Search

A chemical detection system made of a gas sensor array and algorithms intended to monitor human activity was tested in a NASA spacecraft cabin simulator. Such a chemical-based monitoring system, if extended to home settings, would allow the autonomous detection of emergency situations, thereby postponing the moving of elderly people to assisted living facilities and improving their quality of life. Moreover, in contrast to other monitoring systems based on wearable sensors or video cameras, a monitoring system based on measuring changes in air composition induced by human activities would be non-invasive and would not raise privacy concerns when installed in homes. The third generation of the JPL sensor array was adapted in a small, compact and portable system and deployed in a spacecraft-like room for four weeks while volunteers were performing daily routines. The system was able to predict the total number of people and the level of activity performed in the room, while detecting unexpectedly high concentrations of volatiles.

We investigate optimal experiment selection to train a classifier on gas sensor arrays to get the maximal possible performance in a limited number of experiments. In gas sensing, while collecting data for a particular sensor array, one has to choose what gas and concentration level is going to be presented in the next experiment. It is an active decision by the operator selecting the experiments and training the classifiers. Can the algorithm be trained sooner rather than later? Can we minimize the costs of collecting the data in terms of the man-hour of the operator and the expenses of the experiment itself? Active control sampling provides a way to deal with the challenge of minimizing the calibration costs and is applicable to any situation where experimental selection is parameterized by an external control variable. Our results indicate that active sampling strategies can only improve a random selection of experiments over a wide range of concentration of gasses. However, random or uninformed selection is fairly close. Additionally, our active sampling methodology reveals that, when there is no prior knowledge about the range of concentrations to which the sensor will be exposed during real operation, sensor must be calibrated over the entire working range, not just high concentrations. In fact, our results show that it is especially important to include low concentrations in the calibration since the lack of these values would dramatically decrease the performance of the system.

This is the author’s version of a work that was accepted for publication in Pattern Recognition. Changes resulting from the publishing process, such as peer review, editing, corrections, structural formatting, and other quality control mechanisms may not be reflected in this document. Changes may have been made to this work since it was submitted for publication. A definitive version was subsequently published in Pattern Recognition, Vol. 45, Iss. 12, (2012) DOI: 10.1016/j.patcog.2012.06.002, The increasing size and dimensionality of real-world datasets make it necessary to design efficient algorithms not only in the training process but also in the prediction phase. In applications such as credit card fraud detection, the classifier needs to predict an event in 10 ms at most. In these environments the speed of the prediction constraints heavily outweighs the training costs. We propose a new classification method, called a Hierarchical Linear Support Vector Machine (H-LSVM), based on the construction of an oblique decision tree in which the node split is obtained as a Linear Support Vector Machine. Although other methods have been proposed to break the data space down in subregions to speed up Support Vector Machines, the H-LSVM algorithm represents a very simple and efficient model in training but mainly in prediction for large-scale datasets. Only a few hyperplanes need to be evaluated in the prediction step, no kernel computation is required and the tree structure makes parallelization possible. In experiments with medium and large datasets, the H-LSVM reduces the prediction cost considerably while achieving classification results closer to the non-linear SVM than that of the linear case., The authors would like to thank the anonymous reviewers for their comments that help improve the manuscript. I.R.-L. is supported by an FPU Grant from Universidad Autónoma de Madrid, and partially supported by the Universidad Autónoma de Madrid-IIC Chair and TIN2010-21575-C02-01. R.H. acknowledges partial support by ONRN00014-07-1-0741, USARIEM-W81XWH-10-C-0040 (ELINTRIX) and JPL-2012-1455933.