Small Unmanned Aerial Systems Research Articles

Brushless DC Motor Sizing Algorithm for Small UAS Conceptual Designers

Accurately sizing vehicle components is an impactful step in the aircraft design process. However, existing methods of sizing brushless DC (BLDC) motors for small unmanned aerial systems (SUAS) ignore how cooling affects motor size. Moreover, the literature methods do not predict a notional motor’s electrical constants, namely winding resistance, torque constant, and figure of merit. We developed a sizing algorithm that predicts the optimal mass and electrical constants using a combination of sizing, efficiency, and thermal models. The algorithm works for radial-flux BLDC motors with masses up to 800 g. An experimental teardown of seven motors informed the algorithm’s sizing models. The teardown motors varied in mass (24–600 g) and geometry (stator aspect ratio of 1.4–9.0). Validated against an independent catalog of 30 motors, the sizing models predicted mass and resistance within 10% and 20% of catalog specifications, respectively. Validated against experimental data, the full algorithm predicted mass, efficiency, and temperature within 20%, 5%, and 10% accuracy, respectively. The algorithm also captured how lowering mass would increase losses and temperature, which the literature models ignore. The algorithm can help users develop more viable concepts that save costs in the long run.

Imaging sensor band comparison for situational awareness in wildfires

In recent decades, wildfires have become increasingly widespread and hazardous. Dryer, hotter weather combined with more frequent heat waves leave forest areas susceptible to sudden, intense, and fast-growing forest fires. To protect private property and mitigate the damage, Hotshot firefighters are deployed into these dangerous situations. Extensive satellite and aerial platforms possess optical techniques for monitoring wildfire risks and boundary tracking. Small unmanned aerial system (sUAS)-based EO/IR systems provide a solution for real-time, high resolution, targeted response to acquire information critical to the safety and efficacy of wildfire mitigation. Real-time imagery from a sUAS of the position of Hotshots and the progression of the fire boundary would be easily obtained and offer a method of ensuring safe deployment. An ideal sensor system for situational awareness in this environment would be able to image the ambient terrain and firefighters with good contrast while also detecting fire signatures and imaging through the smoke. Longer wavelength infrared bands have demonstrated imaging through the smoke of forest fires. However, near the wildfire where the Hotshots work, they also receive strong radiometric signals from the temperature of the smoke. The emitted signal of the smoke can obscure the line of sight similarly to the scattering effect of wildfire smoke in the visible spectrum. The reflective and emissive components of a wildfire scene are studied and compared in the visible (VIS, 0.4–0.7 µm), shortwave infrared (SWIR, 1.0–1.7 µm), extended SWIR (eSWIR, 2.0–2.5 µm), and longwave infrared (LWIR, 8–14 µm). Both a radiometric model and calibrated field measurements find a band that has the highest, to our knowledge, probability for a continuous line of sight for terrain, firefighters, and fire signatures in a wildfire scene.

Heat Transfer Models and Measurements of Brushless DC Motors for Small UASs

Heat transfer affects a motor’s sizing, its performance, and, ultimately, the overall vehicle’s range and endurance. However, the thermal literature does not have early-stage models for outrunner brushless DC (BLDC) motors found in small unmanned aerial systems (UASs). To address this gap, we have developed a non-dimensional heat transfer model (Nusselt correlation). Parametric experiments of four different-sized BLDC motors under load in Reynolds-matched wind tunnel tests generated data for model correlation. The motors’ aspect ratios (diameter/length) ranged from 0.9 to 1.5. The freestream Reynolds number of the axial flow over the motors ranged from 20,000 to 40,000. The rotational Reynolds number ranged from 10,000 to 20,000. The results showed that aspect ratio had the largest influence on heat transfer, followed by rotational and freestream Reynolds numbers. A steady-state model used the correlation to predict the motor’s ambient temperature differential within 10 K of experimental data. A case study applied the correlation to predict a hypothetical motor’s continuous torque in different environments. The correlation enables conceptual designers to capture thermally-driven trade-offs in early design stages and reduce costly revisions in later stages.

