Search

Artificial Intelligence (AI) has been used in different research areas in aerospace to create an intelligent system. Especially, an unmanned aerial vehicle (UAV), known as a drone, can be controlled by AI methods such as deep reinforcement learning (DRL) in different purposes. Drones with DRL become more intelligent and eventually they can be fully autonomous. In this paper, DRL method supported by real time object detection model is proposed to detect and catch a drone. Additionally, the results are analyzed by comparing the time to catch the target drone in seconds between DRL method, human pilot and an algorithm which directs the drone towards the target position without using any AI method or navigation and guidance method. The main idea is to catch a drone in an environment as fast as possible without crashing any obstacles inside the environment. In DRL method, the agent is a quadcopter drone and it is rewarded in each time step by the environment provided by Airsim flight simulator. Drone is trained to catch the target drone by using DRL model which is based on deep Q-Network algorithm. After training, the tests have been made by the agent drone with DRL model and human pilots to catch stationary and non-stationary target drone. The training and test results show that the agent drone learns to catch target drone which can be a stationary and a non-stationary. In addition. the agent avoids crashing any obstacles in the environment with a minimum success rate of 94%. Also, DRL model performance is compared with the human pilot performances and the agent with DRL model shows better time to catch the target drone. Human pilots struggle to control the drone by using remote controller when catching the target in simulation. However, the agent with DRL model is rarely missing the target when trying to catch the target. This work was funded partially by the AGAUR under grant 2020PANDE00141, the Ministry of Science and Innovation of Spain under grant PID2020-116377RB-C21 and the SESAR Joint Undertaking (JU) project CORUS-XUAM, under grant SESAR-VLD2 101017682. The JU receives support from the European Union’s Horizon 2020 research and innovation programme and the SESAR JU members other than the Union. Special thanks to Dr. Pablo Royo Chic for his assistance to our paper by piloting a drone to catch the target drone.

Under the light of increasing air traffic congestion and reduced environmental impacts, the growing need for tailored approaches has arisen. To coordinate tailored approaches, accurate predictions of descent trajectories must be available. The ability of ground based automation for such endeavours however must be improved. Specifically in the ability to predict where descents begin, also referred to as the Top of descent (TOD). This paper attempts to provide a machine learning approach to TOD location prediction, considering the operational capabilities of modern-day air traffic management by working within the confines of parameters available to controllers and considerations of stakeholders. To identify the common TOD clusters and the resultant average TOD locations representing the controllers macroscopic perspective, an unsupervised learning approach is utilised. Specifically, an ensemble method of DBSCAN-K-means is used. This was compared to the more microscopic pilot controllers perspective and generic decision-making formulas to determine the proximity to expected TOD. Of which the microscopic option has proven to be more reliable in representing the individual Flight Management System (FMS) recommendations or pilots’ decisions on aircraft. In the prediction of TOD location, operationally available factors such as proximity to expected TOD, current flight level, approach heading, ground speed, fleet and aircraft type are used to train a decision tree model. A Singapore flight information region (FIR) case study is conducted to illustrate the methodology. Results of the methodology employed on the inbound A380 fleet consisting a total of 79 flights, yielded a maximum prediction range of 31NM range from actual TOD occurrence. This paper has demonstrates the possibility of using a machine learning approach to predict TOD of a flight considering both prior cruise characteristics and operational considerations of pilots. Accepted version

The Publisher's final version can be found by following the DOI link. Data quality plays a crucial role in the context of smart buildings. Meanwhile, missing data is relatively common in acquired datasets from sensors within the smart buildings. Poor data could result in a big bias in forecasting, control and operational services. Despite the common techniques to handle missing data, it is essential to systematically select the most appropriate approach for such missing values. This paper aims to focus on the lift systems as one of the essential parts in the smart buildings by exploring the most appropriate data imputation methods to handle missing data and to provide its service and allow a better understanding of patterns to issue the correct control actions based on forecasted models. The imputed data is not only investigated statistically but also modelled through machine learning algorithm to explore the impact of selecting inappropriate imputation techniques. Seven imputation techniques deployed on datasets with three level of missing values including 10%, 20% and 30% and the performance of methods examined through the normalized root mean square error (NRMSE) approach. In addition, the interaction between imputation techniques and a machine learning algorithm, namely random forest were examined. Findings from this paper can be employed in identifying an appropriate imputation technique not only within the lift datasets, but smart building context.

