Enhanced Long-Range Network Performance of an Oil Pipeline Monitoring System Using a Hybrid Deep Extreme Learning Machine Model

Abstract

Leak detection in oil and gas pipeline networks is a climacteric and frequent issue in the oil and gas field. Many establishments have long depended on stationary hardware or traditional assessments to monitor and detect abnormalities. Rapid technological progress; innovation in engineering; and advanced technologies providing cost-effective, rapidly executed, and easy to implement solutions lead to building an efficient oil pipeline leak detection and real-time monitoring system. In this area, wireless sensor networks (WSNs) are increasingly required to enhance the reliability of checkups and improve the accuracy of real-time oil pipeline monitoring systems with limited hardware resources. The real-time transient model (RTTM) is a leak detection method integrated with LoRaWAN technology, which is proposed in this study to implement a wireless oil pipeline network for long distances. This study will focus on enhancing the LoRa network parameters, e.g., node power consumption, average packet loss, and delay, by applying several machine learning techniques in order to optimize the durability of individual nodes’ lifetimes and enhance total system performance. The proposed system is implemented in an OMNeT++ network simulator with several frameworks, such as Flora and Inet, to cover the LoRa network, which is used as the system’s network infrastructure. In order to implement artificial intelligence over the FLoRa network, the LoRa network was integrated with several programming tools and libraries, such as Python script and the TensorFlow libraries. Several machine learning algorithms have been applied, such as the random forest (RF) algorithm and the deep extreme learning machine (DELM) technique, to develop the proposed model and improve the LoRa network’s performance. They improved the LoRa network’s output performance, e.g., its power consumption, packet loss, and packet delay, with different enhancement ratios. Finally, a hybrid deep extreme learning machine model was built and selected as the proposed model due to its ability to improve the LoRa network’s performance, with perfect prediction accuracy, a mean square error of 0.75, and an exceptional enhancement ratio of 39% for LoRa node power consumption.

1. Introduction

1.1. Background

Oil leak localization and detection systems aim to protect pipeline networks from damage to protect their surroundings, organisms, and their lives from the risk of pipeline network catastrophes. Oil pipeline leak detection and monitoring systems are important because when a leak happens, they automatically update the operator with detailed alarms so that the necessary instructions and procedures can be organized and ready to mitigate spills and their durations.

There are different groups of leak detection techniques. Some of these technologies detect product leaks by checking the variance between the volume of the steady-state situation of the system and the value of the instant inlet and outlet flow rate [1,2]. Other techniques employ advanced computerized network systems that are used to monitor a list of different conditions of operation. The real-time transient model (RTTM) was selected as the oil pipeline monitoring model in this study due to its many benefits when compared with other techniques, such as a negative pressure wave, pressure point analysis, acoustic sensing, fiber optics, RFID, vapor sampling, capacitive sensing, infrared cameras, etc. The RTTM achieves a high level of accuracy in leak detection and localization, with appropriate leak size estimation, easy maintenance, good performance in both steady- and transient-state cases, fast processing, and the ability to handle a suitable amount of data [3,4,5,6].

Wireless sensor networks (WSNs) are distributed networks that cover multiple small, self-dominant sensing nodes that can sense, compute, and communicate with the main central server [7]. The main function of the end nodes is to work concertedly to gather and observe information from their immediate surroundings. WSNs have gained significant interest and have been employed in various domains, such as ecological observation systems, healthcare systems, the infrastructure of intelligent cities, industrial system control, smart monitoring, and intelligent surveillance systems [8,9].

Machine learning is a branch of artificial intelligence that involves developing algorithms that allow systems to learn from old data and then predict experiences and output results depending on the new real input data when they receive them. The quantity and quality of data used during the training process and selecting the right algorithm to deploy increase the level of accuracy and correctness of the prediction. Environmental monitoring, industrial manufacturing, speech recognition, image identification, Facebook auto-tagging, and email filtering all require machine learning [10].

1.2. Our Contribution

This work presents a novel approach to implementing an integrated, efficient, and reliable oil pipeline monitoring system. It covers the approach of enhancing the accuracy of system functionality, in addition to utilizing machine learning techniques over wireless sensor networks to enhance network performance. The contributions of this work are explained as follows:

• To the best of our knowledge, this work is the first of its kind to introduce an integrated and efficient solution in the field of oil pipeline monitoring.
• The accuracy and efficiency of a real-time transient model (RTTM)-based LoRa WAN oil pipeline leak detection system is maintained and increased in terms of continued monitoring and a quick detection response time by applying machine learning techniques that enhance node power consumption and extend the node battery lifetime.
• The number of hardware resources utilized to implement the wireless network of the proposed system is optimized due to the selection of long-range (LoRa) technology, which offers several benefits over other wireless communication technologies, especially in terms of environmental monitoring systems that deploy the Internet of things (IoT) and low-power, wide-area network (LP-WAN) applications.
• The whole of the LoRaWAN network’s performance, especially in the context of LoRa node power consumption, was significantly improved by implementing different approaches of machine learning techniques using the LoRa network. The hybrid deep extreme learning machine model provided impressive results in improving network performance, e.g., power consumption, packet loss, and packet delay, compared with that of other models.

1.3. Paper Organization

The rest of this article is organized as follows: Section 2 presents a survey of the related literature and the research problems. Section 3 offers a brief overview of the LoRaWAN technology, machine learning techniques, and the main conception of the deep learning approach and presents the simulator that will be used to implement the network of the proposed leak detection and monitoring system (LDMS), in addition to the main libraries and frameworks utilized to implement the FLoRa network and different approaches to artificial intelligence using the LoRa network. Section 4 describes an appropriate leak detection model to achieve the functions of real-time pipeline monitoring, leak detection, and localization system implementation using the LoRaWAN protocol and the implementation of different approaches to artificial intelligence algorithms for enhancing the performance of a wireless network. Section 5 presents the results and a discussion of the research findings. The conclusion and future work are presented in Section 6.

2. Literature Survey

2.1. Related Work

A concentrated survey of the literature, especially that on approaches that employ wireless sensor networks and artificial intelligence for oil pipeline leak detection systems, is presented in this section. In 2015, Mustafa Alper Akkas et al. [11] focused on employing the electromagnetic approach based on wireless underground sensor network (WUSN-based EM) communication techniques for oil pipeline networks. Their research explained a suitable transmission range that is optimal in terms of the intake losses of the compound underground oil pipeline network and determined the efficient communication distance between sensing nodes operating at 315 MHz in an underwater environment. In 2015, Ola E. Elnaggar et al. [12] studied the problems of linear sensor placement in oil pipeline monitoring. They applied two evolutional algorithms to resolve the issue regarding deploying the optimal number of sensors, i.e., ant colony optimization (ACO) and a genetic algorithm (GA), to realize the aims of the deployment procedure and maximize the coverage of the pipeline, thus creating a connected network and extending the lifetime of the whole network.

