Technical Summary: Modulation Feature Enhancement with a Multi-Stage Attention Network for Underwater Acoustic Target Recognition

1. Problem Statement

Underwater acoustic target recognition (UATR) is a critical component of maritime applications but faces significant challenges due to the complex and diverse nature of ship-radiated noise and the scarcity of real-world data. The primary difficulties include:

• Complex Signal Characteristics: Ship-radiated noise consists of mechanical, propeller, and hydrodynamic components. The propeller noise contains crucial modulation features (shaft frequency, blade number), but these exhibit varying modulation strengths across frequency bands. Conventional full-band Demodulation of Envelope on Noise (DEMON) analysis often introduces errors due to this non-uniformity.
• Inadequate Feature Representation: Existing methods often rely on uniform frequency band segmentation or single-stage attention mechanisms that fail to capture the distinct characteristics of features at different network depths. Furthermore, standard attention mechanisms (like Spatial Attention in CBAM) assume consistent spatial positions of key information across channels, which is unsuitable for DEMON spectral features where key information is distributed differently across channels.
• Data Imbalance: Real-world datasets suffer from severe class imbalance (long-tailed distribution) and "hard-easy" sample imbalance due to varying operating conditions and environments. Standard loss functions often lead to models that overfit head classes and underfit tail classes.

2. Methodology

The authors propose a robust deep learning framework comprising three core components: a novel feature extraction and fusion scheme, a Multi-Stage Multi-Type Attention Network (MMATT), and an Adjustable Class-Balanced Focal Loss (ACBFL).

2.1. Feature Extraction and Fusion

To generate high-fidelity 2-D DEMON spectral features, the authors combine Variational Mode Decomposition (VMD) and the 3/2-D spectrum:

• VMD Decomposition: Instead of uniform band-pass filtering, the signal is adaptively decomposed into Intrinsic Mode Functions (IMFs) using VMD. This addresses the non-uniform modulation characteristics of ship noise.
• 3/2-D Spectrum: For each IMF, the 3/2-D spectrum is computed. This high-order statistic suppresses additive Gaussian noise and eliminates non-phase-coupled harmonic components, preserving phase information crucial for DEMON analysis.
• Fusion Strategy: The final 2-D feature is formed by fusing the amplitude spectrum and the squared 3/2-D spectrum. Specifically, the product of the squared 3/2-D spectrum and the amplitude spectrum is added to the amplitude spectrum. This enhances key frequency information while retaining detailed spectral data.

2.2. Multi-Stage Multi-Type Attention Network (MMATT)

The recognition model utilizes a 1-D Convolutional Neural Network (CNN) integrated with a Multi-Stage Multi-Type Attention Mechanism. Unlike previous works that apply a single attention type, MMATT places distinct attention modules at different network depths to adapt to the evolving feature representations:

• Attention Block 1 (Shallow Layer): Employs the Residual Channel-Independent Spectral Attention Mechanism (R-CISAM). This module uses depthwise convolution to apply distinct spectral attention weights to each channel (IMF component) independently, avoiding interference caused by pooling across channels. It includes a residual connection to preserve original features.
• Attention Block 2 (Middle Layer): Employs the Multi-Scale Separate-and-Fuse Spectral Attention Mechanism (MS-SFSAM). This module uses dilated depthwise separable convolutions with multiple dilation rates (4, 8, 16) to capture global contextual relationships and multi-scale information, which are then fused across channels.
• Attention Block 3 (Deep Layer): Employs a standard Channel Attention Mechanism (SE Block) to recalibrate channel-wise features based on global information.

2.3. Adjustable Class-Balanced Focal Loss (ACBFL)

To address class imbalance and hard-easy sample imbalance, the authors propose ACBFL, which combines two strategies:

• Adjustable Focal Loss (AFL): Modifies the standard focal loss by introducing an adjustable parameter $\beta$ to the modulating factor $(1 - \beta p)^\gamma$. This allows the loss function to vary between cross-entropy and original focal loss depending on the task.
• Adjustable Class-Balanced (ACB): Introduces a tunable parameter $q$ to the class weights. The weights are proportional to $(1/m_y)^q$, where $m_y$ is the number of samples in class $y$. When $q=0$, no balancing occurs; when $q=1$, weights are inversely proportional to class frequencies.
• Combined Formulation: The final loss function integrates both adjustments, allowing flexibility to adapt to varying degrees of imbalance in different datasets.

