Employing digital twin to LoRa based forest fire management systems

Abstract

Forest fires are one of the most critical risks threatening ecology. With the increasing average temperature of the world, the number and severity of forest fires have also increased. This has made early detection of forest fires essential. With developing technology, forest fires can be detected with the help of different techniques. Among these techniques, fire detection using sensor networks offers the most effective solutions in terms of both speed and cost. Thanks to the Internet of Things, wireless sensor networks can be established that regularly monitor an environment and collect this data through a large number of affordable internet-connected devices. It is also possible to establish a network with sensor nodes that measure temperature, humidity, etc. for the detection of forest fires. However, these networks have different requirements for each forest. Therefore, managing the networks can be complicated. Digital Twin technology is an innovative simulation alternative that facilitates the work to be done on this system by creating a real-time model of the physical system. It can be used in the optimization of systems, testing of different scenarios, and observation of the system. It can also provide significant benefits for the management of networks created for forest fires. However, it is quite difficult for such networks to provide the continuous and two-way communication demand of digital twins. Forest conditions are also effective in determining the communication technologies of the wireless network to be designed. Short-range and high-power radio modulation technologies are generally not preferred because of the large areas covered by forests and the fact that the geographical location of the forests makes it inconvenient to frequently maintain the devices in the network. Communication technologies developed for low-power wide area networks are ideal communication techniques for forest fire detection management systems. LoRa radio frequency modulation technology is one of the most widely used low-power wide-area network technologies. Nodes with LoRa modules are preferred in such networks due to their advantages of being able to transmit messages over distances of kilometers and providing years of battery life in battery-powered sensor devices. In addition, allowing unlicensed use and operating in unlicensed frequency ranges are also important reasons why it is often chosen in these networks. Detection of forest fires using computer networks is a subject that has been studied extensively in the literature. In previous studies, sensors such as temperature, humidity, carbon monoxide, etc. were used for fire detection. Although long-distance technologies such as LoRa are preferred, short-distance data transmission techniques such as Bluetooth and ZigBee have also been used. The topologies of the networks have been created with various techniques such as mesh, clustered, tree, and star. In addition, studies have been conducted using artificial intelligence to detect false alarms from sensors. Digital twins have the ability to quickly adapt to changes in the system because they create a real-time model of the physical system. They can be used with optimization algorithms to take the necessary actions for the changing conditions of the system. This technique can be groundbreaking for the optimization of dynamic systems such as computer networks. However, some common artificial intelligence techniques used in the development of twins have difficulty making sense of the relational structure in these networks where neighborhoods are important. For this reason, special digital twin technologies developed for computer networks are also called digital twin networks. These twins usually use graph neural networks to understand the relational structure of networks. Graph neural networks are a special data learning technique used to analyze data represented by graphs, such as protein structures, social networks, etc. Although it varies from model to model, the basic working principle of the technique is based on sharing the properties held in nodes and edges with neighboring nodes over several iterations to update their values. Graph neural networks are used in computer network studies in the literature, in routing optimization, network slicing management, and real-time modeling of some metrics of the network. In these studies, while transmitting messages to the twin network models and implementing new configurations, classical optimization techniques work together with deep reinforcement learning techniques for optimization. It has also been studied to use federated learning techniques for training digital twin models. This thesis study proposes adding a forecaster between the network and the twin to facilitate the application of digital twins to the Internet of Things networks. The problem that arises from the inconvenience of providing continuous data flow from these networks and the need for this flow, is overcome by the estimations it makes after the forecaster is trained with the data coming from the network. In order to train the digital twin model, data from situations that prevent data flow such as many collisions and interference is needed. Due to the difficulty of collecting this data from physical networks and the fact that forest fire detection networks consist of hundreds of nodes and cover square kilometers of area, the network was simulated with a custom simulator developed in this study. A clustered network topology was designed, LoRa technology was used for intra-cluster communication, and GPRS technology was used for inter-cluster communication. At the end of the simulation, all packets formed in the simulation were reported. The forecaster is a binary sequence prediction model that estimates whether a node sent a packet or not for each node and each time interval using the packets formed in the simulation. Since sensor nodes sleep most of the time, the undersampling technique was used to balance the dataset. Then, using these estimates, the overall state of the network was extracted for each time interval to be given to the digital twin. Using a graph neural network-based digital twin, the total output of the network was estimated using the network state estimates of the forecaster. Using simulation data, the actual output values were compared with the twin's estimates. The results are generally promising. Especially for small-scale networks, the system's $R^2$ performance is above 97 percent. The MSE values are well below 1. However, as the network scale increases, the system's performance decreases significantly. This is thought to be due to the cumulative accumulation of errors made by the forecaster in the estimation of each node. The system has also been tested with different digital twin and forecaster AI algorithms. Although the system performs similarly or slightly better with these models, it has been observed that this increase will not solve the scalability problem. As a result, integrating digital twins with the Internet of Things is impractical due to the difficulty of meeting the demands of the twins. This study aims to facilitate this integration by proposing a forecaster model. For this purpose, a simulation was designed and the performance of the system was tested. According to the results, it was concluded that this method is suitable, especially for small-scale networks. However, it has also been shown that the system has a scalability problem. In future work, a forecaster module can be designed to make a more holistic prediction of the network, the simulation parameters can be improved to expand the environmental and traffic throughput, and the digital twin can be made to predict the arrival percentages based on the importance of incoming packets instead of the network throughput.