Hyperspectral analysis of powder mixtures

Mixing of powders and granular materials is a critical step in many different industries including the food, pharmaceutical, cosmetic, paper, plastic, and rubber industry. The selection of the optimal mixing procedure to achieve an even product quality is not straightforward. The mixing process will be affected by particle characteristics such as size, shape, density, and adhesion properties that may vary between different components in the same mix.

Good process and quality control require an assessment of both product composition and homogeneity. This is traditionally done by random sampling and lab analysis. To achieve detailed information about the homogeneity of each batch, a high sample frequency is required. By using hyperspectral imaging, sample composition and distribution can be measured directly in the batch in real-time covering a larger sample volume in a more time and cost-efficient manner. This also allows for the detection of impurities and their distribution within the powder.

In this example, we will demonstrate how hyperspectral imaging can be used to measure the purity of a powder with 4 potential contaminants. The goal is to detect the purity of the base constituent independently of the type of contaminant. The data set was additionally used to show how hyperspectral imaging can be used to investigate sample homogeneity and to reveal poor mixing quality.

Figure 1: 5 different powders were investigated. One of the powders was used as the base for the experiment and the other 4 as contaminants (C1-C4). The contaminants were mixed with the base one at a time making up sample series of 16 samples for each contaminant with the concentrations of the contaminants ranging from 0.5 to 25 %. In addition to having different chemistry, the powders also differed in particle properties like size, shape density, and adhesivity.

Figure 2: The samples were imaged in the laboratory with a HySpex Classic SWIR-384 camera (930-2500 nm) operating at a 30 cm distance. The camera has a spatial resolution of 250 µm and a spectral resolution of 5.45 nm. The samples were scanned under the camera using a translation table. Two broadband halogen lamps were used as light sources. The acquired hyperspectral image data were analyzed using the Breeze software, powered by Prediktera. Quantification models were created using Partial Least Squares (PLS) regression validated with full cross-validation. Models were built based on one contaminant at the time in addition to a global model including all samples.

The different powders all exhibit unique spectral properties in the SWIR range, even though some of the powders show similar absorption features (Figure 3). The base shows similar spectral properties as C4, and there is also a high similarity between samples C2 and C3. To achieve a good separation, a good spectral resolution is therefore required.

The spectral differences were sufficient to create quantification models with acceptable errors for all 5 powders. The model quality assessment was based on the correlation coefficient of cross-validation (Q2) and the Root Mean Square Error of Cross-Validation (RMSECV). The models handling one contaminant at a time (Table 1) performed better than the global model (Table 2). The advantage with the global model is, however, that one can predict the content of the base, and hence the purity, without knowing the nature of the contaminant.

 

Figure 3: Average SWIR reflectance spectra of the 5 pure powders.

Table 1. PLS regression models for the base powder including different subsets of samples in the modeling.

Samples included in model	Q2	RMSECV	Principal Components
C1 - series	0.96	1.49	3
C2 - series	0.91	2.06	2
C3 - series	0.92	1.97	5
C4 - series	0.98	1.18	5
All samples	0.87	2.54	7

Figure 4: Left: Spatial distribution of base in a sample containing 75% base and 25% C2. The image is created by using the global PLS model to predict the base content in each pixel. Right: False color SWIR image of a pure base with an enlarged section showing the spatial resolution of the HySpex Classic SWIR camera.