Deletion variation is an important structural variation associated with various diseases such as cancer, Alzheimer's disease, and autism. Due to the complexity of deletion variation and alignment information, existing methods often misclassify a significant number of non-deletion variations as deletion variations, leading to high false-positive rates in detection results and posing challenges for downstream analysis. To address this issue, this paper proposes a Transformer-based deletion variation detection method for third-generation sequencing data, called TDD (Transformer deletion detection). Candidate variation sites are first extracted by analyzing the CIGAR field in the BAM files and the spacing between adjacent reads. Then, the variation intervals are divided into continuous sub-intervals, and feature matrices are constructed for each sub-interval. Subsequently, the feature matrices are encoded using the Encoder module in Transformer and passed to the classification layers to determine the presence of variation. Finally, breakpoint estimation and merging of variation sub-intervals are performed to obtain the final set of deletion variations. Experimental results show that the proposed method achieves higher F1 scores than mainstream tools, reduces false positives in detection results, and thereby improves the detection of deletions.