D. Trommer’s research while affiliated with Schmalkalden University of Applied Sciences and other places

For many application fields a classifier on a mobile device would be interesting. Such application could be a home-sleep-analyzer or an anaesthesiacontrol device for ambulant and stationary surgeries, for example. So the question is: Does a Support Vector Machine (SVM) approach work on a microcontroller platform with restricted resources sufficiently fast to classify complex data in online mode? This paper will describe the implementation and show that the task can be performed without any loss of exactness compared with the results received by a Personal Computer (PC).

Accurate data acquisition and preprocessing is an important part of classification process at any time, not least under low signal noise ratio. During the acquisition of biological signals an overlapping of several signals from different sources can be observed typically. In this paper a method will be shown, which includes raw data acquisition by a sensor array, single source separation by the use of a recursive Independent Component Analysis approach and controlling of a robotic system. The different hard- and software modules to solve the complete task will be presented.

The contribution of trained decision trees to the monitoring and controlling the depth of anaesthesia is evaluated. The application of an adaptive tool is generally necessary to achieve a robust detection of anaesthetic stages. The main components are neural network or fuzzy logic based classifiers which detects the stages of anaesthesia by the features extracted from one frontal EEG channel. This paper deals with the performance tests of additional tree based adaptive components which perform the nonlinear classification task. The classification performances of decision trees and fuzzy rules are compared by means of two different anaesthesia detection problems. The trained decision trees could be successfully used to support the feature pre-selection.

The robust prediction of the depth of anaesthesia which is estimated on features of the brain electrical potential should be performed by an embedded system. According to limited resources the dimensions of the feature space have to be strongly restricted. The decisions of the anaesthetist to start or to stop the operation and the conclusions for further administration of drugs should be supported by the forecasting of anaesthetic depths. This paper deals with an approach for pre-estimating feature tuples by optimized neural networks (NN). The predicted features are used to forecast classes of anaesthesia by means of two different classifiers, linear discriminant analysis (LDA) and also an optimized NN. Results show that the predictability varies between the extracted features. Despite a higher performance of NN, applying the LDA shows more robustness when predicted feature tuples are used for classification.

Kernel based classification, such as support vector machine, is often implemented and evaluated on powerful data processing systems. In this paper we deal with a new approach to process kernel based classification on processing systems with mathematical support on a lower level, in consequence of their operating system or typical application area, like PLC's. We investigate constrains of a PLC and what approximations are necessary for the implementation of feature extraction and classification. It will be shown that implementing a SVM on a PLC is possible because of their polynomial complexity.

Fuzzy rules were designed to classify four different stages of anaesthesia and the wake stage additionally. Feature sets of one frontal EEG channel measured during operations in hospitals were used to create and adapt fuzzy models. The topology and the parameters of the models were optimized by evolutionary algorithms based on three different multi criteria optimization procedures: NSGA II, SPEA II and IBEA. Typical fuzzy models were selected from the amount of trained fuzzy rules contained in the resulting Pareto set. The performance of the fuzzy models to classify and control the depth of anaesthesia was tested by independent data sets.

A swarm of robots are adapted to an unknown environment by reinforcement learning. Each of the robots is able to scan the environment and to measure the distance to the next hindrance. The information collected by the robots are transferred in a wireless mode by the serial Bluetooth protocol to the central unit. This unit performs the backward recursion of Dynamic Programming to get the optimal route from one place to another for each of the robots. The general goal consists of finding the route with the minimal number of steps from the initial to the final point of a two dimensional area. The performance of the approach can be demonstrated by both the simulation and the real robots.

The sleep stage of rapid-eye movement (REM) should be recognized in real-time to evaluate the current sleep quality. A robust recognition on the basis of only one frontal electro-encephalogram (EEG) channel is necessary to give the sleep stage researcher the opportunity to control the following sleep-stages of the patient. A new approach based on a Support Vector Machine was developed. We classify the REM-stage with different data sets. An adapted kernel with RBF (radial basis function) topology was applied. On the base of these we try to implement the algorithm on a microcontroller-platform. By this way a mobile wireless device can diagnose the actual sleep stage and the further sleep-profile can control by medical treatments.