Powertrain Design For Small Scale Parallel Hybrid-Electric Test Bench

Parallel hybrid-electric propulsion systems for small scale unmanned aerial systems (UAS) when tested with an internal combustion engine are susceptible to damage due to increased torque compared to all-electric configurations. The University of Victoria Centre for Aerospace Research has conducted testing and identified several locations in the system for potential upgrades. One of the largest issues identified was the electromagnetic clutch’s inability to handle the torque of the Corvid-50 engine. Thus, a new clutch and powertrain system was specified which is better rated for combustion operation. Similar findings are reported based on testing performed at Ł-Institute of Aviation, where a hybrid-electric powertrain stand experienced torque spikes. The spike amplitude was several times higher than the nominal momentum of the ICE, given in the specification sheets. In result, at some working modes, a strong slippage of the clutch has been observed. It is also of highest importance to propose and test potential methods to minimize momentum influence to the powertrain. The goal is to increase robustness of such hybrid systems, and decrease the overall mass of the system.

Human–machine Teaming with Small Unmanned Aerial Systems in a MAPE-K Environment

The Human Machine Teaming (HMT) paradigm focuses on supporting partnerships between humans and autonomous machines. HMT describes requirements for transparency, augmented cognition, and coordination that enable far richer partnerships than those found in typical human-on-the-loop and human-in-the-loop systems. Autonomous, self-adaptive systems in domains such as autonomous driving, robotics, and Cyber-Physical Systems, are often implemented using the MAPE-K feedback loop as the primary reference model. However, while MAPE-K enables fully autonomous behavior, it does not explicitly address the interactions that occur between humans and autonomous machines as intended by HMT. In this article, we, therefore, present the MAPE-K HMT framework, which utilizes runtime models to augment the monitoring, analysis, planning, and execution phases of the MAPE-K loop to support HMT despite the different operational cadences of humans and machines. We draw on examples from our own emergency response system of interactive, autonomous, small unmanned aerial systems to illustrate the application of MAPE-K HMT in both a simulated and physical environment, and we discuss how the various HMT models are connected and can be integrated into a MAPE-K solution.

Unlocking the Secondary Critical Raw Material Potential of Historical Mine Sites, Lousal Mine, Southern Portugal

A steady supply of mineral raw materials is vital for the transition to a low-carbon, circular economy. The number of active mines in Europe has severely declined over the last century and half, giving rise to many abandoned mining waste sites and corresponding geological heritage. Also, the rise in minerals demand for large-scale deployment of renewable energy requires the continued and steady availability of key minerals. The supply risk associated with unpredicted geopolitical events needs to be eliminated/mitigated. Historical mine waste sites are the answer but evaluating mine waste is a lengthy and costly exercise. The study, undertaken in the Lousal Mine, used small unmanned aerial systems (sUASs) to model and determine mine waste volumes by generating orthomosaic maps with quick, inexpensive, and reliable results. Calculated mine waste volumes between 308,478 m3 and 322,455 m3 were obtained. XRD and p-XRF techniques determined the mineralogy and chemistry of waste, which varied from mineralization and host rocks with hydrothermal alteration and numerous neogenic sulphates (mostly gypsum, rhomboclase, ferricopiapite, coquimbite, and jarosite) related with supergene processes and weathering. The study shows the viability of using these sUASs to successfully model historical mine waste sites in an initial phase and for future monitoring programs.

Autonomous Emergency Landing for Fixed-Wing Aircraft with Energy-Constrained Closed-Loop Prediction

This paper presents a new approach for autonomous motion planning for aircraft suffering from a loss-of-thrust emergency. Specifically, we show how modifications to the Closed-Loop Rapidly exploring Random Trees (CL-RRT) framework combined with controlled energy dissipation can enable rapid and effective kinodynamic motion planning. This CL-RRT Glide algorithm uses closed-loop prediction not only for node connections but also to estimate the remaining energy and prune infeasible paths. This greatly speeds up the search process, which is essential for emergency situations. In addition, we improve the ability of the gliding aircraft to reach a goal position and energy state. We do so by creating a Dissipative Total Energy Control Scheme (TECS). Dissipative TECS enables the glider to lose excess altitude in order to reach a desired energy level. Simulation results illustrate how the proposed methods enable faster motion planning. We also integrate the system into a small unmanned aerial vehicle system and experimentally demonstrate autonomous glide planning and execution during a motor-failure event. This type of algorithm can primarily benefit unmanned aircraft but can also serve to assist pilots in stressful emergency situations.

