Advanced Optimized Quantum Learning for Robust Water Quality Assessment with an IoT Sensor Data

Water quality has been a critical factor in environmental health and public safety in recent times [1]. The worst water quality harmfully affects ecosystems and also leads to biodiversity loss and aquatic life disruption [2, 3]. Contaminants like heavy metals, pathogens and pesticides can lead to natural habitat degradation which affects both flora and fauna [4]. Also, contaminated water risks the health of humans by causing various diseases namely cholera, dysentery and hepatitis. The World Health Organization (WHO) reports that millions of people suffer from waterborne diseases annually which also highlights the urgent need for effective monitoring and management of water resources [5]. Therefore, preserving high water quality standards is essential to save both ecosystems and human populations from harmful pollutants.

The increase in water pollution and health issues highlights the requirement for proactive water quality prediction [6]. The traditional form of water quality evaluation involves periodic sampling and laboratory analysis which are insufficient for timely detection and response. These techniques are not only time-consuming but also provide limited water conditions pictures and fail to capture dynamic changes. Several prevention methods are used to maintain and improve water quality in the priory. Physical methods include filtration and sedimentation that only remove particulate matter from water [7]. Chemical treatments namely chlorination and ozonation that used to disinfect water to avoid pathogens. Biological techniques are used as biofilters and constructed wetlands that involve natural processes to degrade pollutants.

These traditional methods have many limitations, regular maintenance requirements, the impact of potential environmental and the inability to detect and respond to sudden changes in water quality [8]. Therefore, there is an advanced technology need for real-time water quality management solutions by offering continuous monitoring and early warning. This prediction can prevent health hazards and environmental degradation at an earlier stage. Effective prediction also ensures safer water to consume, agriculture and recreational activities.

For an accurate and reliable prediction, the DL has emerged as a powerful tool of it. The DL method can provide an effective solution even in large datasets and complex patterns. It can identify patterns and correlations significantly than the traditional methods [9]. Some of the popular methods of DL used for real-time prediction are Convolutional Neural Networks (CNNs) are effective for spatial data analysis, Recurrent Neural Networks (RNNs) use time-series data to predict future, Long Short-Term Memory Networks (LSTMs) capture long-term dependencies in sequential data, Autoencoders learns an efficient data representation, Generative Adversarial Networks (GANs) provides a generate realistic synthetic data that suitable to simulate and training predictive models, Graph Neural Networks (GNNs) are Capable to handle a relational data respectively. Each method has a unique style of prediction and various performances based on complexity and data nature [10]. Therefore, accurate and reliable water quality predictions can be attained only by using an appropriate model and optimizing it. Despite their effectiveness, these methods often face challenges when dealing with highly complex, high-dimensional datasets. Additionally, their dependency on large, labeled datasets and susceptibility to overfitting can limit their generalizability and practical utility in certain scenarios.

This paper proposes IoT data with an optimized Quantum Neural Network (QNN) for real-time water quality prediction. The system used a network of IoT sensors that were placed in various water bodies to continuously collect water data such as pH, turbidity, dissolved oxygen and temperature. These sensors are used to transmit real-time cloud data to a processor where an advanced computational method of Quantum Neural Networks (QNNs) is used for prediction. The QNN can explore a vast solution space by identifying optimal patterns and correlations simultaneously. To improve performance further, an optimization model is used to fine-tune the QNN parameters to attain a reliable and optimal solution in water quality predictions. The novelty of the proposed system lies in integrating IoT-based real-time water quality monitoring with an optimized QCNN. It uses quantum principles to handle high-dimensional data and metaheuristic optimization for enhanced predictive accuracy and robustness.

The paper is organised as related work in section 2 which carries a literature work on existing systems, section 3 discusses the material and methods of proposed work that includes the proposed architecture and workflow of it, section 4 explores the result and discussion of the proposed methodology with an existing work comparison and section 5 summarise the work with a conclusion.

3.1 Materials

The provided diagram Figure 1 presents the proposed system architecture for collecting IoT sensor data to predict water quality using an optimized QCNN model. The system is composed of several components that play a crucial role in the data collection and analysis process.

