Research On Fault Diagnosis,Prediction And Tolerance Technology Of Advanced Nuclear System Based On Deep Transferable Knowledge

Advanced Modular Nuclear Systems(AMNS)is an important research direction to promote nuclear technology innovation.It mainly adopts modular assembly and distributed design,combined with other renewable energy sources to achieve power supply and multiple applications.In remote areas,it has far-reaching significance for optimizing energy structure and improving power supply.However,the continuous innovation of reactor technology and the complex requirements of application scenarios have brought significant challenges to the reliability and safety of system operation.At the same time,the rapid development of computer technology with artificial intelligence as the core provides new opportunities for the protection and maintenance of AMNS operations,especially the deep learning(Deep Learning,DL)technology,which has a robust characterization of non-linear mapping relationships.Ability is currently widely used in various research fields.This paper first focuses on the complex structure of the AMNS system and the large scale of sensor data and researches the critical technologies of fault diagnosis and prediction by constructing a deep learning network.By analyzing the characteristics of small sample data in the application environment,the characterization form of the input data,the structural features,the optimization scheme of hyperparameter settings are discussed,and the diagnosis and prediction of different AMNS operating accident conditions are realized.After that,in response to the complex and changeable application scenarios of AMNS and the interference of environmental factors,the deep network feature extraction architecture was designed to optimize the noise adaptively.It combines the noise redundancy elimination technology in signal processing to improve the network's anti-noise performance in the virtual environment.At the same time,to solve the critical issues of AMNS historical diagnosis knowledge sharing and model retraining cost optimization,transfer learning(Transfer Learning,TL)technology is used to fully integrate historical diagnosis and prediction knowledge information to establish a diagnosis prediction and fault-tolerance framework based on nuclear knowledge transfer.The system can use the trained deep neural network to effectively realize the training migration and application and combine the output feedback fault-tolerant control strategy to realize the integrated design of the diagnosis and prediction of the accident condition fault-tolerant technology.The simulation results show that the optimized deep network can achieve better diagnosis and prediction accuracy for accident conditions in a noisy environment and has certain advantages than traditional methods;the network after migration learning is better for new platform cases.It has applicability and achieves a good migration effect;the simulation results verify the effectiveness of the designed fault-tolerant framework.In addition,from a complete scientific point of view,this thesis expounds on the overall process of AMNS accident conditions from initial occurrence to further development to initial mastery and elimination;comprehensively mining the system by designing an integrated framework for diagnosis,prediction,and fault tolerance combined with deep learning technology operating status information,to understand the corresponding failure mode in AMNS and its characteristic mechanism.At the same time,from the perspective of the methodology of the thesis,the results of this research can provide a specific reference for the design of deep network optimization and migration technology.