Flight-Validated Electric Powertrain Efficiency Models for Small UASs

Minimizing electric losses is critical to the success of battery-powered small unmanned aerial systems (SUASs) that weigh less than 25 kgf (55 lb). Losses increase energy and battery weight requirements which hinder the vehicle’s range and endurance. However, engineers do not have appropriate models to estimate the losses of a motor, motor controller, or battery. The aerospace literature often assumes an ideal electrical efficiency or describes modeling approaches that are more suitable for controls engineers. The electrical literature describes detailed design tools that target the motor designer. We developed SUAS powertrain models targeted for vehicle designers and systems engineers. The analytical models predict each component’s losses using high-level specifications readily published in SUAS component datasheets. We validated the models against parametric experimental studies involving novel powertrain flight data from a specially instrumented quadcopter. Given propeller torque and speed, our integrated models predicted a quadcopter’s battery voltage within 5% of experimental data for a 5+ min mission despite motor and controller efficiency errors up to 10%. The models can reduce development costs and timelines for different stakeholders. Users can evaluate notional or existing powertrain configurations over entire missions without testing any physical hardware.

Agricultural weed identification in images and videos by integrating optimized deep learning architecture on an edge computing technology

Recent advancements in deep learning (DL)-based model optimization techniques have resulted in better weed identification accuracy. However, optimizing these models to identify weeds in images captured using small unmanned aerial system (UAS) has not been much explored. Moreover, leveraging the optimized model on resource constrained edge platform that could be easily integrated with UAS for real-time weed identification could be of significant advantage in developing precision aerial spraying weed management technology. Therefore, this study proposes YOLO-Spot model that is based on YOLOv7-tiny architecture, has been optimized and reconstructed to identify weeds amongst crop plants in aerial images and videos. The optimized model tends to use a smaller number of trainable parameters and reduced feature map sizes for weed identification. Most of the redundant convolutional layers along with feature map sizes have been reduced coupled with an integration of a novel module re-parameterized convolutional layer (RCL) within the neck component of the network. Furthermore, YOLO-Spot model has been trained on the three image resolutions, 320 × 320, 640 × 640 and 1280 × 1280, and has been named as YOLO-Spot_S, YOLO-Spot_M and YOLO-Spot_L, respectively. Out of all the variants, YOLO-Spot_M model has achieved significant prediction accuracy as compared to other variants and a denser layered model YOLOv7-Base. YOLO-Spot_M model utilizes over 75 % less parameters and 86 % reduced GFLOPs compared to YOLOv7-Base. As per the results, YOLO-Spot_M has outperformed YOLOv7-Base by achieving +1.3 % and +2.7 % overall accuracy and mAP(@0.5), respectively. The optimized architecture utilizes 4X less power (in W) when trained on a normal graphical processing unit (GPU). Moreover, converting YOLO-Spot_M to half-precision (FP16) for resource constrained device deployment (AGX Xavier), led to a +0.6 % accuracy and 5X faster weed recognition accuracy in aerial images and videos during inferencing. Based on the metrics obtained, YOLO-Spot_M model is recommended model that could be integrated with remote sensing technologies for site-specific weed management.

Flow Confinement Effects on sUAS Rotor Noise

The goal of this paper is to investigate the effects of flow confinement on the noise generated by a small unmanned aerial system rotor measured and simulated in a partially closed test room. More specifically, the influence of the flow recirculation in the test room and the consequent blade turbulence interaction on both the tonal unsteady loading noise and the broadband laminar separation noise components are scrutinized. The study also aims at improving our prediction capabilities of drone rotor noise, and deepening our fundamental knowledge of rotor aeroacoustics in transitional boundary layer regimes. Numerical simulations are conducted by means of a scale-resolving wall-modelled lattice Boltzmann solver to calculate the unsteady flow and the noise generated by the same reference configuration used in our previous works, which consists in a rotor operated at a chord-based Reynolds number of about 70000 at 75% of the tip radius, both in hover and axial flow conditions. Results for a rotor in pristine flow conditions are compared to results obtained by simulating the rotor in a confined environment that reproduces the A-Tunnel test chamber of Delft University of Technology. One one hand, our study reveals that previously observed discrepancies between measured and predicted noise spectra in hover conditions are not due to the influence of the flow confinement on the laminar separation mechanisms, as hypothesized in our previous study, rather to a lack of numerical resolution. On the other hand, the new results reveal that, in hover conditions, the dominant effect of the flow recirculation is to increase the unsteady loading tonal noise components, whereas, in axial flow conditions, no significant effect due to the test environment is observed. Some of our conclusive observations are supported by results obtained using a source identification technique herein presented for the first time.

