<p>Track dynamic inspection data, a multi-sensor time series collected regularly by high-speed rail inspection vehicles, is critical for ensuring track safety through anomaly detection. However, these data are characterized by noise, anomalies, complex inter-attribute correlations, and imbalanced sample categories, which pose significant challenges for effective anomaly detection. In this paper, a novel unsupervised anomaly detection method for track dynamic inspection data termed IFA-TransD is proposed, which incorporates attention fusion and an improved transformer-based detection structure. Specifically, the attention fusion module integrates channel attention and temporal attention in parallel to enhance the representation of temporal and channel features within the sequences, which effectively captures the complex relationships between dimensions and highlights the correlation discrepancies associated with anomalies. The proposed improved Transformer-based detection module TransDetect that strengthens the Transformer’s capability to detect anomalies in multi-dimensional dynamic time-series data. Experimental results on both public anomaly detection benchmarks and track dynamic detection datasets demonstrate the effectiveness of the proposed method. Compared to multiple baselines, it achieves competitive performance on the public dataset and outperforms others on the practical track anomaly detection task, achieving a 98% recall and a 90% F1 score.</p>