Efficient anomaly detection is crucial to ensure safe operation of Autonomous vehicles (AVs). This study proposes an interpretable method for detecting anomalies in AV data based on denoising variational Transformer (DVT). Considering the limitation of simple structure, poor learning ability of typical autoencoder network, and inability of learning data distribution of traditional self-attention mechanism, current work develops an innovative variational attention mechanism by incorporating Gaussian prior distribution into attention weight. Further, based on proposed variational attention mechanism, an unsupervised anomaly detection network model for DVT is constructed, which can learn distribution of input and reconstruct original input as output to construct residuals for anomaly detection. Additionally, a residual interpreter is designed to calculate contribution of each input feature to anomaly score to explain anomaly detection results. By employing proposed model, an AV data anomaly detection framework is constructed. The proposed model is experimentally verified using simulation platform and real AV platform, and experimental results demonstrate that proposed DVT model outperforms other unsupervised anomaly methods, improving F1-score value by 8%–50% and area under curve index by 1%–20%. Moreover, the average interpretation accuracy of proposed residual interpreter surpasses that of kernel shapley additive explanations by more than 50%, while significantly reducing computational time by over 99%.