A High Performance Tubular Solid Oxide Fuel Cell for Unmanned Autonomous Systems

Small unmanned aerial systems (SUASs) have a great potential for numerous applications. The SUASs are light-weight, man-portable, and capable of delivering commercial products. Currently, small UASs are powered by lithium-ion batteries, however, flight endurance is limited and requires frequent charging battery due to low capacity and energy density. Numerous research and development efforts have focused on alternative power sources. These alternative power sources include technologies such as solar cells and hydrogen based polymer electrolyte membrane fuel cells. Additionally, polymer electrolyte membrane fuel cells still have the same technical constraints, the source of hydrogen, as it did many years ago. We have developed and examined the operational capabilities of propane fueled solid oxide fuel cell (SOFC) technology can power an sUAS. A self-sustaining miniature tubular SOFC stack provides a unique high power and high energy density for UAS. The SOFC has great potential for commercial and civilian applications, ranging from small scale systems (e.g. small unmanned air vehicle, wireless sensor networks) to large scale system (e.g. electric vehicle, wireless sensor networks, electric vehicles).

On-field noise measurements and acoustic characterisation of multi-rotor small unmanned aerial systems

The noise signature of small Unmanned Aerial Systems (sUAS) is highly influenced by specific operating and weather conditions. Accurate noise assessment of sUAS operations requires field noise measurements that capture all the complexity of sound emission and propagation during realistic flight situations. This paper presents a measurement and analysis framework for the acoustic characterisation of sUAS through the calculation of conventional noise metrics (LAmax and LAE), frequency and directivity features. Using a multi-channel measurement approach, and back-propagating the sound from ground microphone to source, the presented framework allows the calculation of acoustic hemispheres for a selection of acoustic metrics. Important findings are that (i) the framework is robust for a variety of multi-rotor sUAS varying in size and configuration; (ii) broadband noise and tonal noise are the dominant noise sources during flyover for larger sUAS and smaller sUAS respectively; (iii) the maximum noise radiation of the sUAS tested is found in the rear arc of polar directivity; and (iv) angles of maximum radiation of amplitude modulated noise have been found for most sUAS tested at Θ angles±30∘. This work is intended to be relevant in establishing common protocols regarding sUAS acoustic certification between environmental policymakers, stakeholders, and industry.

Simulation of the Effect of Correlated Packet Loss for sUAS Platforms Operating in Non-Line-of-Sight Indoor Environments

The current state of the art in small Unmanned Aerial System (sUAS) testing and evaluation exists mainly in the realm of outdoor flight. Operating small flying sUAS in constrained indoor or subterranean environments places different constraints on their communication links (control links and camera/sensor links). Communication loss in these environments is much more severe due to the proximity of obstacles. This paper examines how correlated packet loss (burst errors) occurring on both the control and camera communication links affects the ability of pilots to fly and navigate small sUAS platforms in constrained Non-Line of Sight (NLOS) environments. A software test bench called AirSim, a UAV simulator, allows us to better understand the effects of correlated packet loss on flyability without damaging multiple sUAS units by flight testing. The simulation was designed to support the design of test methodologies for evaluating the robustness of the communication links and to understand performance without damaging flight tests. Throughout the simulations, it is observed how different levels of packet loss affect the pilot and the number of simulated crashes into the obstacles placed through space. The simulations modeled packet loss both on the video link and the control link to display how packet loss affects ability to pilot and control the sUAS. The utility of using a simulated environment rather than flight testing prevents damage to the fragile and expensive drones being used.

Discrete-Time Incremental Backstepping Control with Extended Kalman Filter for UAVs

In this study, a discrete-time incremental backstepping (DTIBS) controller with an extended Kalman filter (EKF) is proposed for unmanned aerial vehicles (UAVs) with unknown actuator dynamics. The Taylor series and an approximate discrete method are employed, transforming the second-order continuous-time nonlinear system into a discrete-time nonlinear plant with an incremental input form. The incremental control laws are designed using the incremental nonlinear dynamic inversion (INDI) method and the time-delay control (TDC) method. The TDC is introduced to design the control law, eliminating the need for prior knowledge of the control effectiveness matrix involving some unknown aerodynamic coefficients. In addition, the airflow angle and body rotation rate are selected as key system states, and the EKF is used to design a state estimator to estimate the local state of the small unmanned aerial vehicle closed-loop flight control system under strong noise conditions. The effectiveness of the DTIBS control method with EKF is verified through numerical simulation. The results show that the proposed method can effectively estimate the state under the typical noise characteristics of low-cost sensors, and the closed-loop control systems has good tracking performance and can quickly and effectively track sudden commands.

