Competitive Deep-Belief Networks for Underwater Acoustic Target Recognition

Honghui Yang, Sheng Shen, Xiaohui Yao, Meiping Sheng, Chen Wang2018 · DOI

TL;DR

This paper introduces a novel "Competitive Deep-Belief Network" (CDBN) for underwater acoustic target recognition, a challenging problem due to limited labeled data. By integrating competitive learning into a deep-belief network, the CDBN effectively learns discriminative features from both labeled and unlabeled data, significantly improving classification accuracy and reducing reliance on hand-engineered features.

What Problem Does This Paper Address?

Identifying underwater targets, such as ships, using passive sonar systems is crucial for maritime safety and surveillance. These systems rely on analyzing the radiated noise from targets. However, this task is inherently difficult due to several factors:

  • Complex Underwater Acoustics: The underwater environment distorts acoustic signals, making processing challenging.
  • Small Sample Size Problem: Obtaining large, labeled datasets of underwater acoustic targets is expensive and time-consuming. This is a major bottleneck for traditional supervised machine learning methods.
  • Subjectivity of Manual Analysis: Historically, human sonarmen have been relied upon, but their performance can be inconsistent due to fatigue, environmental conditions, and subjective interpretation.
  • Limitations of Traditional Methods: Existing automatic recognition systems often depend on hand-engineered features (e.g., waveform, wavelet, auditory features) and shallow classifiers. This requires significant domain expertise and engineering effort, and their performance can be suboptimal in noisy or complex environments.

The paper aims to overcome these limitations by developing a more robust and accurate automatic recognition system that can effectively learn from limited labeled data and abundant unlabeled data.

Methods

The proposed method, Competitive Deep-Belief Networks (CDBNs), builds upon Deep-Belief Networks (DBNs) by incorporating a competitive learning mechanism. The training procedure involves four key stages:

  1. Unsupervised Pretraining of RBM: A standard Restricted Boltzmann Machine (RBM), specifically a Gaussian-Bernoulli RBM (GB-RBM) to handle real-valued noise, is pretrained using a large amount of unlabeled ship-radiated noise data. This initializes the network's parameters and allows it to learn basic acoustic signal characteristics.
  2. Hidden Unit Grouping: After RBM pretraining, the hidden units are grouped based on their activation scores for different target categories. This step aims to associate hidden units with specific types of acoustic signals.
  3. Competitive Learning for Feature Enhancement: A competitive layer is introduced, connecting the grouped hidden units. This layer employs a mechanism of "intragroup enhancement and intergroup inhibition" (lateral connections with negative weights). This competitive process encourages hidden units within a group to reinforce each other while suppressing units from other groups. The goal is to enhance the discriminative power of the learned features by promoting a self-clustering property. This competitive RBM is termed a Competitive Restricted Boltzmann Machine (CRBM).
  4. Deep Architecture and Fine-tuning: The trained CRBMs are stacked to form a deep CDBN architecture. This deep model is then fine-tuned using a small amount of labeled data with a supervised learning approach (e.g., backpropagation) to optimize its classification performance.

The key novelty lies in the integration of competitive learning within the DBN framework to improve feature discriminability, addressing the small sample size problem by leveraging unlabeled data.

Datasets

The experiments were conducted using sea trial data recorded in the South China Sea.

  • Unlabeled Data: 20,000 samples of ocean ambient noise and radiated noise from vehicles without labels.
  • Labeled Data: 4,000 samples of radiated noise from two classes of vehicles: small boat and autonomous underwater vehicle.
  • 2,800 labeled samples were used for training.
  • The remaining 1,200 labeled samples were used for testing.

The paper does not introduce a new benchmark dataset but uses a real-world dataset with a specific class imbalance and limited labeled data, which is characteristic of the problem domain.

Key Insights and Ideas

  • Competitive Learning for Discriminative Features: The core insight is that by making hidden units "compete" with each other based on their relevance to different categories, the network can learn features that are not only representative but also highly discriminative, effectively separating different target classes.
  • Leveraging Unlabeled Data is Crucial: The success of the CDBN hinges on its ability to learn meaningful representations from abundant unlabeled data before fine-tuning with limited labeled data. This is a powerful strategy for tackling small-sample-size problems in domains where labeled data is scarce.
  • Deep Features Outperform Hand-Engineered Features: The study demonstrates that features learned automatically by the CDBN (especially from deeper layers) exhibit better relevance to labels and lead to higher classification accuracy compared to traditional, manually designed features.
  • Self-Clustering Property: The competitive learning mechanism imbues the learned features with a self-clustering property, meaning that similar data points naturally group together in the feature space, which is beneficial for classification.

Key Results

  • Classification Accuracy: The proposed CDBN method achieved a classification accuracy of 90.89% using only nine features. This is an improvement of 8.95% over the best-performing traditional methods.
  • Feature Efficiency: The CDBN achieved higher accuracy with significantly fewer features compared to traditional methods. For instance, the best CDBN features achieved 90.89% accuracy with 9 features, while traditional features required 36 features to reach 83.42% accuracy.
  • Feature Significance: Features from the deeper layers of the CDBN (Layer2) showed the highest significance index (0.554), outperforming other features, including those from Layer1 (0.392) and all traditional feature sets.
  • ROC Analysis: ROC curves demonstrated that the CDBN features (especially Layer2) achieved a higher Area Under the Curve (AUC) compared to traditional features. Notably, the SVM trained on CDBN features achieved a 100% true-positive rate at a 27.5% false-positive rate, which is highly desirable for ship-radiated noise recognition.
  • Spectrum Reconstruction: The CDBN was shown to effectively reduce random noise and enhance the line-spectrum characteristics of ship-radiated noise, indicating its ability to capture essential target signatures.

Limitations and Future Work

  • Computational Cost: While not explicitly detailed as a limitation, deep learning models, especially those with multiple layers and competitive mechanisms, can be computationally intensive for training.
  • Hyperparameter Tuning: The performance of deep learning models is often sensitive to hyperparameter choices (e.g., number of layers, units per layer, learning rates, competitive learning parameters). The paper focuses on demonstrating the effectiveness of the proposed architecture rather than an exhaustive hyperparameter search.
  • Generalizability to More Classes: The current study focused on two classes of targets. Future work could explore the CDBN's performance with a larger and more diverse set of underwater acoustic targets.
  • Real-time Application: Adapting the CDBN for real-time underwater acoustic target recognition would require further optimization for speed and efficiency.
  • Exploring Other Competitive Mechanisms: Investigating alternative competitive learning strategies or integrating them with other deep learning architectures could yield further improvements.

Summary updated

BibTeX
@misc{yang2018competitive,
  author = {Honghui Yang and Sheng Shen and Xiaohui Yao and Meiping Sheng and Chen Wang},
  title = {{Competitive Deep-Belief Networks for Underwater Acoustic Target Recognition}},
  year = {2018},
  doi = {10.3390/s18040952},
}
Figure