TL;DR
This paper introduces an innovative adaptive 2D convolution kernel construction method for underwater acoustic detection. By leveraging unsupervised learning (an improved k-means algorithm) to extract visual patterns from historical bearing time records (BTRs), the method effectively reduces noise and narrows the main lobe width, significantly improving the detection of weak targets in challenging underwater environments.
What Problem Does This Paper Address?
Underwater acoustic detection, particularly for estimating the Direction of Arrival (DOA) of targets, faces significant challenges. The complex ocean environment is rife with noise, and limitations in acoustic array design can lead to bearing time records (BTRs) with low signal-to-noise ratios (SNRs) and wide main lobes. These issues make it difficult to accurately detect and track weak targets, especially those exhibiting high-speed maneuvers. Traditional methods like Conventional Beamforming (CBF) often struggle to overcome these limitations, and while image processing techniques have been applied, they can lead to blurred details or require prior knowledge of noise levels, which is often unavailable in real-world underwater scenarios. This paper aims to address the need for a robust and adaptive method that can enhance target detection and resolution in noisy underwater conditions without requiring prior assumptions about the environment.
Methods
The core of this paper is the proposed Adaptive 2D Convolution Kernel Construction method. The process involves several key steps:
- BTR Generation: Initially, conventional array beamforming (specifically for Uniform Linear Arrays - ULAs) is used to generate BTRs from passive sonar data, capturing both time and azimuthal information.
- Peak Extraction and Element Size Determination: An open peak extraction method is applied to the BTR to identify potential targets. Crucially, the size of the "elements" (visual units) used for kernel construction is determined based on the main-to-sidelobe distance of the CBF directivity function. This ensures that the kernel focuses on relevant signal features and avoids being misled by noise.
- Adaptive Kernel Construction (KAdaptive): An improved k-means clustering algorithm is employed. This algorithm takes the extracted visual patterns (main lobe, sidelobes, and noise) from historical BTR data and clusters them. The algorithm automatically determines the optimal number of clusters using a novel SSE-based strategy that analyzes multi-order derivatives of the Sum of Squared Errors (SSE). The most representative cluster, corresponding to the main lobe pattern, is then used to construct the adaptive convolution kernel, denoted as KAdaptive.
- 2D Convolution: The generated KAdaptive kernel is then used to perform a 2D convolution operation on the BTR. To handle boundary effects, mirror padding is applied to the BTR before convolution. The kernel is also rotated by 180 degrees to adhere to the mathematical definition of convolution.
- Output: The final output is a processed BTR that has undergone adaptive convolution, aiming for enhanced noise suppression and improved target resolution.
This approach is novel in its use of unsupervised learning to dynamically extract and adapt convolution kernels directly from the target data, rather than relying on fixed, pre-defined kernels.
Datasets
The paper utilizes two types of data for evaluation:
- Simulated Data: Multiple targets were simulated under various low SNR conditions (from 0 dB to -20 dB). The simulation parameters included a specific array configuration (256 elements, 1.5m spacing), signal frequency (200 Hz), and sampling frequency (2000 Hz).
- Real Sea-Trial Data: Data was collected from shallow-water trials (depths 87.1-101m) using a towed horizontal line array with 256 elements spaced 1.5m apart, operating between 20 and 380 Hz. This dataset contained five targets over a 600-second period.
The paper does not introduce new benchmark datasets but rather uses these datasets to compare its proposed method against existing techniques.
Key Insights and Ideas
- Adaptive Kernels for Dynamic Environments: The most significant insight is the power of using adaptive convolution kernels. Unlike fixed kernels that assume a static environment, this method learns from the historical data to create kernels that are tailored to the specific characteristics of the underwater acoustic signals and noise, making it more robust to variations.
- Unsupervised Learning for Feature Extraction: The application of an improved k-means clustering algorithm for extracting "visual patterns" of the main lobe from BTRs is a clever way to leverage unsupervised learning. This bypasses the need for labeled data or manual feature engineering, which is often difficult in this domain.
- Automatic Cluster Number Determination: The novel SSE-derivative-based method for automatically determining the optimal number of clusters is a crucial advancement. It addresses the limitations of traditional elbow methods, leading to more reliable kernel generation.
- Bridging Image Processing and Signal Processing: The paper effectively bridges concepts from computer vision (convolution kernels, clustering) with signal processing (BTRs, beamforming) to solve a challenging underwater acoustics problem.
Key Results
The proposed Adaptive 2D Conv method demonstrated superior performance across various metrics:
- SNR Enhancement: In simulations with multiple targets, the Adaptive 2D Conv method achieved significantly higher SNRs compared to CBF, Theory 2D Conv, and 1D dCv, especially at lower SNRs. For instance, at 0 dB SNR, it achieved SNRs of 26 dB, 22.5 dB, and 25.4 dB for three targets, vastly outperforming other methods.
- Noise Suppression and Resolution: The method effectively suppressed background noise and narrowed the main lobe width. This was evident in both simulated and real sea-trial data, leading to clearer BTRs with better target distinguishability.
- Weak Target Detection: At extremely low SNRs (-20 dB), where other methods failed to distinguish targets from noise, the Adaptive 2D Conv method successfully enhanced weak targets, allowing for clear separation.
- Computational Cost: While having a slightly higher computational cost than simpler filters like Mean and Gaussian Conv, the Adaptive 2D Conv method achieved a per-frame processing time of approximately 0.31 seconds, which is considered sufficient for near real-time applications.
Limitations and Future Work
- Target Type Applicability: The paper notes that the method is primarily suitable for targets that continuously radiate noise (like ships) because the adaptive kernel construction relies on time information. It is not applicable to transient noise sources like acoustic bombs.
- Computational Cost: Although deemed acceptable for near real-time applications, the computational cost is higher than simpler filters.
- Future Work: The authors suggest that further research could explore extending this adaptive approach to other types of underwater acoustic detection and tracking scenarios. Investigating its performance with different array configurations and in even more complex acoustic environments could also be valuable.
Summary updated