Measuring Evapotranspiration Suppression from the Wind Drift and Spray Water Losses for LESA and MESA Sprinklers in a Center Pivot Irrigation System

Wind drift and evaporation loss (WDEL) of mid-elevation spray application (MESA) and low-elevation spray application (LESA) sprinklers on a center pivot and linear-move irrigation machines are measured and reported to be about 20% and 3%, respectively. It is important to estimate the fraction of WDEL that cools and humidifies the microclimate causing evapotranspiration (ET) suppression, mitigating the measured irrigation system losses. An experiment was conducted in 2018 and 2019 in a commercial spearmint field near Toppenish, Washington. The field was irrigated with an 8-span center pivot equipped with MESA but had three spans that were converted to LESA. All-in-one weather sensors (ATMOS-41) were installed just above the crop canopy in the middle of each MESA and LESA span and nearby but outside of the pivot field (control) to record meteorological parameters on 1 min intervals. The ASCE Penman–Monteith (ASCE-PM) standardized reference equations were used to calculate grass reference evapotranspiration (ETo) from this data on a one-minute basis. A comparison was made for the three phases of before, during, and after the irrigation system passed the in-field ATMOS-41 sensors. In addition, a small unmanned aerial system (UAS) was used to capture 5-band multispectral (ground sampling distance [GSD]: 7 cm/pixel) and thermal infrared images (GSD: 13 cm/pixel) while the center pivot irrigation system was irrigating the field. This imagery data was used to estimate crop evapotranspiration (ETc) using a UAS-METRIC energy balance model. The UAS-METRIC model showed that the estimated ETc under MESA was suppressed by 0.16 mm/day compared to the LESA. Calculating the ETo by the ASCE-PM method showed that the instantaneous ETo rate under the MESA was suppressed between 8% and 18% compared to the LESA. However, as the time of the ET suppression was short, the total amount of the estimated suppressed ET of the MESA was less than 0.5% of the total applied water. Overall, the total reduction in the ET due to the microclimate modifications from wind drift and evaporation losses were small compared to the reported 17% average differences in the irrigation application efficiency between the MESA and the LESA. Therefore, the irrigation application efficiency differences between these two technologies were very large even if the ET suppression by wind drift and evaporation losses was accounted for.

Tonality Perception Lessons Learned

The future is expected to bring Advanced Air Mobility (AAM) vehicles, including small unmanned aerial systems (sUAS), urban air mobility (UAM) vehicles and regional air mobility (RAM) vehicles. These manned and unmanned vehicles are propelled by rotors. Rotors tend to generate tonal sound as their blades interact periodically with airflow features. Since people are more sensitive to tonality, including tones, than broad band sound, AAM generated tonality is expected to be an important consideration for design. In this paper several tonality metrics are examined for their ability to explain perceived annoyance of AAM flyover noise as measured by NASA's Rotorcraft Sound Quality Metric 1 (RoQM-1) test. The various investigated metrics use one-third octave band, narrow band, and autocorrelation analysis. It is observed that tonality influences but does not control perceived flyover noise annoyance due to other sound qualities like roughness, consistent with previous work. The metrics are also examined for their ability to explain perceived tonality of sounds generated by IT equipment. The metrics based on autocorrelation are observed to best explain perceived tonality while also being among the best for explaining flyover annoyance. The plausibility of the auditory nervous system as a physiological auto-correlator is discussed.

The concept of sUAS/DL-based system for detecting and classifying abandoned small firearms

Military object detection and identification is a key capability in surveillance and reconnaissance. It is a major factor in warfare effectiveness and warfighter survivability. Inexpensive, portable, and rapidly deployable small unmanned aerial systems (sUAS) in conjunction with powerful deep learning (DL) based object detection models are expected to play an important role for this application. To prove overall feasibility of this approach, this paper discusses some aspects of designing and testing of an automated detection system to locate and identify small firearms left at the training range or at the battlefield. Such a system is envisioned to involve an sUAS equipped with a modern electro-optical (EO) sensor and relying on a trained convolutional neural network (CNN). Previous study by the authors devoted to finding projectiles on the ground revealed certain challenges such as small object size, changes in aspect ratio and image scale, motion blur, occlusion, and camouflage. This study attempts to deal with these challenges in a realistic operational scenario and go further by not only detecting different types of firearms but also classifying them into different categories. This study used a YOLOv2 CNN (ResNet-50 backbone network) to train the model with ground truth data and demonstrated a high mean average precision (mAP) of 0.97 to detect and identify not only small pistols but also partially occluded rifles.

