In this work various aspects of data preprocessing are discussed. Preprocessing of data from multivariate sensor data systems is often necessary to extract relevant information or remove disturbances. Depending on the sensor type different preprocessing techniques have to be used. A number of important problems face the user e.g. drifting sensor data which results in very short-lived calibration models, and complex models with a very high number of variables. A systematic approach to handle and analyse data is necessary. In this work different preprocessing techniques are elaborated to reduce these problems.

Drift is a gradual change in any quantitative characteristic that is supposed to remain constant. Thus, a drifting chemical sensor does not give exactly the same response even if it is exposed to exactly the same environment for a long time. Drift is a common problem for all chemical sensors, and thus needs to be considered as soon as measurements are made for a long period of time. Drift reduction methods try to compensate for the changes in sensor performance using mathematical models and thus maintaining the identification capability of the chemical sensor. The problem with drifting sensor data and thereby short-lived calibration models is overcome using reference samples and smart algorithms utilizing the relation between the reference measurements and the measurements from the samples. This has been studied in two of the papers in this thesis. Two new approaches have been developed and tested using data from real measurements from the electronic nose and tongue.

In industry and science more and more variables are used to describe the process under study which lead to complex models and calculations. In order to increase the interpretability of the measurements, decrease the calculation demand on the computer, and/or to reduce noise an alternative, more compact, representation of the measurement can be made which describes the important features of the measurement well but with a much smaller vector. This data reduction is the topic of three of the papers in this thesis. Two different methods have been used: A double exponential model has been developed to approximate electronic tongue data. The parameters from this model describe the signal and are used as inputs to multivariate models. Secondly, the more general approach wavelet compression with different strategies for selection of wavelet coefficients has also been studied for variable reduction of both electronic tongue data and X-ray powder diffraction data.

This thesis consists of three parts. The first part describes some new algorithms that we have invented for use in the field of sensor technology. Sensor technology is evolving rapidly and new sensors such as electronic noses and tongues have emerged on the market. Most of these sensors are non-specific and needs to be trained before real life usage. We have developed algorithms to ease the training of these sensors. The first algorithm is a superior algorithm (called ODP) for supervised feature extraction. This algorithm outperforms PCA and LDA. It consists of powerful pre-processing - to avoid statistical problems - as well as a method for minimizing the classification error in the variable reduced space. ODP is protected by a patent application. The second new algorithm (called GBP) is used for variable reduction when data consists of sensory panel judgments of samples as either "good" or "bad". GBP is protected by a patent application. The third algorithm part consists of some algorithms for measuring the information content in multi-cluster data thereby facilitating objective statements about the performance of sensors and algorithms.

In the second part of the thesis we try to model the response of the electronic tongue through a first and a second order model. The second order model has five parameters and shows very good fit with experimental data. It may in the future help to compress tongue data as well as doing noise rejection.

The third part consists of one old work that I did on hybrid control systems and is poorly related the rest of the work in this thesis.