Online Learning Framework for Radio Link Failure Prediction in FANETs

In this paper, we consider the problem of prediction of Radio Link Failures (RLF) in flying ad hoc networks (FANETs). Many environmental factors that influence the quality of radio wave propagation are dynamic, and thus, drones must continually learn and update their radio link quality prediction model while they operate online. Online machine learning algorithms can be used to build adaptive RLF predictors without requiring a pre-deployment effort. To predict the RLF, we use an online machine learning algorithm and information gathering by message-passing from the neighbors. We propose an algorithm called ML-Net (Machine Learning and Network algorithm) to predict RLF. To the best of our knowledge, the combination of online machine learning algorithms together with the message-passing algorithm has not been used before. The proposed methodology outperforms the state-of-the-art online machine learning algorithms.


I. INTRODUCTION
U NMANNED aerial vehicles (UAVs), also known as drones or flying robots, have gained significant attention in various real-life applications.The flying ad hoc network (FANET) is established to leverage high-speed communications.However, due to the high mobility of UAVs in FANETs, the network topology may continuously change, making it challenging to establish end-to-end connections.Radio Link Failure (RLF) prediction can help UAVs handle this issue and improve FANET performance to ensure continuous service availability.Accurate prediction of radio link failures (RLF) is critical for ensuring reliable communications in flying ad hoc networks (FANETs).However, link prediction remains challenging due to the dynamic topology and unpredictable mobility patterns in FANETs.Nodes can move in and out of communication range rapidly, leading to frequent link disruptions.The ability to accurately predict impending link failures can enable proactive mitigation strategies.For example, drones could switch to more reliable links ahead of time to prevent packet loss and service interruptions.Link prediction also allows optimizing routing by avoiding unstable links that are about to fail.Furthermore, timely knowledge of upcoming link losses enables adapting transmission parameters to maintain connectivity.The development of link prediction techniques tailored to the FANET environment is therefore essential for efficient network operation and robust aeronautical communications.Machine learning holds promise for developing adaptive, data-driven predictors that can operate in real-time based on local interactions.This motivates our exploration of online learning combined with message passing for highaccuracy RLF prediction in FANETs.
Two essential characteristics should be present in RLF predictors.First, adaptivity is critical since an RLF predictor must cope with quality fluctuations over time, especially for FANET.Second, plug-and-play is crucial since RLF predictors should be applied without requiring any predeployment effort, which might not be feasible for all deployment scenarios, even if the effort is reduced.Online machine learning algorithms can be used to build RLF predictors that are adaptive without requiring predeployment effort.
Most current machine learning approaches are limited to the traditional batch setting, where data is provided in advance to the training process.Model selection and meta-parameter optimization can rely on the full set of data, and training can assume that the data and its underlying structure are static.In contrast, online machine learning involves continuous model adaptation based on constantly arriving data.The underlying distribution of the data, which changes over time, presents some primary challenges and difficulties in the dynamic environment.Old data can become irrelevant or even detrimental to model the current concept.Online machine learning [3] exhibits great potential for performance improvement with the sequential arrival of data and superiority over offline learning, including real-time predictions and lower memory requirements.
This paper explores the possibility of using online machine learning algorithms together with local information gathering by messages to predict RLF.Each drone uses only its local in-formation and information gathered from its neighbor.Specifically, the main contributions of this paper are as follows.
1) Motivated by the characteristics of RLF predictors mentioned above, we propose a link failure prediction model ML-Net, which combines the online machine learning algorithm with a message-passing algorithm.2) To the best of our knowledge, we pioneer the use of the online machine learning method combined with the message-passing algorithm for RLF prediction.3) We conduct simulations to analyze the performance of the ML-Net.We compare the proposed solution with state-of-the-art techniques and perform an analysis of simulation results.The analysis shows that the proposed approach ML-Net achieves better performance than the state-of-the-art methods.The remainder of this paper is organized as follows.Section II discusses recent related studies.Section III describes the theoretical model used to define our problem.The details of the proposed approach are described in Section IV.In Section V, we describe all aspects of the evaluation setup and simulation results.Finally, we summarize our paper with conclusions and suggestions for future research in Section VI.

