Traction transformers are regarded as critical devices for ensuring the reliable operation of traction power supply systems. Existing acoustic fingerprint-based detection methods are driven by fault samples, so their applicability is limited in condition monitoring scenarios of heavy-haul railway traction transformers where fault samples are scarce. To address this issue, an abnormal acoustic fingerprint detection method for traction transformers is proposed by integrating continuous wavelet transform with a spatial-channel fusion autoencoder (SCF-AE) network. Time-frequency feature maps of transformer acoustic fingerprints are obtained by continuous wavelet transform. The feature learning capability of the network is enhanced and its computational complexity is reduced by incorporating spatial-channel fusion convolution into the autoencoder. During training, normal acoustic fingerprints of traction transformers are compressed and reconstructed by the SCF-AE network. During detection, anomalies are identified by determining whether the reconstruction error of an input sample exceeds a predefined threshold. A precision of 88.17% and a recall of 100% are achieved by the proposed SCF-AE method on the dataset, with an average inference time of only 13.3 ms per sample. Superior accuracy and computational efficiency are demonstrated compared with existing methods.