基于Transformer语义迁移的定子热故障诊断  

Transformer-based semantic transfer for stator thermal fault diagnosis

作　　者：姚驰宇 桂晶 李泼 王伟 陈聪 YAO Chiyu;GUI Jing;LI Po;WANG Wei;CHEN Cong(Anhui Huainan Pingwei Power Generation Co.,Ltd,Huainan,Anhui 232089;China Power Hua Chuang Electricity Technology Research Co.,Ltd,Shanghai 200086;China Power Hua Chuang(Suzhou)Electricity Technology Research Co.,Ltd,Suzhou,Jiangsu 215123)

机构地区：[1]淮南平圩发电有限责任公司,安徽淮南232089 [2]中电华创电力技术研究有限公司,上海200086 [3]中电华创(苏州)电力技术研究有限公司,江苏苏州215123

出　　处：《电气技术》2025年第3期59-64,84,共7页Electrical Engineering

摘　　要：汽轮发电机定子冷水系统须保持良好运行状态,以保障发电机的可靠性和安全性。通常采用停机检修或温差阈值的方法进行热故障检测,但无法在发电机运行状态下及时有效地检出故障。为了更精确地发现定子热故障,本文提出基于Transformer架构的温度预测算法,并基于多测点的温度预测结果对未来温差进行估计,建立定子热故障诊断模型。为了缓解故障运行数据较少的问题,本文使用不同核函数的高斯过程生成多种模式的时间序列,并与原数据进行组合,极大地扩充了训练样本空间。最后,利用已有的测试数据开展实验。结果表明,本文所提预测算法的预测效果优于传统自回归积分滑动平均(ARIMA)模型和长短期记忆(LSTM)算法,基于本文所提预测算法建立的故障诊断模型,对运行状态的识别准确率达到91.9%,且具有较高的精度和召回率,从而确保了较低的误报警率和漏报警率。The stator cooling water system of a turbine generator must maintain optimal operating conditions to ensure the reliability and safety of the generator.Typically,thermal faults are detected using methods such as shutdown maintenance or temperature difference thresholds,but these methods cannot effectively detect faults in real time while the generator is in operation.To more accurately identify stator thermal faults,this paper proposes a temperature prediction algorithm based on the Transformer architecture.Using the predicted temperatures from multiple measurement points,the future temperature difference is estimated,and a diagnosis model for stator thermal faults is established.To address the issue of limited fault operation data samples,this paper utilizes Gaussian processes with different kernel functions to generate various types of time series,which are then combined with the original data,significantly expanding the training sample space.Finally,experiments are conducted using existing test data.The results indicate that the predictive algorithm proposed in this paper outperforms traditional autoregressive integrated moving average(ARIMA)and long short term memory(LSTM)algorithms.Moreover,the diagnostic model based on this predictive algorithm achieves an accuracy rate of 91.9%in identifying operational states,while also maintaining high precision and recall rates,ensuring low false alarm and missed alarm rates.

关 键 词：定子热故障 出水温度 TRANSFORMER 高斯过程 

分 类 号：TM311[电气工程—电机] TP18[自动化与计算机技术—控制理论与控制工程]