance. This is confirmed from a literature survey where few real applications of the methodology were found. This Thesis investigates two nonlinear PLS methodologies, in particular focusing on their limitations. Based on these studies, two nonlinear PLS algorithms are proposed. In the first of the two existing approaches investigated, the projections are updated by applying an optimization method to reduce the error of the nonlinear inner mapping. This ensures that the error introduced by the nonlinear inner mapping is minimized. However, the procedure is limited as a consequence of problems with the nonlinear optimisation. A new algorithm, Nested PLS (NPLS), is developed to address these issues. In particular, a separate inner PLS is used to update the projections. The NPLS algorithm is shown to outperform existing algorithms for a wide range of regression problems and has the potential to become a more widely accepted nonlinear PLS algorithm than those currently reported in the literature. In the second of the existing approaches, the projections are identified by examining each variable independently, as opposed to minimizing the error of the nonlinear inner mapping directly. Although the approach does not necessary identify the underlying functional relationship, the problems of overfitting and other problems associated with optimization are reduced. Since the underlying functional relationship may not be established accurately, the reliability of the nonlinear inner mapping will be reduced. To address this problem a new algorithm, the Reciprocal Variance PLS (RVPLS), is proposed. Compared with established methodology, RVPLS focus more on finding the underlying structure, thus reducing the difficulty of finding an appropriate inner mapping. RVPLS is shown to perform well for a number of applications, but does not have the wide-ranging performance of Nested PLS.