&nbsp; &nbsp; Fault Detection and Diagnosis (FDD) is crucial to guarantee a greater safety and economy of Nuclear Power Plants (NPPs). With the development of artificial intelligence technology, data-driven methods have become the core of FDD in NPPs. Despite the advantages of high flexibility and practicability, data-driven methods may be sensitive to the noise in measurement data, which is inevitable as the equipment ages, or under fault conditions in NPPs. However, there are currently few researches on robustness analysis and robustness improvement of data-driven methods under noisy data.&nbsp; &nbsp; In this paper, a FDD model of NPPs based on Random Forest (RF) is established. Firstly, its diagnostic performance on a noiseless data set containing 13 operating conditions (one steady state condition and 12 fault conditions) is analyzed by 10-fold cross-validation. The results show that the average accuracy rate is 99.43%, and the average macro-f1 score is 0.9946, which indicates that its superior diagnostic performance on noiseless data. Next, the diagnostic performance of RF model under noisy data is studied. The results show that the robustness of RF model under noisy data is exceedingly poor. When the noise level is 1% (noise level is defined as the ratio of the standard deviation of the added noise to the standard deviation of the original data), its accuracy has been reduced to 89.82%, the macro-f1 score has been reduced to 0.9155. And with the increase of noise level, the diagnostic performance of RF model decreases sharply, when the noise level is 10%, the accuracy and macro-f1 score are only 69.17% and 0.7020 respectively.&nbsp; &nbsp; In order to improve the robustness of RF model under noisy data, a method named ‘Train with Noisy Data’ (TWND) is proposed in this paper, in which the training set is composed of noiseless training samples and noisy samples of different noise levels. The results show that TWND method can effectively improve the robustness of RF model under noisy data, and the degree of improvement is related to the noise levels of added noisy data. The RF model achieves the best diagnostic performance when noisy data with noise level of 1%-30% is added to the training set, and when diagnosing the noisy data of 10% noise level, the accuracy and macro-f1 score are dramatically raised to 96.55% and 0.9666 respectively. In addition, applying the TWND method will not reduce the performance of RF model on noiseless data, when the noise levels of the noisy data added to the training set are between 1% and 30%.&nbsp; &nbsp; In summary, a FDD model of NPPs based on random forest is constructed in this paper, its performance and robustness under noiseless data and noisy data are analyzed, and a method of adding noisy data into the training set is proposed to greatly improve the robustness of the random forest model under noisy data. This paper can provide reference for robustness analysis and robustness improvement of nuclear power plant fault diagnosis model based on other data-driven methods.

Robustness Analysis and Improvement of Fault Diagnosis Model for Nuclear Power Plants Based on Random Forest

Paper Proceedings: https://asmedigitalcollection.asme.org/ICONE/ICONE28/volume/85246

2021 28th International Conference on Nuclear Engineering (ICONE28)

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.