Abstract: On-load tap changer (OLTC) plays a crucial role in power systems, and accurately identifying its tap positions is important for the stable operation of the power grid, intelligent maintenance, and fault diagnosis. This paper deeply analyzes the vibration signals generated by OLTC under normal operating conditions during different tap position changes and proposes an OLTC tap position identification method that combines particle swarm optimization (PSO), deeply separable convolutional neural network (DSCNN), long short-term memory (LSTM), and self-attention (SA) mechanisms. The vibration signals from the OLTC top cover are first convolved point-by-point to capture cross-channel characteristics, and local features are extracted using DSCNN. LSTM is then used to extract the time-series features of the vibration signals, with the adaptive weighting capability of SA to select key features for classification. Finally, PSO is employed to find the optimal model parameters. The results demonstrate that the proposed method provides better noise resistance and higher tap position identification accuracy compared to mainstream classification algorithms.