In the oil and gas industry, the acoustic detection of internal leaks in pressurized valves is constrained by environmental noise interference, making it difficult for detection algorithms to discern the temporal characteristics of internal leaks, leading to false alarms for valve internal leaks. To address the drawback of identifying key temporal characteristics of internal leak acoustic signals, we propose the feature mode decomposition (FMD)-Transformer-ResNet internal leak detection method. This method utilizes FMD, fully considering the periodicity and impact characteristics of internal leak acoustic signals, to purify and reconstruct the internal leak acoustic source, overcoming the shortcomings of traditional signal decomposition techniques in dealing with internal leak acoustic signals, such as mode mixing. Subsequently, a Gaussian mixture model (GMM) is used to assign sensitivity weights to the internal leak acoustic source signals. These weights enhance the ability of the Transformer to recognize key temporal characteristics of internal leak acoustic signals. Combined with a residual neural network classifier (ResNet), it achieves high-precision classification and identification of internal leak conditions. Experimental results show that the proposed FMD-Transformer-ResNet valve internal leak detection method achieves an optimal recognition accuracy of up to 97% for internal leaks, surpassing the recognition accuracy of similar methods such as CNN-GRU-Attention and CNN-LSTM-Attention. The proposed method forms a new perspective for valve internal leak condition recognition research, focusing on “signal purification, feature weighting, and accurate prediction,” and holds promising engineering application prospects.