Natural experiments are observational studies where the assignment of treatment conditions to different populations occurs by chance "in the wild". Researchers from fields such as economics, healthcare, and the social sciences leverage natural experiments to conduct hypothesis testing and causal effect estimation for treatment and outcome variables that would otherwise be costly, infeasible, or unethical. In this paper, we introduce VAINE (Visualization and AI for Natural Experiments), a visual analytics tool for identifying and understanding natural experiments from observational data. We then demonstrate how VAINE can be used to validate causal relationships, estimate average treatment effects, and identify statistical phenomena such as Simpson's paradox through two usage scenarios., 5 pages, 4 figures, accepted as short paper at IEEE VIS 2021

Accurate forecasting of solar power generation with fine temporal and spatial resolution is vital for the operation of the power grid. However, state-of-the-art approaches that combine machine learning with numerical weather predictions (NWP) have coarse resolution. In this paper, we take a graph signal processing perspective and model multi-site photovoltaic (PV) production time series as signals on a graph to capture their spatio-temporal dependencies and achieve higher spatial and temporal resolution forecasts. We present two novel graph neural network models for deterministic multi-site PV forecasting dubbed the graph-convolutional long short term memory (GCLSTM) and the graph-convolutional transformer (GCTrafo) models. These methods rely solely on production data and exploit the intuition that PV systems provide a dense network of virtual weather stations. The proposed methods were evaluated in two data sets for an entire year: 1) production data from 304 real PV systems, and 2) simulated production of 1000 PV systems, both distributed over Switzerland. The proposed models outperform state-of-the-art multi-site forecasting methods for prediction horizons of six hours ahead. Furthermore, the proposed models outperform state-of-the-art single-site methods with NWP as inputs on horizons up to four hours ahead., 10 pages, 7 figures, accepted for publication in IEEE Transactions on Sustainable Energy

Persistent challenges in making autonomous vehicles safe and reliable have hampered their widespread deployment. We believe that formal methods will play an essential role in the enterprise of ensuring AV safety by providing tools for the modeling, verification, synthesis, and runtime assurance of AV systems. In this paper, we outline the progress we and others have made towards this goal, and the challenges that remain.

This paper studies the demand-capacity balancing (DCB) problem in air traffic flow management (ATFM) with collaborative multi-agent reinforcement learning (MARL). To attempt the proper ground delay for resolving airspace hotspots, a multi-agent asynchronous advantage actor-critic (MAA3C) framework is firstly constructed with the long short-term memory network (LSTM) for the observations, in which the number of agents varies across training steps. The unsupervised learning and supervised learning are then introduced for better collaboration and learning among the agents. Experimental results demonstrate the scalability and generalization of the proposed frameworks, by means of applying the trained models to resolve different simulated and real-world DCB scenarios, with various flights number, sectors number and capacity settings. 10.13039/501100004543-China Scholarship Council

As the primary navigation source, GNSS performance monitoring and prediction have critical importance for the success of mission-critical urban air mobility and cargo applications. In this paper, a novel machine learning based performance prediction algorithm is suggested considering environment recognition. Valid environmental parameters that support recognition and prediction stages are introduced, and K-Nearest Neighbour, Support Vector Regression and Random Forest algorithms are tested based on their prediction performance with using these environmental parameters. Performance prediction results and parameter importances are analyzed based on three types of urban environments (suburban, urban and urban-canyon) with the synthetic data generated by a high quality GNSS simulator.

