OKI Develops AI System For Ship Classification Using Underwater Sounds

OKI has developed ‘ship classification AI system technology’ for the automatic classification of ship types through deep learning of underwater sounds.

This technology makes it possible to continuously and automatically acquire ship classification data, even in environments such as busy ports with high ship traffic and at night, when visual identification using cameras is difficult.

Internal verification experiments by OKI have demonstrated that the system can classify ships with 90% or better accuracy, even with only small amounts of learning data extracted from ship sound data.

Since radio waves are significantly attenuated and light waves are widely scattered under the sea, classification methods generally rely on sound waves, which can propagate underwater for distances ranging from several to thousands of kilometers.

In addition, since underwater sound source, whether an animal or ship, has its own unique characteristics, the type can be identified using sounds. One widely-used method involves humans listening to underwater sounds captured by underwater microphones and visualizing their characteristic frequencies to be classified by humans. However, this method may result in different classification results depending on the skill level of the person concerned.

Drawing on its long history of research into underwater acoustic products, OKI has developed systems that analyze the characteristics of sounds received by underwater microphones.

The newly developed ‘ship classification AI system technology’ uses deep learning, an AI technology, to automatically classify ships based on their underwater sounds. This system technology creates deep learning models from sounds recorded by underwater microphones installed in the sea, then automatically classifies ships based on their frequency characteristics.

This allows classification of ships without relying on human skill levels. Since it requires less human labor than before, it can also address the growing labor-saving demand in recent years.

OKI conducted verification using ship sound data it gathered. This time, the system achieved classification results with an accuracy of 90% or better, based on a deep learning model prepared using approximately four hours of ship sound data. Deep learning models typically require large amounts of learning data to accurately identify sound types.

However, the amount of publicly available underwater sound data is limited, and preparing various ship sound information in advance poses significant challenges in terms of time and cost. Thus, OKI applied certain processing techniques to enable classification through the use of only small amounts of learning data.

The challenges mentioned above were solved by data augmentation, which artificially creates variations for actual ship sound data, and semi-supervised learning, which trains the model using ship sound information from partial data. As a result, the system achieved an accuracy rate of 90% or better, even with small amounts of learning data.

According to Yoichi Kato, Senior Executive Officer and Head of the TOKKI Systems Division, “Going forward, we will seek co-creation partners to gather field data and conduct practical verification with a view to commercializing this technology.”

Press Release

Source: Maritime Shipping News