1.png

Figure 1. Proposed hardware block

To attain dynamic water quality, real-time water data is taken by using four types of sensors that are connected to the Arduino controller such as:

• Temperature Sensor: Measures the temperature of the water that can affect the chemical and biological processes in the water.
• pH Sensor: Determines the acidity or alkalinity of the water which is crucial to assess the suitability of water for various uses. The pH sensor was chosen to measure and monitor water's suitability for various uses, such as drinking, irrigation, and supporting aquatic life.
• Turbidity Sensor: Measures the clarity of the water that indicates the presence of suspended particles that might affect water quality. The turbidity sensor was selected because it provides valuable insights into the presence of harmful contaminants that could degrade water quality.
• Dissolved Oxygen Sensor: Assesses the amount of oxygen dissolved in the water which is an important parameter for aquatic life and overall water health. The dissolved oxygen sensor was chosen due to its importance in assessing the ecological balance of aquatic environments.

3.2 Arduino controller

It collects data from all the sensors processes this data and serves as the central hub for communication with other system components. It handles data input and process signals to execute control commands based on predefined parameters.

3.3 IoT cloud

The controller-processed data is sent to the IoT cloud. This cloud service stores the data to carry a remote monitoring and further analysis. The IoT cloud acts as a repository where historical data are collected and facilitates the training and evaluation of the optimised QNN model for water quality prediction.

3.4 Proposed optimised QNN methods

3.4.1 M-VAE

A VAE is a probabilistic generative model that learns to encode input data into a latent space and then decode from this latent space to reconstruct the original input. In a Modified VAE, enhancements or adjustments are made to improve the model’s performance or adapt it to specific types of data or tasks. The goal of a VAE is to learn a latent space representation of the input data such that the reconstructed data closely resembles the original input. The encoder network maps an input x to a distribution in the latent space. This is typically parameterized as a Gaussian distribution with mean μ and standard deviation σ:

$q(z \mid x)=\mathcal{N} *\left(z ; \mu(x), \sigma^2(x)\right)$            (1)

where, z is the latent variable. To sample from the distribution $q(z \mid x)$, a reparameterization trick is used:

$z=\mu(x)+\sigma(x) .\epsilon$                      (2)

where, $z=\mu(x)+\sigma(x) .\epsilon$ is a random noise vector sampled from a standard normal distribution $\mathcal{N}$ (0,I).The decoder network maps the latent variable z back to the data space, producing a distribution over possible data points:

$x^{\prime}=p(x \mid z)$            (3)

where, $x^{\prime}$ is the reconstructed data. The VAE is trained to maximize the Evidence Lower Bound (ELBO) on the log-likelihood of the data:

$\left.L_{\mathrm{ELBO}}=\mathbb{E}_{q(z \mid x)}[log p(x \mid z)-K L[q(z \mid x)\}\} p(z)\right]$       (4)

where, KL denotes the Kullback-Leibler divergence between the posterior q(z∣x) and the prior p(z), often assumed to be a standard normal distribution N(0,I). Expanding this, the ELBO can be written as:

$L_{\mathrm{ELBO}}=-K L[q(z \mid x) \| p(z)]+\mathbb{E}_{q(z \mid x)}[\log p(x \mid z)]$        (5)

The hyperparameters of VAE include latent space dimensionality, the rate at which the model learns during training, the number of samples processed before the model’s parameters are updated, the number of layers, the number of neurons per layer, and activation functions used in both the encoder and decoder. In modified VAE, these parameters are tuned using a metaheuristic optimizer to achieve higher accuracy in water quality prediction.

Unlike the traditional VAE, M-VAE uses Gaussian distribution in its latent space to capture more complex patterns in the data. The major advantage of using M-VAE for feature extraction is to better handle high-dimensional and complex water quality datasets.

3.4.2 QCNN algorithm

The QCNN method is applied for a data classification which is combined with a CNN model that is given in Figure 2. This method is processed in quantum computing principles to enhance the feature extraction to achieve an efficient classification. The QCNN procedures are explained in the following.

