An IoT-Enabled Real-Time Water Quality Monitoring System with On-Device McCulloch–Pitts Neural Network Classification

Water is a fundamental requirement for human and ecological life. In daily activities, water is used for drinking, cooking, bathing, washing, and in industrial and agricultural processes. However, rapid population growth and industrialization have led to declining water quality in many regions. Poor water quality can serve as a medium for the transmission of diseases, particularly those affecting the human digestive system. Therefore, water quality monitoring has become a crucial concern across multiple aspects of life, from household use to aquaculture practices [1]. Contaminated water may contain toxic compounds, leftover feed, organic materials, and pathogenic substances that can harm fish [2]. The government has established water quality standards through the Indonesian Ministry of Health Regulation No. 492/MENKES/PER/IV/2010, which includes parameters such as pH, temperature, Total Dissolved Solids (TDS), and turbidity [3]. Turbidity itself is not inherently hazardous, but it can indicate the presence of potentially harmful chemical compounds. These concerns drive researchers to develop innovative methods for reliably measuring water quality parameters, shifting from conventional methods to more efficient approaches [4]. One promising approach involves applying machine learning [5] to recognize environmental phenomena and improve water-quality assessment [6].

Several prior studies have focused on measuring basic water parameters [7]. Some have introduced monitoring systems to facilitate water quality observation through custom or existing web servers, enabling data storage and real-time monitoring via personal computers or smartphones [7, 8]. For instance, an IoT-based water monitoring system with pH, turbidity, salinity, temperature, and ultrasonic sensors [9, 10] connected to an Arduino Uno microcontroller has been developed to send data to Firebase and display real-time results in an Android application [5, 11]. However, these systems are often limited to local monitoring and lack robust remote access [9]. Other studies have developed online pH monitoring systems using the PH-4502C module and a web server, providing real-time visualization via a browser interface [12].

IoT-based systems present a modern solution for real-time water quality monitoring. Deploying sensor devices in the environment transforms it into a smart system capable of interacting with other devices via network connections [13, 14]. For example, a system using pH, temperature, and TDS sensors connected via an ESP32 microcontroller can be accessed through the Blynk application [15, 16]. Improving monitoring efficiency and preventing losses caused by fluctuations in water quality, particularly in aquaculture applications [17].

Beyond data acquisition, intelligent methods are needed for classification and decision-making regarding water quality. Artificial Neural Networks (ANNs), especially the McCulloch-Pitts model, can effectively classify water quality status. ANNs are classification algorithms that emulate the working principles of human neural networks [18]. They map input data to target outputs via neurons arranged across input, hidden, and output layers [19]. ANNs are highly effective for modeling complex, non-linear relationships [20]. The McCulloch-Pitts model represents one of the simplest ANN architectures, yet it remains effective for binary processing and basic classification tasks [21].

Intelligent computing methods, including fuzzy logic, have also been applied to water quality classification [22]. However, such methods often face limitations in field efficiency, especially in remote or hard-to-access locations, hindering real-time data acquisition. To overcome these limitations, this study proposes an IoT-based water quality detection and monitoring system using a McCulloch-Pitts Neural Network model. The system automatically collects sensor data and transmits it to a web-based server, enabling monitoring anytime, anywhere on computers or smartphones [23]. By integrating IoT technology with neural network methods, the proposed system aims to improve the efficiency of water quality classification, expand monitoring coverage, and help ensure the availability of safe water in accordance with health standards [24].

The developed water quality detection and monitoring system was tested to verify its functionality and accuracy in real conditions. System performance was evaluated based on sensor accuracy testing and water quality classification in accordance with the Indonesian Ministry of Health Regulation, No. 492/MENKES/PER/IV/2010 [3]. Table 1 shows the reference standard for drinking water quality, as defined by the Ministry of Health.

Table 1. Standard drinking water quality

Before system testing, individual sensor calibration and accuracy tests were conducted by comparing the readings with a commercial standard device (HANNA brand). The results of the sensor testing are shown in Tables 2 to 4. These results indicate that the sensors have an accuracy range between 85% and 97%, which is acceptable for IoT-based water quality monitoring.

