1 Introduction

Detecting multiple analytes is very important for the smooth operation of plants and taking preventive action towards improving the efficiency of plants [1]. Many industrial processes, viz., production of drugs, agriculture, medical, military, chemicals, food products, beverages, etc., involve using several reactants as the input for production [2,3,4,5]. In these industries, the quality of the product is evaluated by its purity and the presence of by-products in it. The progress of the process is also monitored at various critical points in these industries by an offline or online collection of samples as required [6,7,8]. When the species involved in the evaluation are gaseous in nature, either mass spectrometry or gas chromatography is employed, which measures the species offline and is bulky [9, 10]. The methods specified as examples provide a qualitative and quantitative account of the species evaluated [11, 12]. On most occasions, combinations of techniques or methodologies are resorted in providing accurate and reliable values, enabling the operators to continue the process in an uninterrupted fashion [13, 14]. An alternate technique that provides both a qualitative and quantitative analysis of analytes in real-time is the electronic nose, which is based on the human nose methodology [15,16,17,18,19,20]. Designing an electronic nose using a sensor array with differentially selective sensors has the advantage of using responses from all the sensors over a single sensor that adds new dimensions to the observation, which may help to estimate multiple analytes more accurately [21,22,23,24,25,26,27]. Data from the sensor array help distinguish different analytes and measure their concentration using pattern recognition techniques [28,29,30,31,32]. The change in electrical resistance/conductance of semiconducting metal oxides (SMOs) upon exposure to different gases is employed for chemical sensing [33,34,35]. The variations in operating temperature, nature of dopant, and processing methodologies will significantly vary the selectivity toward different analytes due to varying sensing characteristics [36,37,38]. Significant contributions toward the development of an electronic nose based on SMOs are available in various literatures for particular applications in laboratory level [39,40,41]. Depending on the application concerned, the hardware/software of the e-nose alters and calls for investigation, study, design, development, testing/validation, training, and deployment [42, 43]. Such an approach requires the unique design and development of interface electronic modules for data acquisition systems, data banks, data processing, and comparison algorithms for qualitative and quantitative estimation of the species. There are various methods that measure the responses from e-nose that, include signal conditioning and data processing using a multiplexer (MUX), an analog-to-digital converter (ADC), and an 8051 processor and pattern recognition technique for qualitative and quantitative analysis of analytes [44,45,46,47]. Application-oriented e-noses for perfume odour classification, onion quality evolution, low-cost indoor air quality monitoring, determination of environmental odours, detection E-coli in drinking water, etc., were reported [48,49,50,51,52,53,54].

In the nuclear industry, different gaseous by-products are generated during the operation of different nuclear/allied facilities like spent fuel reprocessing plants, waste management plants, heavy water production plants, etc. [55, 56]. A currently off-line analytical technique is being employed for their detection and is required for the investigation of detection and quantification of multiple analytes in real time.

This paper describes an investigation, design, and development of an electronic nose setup for the discrimination and quantification of analytes of interest for nuclear allied areas in an environment of high temperature, and radioactivity. A chemiresisitive sensor array of four SMOs were processed, and developed and corresponding responses vary dynamically with the change of concentration of analytes. The sensor array was loaded in an in-house designed chamber. Hardware was designed and developed to measure the dynamic sensor array responses using signal conditioning and data processing units by exciting all the sensors with low voltages to improve the life of the sensor and sends data to PC through local area network (LAN). Data acquisition software was developed for logging, monitoring, and online/offline visualization of the four channels of data through LAN and had various graphical features to visualize and study the behavior of sensors towards analytes. The developed hardware and software were tested and validated with multiple analytes by operating the sensor array at different temperatures. Different characteristics from sensor array response were evaluated and analyzed by principal component analysis (PCA) and principal component regression (PCR) algorithm that were developed in the software. Different properties of sensor response were studied, and investigated the best feature for the training of the instrument for the qualitative and quantitative analysis of multiple analytes of hydrogen (A1), formaldehyde (A2), and hydrazine (A3). The instrument was taught from the experimental data using pattern recognition algorithms and was able to discriminate and quantify the multiple analytes in t real time.

2 Overall Scheme of Electronic Nose

The scheme (Fig. 1) for an electronic nose consists of a sensor array assembly with a programmable power supply, signal conditioning and data processing unit, ethernet interface, data acquisition, and pattern recognition algorithms for the qualitative and quantification of analytes.

