Open AccessArticle

RANFIS-Based Sensor System with Low-Cost Multi-Sensors for Reliable Measurement of VOCs

Keunyoung Kim

and

Woosung Yang

Department of Robotics, Kwangwoon University, 20 Kwangwoon-ro, Nowon-gu, Seoul 01897, Republic of Korea

Author to whom correspondence should be addressed.

Technologies 2025, 13(3), 111; https://doi.org/10.3390/technologies13030111

Submission received: 27 December 2024 / Revised: 7 February 2025 / Accepted: 3 March 2025 / Published: 7 March 2025

(This article belongs to the Topic Internet of Things Architectures, Applications, and Strategies: Emerging Paradigms, Technologies, and Advancing AI Integration)

Download

Browse Figures

Figure 1
(a) Multi-sensor system structure; (b) a unit multi-sensor module; and (c) a multi-sensor system. "> Figure 2
Sensor module attachment position inside the chamber. "> Figure 3
Setting of VOC measurement sensor system. "> Figure 4
Comparison of normalized Sensor 1 data and reference data. "> Figure 5
ANFIS structure. "> Figure 6
RANFIS structure. "> Figure 7
Before and after outlier correction of Sensor 1 data positions 1, 4, 5, and 8. "> Figure 8
(a) Gradient compensation of Sensor 1 and (b) reconstructed data of Sensor 1. "> Figure 9
REF sensor and ANFIS and RANFIS results for Sensor 1 data. "> Figure 10
REF sensor and ANFIS and RANFIS results for Sensor 2 data. "> Figure 11
REF sensor and ANFIS and RANFIS results for Sensor 3 data. "> Figure A1
Comparison of normalized reference data and (a) Sensor 1, (b) Sensor 2, and (c) Sensor 3. "> Figure A1 Cont.
Comparison of normalized reference data and (a) Sensor 1, (b) Sensor 2, and (c) Sensor 3. "> Figure A2
Graph comparison by offset of (a) Sensor 1 (MQ135), (b) Sensor 2 (MQ138), and (c) Sensor 3 (PID-A15). "> Figure A3
Sensor 1 training error: (a) ANFIS training error for all positions, (b) ANFIS training error for sensors excluding the sensor with the lowest correlation, and (c) ANFIS training error for sensors with adjusted outliers. "> Figure A4
ANFIS results for all sensors. "> Figure A5
ANFIS results for sensors excluding the sensor with the lowest correlation. "> Figure A6
Before and after outlier correction of (a) Sensor 1 data, (b) Sensor 2 data, and (c) Sensor 3 data. "> Figure A7
ANFIS results for sensors with adjusted outliers. "> Figure A8
Gradient compensation of (a) Sensor 2 and (c) Sensor 3 and reconstructed data of (b) Sensor 2 and (d) Sensor 3. ">

Versions Notes

Abstract

This study describes a sensor system for continuous monitoring of volatile organic compounds (VOCs) emitted from small industrial facilities in urban centers, such as automobile paint facilities and printing facilities. Previously, intermittent measurements were made using expensive flame ionization detector (FID)-type instruments that were impossible to install, resulting in a lack of continuous management. This paper develops a low-cost sensor system for full-time management and consists of multi-sensor systems to increase the spatial resolution in the pipe. To improve the accuracy and reliability of this system, a new reinforced adaptive neuro fuzzy inference system (RANFIS) model with enhanced preprocessing based on the adaptive neuro fuzzy inference system (ANFIS) model is proposed. For this purpose, a smart sensor module consisting of low-cost metal oxide semiconductors (MOSs) and photo-ionization detectors (PIDs) is fabricated, and an operating controller is configured for real-time data acquisition, analysis, and evaluation. In the front part of the RANFIS, interquartile range (IQR) is used to remove outliers, and gradient analysis is used to detect and correct data with abnormal change rates to solve nonlinearities and outliers in sensor data. In the latter stage, the complex nonlinear relationship of the data was modeled using the ANFIS to reliably handle data uncertainty and noise. For practical verification, a toluene evaporation chamber with a sensor system for monitoring was built, and the results of real-time data sensing after training based on real data were compared and evaluated. As a result of applying the RANFIS model, the RMSE of the MQ135, MQ138, and PID-A15 sensors were 3.578, 11.594, and 4.837, respectively, which improved the performance by 87.1%, 25.9%, and 35.8% compared to the existing ANFIS. Therefore, the precision within 5% of the measurement results of the two experimentally verified sensors shows that the proposed RANFIS-based sensor system can be sufficiently applied in the field.

Keywords:

multi-sensor system; ANFIS; RANFIS; learning algorithm; low-cost sensing system; VOCs

1. Introduction

With the development of industry and transportation, air pollutants such as O3 and particulate matter have become a serious environmental concern [1,2]. These pollutants not only adversely affect ecosystems and the environment but also have a negative impact on human health. Volatile organic compounds (VOCs) are a group of 37 air pollutants designated by the Ministry of the Environment for control, including health hazards such as acetaldehyde and formaldehyde, as well as ozone precursors such as benzene, toluene, and xylene. VOC measurement is used in a variety of applications, including indoor air quality, fire detection, odor monitoring, and factory emissions [3]. Industrial facilities, such as automotive paint facilities, are one of the major sources of VOC emissions, and it is necessary to accurately measure VOC emissions from these facilities [4,5].