Table 2. Total Dissolved Solids (TDS) sensor accuracy

Table 3. pH sensor accuracy

Table 4. Temperature sensor accuracy

The data generated from the sensors is used as the basis for water quality classification using the Neural Network method with the McCulloch-Pitts model. Water quality is classified into three categories: potable (suitable for drinking), suitable for washing, and unsuitable for daily use. The classification table for each parameter is presented in Table 5. The values of each parameter are entered into a mathematical function, multiplied by their respective weights, to produce an output that is then activated to generate the final classification. Consequently, each parameter classification is represented numerically as follows:

1. Potable = 2
2. Suitable for washing = 1
3. Unsuitable for daily use = 0

Table 5. Water parameter classification

Each parameter value, represented numerically as 0, 1, or 2, is treated as an input to the neuron and multiplied by its corresponding weight (w) to produce an intermediate output. The McCulloch-Pitts model then sums the weighted inputs according to the following formula:

$net=\sum {{x}_{ij}}~{{w}_{ij}}$         (2)

where, x_ij is the input value from the classified water parameter (pH, turbidity, or TDS); w_ij is the weight assigned to the corresponding input.

The resulting net value is then passed through an activation function to determine the final classification output of the system. In this study, the weights for the McCulloch-Pitts neural network were assigned manually as w₁ = 1, w₂ = 0.25, w₃ = 1 and w₄ = 0.25. The water quality neural network structure is shown in Figure 4. The first neuron combines the pH and turbidity (NTU) inputs, each multiplied by its respective weight, to produce an intermediate output. This output is then activated by the following formula:

$f($net 1$)\left\{\begin{array}{llc}2, & \text { jika } & \text { net } 1=2,5 \\ 1, & \text { jika } & 1<\text { net } 1<2<\text { net } 1<2,5 \\ 0, & \text { jika } & \text { net } 1 \leq 1, \text { net } 1=2\end{array}\right.$         (3)

The intermediate output is then activated and passed to the second neuron, which combines it with the TDS input to produce the final water quality classification. The activation function is as follows:

$f($net 2$)\left\{\begin{array}{llc}2, & \text { jika } & \text { net } 2=2,5 \\ 1, & \text { jika } & 1<\text { net } 2<2<\text { net } 2<2,5 \\ 0, & \text { jika } & \text { net } 2 \leq 1, \text { net } 2=2\end{array}\right.$       (4)

The system was tested in three locations in Palu City, Central Sulawesi, Indonesia:

1. Palu Bangga Streat (Location 1)
2. Tondo sub-District (Location 2)
3. Duyu, Tatanga District (Location 3)

The measurement results for Locations 1, 2, and 3 are presented sequentially in Tables 6, 7, and 8, respectively. Meanwhile, the monitoring graphs displayed on the web interface are shown in Figures 5, 6, and 7.

Table 6. Water quality at Location 1

The final water quality classification at Location 2 is unsuitable, due to the very high TDS level, indicating a high concentration of dissolved particles in the water. The turbidity and pH, which range between 6 and 7, are relatively acceptable for drinking water. Therefore, this water quality can be classified as suitable for washing or can be consumed after proper treatment.

The final water quality classification at Location 3 is suitable for washing. The TDS values in the samples are below 500 ppm, with pH values ranging between 7 and 8, indicating that the water is generally in good condition for drinking purposes. However, the turbidity at Location 3 is slightly higher, so the water quality is classified as suitable for washing, although it can still be consumed after appropriate processing.

Figure 5. Water quality monitoring graph at Location 1

Figure 6. Water quality monitoring graph at Location 2

Figure 7. Water quality monitoring graph at Location 3

Some misclassification cases were observed, particularly when turbidity sensor readings were unstable due to calibration drift or environmental interference. High turbidity values occasionally dominated the classification outcome despite acceptable pH and TDS levels. This limitation highlights the sensitivity of threshold-based neural models to sensor noise. Nevertheless, the classification results remain consistent with local water usage practices, where water with high turbidity is typically avoided for direct consumption.