In 2020, Lam-Thanh Tu et al. [13] utilized the deep learning (DL) approach with LoRa technology to achieve the optimal transmitting power, which is useful for maximizing energy efficiency (EE) in LoRa networks. In 2018, Brecht Reynders et al. [14] proposed a novel MAC protocol to enhance the scalability and reliability of LoRaWAN. A lightweight scheduling algorithm was implemented by dividing a huge number of nodes into several groups, while similar values of transmission powers were used in each group to reduce the capture effect. In 2018, Arliones Hoeller et al. [15] proposed message replication and gateways with multiple receiving antennas to accomplish, respectively, the temporal and spatial diversity of a LoRa network. Their study presented the proposed schemes and assessed them through theoretical analysis and computer simulations. The final results showed that LoRa networks are very sensitive to user increases and traffic density, but the proposed technique can enhance performance.

In 2020, Duc-Tuyen Ta et al. [16] developed a LoRaWAN simulator that integrated with the multi-armed bandit and reinforcement learning (RL) algorithms to formulate and find a professional resource allocation solution for LoRa endpoints. In 2021, Inaam Ilahi et al. [17] presented an intelligent deep reinforcement learning algorithm for multichannel resource allocation in dense LoRa networks, known as the LoRa-DR technique, which enhanced LoRaWAN’s packet delivery ratio and ensured lower power consumption; therefore, the lifetime and network capacity were increased. In 2022, Merin Susan Philip et al. [18] proposed a LoRa-based energy consumption model that can predict the amount of energy consumed for each sensing node in a WSN. The study showed that the average energy consumption and lifespan of a battery are affected by a decreasing sensing interval and an increasing spreading factor of the standard parameters of LoRa.

In 2021, Seham Ibrahem Abd Elkarim et al. [19] introduced a smart spreading factor (SF) assignment method that depended on the different structural designs of artificial neural networks to detect collisions and selected optimal SFs to reduce the probability of collisions and improve network performance, addressing issues such as the energy consumption for the LoRa network. In 2022, Benyamin Teymuri et al. [20] introduced a new approach using the multi-armed bandit (MAB) technique to set LoRa end devices’ transmission power parameters in a centralized manner over the network server side, while also improving the energy consumption.

In 2023, Olaide Agbolade et al. [21] presented a LoRaWAN-based flow-rater approach for detecting and localizing leaks in pipelines. They included an investigational setup that simulated a pipeline network with flow rate and pressure sensor measurements. In 2023, Mohammed Jouhari et al. [22] presented a deep reinforcement learning technique to enhance the energy efficiency in wireless LoRa networks that consisted of flying gateways to expand the lifetime of the networks and LoRa end devices. The trained model could efficiently reallocate the spreading factors (SFs) and adjust the transmission powers (TPs) to achieve network performance enhancement. In 2023, Dere S. et al. [23] presented an approach to implementing an IoT monitoring kit for pressure rate logging in pipeline systems. It applied a pressure sensor and GSM model to gather and send pressure information in a real-time manner in order to detect, localize, and monitor leaks through a mobile application service.

In 2023, Ali Loubany et al. [24] presented an efficient algorithm for enhancing the energy consumption of LoRaWAN networks via a compound of gateways by applying the selection of spreading factors and power controlling. In 2022, Chavala Lakshmi Narayana et al. [25] offered a methodology for oil pipeline monitoring over long distances by applying the LoRa communication protocol and sensor node installation along the pipeline. In 2023, Surenther et al. [26] presented a deep-learning-based grouping model approach, called the DL-GMA technique, that enhanced energy utilization in wireless sensor networks. This approach extends the WSN’s lifetime and increases the efficiency of data transmission.

In 2020, Teoh Ji Sheng et al. [27] developed a smart waste management system utilizing a TensorFlow-based deep learning approach with a LoRa communication protocol to build a real-time monitoring system and achieve better performance in the waste management field. In 2024, Salaheddin Hosseinzadeh et al. [28] proposed a scheme that splits feature extraction from regression analysis, reducing the need for training data. They achieved the lowest root-mean-square error gradient by applying boosting based on the decision tree technique. As a result, they enhanced the accuracy of LoRaWAN propagation estimation and improved the analysis of the link budget, interference management, quality of service (QoS), scalability, and energy efficiency in the entire network.

In 2023, Višnja Križanović et al. [29] presented an optimization approach for energy consumption in terms of selecting the proper data collection processes. It was proven that tuning the transmission rate according to the actual size of the packet payload is important to improve the energy efficiency of communication. In 2022, Syed Usama Minhaj et al. [30] presented the new concept of applying two approaches to machine learning for assigning spreading factors (SFs) and the transmitting power of LoRa end devices by using a combination of centralized and decentralized approaches. SFs are allocated to the devices using reinforcement learning for contextual bandit problems; on the other hand, transmission power is allocated centrally by applying machine learning techniques. In 2018, Vignesh Mahalingam Suresh et al. [31] presented a solution that deploys a machine learning approach on the edge of LoRa network devices and enhances the power transmission values through LoRa by compressing the transmitted data at a very high ratio, reaching 500 times that of conventional methods.

In 2021, Lam-Thanh Tu et al. [32] presented a framework that was derived by investing the tools of stochastic geometry and linking the long-range (LoRa) energy efficiency parameters with the density of the end devices and their transmitting power. The result shows the maximization of energy efficiency. In 2020, Inaam Ilahi et al. [33] presented a deep reinforcement learning algorithm based on physical-layer transmission parameters to enhance LoRaWAN performance and guarantee a decrease in packet collisions over the network channel. In 2023, M.D. Rakibul Islam et al. [34] presented a mathematical approach for a multi-hop network that adjusts the Distance Ring Exponential Station Generator (DRESG) framework. This research covers the most important issues, such as interference, power transition limitations, and different environmental conditions. In 2023, Zhang et al. [35] presented an innovative, reliable, energy-efficient model assisted by beamforming with a firmware update over the air (FUOTA) for LoRa networks; this system was named FLoRa, and it featured several techniques, including delta scripting, channel coding, and beamforming. The FLoRa model enhances the network reliability for transmission by up to 1.51× and the energy efficiency by up to 2.65× compared with the present solutions in LoRaWAN.

Table 1. Summary of the most relevant work.

