Real-time water quality monitoring using AI-enabled sensors: Detection of contaminants and UV disinfection analysis in smart urban water systems

2.1 Artificial intelligence-enabled sensor system design for detection of water pollution

This study introduces an innovative spectroscopic sensor system integrated with artificial intelligence (AI), designed for the detection and analysis of water pollution. The system architecture incorporates various components, each engineered for optimized data collection and processing, as visualized in Fig. 1. The principal component of the system is a circuit board containing the AS7265x sensor, which is serially connected to NODEMCU ESP8266 board via an I2C interface cable. configuration is crucial for capturing detailed spectral data from water samples. The data collected by the spectral sensor is transmitted serially to the NODEMCU ESP8266 and then wirelessly to a personal computer via the ESP8266′s Wi-Fi module. This data to transmission process is a critical step for the subsequent storage and which is fundamental to ensuring accuracy inwater quality analysis.

Close

Fig. 1

AI enabled sensor system.

Export to PPT

The components displayed in the visual can be described as follows:

Clean Water Containers (Case 0): The system begins with a container holding clean water samples for spectral measurement. This clean water state provides a reference point for the system's capability to detect pollution.

Contamination Unit (Case 1): In the second phase, contaminants are added to the clean water. This is typically done with a syringe in a precise manner and is used to test the system's ability to detect contamination.

UV Disinfection Unit (Case 2): Finally, the state of water disinfected with UV light is shown. The UV lamp disinfects the water by killing microbial contaminants and demonstrates the sensor system's ability to collect spectral data from the cleaned water.

In each case, the circuit board containing the AS7265x sensor (Durgun, 2024) detects light reflected from the water sample and transmits this data to the NODEMCU ESP8266. This data is then sent to the personal computer via the Wi-Fi module for processing and analysis. In this system, the computer fulfills the tasks of data storage as well as data processing and water quality analysis through AI software. Each of these three scenarios represents specially designed situations to evaluate various aspects of water quality and provides valuable data on how the system design can detect and analyze different states of water.

The sensor system employed in this study consists of an AS7265x sensor serially connected to a NODEMCU ESP8266 board via an I2C interface cable. The AS7265x sensor is pivotal for capturing detailed spectral data from water samples. The collected data is transmitted serially to the NODEMCU ESP8266, and then wirelessly to a personal computer via the ESP8266′s Wi-Fi module. This process is crucial for ensuring accuracy in water quality analysis. Fig. 1 illustrates the system components: Clean Water Containers (Case 0), Contamination Unit (Case 1), and UV Disinfection Unit (Case 2). The AS7265x sensor detects light reflected from the water samples and transmits this data to the NODEMCU ESP8266, which subsequently sends it to a personal computer for storage and analysis. This setup allows for comprehensive data collection and real-time monitoring of water quality.

2.1.1

2.1.1 Multispectral sensor

At the core of our design is the SPARKFUN AS7265X module, a state of the art triple spectroscopic of sensor array that integrates three distinct optical sensors: i) AS72651,ii) AS72652, and iii) AS72653. Each sensor is equipped with a set of six on board optical filters, which together provide comprehensive spectral coverage across the ultraviolet (UV), visible (Vis), and near-infrared (NIR) ranges, from 410 nm to 940 nm. This configuration allows for the precise and sensitive capture of spectral information, which is crucial for the detection and quantification of various water contaminants. As illustrated in the accompanying Fig. 2, the 18 channel spectral response of the AS7265X sensors demonstrates the normalized responsivity of each channel across the monitored wavelength spectrum. Each peak represents the responsivity of a particular channel corresponding to a specific wavelength, with peaks labeled from A to U, denoting wavelengths from 410 nm to 940 nm. The varying heights of these peaks reflect the relative sensitivity of each sensor channel at different wavelengths. This range allows for precise detection and quantification of various contaminants. The spectral data collected is then processed and analyzed to determine water quality.

Close

Fig. 2

18 channel spectral response of AS7265X sensors (Alshami et al., 2024).

