Research And Application Of Transfer Learning In Soft Sensor Of Wet Ball Mill Ld Paoarameter

Ball mills are a basic grinding equipment in the process industries such as mineral processing,power and chemical industries.The efficiency of the industrial equipment is closely related to its internal load.It is of great significance to adjust the loading and feeding amount of the mill in time to make the mill run stably at the optimal load point,improve the quality of the grinding product and the safe and stable operation of the grinding system.Therefore,the accurate detection of the critical load parameters(Material to ball volume ratio,Pulp density,Charge volume ratio)that can characterize the internal working state of the ball mill has a decisive effect on the optimal operation control of the ball mill.The mechanism model limited to the ball mill is complex,with large time lag,large random disturbance,etc.,and the conventional instrument detection method cannot be implemented.Therefore,it is possible to predict the unknown dominant variable by establishing a function mapping relationship between the auxiliary variable and the dominant variable.Soft measurement modeling methods have become one of the key technologies in this research field.However,in the actual production process,there are steel ball wear and sudden addition of steel balls and other operations,resulting in changes in working conditions and changes in data distribution,the futher leading to that the model can not accurately measure the load parameters.In view of the problem of migration of working conditions during the operation of the above wet ball mill,this paper introduces transfer learning into the soft measurement algorithm,focusing on the soft measurement method of load parameters of wet ball mill under multiple working conditions.The main research contents can be summarized as follows:1.Aiming at the problem of the model misalignment caused by the mismatch between the historical data and the real-time data,and fewer samples in real-time working conditions,a soft measurement model based on the migration variable self-encoder-tag mapping is proposed.Firstly,the hidden variable distribution parameters obtained by encoding the target domain data,and then the corresponding hidden variables of the source domain data are fitted,and the transfered data obtained by decoding.Then the similarity measure is used to select the similar samples to construct the label mapping model,and the mapping label is obtained.Finally,transfered data and mapping tags build the final soft-measurement model.2.Aiming at the problem of the model misalignment caused by the mismatch between the historical data and the real-time data,and fewer samples in real-time working conditions,the domain adaptation idea is introduced,and a soft measurement model based on domain adaptive support vector regression is proposed.The feature information contained in a small amount of tagged sample data in the target domain is transferred to correct the source domain model,and the adaptability of the model constructed by source domain data to the target domain data is improved.3.Aiming at the problem of the constraint of high-dimensional data spatial distribution structure is lacking in the domain adaptive support vector regression method,a soft measurement model based on domain adaptive manifold regular support vector regression is proposed.Based on the SVR that suitable for nonlinear data,the DASVR model is constructed by a small number of labeled samples in the target domain,and the manifold regularization term is introduced into the DASVR to constrain the spatial structure of the data.