Automatically discovering a process model from an event log is the prime problem in process mining. This task is so far approached as an unsupervised learning problem through graph synthesis algorithms. Algorithmic design decisions and heuristics allow for efficiently finding models in a reduced search space. However, design decisions and heuristics are derived from assumptions about how a given behavioral description - an event log - translates into a process model and were not learned from actual models which introduce biases in the solutions. In this paper, we explore the problem of supervised learning of a process discovery technique D. We introduce a technique for training an ML-based model D using graph convolutional neural networks; D translates a given input event log into a sound Petri net. We show that training D on synthetically generated pairs of input logs and output models allows D to translate previously unseen synthetic and several real-life event logs into sound, arbitrarily structured models of comparable accuracy and simplicity as existing state of the art techniques for discovering imperative process models. We analyze the limitations of the proposed technique and outline alleys for future work., Comment: accepted at IEEE International Conference on Process Mining (ICPM) 2021, submitted version

The paper investigates one of the approaches based on machine learning methods aimed at finding and identifying similar disciplines. In the research we used two most popular methods of machine learning to process text data - BERT and Doc2Vec. Machine learning was conducted using the datasets of various disciplines with the total of 2,5 million entries. To assess the quality of the developed models, 30 experts from different scientific fields were engaged in the study to evaluate the level of similarity between the disciplines defined by the trained models. Based on the results of the research, both methods trained using identical datasets generated similar outputs. Another algorithm Doc2Vec, trained on a relatively small data sample with 15 000 entries of the target discipline database that included disciplines descriptions and curriculums, showed better results which justifies the need for developing specific solutions for particular tasks. Further development of machine learning methods and models design to solve specific tasks in the educational field will promote digitalization of education within the area of university operations management.

Intrusion detection systems are an important tool for network security. Their efficiency can be improved by implementing Alert Correlation Systems. Such systems are aimed at identifying relationships between alerts themselves and between alert and properties of protected systems. One of the tasks of alert correlation systems is to identify groups of attackers. Solving this task allows to improve the accuracy of determining the threat level of malicious actors and to determine patterns of similarity between attacks. These improvements help in choosing response measures and forensic investigation. To date, there is no universal dataset suitable for testing the effectiveness of any method related to intrusion detection systems, and the most appropriate dataset for the task of attacker group identification has not been selected. The paper considers the existing approaches to the formation of requirements for datasets for use in intrusion detection tasks, and analyzes modern datasets. A list of requirements for datasets is formed for their use in testing methods for identifying groups of attackers based on the specifics of the task. Weights are determined for the requirements, and a usability rating is determined for the modern datasets. An alternative data source is proposed to meet requirements that are poorly addressed by the current datasets.

The generalization performance of deep learning models for medical image analysis often decreases on images collected with different devices for data acquisition, device settings, or patient population. A better understanding of the generalization capacity on new images is crucial for clinicians' trustworthiness in deep learning. Although significant research efforts have been recently directed toward establishing generalization bounds and complexity measures, still, there is often a significant discrepancy between the predicted and actual generalization performance. As well, related large empirical studies have been primarily based on validation with general-purpose image datasets. This paper presents an empirical study that investigates the correlation between 25 complexity measures and the generalization abilities of supervised deep learning classifiers for breast ultrasound images. The results indicate that PAC-Bayes flatness-based and path norm-based measures produce the most consistent explanation for the combination of models and data. We also investigate the use of multi-task classification and segmentation approach for breast images, and report that such learning approach acts as an implicit regularizer and is conducive toward improved generalization., 15 pages, 4 figures

In this paper we build a neural network model to predict prevalence of malaria for a given geographic location and year. We report on our experience of building the most suitable neural network architecture for this problem. We show that both utilizing dropout and Adam optimizer in the network training process is very effective and can lead to a precise model without overfitting issues. Incorporating rainfall data leads to a significant improvement in the precision of the model, highlighting the fact that this is an important factor in the spread of malaria. We then utilize the selected best neural network to predict the outcome of eradicating malaria at given locations. This can help to decide where to use limited resources, like vaccines or insecticides, for the largest possible impact in malaria control.