Author(s)/Year	Technique Used	Problem Addressed	Network Parameter Enhancement
Mustafa Alper Akkaş et al. [11]/2015	Employing the electromagnetic approach based on a wireless underground sensor network.	Determining the efficient communication distance between sensing nodes.	No
Ola E. Elnaggar et al. [12]/2015	Ant colony optimization (ACO) and genetic algorithms (GAs).	Linear sensor placement for oil pipeline monitoring.	No
Lam-Thanh Tu et al. [13]/2020	Deep learning (DL) approach with LoRa technology.	Achieving the optimal transmitting power, which is useful for maximizing energy efficiency in LoRa networks.	Yes
Brecht Reynders et al. [14]/2018	A lightweight scheduling Algorithm.	Reducing the capture effect.	Yes
Arliones Hoeller et al. [15]/2018	Message replication and gateways with multiple receiving antennas.	Temporal and spatial diversity of LoRa networks.	Yes
Duc-Tuyen Ta et al. [16]/2020	LoRaWAN multi-armed bandit and reinforcement learning (RL) algorithm.	Resource allocation solution for LoRa endpoints.	Yes
Inaam Ilahi et al. [17]/2021	Deep reinforcement learning algorithm for multi-channel resource allocation in LoRa networks.	Enhancing the packet delivery ratio (PDR) and ensuring lower power consumption to increase the lifetime and network capacity.	Yes
Merin Susan Philip et al. [18]/2022	Decreasing the sensing interval and increasing spreading factor of the LoRa standard parameters.	Predicting the amount of energy consumed for each node.	Yes
Seham Ibrahem et al. [19]/2021	Smart spreading factor (SF) assignment method that depended on artificial neural networks.	Selecting optimal SFs to reduce collisions and energy consumption.	Yes
Olaide Agbolade et al. [21]/2023	LoRaWAN-pipeline-based flow-rater approach.	Detecting and localizing leaks in pipelines.	No
Mohammed Johari et al. [22]/2023	Deep reinforcement learning and flying gateways of LoRaWAN.	Reallocating the spreading factors (SFs) and adjusting the transmission powers (TPs) to achieve network performance enhancement.	Yes
Dere S et al. [23]/2023	Application of a pressure sensor and GSM model to gather and send pressure information in a real-time manner.	Monitoring kit for pressure rate logging in pipeline systems.	No
Ali Loubany et al. [24]/2023	Selection of spreading factors and power controlling.	Enhancing the energy consumption of LoRaWAN networks with a compound of gateways.	Yes
Chavala Lakshmi Narayana et al. [25]/2022	Applying the LoRa communication protocol and sensor node installation along the pipeline.	Oil pipeline monitoring over long distances.	No
Surenther et al. [26]/2023	Deep-learning-based grouping model approach.	Extending the WSN lifetime and increasing the efficiency of data transmission.	Yes
Teoh Ji Sheng et al. [27]/2020	Deep learning approach using the LoRa communication protocol.	Building a real-time monitoring system and achieving better performance in the waste management field	No
Salaheddin Hosseinzadeh et al. [28]/2024	Feature extraction from regression analysis.	Enhancing the accuracy of LoRaWAN propagation estimation and energy efficiency.	Yes
Vignesh Mahalingam Suresh et al. [31]/2018	Deploying a machine learning approach on the edge of a LoRa network.	Enhancing the transmission Power.	Yes
MD. Rakibul Islam et al. [34]/2023	Mathematical approach for a multi-hop network.	Treating interference and power transition limitations.	No

provides a summary of the most relevant works discussing oil pipeline monitoring systems, specifically the enhancements of LoRaWAN network performance in these and other systems.

2.2. Research Problems

The related work presented in this literature survey covered more than ten years, and the primary focus was either the approach to enhancing oil pipeline leak detection systems or highlighting the LoRa network performance with a limited ratio of enhancement. During this research, the following issues and challenges were identified:

• A real-time transient model (RTTM) was developed as a robust technique for accurate leak detection, while the integration with LoRaWAN for real-time monitoring presents challenges in the achievement of energy efficiency.
• Current LoRa-based monitoring and leak detection systems lack the power optimization needed for continuous and reliable operation. This issue decreases the node lifespan and affects the feasibility of continued leak detection.
• The standard LoRaWAN does not efficiently balance power consumption and the high-rate data transmission required by RTTM-based oil pipeline leak detection systems.
• According to the achievements in related and previous work, this research reports the need for an optimized LoRaWAN model that deploys different aspects of machine learning techniques to reduce the node power consumption with the target of reaching a 40% reduction with respect to the baseline power consumption of the standard LoRa network configuration, as well as the need to enhance the node packet delay and number of packets lost over the entire network, while maintaining the reliability, which is necessary for stable real-time leak detection and localization operations.

3. Proposed Model Techniques

3.1. LoRa Overview: What Is LoRa?

LoRa refers to a long-range wireless communication technique introduced by the LoRa Alliance and developed by Semtex. It is used in wireless metropolitan area networks (WMANs). LoRa was proposed to support wireless network services for devices with limited resources, such as battery-operated devices and those with low data rates that do not exceed 100 kbps, and it has a long-distance coverage of up to 20 km [36,37,38].

LoRa is considered a low-power, long-coverage-range wireless communication technique that transfers small-sized data over long distances. These important features make it a preferred choice for IoT industries. LoRaWAN is an application layer that calculates how devices utilize the LoRa hardware. In addition to this concept, various challenges related to LoRa are being overcome in different fields, such as those associated with weather conditions, pollution, and disasters, as well as operations related to measuring energy monitoring and consumption or production in several renewable energy systems [39].

LoRa is preferred for its simplicity because it is considered a cost-effective solution, so it is broadly utilized in smart system engineering, transportation, agriculture, household appliances, and the oil and gas industry. In the LoRa network, the desired data are sent over longer distances than those with other wireless communication technologies, such as Bluetooth, Wi-Fi, and Zigbee. The standards of the LoRaWAN configuration ensure that it provides integrity with different devices; this is a key reason why the LoRa technology was so quickly integrated into smart systems and the IoT field. A LoRa network links densely connected networks, which consist of a huge number of endpoint devices, and this makes it able to observe, detect, and monitor information over a huge number of nodes in a specialized manner. The structure of LoRaWAN has four components: end-point devices, gateways, an application server, and a network server. In addition, the LoRa network’s stack protocol presents and focuses on the layers involved in data transmission, such as the physical layer, the data link layer, and the application layer [40]. The LoRa network structure and stack protocol are shown in Figure 1.

LoRa Transmission Parameters

The endpoints of LoRa configure and employ several attributes, such as separating factors, bandwidths, the correction rate, and the transmitting power, which result in many opportunities for variation. A LoRa network’s efficiency is significantly affected by the setting of these attributes; calculating the appropriate configuration and adjusting the value of energy transmitted while still providing the necessary range of communication performance is an important challenge that should be addressed. Therefore, it is vital to select transmission attributes that fit with the LoRa network.