To accurately measure VOCs, flame ionization detectors (FIDs) are commonly used as recognized VOC detectors. Although FIDs provide high accuracy and a wide measurement range, due to costs, environmental maintenance, and the safety requirements of hydrogen supply for FIDs, it is difficult to install them in factories [6]. Therefore, they rely on rental or door-to-door measurements. Therefore, to solve these problems, a full-time self-measurement system utilizing low-cost gas sensors such as photo-ionization detectors (PIDs) and metal oxide semiconductors (MOSs) systems can be considered an alternative.

PID sensors are characterized by high sensitivity and relatively low cost, and MOS sensors are attracting attention for their affordability and reasonable lifetime [6,7]. However, low-cost gas sensors have limitations such as unstable measurement accuracy due to initialization, environmental factors such as temperature and humidity, and nonlinear data due to sensor aging [8,9,10]. Therefore, it is difficult to obtain reliable measurement values when using low-cost sensors.

Additionally, in automotive paint facilities, emission control systems are designed to minimize VOC emissions. However, the airflow inside an exhaust system varies depending on the installation location, shape, internal wind speed, and particle size in the vent, making it challenging for sensors to provide stable and accurate measurements [11,12,13]. Therefore, in this study, we aim to improve the accuracy of the measurement by deploying a multi-sensor system to enhance spatial resolution.

To improve this, researchers have been utilizing multi-sensor and artificial intelligence (AI) algorithms. Previous studies have constructed sensor arrays with various MOS sensors and used backpropagation neural networks (BPNN-DT) or partial least squares (PLS) models to predict VOC concentrations [14,15], but these approaches were mainly performed in a confined laboratory environment, which limits their applicability in real factory environments. Multi-linear regression (MLR) and artificial neural network (ANN) models have been used in studies to calibrate accurate data in complex environments such as factory exhaust, and ANNs in particular have the advantage of modeling nonlinear data [16,17]. However, ANN models rely on high-quality training data and can perform poorly on noisy data. Preprocessing such training data to detect and remove outliers plays an important role, and statistical analysis such as standard deviation, median absolute deviation, and interquartile range (IQR) has been widely used to determine outliers [18].

To overcome these limitations and improve the accuracy of VOC measurement, this paper proposes a multi-sensor system based on the reinforced adaptive neuro fuzzy inference system (RANFIS). The system uses low-cost VOC sensors at many positions to improve the spatial resolution in gas exhaust vents. It reliably calibrates sensor data using the RANFIS model, which combines fuzzy logic and a process similar to the feature of reinforcement learning. The model effectively handles nonlinear multi-sensor data, analyzes outliers to detect sensor anomalies, and corrects noise in the data. This improves the accuracy of VOC measurement and enables a system that can systematically manage VOC emissions in a factory around the clock. The RANFIS-based sensor system is an economical and efficient alternative to expensive equipment and is expected to contribute to real-time VOC monitoring in industrial sites.

2. Materials and Methods

2.1. Sensor Module

In this study, a unit multi-sensor module including three low-cost VOC sensors and a temperature and humidity sensor was fabricated for VOC emission measurement. Table 1 shows the low-cost VOC sensors included in the multi-sensor module and the reference (REF) sensor for precise measurements. The MOS sensor measures the change in electrical resistance in response to VOCs, and the PID sensor uses a UV light source to ionize gas molecules to measure their concentration. The sensors used to build the unit multi-sensor module are MQ135 (0–1000 ppm), MQ138 (0–500 ppm), and PID-A15 (0–4000 ppm, 100 ppb resolution), and the high-end PID meter FIX800 was used as a reference sensor for comparing measurements from low-cost sensors. In particular, as a PID sensor, it offers higher reliability and provides precise data due to its superior sensitivity and advanced filtering. A PID operates by using high-energy ultraviolet (UV) light to ionize gas molecules. This mechanism allows PID sensors to detect low concentrations of VOCs and other hazardous gases with high accuracy and fast response times.

As shown in Figure 1a, the multi-sensor module is designed to operate based on the CORTEX-M3 MCU, with the sensors on the top of the PCB and the power and communication components inside the case. The multi-sensor module is connected to the control PC via CAN communication and is configured to allow real-time data transmission and reception between multi-sensor systems. The host MCU sends CAN messages, the client MCU receives and processes the sensor data and sends it to the control PC, and the data are stored in the DB.

For this purpose, a multi-sensor system, as shown in Figure 1, for VOC measurement was constructed to obtain data and validate the proposed method in a real environment. Figure 1b shows the structure of the developed unit multi-sensor module, and Figure 1c shows the configuration of the multi-sensor system consisting of eight sensor modules installed.

2.2. Experiments

2.2.1. Setup

The testbed was constructed to allow the introduction of zero gas and VOCs into the chamber. High-purity nitrogen (N2, 99.9%) was used as the zero gas, and toluene was used as the VOC input gas. Toluene is one of the most common VOCs emitted from automotive paint facilities, along with butyl acetate and xylene [4].

Several studies that simulated workplace emission prevention facilities conducted experiments using a chamber similar to Figure 2, reflecting the decrease in wind speed and pressure at exhaust after passing through the activated carbon filter [24,25]. In this study, a similar chamber was constructed to perform preliminary experimental verification of the proposed measurement system. The chamber was designed to allow analysis by placing sensors at multiple locations. Positions 1–8 in Figure 2 show the locations of the sensor modules attached to the designed chamber.

