Scholarlink

Purpose <p>Beam tuning at the high-intensity heavy-ion accelerator facility (HIAF) remains heavily reliant on operator experience. This study proposes a multi-task Transformer ensemble method to predict detector operation sequences and provide real-time recommendations for the next operation during beam tuning, with the aim of supporting future automated operation and maintenance.</p> Methods <p>Bit-level state-change detection of HIAF machine protection system (MPS) programmable logic controller (PLC) data yielded 2108 operation events spanning ten patterns, which were validated by unsupervised clustering. A multi-task Transformer with 247-dimensional multi-level features jointly predicts the next device, device type, location region, operation pattern, and anomaly status. Rare-class merging, embedding-layer Mixup, and a weighted Transformer ensemble were employed to mitigate data sparsity, class imbalance, and overfitting, and sequence-wise grouped cross-validation was used to prevent sliding-window leakage.</p> Results <p>Using sequence-wise grouped fivefold cross-validation, in which all sliding-window samples generated from the same original operation sequence were assigned to the same fold, the deployed seven-model Mixup ensemble achieved Top-5 accuracy of 82.3% ± 1.9%, Top-1 accuracy of 53.7% ± 2.4%, and anomaly-detection AUC of 87.8% ± 6.4%.</p> Conclusion <p>The proposed method has been deployed as a stateful incremental prediction API, providing a practical basis for real-time HIAF beam-tuning assistance and future automated operation and maintenance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A multi-task transformer-based prediction model for HIAF detector operation sequences

  • 期刊论文

  • 发表日期：

    2026年7月18日

查看全文
  • Yuan Ye,
  • Zhang Jianchuan,
  • Wei Yuan,
  • Du Xiying,
  • Li Zhixue,
  • Zhu Guangyu,
  • Xie Hongming,
  • Wu Junxia

摘要

Purpose

Beam tuning at the high-intensity heavy-ion accelerator facility (HIAF) remains heavily reliant on operator experience. This study proposes a multi-task Transformer ensemble method to predict detector operation sequences and provide real-time recommendations for the next operation during beam tuning, with the aim of supporting future automated operation and maintenance.

Methods

Bit-level state-change detection of HIAF machine protection system (MPS) programmable logic controller (PLC) data yielded 2108 operation events spanning ten patterns, which were validated by unsupervised clustering. A multi-task Transformer with 247-dimensional multi-level features jointly predicts the next device, device type, location region, operation pattern, and anomaly status. Rare-class merging, embedding-layer Mixup, and a weighted Transformer ensemble were employed to mitigate data sparsity, class imbalance, and overfitting, and sequence-wise grouped cross-validation was used to prevent sliding-window leakage.

Results

Using sequence-wise grouped fivefold cross-validation, in which all sliding-window samples generated from the same original operation sequence were assigned to the same fold, the deployed seven-model Mixup ensemble achieved Top-5 accuracy of 82.3% ± 1.9%, Top-1 accuracy of 53.7% ± 2.4%, and anomaly-detection AUC of 87.8% ± 6.4%.

Conclusion

The proposed method has been deployed as a stateful incremental prediction API, providing a practical basis for real-time HIAF beam-tuning assistance and future automated operation and maintenance.