• Spreading Factor: The spreading factor calculates how much a signal spreads in time. The SF is obtained by dividing the rate of the chip over the rate of the symbol; in other words, the SF refers to the number of bits represented by a specific symbol. When the value of the SF increases, the signal-to-noise ratio (SNR) also increases, resulting in increased sensitivity and range. The LoRaWAN spreading factor standards range between 7 and 12, indicating that a higher value of the spreading factor improves the limit of the communication range by scaling the reactivity for the receiver side, while the data rate is reduced. Thus, higher SF values lead to high LoRa node power consumption.
• Bandwidth: The bandwidth denotes the range of various frequencies that the signal or a communication channel can hold. It is typically calculated in hertz (Hz) and refers to the difference between the highest and lowest frequencies that can travel over a channel. Typical bandwidth ranges for the LoRa network are 125 kHz, 250 kHz, and 500 kHz. A wide range of bandwidths can lead to a higher data rate but at the same time, reduce the coverage range due to increasing noise levels.
• Coding Rate: The LoRa custom error-coding approach improves the range of wireless connections. The outcomes of the CR provide extra information that should be carried over the physical layer of the LoRa payload, and this is adjusted using the coding rate attribute. The coding rate refers to the level of redundancy added to the transmitted data for error correction. LoRa provides four CR configurations: 4/5, 4/6, 4/7, and 4/8. A lower value of the CR, e.g., 4/5, means less redundancy, which leads to higher throughput but lower robustness to interference, while a higher CR provides improved error correction abilities but at the cost of the data rate.
• Transmission Power: The transmission power is a crucial approach for the delivery of data packets in the LoRa network, so it should be adjusted properly. Making the value of the transmission power low will extend the life of the battery while making the signal range shorter, and vice versa. The most common range of TP values is between 2 dBm and 20 dBm for most LoRa devices, making balancing the TP important for optimizing performance [41].

3.2. Machine Learning: A General Overview

Machine learning is an area of research that is important and rich in information and applications in the IoT field; it involves giving computer programs the ability to enhance their performance depending on old experiences. In general, machine learning consists of different stages, such as data processing, training a model on specific data, and finally, testing the model to evaluate the system’s performance [42].

To accomplish this approach, different machine learning techniques are reviewed to evaluate the data prediction accuracy by finding relationships and extracting features between various huge amounts of data. When using a regression approach, the random forest algorithm is considered one of the most powerful ensemble learning techniques, and it combines multiple groups of decision trees to make predictions. The random forest has been compared with other regression algorithms, such as support vector regression (SVR), linear regression, k-nearest neighbors (k-NN), the decision tree (DT), and gradient-boosting machines (GBMs) [43,44].

Although the random forest algorithm suffers from several limitations, such as intensive computational load, difficulties in fine-tuning for certain high-precision tasks, and greater memory consumption, it has proven to be effective for many factors, such as handling high-dimensional datasets with large and nonlinear features, providing interpretability, demonstrating robustness to overfitting, and handling missing values [45,46].

Table 2. A brief comparison of machine learning techniques.

Regression Technique	Description	Strengths	Weaknesses	How Random Forest Is Better
Random Forest	Ensemble learning technique joining multiple decision trees in which each specific tree is trained on a random subset of the data. The predictions are averaged.	Nonlinear data handling used to reduce overfitting; robust to outlier values, performs well with large datasets.	It can be expensive due to its computation.	Perfect for complex and large, nonlinear datasets; reduces overfitting.
Linear Regression	Making predictions depends on a linear relationship between features.	Simple, understandable, working on linearly related data.	Sensitive to outlier values, inadequate performance with datasets with nonlinear relationships.	RF can handle nonlinearity, and it performs better with outliers.
Support Vector Regression	Determining a hyperplane in a high-dimensional dataset that fits the data perfectly within a particular boundary of tolerance.	Efficient with small datasets, performs well in high-dimensional spaces, supports good generalization with unseen data.	Needs extensive computation, is difficult to tune, and its performance is reduced with huge datasets.	RF is more accurate and faster with large datasets, tuning easily. Better performance for nonlinear and large datasets.
k-Nearest Neighbors	Estimating the targeted value based on calculating the average values of the k-nearest data points in a certain dataset.	Simple and instinctive, performs well with small datasets and limited numbers of features.	Performance is lower with huge datasets, and it is very sensitive to noisy data and unrelated features.	RF is stronger regarding noise and unrelated features, and it has better achievements on complex and large datasets.
Decision Trees	A simple technique that divides data into multiple branches to make estimations based on feature values.	Easy to understand, can handle nonlinear relations, and performs well when the dataset is small.	Vulnerable to overfitting, sensitive, and unstable with small changes in the data.	RF performs better than DT for overfitting, reducing variance, and enhancing stability through averaging.

presents a brief review of the most efficient regression techniques, focusing on the weaknesses and strengths of each one and why the random forest algorithm was selected [47].

While machine learning has innovative applications in wireless communication for enhancing network performance parameters, such as efficiency and security requirements, there are still limitations in its ability to address resource management and help in decision making in situations that depend on IoT applications within LoRaWAN. However, machine learning algorithms have a clear ability to enhance the efficiency of LoRaWANs through dynamic resource allocation, network traffic prediction, interference modification, and energy savings, ultimately refining the network’s reliability, capacity, and power consumption [48,49].

The deep extreme learning machine (DELM) presents several distinct benefits over traditional machine learning methods, neural networks, and other deep learning approaches. Unlike classical machine learning, which often requires manual feature extraction and iterative fine-tuning of the parameters, the DELM leverages random weights for the initialization of the hidden layer and initiative weights for the computation of the output. These features not only make the implementation simpler but also decrease the training times impressively. Unlike conventional neural networks, which may suffer from long and extensive training processes as a result of the backpropagation process, DELM is a one-shot learning method that decreases the computational overhead while maintaining competitive accuracy in tasks such as regression and classification. Furthermore, the DELM’s ability to study hierarchical representations through multiple levels of layers improves its ability to capture complicated patterns in data, making it suitable for complicated datasets in various domains such as natural language processing, image recognition, and the analysis of sensor data.

DELM models suffer from different limitations in terms of random weight initialization, and they may require difficult and careful configuration to achieve consistent performance. In addition, sometimes, their accuracy can be limited by the lack of a backpropagation technique, which is discarded to speed up the training process. They stand out for their scalability, efficiency, and forceful performance, making them effective tools in recent machine learning research and applications [50].

3.3. Network Simulator

In this section, we explain the simulation environment and several tools that will be selected for modeling and analyzing the FLoRa network’s performance. An overview of the network simulator and several frameworks and libraries that will be selected is presented below.

• OMNeT++: OMNeT++ is an abbreviation for object module networks simulator. It is a flexible and expandable network simulator that depends on the C++ language and its libraries. OMNeT++, an object-oriented modular framework for simulating discrete-event networks, boasts a generic architecture. This flexibility allows it to tackle a wide range of problems, such as modeling communication networks for both wired and wireless situations, protocol schemes and analysis, queuing network performance, distributed hardware systems and multiprocessors, validation of hardware architecture, and complex software performance evaluation. These components developed in C++ are built hierarchically, and simple components can be collected to build compound models, which are used in a high-level network description (NED) language. OMNeT++ can support different technologies, such as Wi-Fi, ZigBee, LTE, and LoRa [51].
• FLoRa: FLoRa is the LoRa framework; it is a software framework for deploying end-to-end simulations of long-coverage-range networks. FLoRa is based on the OMNeT++ simulator and INET framework. FLoRa ensures the creation of LoRa networks with modules such as LoRa nodes, gateways, and network and application servers. The concept of an application in OMNeT++ can be deployed using independent modules that are linked to the network server. The network server and LoRa nodes support effective system management configuration and parameter adjustment by applying the adaptive data rate (ADR) technique. In addition, the statistics on energy consumption are composed and calculated for every node [51].
• Scikit-learn, often called sklearn, is a common open-source machine learning library used with Python. It is constructed on top of Matplotlib, NumPy, and SciPy and provides an efficient and simple tool for data analysis and mining tasks, with vital tools and algorithms for tasks ranging from regression and classification to clustering and dimensional reduction. In addition, pandas is commonly utilized for data cleaning, manipulation, handling, transformation, and analyzing various domains more flexibly; it is used in finance, statistics, economics, etc. From the perspective of network parameter enhancement, it is important to combine and merge huge datasets and implement various data operations, such as data cleaning, handling missing data, duplication removal, reshaping, and data transformation [52].
• Extreme Learning Machine: The high-performance extreme learning machine library is a Python-dependent implementation of the extreme learning machine (ELM) algorithm. It is considered to be fast, efficient, and capable of handling huge datasets. It is expressly preferred for tasks such as regression and classification approaches, and it provides approaches to both neural networks and single-layer and multiple-layer ELMs, permitting users to create complex models depending on the problem’s requirements [53].

