To address the issue of low testing accuracy caused by noise spectrum aliasing in ultrasonic testing, a data-model-driven method for high-quality extraction of ultrasonic detection signals is proposed. This approach combines empirical mode decomposition (EMD) with a component clustering index to pre-process the ultrasonic signal, reducing the impact of noise on signal extraction. Using the Gaussian echo model, time-frequency transformation, spectral Gaussian fitting, and the artificial bee swarm algorithm, the model parameters of the pre-processed signals are accurately estimated. The Gaussian echo model is then reconstructed to achieve high-quality signal extraction. Simulation results demonstrate that this method can extract ultrasonic signals with an SNR as low as 4. 56 dB, improving the mean SNR of the extracted signals to 28. 71 dB—significantly outperforming common methods such as EMD ( SNR = 9. 82 dB) and variational mode decomposition ( SNR = 11. 07 dB). Furthermore, ultrasonic testing experiments confirm the method′s ability to extract high-quality signals even under noise spectrum aliasing interference.