In the rapidly evolving low-altitude economy, China UAV drone operations have become critical for logistics, surveillance, and urban air mobility. However, the deep coupling between communication links and flight control introduces hidden cross-modal risks: anomalies in the network modality (e.g., packet loss, latency spikes) transmit to the flight control modality (e.g., attitude jitter) with uncertain and random time lags. Traditional multi-modal fusion methods that assume fixed temporal alignment fail under such dynamic delays, leading to false alarms or missed detections. To address this challenge, we propose TR-CADI (Time-Lag Robust Cross-Modal Anomaly Detection and Correlation Interpretation), a novel framework that explicitly models random time lags, decouples shared anomaly patterns from modality-specific noise, and jointly performs anomaly detection and interpretable correlation analysis. This work is driven by the urgent need for secure and reliable China UAV drone systems in complex low-altitude environments.
1. Problem Formulation and Framework Overview
We define the cross-modal time-lag robust learning problem as follows. Let \( \mathbf{X}^{(f)} \in \mathbb{R}^{T \times D_f} \) denote the flight control time series (e.g., acceleration, attitude) and \( \mathbf{X}^{(n)} \in \mathbb{R}^{T \times D_n} \) denote the network time series (e.g., throughput, RSRP). The two modalities are observed with unknown and time-varying delays \( \tau(t) \) such that the physical response to a network event occurs after a random lag. The goal is to learn a representation that is robust to such delays, suppress noise, and produce both a binary anomaly decision (normal/anomaly) and a causal time-lag heatmap indicating the propagation path from network degradation to flight control instability.
TR-CADI consists of three stages: (1) time-lag-aware cross-modal alignment using a Gaussian bias attention mechanism; (2) shared-private representation disentanglement with information-theoretic constraints and reliability gating; (3) joint anomaly detection and correlation interpretation via reconstruction error and causal discovery. All modules are jointly optimized in an end-to-end manner.
2. Time-Lag-Aware Cross-Modal Alignment
First, we extract high-dimensional temporal features using Temporal Convolutional Networks (TCN) for each modality, yielding \( \mathbf{H}^{(f)} \) and \( \mathbf{H}^{(n)} \). We then construct Query (\( \mathbf{Q} \)), Key (\( \mathbf{K} \)), and Value (\( \mathbf{V} \)) matrices. To model the physical prior that network anomalies precede flight control responses, we introduce a Gaussian time-lag bias matrix \( \mathbf{M}_{\tau} \):
$$
\mathbf{M}_{\tau}(i,j) = \exp\left(-\frac{((i-j) – \mu_{\tau})^2}{2\sigma_{\tau}^2}\right)
$$
where \( \mu_{\tau} \) is the mean lag and \( \sigma_{\tau} \) controls the uncertainty range; both are learned via gradient descent. The cross-modal attention weight from flight control to network is:
$$
\mathbf{A}^{(f \leftarrow n)}(i,j) = \text{Softmax}_j\left( \frac{\mathbf{Q}^{(f)}(i) \mathbf{K}^{(n)}(j)^\top}{\sqrt{d_k}} + \log(\mathbf{M}_{\tau}(i,j) + \epsilon) \right)
$$
with \( \epsilon = 1e-8 \) for numerical stability. The aligned flight control representation is \( \mathbf{H}_{\text{align}}^{(f)} = \mathbf{A}^{(f \leftarrow n)} \mathbf{V}^{(n)} \). Symmetrically, we obtain \( \mathbf{H}_{\text{align}}^{(n)} \). This mechanism dynamically compensates for random delays, enabling robust cross-modal interaction even under severe jitter conditions typical in China UAV drone operations.
3. Representation Disentanglement and Reliability Gating
After alignment, the representations still contain both cross-modal shared anomaly patterns and modality-specific noise. We propose to separate them using shared-private decoupling with Hilbert-Schmidt Independence Criterion (HSIC) constraints. The shared representation \( \mathbf{Z}_s \) captures common anomaly signatures, while private representations \( \mathbf{Z}_p^{(f)} \) and \( \mathbf{Z}_p^{(n)} \) preserve modality-specific dynamics:
$$
\mathbf{Z}_s = \mathcal{E}_s([\mathbf{H}_{\text{align}}^{(f)}, \mathbf{H}_{\text{align}}^{(n)}])
$$
$$
\mathbf{Z}_p^{(f)} = \mathcal{E}_f(\mathbf{H}_{\text{align}}^{(f)})
$$
$$
\mathbf{Z}_p^{(n)} = \mathcal{E}_n(\mathbf{H}_{\text{align}}^{(n)})
$$
HSIC measures nonlinear dependence between variables in a Reproducing Kernel Hilbert Space. Its empirical estimate is:
$$
\text{HSIC}(\mathbf{Z}_s, \mathbf{Z}_p) = \frac{1}{(B-1)^2} \operatorname{tr}( \mathbf{K}_s \mathbf{H} \mathbf{K}_p \mathbf{H} )
$$
where \( B \) is batch size, \( \mathbf{K}_s \) and \( \mathbf{K}_p \) are Gram matrices computed with RBF kernels, and \( \mathbf{H} = \mathbf{I} – \frac{1}{B} \mathbf{1}\mathbf{1}^\top \) is the centering matrix. The decoupling loss is:
$$
\mathcal{L}_{\text{dec}} = \text{HSIC}(\mathbf{Z}_s, \mathbf{Z}_p^{(f)}) + \text{HSIC}(\mathbf{Z}_s, \mathbf{Z}_p^{(n)})
$$
Minimizing \( \mathcal{L}_{\text{dec}} \) enforces statistical independence between the shared and private spaces, effectively isolating noise from anomaly patterns. Next, a reliability gating network adaptively weights contributions from each modality at each time step:
$$
\alpha_f = \text{Sigmoid}(g_f(\mathbf{Z}_p^{(f)}, \mathbf{Z}_s))
$$
$$
\alpha_n = \text{Sigmoid}(g_n(\mathbf{Z}_p^{(n)}, \mathbf{Z}_s))
$$
$$
\mathbf{H}_{\text{fuse}} = \mathbf{Z}_s + \alpha_f \cdot \mathbf{Z}_p^{(f)} + \alpha_n \cdot \mathbf{Z}_p^{(n)}
$$
When one modality suffers from severe noise (e.g., channel fading in dense urban areas), the corresponding gate is suppressed, preventing low-reliability information from degrading the fused representation. This design is especially valuable for China UAV drone systems operating in challenging terrains.