4. Network Implementation and Learning Algorithm

This section investigates the simulation environment, techniques, and implementation tools used to build and analyze the FLoRa network for oil pipeline leak detection and localization within the OMNeT++ framework. We will explore how OMNeT++ can be integrated with several network libraries and how Python libraries such as Keras and TensorFlow can be utilized to implement and analyze deep learning algorithms in real-time network scenarios.

4.1. Default LoRa Network

The oil pipeline leak detection and localization monitoring system will be deployed in OMNeT++, based on the Inet and FLoRa frameworks, to cover the network part of a pipeline with a length of 200 km.

An oil pipeline monitoring system is proposed to calculate a leak’s volume and detect the specific location of a leak in a pipeline. Generally, it consists of two parts; the first is the endpoint-sensing part, which divides the pipeline into 100 segments. Each segment has a group of sensors deployed in the inlet and outlet parts of a pipeline segment, e.g., a pressure sensor, temperature sensor, and flow meter. Each specific sensor is used to read a certain value of the oil state, and each group of sensors is connected with LoRa nodes, which are used to collect the sensed data and transmit them to the central server. The second part of the proposed monitoring system is the real-time transient model (RTTM) technique, which is located on the LoRa server side, and it is responsible for mathematical calculations to achieve real-time monitoring and the detection of leak points.

The network design of the proposed system is described as follows:

• The LoRa network consists of 100 LoRa nodes to cover the whole oil pipeline network.
• Each pair of sequential nodes in the pipeline network is separated by two kilometers. This distance refers to a single pipeline segment supplied with two sensors for the inlet and outlet.
• The network comprises ten LoRa gateways, separated over the oil pipeline network, and each one boasts a wireless coverage radius of twenty kilometers.
• Every ten nodes attach wirelessly to a single LoRa gateway. These nodes work as sensors used to transmit the instant parameters of the oil for each pipeline segment, such as the pressure (Pt), instant time, flow rate (M), and temperature (Temp) (for both the inlet and outlet), as well as the node identifier.
• LoRa gateways act as bridges between endpoints and the LoRaWAN server, connecting via the cloud. They are responsible for managing connections, establishing and coordinating the connected nodes, and ultimately, forwarding the sensed data to the LoRa server for further processing.
• The function of the LoRa server is to continuously monitor the pipeline status within the simulation. On the server side, a real-time transient model is deployed to calculate the leak locations and quantities in the pipeline. To achieve this, it calculates the discrepancies, which are represented by X, which is, depending on the calculation, the difference between the values of the inlet simulated and measured (sensor) flow rate and the calculated values of the inlet flow rate, and Y refers to the same parameters but for outlet flow rates in the pipeline segment via a specific mathematical equation.

$$X=MIn\left(cal\right)-MIn\left(mes\right)$$

(1)

$$Y=MOut\left(cal\right)-MOut\left(mes\right)$$

(2)

where

• $\mathrm{M}\mathrm{I}\mathrm{n}\left(\mathrm{c}\mathrm{a}\mathrm{l}\right)$ = calculated flow rate for the inlet;
• $\mathrm{M}\mathrm{I}\mathrm{n}\left(\mathrm{m}\mathrm{e}\mathrm{s}\right)$ = simulated and measured flow rate for the inlet;
• $\mathrm{M}\mathrm{O}\mathrm{u}\mathrm{t}\left(\mathrm{c}\mathrm{a}\mathrm{l}\right)$ = calculated flow rate for the outlet;
• $\mathrm{M}\mathrm{O}\mathrm{u}\mathrm{t}\left(\mathrm{m}\mathrm{e}\mathrm{s}\right)$= simulated and measured flow rate for the outlet.

7.

The leak flow rate value M (Leak) is calculated by using the following equation:

$$M\left(Leak\right)=X-Y$$

(3)

8.

The leak position value in meters (m) X (Leak) is calculated by using the following equation:

$$X\left(Leak\right)=(-Y/M) \ast Le+Nu \ast Le$$

(4)

where the parameter Le is the pipeline segment length, and Nu refers to the number of LoRa nodes. The RTTM continuously utilizes the outcomes of various sensor measurements, such as the pressure, temperature, and flow rate; these are applied in a series of equations on the LoRa server to provide the necessary values for the leak calculations. This scenario validates a successful deployment of real-time LoRa technology selected for the proposed system due to its efficiency in terms of low power consumption and long-range communication and its presentation of distinct benefits over competing technologies [54].

Table 3. Sample of the output of the OMNeT++ oil-pipeline-based LoRa network monitoring system.

Node No	Time (s)	M In (kg/s)	Pt In (Bar)	Temp In (c)	M Out (kg/s)	Pt Out (Bar)	Temp Out (c)	M (Leak) (Kg/s)	X (Leak) (m)
12	18, 6.78	145.33	59.99	39.96	145.33	59.87	39.64	0	0
29	60,691.78	255.66	59.97	39.91	137.25	59.74	38.88	118.40	58,136.38
59	35,146.78	512.34	59.89	39.94	132.41	59.18	38.87	379.93	118,068.43
99	86,401.78	146.85	59.99	39.85	142.10	59.87	38.55	4.74	199,360.54

presents a sample of the output of the OMNeT++ oil-pipeline-based LoRa network monitoring system,

Table 4. General configuration of the OMNeT++ LoRa network.

Parameters	Value
Energy Detection	−11 dBm
Number of Nodes	100
Number of Gateways	10
Node Spacing	2000 m
Gateway Coverage	20,000 m
Simulation Time	1 day
Initial Spreading Factor (SF)	12
Initial Transmission Power (TP)	14 dBm
Initial Bandwidth (BW)	125 kHz
Initial Coding Rate (CR)	4
General.sigma	3.57
Ipv4Delay.config	Cloud Delay.xml

explains the basic configuration of the OMNeT++ LoRa network, and Figure 2 shows the network design of the proposed system.

