Active sonar automatic target recognition using explainable artificial intelligence

SOund NAvigation Ranging, or sonar, target recognition is the process of transmitting and receiving sound scattered from a target and performing processing on the signal to identify some parameter about a target (e.g., class, shape, distance, etc.,). An active sonar system emits sound waves to a target of interest and records the backscattered echo. The transmit signal encounters distortion due to environmental parameters, noise, and additional obstructions within the propagation path. The backscattered response (echo) is dependent on unknown target parameters such as size, geometry, and material properties. The culmination of these effects result in a received signal that contain entangled target-specific information. The target-specific information can be used for classification.

Two main approaches have been employed for disentanglement and classification of active sonar signals. The first approach utilizes statistical or signal processing techniques for classification. The second approach employs machine learning techniques that may achieve expert-level accuracy but may not be interpretable. The latter approach relies on large datasets that are costly and challenging to obtain. Feature extraction is a technique employed to disentangle the target-specific information prior to classification. Information of acoustic physics can be used to make informed decisions for feature extraction algorithms facilitating feature interpretation and increased classification accuracy.

Novel feature extraction techniques informed by the physical domain and validated across machine learning classifiers are presented. The classification results and trained classifiers are examined to explain classifier choices, known as explainable artificial intelligence. Specifically, novel feature representations are presented for classification of three experimentally collected datasets and one simulated dataset. The innovation presented is the unified computational framework and iterative approach to employing physically-informed feature extraction techniques, validation through machine learning classifiers, and feature interpretation. Methods may be used to create interpretable feature representations in classification pipelines for a diverse portfolio of active sonar targets.