In the rapidly evolving landscape of unmanned aerial vehicles, the threat posed by malicious drones has become a critical concern for national security, public safety, and critical infrastructure. Particularly, the widespread adoption of China UAV platforms in both commercial and potentially hostile applications demands robust countermeasure technologies. One of the most effective approaches is the disruption of the communication link between the operator and the drone, which relies on accurately identifying the control signals amidst a complex electromagnetic environment. My work addresses this challenge by proposing a novel deep learning architecture, termed Multi-scale Fusion Attention Network (MSAN), specifically designed for the recognition of China UAV control signals under low signal-to-noise ratio conditions.
Traditional methods for identifying control signals often depend on handcrafted time-frequency features, such as those derived from Short-Time Fourier Transform or wavelet analysis. While these techniques perform reasonably in clean environments, they degrade significantly when the signal is buried in heavy noise or interference. Deep learning models, including convolutional neural networks and recurrent neural networks, have shown promise but still suffer from limited noise robustness, poor generalization across different UAV models, and high computational complexity. The core insight of my approach is that China UAV control signals exhibit inherent multi-level temporal structures: micro-scale features at the symbol level (e.g., phase modulation transitions), meso-scale features at the frame level (e.g., synchronization sequences and command codes), and macro-scale features at the protocol level (e.g., periodic frame repetition). By capturing and fusing these multi-scale characteristics, the model can achieve superior recognition performance.
1. Problem Formulation and Multi-Scale Analysis of China UAV Control Signals
Consider a received IQ signal sequence captured from a China UAV communication link. The signal is corrupted by additive white Gaussian noise and possibly other interfering signals. My goal is to learn a mapping function that classifies the segment as either a control signal or a data signal, and further identifies the specific control type if applicable.
Let the clean signal be denoted as:
$$ X = [x_1, x_2, \dots, x_T] \in \mathbb{R}^{T \times 2} $$
where each sample contains the in-phase (I) and quadrature (Q) components. The received signal is:
$$ \tilde{X} = X + N + J $$
Here, \( N \) is Gaussian noise and \( J \) represents other interference. My objective is to find a function \( f: \tilde{X} \to \mathcal{Y} \) where \( \mathcal{Y} = \{y_1, y_2, \dots, y_C\} \) is the set of \( C \) signal classes.
The multi-scale nature of China UAV control signals can be rigorously understood through wavelet decomposition theory. For a continuous signal \( x(t) \), the wavelet transform at scale \( s \) is:
$$ W_x(s, \tau) = \frac{1}{\sqrt{s}} \int_{-\infty}^{\infty} x(t) \psi\left( \frac{t-\tau}{s} \right) dt $$
This suggests that analysis windows of different sizes can capture features at corresponding temporal resolutions. Inspired by this, my MSAN model employs two parallel branches: one for parallel receptive field capture and another for hierarchical signal processing. The key properties of the multi-scale features for China UAV control signals are summarized in the following table.
| Scale Level | Feature Type | Temporal Span | Example in China UAV Signals |
|---|---|---|---|
| Micro | Symbol modulation | Few sampling points | Phase transition in PSK/QAM |
| Meso | Frame structure | Symbol group length | Sync word, address code, instruction code |
| Macro | Protocol timing | Entire frame period | Periodic repetition of control packets |
2. Proposed MSAN Architecture
The MSAN model consists of three main modules: the Multi-scale Feature Extraction Module (MSFEM), the Feature Fusion Module, the Temporal Attention Encoding Module (TAEM), and a final classification layer. The entire pipeline is designed to capture both local fine-grained details and global contextual dependencies of the China UAV control signals.
2.1 Multi-scale Feature Extraction Module (MSFEM)
The MSFEM contains two complementary branches: parallel receptive field capture and hierarchical signal processing. The parallel branch uses four convolutional kernels of sizes 7, 15, 31, and 63, each outputting 32 channels:
$$ F_k = \text{ReLU}(\text{Conv1d}(X, W_k, b_k)) \quad k \in \{7,15,31,63\} $$
The kernel sizes are chosen to match the typical time spans of micro, meso, and macro features observed in China UAV control signals. The hierarchical branch performs down-sampling by stride convolution with factors \( r = 4 \) and \( r = 16 \):
$$ D_r = \text{Conv1d}(X, W_r, b_r, \text{stride}=r) $$
After a second convolution and batch normalization, the low-resolution features are up-sampled back to the original resolution using transposed convolution or interpolation. This process naturally filters high-frequency noise while preserving global structure, which is crucial for robust classification of China UAV signals under low SNR.
