To address the challenges of low accuracy and data imbalance in anomaly detection for single-domain data, a novel anomaly detection method based on the fusion of multi-transformation domain features is proposed. The method first performs EMD (Empirical Mode Decomposition) decomposition on the original data to obtain modal domain data and trend domain data. By extracting multi-domain information from the original domain, modal domain, and trend domain, features are extracted from these domains. These features are then fused to enhance the model’s ability to detect anomalies. Through comparative evaluation in credit card fraud detection and pipeline leakage detection experiments, the IR-TCN (Intrinsic Mode Function and Residual-Temporal Convolutional Network) model achieved an AUPRC (Area Under the Precision-Recall Curve) of 0.836 in the credit card fraud case and an accuracy of 0.9882 in the PVC (Polyvinyl Chloride) pipeline leakage detection. The results show that the multi-transformation domain feature fusion strategy significantly improves the AUPRC for anomaly detection in highly imbalanced data samples and enhances accuracy in more balanced data samples. This method is suitable for anomaly detection tasks involving extremely imbalanced or complex data.