The increasing use of Unmanned Autonomous Vehicles (UAVs) or drones across various sectors raises concerns about uncooperative drones that pose risks to critical infrastructure. Understanding drone intentions is crucial for risk assessment and countermeasure implementation. Drone intentions are inherently intangible and unobservable, making direct inference challenging. Existing methods often rely on observable data and expert-defined features, leading to unreliable inferences due to biases. This research addresses this limitation by developing a novel CPhy-ML approach that leverages the power of deep learning while integrating flight physics and control principles to enhance the robustness and accuracy of intention inference. The proliferation of affordable drone technology has amplified the potential for malicious use, necessitating effective methods for detecting and mitigating threats. The challenge lies in distinguishing malicious intentions from benign behavior, avoiding excessive false positives or unwarranted trust in autonomous systems. Existing intention classification methods using expert-defined features struggle with scalability and suffer from cognitive biases when applied to complex drone movements. While data-driven methods focusing on trajectory prediction exist, they often neglect crucial information embedded in flight physics. The CPhy-ML framework aims to overcome these limitations by combining the strengths of data-driven learning with the constraints of flight dynamics and control measurements. This integrated approach allows for more reliable intention inference by connecting the drone's purpose to its observed mission profile and control strategies. The use of a rich and diverse dataset further enhances the generalization capability of the model, ensuring robust performance across a range of scenarios.

Prior work on drone intention classification primarily uses expert-knowledge-based methods defining low-dimensional behavioral features from simple motion dynamics. These methods often rely on geofencing, expert traffic rules, or drone flight constraints but lack scalability for classifying diverse intentions. In contrast, intention inference methods focus on predicting future trajectories using data-driven learning models. While these models use snapshot data to cluster drone attributes, they largely ignore the continuous flight physics, which offers valuable information about mission profiles and intentions. Although physics-informed models have shown promise in improving the learning capabilities of data-driven methods, there's a gap in uncovering the hidden nature of intention and the complex capabilities of drones. This paper bridges this gap by proposing a novel approach.

The proposed CPhy-ML framework infers drone intentions through two complementary definitions: trajectory intention (purpose of use and future trajectory) and reward function intention (hidden motivation behind control design). The framework uses a rich dataset generated from open-access sources and real-world drone data, including synthetic radar data, to train the models. The dataset includes four trajectory intention classes (mapping, point-to-point, package delivery, and perimeter flights) and two reward function classes (normal and anomalous trajectories). The CPhy-ML architecture (Fig. 1) consists of three main components. First, a hybrid classifier, incorporating a Convolutional Bidirectional LSTM with Attention (CBLSTMA) network and a deep LSTM autoencoder, classifies trajectory intention and detects anomalies. Second, a Deep Mixture of Experts (DMOE) network predicts the drone's future airspace occupancy. Each expert in the DMOE is a multi-input CNN, weighted by the hybrid classifier's output probabilities. Third, a Reservoir Computing (RC) network enhanced by a physics-informed model predicts trajectories for anomalous behavior. For reward intention inference, an off-policy model-based reward-shaping Inverse Reinforcement Learning (IRL) architecture is employed. This architecture infers the hidden reward function (a quadratic function of states and control inputs) based on sampled trajectories and a linear drone model. The framework utilizes sub-trajectory and summary features extracted from telemetry data and synthetic radar data to improve real-time detection capabilities. The use of a dynamic mode decomposition with control (DMDc) provides a linear model for noise suppression and facilitating prediction analysis. The DMDc model is incorporated into the CPhy-ML framework to improve the accuracy and robustness of the trajectory prediction. A linear quadratic regulator (LQR) controller is used to obtain the control input which is then used in the prediction model to improve its accuracy and stability. A detailed explanation of each component, including the loss functions, weight calculations, and algorithm details, is provided in the supplementary materials.

The CPhy-ML framework achieves state-of-the-art performance in drone intention inference. The hybrid classifier shows high accuracy (94.51% validation, 97.75% testing) in classifying trajectory intentions and successfully detects novel or unseen trajectories. The DMOE network improves trajectory regression results (R² = 0.5105 validation, 0.7482 testing), demonstrating its ability to capture trajectory variability. In trajectory prediction, the physics-informed reservoir computing (PIRC) method significantly outperforms standard RC methods, reducing mean squared error (MSE) and demonstrating robustness to noise in real-world scenarios. The reward-shaping IRL algorithm accurately infers the hidden reward function, significantly reducing the root mean squared spectral norm error (RMSSNE) from 3.3747 to 0.3229 compared to conventional IRL approaches. The inferred reward function effectively identifies anomalous trajectories based on tracking errors and control input violations. The study also shows that the richness of the trajectory data is crucial for accurate model generalization. The DMDc-LQR approach effectively estimates drone trajectories with noise suppression, which is essential for enhancing the accuracy of the machine learning model.

The CPhy-ML framework successfully addresses the challenge of inferring drone intentions by combining data-driven learning with flight physics and control information. The results demonstrate its superior performance over existing methods in both intention classification and prediction. The integration of physics-informed models enhances the robustness and generalization capabilities of the deep learning models, leading to improved accuracy and reliability. The ability to infer the hidden reward function provides valuable insights into the drone's control objective and potential malicious behavior. The findings contribute significantly to the development of advanced counter-drone technologies, providing more accurate and reliable methods for detecting and mitigating threats. The framework's ability to handle both normal and anomalous trajectories makes it suitable for real-world applications.

This research presents a novel CPhy-ML framework for uncovering drone intentions, achieving state-of-the-art performance in trajectory prediction and reward function inference. The integration of flight physics and control information significantly improves the robustness and reliability of the system. Future research could explore incorporating more sophisticated control models, expanding the range of intention classes, and developing methods for real-time implementation of the CPhy-ML framework.

The CPhy-ML framework's performance is dependent on data richness and variability. Limited data diversity may hinder accurate generalization, particularly in trajectory bounding box prediction. The DMOE network's prediction time increases with the number of intention classes. The trajectory prediction performance of the RC network is affected by the richness of the training data. The control-based intention inference method is limited to short-term predictions and may be affected by noise in the control inputs. Future work should address these limitations by exploring data augmentation techniques, improving the efficiency of the DMOE network, enhancing the robustness of the RC network, and developing noise-reduction techniques for control-based inference.