Figure 3 shows the actual built testbed experimental environment. The gas flow control was performed using a mass flow controller (MFC) from Bronkhorst, with a maximum flow rate of 215 mL/min. The MFC precisely controls the flow rate of the incoming zero gas, allowing the air velocity and pressure inside the chamber to be adjusted. To relieve the pressure at the point where the zero gas meets the VOCs, a bypass was designed in the mixing path of the two gases with a check valve connection to prevent excessive pressure.

In the experiments, toluene in liquid form as a VOC input was placed in a 60 mL syringe and injected precisely via a syringe pump. The injected toluene was vaporized and used in the experiment before entering the chamber. A total of eight multi-sensor modules were attached to the inside of the chamber, and the multi-sensor system measured VOC concentrations at 2 s intervals. At the end of the chamber, a reference sensor called FIX800 was installed to verify and calibrate the VOC concentration. This testbed-based experimental environment was operated to precisely measure the concentration changes of VOCs and analyze the response characteristics of the sensor.

The data acquired from the sensor modules are efficiently managed through a Python-based data collection and processing program. The VOC data measured by the multi-sensor system is stored in MariaDB in real time to ensure data stability and security. The dataset was generated by setting the VOC measurement data of the sensor module as the input and the VOC data measured by the reference sensor as the answer value. The experiments were conducted twice for 8 h and 7 h, respectively, and a total of 14,631, and 12,629 data points were obtained. The first experiment was utilized to train the model based on the sensor data, while the second experiment was configured separately to evaluate the performance of the trained model.

2.2.2. Data Characteristics

A normalized comparison of the VOC measurement data from each multi-sensor module within the chamber revealed that the measurements varied depending on the sensor position attached to the chamber. As shown in Figure 4 and Table 2, and Appendix A (Figure A1 and Figure A2, Table A1 and Table A2), the graphs for Position 1 and Position 5 of Sensor 1 attached to the top of the chamber exhibited similar responses. While the correlation coefficient was lower compared to the other sensors, some minor differences were observed. The graph plots for Positions 4 and 8 attached to the right side of the chamber were similar to each other. This clearly shows the difference in sensor response depending on the attachment position in the chamber and highlights the importance of thoroughly evaluating sensor placement for accurate measurements.

The results of this analysis show that even the same type of sensor can produce different measured values depending on where it is attached, due to the spatial variability of sensor characteristics and the non-homogeneous distribution of gas concentrations in the exhaust vent. This highlights the importance of considering spatial and monitoring variability, especially in setups where adjacent sensors are not placed. Therefore, precise calibration that considers the offset and location characteristics of each sensor is necessary to improve the accuracy of the measured values. However, due to the feature of gas measurement, precise calibration of multiple sensors individually is time-consuming and not economically feasible. Even if it were possible, it is almost impossible to expect reliable and accurate measurements in in-pipe measurements, where it is difficult to guarantee that the calibrated and tuned sensor modules are identical in one or two specific cases [11,12,13]. To overcome these issues, other calibration and judgment methods, such as suitable AI algorithms, are needed.

2.3. ANFIS Model

A fuzzy inference system (FIS) is a system that processes uncertain or ambiguous data and supports decision making based on fuzzy logic. Unlike binary logic, fuzzy logic allows for intermediate states between “true” and “false,” giving a numerical representation of ambiguous human language. This allows for the analysis and control of complex nonlinear systems.

The adaptive neuro-fuzzy inference system (ANFIS), proposed by JANG, is a hybrid system that combines neural networks and fuzzy inference systems to perform data-driven learning and rule-based modeling simultaneously [26]. Figure 5 shows the structure of the ANFIS, which is suitable for solving complex nonlinear problems and is widely used in prediction, classification, control systems, etc.

2.4. RANFIS Model

ANFIS automatically models nonlinear systems and can be precisely calibrated for each sensor. However, the dataset itself contains noise and outliers, which reduces accuracy during training. To solve this problem, we propose the RANFIS model. The RANFIS model is an algorithm that adds two layers in front of Layer 1 of the ANFIS model to analyze and remove outliers and analyze gradients to compensate for data continuity. The RANFIS structure is shown in Figure 6.

When multiple inputs come in in real time for a multi-sensor system, Layer 1 uses the IQR of the data coming in at a point in time to determine outliers. Use Equations (1) and (2) to calculate the IQR of real-time data. And use Equation (3) to identify outliers. Using Equation (4), the data from a sensor identified as an outlier were replaced with the nearest valid value.

m_{i} = \frac{1}{N} \sum x_{i j}

(1)

σ_{i} = \sqrt{\frac{1}{N} \sum {(x_{i j} - m_{i})}^{2}}

(2)

O u t l i e r s r_{i} = {x_{i j} | x_{i j} > m_{i} + τ \cdot σ_{i} o r x_{i j} < m_{i} - τ \cdot σ_{i}}

(3)

O_{i}^{1} = \{\begin{matrix} x_{i j} & m_{i} - τ \cdot σ_{i} < x_{i j} < m_{i} + τ \cdot σ_{i} \\ {a r g m i n}_{x} |r_{i} - x_{\hat{i} j}| & O u t l i e r s (i \neq \hat{i}) \end{matrix}

(4)

where x is the input data, i is the sensor location, j is the amount of data in the dataset, N is the number of sensors,

m_{i}

is the meaning of the sensor data at a point in time,

σ_{i}

is the standard deviation,

r_{i}

are outliers, τ is the threshold for analyzing outliers, and

\hat{i}

is the sensor location excluding i.