Spatial corn canopy temperature extraction: How focal length and sUAS flying altitude influence thermal infrared sensing accuracy

Irrigation water management starts with quantifying irrigation prescriptions based on crop water requirements at a spatial scale. The most accepted and effective method for estimating this water need is to use the canopy temperature of the plants. The use of small unmanned aerial systems (sUAS) with a thermal infrared camera has long been established as an effective method of measuring plant canopy temperatures at a spatial scale. However, concerns still exist about the accuracy of canopy temperatures collected by these systems and how imagery collection altitude or camera viewing angle affects the temperature estimation. To address these concerns, this study was designed to evaluate the effects of flying altitude of airborne thermal infrared (TIR) imagery and camera view angle on corn-canopy temperature sensing accuracy and image quality. Three different thermal cameras with focal lengths of 9 mm, 13 mm, and 19 mm were used, and each camera was flown at an altitude of 30 m, 50 m, and 70 m. Thermal cameras were mounted on an sUAS and flown over a 6000 m2 cornfield. The sUAS was flown at a speed of 3 m/s on autopilot mode guided by pre-planned missions. The thermal camera was triggered once per second, and imagery was geotagged using an external GPS unit. Images obtained using different combinations of focal lengths and flight altitudes were processed and converted into orthomosaics. Calibration was conducted using ground-based temperature reference panels and temperature maps were created from the orthomosaics. The orthomosaics were examined for the accuracy of the corn-canopy temperature sensing, ability to differentiate between hot and cold surfaces, ease of image stitching/developing orthomosaics, geometric accuracy, image quality, and spatial resolution. The results indicated that with the combination of appropriate camera focal length, altitude, and image calibration techniques, a canopy temperature map of crops with a temperature error of less than 2 °C from the actual canopy temperature can be produced. A narrow-angle thermal camera flying at low altitudes (<50 m) was found to be the least suitable combination for corn canopy temperature sensing. The most appropriate combination for temperature estimation of corn canopies was with a 13 mm focal length camera flying at an altitude of 50 m above ground level.

Blade-Wake Interaction Noise for Small Hovering Rotors, Part I: Characterization Study

This work illustrates the use of artificial neural network modeling to study and characterize broadband blade-wake interaction noise from hovering small unmanned aerial systems rotors subject to varying airfoil geometries, rotor geometries, and operating conditions. Design of experiments was used to create input feature spaces, and a high-fidelity strategy was implemented at the discrete data points defined by the input feature spaces to design airfoils and rotor blades, predict the unsteady rotor aerodynamics and aeroacoustics, and isolate the blade-wake interaction noise from the acoustic broadband noise. A metric for the blade-wake interaction noise was developed, and the ANOPP2 Artificial Neural Network Tool was used to identify an optimal prediction model for the nonlinear relationship between the input features and the metric for blade-wake interaction noise. This optimal artificial neural network was then validated over training/test data and exhibited prediction accuracy over 91% for data previously unseen by the model. A sensitivity analysis was conducted, which showed that input features that directly modify the thrust coefficient had a dominant effect over blade-wake interaction noise. The optimal prediction model along with aerodynamic simulations were used to further study the effect of varying input features on blade-wake interaction noise, and three types of blade-wake interaction noise were identified.

Detecting Anomalies in Small Unmanned Aerial Systems via Graphical Normalizing Flows

<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">With the increasing deployment of small Unmanned Aerial Systems (sUAS) on various tasks, it becomes crucial to analyze and detect anomalies from their flight logs. To support research in this area, we curate DLA, the first real-world time series anomaly detection dataset in the domain of sUAS, which contains 41 sUAS flight logs annotated with various types of anomalies. As anomalies tend to occur in low-density areas within a distribution, we propose Graphical Normalizing Flows (GNF), a graph-based autoregressive deep learning model, to perform anomaly detection through density estimation. GNF contains (1) a temporal encoding module using a transformer to capture the temporal dynamics, (2) an inter-feature encoding module leveraging graph representation learning on a Bayesian network to model the statistical dependencies among time series features, and (3) a density estimation module with Normalizing Flows. Extensive experiments have demonstrated GNF’s superior anomaly detection power on DLA compared with state-of-the-art baselines.</i>