Sodium-cooled fast reactors (SFRs), expected to improve the efficiency of nuclear fuel utilization, use liquid sodium as a reactor coolant, which makes it possible to breed nuclear fuels while generating electricity. To increase the safety of SFRs, it is important to develop a method to detect the generation of bubble jet flow in the early stage and identify the states of it when a heat transfer tube is damaged in a steam generator. For this issue, we propose a novel state sensing method with convolutional neural networks (CNNs) and time-frequency representations (TFRs). This study consists of three phases. First, using water and air as simulant fluids to perform the proof of concept, pipe flow sound and bubble jet flow sound are acquired, each of which simulates normal and anomaly sound. Second, three TFRs are extracted from raw signals based on short-time Fourier transform (STFT), continuous wavelet transform (CWT), and synchrosqueezed wavelet transform (SWT). Third, typical CNNs including AlexNet, VGG16, and ResNet18 are introduced for the identification of pipe flow sound and three types of bubble jet flow sound. Also, t-distributed stochastic neighbor embedding (t-SNE) is applied to visualize the learning process of CNNs. As a result, the model combining ResNet18 and STFT reaches the highest accuracy of 99.20 ± 0.31% and correctly identifies 1984 out of 2000 test data. These results demonstrate that our proposed method based on the acoustic identification of deep learning has great potential to sense the states of bubble jet flow in actual SFRs.