2.png

Figure 2. QCNN architecture

Preparing the quantum state

Initially preparing the data for quantum processing is executed. Consider quantum systems with a limited number of qubits available that processed a dimensionality data reduction in qubits. Assume every data sample is an n-dimensional vector and also minimise the dimensionality of the data sample to m dimensions to fit within the quantum system's capacity. Next, the minimised data vector is normalized in range [0,1] as $x=[x 1, x 2, \ldots, x m]$ and then converted into angle information suitable for quantum encoding using $\alpha i=\pi x i$ that resulted as angle vector $\alpha=\lceil\alpha 1, \alpha 2, \ldots, \alpha m\rceil$. By encoding IoT data into a quantum state, angle data (α) to measure angle rotation for quantum gates. Specifically, each α is used as the rotation angle for a Ry gate for every qubit. The Ry gate is a quantum gate that rotates the qubit around the y-axis of the Bloch sphere by the specified angle. Here initial quantum state ∣0⟩ is transformed into a quantum state that represented as $|\phi d a t a\rangle=\bigotimes_{i=1}^m R y(\alpha i)|0\rangle$.

Quantum convolutional layer

This layer applies quantum circuits to the encoded quantum state and acts as a quantum filter to extract features in quantum namely superposition and entanglement. Multiple quantum convolution kernels are applied to the quantum state that is configured with various parameters to capture different patterns and features within the data. The quantum convolutional layer in the QCNN architecture used the principles of quantum mechanics like superposition and entanglement for feature extraction. The quantum convolutional layer applies quantum circuits to the quantum state which allows the model to encode complex features in a more compact and efficient manner. The advantage of quantum convolutional layers lies in their ability to perform simultaneous processing of multiple states which enables them to extract richer and more diverse features from water quality data with fewer parameters than classical models.

Quantum pooling layer

Next, the data moves to quantum pooling which minimises the feature maps obtained from the quantum convolutional layer which summarises the extracted features to maintain significant data. This process manages the system’s complexity and prepares the data for further processing. The quantum pooling layer performs a dimensionality reduction operation similar to conventional pooling in CNNs. But, it also uses the concept of quantum parallelism. In this layer, the quantum states obtained from the quantum convolutional layer are processed through quantum pooling circuits. The primary advantage of quantum pooling is its ability to reduce the number of qubits needed. It is also used to hold the essential characteristics of the data.

Measurement and classical processing

Then the conversion of quantum data into classical data is processed. The measurement process collapses the quantum states into classical values and provides a new feature vector which is a captured data extraction of quantum layers.

Fully connected classical layer

Converted data is fed into a fully connected classical layer for a final classification task. It trains the feature vectors by using backpropagation methods to achieve higher efficiency and effectiveness in training methods.

Training with backpropagation

The backpropagation is used to train HCNN by using systems like TensorFlow Quantum that facilitate a quantum circuit’s numerical simulation in classic systems and implement a gradient descent in quantum systems. The adjoint differentiation model is employed to compute gradients in the quantum to attain an efficient quantum circuit training model.

To enhance the accuracy of the QCNN method of prediction, the hyperparameter of QCNN is tuned by using a Siberian Tiger Optimization (STO) algorithm [21, 22]. Table 1 lists the parameters of the QCNN method.

These parameters can be tuned using the Siberian Tiger Optimization (STO) algorithm to find the optimal configuration for the QCNN.

Table 1. QCNN hyperparameters

3.4.3 STO method for QCNN parameter tuning

It is motivated using Siberian tigers’ characteristics that involve hunting prey and fighting brown bears. Compared to other metaheuristic optimization algorithms, the STO has a faster convergence rate and can easily escape local minima. The mathematical evaluation of these behaviours with an iterative optimization to identify an optimal result for complex issues. This method can effectively fine-tune the parameters of a QCNN to classify the data.

Initialization

Initially, the population of candidate solutions is initialised that represents a set of hyperparameters for the QCNN.