Motor imagery (MI)-based brain-computer interface (BCI) systems are being increasingly employed to provide alternative means of communication and control for people suffering from neuro-motor impairments, with a special effort to bring these systems out of the controlled lab environments. Hence, accurately classifying MI from brain signals, e.g., from electroencephalography (EEG), is essential to obtain reliable BCI systems. However, MI classification is still a challenging task, because the signals are characterized by poor SNR, high intra-subject and cross-subject variability. Deep learning approaches have started to emerge as valid alternatives to standard machine learning techniques, e.g., filter bank common spatial pattern (FBCSP), to extract subject-independent features and to increase the cross-subject classification performance of MI BCI systems. In this paper, we first present a review of the most recent studies using deep learning for MI classification, with particular attention to their cross-subject performance. Second, we propose DynamicNet, a Python-based tool for quick and flexible implementations of deep learning models based on convolutional neural networks. We show-case the potentiality of DynamicNet by implementing EEGNet, a well-established architecture for effective EEG classification. Finally, we compare its performance with FBCSP in a 4-class MI classification over public datasets. To explore its cross-subject classification ability, we applied three different cross-validation schemes. From our results, we demonstrate that DynamicNet-implemented EEGNet outperforms FBCSP by about 25%, with a statistically significant difference when cross-subject validation schemes are applied.

The new public health crisis threatens the entire world. The infectious virus SARS -CoV-2 (COVID-19) is a new virus that has spread rapidly across the world. Coronavirus spreads faster than the early virus, SARS-CoV and MERS-CoV, the first major beta-coronavirus in the human respiratory system. There are approximately 9,033,508 reported cases worldwide. CT scans have been used to diagnose suspicious negative and positive cases. To detect infected cases and mortality rates, many epidemiological models are being used. Therefore, Machine Learning techniques play a vital role in effective prediction. This technology has been used to extract information on large data sets and predictive performance analysis. As a result, a variety of Machine Learning prediction techniques were employed. This paper aims at determining which Classification method performs a high accuracy rate for the collected data samples of COVID-19 positive cases. The Support Vector Machine (SVM) gives 85% of accuracy, K-Nearest Neighbor gives 80% accuracy, and Naive Bayes (NB) text classifier method gives 65% accuracy.

Recently, the coronavirus 2019 or COVID-19, which originated in China, has spread to other countries' population. It is critical to evaluate an automated detection system for rapid alternative prediction and diagnosis in order to reduce the impact of COVID-19. Because of the constant increase in cases, there are fewer COVID-19 available kits than are required for testing in hospitals. Deep learning methods are evolving to provide outstanding performance in the medical field. Deep learning inspired by brain structure is referred to as machine learning. This paper provides an overview of COVID-19’s detection applications based on deep learning. Furthermore, a comprehensive review of the literature on deep leaning in COVID-19 disease has been illustrated. The proposed research study shows that in spite of presence of issues in medical database, where the transfer method can be used effectively.

As an increasing volume of international trade activities around the world, the amount of cross-boarder import declarations grows rapidly, resulting in an unprecedented scale of potentially fraudulent transactions, in particular false commodity code (e.g., HS Code). The incorrect HS Code will cause duty risk and adversely impact the revenue collection. Physical investigation by the customs administrations is impractical due to the substantial quantity of declarations. This paper provides an automatic approach by harnessing the power of machine learning techniques to relief the burden of customs targeting officers. We introduced a novel model based on the off-the-shelf embedding encoder to identify the correctness of HS Code without any human effort. Determining whether the HS Code is correctly matched with commodity description is a classification task, so the labelled data is typically required. However, the lack of gold standard labelled data sets in customs domain limits the development of supervised-based approach. Our model is developed by the unsupervised mechanism and trained on the unlabelled historical declaration records, which is robust and able to be smoothly adapted by the different customs administrations. Rather than typically classifying whether the HS Code is correct or not, our model predicts the score to indicate the degree of the HS Code being correct. We have evaluated our proposed model on the ground-truth data set provided by Dutch customs officers. Results show promising performance of 71% overall accuracy.

