The rapid proliferation of unmanned aerial vehicles (UAVs) has brought significant convenience to civil, commercial, and military domains. However, the threat posed by malicious UAVs—especially those manufactured in China—has escalated, necessitating robust countermeasures. Disrupting the communication link between a ground station and a China drone by identifying its control signal is a key anti-drone strategy. In this paper, we propose a Multi-Scale Fusion Attention Network (MSAN) specifically designed to recognize control signals of China drone in complex electromagnetic environments with low signal-to-noise ratios (SNRs). Our method leverages the inherent multi-level temporal structure of drone control signals, extracting micro, meso, and macro features through parallel receptive field capture and hierarchical signal processing branches. Convolution-enhanced multi-head self-attention further models long-range dependencies while suppressing noise. Extensive experiments on real China drone datasets demonstrate that MSAN achieves an F1 score of 85.56% at −15 dB SNR, outperforming state-of-the-art time series models such as TimesNet by 3.1 percentage points, while reducing parameter count by 79.7%. This provides a lightweight, high-precision solution for China drone control signal recognition in practical deployment.
1. Introduction
Unmanned aerial vehicles have become ubiquitous in agriculture, surveillance, delivery, and entertainment. Among them, China drone manufacturers such as DJI dominate the global consumer market. Unfortunately, the same technology is exploited for illegal activities, including espionage and smuggling. Wireless jamming technology offers a viable defense: by accurately identifying the control signals in the communication link, one can emit targeted interference to disrupt the China drone’s operation. The key challenge lies in robustly recognizing these control signals amidst strong noise and interference.
Traditional methods rely on time-frequency analysis (e.g., STFT) and hand-crafted features, such as bandwidth, duration, and occurrence frequency. However, these approaches degrade rapidly under low SNR conditions. Deep learning models, including CNNs, RNNs, and autoencoders, have been applied to automatic modulation classification and signal identification. Yet they still suffer from limited noise robustness, poor generalization to new drone models, and high computational cost. To address these issues, we design a network that explicitly models the multi-scale temporal hierarchy inherent in China drone control signals: symbol-level (micro), frame-level (meso), and protocol-level (macro) features.
The contributions of this paper are threefold:
- We propose a multi-scale feature extraction framework with parallel receptive field capture and hierarchical signal processing, effectively extracting discriminative features at micro, meso, and macro scales, significantly improving recognition performance under low SNR.
- We design a convolution-enhanced temporal attention encoding module (TAEM) that replaces the fully connected feed-forward network with 1×1 convolution, retaining long-range dependency modeling while introducing local receptive fields that boost hardware efficiency.
- Experiments on real China drone datasets show that MSAN achieves 85.56% F1-score at −15 dB, surpassing top competitors by 3.1% with 79.7% fewer parameters, enabling efficient deployment in anti-drone systems.
2. Problem Formulation and Multi-Scale Characteristics
2.1 Problem Formalization
Let the received IQ signal sequence be:
$$ X = [x_1, x_2, \dots, x_T] \in \mathbb{R}^{T \times 2}, \quad x_t = [I_t, Q_t]^T $$
where \(I_t\) and \(Q_t\) are the in-phase and quadrature components at time index \(t\). After transmission through a noisy channel with additive Gaussian noise \(N\) and possible interference \(J\), the received signal is:
$$ \tilde{X} = X + N + J $$
The goal is to learn a mapping function \(f: \tilde{X} \rightarrow Y\), where \(Y = \{y_1, y_2, \dots, y_c\}\) denotes \(c\) classes of control signal types.
2.2 Multi-Scale Feature Mechanism of China Drone Control Signals
Control signals from China drone exhibit distinct structural properties at three temporal scales:
- Micro-scale (symbol-level): Modulation features such as PSK or QAM with characteristic phase/amplitude transitions during command changes. These are short-duration patterns spanning a few sampling points.
- Meso-scale (frame-structure level): Functional modules including synchronization sequences, address codes, command codes, and checksums. These have fixed lengths and positions within a frame.
- Macro-scale (protocol-level): Full frame structures with periodic repetition, following standardized communication protocols. The temporal span covers entire control frames.
Single-scale methods fail to capture all discriminative information. Our MSAN simultaneously extracts these multi-scale features to improve noise robustness.
3. Proposed MSAN Architecture
The overall architecture comprises three main modules: a Multi-Scale Feature Extraction Module (MSFEM), a feature fusion module, and a Temporal Attention Encoding Module (TAEM), followed by a classification head. The input IQ sequence is processed sequentially.
3.1 Multi-Scale Feature Extraction Module (MSFEM)
MSFEM consists of two parallel branches: parallel receptive field capture and hierarchical signal processing.