1. Population Matrix (X): Initialize the tiger’s positions randomly that represent a candidate solution in each space.
2. Best Solution (X_best): identify the best solution of an objective function value.

Phase 1: Prey Hunting

This phase explores the hunting behaviour of tigers in search space extensively. It selects a prey position with a better objective function value and simulates an attack, resulting in significant changes in its position.

1. Position Update:

$x_{\mathrm{i}, \mathrm{j}}^{p 1 s 1}=x_{\mathrm{i}, \mathrm{j}}+\operatorname{rand}\left(M_{\mathrm{i}, \mathrm{j}}-I_{\mathrm{i}, \mathrm{j}} x_{\mathrm{i}, \mathrm{j}}\right)$          (6)

where, rand is a parameter that varies from zero to one, M_i,j is a member, j is the dimension and I_i,j are coefficient varies from zero to two.

2. Final Position Update:

$x_{\mathrm{i}, \mathrm{j}}^{p 1 s 2}=x_{\mathrm{i}, \mathrm{j}}+\frac{{rand}\left(u l_j-l l_j\right)}{t}$       (7)

where, t is the iteration number, and ul and ll denote the lower and upper limits for optimization.

Phase 2: Fighting with a Bear

This phase simulates tigers fighting brown bears. Tigers select a bear position and calculate a new position based on the bear's position. It can be expressed as follows:

1. Position Update:

$x_{\mathrm{i}, \mathrm{j}}^{p 2 s 1}=\left\{\begin{array}{c}x_{\mathrm{i}, \mathrm{j}}+\operatorname{rand}\left(x_{k, \mathrm{j}}-I x_{\mathrm{i}, \mathrm{j}}\right), F_k<F_{\mathrm{i}} \\ x_{\mathrm{i}, \mathrm{j}}+\operatorname{rand}\left(x_{\mathrm{i}, \mathrm{j}}-x_{\mathrm{k}, \mathrm{j}}\right), E L S E\end{array}\right.$             (8)

2. Final Position Update:

$x_{\mathrm{i}, \mathrm{j}}^{p 2 s 2}=x_{\mathrm{i}, \mathrm{j}}+\frac{\operatorname{rand}\left(u l_j-l l_j\right)}{t}$  (9)

Iterative Process

The STO algorithm iteratively processes these stages to reach an optimal solution. The objective function is set as a function of the accuracy or error rate of the model. The pseudocode of the proposed optimization is given below:

Pseudocode of optimised QCNN

1. Sets the hyperparameters of the QCNN model based on the provided candidate solution.
2. Train specified hyperparameters.
3. Evaluate the trained model on a validation dataset and calculate the accuracy.
4. Represents the accuracy of the trained model on the validation dataset.
5. Returns the validation accuracy as an objective value to maximize during optimization.
6. Randomly initialize the population of candidate solutions.
7. Select candidate solutions based on their fitness scores.
8. Select a prey position for the tiger.
9. Update the tiger's position in the first stage.
10. Finalize the tiger's position update.
11. Select a bear position for the tiger.
12. Update the tiger's position in the first stage of the bear fight.
13. Finalize the tiger's position update during the bear fight.
14. Output the best solution and its accuracy.

By using an optimized QCNN model, the system provided an efficient water quality prediction with advanced quantum computing even in complex water quality data. It also carried reliable and more accurate prediction outcomes. The same iterative steps were followed to tune the hyperparameters of VAE.

3.5 Display alert

The predicted quality data outcomes are transmitted to the IoT cloud where the controller is also linked to a display alert system. This component provides real-time alerts and notifications in the system to ensure immediate awareness of any important changes or issues in water quality parameters. The alerts help in prompt decision-making and corrective actions.

Therefore, overall, an integrated system continuously monitors the quality of water with an integration of IoT technology and predictive analysis using an optimised QCNN model. Thus, the proposed optimised QCNN model is effectively trained with the historical IoT data and optimized to accurately predict future water quality. This method helps to ensure water safety to provide a quality environment.