In complex industrial environments, the signals generated by inclusions in molten aluminum when passing through the electric sensing zone are often temporally overlapped with strong industrial noise. This overlap reduces the detection accuracy of systems that rely solely on time-domain parameters for denoising, thereby affecting the quality and performance of metal products. To address this issue, this paper investigates signal analysis and identification for an electric-sensing-zone-based detection system. A signal processing method based on a multi-branch residual attention convolutional neural network (MRACNN) is proposed. Wavelet transform is first employed to extract the time-frequency features of transient signals, so that the differences in frequency-band distribution between inclusion signals and industrial noise can be effectively characterized. The extracted features are then fed into the MRACNN for valid signal identification and screening. Industrial experiments showed that the proposed method improved the detection range and statistical reliability of ESZ-based inclusion measurement, extending the effective detection limit approximately from 65 μm to 20 μm. Comparison of online results with offline PoDFA measurements showed consistent cleanliness grades. These results demonstrate that the proposed software-based signal analysis method can overcome the limitation of hardware measurement resolution under strong industrial noise conditions.