An automated multicomponent system for measuring the content of monatomic phenolic compounds in alcohol solutions has been developed. To assure the quality control of widespread alcohol solutions, including medications and strong alcoholic beverages, a new multi-sensor automated quality monitoring system has been proposed. The system is based on a multilayer artificial neural network for the digital processing of sensor signals. Taking into account the cross-sensitivity of the sensors and the application of selective sensors reduce the error in determining the concentration of light phenolic compounds. The sensor signal processing process is complemented by parallel channels for calculating the ethyl alcohol concentration as well as the illumination and temperature values by linearly converting the output signals. Simulation models using sensors based on the electronic theory of impurity sorption on the surface of semiconductors were acquired.
The application of semiconductor sensors as working elements in multi-sensor
systems makes it possible to effectively solve typical problems of modern analytical
chemistry - to analyze the compositions of liquids and gases. Such analyzing systems,
also known as "electronic tongues" and "electronic noses," work with chemical
sensors of different operating principles and biosensors in combination with
multistep mathematical algorithms for data processing. Therefore, the development and
implementation of new analytical systems is an urgent technological challenge and is
fully in line with global trends. The practical principle of the electronic nose is best
explained through its natural counterpart. Volatile compounds reach the olfactory
epithelial tissues, where they reach the olfactory receptors. This generates a
corresponding neuron, which transmits information about this fact to the brain. In it,
the information of neurons is consolidated and structured according to existing
patterns that enable the subject to recognize a specific odor [
Preferably, there is a need to make extensive application of artificial electronic
sensor systems during manufacturing products on an industrial scale. In particular, the
control and monitoring of the processes of growing, maturing and spoiling organic
raw materials are extremely important and valuable for ensuring the safety and quality
of food products [
The presence and concentration of light volatile compounds in alcoholic beverages
are also monitored by means of multi-component semiconductor gas analyzers.
Semiconductor sensor systems provide the ability to control the concentrations of
several lightweight volatile compounds. But concentration measurement gives a
relatively high error, which limits their scope. This error is related to the
crosssensitivity of the gas analyzers. One method of reducing this error is to design and use
highly selective sensor systems. They are an important component of the concept of
"lab-on-a-chip" or "micro-total-analysis". This is the concept of making miniature
devices that enable the sequencing of chemical processes or chain chemical reactions
on a single chip with a small area. In addition, state-of-the-art lab-on-a-chip devices
provide real-time chemical or quantitative chemical analysis [
Semiconductor gas-sensitive sensors have a low response time, but they are also
sensitive to other physical parameters, including light, temperature, humidity, and
background concentrations in the air of other organic volatile compounds. The effect
of these factors has a significant effect on the measuring signal. Currently, several
technological approaches are being used to minimize the impact of these factors.
These approaches include the method of temperature compensation for gas-sensitive
sensors and the pulse mode of sensor heating. By applying these methods, the effect
of evaporation of water and ethanol on the sensitivity of the sensors was significantly
reduced [
However, even such sensors also have low selectivity and significant measurement error. The responses of each low-selective sensor are slightly dependent on the type of adsorbent. One option to increase selectivity is to increase the number of sensors and create their array. The set of multiple responses to the sensory system forms a unique imprint for each substance composition. The artificial neural network apparatus can be used to detect the composition of volatile compounds in sensor systems. A computer program can evaluate the signal pattern and can compare the flavors of different samples. The measurement results are comparative, not quantitative, and therefore presented as an "imprint".
Currently, artificial neural networks are used in many fields of science and
technology. They are powerful tools for solving complex problems whose answers are
not obvious and cannot be determined by classical mathematical algorithms. Neural
networks have been increasingly used since the introduction of the backpropagation
algorithm [
Low-accuracy electronic sensor systems were used for pattern recognition and selectivity. Artificial neural networks can also be effectively used to increase the selectivity of semiconductor sensor systems. Therefore, the creation of multicomponent gas analyzers based on neural networks that provide simultaneous quantitative analysis to determine the concentrations of components, increase selectivity and reduce sensitivity to environmental factors are and remain relevant technical problems of analytical chemistry. 2
A block diagram of a multicomponent analyzer measuring complex for volatile compounds is shown in Fig. 1.
The measuring chamber is shown in fig. 2.
Nowadays, the most promising and convenient are sensor systems based on
compact digital devices - microcontrollers equipped with built-in ADCs. The complex
consists of sensors, sensitive to light volatile compounds and sensors of
environmental parameters, microcontrollers of control and preprocessing of signals,
and also the communication microcontroller of communication. The microcontroller
(1) checks the illumination brightness of the sensor plate in an assigned intensity
range. Brightness values are obtained from the light sensor. Brightness ratios and
corresponding photoresponses of different samples are stored by the neural network.
ADC microcontrollers convert analog sensor signals into digital signals, perform
signal correction and normalization, control sensor backlight modes with optical
recreation and change of temperature measurement modes. The communication
microcontroller receives data from ADC microcontrollers and processes the data
according to algorithms implemented by the artificial neural network. Additionally,
software and hardware encryption of the accumulated data may be implemented [
When using a 4-channel barrier sensor, a bias voltage should be applied to the p-n junctions. The photocurrent is determined by the rate of superficial recombination of charge carriers. Each channel has a distinct sensitivity to analytes. The equivalent circuit of the sensor is shown in fig. 3.
Since the sensor measurement data also depend on the illumination brightness of the p-n junctions, the illumination intensity varies linearly with the DAC converter in the specified range during the measurements. There may be several photocurrent extremes in the range of illumination intensity, which could be used as additional informative parameters. 3
Experiments with a set of substances of light alcohol compounds were carried out. The evaporation of the substances was above the sensors of the electronic nose. The sensors were illuminated by the incident light of a given intensity. The results of photocurrent measurements from 4 electronic nose sensors are shown in Fig. 4. 0 100 200
300 time, ms 400 500 600 Fig. 4. Results of photocurrent measurements (sample # 1).
