A Novel ML-Powered Nanomembrane Sensor for Smart Monitoring of Pollutants in Industrial Wastewater

Abstract

This study presents a comprehensive analysis aimed at validating the use of an innovative nanosensor based on graphitic nanomembranes for the smart monitoring of industrial wastewater. The validation of the potential of the nanosensor was carried out through the development of advanced analytical methodologies, a direct experimental comparison with commercially available electrode sensors commonly used for the detection of chemical species, and the evaluation of performance under conditions very similar to real-world field applications. The investigation involved a series of controlled experiments using an organic pollutant-benzoquinone-at varying concentrations. Initially, data analysis was performed using classical linear regression models, representing a conventional approach in chemical analysis. Subsequently, a more advanced methodology was implemented, incorporating machine-learning techniques to train a classifier capable of detecting the presence of pollutants in water samples. The study builds upon an experimental protocol previously developed by the authors for the nanomembranes, based on electrochemical impedance spectroscopy. The results clearly demonstrate that integrating the nanosensor with machine-learning algorithms yields significant performance. The intrinsic properties of the nanosensor make it well-suited for potential integration into field-deployable platforms, offering a real-time, cost-effective, and high-performance solution for the detection and quantification of contaminants in wastewater. These features position the nanomembrane-based sensor as a promising alternative to overcome current technological limitations in this domain.

Keywords: data analysis; environmental monitoring; industrial wastewater; machine learning; nanosensor; pollutant detection; smart monitoring.

Figure 1

Sensors used in this study: (a) commercial SPEs (image sourced from the ItalSens website [52]); and (b) nanomembranes and custom PCB interface circuit for signal acquisition.

Figure 2

SEM images of the fabricated films: (a) with the magnification of 601; and (b) magnification 45,000.

Figure 3

Experimental setup: (a) for CV, CA, and EIS on commercial electrode-based sensors; (b) for time-domain EIS on nanomembranes.

Figure 4

Preliminary results obtained for each technique at different concentrations: (a) CV; (b) CA; and (c) EIS. For CA, results are reported on a logarithmic scale for greater readability. The different colors used for the curves indicate the concentrations tested.

Figure 5

Absolute time-domain impedance response of nanomembranes at (a) 0 mM; (b) 0.1 mM; (c) 1 mM; and (d) 10 mM. The color scale indicated refers to the different analysis frequencies tested.

Figure 6

Normalized time-domain impedance response of nanomembranes at (a) 0 mM, (b) 0.1 mM, (c) 1 mM, and (d) 10 mM. The color scale indicated refers to the different analysis frequencies tested.

Figure 7

Reproducibility and limitations of the linear regression model for nanomembranes: (a) average normalized response $M_{norm}\left(1\mathrm{MHz}\right)$ at different concentrations with relative standard deviation with a confidence level of 99.7 %, used for model construction; (b) validation of the regression model with the associated uncertainty. The results highlight the strong overlap among concentration responses and the incompatibility of a linear regression framework for pollutant quantification.

Figure 8

Accuracy of the classification obtained from the conducted analyses. In particular, it reports, for each measurement technique and for each algorithm tested, the percentage of accuracy for each concentration tested and the total average accuracy. A color map has been added to make the matrix easier to understand, with green representing maximum accuracy and red representing minimum accuracy.

Figure 9

Classification accuracy using a 2-class model (detection-only). In particular, it reports, for each measurement technique and algorithm tested, the accuracy percentage for each class tested and the total average accuracy. A color map has been added to make the matrix easier to understand, with green representing maximum accuracy and red representing minimum accuracy.

Figure 10

Classification accuracy across different observation windows: (a) 4-class quantification, and (b) 2-class detection. In particular, the average accuracy percentage obtained for each training and test time is reported. A color map has been added to make the matrix easier to understand, with green representing maximum accuracy and red representing minimum accuracy.

Figure 11

Accuracy of classification with respect to different noise levels considered. Shows the confusion matrices as the noise levels increase: (a) 4-class quantification and (b) 2-class detection. Noise levels considered are [0, 5, 10, 25, 35, 50] times the standard deviation ($\sigma$). A color map has been added to make the matrix easier to understand, with green representing maximum accuracy and red representing minimum accuracy.