报告详情
Cross-Power-Level Fault Diagnosis of Nuclear Power Plants Based on Source-Free Global-Local Clustering Adaptation
编号：39
访问权限：仅限参会人
更新：2026-10-21 21:50:57 浏览：5次
口头报告
摘要
With the increasing application of artificial intelligence and digital instrumentation and control systems in nuclear power systems, data-driven intelligent fault diagnosis has developed rapidly over the past decade. However, most existing methods assume that sufficient training data are available and that the training and test data follow the same distribution. In actual operation, the power level of a nuclear power system is adjusted according to operating demands, leading to a significant distribution discrepancy between the historical operation data (source domain) and the data at new power levels (target domain), so that models trained on limited power levels degrade markedly at unseen power levels.
Conventional domain adaptation usually aligns the feature distributions of both domains simultaneously to improve the target-domain diagnosis performance, but it requires access to the source-domain data during adaptation, which is often impractical in nuclear power systems owing to privacy and security requirements, the high cost of joint training and long-term retention, and the delivery of pre-trained models without training datasets. Therefore, this paper performs cross-power-level fault diagnosis using only the source-domain pre-trained model and completely unlabeled target-domain data during adaptation.
For cross-power-level fault diagnosis in nuclear power systems, this paper adopts a source-free global-local clustering adaptation method that, without changing the network architecture or updating the source-domain classifier, completes the adaptation through global class-structure modeling and a local neighborhood-consistency constraint, thereby improving the discriminability of target-domain features.
Seven cross-power-level tasks are constructed from 20 typical conditions, with the source domain fixed at 100% power and target domains covering 92%, 90%, 85%, and 80% power; each task is repeated with five random seeds. The tasks vary either the number of diagnosed classes or the target-domain power gap to examine their separate influences on diagnosis performance.
Experimental results show that, compared with the source-domain pre-trained model, the proposed method raises the average diagnostic accuracy over the seven tasks from 89.56% to 94.16%. With a fixed power gap and a varying number of classes, the average accuracy is 93.69%; with a fixed class set and a varying power gap, it is 94.77%, indicating that the method adapts well to both variations.
These results indicate that, under the constraint that operation data cannot leave the domain, the proposed method completes cross-power-level adaptation without accessing source-domain raw data, satisfying privacy and security requirements while reducing data retention cost, and provides a feasible approach for the engineering application of source-free fault diagnosis in nuclear power systems.
Conventional domain adaptation usually aligns the feature distributions of both domains simultaneously to improve the target-domain diagnosis performance, but it requires access to the source-domain data during adaptation, which is often impractical in nuclear power systems owing to privacy and security requirements, the high cost of joint training and long-term retention, and the delivery of pre-trained models without training datasets. Therefore, this paper performs cross-power-level fault diagnosis using only the source-domain pre-trained model and completely unlabeled target-domain data during adaptation.
For cross-power-level fault diagnosis in nuclear power systems, this paper adopts a source-free global-local clustering adaptation method that, without changing the network architecture or updating the source-domain classifier, completes the adaptation through global class-structure modeling and a local neighborhood-consistency constraint, thereby improving the discriminability of target-domain features.
Seven cross-power-level tasks are constructed from 20 typical conditions, with the source domain fixed at 100% power and target domains covering 92%, 90%, 85%, and 80% power; each task is repeated with five random seeds. The tasks vary either the number of diagnosed classes or the target-domain power gap to examine their separate influences on diagnosis performance.
Experimental results show that, compared with the source-domain pre-trained model, the proposed method raises the average diagnostic accuracy over the seven tasks from 89.56% to 94.16%. With a fixed power gap and a varying number of classes, the average accuracy is 93.69%; with a fixed class set and a varying power gap, it is 94.77%, indicating that the method adapts well to both variations.
These results indicate that, under the constraint that operation data cannot leave the domain, the proposed method completes cross-power-level adaptation without accessing source-domain raw data, satisfying privacy and security requirements while reducing data retention cost, and provides a feasible approach for the engineering application of source-free fault diagnosis in nuclear power systems.
关键词
Nuclear Power Systems; Fault Diagnosis; Cross-power-level; Source-free Domain Adaptation
报告人
Mingxu Dong
Master's Student Harbin Engineering University稿件作者
全部评论
重要日期
-
会议日期
10月30日
2026
至11月01日
2026
-
09月30日 2026
初稿截稿日期
主办单位
Harbin Engineering University
State Key Laboratory of Marine Thermal Energy and Power
Heilongjiang Nuclear Society
State Key Laboratory of Marine Thermal Energy and Power
Heilongjiang Nuclear Society
发表评论