4.2. Random Forest Model

In this section, we will present the random forest algorithm as an important machine learning technique for building a prediction model to enhance the LoRa network’s performance. Ensemble learning is a machine learning approach that combines multiple individual models to build a more influential predictive model. The essential concept is that by combining the predictions of several models, the ensemble process can achieve better accuracy and strength than the single model when processed on its own.

The process of data transmission in the LoRA network infrastructure of the oil pipeline monitoring system happens when LoRa endpoint devices send information to the server via separate gateways within the oil pipeline. The random forest model is connected with the FLoRa server to process the LoRa network parameters, which are related to the flow of incoming data and improve the network output performance depending on the trained dataset, which is created by collecting the LoRa network parameters from OMNeT++ by running the LoRa network in multiple scenarios with different network transmission parameters and recording its corresponding output performance for seven continuous days.

The main goal of using the random forest algorithm is to predict and improve the network output performance by implementing the regression task based on building multiple decision trees and combining them to obtain a more accurate and steady-state prediction. To achieve the best performance, the random forest model is adjusted with 100 decision trees, the maximum depth of the trees is 15, the minimum number of samples required to split an internal node is 5, and the minimum number of samples per leaf to adjust the minimum sample number required at a leaf node is 2. Efficiently, it obtained the underlying patterns in the data, providing a hard baseline for the proposed model.

Table 5. Summary of dataset features and components.

Parameter	Value/Range	Unit	Description
Number of LoRA gateways	10	-	The total number of LoRa gateways in the simulation.
Number of LoRa nodes	100	-	The total number of LoRa nodes in the simulation.
LoRa server	1	-	Number of LoRa servers managing the network.
Spreading factor (SF)	7, 8, 9, 10, 11, 12	-	Used to adjust the energy consumption and transmission range.
Payload (PL)	26–130	Bytes	Size of the data transmitted by LoRa nodes.
Bandwidth (BW)	125	kHz	Communication bandwidth.
Transmission power (TP)	11, 14, 20	dBm	Transmission power used by LoRA nodes.
Coding rate (CR)	4/5	-	Error correction rate used during the transmission.
Noise figure (NF)	6	dB	Impact of noise on communication quality.
Low data rate optimization	0, 1	-	Communication optimization at lower data rates.
Packet energy consumption	0.01–0.2	Joules	Total energy consumed for transmitted packets.
Power consumption during the idle state	0.001–0.05	mW	Consumption of power by each node in the idle state.
Signal-to-noise ratio range for training	−19 to −6	dB	Signal-to-noise ratio ranges utilized for training tasks.
Received signal strength indicator	−120 to −80	dBm	Calculated at each node to evaluate the signal strength.
Queuing time	5–50	ms	The waiting time spent in the queue until a packet is transmitted.
Number of collisions	0–5 per second	Count	Number of packet collisions affecting retransmission energy consumption.
Packet loss	0–10%	Packets	The ratio of packet loss during transmission.
Data rate	0.3–50	kbps	The data rate used by LoRA nodes.
Distance range for training	110	km	The distance range within which nodes communicate during training.
Bit error rate	1 × 10−14–1 × 10−5	-	Bit error rate for estimating the level of error during training.
Number of iterations	7 × 106	-	The count of iterations achieved during model training.
Signal-to-noise ratio threshold	−7.5, −10, −12.5, −15, −17.5, −20	dB	Signal-to-noise ratio threshold utilized for transmission decisions.
Maximum payload size	242, 115, 51	Bytes	The maximum payload size that can be handled by the LoRa node.
Transmission delay	5–100	ms	Time delay from transmission to receipt operations.
Retransmission count	0–3	Count	Number of times packets are retransmitted due to collision or error.

explains the features of the dataset’s components. The mathematical approach is described in the following equations.

$$\widehat{y}={\displaystyle \frac{1}{M}}{\Sigma }_{m=1}^M{T}_m(x)$$

(5)

where

• M is the number of trees in the forest;
• T_m (X) is the output of each decision tree for the input (x);
• $\widehat{\mathrm{y}}$ is the predicted output.

$${PC}_{pred}={\displaystyle \frac{1}{M}}{\Sigma }_{m=1}^M{T}_m(XPC)$$

(6)

where ${\mathrm{P}\mathrm{C}}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{d}}$ refers to the value of the predicted power consumption, M is the number of trees in the forest, Tm is the output of each decision tree, and XPC represents the input feature vectors of each decision tree related to power consumption.

$${PD}_{pred}={\displaystyle \frac{1}{M}}{\Sigma }_{m=1}^M{T}_m(XPD)$$

(7)

where ${\mathrm{P}\mathrm{D}}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{d}}$ refers to the value of predicted delayed packets, M is the number of trees in the forest, Tm is the output of each decision tree, and XPD represents the input feature vectors of each decision tree for the input (x) related to the packet delay.

$${PL}_{pred}={\displaystyle \frac{1}{M}}{\Sigma }_{m=1}^M{T}_m(XPL)$$

(8)

where ${\mathrm{P}\mathrm{L}}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{d}}$ refers to predicted lost packets, M is the number of trees in the forest, Tm is the output of each decision tree, and XPL represents the input feature vectors of each decision tree for the input (x) related to the packet loss [44].

4.3. Deep Extreme Learning Machine Model

In this section, we will explain the effect of deep learning algorithms on the enhancement of LoRa network parameters and try to obtain better performance than that obtained with the random forest model. Deep extreme learning machine (DELM) is a branch of machine learning that presents the ability to analyze complex patterns and find the relationships within data patterns more accurately and faster than traditional machine and deep learning techniques because it utilizes random weights during the initialization of the hidden layer, requiring no reiterative backpropagation for training, making it well suited for optimizing and predicting network performance.

A DELM model is deployed to perform analysis on the dataset, which includes the incoming data packet information and the LoRa network parameters. By processing the features in parallel advanced neural network architectures, deep learning algorithms can intelligently regulate node power levels based on previous experience and different factors, e.g., traffic volume, network congestion, transmission power, bandwidth, the spreading factor, coding rate, queuing time, data rate, and the proximity of other devices, as illustrated in Table 4.

This intelligent power management strategy minimizes energy waste and extends node battery life, in addition to enhancing the overall network efficiency. The input layer, which contains 25 densely connected neurons, receives data representing various network parameters. These data are then passed through a sequence of hidden layers where patterns in the input data are identified and extracted. The first hidden layer and second hidden layer act as the root of the network architecture, where complex transformations and computations occur to optimize and enhance the network parameters within these layers.

The DELM model learns complicated relationships between input features and target outputs. Deploying the tanh activation function, it is an innovative technique used to transform the network’s inputs, improving the ability of the model to catch complex patterns without repetitively adjusting the internal weights and biases, leading to a quick process and accuracy in the desired results.