II. RELATED WORK
In this study, we propose the use of online machine learning algorithms combined with message-passing approach for predicting Radio Link Failure (RLF) in communication networks.Several recent studies on RLF suggest using online machine learning algorithms for prediction.
In [1], the authors studied the link quality prediction for wireless mesh networks.They performed a performance analysis of four state-of-the-art algorithms for link quality prediction and proposed a new hybrid online algorithm for link quality prediction based on this analysis.
The authors of [7] presented an adaptive link estimator (TALENT) that uses online learning.They argued that TALENT adapts to network dynamics better than statically trained models without the need for advance data collection for training the model before deployment in a real system.The solutions follows the performance of the autoencoders very closely with a tiny margin on very bad, very good and intermediate link quality classes.
The study [8] introduced a framework to reduce energy and network capacity overhead expenses by incorporating active learning to selectively label only a portion of the samples from the data stream.The framework also uses incremental training batches to conserve labeling resources and updates the batches using change detection and forgetting mechanisms to mitigate concept drift.Experimental results showed that the framework reduces label queries by up to 21.5% and prediction error by up to 9% after periods of concept drift.
As a part of their work in [12], the authors looked at the problem of predicting channel quality between vehicles in terms of path loss, which shows strong fluctuations over time due to the highly dynamic nature of vehicular environments.
They proposed a framework for a data-driven path loss prediction model that combines the changepoint detection method and online learning.The evaluation of the proposed framework was done using real-world datasets.
Machine learning methods were used in [10] to predict the short-term evolution of link quality for switching to a better link for data transmission.The problem was modeled as a game of prediction based on experts' advice, using the Link-Quality Indicator (LQI) metric.A decision-maker predicts the LQI values, called a forecaster, who receives advice from several experts.To predict values close to actual LQI values, the forecaster can learn how to adapt its strategy.As a general model, the proposed learning and prediction model can be easily adapted to different link-quality metrics or prediction methods.
In summary, some recent works have used online machinelearning algorithms to predict RLF and took into consideration the adaptability issue, see [2].Based on these previous studies, we propose the use of a novel method based on online machine learning with a combination of neighbor information gathering by a message-passing solution.

III. SYSTEM MODEL AND PROBLEM FORMULATION
We consider a set of n UAVs (denoted as S) that are deployed arbitrarily in a given area and may be located in any position in R 3 .The network imposed on S is connected.
We assume that the transmission power of all nodes is fixed.We denote the transmission power of node i at time t as P t i .The set of nodes i's neighbors at time t depends on the current positions and channel gain.Based on [13], we can approximate the path loss on link ij as follows: where G l is the product of the transmitter and receiver antenna field radiation patterns of LOS transmissions, d t is the distance between a transmitter and its corresponding receiver at time t, and λ is the operating wavelength.Thus, the path loss is related to the transmission distance when transmitting radio signals over a specific channel through a specific antenna.Let N t i be the set of neighbors of node i at time t.Formally, N t i is defined as: where θ is a predefined parameter of the network.A network is defined as a directed graph G t = (S, E t ), where node i ∈ S has a directed edge (link) to each of its neighbors according to Equation 2. The edge set E t is the union of all directed edges among the nodes in time t, E t = i∈S E t i .At each timestamp t, when node i receives a message from node j, node i collects a set x t of metrics (features) that describe the communication with node j at time t.Node i adds x t to a multivariate time series that represents features of communication with node j, denoting this multivariate time series as X t ij = {x 1 , . . ., x t }.Now we can formally define our problem.Given the current X t ij , we wish to determine whether the edge e ij will fail or not in the following timestamp (namely, at t + 1).