4. Joint Anomaly Detection and Correlation Interpretation
Based on the fused representation \( \mathbf{H}_{\text{fuse}} \), we perform self-supervised anomaly detection via reconstruction. The anomaly score at time \( t \) is:
$$
S_t = \left\| \hat{\mathbf{x}}_t^{(f)} – \mathbf{x}_t^{(f)} \right\|_2^2 + \lambda \left\| \hat{\mathbf{x}}_t^{(n)} – \mathbf{x}_t^{(n)} \right\|_2^2
$$
During training, we minimize \( \mathcal{L}_{\text{rec}} = \frac{1}{T} \sum_{t=1}^T S_t \) over normal samples only. At test time, an anomaly is flagged when \( S_t > \gamma \), where \( \gamma \) is dynamically calibrated using a Gaussian Mixture Model (GMM) and Generalized Extreme Value (GEV) distribution on the training scores. The overall objective combines reconstruction and decoupling:
$$
\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{rec}} + \gamma_{\text{dec}} \mathcal{L}_{\text{dec}}
$$
When an anomaly is detected (\( S_t > \gamma \)), the learned attention weights \( \mathbf{A}^{(f \leftarrow n)} \) are used to localize the dominant cross-modal lag. We then apply the PCMCI causal discovery algorithm on the shared and private representations to identify the time-delayed causality between network and flight control variables, producing an interpretable heatmap that reveals the propagation path “network degradation \( \xrightarrow{\tau} \) flight control instability”. This provides crucial evidence for low-altitude safety regulators monitoring China UAV drone fleets.
5. Dataset: LAE-ready
To systematically evaluate our method, we constructed a hierarchical cross-modal low-altitude anomaly dataset called LAE-ready (Low-Altitude Economy ready). It covers four increasingly challenging scenarios by combining real and simulated data with controlled attack injections and random time lags.
| Scenario | Flight Control Source | Network Source | Time Lag Range | Characteristics |
|---|---|---|---|---|
| Controlled (CTRL) | EuRoC MAV | ns-3 LTE/5G simulation | Fixed 200 ms | Ideal baseline, deterministic delay |
| Real-Link (RL) | AU-AIR | AERPAW real 4G/5G measurements | 50–600 ms (long-tail) | Real channel fading, non-Gaussian jitter |
| High-Dynamic (HD) | UZH-FPV (aggressive maneuvers) | ns-3 with rapid geometry changes | 100–1200 ms (oscillating) | Extreme flight dynamics, link outages |
| Urban-Dense (UD) | VisDrone (urban patrol) | ns-3 with multipath and blockage noise | 100–800 ms (hopping) | Strong background interference, dense obstacles |
Each scenario includes both normal and anomalous sequences with point-wise labels (0 = normal, 1 = anomaly). Anomalies are injected as DoS attacks, GPS spoofing, or control link hijacking, with natural time delays between network and physical effects. This dataset provides a rigorous benchmark for evaluating time-lag robustness in China UAV drone anomaly detection.