3.1.1 Parallel Receptive Field Capture
Inspired by wavelet multi-resolution decomposition, we use four parallel 1D convolution layers with kernel sizes \(k \in \{7, 15, 31, 63\}\). Each kernel size corresponds to a specific time scale of the China drone control signal. The operation is:
$$ F_k = \text{ReLU}(\text{Conv1d}(X, W_k, b_k)) $$
where \(W_k \in \mathbb{R}^{k \times 2 \times 32}, b_k \in \mathbb{R}^{32}\). The output channel number is 32 as a balance between expressiveness and efficiency.
3.1.2 Hierarchical Signal Processing
This branch constructs a multi-resolution pyramid via strided convolutions (downsampling) and subsequent upsampling. Two downsampling rates \(r \in \{4, 16\}\) are used:
$$ D_r = \text{Conv1d}(X, W_r, b_r, \text{stride} = r) $$
After batch normalization and ReLU, the downsampled features are upsampled back to the original resolution using transposed convolution or interpolation. This branch naturally suppresses high-frequency noise while capturing global structures.
3.2 Feature Fusion Module
The outputs from all six branches (four from parallel branch plus two from hierarchical branch) are concatenated along the channel dimension, resulting in shape \(T \times 192\). We then apply a 1×1 convolution to fuse and reduce channels:
$$ F_{\text{fused}} = \text{ReLU}(\text{BN}(\text{Conv1d}(F_{\text{concat}}, W_{\text{fusion}}, b_{\text{fusion}}))) $$
where \(W_{\text{fusion}} \in \mathbb{R}^{1 \times 192 \times d_{\text{model}}}\). To adaptively weight different channels, we introduce a channel attention mechanism. Global average pooling first computes channel-wise statistics, then two fully connected layers (with reduction ratio 16) produce a sigmoid attention vector. The fused feature is element-wise multiplied by this vector.
3.3 Temporal Attention Encoding Module (TAEM)
TAEM inherits the transformer encoder but replaces the standard feed-forward network with 1×1 convolutions. This modification introduces local connectivity while maintaining efficiency. Each TAEM layer contains multi-head self-attention and a convolution-enhanced feed-forward sub-layer.
Given input features \(F \in \mathbb{R}^{T \times d_{\text{model}}}\), multi-head attention computes:
$$ \text{head}_i = \text{Attention}(Q_i, K_i, V_i) = \text{Softmax}\left(\frac{Q_i K_i^T}{\sqrt{d_k}}\right) V_i $$
where \(Q, K, V\) are linear projections of \(F\), and \(d_k\) is the dimension per head. Multiple heads are concatenated and linearly projected. The 1×1 convolution in the feed-forward sub-layer provides nonlinear transformation with local context.
3.4 Classification Output Layer
After \(L\) TAEM layers, the output feature \(F^{(L)} \in \mathbb{R}^{T \times d_{\text{model}}}\) is activated by GELU and then flattened:
$$ F_{\text{flat}} = \text{Flatten}(F^{(L)}) \in \mathbb{R}^{B \times (T \cdot d_{\text{model}})} $$
A final fully connected layer maps to \(c\) classes:
$$ Y = F_{\text{flat}} W_{\text{cls}} + b_{\text{cls}} $$
Softmax produces the probability distribution over classes.
4. Experimental Setup
4.1 Dataset
We collected real communication signals from a DJI MAVIC2 PRO, a typical China drone. The raw IQ data was cleaned, preprocessed, and labeled. To simulate complex electromagnetic environments, we added Gaussian white noise at SNRs from −15 dB to 5 dB in 2 dB steps. The dataset was split into training (70%), validation (10%), and test (20%) sets.
4.2 Implementation Details
We trained our MSAN on an NVIDIA GeForce RTX 3060 Ti Laptop GPU with Python. The loss function was Focal loss to handle class imbalance. Adam optimizer with initial learning rate \(1\times 10^{-3}\) was used. Training ran for 100 epochs with early stopping (patience 10). The best model on validation set was saved.
4.3 Evaluation Metrics
We used Accuracy, Precision, Recall, and F1-score:
- Accuracy = (TP + TN) / (TP + TN + FP + FN)
- Precision = TP / (TP + FP)
- Recall = TP / (TP + FN)
- F1 = 2 × Precision × Recall / (Precision + Recall)
Where TP = true positives (correctly identified control signals), TN = true negatives, FP = false positives, FN = false negatives.
