Research On Process Monitoring Method Of Grinding-classification Based On Semi-supervised Learning

The grinding-classification process is a key section in the production of mineral processing industry,the running state of it directly affects the important production indexes such as output,energy consumption and concentrate recovery rate.Therefore,the fault monitoring of grinding-classification process has received much attention.However,as a result of the complexity and uncertainty of the grinding-classification process,the traditional process monitoring method has encountered difficulties.Therefore,based on the analysis of the existing process monitoring methods,combined with the actual situation of modern grinding-classification process has widely used the computer and instrument technology,with full use of a lot of operation data of grinding-classification process and knowledge from expert,the research on the method of fault monitoring for grinding classification process is started,which is based on semi-supervised learning,the main contents include:(1)Study the current situation and development trend of data-driven process monitoring methods and focus on analyzing the basic principles and main problems of process monitoring methods of machine learning.(2)For local and global consistency is a transductive semi-supervised learning method with bad real-time and not suitable for online monitoring,by introducing the linear mapping function,proposes an inductive process monitoring method based on local and global consistency.The method uses a small amount of labeled data and a large number of unlabeled data,and the semi-supervised learning model is used to monitor the process.This method not only can identify the fault accurately,but also can directly determine the specific fault types.(3)A new process monitoring method based on semi-supervised learning is studied to solve the problem of unknown fault in the grinding-classification process.The method uses a unique state label matrix and combines with optimal calculation analysis,can be used to realize the accurate identification of typical faults and unknown faults.(4)In semi-supervised learning,the label has a guiding effect in learning,once an error occurs,the semi-supervised learning will be degraded.Aiming at the problem of label noise,based on the above method,the new method has some fault-tolerance ability to the noise in the label by improving the constraint of the fitting term.(5)Through the simulation experiments of the actual operation data of the grinding-classification process,the validity and practicability of the above methods in identifying the unknown faults and handling the label noise are proved.