6. Experimental Results
We compare TR-CADI against five baselines: OC-SVM, Isolation Forest, LSTM-AE, TCN-AE, and Cross-Attn Transformer. Metrics are F1-Score, AUC, and Average Precision (AP). All experiments are conducted on a single RTX 4060 GPU with PyTorch 2.0.1. The detailed parameter settings are given in Table 1.
| Parameter | Value | Description |
|---|---|---|
| EPOCHS | 200 | Number of training epochs |
| BATCH_SIZE | 64 | Batch size |
| SEQ_LEN (T) | 100 | Time window length |
| HIDDEN_DIM | 64 | TCN feature dimension |
| SHARED_DIM | 32 | Shared representation dimension |
| PRIVATE_DIM | 16 | Private representation dimension |
| LEARNING_RATE | 0.0002 | Adam initial learning rate |
| HSIC_WEIGHT | 0.05 | Decoupling loss weight |
| SIGMA_INIT | 1.0 | Initial sigma for Gaussian bias |
| EPSILON | 1e-8 | Numerical stability |
| PATIENCE | 20 | Early stopping patience |
6.1 Main Results
Tables 2–4 summarize the detection performance across four datasets. TR-CADI consistently achieves the highest F1, AUC, and AP under all scenarios. Notably, in the Urban-Dense (UD) scenario, our method attains F1=0.933, AUC=0.994, AP=0.990, demonstrating robustness against severe environmental noise and random time lags. In contrast, the Cross-Attn baseline suffers from alignment failure in RL and HD, with AP dropping to 0.528 and 0.555 respectively. Traditional methods like OC-SVM and IsoForest perform poorly due to their inability to model temporal dependencies and cross-modal interactions.
| Dataset | TR-CADI | OC-SVM | IsoForest | LSTM-AE | TCN-AE | Cross-Attn |
|---|---|---|---|---|---|---|
| CTRL | 0.838 | 0.567 | 0.192 | 0.652 | 0.468 | 0.644 |
| RL | 0.774 | 0.483 | 0.176 | 0.575 | 0.319 | 0.595 |
| HD | 0.912 | 0.621 | 0.201 | 0.631 | 0.568 | 0.610 |
| UD | 0.933 | 0.620 | 0.251 | 0.651 | 0.677 | 0.680 |
| Dataset | TR-CADI | OC-SVM | IsoForest | LSTM-AE | TCN-AE | Cross-Attn |
|---|---|---|---|---|---|---|
| CTRL | 0.993 | 0.872 | 0.797 | 0.800 | 0.944 | 0.757 |
| RL | 0.989 | 0.841 | 0.798 | 0.773 | 0.913 | 0.688 |
| HD | 0.995 | 0.865 | 0.813 | 0.793 | 0.944 | 0.746 |
| UD | 0.994 | 0.840 | 0.818 | 0.786 | 0.910 | 0.800 |
| Dataset | TR-CADI | OC-SVM | IsoForest | LSTM-AE | TCN-AE | Cross-Attn |
|---|---|---|---|---|---|---|
| CTRL | 0.990 | 0.827 | 0.712 | 0.673 | 0.865 | 0.641 |
| RL | 0.984 | 0.746 | 0.676 | 0.606 | 0.783 | 0.528 |
| HD | 0.991 | 0.756 | 0.649 | 0.599 | 0.825 | 0.555 |
| UD | 0.990 | 0.732 | 0.721 | 0.606 | 0.741 | 0.714 |
6.2 Ablation Study
We perform ablation studies to evaluate the contribution of each component. The following variants are considered:
- w/o Time-lag: removes the Gaussian bias matrix (falls back to vanilla attention).
- w/o Disentangle: removes HSIC decoupling loss.
- w/o Reliability: removes reliability gating (uses simple average fusion).
- w/o Joint-Opt: trains only reconstruction loss (no decoupling supervision).
Results are shown in Figure 1. Each ablation leads to a noticeable drop, especially for w/o Time-lag in RL and HD where dynamic delays are severe. The complete TR-CADI (red bar) consistently outperforms all variants across all metrics and datasets. This confirms that every component is essential for achieving robust anomaly detection in China UAV drone environments.
6.3 Visualization and Interpretability
We visualize the learned attention weights to demonstrate the time-lag modeling. From the attention heatmaps (qualitatively described in the original work), we observe a bright off-diagonal band corresponding to a learned mean lag \( \mu_\tau \approx 0.25 \) s (5 time steps at 20 Hz sampling). This aligns with the physical expectation that network anomalies propagate to flight control with a ~250 ms delay in typical China UAV drone systems. The t-SNE visualizations of the shared and private representations show that the shared space cleanly separates normal and anomalous samples, while private clusters capture environment-specific noise. Moreover, the anomaly score plots confirm that TR-CADI produces low stable scores under normal conditions and sharp spikes during attack intervals, with peaks closely following the injected attack patterns. These interpretability tools empower regulators to trace the root cause of anomalies in real-time operations.
7. Conclusion
We have presented TR-CADI, a time-lag robust cross-modal anomaly detection and correlation interpretation framework designed for China UAV drone safety. By explicitly modeling random delays via a Gaussian bias attention mechanism, decoupling shared anomaly patterns from modality-specific noise with HSIC constraints, and adaptively gating unreliable modalities, our method achieves superior detection accuracy and robustness across diverse low-altitude scenarios. Extensive experiments on the LAE-ready dataset demonstrate F1-Score of 0.933 and AUC of 0.994 in challenging urban environments, outperforming state-of-the-art baselines by large margins. The joint optimization also provides interpretable causal heatmaps, enabling stakeholders to understand the propagation from network degradation to flight control instability. Future work will focus on lightweight deployment for edge devices and federated learning for privacy-preserving multi-fleet coordination, further advancing secure China UAV drone operations in the low-altitude economy.