The practical significance of time series anomaly detection in industrial equipment monitoring is substantial, financial system security, and space science exploration. Current approaches can be roughly divided into two categories: point-based approaches, which struggle to capture complex temporal dynamics, and correlation-based approaches, which can model relationships between time points but lack sufficient focus on anomalous regions. Recently, Transformer models have demonstrated potential in time series analysis by unifying global representation and long-range dependency modeling. In particular, Anomaly Transformer quantifies association bias to differentiate between normal and anomalous points, attaining cutting-edge results on several benchmark datasets. Based on the existing model, this paper proposes an enhanced Anomaly Transformer. The proposed improvements include designing learnable positional embeddings to improve the way temporal position data is represented, improving the token embedding module by incorporating multi-layer convolutions and residual structures to enhance feature extraction capabilities, and integrating these optimizations into the multi-head attention mechanism. Additionally, an anomaly weight parameter is introduced into the anomaly attention module, enabling the model to calculate previous links by concentrating more on anomalous locations. Results from experiments show that the suggested methodology performs noticeably better than current state-of-the-art techniques on a variety of datasets, highlighting its superiority and robustness.