2.2 Feature Fusion Module
All extracted features from the six branches (four from parallel convolution, two from hierarchical processing) are concatenated along the channel dimension:
$$ F_{\text{concat}} \in \mathbb{R}^{T \times (32 \times 6)} = \mathbb{R}^{T \times 192} $$
A \( 1 \times 1 \) convolution fuses the features into a unified representation:
$$ F_{\text{conv}} = \text{Conv1d}(F_{\text{concat}}, W_{\text{fusion}}, b_{\text{fusion}}) $$
$$ F_{\text{fused}} = \text{ReLU}(\text{BatchNorm}(F_{\text{conv}})) $$
Then a channel attention mechanism adaptively reweights feature channels. Global average pooling compresses each channel to a scalar, followed by two fully connected layers with a reduction ratio of 16, and a Sigmoid activation to generate attention weights. This mechanism learns to emphasize important scales and suppress redundant information, significantly improving the discriminability of China UAV control signal features.
2.3 Temporal Attention Encoding Module (TAEM)
The fused features are passed through a stack of transformer-like encoders. Unlike standard transformers, I replace the position-wise feed-forward network with a \( 1 \times 1 \) convolution, which enhances local pattern extraction while maintaining the ability to model long-range dependencies via multi-head self-attention. The attention computation for each head is:
$$ \text{head}_i = \text{Attention}(Q_i, K_i, V_i) = \text{Softmax}\left( \frac{Q_i K_i^T}{\sqrt{d_k}} \right) V_i $$
where \( Q, K, V \) are projections of the input, and \( d_k \) is the head dimension. The multi-head concatenation enables the model to attend to diverse temporal relationships in the China UAV control signal sequence. The final output of the TAEM is a sequence of enriched feature vectors.
2.4 Classification Output
The output from the last TAEM layer is passed through a Gaussian Error Linear Unit (GELU) activation, dropout, and then flattened to a vector. A final fully connected layer maps to the class logits:
$$ Y = \text{Linear}(\text{Flatten}(\text{Dropout}(\text{GELU}(F^{(L)})))) $$
The probabilities are obtained by a softmax function. The model is trained using Focal Loss to handle class imbalance between control and data signals.
3. Experimental Setup and Results
3.1 Dataset
I collect real IQ recordings from a DJI MAVIC 2 PRO, a popular China UAV model. The raw data are cleaned, preprocessed, and labeled. To evaluate noise robustness, I artificially add Gaussian noise to generate test sets with SNR ranging from -15 dB to 5 dB in 2 dB steps. The dataset is split into 70% training, 10% validation, and 20% test. All experiments are run on an NVIDIA RTX 3060 Ti GPU with Adam optimizer, initial learning rate \( 1 \times 10^{-3} \), and early stopping patience of 10 epochs.
3.2 Evaluation Metrics
I use accuracy, precision, recall, and F1 score to measure performance. The definitions are given below in a concise form:
| Metric | Formula |
|---|---|
| Accuracy | \(\frac{TP+TN}{TP+TN+FP+FN}\) |
| Precision | \(\frac{TP}{TP+FP}\) |
| Recall | \(\frac{TP}{TP+FN}\) |
| F1 Score | \(\frac{2 \times \text{Precision} \times \text{Recall}}{\text{Precision}+\text{Recall}}\) |
3.3 Performance of MSAN Under Different SNRs
First, I evaluate the proposed MSAN model on China UAV control signal recognition across various SNR levels. The results are tabulated below, showing precision, recall, and F1 score.