After multiple stages of testing and tuning the DELM neural network, we obtained the best results with two hidden layers—the first with 50 neurons and the second with 35 neurons. The output layer, involving three neurons, provides the final prediction for the network enhancement parameters, and this is based on the most accurate processed input data. Figure 3 describes the DELM neural network structure. The mathematical approach of the DELM model is described in the following equations [55].

$$H_1=\sigma (W_{1}X+b_1)$$

(9)

where

• W1 is a weight matrix that is randomly initialized for the first hidden layer;
• X = {${\mathrm{x}}_1$.${\mathrm{x}}_2$…, ${\mathrm{x}}_{\mathrm{n}}$};
• b1 is the bias vector;
• σ is the tanh activation function.

$$H_2=\sigma (W_2{H}_1+b_2)$$

(10)

where

• W2 is a weight matrix that is randomly initialized for the second hidden layer;
• ${\mathrm{H}}_1$ is the output of the first hidden layer, and b2 is the bias vector.

$$\beta ={H^{+}}_{2}Y$$

(11)

where

• β is the output weight matrix;
• ${{\mathrm{H}}^{+}}_2$ is the Moore–Penrose pseudoinverse of the second hidden layer output;
• Y includes the actual values of node power consumption, packet delay, and loss coming from the dataset.

The DELM model regulates its weights for each metric by training with these specific target values to obtain accurate predictions in each case.

$$Ypred=f(H_2 \ast \beta )$$

(12)

where

• ${\mathrm{H}}_2\mathrm{i}\mathrm{s}$ the second hidden layer’s output;
• $\mathrm{f}$ is the linear activation function applied for the output layer;
• Y pred is the predicted output, representing the predicted enhanced network parameters.

We used the mean square error as the cost function to measure the difference between the predicted and actual values as a measure of the model’s accuracy [54].

$$\mathrm{MSE}={\displaystyle \frac{1}{\mathrm{n}}}{\mathsf{\Sigma }}_{\mathrm{i}=1}^{\mathrm{n}}({{\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}})}^2$$

(13)

where

• ${\mathrm{y}}_{\mathrm{i}}$ is the actual value of the network parameter;
• $\widehat{{\mathrm{y}}_{\mathrm{i}}}$ is the predicted value from the model;
• $\mathrm{N}$ is the number of data points in the dataset.

The mathematical approach used to calculate the predicted enhanced parameters of the LoRa network are described below:

$${PC}_{pred}=f\left(H_2\beta \right)$$

(14)

where f represents the linear activation that guarantees continuous predicted output, $H_2$ is the final hidden layer’s output, and β represents the calculated output weights depending on the training data for power consumption.

$${PD}_{pred}=g\left(H_2\beta \right)$$

(15)

where g represents the linear activation function for the output layer, $H_2$ is the final hidden layer’s output, which is related to the predicted packet delay, and β represents the calculated output weights depending on the training data for the packet delay.

$${PL}_{pred}=h\left(H_2\beta \right)$$

(16)

where h is the linear activation function used to predict the continuous packet loss, ${\mathrm{H}}_2$ is the output of the final hidden layer, which is related to the predicted packet loss, and β represents the output weights, which are calculated depending on the training data for packet loss.

4.4. Hybrid DELM Model

The hybrid DELM model combines the deep extreme learning machine model and the random forest model. It utilizes the DELM’s ability to learn complex nonlinear relationships due to its deep behavior and flexibility in neuron configuration, with fast processes according to its feature of random weight initialization and the strength of the random forest technique to decrease overfitting by bagging and averaging multiple decision trees, making it strong with noisy data and effective for high-dimensional datasets. The random forest model efficiently obtains the underlying patterns in the data, providing a hard baseline for the proposed model. Simultaneously, the DELM was employed due to its rapid learning abilities, employing a two-hidden-layer structure, with neurons activated by the hyperbolic tangent activation function, which smoothed the modeling of complex relationships among the nonlinear features. The predictions that were achieved from both models were averaged to formulate a comprehensive output.

The hybrid approach achieved a great reduction in the mean squared error (MSE) to 0.9, indicating its effectiveness in enhancing the accuracy of predictions for LoRa network performance in terms of power consumption, packet loss, and packet delay, depending on an accurate understanding of the complex relationships between the selected input features, such as the spreading factor, transmission power, coding rate, and bandwidth, which led to a perfect enhancement.

Finally, the proposed model is integrated with a random search optimization technique, and they work together to select the best LoRa network node parameters, such as the spreading factor, bandwidth, coding rate, and transmission power, and the LoRa server is informed to reconfigure the LoRa node parameters. However, the hybrid DELM presents some limitations in different aspects, such as increasing the complexity by combining two models as one unit, which leads to more computational processing and tuning, and it may require intensive memory and more execution time to deploy.

The hybrid DELM approach is absolutely valuable for systems with complex and nonlinear relationships, especially when the accuracy and processing time are critical, such as in wireless sensor networks or real-time monitoring applications. The integration of these advanced machine learning techniques not only improved the LoRaWAN resource utilization but also contributed to the improvement of the overall performance of the monitoring system. The pseudocode of the hybrid DELM model is provided in Algorithm 1. Figure 4 describes the workflow of the DELM model and how it can be integrated with LoRaWAN to enhance its performance. The mathematical equations that were employed to evaluate the predicted network performance are described in the following equations:

$$PC={\sum }_{i=0}^{N}P\left(i\right)$$

(17)

where

• $\mathrm{P}\mathrm{C}$ is the total power consumption in the network;
• $\mathrm{P}\left(\mathrm{i}\right)$ is the power consumption of node (i);
• $\mathrm{N}$ is the total number of nodes in the network.

$$PD={\displaystyle \frac{\sum _{i=0}^{N}T\left(i\right)}{N}}$$

(18)

where

• $\mathrm{P}\mathrm{D}$ is the average packet delay;
• $\mathrm{T}\left(\mathrm{i}\right)$ is the delay of the packet (i);
• $\mathrm{N}$ is the total number of nodes in the network.

$$PL={\sum }_{i=0}^{N}L\left(i\right)$$

(19)

where

• $\mathrm{P}\mathrm{L}$ is the total packet loss;
• $\mathrm{L}\left(\mathrm{i}\right)$ is the number of lost packets of node (i);
• $\mathrm{N}$ is the total number of nodes in the network.

$$ER={\sum }_{i=0}^N\left({\displaystyle \frac{AV\left(i\right)-PV\left(i\right)}{AV\left(i\right)}}\right) \ast 100$$

(20)

where

• $\mathrm{E}\mathrm{R}$ is the total percentage of the model enhancement ratio;
• $\mathrm{A}\mathrm{V}\left(\mathrm{i}\right)$ is the actual value of node i;
• $\mathrm{P}\mathrm{V}\left(\mathrm{i}\right)$ is the predicted value of node i.

