BioNet-A: Ultrasonic echo representation network for target discrimination using active SONAR

Ultrasonic echoes provide critical cues for object perception, yet their millisecond duration and frequency-specific structure violate assumptions of conventional spectrogram-based convolutional models. Existing biomimetic front-end systems, including auditory spectrograms, cortical wavelets, and the biomimetic BioNet, either under-utilize model capacity on brief echoes or impose frequency shift-invariance suited to vocalizations but detrimental for echo discrimination. This study introduces BioNet5-A, a biomimetic encoder optimized for ultrasonic echoes. BioNet5-A is derived from an autoencoder pretrained on bat vocalizations and incorporates three architectural innovations: (1) spectrotemporal attention to concentrate capacity on the ∼2 ms echo segment and relax frequency-axis invariance; (2) multi-sized convolution/transposed-convolution modules that capture echo structure across multiple scales; and (3) a symmetric, weight-tied encoder–decoder to stabilize training and regularize the biomimetic code. Using a controlled ultrasonic dataset spanning multiple objects and noise conditions, BioNet5-A consistently outperforms auditory spectrogram, cortical wavelet, and the conventional BioNet, and shows improved clustering, discrimination, and robustness. Additionally, representations by the model remain compact and interpretable, aligning with bat midbrain physiology. These results position BioNet5-A as a practical front end for biosonar-inspired sensing and ultrasonic applications.