3. Key Contributions

The paper claims the following specific contributions:

1. Novel Feature Extraction: A 2-D DEMON spectral feature extraction and fusion method based on VMD and the 3/2-D spectrum, which effectively captures modulation envelope information and suppresses noise.
2. MMATT Architecture: A recognition network integrating a 1-D CNN with a Multi-Stage Multi-Type Attention Mechanism that utilizes different attention modules at different network depths.
3. New Attention Mechanisms: The proposal of two specific mechanisms:
   • R-CISAM: For channel-independent spectral weighting in shallow layers.
   • MS-SFSAM: For multi-scale feature enhancement in middle layers.
4. ACBFL: An Adjustable Class-Balanced Focal Loss that alleviates the negative effects of long-tailed data distribution while offering flexible adaptability.
5. Empirical Validation: Comprehensive experiments on the real-world ShipsEar dataset demonstrating the effectiveness of the proposed components.

4. Experimental Results

Experiments were conducted on the ShipsEar dataset (11 ship classes + natural noise) using 5-fold cross-validation. The proposed full model (VMD-Fusion-CNN-MMATT-ImFL) was compared against baselines and ablated variants.

• Feature Extraction: The proposed VMD-Fusion-CNN achieved an Overall Accuracy (OA) of 89.50%, F1-score of 84.90%, and Average Accuracy (AA) of 82.58%. This significantly outperformed the traditional DEMON approach (Trad-CNN: 73.21% OA) and the band-pass filtered approach (Filter-CNN: 85.07% OA).
• Model Performance: The full model achieved 91.68% OA, 90.04% F1-score, and 89.87% AA.
  • Compared to the baseline (Trad-CNN), this represents improvements of 18.47% in OA, 26.25% in F1-score, and 29.49% in AA.
  • Compared to the model using only the improved features and network (VMD-Fusion-CNN-MMATT-CE), the addition of ACBFL (ImFL) further improved performance, confirming its efficacy in handling class imbalance.
• Ablation Studies:
  • Removing any of the three attention blocks (R-CISAM, MS-SFSAM, or SE) resulted in performance degradation, confirming the necessity of the multi-stage, multi-type design.
  • Swapping the order of attention blocks also reduced performance, validating that specific attention types are best suited for specific network depths (local details in shallow layers, multi-scale context in middle layers, channel recalibration in deep layers).
  • The multi-scale approach (using dilation rates 4, 8, 16) outperformed models using single dilation rates.
• Loss Function Analysis: Models with adjustable parameters ($\beta$ and $q$) in ACBFL outperformed those with fixed parameters (0 or 1), demonstrating the value of the proposed flexibility.

5. Significance and Claims

The authors claim that their work provides a comprehensive solution to the challenges of underwater acoustic target recognition by addressing feature representation, network architecture, and data imbalance simultaneously.

• Feature Robustness: The combination of VMD and 3/2-D spectrum provides a more accurate and robust representation of modulation information than traditional methods, effectively handling the non-uniform modulation of ship noise.
• Architectural Innovation: The MMATT framework demonstrates that applying diverse attention mechanisms at different network depths yields superior performance compared to single-type or single-position attention mechanisms.
• Practical Applicability: The ACBFL offers a flexible tool for handling the inherent class imbalance in real-world acoustic datasets without requiring complex resampling techniques, making the model more suitable for practical deployment where data distributions vary.
• Performance: The experimental results on the ShipsEar dataset validate that the proposed framework significantly enhances recognition accuracy, robustness, and the ability to recognize minority classes (tail classes) compared to existing state-of-the-art methods.