Spectroscopy, in general, is a technique suitable for determining the chemical composition and certain physical properties of the inspected body, based on the analysis of the interaction of optical radiation with the molecular and atomic structures of that body since such interaction at each wavelength of the spectrum is usually substance-specific.
In all cases, the practical use of spectroscopy involves developing machine learning models that convert spectral information into traditional parameter values, whether quantitative or qualitative.
Depending on the specific type of interaction of the radiation with the substance, as well as the optical range used, we can speak of different spectroscopies: fluorescence, Raman, emission, absorption… as well as UV, VIS, NIR, SWIR, MWIR, LWIR, THz… ranges.
Absorption NIR spectroscopy – using optical radiation in the 700 to 2500 nm range – is the most popular because it optimally combines information on the chemical composition for a wide variety of materials, technological maturity – translatable into practical reliability – and prices compatible with affordable paybacks.
Spectroscopy -and especially NIR spectroscopy– is becoming a versatile and robust tool in what is known as PAT (process analytical technologies), that is, the control of production processes and the quality of the finished product based on knowing, in real time and online, the composition and physical properties of what is intended to be produced and not only having information on process conditions.
For this reason, its applicability ranges from the exact determination of water content to the determination of the final mixing point, not to mention the degree of curing (polymerization), the identification of materials for separation, the validation of raw materials or even the detection of anomalies due to accidents or fraudulent practices.
The obvious advantage is the time needed to get the result and therefore to make a decision regarding the ongoing process. With a spectroscopic probe the result is available in fractions of a second, while laboratory ones can take from hours to days.
Also, as the determination is fully automatic, distorting factors in traditional analyses inherent to human intervention, both in sampling and analysis, are eliminated.
In general, NIR spectroscopy is not suitable for the determination of concentrations below 0.1% by weight, with 0.5% by weight being a fairly frequent threshold in applications of this technology when used in-line.
Raman spectroscopy uses a different mechanism than absorption: the sample is irradiated with a very intense light source (laser) and the radiation generated by the interaction of the laser with the substance is captured. The Raman signal is usually very weak, but, on the other hand, the Raman spectrum of each substance has a much more defined or characteristic profile than its absorption profile. As a result, it is an ideal means of identifying substances and quantifying their content when the contents are relatively low, as it is more specific than absorption spectroscopy (NIR, for example).
Like NIR spectroscopy, Raman probes are ideal media for use as PAT tools. Unlike NIR spectroscopy, where water can be a considerable interfering factor, for the Raman signal water is not a problem. Both spectroscopies can be complementary in terms of possible applications: given a certain practical case, the results achievable by both technologies have to be compared in order to select the most suitable one, since the factors to be considered are not only metrological.
Artificial vision based on deep learning (neural networks) mimics, in a way, the way the human brain processes visual information, so, as with the brain, it is essential that there is a learning process first. To this end, a set of images sufficiently representative of the cases to be detected must be available to “train” the neural network, i.e., the network must be able to “interpret” or “recognize” the characteristics of interest. Once the constructed network has been validated, it is used to inspect each new image, the output of the network being a label that identifies the image as belonging to one of the pre-established classes. In practice, a neural network is a set of numerical values, collected in a digital file, specific to a particular task, so that dedicated software takes care of “applying” such a network to each image in order to obtain the label corresponding to the class to which the image appears to belong.
Traditional computer vision algorithms are deterministic. This means that, by means of mathematical functions, either generic or specific, each image is analyzed looking for the features of interest. This approach limits its application to very specific cases, in which the features are very marked or evident with respect to the rest of the image elements, and always requires the intervention of human experts. In the real world, circumstances favorable to the use of deterministic algorithms are not at all frequent. On the contrary, there are many interfering elements that greatly affect the accuracy of the result. Deep learning, on the other hand, does not start from more or less realistic assumptions, but, in the learning phase, it automatically adapts to the circumstances, i.e. it is less sensitive to unforeseen variations, and this can be achieved without the need for theoretical assumptions or the work of an expert.
The most frequent application is the detection of morphological defects in order to activate a rejection system that discards those products labeled by the neural network as defective. For example, detection of foreign bodies or defective units in food, intolerable chromatic differences in parts, textiles or food products or the objective quantification of faults or defects in order to calculate a fair price for the various degrees of quality of a product.