Introduction to many prediction variables by examples PCR

 About 15 years ago the Danish Meat Research Institute started to develop the first full automatic classification equipment the Classification Centre (CC). The background for this rather ambitious project was a realization of the fact that operators influence the measurements when they insert the probe using manual probe equipment.


The solution was mechanic. The measurements were still made by insertion of probes, but now automatically. Subsequent to the classification, the carcass will automatically be branded with health certificates and brands indicating the lean meat content (or SEUROP grade).


The equipment consists of several mechanical items to fix the carcass and to measure some characteristics of the carcass to position the probes. 7 probes are inserted into the shoulder, back and ham and 7 measurements are made.

Each measurement consists of about 200 data, which describe how meat and fat reflect light along a line of about 10 cm. The information is interpreted by use of a neural network into 7 measurements of fat thickness and 3 measurements of the loin thickness.


The next step is to validate the measurements and for this purpose the close correlation between the measurements is used from experience we know that the thickness of fat vary almost in the same way along the carcass but at different levels. If one thickness deviates improbably from the other measurements, it will be replaced by a predicted thickness.



In this sketch the different steps are shown. First, the measurements of length to position the probes and after that the insertion of probes. The validation has two actions. If too many deviations appear (i.e. 3 or more), the probes will be pushed 2 cm upwards and inserted into the carcass again. This happens in about 10% of the cases. This part of the system is regarded as the equipment and the 10 results form the interface to the system we look at in our common Regulation in the EU.


Now we have 10 results plus the weight. This contains all we know about the carcass correlated to leanness. In the previous step the validation step - we utilized the correlation between the data. In the next step the prediction step - we want data with different information.


Therefore, we use a so-called principal component analysis. The analysis can easily be illustrated in two dimensions. When two variables are correlated they form an ellipse.


If we push the system of co-ordinates such that the origin is in the centre and the x-axis along the main axis of the ellipse, then we can express each point by independent variables. The new x-variable the first principal component - expresses the main part of the variation.


If we do the same in 11 dimensions, we can reduce the number of variables to three or four principal components, which contain the main part of the original variation. These few independent variables are used to calibrate the equipment by linear regression and finally, an equation for predicting the lean meat percentage can be expressed by the original measurements.


The experiences.


We have a very precise equipment and an equipment independent of the operators. However, it discloses other dependencies.

A lot of effort has been made to identify sources of error. The influence from the knife mounted at the point of the probe and the influence from the slaughter process (weak or strong surface treatment of the skin) are the most important.


A system to monitor the measurements is necessary. The system has to include all measurements and not only the meat percentage. If you have a problem, you have to identify the reason for the problem, which can be one of the probes or one of the items, which measure the length of the carcass etc.


The economy - of both the pig producers and of the slaughterhouses - is very closely related to the meat percentage. Every week the slaughterhouses compare their own average results with others and they expect that the measurements are highly reliable. Nationally, the trueness or the exact accordance with the reference seems not to be the most important issue, but a possible bias between abattoirs and the precision of the measurements is looked on as very important.