The Research On Machine Learning Theory And Method For Thermal Process Monitoring And Diagnosis

Thermal process monitoring and diagnosis of modern large-scale coal-fired power units are important for energy conservation and emission reduction.It is also important for promoting our country’s power industry to a new power system with recycled energy as the main body.With the development of information technology,massive operating data provides a solid foundation for developing data-driven machine learning methods,which profoundly affect the development of thermal process monitoring.This paper focuses on the machine learning theory and method,especially the multivariate statistical analysis,and carries out studies on the monitoring and diagnosis of thermal processes.The main research contents are arranged as follows:（1）Aiming at the dynamical issues in thermal processes,a dynamic predictable feature analysis method is proposed considering time-series dependencies among process paramaters.First,the reliability of modelling thermal processes using dynamic latent variable analysis is revealed from static and dynamic perspectives,respectively.Second,combined with the k NN method and graph embedding technology,a graph-based predictable feature analysis（GPFA）model is established.In GPFA,conditional covariance is employed to characterize the predictability of dynamic latent variables.Then,an evaluation index termed prediction information（PI）based on the information entropy theory is proposed to optimize the model parameters,and!statistics and SPE are used to monitor the thermal process.Finally,case studies on the Tennessee Eastman process and a flue gas waste heat utilization system are conducted to demonstrate the superiority of GPFA over other approaches in terms of fault detection performance.（2）Aiming at the uncertainty of thermal data caused by measurement noise,a probabilistic predictable feature analysis（PPFA）method is proposed.First,a high-order state space function based on the multi-step autoregressive model of latent variables is conducted to characterize the dynamic system.In PPFA,the distribution of noises is captured in a probabilixtical manner,and thus a full probabilistic interpretation of measurement noises is obtained.Second,an improved EM algorithm and a forward-backward algorithm based on the Kalman filter are used to optimize the parameters,and a dynamic model based on the PPFA method is obtained.Then,a new dynamic statistic is proposed as the complement of!statistic and to reveal the fluctuation of a dynamic system,and the overall framework of fault monitoring is proposed.The validity of the proposed algorithm is proved by its application on the three-phase flow facility（TPFF）and the real fault cases of a medium-speed coal mill in a subcritical 600MW coal-fired power plant.（3）Aiming at coping with the heavy-tailed characteristic of measurements when contaminated with outliers,a robust probabilistic predictable feature analysis（RPPFA）method is proposed.First,a form of Gaussian-Student t mixture distribution is designed to explain the heavy-tailed characteristic of measurement noises,which overcomes the problem of poor robustness of a single Gaussian hypothesis to outliers.Second,for the estimation of model parameters under the condition of mixed distribution,three conjugate distribution forms of Gamma distribution,Beta distribution and Bernoulli distribution are introduced respectively,which simplifies the process of parameter calculation.Then,in the forward and backward algorithm of the E step in the EM algorithm,a weight correction coefficient modifying the Kalman gain is introduced,and the parameters of the forward recursion are corrected.Finally,through the application of TPFF and a medium-speed coal mill,the superiority and robustness of the proposed RPPFA model for thermal process monitoring are proved.（4）Aiming at the problem that the complex data information of multiple operating modes cannot be covered by the model established based on the single operating data,a multi-mode probabilistic predictable feature analysis（MPPFA）method is proposed.First,the deep autoencoder（DAE）and the Gaussian mixture model（GMM）are fused,with which DAE extracts the low-dimensional features,and GMM is further used to classify the features.Second,the original data is divided into sub-datasets representing different operating modes,and local PPFA models are established respectively,and the probability weighting method is used to aggregate them into MPPFA.Then,Bayesian theory is introduced to extend the deterministic reconstruction model based on deep autoencoders to a probabilistic mapping model.Five monitoring statistics,"!,",",_#!and_#are proposed for process monitoring,and the contribution indices,_$^%"&!and_$^%"&"are proposed for fault isolation.Through the application of TPFF and the coal mill,the adaptability and effectiveness of the proposed MPPFA method for monitoring and diagnosis of the multi-mode thermal process are verified.（5）In the thermal process monitoring and diagnosis system,the proposed algorithms are integrated and an application system is designed and developed.A complete development plan is given from the system architecture,system deployment,and system core functions.The system is verified based on the case analysis of practical faults in a coal-fired unit.The results show that the development and application of online monitoring and diagnosis system can effectively improve the operation level of thermal processes.Based on the system,the operation adjustment of the coal-fired unit can be completed in time,and the early warning and diagnosis of the system equipment can be realized.