Abstract In this paper, an indirect flow pattern recognition method based on time-frequency analysis and neural networks is proposed to investigate the flow patterns in the narrow rectangular channel under heating and non-inertial conditions. Firstly, the adaptive optimal kernel algorithm is utilized to analyze on the typical pressure signal and convert it into time-frequency spectrograms. Then based on the concept of transfer learning strategy, convolutional neural networks are applied as feature extractors to classify flow patterns by the spectrogram images. The proposed method is verified by the visualized flow boiling experiment data. The results show that the adaptive time-frequency algorithm can effectively reflect the characteristics of different flow pattern signals, and several chosen neural network models show high recognition accuracy after training. Among them, VGG-16 network with small convolution kernels and strong transferability has the highest recognition rate. In addition, the network based on data of static conditions remains identifying more than 75% spectrograms of rolling conditions, exhibiting the generalization ability of the method under different flow conditions.