TL;DR
This paper introduces a novel deep learning model that mimics the human auditory system to classify ship types based on their radiated noise. By incorporating auditory-like mechanisms, the model achieves state-of-the-art accuracy, demonstrating the potential of bio-inspired AI for underwater acoustic analysis.
What Problem Does This Paper Address?
Monitoring underwater noise is crucial for understanding the impact of human activities, particularly shipping, on marine environments. Accurately identifying ship types from their radiated noise is challenging due to complex noise generation mechanisms and the unpredictable nature of underwater sound propagation. Traditional machine learning methods often rely on manually engineered features, which can be suboptimal and fail to capture the full complexity of the acoustic signals. This paper aims to overcome these limitations by developing a more robust and effective ship type classification system using a deep learning approach inspired by the human auditory system.
Methods
The core of this work is a novel convolutional neural network (CNN) architecture designed to emulate the human auditory pathway. The model is structured into two main components:
- Cochlea Model: This initial stage processes the raw acoustic signal. It employs a time convolutional layer with dilated auditory filters (gammatone filters) to decompose the signal, mimicking the frequency separation performed by the cochlea's basilar membrane. A time-frequency conversion layer then transforms these decomposed signals into a representation akin to a spectro-temporal excitation pattern, similar to how hair cells in the cochlea process vibrations.
- Auditory Center Model: This deeper part of the network further refines the features. It includes a supervised auditory feature recalibration mechanism to selectively emphasize important auditory features. The subsequent multistage auditory center uses a deep architecture with inception and residual connections, inspired by advancements in computer vision, to learn complex spectro-temporal patterns. This hierarchical structure allows the network to process information at different levels of abstraction, much like the auditory cortex. A key innovation is the use of global average pooling at the final stage, which significantly reduces the number of parameters compared to traditional fully connected layers, mitigating overfitting.
Datasets
The experiments were conducted using acoustic data acquired by the Ocean Networks Canada observatory. The dataset comprises underwater acoustic data from ships within a 2 km radius, with recordings lasting 5 to 20 minutes. For classification, the data was segmented into 3-second samples, further divided into 1-second frames. The dataset was resampled to a 16 kHz sampling frequency. The classification experiments were performed on ocean background noise and four specific ship types: Cargo, Passenger ship, Tanker, and Tug. Approximately 140,000 training samples and 82,000 validation samples were collected over 29 months.
Key Insights and Ideas
- Bio-inspired AI for Acoustics: The paper strongly advocates for leveraging the human auditory system's processing strategies as a blueprint for designing effective underwater acoustic analysis models.
- Dilated Convolutions for Auditory Filters: Using dilated convolutions with gammatone filters in the initial stage allows for a wider receptive field without increasing parameters, effectively capturing broader frequency components.
- Deep Hierarchical Processing: The multistage auditory center, inspired by both auditory pathways and modern deep learning architectures (Inception-ResNet), demonstrates that hierarchical feature extraction is crucial for complex acoustic pattern recognition.
- Parameter Reduction via Global Average Pooling: Replacing fully connected layers with global average pooling is a significant architectural improvement that combats overfitting and simplifies the model.
- Robustness to Operative Conditions: The model shows resilience to variations in ship speed, course, and distance, suggesting its practical applicability in real-world, dynamic ocean environments.
Key Results
The proposed model achieved a classification accuracy of 87.2% on four ship types and ocean background noise. When evaluated on recordings using majority voting, the accuracy increased to 94.75%. This performance significantly outperforms previous methods, including waveform features, MFCCs, wavelet features, and an existing auditory-inspired CNN, which achieved accuracies ranging from 68.2% to 83.2%. The study also found that increasing the distance between the ship and the hydrophone negatively impacts recall rates for most ship types, though the model remains robust.
Limitations and Future Work
The authors acknowledge that the recall rate for Cargo, Passenger ships, and Tanker ships decreases with increasing distance from the hydrophone. While the model shows robustness to operative conditions, further investigation into the influence of SOG and COG could be beneficial. The paper suggests that the proposed method could facilitate the development of "smart hydrophones" capable of not only measuring but also alerting to specific underwater acoustic events, making ocean listening easier for researchers. Future work could explore extending this approach to a wider variety of ship types and more complex acoustic environments.
Summary updated