Fig. 1
figure 1

Scheme depicting the overall block diagram of an electronic nose for the analysis of multiple analytes

Full size image

The sensor array consists of four sensors from S1 to S4 with different SMOs. The multichannel hardware measures the dynamic response from four channels. The qualitative and quantitative analyses of multiple analytes were executed by applying a pattern recognition algorithm on sensor array data.

3 Sensor Array and Housing

3.1 Sensing Materials Deployed in Current Study

The SMO gas sensor works on the principle of chemoresistance. Four different SMO sensors, viz., indium oxide (S1), tin oxide (S2), zinc oxide (S3), and chromium niobate (S4), were prepared and processed. The sensors of different materials were designed and developed for the sensing of A1, A2, and A3 analytes. Among them, CrNbO4 is a p-type SMO, while others are n-type. The resistance of n-type material decreases in the presence of a reducing analyte while, it increases for a p-type material.

3.2 Design and Development of Semiconductor Metal Oxide Sensors

A screen printing technique was employed to develop the sensor for the current investigation. In this technique, the sensing material, and heater were prepared, processed and printed on the sensor base in the specific pattern using print screen mesh followed by heat treatment [57,58,59,60,61,62]. The details of in-house development and fabrication of semiconductor metal oxide sensors of different materials using screen printing are described in Sect. 1 of supplementary material.

3.3 Design and Development of Sensor Array Housing

The sensors’ housing was designed and developed to accommodate four sensors. The details of mechanical drawing (SF2.1) are discussed in Sect. 2 of the supplementary material. The thick films of S1, S2, S3 and S4 were loaded in the housing. The typical sensor array housing for the four sensors of array is presented in Fig. 2.

Fig. 2
figure 2

Photograph of sensor array housing of four different sensors

Full size image

All four sensors' heaters were connected to the multi-channel programmable DC power supply and maintained at optimized temperatures. The gaseous analytes were injected through a leak-tight septum, and the instrument developed in-house measured corresponding sensor array responses based on the constant voltage method. The different features (sensitivity, differential signal rise, and recovery times) were derived from the responses and considered for analysis.

3.4 Methodology for Assessing the Performance of the Sensor

Different concentrations of analytes A1, A2, and A3 ranging from 10 to 500 ppm were considered for the current study. The analytes were injected by a Hamilton microliter gas tight syringe through a septum. After attaining a steady response due to analytes injection, the housing valve was left open for natural air diffusion to replenish fresh ambiance in the chamber without any forced carrier flow. The sensor response changed from the baseline when the analyte was injected and retraced back to the baseline when the chamber outlet was open to air.

4 Design and Development of Hardware for Measurement of Sensor Array Responses

The hardware was designed and developed to measure the dynamic responses from the four sensors. The in-house developed hardware provides a low excitation voltage of 100 to 300 mV across the sensor, which improves the sensor life. This hardware also offers high accuracy, stability, input impedance, reliability, etc. Typically, the range of sensor response is 10 to 10000 kΩ towards analyte concentrations from 10 to 500 ppm. As the response of the sensor changes with the concentration of the analytes, the hardware was designed for broad dynamic characteristics of sensors towards analytes and low excitation voltage across the sensors so that the life of the sensor will not degrade. The developed hardware excites all the sensors of the array and measures the responses of the sensor using 8051 microcontroller-based signal conditioning and data processing unit. The signal conditioning unit processes all the sensors response and converts them into engineering units using algebraic equations by calibration. The hardware sends the array data through LAN, where data can be acquired and stored in a file for post-analysis of sensors towards analytes. The hardware was simulated and optimized using multisim software.

The circuit diagram of 4-channel hardware based on constant voltage and corresponding photograph is presented in Fig. 3.

Fig. 3
figure 3

Circuit diagram and photograph of multichannel hardware developed

Full size image

5 Software Scheme

5.1 Data Acquisition Software for the Sensor Array of Four Sensors

Data acquisition software is used to store the sensor data in a file for post-analysis [63,64,65]. In the current study, the protocol of the in-house developed hardware is different and is required to process and acquire the sensor array data of four sensors into the PC remotely in a file through LAN. Typically, the user can program the number of average data points, delay time between the samplings, file name to save the data, etc. The quad-channel data acquisition software is presented in Fig. 4. The software configures and initiates the LAN communication with the hardware, and the acquired data was saved in a user-chosen file. The data received was processed and converted into the response of the sensor. The software was developed to communicate with the hardware using suitable commands, processes using suitable string conversion functions, and error handling for an error-free and generous program flow. The error handling was executed into the main program in such a way that any error would not lead to a crash or hanging of the program. The program displays the sensors’ online time-series data. The program was tested and validated with standard inputs. The acquired sensor array data saved in the file can be directly imported into the text file for further processing. The software also provides the off-line graph as the popup window, where the user can select the desired channel to view a large amount of data.