Weaknesses in computer systems such as faults, bugs and errors in the architecture, design or implementation of software provide vulnerabilities that can be exploited by attackers to compromise the security of a system. Common Weakness Enumerations (CWE) are a hierarchically designed dictionary of software weaknesses that provide a means to understand software flaws, potential impact of their exploitation, and means to mitigate these flaws. Common Vulnerabilities and Exposures (CVE) are brief low-level descriptions that uniquely identify vulnerabilities in a specific product or protocol. Classifying or mapping of CVEs to CWEs provides a means to understand the impact and mitigate the vulnerabilities. Since manual mapping of CVEs is not a viable option, automated approaches are desirable but challenging. We present a novel Transformer-based learning framework (V2W-BERT) in this paper. By using ideas from natural language processing, link prediction and transfer learning, our method outperforms previous approaches not only for CWE instances with abundant data to train, but also rare CWE classes with little or no data to train. Our approach also shows significant improvements in using historical data to predict links for future instances of CVEs, and therefore, provides a viable approach for practical applications. Using data from MITRE and National Vulnerability Database, we achieve up to 97% prediction accuracy for randomly partitioned data and up to 94% prediction accuracy in temporally partitioned data. We believe that our work will influence the design of better methods and training models, as well as applications to solve increasingly harder problems in cybersecurity., Comment: Under submission to KDD 2021 Applied Data Science Track

Cloud computing has become a de facto choice of many individuals and enterprises for computing solutions. In the last few years, many cloud providers appear in the market that offers the same services. It is a trivial job to choose an optimal service best suited for organisations in such a massive arms race of service providers. Existing consumer experience could help significantly build a holistic perception of their experiences that ultimately influence service adoption decisions. Sentiment analysis is an effective tool to understand consumer experience about the product or service. The sophisticated sentiment analysis could help businesses to gain a better insight and respond proactively to consumer issues. There are various methods for sentiment analysis that produces ideal results under different conditions. Therefore, it is very important to choose the right method to predict consumer's sentiment for a greatest result. In this paper we analyse the sentiment prediction accuracy of widely used neural network methods - recurrent neural network (RNN), long short-term memory (LSTM) and gated recurrent network (GRU). We use software as a service (SaaS) dataset having 6258 reviews. From analysis results we find that GRU outperforms the LSTM and RNN methods.

Explainable Artificial Intelligence (XAI) aims at introducing transparency and intelligibility into decision-making process of AI systems. Most of the work in this area is focused on supervised machine learning tasks such as classification and regression. Unsupervised algorithms such as clustering can also be explained with existing approaches. This is most often achieved by explaining a classifier trained on cluster data with cluster labels as a dependant variable. However, with such a transformation the information about cluster shape and distribution is lost, which may lead to wrong interpretation of explanations. In this paper, we introduce a method that aids end experts in cluster analysis with human-readable rule-based explanations. We use state-of-the-art explanation mechanism on the multidimensional bounding boxes that represent arbitrarily-shaped clusters. We demonstrate our approach on reproducible synthetic datasets.

The exceptional progress in the field of machine learning (ML) in recent years has attracted a lot of interest in using this technology in aviation. Possible airborne applications of ML include safety-critical functions, which must be developed in compliance with rigorous certification standards of the aviation industry. Current certification standards for the aviation industry were developed prior to the ML renaissance without taking specifics of ML technology into account. There are some fundamental incompatibilities between traditional design assurance approaches and certain aspects of ML-based systems. In this paper, we analyze the current airborne certification standards and show that all objectives of the standards can be achieved for a low-criticality ML-based system if certain assumptions about ML development workflow are applied.