Algorithm 1: Pseudocode of the proposed hybrid DELM model

Input: Dataset with network parameters (SF, TP, BW, CR, PC, PD, PL, etc.) Output: Predicted power consumption, packet delay, packet loss

1. Initialization 1.1 Import necessary libraries (pandas, NumPy, scikit-learn, random forest, hpelm, matplotlib). 1.2 Load the dataset. 1.3 Initialize the target network parameters (power consumption, packet delay, packet loss). 2. Data Preprocessing and Feature Extraction 2.1 Drop unwanted columns from the dataset. 2.2 Data filtering based on specific conditions (deleting rows with zero values in key parameters). 2.3 Filling missing values in related columns. 2.4. Divide the dataset into two sets: training and testing. 2.5 Extract relevant features (SF, TP, BW, CR, etc.). 2.6 Define target variables for optimization (power consumption, packet delay, packet loss). 3. Model Architecture 3.1 Explain input layer features. 3.2 Set two models for the hybrid approach: Random forest model for the baseline point of prediction. DELM model for achieving enhanced prediction. 3.3 Set up the random forest model with optimal hyperparameters (100 trees). 3.4 Initialize the DELM model with two hidden layers: Add the first hidden layer: 16 neurons; Add the second hidden layer: 8 neurons (main enhancement layer); Add an output layer with 1 neuron (predicted enhancement). 4. Hybrid DELM Model Training 4.1 Train the random forest model on the training data. 4.2 Apply bootstrapping on the training data. 4.2.2 Build decision trees based on the split features. 4.2.3 Aggregate predictions depending on individual trees (bagging). 4.3 Generate baseline predictions for power consumption, packet delay, and packet loss. 4.4 Assess the baseline performance using cross-validation. 4.5 Initialize weights and biases randomly in the DELM model. 4.6 For each cross-validation fold: 4.6.1 Train the DELM model on the training dataset. 4.6.2 Apply forward propagation on the input data to calculate predictions. 4.6.3 Validate the model with the testing set for each fold. 4.7 Combine the predictions from the random forest and DELM using an ensemble approach. 4.8 Average the outputs of both models by applying weighted averaging based on the validation results. 4.8 Select the combination of parameters (SF, TP, BW, CR) with the lowest MSE as the optimized parameters for the best prediction. 5. Model Validation 5.1 Evaluate model performance using the mean squared error. 5.2 Compare the predicted and target values to calculate the enhancement ratio. 6. OMNeT++ simulation integration 6.1 Connect the OMNeT++ simulation with the hybrid DELM model. 6.2 For each data packet in real time: 6.2.1 Extract features. 6.2.2 The hybrid model predicts the output depending on the optimized parameters. 6.2.3 Update the network with the optimized parameters (SF, TP, BW, CR). 7. Output 7.1 Display the enhanced network parameters. 7.2 Save results for analysis.

5. Results and Discussion

In this section, we will discuss the output performance results of the wireless LoRa network for the oil pipeline leak detection and localization system in different scenarios. Although the proposed RTTM-based LoRaWAN monitoring system achieved the desired objective with the default configuration of the LoRa network in terms of detecting the leak quantity and location, in addition to providing real-time monitoring of oil pipeline status, such as the pressure, flow rate, and temperature in an accurate manner, we were motivated to maintain the system’s efficiency over time so that it could work with the same efficiency by enhancing the node energy consumption and prolonging the battery life by utilizing the artificial intelligence techniques. The implementation of machine learning techniques presents a great solution that utilizes the strength of prediction, hyperparameter tuning, and optimization to enhance the output of the LoRa network with the best combination of network parameters, such as the spreading factors, transmission power, bandwidth, and coding rate.

Firstly, the default LoRa network was run, according to our proposed design, with 100 LoRa nodes within the OMNeT++ simulation. The results were tested and analyzed to calculate the total network output performance with network parameters of SF = 12, TP = 14, BW = 125 kHz, and CR = 4. The second scenario involved implementing the random forest regression model on a real-time dataset that was collected from the LoRa network. The total predicted LoRa node power consumption was enhanced by 19% compared with the default LoRa network performance with suggested network parameters of SF = 12, TP = 12, BW = 125 kHz, and CR = 3, with an accuracy of 83 in terms of the MSE.

Thirdly, the DELM model was executed, and the total enhancement ratio of the LoRa node power consumption with the proposed parameters of SF = 11, TP = 10, BW = 250 kHz, and CR = 3 was calculated, reaching 29% compared with that of the default LoRa network performance, and the model accuracy decreased to 27. Finally, the hybrid DELM model was implemented, and a perfect enhancement ratio was achieved with network parameters of SF = 9, TP = 8, BW = 250 kHz, and CR = 2 for the total predicted LoRa node power consumption, reaching 39% with a model accuracy of 0.75. We also analyzed the other output results of OMNET++, such as time period intervals and the number of transmitted packets, to calculate the node packet delay and packet loss and to evaluate the network performance. We obtained different enhancement ratios for packet delay and packet loss by applying the proposed machine learning models; with the hybrid DELM model, we achieved 33% and 31% for the total packet delay and packet loss, respectively.

All the results are described in

Table 6. Brief comparison of LoRa network performance.

Parameters	Default LoRa	Random Forest	DELM	Hybrid DELM
Total Power Consumed (mWh)	4332	3509	3033	2686
Packet Delay (ms)	85,472	70,941	63,249	57,266
Total Packet Loss	502	427	387	346
Power Consumption Enhancement Ratio	-	18%	29%	39%
Packet Delay Enhancement Ratio	-	17%	26%	33%
Packet Loss Enhancement Ratio	-	14%	23%	31%

, as shown below. The final results reveal that the hybrid DELM model presents a perfect enhancement of LoRa network performance.

Table 7. Performance analysis of LoRa node no. 3.

Model	SF	TP	BW	CR	Power Consumed (mWh)	Packet Delay (ms)	Packet Loss
Default LoRa	12	14	125	4	43.18	85,448	73
Random Forest	12	12	125	3	34.98	70,921	62
DELM	11	10	250	3	29.97	63,231	56
Hybrid DELM	9	8	250	2	26.37	57,250	50

explains the effect of the proposed models on the LoRa node performance; node no. 3 was randomly selected for presentation. Figure 5 shows the enhancement of LoRa network performance by applying the hybrid DELM in comparison with the DELM model, random forest, and default LoRa network.

6. Conclusion and Future Work

In this research paper, an integrated oil pipeline monitoring solution based on wireless sensor networks was implemented. LoRaWAN was selected as the network infrastructure for the proposed model. The objective of this work was focused on two approaches; the first was the implementation of a real-time oil pipeline monitoring system by utilizing a real-time transient model with LoRaWAN technology to achieve high accuracy in terms of detecting the leak quantity and location while taking advantage of the abilities of LoRa to implement the proposed system with the fewest hardware resources. Second, different machine learning techniques were deployed to enhance the LoRa network performance, including the random forest algorithm, which enhanced the LoRa node performance by 18%, 17%, and 14% in terms of power consumption, packet delay, and packet loss, respectively, compared with the results for the default LoRa network. The DELM model achieved better results, with enhancement ratios of 29%, 26%, and 23%, respectively.

Finally, the hybrid DELM model was selected due to its effectiveness and efficiency in achieving the objective of the study; it achieved a LoRa node power reduction of 39%, with a model accuracy of 0.75. Looking ahead, further research and efforts should be focused on developing the hybrid DELM model to enable it to overcome some limitations, including the computational process, execution time, and intensive memory requirements, in order to achieve more advanced network optimization tasks.