IV. ML-Net
In this section, we describe the proposed ML-Net approach.We combine the machine online learning algorithm with the network structure algorithm to cope with RLF prediction in the next timestamp.We suggest using the expression S(X t ij , m t i←j ) which approximates the probability that the edge e t ij will fall in the next timestamp.
where 0 ≤ α ≤ 1, m t i←j is the message sent by j to i including the belief of node j about v t ji , and P r(v is the probability that edge e ij will fail in timestamp t + 1 given multivariate time series X t ij .We use on the shelf online machine learning algorithm to calculate P r(v t+1 ij = −1|X t ij ) (in Section V-A we will broadly describe the proposed algorithms).
Due to the fact that the G t = (S, E t ) is a directed graph, then the following scenario may occur: the edge e t i,j exists but the edge e t j,i does not exist.In this scenario, m t i←j may arrive but m t j←i cannot arrive.In Figure 1 we can see an example of this scenario.Figure 2 showcases a random graph with 10 nodes (A-J), representing entities in a system.The directed edges between nodes symbolize the flow of messages, demonstrating the operation of the message-passing algorithm.The labels on the arrows, m t ij , represent the messages passed from node i to node j at time t.This graph serves as a visual representation of the algorithm's functioning in a system with multiple interacting components.if msg r eceived(j, t) then 6: end if The ML-Net algorithm shows the key steps for combining online machine learning with message passing to predict link failures.For each neighbor j, node i first checks if a message was received from j at time t.If so, i updates its belief about the i → j link failing using both its own prediction and j's belief from the message.This allows propagating information through the network.If no message is received, i relies only on its own link prediction.Finally, the algorithm returns the combined prediction score S for each link i → j based on the updated beliefs.This demonstrates how ML-Net leverages both local online learning models and information exchange with neighbors to achieve accurate real-time RLF prediction in dynamic FANET environments.

A. Notations
The notations used in this paper are summarized in Table I.Also, Table I I: Summary of notations used in this paper V. EXPERIMENTAL SETUP Simulation is the most reliable and cost-effective approach to examining the performance of real-world problems.For performance evaluation of the proposed technique, OMNET++ [15] is used as the primary tool for FANET environment generation and routing protocol implementation.This section describes the data generation process.Let K represent the set of the source and destination pairs within the network that should communicate.
We simulated the network for |S| = 20 and for |K| = 3.In Figure 3 we can see an example of a network with 20 UAV's and |K| = 3.
The routing protocol we chose to use is AODV (Ad-Hoc On-Demand Distance Vector) [14].This protocol is designed for wireless and mobile ad hoc networks and has been broadly adopted for FANETs.Next, we present the schemes for comparison and the evaluation metrics.

A. Schemes for Comparison
In the following, we briefly describe the state-of-the-art online machine learning algorithms and their hyperparameters used in this paper to compare with ML-Net.These algorithms were selected as the comparison baseline solutions due to their advanced nature.
1) The Extremely Fast Decision Tree (EFDT) [9]: EFDT constructs a tree incrementally.The EFDT seeks to select and deploy a split as soon as it is confident the split is useful and then revisits that decision, replacing the split if it becomes evident that a better split is available.The EFDT learns rapidly from a stationary distribution, and eventually, it learns the asymptotic batch tree if the distribution from which the data are drawn is stationary.We chose a EFDT with the following hyperparameters: grace period of 200 and split confidence of 1e −7 .2) Adaptive Random Forest (ARF) [4]: ARF algorithm for the classification of evolving data streams enables the Random Forests algorithm for evolving data stream learning.There are effective resampling methods and flexible operators in ARF that can cope with different types of concept drifts without requiring complex optimizations for different data sets.We chose a ARF with the following hyperparameters: grace period of 50, lambda is 6, max features are 3, split confidence of 0.01, and tie threshold of 0.05.
3) Hoeffding Tree Classifier (HTC) [6]: This method involves determining an upper bound for the learner's loss based on the number of examples used in each step of the algorithm.In this algorithm, the number of examples required for each step is minimized while ensuring the model performance is not significantly different from the one obtained from an infinite dataset.We chose a HTC with the following hyperparameters: grace period of 200, min samples reevaluate of 20, split confidence of 1e −7 , and tie threshold of 0.05.[5]: The Streaming Random Patches algorithm is a machine learning technique that involves randomly selecting subsets or "patches" of data from a continuous stream of input.These patches are used to train and update a model iteratively over time, allowing the model to adapt and learn from new incoming data.Authorized licensed limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

B. Evaluation Metric
We use the following performance metrics to evaluate the performance of different machine learning algorithms: 1) Precision: The ratio of the number of true positives over the total number classified as positive.In the context of our problem, it is the ratio of correctly predicted link failures versus the total number of link failures predicted.The precision value is computed as follows: where P is the precision value, T P is the number of "true positives," and F P is the number of "false positives." 2) Recall: The ratio of the number of data points associated with link faults correctly classified over the total the number of data points associated with link faults that have occurred.The recall value is given by 3) F 1 -Score: The F 1 -Score is the harmonic average of the precision and recall values.It takes a value in [0, 1].Higher the value of F 1 -Score, the better the performance of the machine learning technique.It is computed as follows: 4) Cohen's kappa-κ : Cohen's kappa measures the agreement between two raters who each classify N items into C mutually exclusive categories the Cohen's kappa formula can be written as:

