NIS Sorter - Initial Accuracy Testing

Jump to main NIS Sorter/Scanner page.

I have started calibrating and characterising the machine recently. The following table lists the traits I am aiming for and their current status.

These figures are still a bit rubbery as I’m still going through the process so these will be updated from time to time as new data becomes available. Stats in these tables is based on P=0.05, N=20.

Raw Measurement Accuracy

This first table shows the error for the various raw measurements the machine makes.

Parameter	Measure Unit	Measure Error (+/-) Single Nut	Measure Error (+/-) 20 Nut Batch Average*	Comment
Whole Nut - Colour	8 bit RGB	2.5 (2%)	0.5	Needs transform to standard colour scale
Whole Nut - Size Average Diameter Minimum Diameter Maximum Diameter Min Enclosing Diam	mm mm mm mm	0.20 0.50 0.50 0.30	0.05 0.15 0.15 0.10	= required slot opening for separation = required circle opening for separation
Whole Nut - Mass Whole Nut - Moisture Mass	gr gr	0.15 0.08	0.05 0.02	+/-0.1gr is possible in a half-speed high resolution mode. Experimental
Whole Nut - Calculated Metrics Volume Specific Gravity Shape Moisture Content	cm^3 gr/cm^3 mm/mm w/n	0.25 0.05 0.02 2%	0.05 0.02 0.01 0.5%	Volume / Mass Min/Max, Prolateness, Convex Deviation Experimental

• Indicative of single source batch. Depends on the in-batch variability, some scenarios are higher than others.

Classification & Severity Accuracy

This tables shows the accuracy for various high level categories of defect.

Category	% False Positives	% False Negatives	Severity Error	Comment
Whole Nut - Categories Normal Immature Sun Faded Aged/Rotten	<1%* - - <1%	<1%* - - <1%	5% - - 5%	* these can increase as other categories are added calibrating calibrating Age Faded & Rotten mix a bit
Defect - Categories				calibrating

* Perfectly 'Normal' nuts are Severity = 0

Rapid KR

The machine is also capable of doing a Rapid Kernel Recovery test, where either wet or dry nuts are cracked and placed in the machine, and KR is calculated a few seconds later. This can be done as additional function of the sorter with minor modifications, or a compact single purpose version can be built for in-field use.

This table shows accuracy for five NIS traits that the Rapid KR mode can measure.

Trait	Single Nut Accuracy	In-Tree Variability	20 Nut Batch Accuracy*	100 Nut Batch Accuracy*
Nut Diameter (mm)	0.6	3.5	0.8	0.4
Nut Mass (gr)	0.5	3.0	0.7	0.3
Kernel Diameter (mm)	0.5	2.7	0.6	0.3
Kernel Mass (gr)	0.2	1.0	0.2	0.1
Kernel Recovery	1.5%	4.4%	1.0%	0.5%

* There is substantial in-tree variability for these traits and it is the dominant source of error in the batch values. It can be ameliorated by increasing sample size. It is worth noting that these are roughly the same values you would expect from the standard 48hr test because it is also affected by the in-tree variability.

Visual Separation of Varieties

Finally the machine can also separate between different varieties of nuts based on the NIS appearance.

This table shows the separation accuracy for four common varieties.

Mixing From\To	A4	A16	A38	H246	Overall Error
A4	-	< 1%	< 1%	< 1%	1%
A16	< 1%	-	2%	11%	10%
A38	< 1%	2%	-	5%	7%
H246	< 1%	14%	4%	-	11%

This is very much a work in progress at this stage, I will be looking at a larger group of varieties in the near future.