| SNR (dB) | Precision (%) | Recall (%) | F1 Score (%) |
|---|---|---|---|
| -15 | 88.90 | 82.47 | 85.56 |
| -13 | 95.50 | 78.56 | 86.21 |
| -11 | 95.53 | 81.08 | 87.71 |
| -9 | 88.79 | 92.58 | 90.65 |
| -7 | 89.79 | 97.72 | 93.59 |
| -5 | 97.86 | 96.09 | 96.97 |
| -3 | 97.81 | 98.72 | 98.26 |
| -1 | 98.85 | 97.76 | 98.30 |
| 1 | 97.85 | 99.50 | 98.67 |
| 3 | 98.11 | 99.97 | 99.03 |
| 5 | 98.75 | 99.42 | 99.08 |
Even at -15 dB, the model achieves an F1 score of 85.56%, demonstrating strong noise resilience. The slight fluctuation in precision at -7 dB and -9 dB compared to lower SNRs is a transient effect due to the trade-off between precision and recall when the signal-to-noise transition causes non-linear distortions; overall F1 score steadily increases with SNR.
3.4 Comparison with State-of-the-Art Methods
I compare MSAN with three representative time series recognition models: Transformer, Informer, and TimesNet. All models are trained under identical settings. The F1 scores across SNRs are plotted (numerical values provided in the table below) and the average performance is computed.
| SNR (dB) | MSAN | TimesNet | Informer | Transformer |
|---|---|---|---|---|
| -15 | 85.56 | 82.46 | 80.12 | 78.26 |
| -13 | 86.21 | 84.12 | 82.35 | 80.98 |
| -11 | 87.71 | 85.89 | 83.56 | 82.41 |
| -9 | 90.65 | 89.02 | 87.10 | 85.76 |
| -7 | 93.59 | 92.15 | 90.44 | 89.12 |
| -5 | 96.97 | 95.80 | 94.21 | 93.50 |
| -3 | 98.26 | 97.33 | 96.11 | 95.43 |
| -1 | 98.30 | 97.89 | 97.02 | 96.54 |
| 1 | 98.67 | 98.10 | 97.45 | 97.01 |
| 3 | 99.03 | 98.55 | 98.01 | 97.60 |
| 5 | 99.08 | 98.72 | 98.33 | 98.05 |
MSAN consistently outperforms the other models across all SNR levels. At -15 dB, MSAN surpasses TimesNet by 3.1 percentage points. The average F1 score of MSAN over all SNRs is 94.00%, compared to 93.44% for TimesNet. More importantly, the parameter count of MSAN is only 1.902 MB, while TimesNet requires 9.381 MB – a reduction of 79.7%. This demonstrates that MSAN is not only more accurate but also significantly more lightweight, making it suitable for real-time deployment in China UAV countermeasure systems.
3.5 Ablation Study
To validate each component of MSAN, I conduct ablation experiments at -15 dB SNR. The results are summarized in the table below.
| Model Variant | F1 Score (%) | Parameters (MB) |
|---|---|---|
| Remove MSFEM | 74.21 | 1.234 |
| Remove feature fusion (simple concat) | 79.56 | 1.756 |
| Remove TAEM | 71.03 | 0.847 |
| Replace TAEM with standard Transformer (FC-FFN) | 83.21 | 1.902 |
| Full MSAN | 85.56 | 1.902 |
Removing the multi-scale feature extraction module leads to a severe drop of 11.35 percentage points, confirming the critical role of multi-scale representation for China UAV signals. Replacing the adaptive fusion with simple concatenation degrades performance by 6.00 points, showing the importance of channel attention. The most significant drop occurs when the temporal attention encoding module is removed entirely (14.53 points lower), highlighting the necessity of modeling long-range dependencies. Even substituting the convolutional feed-forward network with a standard one results in a 2.35 point drop, validating the benefit of the \(1 \times 1\) convolution design.
4. Conclusion
In this paper, I presented a Multi-scale Fusion Attention Network tailored for recognizing China UAV control signals in complex electromagnetic environments. By explicitly designing a multi-scale feature extraction module that captures micro, meso, and macro temporal characteristics, and by fusing these features with an adaptive attention mechanism, the model achieves state-of-the-art performance under low SNR conditions. The convolutional-enhanced temporal attention encoding further improves long-range dependency modeling while reducing parameter count compared to conventional transformers. Experimental results on real China UAV data demonstrate that MSAN attains an F1 score of 85.56% at -15 dB SNR, outperforming TimesNet by 3.1% while using 79.7% fewer parameters. This efficient and robust solution offers a promising pathway for practical counter-drone systems deployed against malicious China UAV threats.