C. Features
This section provides an exploratory data analysis of key features in the communication system.Visualizations are utilized to understand feature distributions and relationships in the data.6) angle between: This feature represents the angle between the transmitter and receiver.The relative orientation between these devices can affect the signal strength and thus the link quality.7) Link: This is the target variable that we want to predict.
It is a binary variable representing the quality of the communication link, with '1' indicating a good link and '0' indicating a bad link.This allows us to frame the problem as a binary classification task.
These features provide a comprehensive description of each signal reception event, capturing information about the transmitter, signal quality, received signal power, and relative device orientation.This enables construction complex models to predict the communication link quality.
Figure 4 shows the histogram that presents the distribution of the logarithm of the received Power Reception.The x-axis shows the log power values, and the y-axis shows the frequency.The kernel density estimate smooth line approximates the probability density function.
The log transformation handles skewed data by making the distribution more symmetric.The log-transformed power distribution appears Gaussian, concentrated around -26.
The KDE peak indicates the most frequent log power value.The distribution spread provides insights into power variability.This transformation can help modeling techniques assume normality.
Figure 5 is a histogram that shows the distribution of angles between transmitters and receivers.The x-axis represents the angle from 0 to 180 degrees.The y-axis represents the frequency.This reveals common orientations in the simulation.
The overlaid Kernel Density Estimate (KDE) approximates the distribution.It assists in identifying shape characteristics like peaks.These insights are key to understanding angle variability and its potential impact on system performance.
The heatmap presented in Figure 6 illustrates feature correlations.Each cell shows the correlation coefficient between feature pairs.Cool colors indicate negative correlations.Warm

D. Evaluation and Numerical Results
We evaluate the proposed method using data collected from simulations implemented in OMNET++ [15].As part of our simulation, we randomly deployed 20 drones and randomly selected three pairs of source destinations from these drones as shown in Figure 3 for an example.The drones run the AODV routing protocol.
Each node participating in communication extracts the following parameters (features) when receiving a message from its neighbor: reception power, transmission power, SINR, modulation, location, and orientation.Features collection does not require particular messages.Simulation runs for 3000 seconds, during which the link failure is collected.
Table II summarizes the parameters used by us in the simulation.
We first used the schemes from Section V-A to predict RLF so that we could compare our method with their performance.The compression mechanism is implemented using [11].We calculate P r(v t+1 ij = −1|X t ij ) using all the schemes for  Figure 7 shows the performance of ML-Net as a function of α.The x-axis represents the time of the simulation, and the y-axis represents the accuracy of the RLF prediction.As we can learn from the figure, the accuracy of the RLF prediction generally increases as we increase the value of α.This is because a larger value of α gives more weight to the node's own information and less to its neighbor's information.We can also see that when α is very small, the algorithm is overly influenced by the neighbors' information, and the predictions may not be accurate enough.On the other hand, when α is very large, the algorithm may not be able to adapt quickly to changes in the network topology due to a lack of influence from the neighbors' information.Based on Figure 7, we chose the alpha value to be 0.6, as it strikes a balance between the node's own information and its neighbor's information and achieves the best accuracy value.We provide a detailed explanation of the performance analysis of ML-Net with different online learning algorithms in calculating P r(v t+1 ji = −1|X t ji).To evaluate the performance of our proposed approach, we compare it with other

Fig. 3 :
Fig. 3: Example of network with 20 nodes and 3 pairs of source and destination.

Fig. 7 :
Fig. 7: α selection for ML-Net Algorithm 1 ML-Net Algorithm 1: Input: Node i, current time t, set of neighbors N t

TABLE
contains further notation defined below in the paper.
ijEdge between node i to j in time t m t i←j Message sent by j to i (contains v t ji ) 1) Time: This feature represents the elapsed time since the start of the simulation or experiment.It captures the dynamics of the communication system over time.2) IdTransmitter: This categorical feature identifies the transmitter for each signal received.It allows the model to capture any specific characteristics related to individ-

TABLE II :
Simulation Configuration