Using Equations (5) and (6), Layer 2 analyzes the gradient of the incoming data at a point in time and the previous data to detect gradient analysis based outlier (GAO) of Equation (5) compared to the average of the rate of change of the multi-sensor system. Then, the GAO’s gradient values are replaced with the nearest gradient values to correct the continuity of the data. Additionally, considering the characteristic of gas flow, the reliability of incoming sensor data is evaluated as a quality score (QS) in this work. Through GAOs, the QS is calculated based on the ratio of GAOs to cumulative data.

G A O {\tilde{x}}_{i j} = \{x_{i j}| \frac{\nabla x_{i j}}{\sum \nabla x_{i j}} < 0}

(5)

O_{i}^{2} = \{\begin{matrix} {\hat{x}}_{i j} & G A O (Q S \geq R) \\ {a r g m i n}_{x} |\nabla {\tilde{x}}_{i j} - {\nabla x}_{\hat{i} j}| & G A O (Q S < R) \\ x_{i j} & O t h e r \end{matrix}

(6)

where x is the input data, i is the sensor location, j is the number of data points in the dataset,

\tilde{x}

is GAO,

{\hat{x}}_{i j}

are the data from the sensor that have the lowest QS, ∇

{\tilde{x}}_{i j}

is the gradient of GAO with the previous time point, ∇

x_{\hat{i} j}

are the gradient normal values, and

\hat{i}

is the sensor location excluding i.

3. Results

3.1. Evaluation of the ANFIS Model

The ANFIS model was trained through the datasets measured by the reference sensor and low-cost multi-sensor system. Then, the VOC concentration is estimated in terms of the trained ANFIS model with the multi-sensor system inside the chamber. Sensors 1, 2, and 3 were used as inputs for three separate datasets, and FIX800, a high-cost PID-based device, was used as the reference sensor for the datasets. In the preprocessing step, an IIR filter of order 5 and a cutoff frequency of 1 was applied to all datasets using Equation (7), normalized by min-max scaling, and used with sensor modules at different locations. For data learning, the membership function is Gaussian bell-shaped and the epoch is set to 100.

y [n] = \sum_{k = 0}^{M} b_{k} x [n - k] - \sum_{k = 1}^{N} a_{k} y [n - k]

(7)

Here, x[n] is the current input signal and y[n] is the current output signal,

a_{k}

and

b_{k}

are the feedback and feedforward coefficients, and M and N are the order of the coefficients.

First, we trained the ANFIS using sensor data from all locations, and in the second training, for each sensor, we excluded the sensor with the lowest correlation with the reference and trained the ANFIS on the remaining seven sensors. As shown in Table 2, Sensor 1 had the lowest correlation of −0.563 at Position 8. Sensor 2 had the lowest correlation of 0.604 at Position 2. Sensor 3 had the lowest value of 0.631 at Position 6. After removing the data from these locations, we trained the ANFIS model. Finally, one common method for determining outliers is a statistical method using the interquartile range (IQR) [16]. We used this method to determine the outliers in the above dataset, calibrated by replacing each outlier with the nearest normal value, and trained the ANFIS model.

The trained models were validated for performance using RMSE and MAPE as evaluation metrics using Equations (8) and (9):

R M S E (R o o t M e a n S q u a r e E r r o r) = \sqrt{\frac{1}{n} \sum {(y_{i} - \hat{y_{i}})}^{2}}

(8)

M A P E (M e a n A b s o l u t e P e r c e n t a g e E r r o r) = \frac{1}{n} \sum |\frac{y_{i} - \hat{y_{i}}}{y_{i}}| \times 100

(9)

where n is the number of data points,

y_{i}

is the raw data, and

{\hat{y}}_{i}

is the predicted data.

After training using sensor data from all locations, compared with Table 3 and Figure A4), the RMSE values are 30.41, 20.32, and 9.64 for Sensor 1, 2, and 3, respectively, with Sensor 3 performing the best. The ANFIS results excluding the sensor with the lowest correlation with the reference sensor are shown in Table 3 and Figure A5. We can see that the error is reduced compared to regular ANFIS training. These results suggest that the training dataset contains elements that interfere with learning. These elements can act as noise or outliers in the overall dataset, which not only degrades the performance of the learning model but also disrupts the continuity of the data, causing predictions to differ significantly from the correct answer. Therefore, a method for identifying and correcting outliers is needed.

In the above dataset, outliers are identified, as shown in Figure 7 and Figure A6, corrected by replacing each outlier with the nearest normal value, and the results of training the ANFIS model are shown in Table 3 and Figure A7. The RMSE of each sensor shown in Table 3 is 23.49, 12.60, and 5.70, which shows the importance of outlier correction to improve accuracy.