In this paper the general susceptibilities of Unmanned Aerial Vehicles (UAVs) against modern-cyber threats are explored and potential solutions proposed. This is achieved by applying digital-twin architectures and data-driven methods to UAVs to facilitate identification of real-time intrusions and anomalies. These concepts are validated by performing novelty detection on open access UAV flight data with GPS spoofing attacks, which represents a typical system use-case. Multiple machine learning models are trained to demonstrate the feasibility of detecting modern cyber-intrusions and anomalies using the digital-twin architecture. This includes both classical and deep learning techniques to help identify the most suitable model types for the proposed design. The overall results are positive and help highlight the potential of digital-twin architectures for the UAV contexts.

Contrastive learning has revolutionized self-supervised image representation learning field, and recently been adapted to video domain. One of the greatest advantages of contrastive learning is that it allows us to flexibly define powerful loss objectives as long as we can find a reasonable way to formulate positive and negative samples to contrast. However, existing approaches rely heavily on the short-range spatiotemporal salience to form clip-level contrastive signals, thus limit themselves from using global context. In this paper, we propose a new video-level contrastive learning method based on segments to formulate positive pairs. Our formulation is able to capture global context in a video, thus robust to temporal content change. We also incorporate a temporal order regularization term to enforce the inherent sequential structure of videos. Extensive experiments show that our video-level contrastive learning framework (VCLR) is able to outperform previous state-of-the-arts on five video datasets for downstream action classification, action localization and video retrieval. Code is available at https://github.com/amazon-research/video-contrastive-learning., Comment: Code is publicly available at: https://github.com/amazon-research/video-contrastive-learning

AI technology has made remarkable achievements in computational pathology (CPath), especially with the help of deep neural networks. However, the network performance is highly related to architecture design, which commonly requires human experts with domain knowledge. In this paper, we combat this challenge with the recent advance in neural architecture search (NAS) to find an optimal network for CPath applications. In particular, we use differentiable architecture search (DARTS) for its efficiency. We first adopt a probing metric to show that the original DARTS lacks proper hyperparameter tuning on the CIFAR dataset, and how the generalization issue can be addressed using an adaptive optimization strategy. We then apply our searching framework on CPath applications by searching for the optimum network architecture on a histological tissue type dataset (ADP). Results show that the searched network outperforms state-of-the-art networks in terms of prediction accuracy and computation complexity. We further conduct extensive experiments to demonstrate the transferability of the searched network to new CPath applications, the robustness against downscaled inputs, as well as the reliability of predictions.

Virtual Reality (VR) is a promising technology platform for immersive visual analytics. However, the design space of VR analytics interface design is vast and difficult to explore using traditional A/B comparisons in formal or informal controlled experiments— a fundamental part of an iterative design process. A key factor that complicates such comparisons is the dataset. Exposing participants to the same dataset in all conditions introduces an unavoidable learning effect. On the other hand, using different datasets for all experimental conditions introduces the dataset itself as an uncontrolled variable, which reduces internal validity to an unacceptable degree. In this paper, we propose to rectify this problem by introducing a generative process for synthesizing clustered datasets for VR analytics experiments. This process generates datasets that are distinct while simultaneously allowing systematic comparisons in experiments. A key advantage is that these datasets can then be used in iterative design processes. In a two-part experiment, we show the validity of the generative process and demonstrate how new insights in VR-based visual analytics can be gained using synthetic datasets.

In order to satisfy the ever-growing Quality of Service (QoS) requirements of innovative services, cellular communication networks are constantly evolving. Recently, the 5G Non-Standalone (NSA) mode has been deployed as an intermediate strategy to deliver high-speed connectivity to early adopters of 5G by incorporating Long Term Evolution (LTE) network infrastructure. In addition to the technological advancements, novel communication paradigms such as anticipatory mobile networking aim to achieve a more intelligent usage of the available network resources through exploitation of context knowledge. For this purpose, novel methods for proactive prediction of the end-to-end behavior are seen as key enablers. In this paper, we present a first empirical analysis of client-based end-to-end data rate prediction for 5G NSA vehicle-to-cloud communications. Although this operation mode is characterized by massive fluctuations of the observed data rate, the results show that conventional machine learning methods can utilize locally acquirable measurements for achieving comparably accurate estimations of the end-to-end behavior.