Handheld NIR Spectral Sensor Module Based on a Fully-Integrated Detector Array

Abstract

For decades, near-infrared (NIR) spectroscopy has been a valuable tool for material analysis in a variety of applications, ranging from industrial process monitoring to quality assessment. Traditional spectrometers are typically bulky, fragile and expensive, which makes them unsuitable for portable and in-field use. Thus, there is a growing interest for miniaturized, robust and low-cost NIR sensors. In this study, we demonstrate a handheld NIR spectral sensor module, based on a fully-integrated multipixel detector array, sensitive in the 850-1700 nm wavelength range. Differently from a spectrometer, the spectral sensor measures a limited number of NIR spectral bands. The capabilities of the spectral sensor module were evaluated alongside a commercially available portable spectrometer for two application cases: to quantify the moisture content in rice grains and to classify plastic types. Both devices achieved the two sensing tasks with comparable performance. Moisture quantification was achieved with a root mean square error (RMSE) prediction of 1.4% and 1.1% by the spectral sensor and spectrometer, respectively. Classification of the plastic type was achieved with a prediction accuracy on unknown samples of 100% and 96.4% by the spectral sensor and spectrometer, respectively. The results from this study are promising and demonstrate the potential for the compact NIR modules to be used in a variety of NIR sensing applications.

Keywords: integrated photonics; near-infrared; portable devices; sensors; spectral sensing.

Figure A1

Photograph showing the appearance of samples measured for the plastic classification experiment. (a) type 1 PET; (b) type 2 HDPE (dark blue sample that resulted in low reflectance is circled in red); (c) type 4 LDPE; (d) type 5 PP; (e) type 6 PS; and (f) type 6 foam. The green stars indicate the samples that were used in the test set.

Figure A1

Photograph showing the appearance of samples measured for the plastic classification experiment. (a) type 1 PET; (b) type 2 HDPE (dark blue sample that resulted in low reflectance is circled in red); (c) type 4 LDPE; (d) type 5 PP; (e) type 6 PS; and (f) type 6 foam. The green stars indicate the samples that were used in the test set.

Figure A2

Representative measurements prior to normalization, of colored HDPE (type 2) samples taken by the SpectraPod (a) and NIRscan (b). The measurement with the lowest photocurrent and intensity resulted from the dark blue sample (circled in red in Figure A1b).

Figure 1

The standalone handheld SpectraPod modules (physical dimension: 8.2 × 8.2 × 3.4 cm³) and their response curves. (a) SpectraPod with cuvette holder attachment used in the moisture quantification experiment; (b) the basic SpectraPod module for acquiring reflectance measurements from plastic samples. The response curves of the SpectraPod containing chip 1 (c) and chip 2 (d).

Figure 2

The raw photocurrent and intensity measurements collected using the (a) SpectraPod and (b) NIRscan, (c) the transformation of SpectraPod measurements after normalization by the sum of all spectral points within each measurement and (d) the transformation of NIRscan measurements after applying the standard normal variate method. One measurement corresponding to the minimum, median and maximum moisture content is highlighted by the bold line in each plot and their corresponding MC is indicated in the legend.

Figure 2

The raw photocurrent and intensity measurements collected using the (a) SpectraPod and (b) NIRscan, (c) the transformation of SpectraPod measurements after normalization by the sum of all spectral points within each measurement and (d) the transformation of NIRscan measurements after applying the standard normal variate method. One measurement corresponding to the minimum, median and maximum moisture content is highlighted by the bold line in each plot and their corresponding MC is indicated in the legend.

Figure 3

The predicted moisture content of two types of rice grains, obtained via PLS regression using normalized photocurrents from the SpectraPod (a) and NIRscan (b), compared to their expected values, obtained via the oven-drying method. The number of latent variables (LVs), the root mean square error of calibration (RMSEC) and prediction (RMSEP) and the coefficient of determination for the calibration (R²C) and prediction (R²P) are indicated.

Figure 4

The average measurement of each plastic type measured by the SpectraPod (a) and NIRscan (b), following normalization by the sum of all spectral points within each measurement.

Figure 5

Results of plastic type classification using the SpectraPod (top, a–c) and NIRscan (bottom, d–f) measurements. Confusion matrices display the outcome of PLS-DA classification on the calibration (left: relative frequency) and test sample set (middle: relative frequency, right: absolute frequency). One hundred percent accuracy was obtained by the SpectraPod for both cross-validation and test set prediction. Accuracies of 96.9% and 96.4% were obtained by the NIRscan for cross-validation and test set prediction, respectively.

Figure 5

Results of plastic type classification using the SpectraPod (top, a–c) and NIRscan (bottom, d–f) measurements. Confusion matrices display the outcome of PLS-DA classification on the calibration (left: relative frequency) and test sample set (middle: relative frequency, right: absolute frequency). One hundred percent accuracy was obtained by the SpectraPod for both cross-validation and test set prediction. Accuracies of 96.9% and 96.4% were obtained by the NIRscan for cross-validation and test set prediction, respectively.