The rapid proliferation of Unmanned Aerial Vehicles (UAVs), commonly known as drones, within urban low-altitude airspace presents a paradigm shift for logistics, infrastructure inspection, and public services. However, this integration introduces significant safety and security challenges. The complex urban environment, characterized by dense infrastructure, dynamic air traffic, and vulnerable populations, amplifies the potential consequences of UAV drone incidents. A critical step towards ensuring the safe scalability of urban air mobility is a thorough, data-driven understanding of the root causes behind UAV drone accidents and public safety events. This study aims to construct a comprehensive risk identification framework by analyzing historical UAV drone accident data, employing a hybrid methodology of Fault Tree Analysis (FTA) and Bayesian Networks (BN) to systematically identify, quantify, and rank the key risk factors contributing to urban UAV drone mishaps.
The operational safety of a UAV drone is influenced by a multifaceted interplay of factors. Prior research has explored various aspects, such as third-party risk modeling, focusing on ground collision probabilities, noise, and privacy concerns. Other studies have delved into collision risk with static and dynamic obstacles, the impact of meteorological conditions, and the vulnerabilities associated with communication and navigation systems like GPS denial. While these contributions are valuable, many approaches focus on isolated risk domains. There remains a need for an integrated framework that synthesizes human, technical, environmental, and procedural factors into a unified model derived from actual UAV drone accident data. This study addresses this gap by building a holistic risk factor taxonomy from real-world cases and applying advanced probabilistic reasoning to discern the most critical failure paths.
1. Analysis of UAV Drone Accident Cases and Risk Factor Extraction
To establish a foundation rooted in empirical evidence, this study compiled and analyzed UAV drone safety incident reports spanning from 2017 to 2024. Data was sourced from publicly available aviation safety databases and incident reporting platforms, resulting in a curated dataset of 75 significant civilian UAV drone accident records. Each report was scrutinized to extract the primary and contributing causes of the incident.
Through iterative analysis, 23 distinct fundamental risk factors were identified and categorized. The frequency of each factor’s occurrence within the dataset was calculated. These factors, denoted as B1 through B23, form the basic events for subsequent modeling. The occurrence frequency provides an initial, data-driven prior probability for each risk factor, calculated as:
$$Q_i = \frac{N_i}{N_r \times T}$$
where $Q_i$ is the statistical probability (prior) of basic event $i$, $N_i$ is the count of accidents where factor $i$ was present, $N_r$ is the total number of analyzed accident reports (75), and $T$ is the time span in years (8). The extracted risk factors and their observed frequencies are summarized in the table below.
| Factor Code | Risk Factor Description | Frequency | Prior Probability ($Q_i$) |
|---|---|---|---|
| B1 | Violation of flight regulations (e.g., no license, beyond visual line of sight improperly) | 18 | 0.0300 |
| B2 | Flight within a restricted or no-fly zone | 9 | 0.0150 |
| B3 | Loss of control link / Signal lost | 1 | 0.0017 |
| B4 | Illegal filming / Privacy invasion | 5 | 0.0083 |
| B5 | Personnel carelessness / Pre-flight oversight | 5 | 0.0083 |
| B6 | Software system bug or freeze | 1 | 0.0017 |
| B7 | Personnel operational error during flight | 12 | 0.0200 |
| B8 | Lack of safety awareness / Insufficient training | 7 | 0.0117 |
| B9 | Communication datalink failure | 6 | 0.0100 |
| B10 | GPS signal loss or spoofing | 3 | 0.0050 |
| B11 | Battery failure (e.g., sudden discharge, fire) | 2 | 0.0033 |
| B12 | Motor / Propulsion system failure | 2 | 0.0033 |
| B13 | Structural damage / Mechanical fatigue | 1 | 0.0017 |
| B14 | Payload (camera, sensor) malfunction | 1 | 0.0017 |
| B15 | Airspace conflict with another UAV drone or aircraft | 3 | 0.0050 |
| B16 | Poor or unrealistic flight path planning | 7 | 0.0117 |
| B17 | Lack of emergency contingency procedures | 1 | 0.0017 |
| B18 | Obstacle (building, tower, tree) collision | 10 | 0.0167 |
| B19 | Bird strike | 2 | 0.0033 |
| B20 | Adverse wind conditions (gust, turbulence) | 6 | 0.0100 |
| B21 | Precipitation (rain, snow) | 2 | 0.0033 |
| B22 | Low visibility / Fog | 3 | 0.0050 |
| B23 | Electromagnetic interference (EMI) | 1 | 0.0017 |
2. Construction of a FTA-BN Hybrid Risk Identification Model for UAV Drone Operations
To move beyond descriptive statistics and model the causal relationships and probabilistic interactions among these risk factors, a hybrid FTA-BN methodology is adopted. FTA provides a structured, top-down approach to logically connect basic events to an undesired top event. BN, derived from the FTA, enables powerful probabilistic reasoning, including diagnostic inference and sensitivity analysis, under conditions of uncertainty—a hallmark of complex UAV drone operations.