3.2. Evaluation of the RANFIS Model

In Section 2, datasets were generated using Sensor 1, Sensor 2, and Sensor 3 of the eight multi-sensor modules attached to the chamber as input data, and the FIX800 installed at the end of the chamber as the answer value. In the preprocessing step, all datasets were subjected to an IIR filter of order 5 and a cutoff frequency of 1 and normalized with min-max scaling to be used as input with sensor modules at different locations. As in Section 3.1., the same dataset was used for training and evaluation. RMSE and MAPE were used as evaluation metrics.

Table 4 shows the outlier rate for each sensor. Some sensors show a high outlier rate due to sensor malfunctions, resulting in low reliability.

Table 5 shows the proportion of GAO in the entire dataset. If the QS is greater than or equal to R (see Equation (6)) at a point in time as Sensor 1 data of Pos 5 in Table 5, the data are newly replaced with the sensor data (Sensor 1 data of Pos 3) that have the lowest QS. In this work, R is set at 40%. As shown in Figure 8 and Figure A8, the outliers identified through gradient analysis were replaced with the nearest valid gradient values.

According to Figure 9 and Table 6, Sensor 1 is the cheapest sensor for the MOS method, and the ANFIS model results did not track the REF well and showed offset differences, with an RMSE of 29.215 and the largest prediction error. On the other hand, the RANFIS model has an RMSE of 3.758, which is 87% more accurate than ANFIS. In Figure 10 and Table 6, the RMSE of the ANIFS model trained on Sensor 2 is 15.654, and the RANFIS model trained on Sensor 2 is 11.594, which is a 26% improvement and more accurately tracks the shape of the REF. In Figure 11 and Table 6, for Sensor 3, both ANFIS and RANFIS models performed well, with ANFIS having an RMSE of 5.854 and RANFIS having an RMSE of 4.837, respectively, and with RANFIS having a slightly lower error.

4. Discussion

In this study, we developed a smart multi-sensor system utilizing low-cost VOC sensors and applied the RANFIS model to calibration measurement data, evaluating its potential as a substitute for expensive VOC sensors. The experimental results demonstrated that the RANFIS model outperformed the ANFIS model overall. This was particularly evident when utilizing MQ135 and PID-A15 sensors, where the RMSE values decreased to 3.76 and 4.84, respectively, signifying enhanced prediction accuracy. These findings imply that the RANFIS model exhibits resilience in nonlinear systems and maintains consistent performance even in the presence of outliers.

To effectively monitor VOC emissions in industrial environments such as automotive paint facilities, it is imperative to address the challenge of measurement accuracy degradation caused by variations in airflow and particle size within exhaust systems. In addressing this challenge, this study proposes the implementation of a multi-sensor system to enhance spatial resolution and improve reliability through outlier correction. Sensor data outliers were detected using the IQR and GAO methods. Outlier data were replaced with the nearest normal values to maintain data continuity.

Furthermore, during the process of real-time data collection in the multi-sensor system, QS was employed for the quantitative evaluation of sensor reliability. The experimental results revealed that a higher GAO ratio led to decreased data quality, and the substitution of sensor data with more reliable readings when the QS exceeded a set threshold (R = 40%) proved to be effective. It is noteworthy that even when utilizing low-cost MOS sensors such as MQ135 and MQ138, a high level of prediction accuracy was guaranteed.

However, since this study employed data collected under specific experimental conditions, additional data collection and validation are required for broader industrial applications. Further research is needed to assess the impact of factors such as abrupt changes in VOC concentration and sensor aging over long-term use on model performance. Additionally, a method of optimizing a similar process with reinforcement learning designed in this work could be considered to enhance the real-time learning and adaptive capabilities of the RANFIS model.

In conclusion, the proposed smart multi-sensor system and RANFIS model demonstrated reliable concentration prediction using low-cost VOC sensors. This study suggests the feasibility of building an economical and efficient VOC monitoring system, with the potential to expand its applications across various industrial environments in the future.

5. Conclusions

In this paper, we built a smart multi-sensor system using a multi-sensor module consisting of low-cost VOC sensors and trained a RANFIS model based on the collected data to verify its potential to replace expensive VOC sensors.

5.1. Summary

To propose an alternative to the inefficient and expensive VOC sensing system, we developed a smart multi-sensor system to overcome the limitations of low-cost sensors such as PID and MOS, which have unstable measurement accuracy due to initialization errors, nonlinearity, sensor aging, etc. This system consists of multi-sensor modules, including multi-sensors deployed at multiple locations, to enhance spatial resolution and improve measurement accuracy by integrating with the RANFIS model.

The results showed that the RANFIS model outperformed the ANFIS overall, especially when trained with MQ135 and PID-A15, with an RMSE of 3.76 and 4.84, respectively. This proves that the RANFIS is robust to nonlinear systems and robust to data containing outliers and that it is a suitable model for predicting VOC concentrations. In particular, the prediction accuracy was maintained close to the correct value even with the low-cost MOS sensors MQ135 and MQ138, suggesting the possibility of an economical and efficient VOC monitoring system.

5.2. Study Limitations

However, although the proposed sensor system with the RANFIS model exhibits the potential to efficiently monitor VOC concentrations through experimental verification, this study is limited by its experimental evaluation in a controlled environment only. The current experimental setup primarily focuses on controlled conditions, which may not fully account for the complex and dynamic conditions encountered in environmental conditions.

