Research And Application Of Transfer Learning In Soft Sensor Of Wet Ball Mill Load Parameters

Wet ball mill is a crucial infrastructure in process industry,especially in power,mineral processing,chemical and metallurgical fields.However,because of the influence of nonlinear,strong coupling,large time delay and physical condition restriction on the running process of the ball mill.It lacks effective measurement methods for key parameters(material to ball volume ratio,pulp density and charge vo lume ratio),which can express the internal load of wet ball mill.Thus,there are many problems such as high energy consumption,difficult to realize the optimization control of grind ing process and the existence of hidden safety problems,which seriously restrict the improvement of the automation level of this kind of basic equipment.The soft sensing modeling method is an effective solution to solve the above problems in industry by estab lishing the function mapping relation between the auxiliary variables and the leading variables to realize the prediction of the unknown leading variables.However,when the change of the production task,the restructure of the equipment,the material and the operating environment change,the ball mill will run in the multi mode and multip le conditions environment,which will result in the change of the data distribution of the working condition(target domain)and the modeling condition(source domain),which makes the trad itional soft measurement model not accurately predicted.Aiming at the problem that it is difficult to model and predict the unknown modal lead ing variab les in the process industry,this paper introduces the transfer learning strategy into soft sensor,focusing on the soft sensor of the load parameters of the wet ball mill under the multi mode unknown mode.The main contents can be summed up as follows :(1)In order to solve the problem that It is difficult for a wet ball mill to get the label of the dominant variab le when working under multip le conditions,a jo int distribution adaption algorithm is introduced,which can match the marginal distribution and conditional distribution in the process of dimensionality reduction.And the load parameters of wet ball mill are modeled by soft sensor from multi-task and multi-source domain respectively.(2)In the case of only a small number of labels in the target domain,first,a jo int maximum mean discrepancy regular term is introduced in the deep autoencoder network to reduce the difference in the marginal distribution between source domain and target domain.Next,a small number of target domain labels are used to construct the inter domain label mapping model.Finally,combined with the above two models to predict the load parameters of ball mill.(3)In view of the fact that there is only a small amount of label in the working condition,the domain adaptive extreme learning mchine is introduced to transfer the knowledge and model from source domain,and the soft sensor model of the load parameter of the wet ball mill is established by using the regular term of manifo ld which can maintain the geometric structure of the data.(4)In order to solve the problem of distributed processing and modeling of large industrial data,the parallel imp lementation algorithm of domain adaptive extreme learning machine based on distributed processing platform of cluster server is studied,and its efficiency is verified by the soft sensor modeling of ball mill load parameters.