ABSTRACT
On the basis of these samples from different cattle stables the possibility of recognition by the multisensor array was tested. The different valuation methods [5,6] were also compared based on this example. When assessing this example, it must be considered that the measurement points marked with an X come from the same beef bull stable, but the samples were taken four weeks later. These samples differ from those which form the PCA-plot, because the composition of the samples is changed. The difference is caused by factors such as meteorological fluctuations, altered feeding, and weight increase of the animals. For the unmonitored methods, only the sensor signals are relevant. These signals classify the different measurements in groups based on the measurement data. With the monitored methods, the groups must be known before valuation. During the actual valuation, the measurement values are assigned to the groups. Of the above described methods, those that were selected for data valuation are included in the FOX 4000 software package used. The represented Plot (figure 4) shows the result of principal component analysis (PCA), an unmonitored, model-based valuation method. Here, 70% of the measurements of the unknown sample (X) are assigned to the right group (beef bull stable B). Valuation with discriminating factor analysis (DFA), a monitored, model-based method, enabled 90% of the newer measurements to be matched to the corresponding measurements that were four weeks older. The highest recognition rate of
95% was achieved by a trained neural network with backpropagation architecture. The results of the comparison of the valuation methods are again summarized in table 1.