Specifically, the study does not incorporate environmental data features such as fluctuations in temperature, humidity, and other VOCs, which can significantly influence sensor readings and VOC detection accuracy. These factors, which vary over time and geographical location, can directly or indirectly affect sensor data, causing it to exhibit unexpected behavior that is not observed under controlled conditions. In addition, natural environments often present challenges, such as exposure to direct sunlight, wind, and atmospheric contaminants, which can affect sensor performance and calibration.

In future research, we will apply the RANFIS model optimized through learning that considers various environmental variables to a paint booth workplace to verify its performance. This will allow us to assess how the RANFIS model adapts to changing environmental conditions and evaluate the field applicability of the proposed sensor monitoring system. In addition, we will consider external environmental factors to ensure that the proposed sensor monitoring system works reliably and effectively in the field.

Author Contributions

Conceptualization, K.K. and W.Y.; methodology, K.K. and W.Y.; software, K.K.; validation, K.K.; formal analysis, K.K. and W.Y.; investigation, K.K.; resources, K.K. and W.Y.; data curation, K.K.; writing—original draft preparation, K.K.; writing—review and editing, K.K. and W.Y.; visualization, K.K.; supervision, W.Y.; project administration, W.Y.; funding acquisition, W.Y. All authors have read and agreed to the published version of the manuscript.

Funding

This research was supported through the “R&D Project for Intelligent Optimum Reduction and Management of Industrial Fine Dust” funded by the Korean Ministry of Environment (MOE) (2480000134) and conducted by the Research Grant of Kwangwoon University in 2022.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

Data are contained within the article.

Conflicts of Interest

The authors declare that they have no conflicts of interest.

Appendix A

Figure A1a,b provides a comparison of the normalized graphs of Sensor 2 and Sensor 3, respectively, but the differences are not as pronounced as for Sensor 1. Figure A2 shows the measured and initial values for Sensors 1, 2, and 3, and Table A1 and Table A2 show the correlation coefficients with the reference sensor for each location and the initial values of the experiment. In particular, for Sensor 1, the correlation coefficient of the sensors attached to the top and right side is relatively low, while the correlation coefficient of the sensors attached to the bottom and left side is high. In the case of the bottom and left sensors, the graph shape follows the data of the reference sensor (FIX800) well. However, when comparing the initial values before normalization, it is confirmed that there are offset differences between each sensor value even though the graph shape is similar.

Figure A1. Comparison of normalized reference data and (a) Sensor 1, (b) Sensor 2, and (c) Sensor 3.

Figure A2. Graph comparison by offset of (a) Sensor 1 (MQ135), (b) Sensor 2 (MQ138), and (c) Sensor 3 (PID-A15).

Table A1. Sensor 2 (MQ138) and REF sensor correlation coefficient and initial value by sensor position.

Sensor Position		Correlation Coefficient	Initial Value [V]
Up	Position 1	0.84	0.82
Left	Position 2	0.60	0.69
Down	Position 3	0.83	1.43
Right	Position 4	0.82	0.73
Up	Position 5	0.68	1.11
Left	Position 6	0.87	0.82
Down	Position 7	0.68	1.24
Right	Position 8	0.64	0.94

Table A2. Sensor 3 (PID-A15) and REF sensor correlation coefficient and initial value by sensor position.

Sensor Position		Correlation Coefficient	Initial Value [V]
Up	Position 1	0.98	0.00
Left	Position 2	0.97	0.02
Down	Position 3	0.99	0.07
Right	Position 4	0.99	0.01
Up	Position 5	0.99	0.01
Left	Position 6	0.63	0.00
Down	Position 7	0.99	0.00
Right	Position 8	0.99	0.00

Figure A3 shows the training error for the above three training runs for the three sensors, and we can see that the training error decreases and converges.

Figure A3. Sensor 1 training error: (a) ANFIS training error for all positions, (b) ANFIS training error for sensors excluding the sensor with the lowest correlation, and (c) ANFIS training error for sensors with adjusted outliers.

Figure A4. ANFIS results for all sensors.

Figure A5. ANFIS results for sensors excluding the sensor with the lowest correlation.

Figure A6. Before and after outlier correction of (a) Sensor 1 data, (b) Sensor 2 data, and (c) Sensor 3 data.

Figure A7. ANFIS results for sensors with adjusted outliers.

Figure A8. Gradient compensation of (a) Sensor 2 and (c) Sensor 3 and reconstructed data of (b) Sensor 2 and (d) Sensor 3.

