Study On Soft Sensor Of Wet Ball Key Parameters Based On Multi-Task And Transfer Learning

The ball mill is an important basic equipment which be used in chemical,electric power,construction and other industries,but it is also a high energy-consuming equipment.Its energy consumption and efficiency are closely related to the mill internal load parameters.Therefore,measuring ball mill load parameters accurately is the key to achieve optimal control and energy saving.Due to the features such as rotation and continuous operation of the mill,various direct measurement methods are difficult to implement and the internal load parameters cannot be obtained accurately.In the actual industrial production,the indirect measurement methods based on the data-driven is often used in order to improve the measuring accuracy.However,the traditional soft sensor method neglects the correlation among multiple load parameters of the wet ball mill,which leads to the poor prediction accuracy of the model.So,this study proposes a multi-task regularized extreme learning machine.Multiple load parameters are linked through multi-task learning for the purpose of mining hidden information between multiple parameters.This method improves the prediction performance of the model.In addition,due to the change of ore composition and production plan,ball wear and tear and so on in the actual operation of wet ball mill,it will lead to change the production process conditions and data distribution transfer between auxiliary variables and dominant variables to be tested.The multi-task regularized extreme learning machine cannot obtain satisfactory results when dealing with multi-condition problems.For this purpose,the transfer learning is introduced into the wet mill load parameter measurement,which is used to solve the multiple working conditions of the mill and improve the prediction accuracy of the model.The main research of this article includes the following aspects:1)When the extreme learning machine predicts the wet ball mill load parameters,it ignores the correlation between the load parameters leading to the problem that the prediction accuracy is low.A multi-task regularized extreme learning machine algorithm is proposed and experiments are performed on the wet ball mill dataset.The results show that this algorithm can effectively improve the model’s prediction accuracy and generalization ability.2)When a small number of labeled samples exist in target domain under ball mill operating condition transformation,a domain adaptation algorithm based on a semi-supervised extreme learning machine is introduced.This model minimizes the source and target domain data prediction errors,and adopt a feature transformation matrix which make the source and target domains sharing some knowledge under this feature transformation.This method can effectively solve ball mill load parameters soft measurement problem of a small amount of labeled samples in the target domain under multiple operating conditions..3)When new operating conditions hasn’t labeled sample,a domain adaptation algorithm based on deep extreme learning machine self-encoder is proposed.This model introduces a feature transformation matrix as a hidden variable.Using the hidden variable to map the source and target domain data to a common low-dimensional sub-feature space.This method can effectively solve ball mill load parameters soft measurement problem of no labeled samples in the target domain under multiple operating conditions.