Export to PPT

The spectral response graph provides an insight into the sensors’ capabilities to discrim- inate between different wavelengths, which is essential for identifying and analyzing a wide range of substances present in water samples. For instance, the peaks in the blue region (410 nm–485 nm) are critical for detecting substances that absorb or scatter light in the UV spectrum, while the red and NIR peaks (610 nm–940 nm) are indicative of the sensors’ ability to detect organic and inorganic matter that typically absorbs light in the longer wavelengths.The AS7265X sensors’ design enables the concurrent detection of multiple wavelengths, allowing for the simultaneous analysis of various chemical and biological constituents in water. This multiplexed approach not only increases the efficiency of the detection process but also significantly enhances the analytical capabilities of the system, making it a robust tool for comprehensive water quality assessment.

2.1.3

2.1.3 Machine learning models

To classify the water samples into categories of clean water, contaminated water, and UV-treated contaminated water, three machine learning models were employed: Random Forest, Support Vector Machines (SVM), and Neural Networks.

The Random Forest model used in this study consists of 100 decision trees. Each tree is trained with a random subset of the data to ensure robustness and reduce overfitting. The maximum depth of each tree is set to 10, and the criterion used for splitting nodes is 'gini impurity'. Hyperparameters were optimized using grid search cross-validation. The key hyperparameters tuned include the number of trees, maximum depth, and the criterion for node splitting. The spectral data collected was split into training (70 %) and testing (30 %) datasets. The model was trained using the training dataset, and its performance was evaluated using the testing dataset. The accuracy, precision, recall, and F1-score metrics were used to assess the model's performance.

The SVM model employed a radial basis function (RBF) kernel, which is effective in handling non-linear data. The regularization parameter (C) was set to 1.0, and the kernel coefficient (gamma) was set to 'scale'. Hyperparameters were tuned using a grid search with cross-validation to find the optimal values for C and gamma. Similar to the Random Forest model, the spectral data was split into training and testing datasets. The SVM model was trained on the training data and evaluated on the testing data using accuracy, precision, recall, and F1-score metrics.

The neural network model used in this study is a feedforward neural network with three hidden layers. The architecture includes an input layer with 18 nodes, corresponding to the 18 spectral channels, three hidden layers with 64, 32, and 16 nodes, respectively, and an output layer with 3 nodes representing the classification categories. The ReLU activation function was used for the hidden layers, and the softmax activation function was used for the output layer. The model was trained using the Adam optimizer with a learning rate of 0.001. The training process included 100 epochs, and early stopping was implemented to prevent overfitting. Hyperparameters such as the number of hidden layers, nodes per layer, and learning rate were tuned to optimize performance. The data was divided into training and testing sets as described for the other models. The neural network was trained on the training dataset and evaluated on the testing dataset using the same performance metrics: accuracy, precision, recall, and F1-score.

The combined flow diagram for the machine learning models is shown in Fig. 3. This diagram illustrates the data flow from collection to deployment for all three models (Random Forest, SVM, and Neural Networks).

Close

Fig. 3

Combined flow diagram for machine learning models.

Export to PPT

2.2 Sample preparation

Commercial drinking water samples were used for this experiment. Drinking water sam- ples containing microorganisms were also taken for this experimental analysis. The samples were taken into the sample tube and measured. In the first stage, a clean drinking water sample weighing about 100 ml was taken in a silver bowl, placed on top of the waterfilled bowl with the distance between the sensor system and the water 1 in. to obtain better observations. Then the relevant spectral readings for the clean drinking water sample are observed. Then, it was prepared by adding the dirty water sample into the clean drinking water sample. And spectral measurements were made. Then, the contaminated water sample was placed in the environment where the closed vessel type UV reactor was located. And the microorganisms in the dirty water were disinfected with UV light. The disinfected water sample was again spectral measurements.

