Prediction Engine

The Prediction Engine is a PAT tool that enables (real-time) prediction and classification of hyperspectral data.

The Prediction Engine supports hyperspectral cameras as a tool of PAT (process analytical technology). Spectroscopy related experts are enabled to benefit from spatially resolved predictions – and this without the need of changing their familiar chemometric analysis software!

Chemometric modelling is supported by recognized analysis software like “The Unscrambler” from Camo. Once configured, a Prediction Engine performs chemometric predictions on hyperspectral data – even in industrial real-time.

Like all other products of Perception Park also the Prediction Engine is based on the Perception Core technology. As a consequence, instrument standardization, hyperspectral pre-processing, feature extraction & operation, post-processing of prediction images, etc. are supported for at-line or in-line purposes.

Together with the PC software suite Perception Studio, also users from Science & Education are enabled to investigate Hyperspectral data for off-line purposes.

Hyperspectral data processing – Functional Modules:

The following illustration allows an overview based on an abstraction through functional modules.

Instrument Abstraction

Dependent on the acquisition technology, data are provided by the instrument (camera) in different formats. This functional module makes an instrument compatible by abstraction of its acquisition technology.

Instrument Standardization

Interferences caused by the acquisition technology are suppressed and hyperspectral data are corrected. These disorders are depending on the chosen instrument hardware. Such disorders could be:

read-out noise
dark current noise
defect pixels
wavelength calibration
...

Disturbing Influence Suppression

Interferences like the non-uniformity of illumination, which are caused by the measuring setup, are suppressed and data are corrected with regards to application relevant needs.

Hyperspectral Pre-processing

Application of typical (scientifically and industrially established) pre-processing methods to hyperspectral data like filtering, derivative, normalization, etc.

Hyperspectral Feature Extraction and Operation

By Hyperspectral features, information hidden in a spectral curve is described by a single value per pixel. This process leads to dimensional reduction e.g. from a 3 dimensional hyperspectral cube to a 2 dimensional feature image.

Advantages:

The description of information by feature values is more specific and more understandable compared to investigation into a spectral curve.
Feature operation often leads to the calculation of new (dependent) features as a function of base features. Often the result is a decision or a class-information like: “spectra are oversteered” or “a spectrum is similar to a referenced material”.