Fig. 4
figure 4

Quad-channel data acquisition software for the sensor array of four sensors

Full size image

5.2 Development of Pattern Recognition Algorithm

Earlier, authors implemented PCA to recognize analytes from responses from a 3-sensor array using an astable multivibrator based instrument [66]. In the current study, the pattern recognition algorithm was extended to process the data from a 4-sensor array along with the development of PCR for quantitative analysis. The details of the development of PCA and PCR algorithms in the software for the sensor array of four sensors are provided in Sect. 3 of the supplementary material.

6 Gas Sensing Studies with 4-Sensor Array Towards Multiple Analytes

Typically, semiconducting metal oxides chemisorb oxygen at high temperature that reacts with the analyte gases and alters the charge carrier density, manifesting as the electrical resistance changes.

The various experiments were conducted with different concentrations of analytes A1, A2, and A3 at different temperatures, and measured corresponding dynamic responses. The sensor response was measured as a function of the temperature in range of 548 to 598 K with an interval of 25 K. The optimum temperature required for the operation of sensor towards the highest sensitivity was found to be 325 ºC. A typical variation of sensitivity with temperature for the sensor was presented in Fig. 5. All sensors were operated at the same temperature to minimize the process variable-dependent relation among the responses.

Fig. 5
figure 5

The variation in sensitivity of the sensor at different temperatures

Full size image

The sensor array with sensors S1 to S4 was evaluated in the air at an operating temperature of 325 °C. Initially, the baseline stability was recorded continuously, and a typical recording for about eight hours is presented in Fig. 6, with corresponding means and standard deviations shown in the inset. The testing of the sensor array with analytes A1 to A3 in the concentration range from 10 to 500 ppm individually was followed by measuring the corresponding responses using developed instrumentation. All the analyte injections were repeated thrice for concurrence, and the average value was considered for analysis.

Fig. 6
figure 6

Typical baseline stability of S1, S2, S3 and S4 sensors with mean and standard deviation

Full size image

The rise time is described as the time interval over which response of the sensor towards gas reaches a fixed percentage (usually 90%) of the final value when the sensor is exposed to the gas at a full scale concentration. Recovery time is the time interval over which the sensor response towards gas reduces to 10% of the saturation value when the sensor is exposed to full scale concentration of the gas and then placed in clean air ambient.

The typical responses of S3, S1, and S4 toward different concentrations of A1, A2, and A3, respectively, are shown in Fig. 7(a–c).

Fig. 7
figure 7

a Typical response of S3 towards A1. b Typical responses of S4 towards A2. c Typical responses of S1 towards A3

Full size image

The variations in different characteristics, viz., sensitivity, differential signal, rise time, and recovery time derived from the responses of the sensors are detailed in Figs. SF4.1, SF4.2, and SF4.3 of Sect. 4 of the supplementary material.

The developed sensor array will be used in chemical handling of reprocessing facilities of nuclear reactor, where the gas sampling interval will be about 30 min to qualify the air ambience. The recovery time of 5–7 min will not pose any problem of interference between the successive samples for the intended application in real-time monitoring.

7 Application of Pattern Recognition Algorithms

7.1 Qualitative Discrimination of Analytes

Among different characteristics derived from the sensor array responses, the sensitivity of four sensors from S1 to S4 towards multiple analytes (A1 to A3) with different concentrations is presented in Fig. 8.

Fig. 8
figure 8

Sensitivity of array of four sensors of S1, S2, S3 and S4 sensor towards A1, A2 and A3 measured from the developed instrument

Full size image

In the current study, the features matrices consisting of 33 observations (analytes with varying concentrations) with four variables (sensors) from the experiments were analyzed using the PCA module described in Sect. 5. The PCA was carried out on different properties of sensor response for the qualitative analysis of analytes, and the corresponding all Eigen values and principal components were computed and saved in the database for training the instrument.

The selectivity of the sensors towards different analytes summarizing the output from Fig. 8 was presented in Table 1. The details of sensitivity vs. gas concentration are described in the supplementary material of Sect. 4.

Table 1 The selectivity of the sensors at 325 °C operating towards different analytes (✔ indicates higher selectivity)
Full size table

A typical scree plot to screen the number of principal components (PCs) for sensitivity data (Fig. SF5.1, Sect. 5 of supplementary materials) indicates that the 1st and 2nd PCs sufficiently show the maximum variation in data, and the corresponding score plot is shown in Fig. 9.