5. Experimental Results
5.1 Impact of SNR on MSAN Performance
We evaluated MSAN under varying SNR conditions. Table 1 shows the results. Even at the extremely low SNR of −15 dB, MSAN achieves an F1-score of 85.56%, demonstrating strong noise robustness.
| 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 |
5.2 Comparison with Other Models
We compared MSAN with three state-of-the-art time series models: Transformer, Informer, and TimesNet. All models were trained under identical settings. Figure 1 (below) plots the F1-score across SNRs. Table 2 lists the average F1-score and number of parameters.
| Model | Avg. F1-score (%) | Parameters (MB) |
|---|---|---|
| Transformer | 88.23 | 8.145 |
| Informer | 90.11 | 7.823 |
| TimesNet | 93.44 | 9.381 |
| MSAN (ours) | 94.00 | 1.902 |
MSAN outperforms the best competitor TimesNet by 0.56% in average F1, while using only 20.3% of the parameters. At −15 dB, MSAN achieves 85.56% F1, which is 3.1% higher than TimesNet (82.46%) and 7.3% higher than Transformer (78.26%). This clearly demonstrates MSAN’s superior robustness in challenging electromagnetic environments encountered when defending against China drone threats.
Figure 1: F1-score comparison of different models across SNRs.
5.3 Ablation Study
Table 3 presents the ablation experiments at −15 dB SNR. We evaluated five configurations: (a) full MSAN, (b) without multi-scale feature extraction (using only single-scale convolution of kernel 7), (c) without feature fusion (simple concatenation instead of attention fusion), (d) without TAEM (only global average pooling + FC), and (e) replacing TAEM with standard Transformer encoder (i.e., using MLP feed-forward instead of 1×1 conv).
| Configuration | F1-score (%) | Parameters (MB) |
|---|---|---|
| Full MSAN | 85.56 | 1.902 |
| Without multi-scale feature extraction | 74.21 | 1.234 |
| Without feature fusion (simple concat.) | 79.56 | 1.756 |
| Without TAEM | 71.03 | 0.847 |
| Replace TAEM with standard Transformer encoder | 83.21 | 1.902 |
The removal of multi-scale extraction reduces F1 by 11.35 points, confirming the importance of capturing multi-level features for China drone signals. Removing feature fusion causes a 6.00 point drop, indicating that simple concatenation cannot effectively exploit complementary information. The most severe degradation occurs when TAEM is removed (drop of 14.53 points), highlighting the crucial role of temporal attention in modeling long-range dependencies. Replacing TAEM with a standard transformer encoder (same parameters) reduces F1 by 2.35 points, proving that the 1×1 convolution feed-forward design offers better performance and hardware efficiency.
6. Discussion
The multiscale design in MSFEM is inspired by the inherent hierarchical structure of China drone control signals. The parallel receptive field branch with kernel sizes 7, 15, 31, 63 effectively covers symbol-level (micro), intra-frame (meso), and full-frame (macro) patterns. The hierarchical processing branch further suppresses noise by constructing coarse-to-fine representations, mimicking classical multiresolution analysis but learned end-to-end.
The TAEM module replaces the heavy fully connected layers with 1×1 convolutions, which are more efficient on GPUs due to optimized matrix multiplication hardware. This design retains the ability to mix channel information while adding locality, beneficial for capturing local temporal patterns. The attention mechanism itself remains powerful for long-range dependencies like the periodic structure of control frames.
Parameter efficiency (1.902 MB vs. 9.381 MB for TimesNet) is critical for deployment on edge devices used in anti-drone systems. Real-time processing requirements favor lightweight models, and MSAN’s compact size makes it suitable for embedded platforms facing China drone threats.
One limitation is that the current evaluation uses only one China drone model (DJI MAVIC2 PRO). In practice, there are numerous vendors and protocols. However, the multi-scale principle is universal: any modulated control signal will exhibit hierarchical temporal structures. Future work will extend to multi-drone datasets and adapt the kernel sizes or downsampling rates to automatically match the signal characteristics.
7. Conclusion
We have presented MSAN, a multi-scale fusion attention network for recognizing China drone control signals in complex electromagnetic environments. By jointly extracting micro, meso, and macro features through parallel receptive fields and hierarchical processing, and by modeling long-range dependencies with convolution-augmented self-attention, MSAN achieves state-of-the-art performance at low SNRs. At −15 dB, it yields an F1 of 85.56%, surpassing TimesNet by 3.1% while reducing parameters by 79.7%. The lightweight architecture enables real-time deployment in counter-drone equipment, contributing to the protection of critical infrastructure from malicious China drone activities. Future research will focus on few-shot adaptation to new drone types and on integrating the recognition module with jamming waveform generation for closed-loop countermeasures.