Evidence Driven Condition Monitoring Method Under Big Data With Application Of Bolier Water Wall

The efficient and stable operation of large units is directly related to the economic benefits of thermal power generation enterprises and the safe operation of the power system.Therefore their sudden failure can bring a huge negative impact on the economic benefits of the enterprise,and even more,endanger the lives of the workers operating on site.Researchers have proposed a large number of condition monitoring methods based on different mechanisms to monitor the operating condition of individual thermal equipment in thermal power plants in real time and to provide early warning of possible abnormal conditions of the equipment.However,with the development of data collection techniques,the availability of historical operating data of the relevant equipment has increased exponentially,which poses two challenges for the existing condition monitoring methods.The first is that as the amount of available data increases,the amount of imprecise and uncertain information between different operating conditions of the equipment implied in the operating data increases,and the existing methods cannot describe both types of incomplete information;the second is that the increase in the amount of available data increases the computational load of the computer,leading to the problem of overflowing computational memory and long running time due to the high complexity of the algorithm.To address the above two challenges,this thesis proposes two sets of big data evidence-driven condition monitoring modeling methods for different computing resource availability with the help of the advanced Apache Spark parallel computing framework,both of which can fully describe the imprecise and uncertain information among the operating states hidden in a large amount of historical operating data and can maximize the use of all available The application examples on boiler water-cooled walls show that the proposed method can quickly and accurately complete the tasks of equipment health monitoring and abnormal condition warning within a reasonable computing time,which is of great practical engineering significance to the safe operation of thermal power units.Specifically,this thesis carries out the following main research and innovations.1、For the offline engineering problem of how to mine the equipment operating conditions existing in a large amount of equipment historical operating data and describe the imprecise and uncertain information between different operating conditions,a method of using distributed evidential clustering to build an equipment big data evidence base is proposed.The method fills the gap that existing clustering algorithms cannot describe the imprecision and uncertainty of device states at the same time,and breaks the computational bottleneck of existing clustering algorithms in the face of large data.The results on a commonly used machine learning simulation dataset show that the proposed distributed evidence clustering algorithm has superior clustering performance as well as efficient scaling efficiency.The results of constructing a real engineering problem with an evidence base on a large amount of historical operating data of boiler water-cooled walls show that the proposed distributed evidence clustering algorithm accurately determines the number of water-cooled wall operating states implied in the big data within an acceptable operating time,and reasonably completes the classification of water-cooled wall operating states at each historical moment while fully describing the imprecise and uncertain information between the classes,providing a This provides high-quality supervised training data for the construction of a big data evidence-driven water-cooled wall condition monitoring model.2.To address the online engineering problem of how to diagnose the real-time operating state of equipment and achieve real-time abnormal state warning when large-scale computing resources can be used online for a long and stable period of time,a method of using distributed evidence K-nearest neighbor classification to build a real-time equipment state diagnosis model is proposed.The method is able to make full use of the imprecise and uncertain information among the operational states described in the constructed evidence base of equipment big data,while overcoming the drawbacks of the evidence K-nearest neighbor classification algorithm in terms of high complexity of parameter estimation and slow generation of classification results in the face of a large amount of available supervised training data.The results on commonly used machine learning simulation datasets show that the proposed distributed evidence K-nearest neighbor classification algorithm has superior classification accuracy as well as scaling efficiency.The results on the actual engineering problem of establishing a watercooled wall condition monitoring model show that the proposed distributed evidence K-nearest neighbor algorithm can flexibly diagnose the real-time operating status of the water-cooled wall and complete the abnormal condition warning of the water-cooled wall in a timely manner.3.For the offline engineering problem of how to maximize the reduction of the amount of supervised data of the constructed equipment when the distributed evidence K-nearest neighbor condition monitoring model cannot be constructed due to the inability to use large-scale computing resources online for a long and stable period of time,a method of streamlining the size of the equipment big data evidence base using distributed evidence sample selection is proposed.The method is able to maximize the elimination of redundant data from the device big data evidence base without degrading the performance of real-time device condition monitoring.Results on a commonly used machine learning simulation dataset show that the proposed distributed evidence sample selection algorithm has a powerful ability to reduce data size as well as efficient scaling efficiency.The results in the practical engineering problem of streamlining the water-cooled wall big data evidence base show that the proposed distributed evidence sample selection method is able to retain only those data that contribute the most to the building of water-cooled wall condition monitoring models,providing high-quality streamlined supervised training data for the building of water-cooled wall condition monitoring models with limited computational resources.4.To address the online engineering problem of how to build an equipment condition monitoring model with comparable performance to distributed evidence K-nearest neighbors by improving the existing evidence K-nearest neighbor algorithm based on a streamlined equipment big data evidence base without long-term and stable online use of large-scale computing resources,we propose a method for building a real-time equipment condition diagnosis model using sparse reconstructed evidence K-nearest neighbor classification.method.The sparse reconstruction evidence K-nearest neighbor algorithm can overcome the "distance concentration" problem caused by the over-sensitivity to hyperparameter K and the use of Euclidean distance in the traditional EKNN algorithm,and thus improve the classification accuracy.The results on a commonly used machine learning simulation dataset show that the proposed sparse reconstruction evidence K-nearest neighbor classification algorithm has superior classification performance than the existing variant evidence K-nearest neighbor classification algorithm.In the results of building a water-cooled wall condition monitoring model for practical engineering problems,it is shown that the proposed sparse reconstruction evidence K-nearest neighbor condition monitoring method not only flexibly diagnoses the realtime operating status of the water-cooled wall,but also has an early warning performance that does not significantly lag behind the distributed evidence K-nearest neighbor condition monitoring method.