References

Zheng, K.; Zhao, S.; Yang, Z.; Xiong, X.; Xiang, W. Design and implementation of LPWA-based air quality monitoring system. IEEE Access 2016, 4, 3238–3245. [Google Scholar] [CrossRef]
Badjagbo, K.; Sauvé, S.; Moore, S. Real-time continuous monitoring methods for airborne VOCs. TrAC Trends Anal. Chem. 2007, 26, 931–940. [Google Scholar] [CrossRef]
Schütze, A.; Baur, T.; Leidinger, M.; Reimringer, W.; Jung, R.; Conrad, T.; Sauerwald, T. Highly sensitive and selective VOC sensor systems based on semiconductor gas sensors: How to? Environments 2017, 4, 20. [Google Scholar] [CrossRef]
Song, M.Y.; Chun, H. Species and characteristics of volatile organic compounds emitted from an auto-repair painting workshop. Sci. Rep. 2021, 11, 16586. [Google Scholar] [CrossRef] [PubMed]
Eklund, B.M.; Nelson, T.P. Evaluation of VOC emission measurement methods for paint spray booths. J. Air Waste Manag. Assoc. 1995, 45, 196–205. [Google Scholar] [CrossRef] [PubMed]
Xu, W.; Cai, Y.; Gao, S.; Hou, S.; Yang, Y.; Duan, Y.; Wu, J. New understanding of miniaturized VOCs monitoring device: PID-type sensors performance evaluations in ambient air. Sens. Actuators B Chem. 2021, 330, 129285. [Google Scholar] [CrossRef]
Khan, S.; Le Calvé, S.; Newport, D. A review of optical interferometry techniques for VOC detection. Sens. Actuators A Phys. 2020, 302, 111782. [Google Scholar] [CrossRef]
Kim, J.N.; Kim, H.J. A Chemoresistive Gas Sensor Readout Integrated Circuit with Sensor Offset Cancellation Technique. IEEE Access 2023, 11, 85405–85413. [Google Scholar] [CrossRef]
Hirobayashi, S.; Kimura, H.; Oyabu, T. Dynamic model to estimate the dependence of gas sensor characteristics on temperature and humidity in environment. Sens. Actuators B Chem. 1999, 60, 78–82. [Google Scholar] [CrossRef]
Kwon, Y.M.; Oh, B.; Purbia, R.; Chae, H.Y.; Han, G.H.; Kim, S.W.; Baik, J.M. High-performance and self-calibrating multi-gas sensor interface to trace multiple gas species with sub-ppm level. Sens. Actuators B Chem. 2023, 375, 132939. [Google Scholar] [CrossRef]
Yang, W.; Kuan, B. Experimental investigation of dilute turbulent particulate flow inside a curved 90° bend. Chem. Eng. Sci. 2006, 61, 3593–3601. [Google Scholar] [CrossRef]
Taheri, A.; Khoshnevis, A.B.; Lakzian, E. The effects of wall curvature and adverse pressure gradient on air ducts in HVAC systems using turbulent entropy generation analysis. Int. J. Refrig. 2020, 113, 21–30. [Google Scholar] [CrossRef]
Naik, S.; Bryden, I.G. Prediction of turbulent gas-solids flow in curved ducts using the Eulerian-Lagrangian method. Int. J. Numer. Methods Fluids 1999, 31, 579–600. [Google Scholar] [CrossRef]
Chen, Z.; Zheng, Y.; Chen, K.; Li, H.; Jian, J. Concentration estimator of mixed VOC gases using sensor array with neural networks and decision tree learning. IEEE Sens. J. 2017, 17, 1884–1892. [Google Scholar] [CrossRef]
Wolfrum, E.J.; Meglen, R.M.; Peterson, D.; Sluiter, J. Metal oxide sensor arrays for the detection, differentiation, and quantification of volatile organic compounds at sub-parts-per-million concentration levels. Sens. Actuators B Chem. 2006, 115, 322–329. [Google Scholar] [CrossRef]
Topalović, D.B.; Davidović, M.D.; Jovanović, M.; Bartonova, A.; Ristovski, Z.; Jovašević-Stojanović, M. In search of an optimal in-field calibration method of low-cost gas sensors for ambient air pollutants: Comparison of linear, multilinear and artificial neural network approaches. Atmos. Environ. 2019, 213, 640–658. [Google Scholar]
Spinelle, L.; Gerboles, M.; Villani, M.G.; Aleixandre, M.; Bonavitacola, F. Field calibration of a cluster of low-cost available sensors for air quality monitoring. Part A: Ozone and nitrogen dioxide. Sens. Actuators B Chem. 2015, 215, 249–257. [Google Scholar]
Yang, J.; Rahardja, S.; Fränti, P. Outlier detection: How to threshold outlier scores? In Proceedings of the International Conference on Artificial Intelligence, Information Processing and Cloud Computing, Wuhan, China, 15–17 December 2019.
Winsen Sensor—MQ135 Semiconductor Sensor for Air Quality. Available online: https://kr.winsen-sensor.com/product/mq135-semiconductor-sensor-for-air-quality/ (accessed on 27 December 2024).
Winsen Sensor—MQ138 VOC Gas Sensor. Available online: https://kr.winsen-sensor.com/product/mq138-voc-gas-sensor/ (accessed on 27 December 2024).
Alphasense—PID Sensors. Available online: https://www.alphasense.com/products/view-by-target-gas/pid-sensors (accessed on 27 December 2024).
Wandi Co., Ltd. Available online: https://wandi.co.kr/main/page.html?pid=103 (accessed on 27 December 2024).
LDAR Tools. Available online: https://ldartools.com/phx42fid/ (accessed on 27 December 2024).
Kim, K.S.; Kim, S.; Jun, T.H. Activated carbon-coated electrode and insulating partition for improved dust removal performance in electrostatic precipitators. Water Air Soil Pollut. 2015, 226, 367. [Google Scholar]
Kim, S.Y.; Yoon, Y.H.; Kim, K.S. Performance of activated carbon-impregnated cellulose filters for indoor VOCs and dust control. Int. J. Environ. Sci. Technol. 2016, 13, 2189–2198. [Google Scholar] [CrossRef]
Jang, J.S. ANFIS: An adaptive-network-based fuzzy inference system. IEEE Trans. Syst. Man Cybern. 1993, 23, 665–685. [Google Scholar] [CrossRef]

Figure 1. (a) Multi-sensor system structure; (b) a unit multi-sensor module; and (c) a multi-sensor system.

