TL;DR
This paper introduces a novel multimodal fusion approach for accurately detecting and counting multiple underwater sound sources using a single Acoustic Vector Sensor (AVS). By decomposing the AVS signal into multiple modes and then applying density-based clustering, the method significantly improves source discrimination capabilities, especially in challenging underdetermined scenarios.
What Problem Does This Paper Address?
Estimating the direction of arrival (DOA) and the number of sound sources is crucial for underwater acoustic systems. Traditional methods using acoustic vector sensors (AVS) often struggle when multiple sources are present, as the sensor's ability to distinguish them diminishes with increasing source count. This is particularly problematic in "underdetermined" scenarios where the number of sources exceeds the number of independent measurements. Existing techniques like the conjugate cross-spectrum method can provide DOA histograms, but accurately counting sources from these histograms becomes difficult as the signal-to-noise ratio decreases and source overlap increases. The paper aims to overcome this limitation by enhancing the discriminative power of a single AVS in multi-source environments.
Methods
The proposed method employs a three-stage multimodal fusion strategy:
- Intrinsic Time-Scale Decomposition (ITD): The raw signal from the AVS (sound pressure and two orthogonal vibration velocity components) is decomposed into multiple "modes" based on frequency bands. The core idea here is that by separating the signal into different frequency components, the number of dominant sources within each mode is reduced, thereby increasing the "W-disjoint orthogonality" (WDO) and improving the ability to distinguish sources within that mode.
- Density Peak Clustering (DPC) with Gap-Based Method: For each decomposed mode, the DOA of sound sources is estimated using the conjugate cross-spectrum of the sound pressure and vibration velocity signals. Subsequently, DPC, a clustering algorithm that identifies cluster centers based on local density and minimum distance, is applied to these DOA estimates. This step helps to identify potential source directions and count sources within each individual mode.
- Multimodal Fusion using DBSCAN: Since the DOAs obtained from individual modes might be slightly different or contain irrelevant estimates, a final fusion step is necessary. The Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm is used to cluster the DOA estimates from all modes. DBSCAN is effective at identifying dense regions of points and can handle noise, allowing it to merge DOAs belonging to the same source across different modes and provide a final, robust estimate of the number of sources and their DOAs. The paper also mentions a modified DBSCAN that adaptively determines its parameters (EPS and M) to improve robustness.
Datasets
The experiments were conducted using real-world data collected in Fuxian Lake. The sound sources were generated by four small sailing boats moving in uniform motion at a distance of 1-2 km from the AVS. The initial directions of the boats were set to be within four degrees of 40°, 140°, 210°, and 320° from the AVS. This experimental setup serves as the primary dataset for evaluating the proposed algorithm.
Key Insights and Ideas
- Decomposition Enhances Discrimination: The fundamental insight is that decomposing a complex multi-source signal into multiple frequency-based modes (using ITD) effectively reduces the complexity within each mode, making it easier for subsequent algorithms to distinguish individual sources.
- Complementary Strengths of Clustering Algorithms: The paper strategically combines DPC and DBSCAN. DPC is used for initial source identification within each mode, while DBSCAN provides a robust global fusion mechanism to consolidate results across modes and handle potential outliers or interference.
- Bridging Gaps in Source Traces: A significant finding is that when a source's trace (its estimated DOA over time) is interrupted in one mode due to insufficient dominant time-frequency points, it can often be recovered and made continuous by observing it in other modes. This highlights the benefit of multimodal fusion for creating more complete source tracking.
- Adaptive DBSCAN for Robustness: The introduction of an adaptive DBSCAN parameter selection method addresses a known sensitivity of DBSCAN, making the overall system more reliable in real-world scenarios.
Key Results
The paper demonstrates that the multimodal fusion algorithm significantly improves source detection and counting accuracy compared to methods without fusion.
- Continuous Source Traces: Figures 5a-c show that the multimodal fusion algorithm produces more continuous traces for the four boats compared to the non-fusion method (Figure 5d), especially in areas where interference from other boats would otherwise cause discontinuities.
- Improved Source Counting Accuracy: Figure 6 shows quantitative distributions of source numbers. When using four modes, the algorithm achieved a source counting accuracy of 46% of the time, a notable improvement over the 32% accuracy achieved without multimodal fusion. The accuracy further increased with the number of modes used.
- Accurate Source Number Identification: The quantitative distribution of source numbers showed a clear peak corresponding to the true number of sources (four in this experiment) when the multimodal fusion algorithm was applied, indicating better estimation of the source count.
Limitations and Future Work
- Parameter Sensitivity: While the paper introduces an adaptive DBSCAN, it acknowledges that DBSCAN is generally sensitive to parameter selection (EPS, M). The adaptive method aims to mitigate this, but further research might explore even more robust parameter-free or self-adaptive techniques.
- Computational Complexity: The ITD decomposition and subsequent clustering steps can be computationally intensive. While the paper focuses on accuracy, future work could explore optimizations for real-time applications.
- Performance with More Sources: The paper demonstrates success with four sources. Investigating the algorithm's performance with a larger number of closely spaced or interfering sources would be a valuable next step.
- Real-World Environmental Factors: The experiments were conducted in a lake. Evaluating the algorithm's performance in more complex underwater environments with varying noise conditions, reverberation, and multipath propagation would be beneficial.
Summary updated