Machine learning based fault diagnosis methods in nuclear power plants (NPPs) often exhibit lack of labeled fault data in sufficient quantity, thus posing a significant challenge for the supervised fault diagnosis. This paper presents a novel, uncertainty-aware, and unsupervised fault diagnosis framework for pressurized water reactor (PWR) type NPP. For this purpose, it integrates the k-nearest neighbors (KNN) method with shapley additive explanations (SHAP) analysis, not only to enhance the model interpretability but also to ensure its reliability. The KNN model, trained on steady-state operational data, firstly detects anomalies, while SHAP analysis subsequently performs the fault classification in an unsupervised manner. For the purpose of fault classification, feature contributions are computed for each detected anomaly, thus providing an interpretable features pattern. This pattern information is further analyzed and transformed into an entropy-based metric, which is calculated from the normalized SHAP values. This metric enables a transparent classification between the sensor and process faults. Regarding the sensor faults, this methodology further provides the identification of the faulty sensor through the analysis of the highest contributing feature. In this way, explainable artificial intelligence (XAI) has been utilized not only for model interpretation but has also played its role in the fault diagnosis process. Additionally, a bootstrap resampling approach has been incorporated for the estimation of model uncertainty, thus providing the confidence measures that support safer and risk-informed decisions. Performance of the proposed methodology has been evaluated against various sensor and process faults, not only through clean data but also against noisy data. An inter-comparison among various anomaly detection methodologies within the proposed framework has also been presented. The testing results demonstrate that the proposed method achieves 99.56% ± 0.051% (mean ± standard deviation) classification accuracy across 100 bootstrap iterations, and also shows resilience under noisy conditions. This fully unsupervised, interpretable, and uncertainty-informed framework thus offers a significant advancement for reliable fault diagnosis in safety-critical nuclear power plant operations.
扫码关注我们
求助内容:
应助结果提醒方式:
