Joint caliber and transient acoustic signals classification with wavelet domain analysis

Transient acoustic signal classification has been widely applied in military reconnaissance, urban security, and field law enforcement due to its strong target relevance, high concealment, and low deployment cost. However, the diversity of waveforms in complex noisy environments and the requirements for real-time performance and lightweight design on edge devices greatly limit its practical application. To address this, this paper proposes an end-to-end transient acoustic signal classification framework based on multi-scale wavelet features and multi-task learning. The framework adopts a hard-shared convolutional backbone network to jointly perform caliber classification and waveform classification tasks, significantly reducing model redundancy and improving classification efficiency. This work introduces a multi-window (1 ms, 3 ms, 6 ms) wavelet analysis method that effectively extracts transient structural features of signals at different time resolutions and builds a multi-task model based on this. In addition, various adversarial perturbations, such as time masking, and gain transition, are designed to evaluate the model’s resilience to real-world acoustic distortions. Experimental results demonstrate that the proposed method maintains excellent classification accuracy and generalization capability under various interference conditions, showing strong engineering practicality and deployment potential.