Research On Condenser Fault Diagnosis Based On PSO-DDSAEN

Condenser is one of the most important auxiliary equipments in power plant steam turbine units.Its operation status directly affects the safety and economy of the whole power plant.Therefore,the research on condenser fault diagnosis method has very important theoretical significance and engineering practical value.With the complication and precision of modern power plant equipment,the complexity of condenser faults is also increasing.Its fault data are characterized by complexity,nonlinearity and high noise.Traditional condenser fault diagnosis methods are mostly based on shallow machine learning models,which have certain limitations on the feature learning ability of complex faults,poor scalability of diagnosis performance,and often lead to problems such as falling into local optimization and over-fitting.Compared with the shallow machine learning model,deep learning has strong feature learning ability,and can dig the hidden information in more depth for complex original data.Therefore,based on Deep Auto-Encoder Networks in deep learning and the condenser equipment in power plants,this paper introduces the deep learning algorithm into condenser fault diagnosis for the first time.Aiming at the practical problems of difficult data marking,scarce fault data and high noise in power plant condenser operation,a condenser fault diagnosis method based on Deep Denoising Sparse Auto-Encoder Networks(DDSAEN)model is proposed.As a kind of deep neural network,DDSAEN uses unsupervised learning method to learn the greedy features of a large number of unlabeled data layer by layer,and then uses a small number of labeled data to fine-tune the model.The introduction of sparsity constraints in the network model makes the proposed method still have strong learning ability for scarce fault data,improving feature representation and enhancing feature sparsity.At the same time,with the addition of denoising strategy,the proposed method can still extract more characteristic features from fault data withdifferent noise levels and enhance the robustness of the features.Through different experiments,the performance of the DDSAEN diagnostic model is tested,which verifies the strong feature learning ability of the method.Compared with the shallow machine learning model,the method has certain advantages and better diagnostic performance when facing unlabeled data,imbalance,high noise and other conditions.At the same time,in the construction of the DDSAEN model,this paper analyzes and studies the influence of key parameters such as network layer number,hidden layer node number,iteration number,sparse parameter,noise figure,etc.on the diagnostic performance of the DDSAEN model through different experiments and comparisons.Finally,in view of the complex structure of the DDSAEN model,the parameter determination method of its manual experiments is stochastic,time-consuming and inefficient.In this paper,Particle Swarm Optimization(PSO)is adopted to automatically optimize the key parameters of the network model,and a condenser fault diagnosis method based on PSO-DDSAEN is proposed.Through comparative experimental analysis and research,the effectiveness of the improved method is verified,and the method has better diagnostic performance and intelligence.