2.3 Spectral data collection

In this segment, we delve into the methodology employed to amass spectral readings from both drinkable and contaminated water specimens. The initial step involves the preparation of the samples. Once the experimental apparatus is fully set up and connected, the spectral readings are captured using the sensor. This is achieved by running the data collection software available in the Arduino IDE library. Data transmission is facilitated by the I2C bus, which sends the information to the ESP8266 microcontroller board. Subsequently, this board forwards the spectral readings to a desktop computer using a wireless connectivity module. In total, spectral information was collected from 1285 different water specimens. The multispectral sensor apparatus operates with a time interval of 150 ms between sam- plings, translating to a frequency of 6 readings every second for all the spectral data of drinkable and tainted water samples.

2.4 Bacterial growth conditions

Escherichia coli ATCC (American Type Culture Collection) 25,922 bacteria was used. The bacteria were grown aerobically at 120 rpm in a 37  ℃ environment inside a 100 ml nutrient broth for one night. Afterwards, the bacteria grown were centrifuged 3939×xg for 10 min at a temperature of 20 ℃ and left to wait in 100 ml phosphate buffered saline solution. Finally, the initial population of 10³ CFU/ml was diluted for use in the experiment. Afterwards, 24 samples were prepared by diluting as 10⁶  CFU/ml. The same dilution coefficients were used for preparing the control groups for each experiment which were not exposed to light (Durgun and Karaman,2019).

3.1 Spectral analysis of water samples: implications for quality and safety

This study presents a spectral analysis of different water samples, comparing the characteristics of clean water (Case 0), E. coli contaminated water (Case 1), and UV-disinfected water (Case 2). The findings highlight significant indicators of water quality and safety, reflected through changes in optical properties of water.

In the clean water scenario (Case 0), the spectrum typically displayed low intensity values across all wavelengths. This is consistent with the transparency and minimal contamination of the water, showing no significant absorption or reflection characteristics. In contrast, the E. coli contaminated water sample (Case 1) displayed a notable increase in intensity values at certain wavelengths. This increase can be interpreted as E. coli absorbing or reflecting more light at these wavelengths, indicating a clear alteration in the optical properties of the water due to contamination. Particularly noteworthy are the spectral characteristics observed in the UV-disinfected water sample (Case 2). Post UV disinfection, the spectral profile reverted to a pattern similar to that of clean water (Case 0). This finding suggests that UV light effectively eradicated E. coli bacteria, restoring the water’s optical properties to a state akin to clean water. This result serves as a strong testament to the efficacy of the UV disinfection process (Fig. 4).

Close

Fig. 4

Spectral analysis of water quality.

Export to PPT

The implications of these findings suggest that spectral analysis can be a potent tool in assessing the quality and safety of water. The use of such analyses for evaluating contamination in water samples and the efficacy of disinfection processes can make significant contributions to the management and preservation of water resources.

3.2 Detailed spectral analysis: distinguishing water quality at key wavelengths

This study delves into the spectral characteristics of water samples at specific wavelengths, notably at 610 nm and 860 nm, to discern the optical variations corresponding to different states of water quality: clean water (Case 0), E. coli-contaminated water (Case 1), and UV-disinfected water (Case 2). The findings are illustrated in Fig. 5, which presents the normalized intensity values at these wavelengths for each case.

Close

Fig. 5

Normalized intensity at 610 nm and 860 nm for different water samples.

Export to PPT

Analysis at 610 nm Wavelength: The graph depicting normalized intensity at 610 nm demonstrates a significant difference between Case 0 and Case 2, underscoring the effectiveness of UV disinfection in reverting the optical properties of contaminated water to a state resembling that of clean water. Additionally, an increase in intensity is observed for Case 1, indicative of the alteration in optical properties due to E. coli contamination. This wavelength proves to be particularly insightful for observing the changes brought about by contamination and subsequent disinfection processes.

