Fault Dianosis On Nuclear Power Circuit System Based On Deep Learning

Ensuring the safe and stable operation of nuclear power units is the most critical issue in the field of nuclear power,and the use of fault diagnosis technology can significantly improve the continuous and reliable operation time of nuclear power units and avoid the occurrence of major faults as much as possible.Therefore,the research of fault diagnosis for nuclear power units is of great practical significance and application value.Deep learning technology is widely used in fault diagnosis research of industrial systems due to its powerful capability in data classification and feature extraction,but the traditional fault diagnosis methods are limited by the problems of data acquisition due to the scarcity of nuclear power fault data and expensive fault experiments,and the application of fault diagnosis models due to the noise interference and time-variable dimensional coupling of the actual state data of nuclear power units,which are difficult to be grounded.It is difficult to apply the traditional fault diagnosis methods to the actual industrial site.Therefore,this paper investigates the deep learning-based fault diagnosis method of nuclear power unit two-loop system from two perspectives of data acquisition and method application,and the main research contents and results are as follows.Construct a nuclear power failure data warehouse based on digital twin.Using a modular modeling approach,a digital twin for the MSR system and a digital twin for the condenser of a nuclear power unit are established on the APROS integrated simulation platform to realize high-precision simulation of its steady-state process and dynamic process.Based on the digital twin system,four typical failure models are established for the MSR,including uneven flow rate,breakage,heat transfer deterioration and valve characteristic change,and five typical failure models are established for the condenser system,including abnormal operation of the pumping equipment,poor heat transfer effect of the condenser,insufficient circulating water,influence of heat source and poor vacuum tightness.On this basis,the corresponding fault data are collected to build a fault data warehouse for the steam-water separation and reheat system and condenser system,so as to solve the problem of data acquisition for nuclear power failure data scarcity and expensive fault experiments.An offline fault diagnosis model for nuclear power two-loop systems is constructed based on deep residual networks.Since the unique residual connection structure of the deep residual network enables it to build a deep network and effectively improve the feature extraction capability,and the two-dimensional convolutional operation of the convolutional layer enables it to effectively process the time-variant two-dimensional image data,the network structure of the fault diagnosis model is designed based on the deep residual network.The number of residual blocks and the number of convolutional kernels in the convolutional layer inside each residual block are optimally designed using an artificial bee colony algorithm to achieve the optimal network structure with the highest model fault diagnosis accuracy and the lowest total number of computational parameters.The channel attention mechanism and spatial attention mechanism are added to enable the network to learn important features automatically,so as to speed up the network training and improve the fault diagnosis accuracy.Then,sliding window,edge-zeroing and normalization operations are performed on the fault datasets of MSR system and condenser system to obtain a uniform size fault dataset for network training and verification.Finally,a fault diagnosis model A-DRNS based on attention mechanism and deep residual network is established in this paper,which is compared with the traditional DRSN,VGG and MLP models.The results show that the A-DRSN model designed in this paper has faster training convergence,higher fault diagnosis accuracy and better resistance to noise interference.Building an online fault diagnosis model for nuclear power two-loop systems based on incremental learning.In order to make the traditional static,offline fault diagnosis model have the ability to continuously learn new tasks,this paper designs a C-i LP incremental learning algorithm in terms of both model structure and update method.For the model structure,the softmax classifier in the A-DRSN model is changed to a class-averaged classifier that does not need to change the network structure when learning a new task.For the update method,the fine-tuning-joint training method based on the replay strategy is used to build a training data set with a fixed number of samples;the knowledge distillation method based on the regularization strategy is used to add distillation loss terms of old knowledge in the loss function;the fixed feature space parameters method based on the parameter isolation strategy is used to isolate and protect the effective features learned by the old task.Comparing with the traditional incremental learning algorithms i Ca RL,Lw F and Pack Net,the results show that the C-i LP algorithm designed in this paper has a smoother degradation of model performance and higher fault diagnosis accuracy when learning new tasks.The fault diagnosis research based on deep learning in this paper can effectively solve the problems of data acquisition and method application of traditional fault diagnosis methods,and can be applied to the fault diagnosis of nuclear power secondary circuit system to ensure the safe operation of nuclear power units.