Table 7. Water quality at Location 2

Table 8. Water quality at Location 3

• Sample 1 Analysis

The first sample shows significant sensor value fluctuations due to unstable network conditions. The measured TDS is low with pH between 5–6, but the turbidity reading experienced calibration errors. In general, the water could still be used after treatment, but the neural network model classified it as not suitable.

1. Neural Network Formulation:

$net1=x\_\text{pH}{{w}_{1}}+x\_\text{NTU}{{w}_{2}}$         (5)

$net2={{\hat{y}}_{1}}{{w}_{3}}+x\_\text{TDS}{{w}_{4}}$          (6)

2. Calculation results:

$\begin{aligned} & \text { net } 1=(1.0 \times 1.0)+(0.0 \times 0.25 \times)=1.0 \rightarrow f(\text { net } 1)=0.0\end{aligned}$         (7)

$\begin{aligned} \text { net } 2= & (0.0 \times 1.0)+(2.0 \times 0.25 \times)=0.5  \rightarrow f(\text { net } 2)=0.0(\text { Not suitable })\end{aligned}$       (8)

Interpretation: The final classification is 'Not Suitable' because the NTU error reduced the output. Practically, the water could still be used for washing or even drinking after proper treatment.

• Sample 2 Analysis

The second sample shows stable monitoring results with good network conditions. The pH value is between 6–7 (good), NTU is good, but TDS is very high. The final neural network output classified the water as not suitable.

1. Neural Network Formulation:

$n e t 1=x_{\_} \mathrm{pH} \cdot w_1+x_{\_} \mathrm{NTU} \cdot w_2$        (9)

net $2=\hat{y}_1 \cdot w_3+x_{-}$TDS $\cdot w_4$        (10)

2. Calculation results:

$\begin{gathered}\text { net } 1=(2.0 \times 1.0)+(0.0 \times 0.25 \times)=2.5  \rightarrow f(\text { net } 1)=2.0\end{gathered}$          (11)

$\begin{aligned} & \text { net } 2=(2.0 \times 1.0)+(0.0 \times 0.25 \times)=2.0  & \rightarrow f(\text { net } 2)=0.0(\text { Not suitable, exact threshold })\end{aligned}$       (12)

Interpretation: Although pH and NTU are good, the very high TDS dominates the final decision. The net2 value being at the threshold produced a strict 'Not Suitable' classification.

• Sample 3 Analysis

The third sample provides stable data with TDS < 500 ppm and pH 7–8 (good), but NTU is slightly higher. The neural network classified the water as suitable for washing.

1. Neural Network Formulation:

$n e t 1=x_{-} \mathrm{pH} \cdot w_1+x_{-} \mathrm{NTU} \cdot w_2$        (13)

$net2=\hat{y}_1 \cdot w_3+x_{-}$TDS $\cdot w_4$      (14)

2. Calculation results:

$\begin{gathered}\text { net } 1=(2.0 \times 1.0)+(1.0 \times 0.25 \times)=2.25  \rightarrow f(\text { net } 1)=1.0\end{gathered}$        (15)

$\begin{aligned} & \text { net } 2=(1.0 \times 1.0)+(2.0 \times 0.25 \times)=1.5  & \quad \rightarrow f(\text { net } 2)=1.0(\text { Suitable for washing })\end{aligned}$         (16)

Interpretation: The combination of good pH and TDS was limited by higher NTU, resulting in the final classification 'Suitable for Washing'. Practically, the water could still be consumed after turbidity treatment.

This work was funded by the Tadulako University Excellence Research Grant (Contract No. 0604/UN28.16/AL.04/2025). The authors also wish to express their gratitude to the Power Electronics and Control System Laboratory at the Faculty of Engineering, Tadulako University, for providing the necessary research facilities and support.