-0.2 -0.3 Am -0.4 ,t ren -0.5 r u tco -0.6 o hP -0.7 -0.8 -0.9
sample # 1 sample # 2 sample # 3 Channel #3 sample # 1 sample # 2 sample # 3 2000 y it s n e d t n e rru1000 c o t o h P
0 ity 3000 s n e d ten 2000 r r u c o tho 1000 P 0 itsy 2000 n e d t n e r r u c o toh 1000 P 0
The dependencies of the sensor data from different samples are shown in Fig. 4.. The normalized values of the photocurrent are shown on the X-axis.
The currents density passed into the measurement sampling is shown on the Yaxis, Fig.5. Photocurrents from different channels are shown in different colors. As can be seen in the Fig. 5., the values on the sensors from different channels overlap.
0.4 0.5 Normalized Photocurrent
During the measurements of different samples of alcoholic beverages, it turned out that the results of the sensors for the related samples have similar values. The expected correlation of the sensors with the specimens of the test substances (cognacs) was not observed. Obviously, under such conditions, it is advisable to use an artificial neural network to classify substances. The multilayered, fully connected feedforward neural network architecture has been selected. The artificial neural network receives sets of values of variables, specifically the values of equivalent concentrations of the mixture components of volatile compounds from optical and resistive sensor matrices. The normalized values of the concentrations of the components of the volatile compounds are calculated for each set of input variables. Values depend on the normalized output signals of the sensors. The input layer of the neural network consists of 8 input neurons. Input signals were normalized to 1 and reflected 4 channels of photocurrents from electronic nose sensors. The neural network contains three hidden layers, each of which has 10 neurons. The output layer has 6 neurons. The structure of the neutron network is presented in Fig. 6. This artificial network allows increasing the required number of neurons in layers.
Artificial neural network training was carried out by the method of training with the teacher. To determine the minimum of the error function, the backpropagation method in stochastic gradient descent algorithm was used. The algorithm of artificial neural network training is shown in Fig. 7. During training, the signals In, Out - are the input and output vectors of the neural network, respectively. The W array is a matrix of neural network weights.
The user interface of the sensor software with the neural network implementation is shown in Fig. 8.
Fig. 8. Program interface
The “New System” button allows the user to configure the neural network. During configuration, it is possible to specify the number of input, output neurons and the number of neurons in each of the 3 intermediate layers. The “Set I / O”, “Save I / O”, “Set Value”, and “File I / O” buttons define the input vectors of the input and output neurons (frames). The process of training the neural network begins with the “Train” button. The Empty Weights button sets the weight of the Wij training to its initial, random position. The “Initialize Weights” button changes the learning scales and exits the system from the point of local minimum error.
The values of the input and output neurons in the frame are represented at the top left in the form of a matrix of rectangles. The color of the rectangles is determined by the value of the corresponding neurons. The scale for matching (and changing) the color and meaning of the neurons are shown at the bottom right.
In Fig. 8. it is shown that the neural network has finished the training and the schedule of the error change can be observed below.
The experiment was carried on three sample substances. For each substance, 51 measurements of photocurrent were executed, and therefore 153 measurements of photocurrent from three samples were executed, in the future, these measurements are called frames. The influence of time and the influence of the measurement history from the sensors on the current measurement result were not taken into account. A sample of 99 frames was randomly selected from 153 frames, 33 frames from each sample for neural network training. The remaining 54 frames, 16 frames from each sample are left as controls to test the operation of the already trained neural network. The neural network was trained for about 200 epochs, after which the total error was less than = 0.01, which was admitted a positive result. After training, the system was tested on control frames. The test gave 98% correct results. In 2% of cases, the nonauthentic odor (with indicators of 50-92%) was the main one, while the authentic odor (with indicators of 8-50%) took second place. The result demonstrates the possibility of a system error (due to a lack of training frames) and demonstrates the need to consider not only the winners (first place) but also those source neurons that took second place.
In turn, among 98% of the correct results, 2% of cases were the authentic odor (8293%), while the non-authentic odor (7-18%) ranked second. In the rest (96%) of control shots, the authentic smell was the undisputed leader (more than 99% probability). 6
During creating automated multicomponent systems for measuring the concentration of mixtures of volatile compounds, the processes occurring on the surface of the semiconductor sensitive layer as the temperature, light, and humidity of the environment are taken into account.
The use of multicomponent low-selection sensors gives the possibility to significantly reduce measurement errors in determining the concentrations of light mixtures, including when exposed to external disturbances.
For the training of neural networks, it is necessary to create and develop specialized mathematical models that describe the transformation processes in the used sensors. With the help of a neural network, sensor parameters can be normalized and calibrated to increase selectivity and to predict complex integral features in alcohol solution samples.
When considering a component that is recognized by a neural network, not only the winner neurons from the set of source neurons but also the source neurons that received the second result should be considered.
An increase in the number of measurement samples leads to an increase in neural network complexity, as well as an increase in the number of sensors needed to accurately classify the test substances.
Many substances have a specific range of odors that can confuse the neural network and generate incorrect neuron at the output, and therefore the correct classification of substances by constituents is an important technical task.
In the multi-component analytical system implemented, optimal conditions for recognizing the concentration of monatomic phenolic compounds in alcohol solutions were determined and formulated.
The main areas of further research are the development of methods for determining the age of strong drinks, including the justification for the use of markers and their linear combinations to improve the accuracy and reliability of the result, as well as the transfer of the measuring complex to the platform of programmable digital integrated circuits FPGA.