This instrument application was developed using a neural network as the core concentration predictor for alcohol in a process flow cell.  Spectra was taken using a StellarNet NIR model, miniature fiber optic spectrometer. Just 60 samples were cataloged with known concentrations using a 10cm path length. 10 samples were reserved for testing. The range of concentrations were evenly distributed from 1 to 76.5% per volume using grain alcohol.

A neural network was trained using BrainMaker to provide 100% accuracy with both the training and the test set. Training tolerance statistics were tightened until the 100% accuracy could not be achieved. A training run lasted for 45 minutes using a 66MHz 486 computer. All sample spectra was saved as second derivative of absorbance.

A realtime neural network predictor was developed to read the net file output from the BrainMaker desktop trainer. This part was integrated into the SpectraWiz spectrometer operating software.  The software was setup to read a second spectrometer channel which could monitor the lamp excitation to the flow cell as in a standard dual beam spectrophotometer. Using the same lamp the first channel would calculate the alcohol absorbance in a 10cm flow cell. The 2 absorbance calculations were subtracted then converted to second derivative.  This corrected for any drift in the lamp output over time.

The real-time spectral data was then presented to the neural network runtime predictor. The predicted results were then displayed as a digital number between 0 and 100%.. Predictor outputs were available many times per second. Accuracy appeared to be about 1%. No attempts were made to optimize the network, although genetig training options were available to optimize network performance. Colors did not appear to affect the readings as the selected wavelength region was in the NIR.