2.1 Development of the Fault Tree (FTA)
The top event (T1) is defined as the occurrence of a “UAV Drone Public Safety Incident.” This encompasses any event where UAV drone operation leads to actual harm or high potential for harm, including collisions, crashes onto people/property, unauthorized incursions into sensitive airspace, illegal surveillance, and other acts that threaten public security.
The 23 basic events (B1-B23) were logically grouped into eight intermediate event categories based on their nature:
- M1: Regulation & Behavioral Risk: B1, B2, B4.
- M2: Operational & Technical Risk: B3, B5, B6, B7, B8.
- M3: Natural Environment: B20, B21, B22.
- M4: Operational Environment: B18, B19, B23.
- M5: Human Risk: M1, M2. (A fusion of regulation and operational aspects).
- M6: UAV Drone Hardware Failure Risk: B9, B10, B11, B12, B13, B14.
- M7: Flight Management Risk: B15, B16, B17.
- M8: Flight Environment Risk: M3, M4.
The fault tree logic asserts that the top event T1 occurs if any one of the direct causes M5, M6, M7, or M8 occurs. Similarly, each intermediate event is the logical OR combination of its constituent basic or other intermediate events. The probability of an OR gate output $P_{out}$ given $n$ input events with probabilities $P_i$ is calculated as:
$$P_{out} = 1 – \prod_{i=1}^{n} (1 – P_i)$$
Using this formula and the prior probabilities $Q_i$ from the data, the probabilities for all intermediate events and the top event were computed. The initial FTA quantification highlighted that human risk (M5) had the highest probability of occurrence among direct causes, followed by UAV drone hardware failure risk (M6). This preliminary finding underscored the predominant role of human factors in UAV drone safety but required the dynamic reasoning capabilities of a BN for deeper validation and factor prioritization.
2.2 Transformation from FTA to Bayesian Network (BN)
The FTA was mapped to an equivalent BN to enable advanced inference. In this transformation:
- Each event in the FTA (basic, intermediate, top) becomes a node in the BN.
- The logical OR gates define the conditional probability tables (CPTs) for the parent nodes. For a node with $m$ parents connected via an OR gate, the probability of the node being in state “Yes” (event occurred) is $1 – \prod(1 – P_{\text{parent}=\text{Yes}})$.
- The directed edges flow from basic events upwards to the top event, representing causality.
The resulting BN structure consists of 32 nodes: 23 root nodes (B1-B23), 8 intermediate nodes (M1-M8), and 1 leaf node (T1). Each node has a binary state: “Yes” or “No.” The initial CPTs were populated based on the FTA logic and the data-informed prior probabilities.
2.3 Parameter Learning and BN Model Finalization
To refine the model and ensure its parameters accurately reflect the dependencies in the accident data, parameter learning was performed using the Expectation-Maximization (EM) algorithm. This algorithm is particularly suited for situations with incomplete data, as it iteratively estimates the most likely values for the CPT entries given the observed dataset of 75 cases, where each case is a pattern of occurred risk factors. The software GeNIe was utilized for this learning process and for subsequent model analysis. The finalized BN model provides a robust computational representation of the probabilistic relationships between all risk factors and the ultimate UAV drone public safety incident.
3. Key Risk Factor Inference Using the FTA-BN Model
The primary advantage of the BN model is its capability for bidirectional probabilistic reasoning. We employed diagnostic (backward) reasoning and sensitivity analysis to identify the most critical risk factors.
3.1 Diagnostic Reasoning (Backward Analysis)
Diagnostic reasoning answers the question: “Given that a UAV drone public safety incident (T1) has definitely occurred, which risk factors are most likely to have caused it?” This is computed by setting the state of T1 to “Yes” (probability = 100%) and updating the probabilities of all other nodes in the network (calculating their posterior probabilities).
The change in probability for each basic event, from its prior ($Q_i$) to its posterior given T1=Yes, is a strong indicator of its diagnostic importance. The percentage change rate is calculated as:
$$\text{Change Rate} = \frac{\text{Posterior Probability} – \text{Prior Probability}}{\text{Prior Probability}} \times 100\%$$
The top 10 basic events ranked by their posterior probability and their corresponding change rates are presented in the following tables. These factors become the prime suspects when a UAV drone incident is investigated.