Figure 2. Sensor module attachment position inside the chamber.

Figure 3. Setting of VOC measurement sensor system.

Figure 4. Comparison of normalized Sensor 1 data and reference data.

Figure 5. ANFIS structure.

Figure 6. RANFIS structure.

Figure 7. Before and after outlier correction of Sensor 1 data positions 1, 4, 5, and 8.

Figure 8. (a) Gradient compensation of Sensor 1 and (b) reconstructed data of Sensor 1.

Figure 9. REF sensor and ANFIS and RANFIS results for Sensor 1 data.

Figure 10. REF sensor and ANFIS and RANFIS results for Sensor 2 data.

Figure 11. REF sensor and ANFIS and RANFIS results for Sensor 3 data.

Table 1. Low-cost sensors in multi-sensor module and reference sensors.

Low-Cost Sensor				REF Sensor
	Sensor 1	Sensor 2	Sensor 3	REF Sensor
Model	MQ135 [19]	MQ138 [20]	PID-A15 [21]	FIX800 [22]	phx42 [23]
Method	MOS	MOS	PID	PID	FID
Range [ppm]	0 to 1000	0 to 500	0 to 4000	0 to 1000	0 to 100,000
Average Time [s]	2	2	2	1	1
Cost [$]	2.31	34.82	564.59	3421.53	20,557.94
Other Features	Preheat time 24 h Error -	Preheat time 24 h Error -	Output $variation with temperature \pm 5$ %	Resolution 0.1 ppm $Error \pm 3$ %	Resolution 1 ppm $Error \pm 0.2$ %

Table 2. Sensor 1 (MQ135) and REF sensor correlation coefficient and initial value by sensor position.

Sensor Position		Correlation Coefficient	Initial Value [V]
Up	Position 1	0.73	0.55
Left	Position 2	0.95	0.20
Down	Position 3	0.91	0.06
Right	Position 4	−0.28	0.36
Up	Position 5	0.57	0.66
Left	Position 6	0.96	0.03
Down	Position 7	0.85	0.42
Right	Position 8	−0.56	0.16

Table 3. ANFIS results.

SENSOR	Sensor 1 (MQ135)		Sensor 2 (MQ138)		Sensor 3 (PID-A15)
SENSOR	RMSE	MAPE	RMSE	MAPE	RMSE	MAPE
All	30.41	38.83	20.32	20.25	9.64	9.47
Remove Lowest Correlation	30.35	36.05	16.39	16.25	5.87	7.49
Adjusted Outlier	23.49	24.45	12.60	15.36	5.70	4.79

Table 4. Outlier proportion of the entire dataset.

	Pos 1	Pos 2	Pos 3	Pos 4	Pos 5	Pos 6	Pos 7	Pos 8
Sensor 1	0.00	0.00	0.00	0.12	0.14	0.00	0.00	0.60
Sensor 2	0.04	0.28	0.01	0.25	0.06	0.01	0.10	0.18
Sensor 3	0.20	0.01	0.00	0.00	0.00	0.55	0.04	0.00

Table 5. Quality score according to the GAO proportion of the entire dataset.

	Pos 1	Pos 2	Pos 3	Pos 4	Pos 5	Pos 6	Pos 7	Pos 8
Sensor 1	0.29	0.04	0.01	0.23	0.41	0.09	0.10	0.35
Sensor 2	0.06	0.06	0.07	0.09	0.09	0.04	0.08	0.12
Sensor 3	0.07	0.06	0.03	0.04	0.04	0.12	0.04	0.04

Table 6. Comparison of RMSE and MAPE between ANFIS and RANFIS results.

Sensors	Sensor 1 (MQ135)		Sensor 2 (MQ138)		Sensor 3 (PID-A15)
Model	ANFIS	RANFIS	ANFIS	RANFIS	ANFIS	RANFIS
RMSE	29.22	3.76	15.65	11.59	5.85	4.84
MAPE	33.45	3.37	15.53	12.07	7.44	5.93

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

© 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).

Share and Cite

MDPI and ACS Style

Kim, K.; Yang, W. RANFIS-Based Sensor System with Low-Cost Multi-Sensors for Reliable Measurement of VOCs. Technologies 2025, 13, 111. https://doi.org/10.3390/technologies13030111

AMA Style

Kim K, Yang W. RANFIS-Based Sensor System with Low-Cost Multi-Sensors for Reliable Measurement of VOCs. Technologies. 2025; 13(3):111. https://doi.org/10.3390/technologies13030111

Chicago/Turabian Style

Kim, Keunyoung, and Woosung Yang. 2025. "RANFIS-Based Sensor System with Low-Cost Multi-Sensors for Reliable Measurement of VOCs" Technologies 13, no. 3: 111. https://doi.org/10.3390/technologies13030111

APA Style

Kim, K., & Yang, W. (2025). RANFIS-Based Sensor System with Low-Cost Multi-Sensors for Reliable Measurement of VOCs. Technologies, 13(3), 111. https://doi.org/10.3390/technologies13030111

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Menu