Analysis at 860 nm Wavelength: At 860 nm, the disparity in intensity between Case 0 and Case 1 is less pronounced yet discernible, suggesting that E. coli contamination exerts a noticeable effect at this wavelength as well. The results at 860 nm further reinforce the findings at 610 nm, providing a comprehensive understanding of the impact of bacterial contamination and UV treatment on water’s optical characteristics. These observations demonstrate the utility of spectral analysis in evaluating water quality and safety. By focusing on specific wavelengths where significant differences are observed, it is possible to effectively identify and quantify changes in water properties due to contamination and treatment processes. This approach offers a powerful tool for the assessment and management of water resources, ensuring their safety and suitability for various uses.

3.3 Comparative analysis of machine learning models in real-time spectral water quality assessment

This study embarked on a comparative analysis of three prominent machine learning models—Random Forest, Support Vector Machines (SVM), and Neural Networks—for the classification of water samples into three distinct categories: clean water, contaminated water, and UV-treated contaminated water. The classification was based on spectral measurements, with the primary objective of identifying the most effective model for real-time water quality assessment through spectral data analysis.

Random Forest Model: This model exhibited a remarkable accuracy of 100 %, showcasing its exceptional capability in handling the spectral data for water quality classification. Its robustness is underscored by perfect scores in precision, recall, and F1-score metrics. These results indicate the Random Forest model’s reliability and potential as an essential tool for accurate water quality assessment.

Support Vector Machines (SVM) and Neural Networks Models: Both these models also achieved an impressive accuracy of 100 %, closely paralleling the performance of the Random Forest model. Despite minor variances in individual classifications, these models were highly effective in interpreting spectral data, evidenced by their near-perfect precision and recall rates.

The outcomes of this comparative analysis underscore the potential of utilizing advanced machine learning techniques for real-time water quality monitoring. Each model tested demonstrated high accuracy and reliability, with the Random Forest model slightly leading due to its flawless classification capability. These findings are pivotal for the development of efficient, real-time water quality monitoring systems, which can be integrated into edge computing devices for on-site assessments. Future research endeavors will aim to further optimize these models for enhanced resource efficiency and processing speed, ensuring their suitability for deployment in practical water quality monitoring scenarios.

3.4 Microbial analysis of water samples: E. coli as a contaminant indicator

In this study, Escherichia coli ATCC 25922 strain served as the biological contaminant to assess water quality across different cases: Case 0 (clean water), Case 1 (contaminated water), and Case 2 (UV-treated contaminated water). The bacteria were cultured aerobically at 120 rpm in a 37 °C environment in a 100 ml nutrient broth overnight. Post cultivation, the bacteria were centrifuged at 3939×g for 10 min at 20 °C and resuspended in 100 ml phosphate buffered saline solution. The initial bacterial population was set to 10³ CFU/ml and diluted to 10⁶ CFU/ml for the experiment. Control groups for each case were prepared using the same dilution coefficients and were not exposed to light.

After 24 h, the microbial growth in Petri dishes was visually documented, providing a graphical representation of the bacterial proliferation for each case. The graphical outcomes, displayed in Fig. 6, showcase the clarity of clean water with no bacterial growth (Case 0), the high density of bacterial colonies in contaminated water (Case 1), and the substantial reduction of bacterial presence in UV-treated water (Case 2). These results visually corroborate the spectral analysis findings, where Case 1′s significant increase in intensity at certain wavelengths indicates the presence of contamination. Similarly, the reduced bacterial growth in Case 2 aligns with the spectral data suggesting effective UV disinfection. The visual evidence from the Petri dishes provides a compelling complementary tool to spectral analysis for water quality assessment.

Close

Fig. 6

Microbial cultures of water samples illustrating E. coli growth in different cases after 24 hours.

Export to PPT

The study’s microbial component enhances the understanding of water contamination and the efficacy of UV treatment, serving as a crucial step towards developing comprehensive water quality monitoring systems. In addition to spectral data, other metrics such as total dissolved solids (TDS), pH levels, and microbial contamination were also used to evaluate water quality. These metrics provide a comprehensive view of water quality changes. For instance, TDS and pH levels are critical for assessing the chemical composition and contamination levels of water. The integration of these metrics enhances the robustness of the water quality assessment.