Fig. 9
figure 9

Scores plot towards sensitivity for the discrimination of A1, A2 and A3

Full size image

Similar analysis with other feature data matrices, viz., differential signal, rise time, and recovery time, resulted in corresponding patterns in 2D-PC space (Figs. SF5.5, SF5.6, and SF5.7 in Sect. 5 of the supplementary material). The percentages of corresponding PCs were tabulated accordingly in Table ST1 of supplementary material. Among the feature matrices considered, the sensitivity matrix could discriminate all three analytes without overlapping with 95.1% (1st PC 87.6%, 2nd PC 7.5%) of variation that discriminates A1, A2, and A3 in real time (Fig. 9).

7.2 Quantitative Analysis of Analytes

7.2.1 Data Training

From the sensitivity data matrix constructed for PCA, 70% of samples were considered for training the instrument, and the remaining 30% was taken for the validation of the PCR model.

The PCA scores output of training data was fetched as input for PCR, which gives more than 95% discrimination of analytes. The regression coefficient matrix (β) is related to the concentration matrix (C) through Eq. (1) and was computed. The estimated β values were stored in the database. The typical β values for different analytes measured from an in-house developed instrument are presented in Table 2.

$${\upbeta } = \left( {{\text{Score}}^{\text{T}} \cdot {\text{Score}}} \right)^{ - 1} \cdot \left( {{\text{Score}}^{\text{T}} \cdot {\text{C}}} \right)$$
(1)
Table 2 Typical computed β constant for estimation of quantification of analytes
Full size table

For the quantification of analytes, the concentration of unknown can be computed using Eq. (2).

$${\text{C}} = {\upbeta }_0 + {\upbeta }_1 {\text{PC}}_1 + {\upbeta }_2 {\text{PC}}_2$$
(2)

7.2.2 Validation of Electronic Nose

The performance of the electronic nose setup was evaluated with the remaining 30% sample data that corresponds to the 6 × 2 data matrix. A good linear correlation between the measured and the standard concentrations of hydrogen (Fig. 10) supports the validation of the PCR model in quantifying the analytes with 4-sensor array setup.

Fig. 10
figure 10

Typical graph between measured and standard concentration of hydrogen

Full size image

The linear fit details of all analytes A1, A2, and A3 between standard and measured concentrations are presented in Table 3. The developed setup is able to quantify the analytes with the accuracies mentioned in Table 2.

Table 3 PCR fit parameters obtained for different analytes using instrument developed
Full size table

For the study of the sensor's life, the individual sensor’s materials performances were studied in detail for more than three months each. In the current study, the new sensor films of In2O3, SnO2, ZnO and CrNbO4 compositions were developed and studied for qualitative (PCA) and quantification analysis (PCR) of multiple analytes towards the nuclear reactor/allied facilities.

8 Conclusions

An electronic nose setup was investigated, designed and developed for the qualitative and quantitative analysis of A1 (H2), A2 (HCHO) and A3 (NH2NH2) using a sensor array of four sensors S1 (In2O3), S2 (SnO2), S3 (ZnO) and S4 (CrNbO4) and pattern recognition algorithm for the nuclear reactor/allied facilities. The sensors of different materials were prepared, processed, and developed to sense the multiple analytes generated in nuclear reactors. The instrumentation was developed in such way that it provides a low excitation voltage across the sensors to improve life of the sensor. The various experiments were carried out at different temperatures with different concentrations of multiple analytes. The PCA (principal component analysis) and PCR (principal component regression) of the pattern recognition algorithm was developed in the software. The algorithms were executed on different sensing features of sensor response (viz., sensitivity, differential signal, rise, and recovery time) measured from the developed instrument and investigated that sensitivity was the best among for the discrimination of analytes under study. The electronic nose was trained from the experimental data and evaluated for the quantitative analysis of these analytes using PCR. The measured and estimated values found a good correlation with a minimum regression coefficient of determination R2 = 0.997. The instrumentation setup was trained from the experimental data using a pattern recognition algorithm and is able to discriminate and quantify the multiple analytes of hydrogen (H2), formaldehyde (HCHO), and hydrazine (NH2NH2) with a maximum variance of 95.1% and an average accuracy of 98.1% respectively. This low-voltage excitation instrumentation would help extend the lifetime of the sensor, which would be undertaken as a part of the long-term performance of the sensor array on field applications. The customized instrument setup can be modified/tuned for other chemiresisitive- based sensor arrays for agriculture, medical, food, military, pharma industry applications, etc.