Engineering Innovation in
Digital Defense, Aerospace & Homeland Security
We created a decision-support platform based on AI algorithms to process audio tracks acquired in an underwater setting and to perform object classification.
Approach & Solution
Approach
Acoustic signals detected in the marine environment are characterised by the presence of various sources of noise and distortion, which are superimposed on the 'useful' signal for the identification of the emitting object. This is a complex factor in the classification process, especially if little information is available or only the acoustic signal detected by the sonar sensor.
The most common causes of noise are due to the noise of the marine environment itself, which is amplified by specific characteristics such as: the morphological conformation and type of seabed, the presence of interference sources, especially in the case of high traffic density, and the noise of the observation platform itself (self-induced noise).
The most common causes of noise are due to the noise of the marine environment itself, which is amplified by specific characteristics such as: the morphological conformation and type of seabed, the presence of interference sources, especially in the case of high traffic density, and the noise of the observation platform itself (self-induced noise).
Solution
The solution implemented consists of two functional components:
The two components operate independently, so that the performance of one is not compromised by the performance of the other; periodic synchronisation of the respective databases is foreseen.
- a "peripheral" component, which is installed on board the submarine unit and receives the acoustic traces acquired with the on-board sonars, and after an initial pre-processing phase performs the classification of the traces using specific machine learning algorithms
- a "central" component installed at a validation and enhancement station, on which users validate the results obtained with the AI algorithms and perform the training of the machine learning networks employed.
The two components operate independently, so that the performance of one is not compromised by the performance of the other; periodic synchronisation of the respective databases is foreseen.
Results
Process underwater audio tracks
Establish type of detected target
Multilevel target classification (family and class)
Degree of recognition
>70%
Training for performance improvement
Project value
Process performance
Innovation
Enabling Technologies
AI Advanced Analytics
Project Team
Defense, Aerospace & Homeland Security