Monitoring industrial reactors using ultrasonic tomography requires solving a strongly nonlinear and ill-posed inverse problem based on limited, noisy boundary measurements. This paper proposes a novel hybrid deep learning architecture that utilizes multi-domain signal processing via the Hilbert transform, Fourier transform, and S-transform (HFS) approximation. Instead of processing raw time-of-flight sequences directly, they were converted into three-channel descriptors, and the training dataset was explicitly augmented with 5% Gaussian noise to make the model resilient to real-world measurement disturbances. Evaluation proved the significant superiority of the proposed approach over the baseline model analyzing raw signals. The HFS architecture reduced the Mean Squared Error from 0.241 to 0.128 and substantially increased the Pearson Correlation Coefficient from 0.624 to 0.863. This confirms that the proposed approach yields higher spatial localization fidelity, better preserves the geometry of submerged objects, and effectively suppresses background artifacts. The proposed framework provides a reliable and stable data-driven tomographic imaging tool for multiphase processes.