Process Monitoring Of Grinding-Classification Based On Robust Semi-Supervised Learning

The characteristic of modern industry is that a large amount of process data is stored in the production process,but how to use the data is a challenge.Against this background,data-driven process monitoring and fault diagnosis methods have been widely studied and applied.However,in many data-driven methods,the combination of production experience and expert knowledge with unmarked process history data is not researched very well.On the premise of retaining a large number of untagged process historical data,this thesis combines production experience and expert knowledge for process monitoring and fault diagnosis research.The main research contents include:(1)Firstly,the research status of data-driven process monitoring method is analyzed,and the overall monitoring strategy of this thesis is proposed.At present,the method of process monitoring based on statistical analysis can not take advantage of expert knowledge and the workload of supervised learning method is too large and a large number of unmarked process data can not be used,therefore,we confirm the semi-supervised learning method as the process monitoring method studied in this paper.(2)The semi-supervised learning process monitoring method for attribute noise is studied.In the actual grinding-classification production process,due to the external environment,equipment internal vibration,sensor sensitivity and other factors,making the production data attached to a variety of noise,seriously affecting the monitoring results of the process.Since the L1 norm regularization technique has many applications in removing noise,a semi-supervised learning model based on L1 norm regularization is established,which reduces the influence of the noise in industrial data on the process monitoring to a certain extent.A small amount of markup data is combined with a large amount of unlabeled historical process data to construct a neighborhood weighted graph,and then an optimization objective function is established.After optimizing the objective function,a state label of the monitoring data can be obtained.The monitoring and simulation of the data of grinding-classification proves the robustness and validity of the method(3)The semi-supervised learning process monitoring method for label noise is studied.In semi-supervised learning model,sample marking is a key process because the label has a guiding role in the subsequent learning process.If the marking is wrong,the performance of the algorithm will be degraded.However,due to lack of knowledge or man-made negligence,marking errors are sometimes unavoidable.Therefore,the study of how to deal with this problem at the algorithm level can weaken the influence of label noise on the process monitoring accuracy to a certain extent.The monitoring and simulation of the data of grinding-classification proves the robustness and validity of the method.