| Factor Code | Risk Factor Description | Prior Probability | Posterior Probability |
|---|---|---|---|
| B1 | Violation of flight regulations | 0.0300 | 0.2987 |
| B7 | Personnel operational error | 0.0200 | 0.1751 |
| B2 | Flight in a no-fly zone | 0.0150 | 0.1711 |
| B8 | Lack of safety awareness | 0.0117 | 0.1106 |
| B4 | Illegal filming | 0.0083 | 0.1087 |
| B16 | Poor flight path planning | 0.0117 | 0.0907 |
| B9 | Communication datalink failure | 0.0100 | 0.0904 |
| B20 | Adverse wind conditions | 0.0100 | 0.0795 |
| B5 | Personnel carelessness | 0.0083 | 0.0779 |
| B15 | Airspace conflict | 0.0050 | 0.0599 |
| Factor Code | Risk Factor Description | Change Rate |
|---|---|---|
| B4 | Illegal filming | 1210% |
| B15 | Airspace conflict | 1098% |
| B2 | Flight in a no-fly zone | 1041% |
| B1 | Violation of flight regulations | 896% |
| B8 | Lack of safety awareness | 845% |
| B5 | Personnel carelessness | 839% |
| B9 | Communication datalink failure | 804% |
| B12 | Motor failure | 800% |
| B11 | Battery failure | 796% |
| B10 | GPS signal loss | 791% |
3.2 Sensitivity Analysis
Sensitivity analysis measures the degree to which uncertainty in each input node (risk factor) contributes to the uncertainty in the target node (T1). In GeNIe, this is often visualized by the intensity of a node’s color, where darker shades indicate a greater influence on the target probability. The analysis conducted on our UAV drone risk BN confirmed the findings from diagnostic reasoning and added further nuance. Nodes corresponding to human-risk factors—particularly B1 (Regulation Violation), B2 (No-fly Zone Flight), B7 (Operational Error), B8 (Safety Awareness), B5 (Carelessness), and B4 (Illegal Filming)—exhibited the darkest shading, signifying the highest sensitivity. Among technical factors, B9 (Communication Failure), B11 (Battery), and B12 (Motor) showed the next highest level of influence on the likelihood of a UAV drone public safety incident.
3.3 Synthesis of Key Risk Factors
By synthesizing the results from the FTA quantification, diagnostic reasoning, and sensitivity analysis, a clear hierarchy of key risk factors for urban UAV drone operations emerges. The most critical factors, which exhibit high posterior probability, high diagnostic change rate, and high sensitivity, are predominantly related to human and regulatory conduct. The convergence of evidence points to the following as the paramount risk factors:
- B1: Violation of Flight Regulations – The single most probable and influential factor.
- B2: Flight within a No-Fly Zone – A specific, high-consequence regulatory violation.
- B4: Illegal Filming / Privacy Invasion – A key motivator for malicious or negligent UAV drone use with high diagnostic relevance.
- B5 & B8: Personnel Carelessness and Lack of Safety Awareness – Foundational human factors enabling other errors.
- B7: Operational Error During Flight – A direct execution failure.
Following these, the most significant technical/hardware factors for the UAV drone itself are:
- B9: Communication Datalink Failure – Leading to loss of control.
- B11 & B12: Battery and Motor/Propulsion System Failure – Critical power and thrust system failures.
Significant environmental and procedural factors include:
- B15: Airspace Conflict – Indicative of traffic management challenges.
- B20: Adverse Wind Conditions – A major environmental hazard for small UAV drones.
4. Conclusion and Implications for UAV Drone Safety Management
This study developed a data-driven, probabilistic framework for identifying key risk factors in urban UAV drone operations by integrating Fault Tree Analysis with Bayesian Networks. The analysis of 75 real-world accident cases provided an empirical foundation, leading to a structured model of 23 risk factors. The hybrid FTA-BN approach enabled not only the logical structuring of causality but also powerful diagnostic and sensitivity analyses to quantify the contribution of each factor.
The conclusive finding is that human and regulatory risk factors are the dominant contributors to UAV drone public safety incidents, substantially outweighing purely technical failures. The key risk factors identified—violation of regulations, incursion into no-fly zones, illegal surveillance, and personnel carelessness—highlight areas requiring urgent and focused mitigation. For regulators and airspace managers, this underscores the critical importance of robust enforcement, geo-fencing compliance, and pilot education and certification. For UAV drone manufacturers and operators, it emphasizes the need for reliable communication links, battery management systems, and propulsion system integrity, as these are the leading technical failure modes.
The proposed FTA-BN model serves as a dynamic risk assessment tool. It can be updated with new accident data to refine probabilities and can be adapted to assess the risk impact of proposed safety interventions, such as mandatory detect-and-avoid systems or enhanced pilot training protocols. Future work should focus on expanding the accident dataset for greater statistical robustness, integrating real-time operational data (e.g., weather, traffic density) for proactive risk monitoring, and exploring machine learning techniques to automatically identify emerging risk patterns from incident reports. By prioritizing resources and strategies towards mitigating these key risk factors, the safe, scalable, and sustainable integration of UAV drones into the urban airspace ecosystem can be significantly advanced.