Research And Application Of Set Membership Estimation In The Soft Sensor Modeling For Ball Mill Material Level

Ball mill is the key equipment in the process of industrial production,which is widely used in metallurgy,electric power,mineral processing,chemical industry and so on.The economy is related to the internal material level,low material level leads to low efficiency and energy utilization rate,high material level easily lead to ball mill grinding,and potential safety hazard creats the potential for security problems.Therefore,the accurate measurement of the material level in the cylinder is very important for the optimal control of ball mill.However,due to the characteristics of the ball mill,it is very difficult to measure the position of the ball mill directly in the actual operation,therefore,the data driven modeling method is used to establish the soft sensor model,and the level are estimated by inputting the auxiliary variables related to the material level.There are many traditional soft sensor modeling methods are widely used in the modeling of ball mill material level,such as support vector machines,partial least squares,neural networks and principal component regression analysis.As a kind of neural network,the extreme learning machine(ELM)with its simple and efficient training mechanism,to avoid the process of reverse trimming of traditional neural network,so as to improve the learning speed and generalization of the model,and is widely used.However,the output of the hidden layer of the feed forward neural network is randomly selected by a certain probability,which leads to the randomness of the trained model is very large and the prediction results are unstable.In addition,the neural network structure of ELM single hidden layer also limits the ability of its feature extraction.In order to better learn the abstract feature representation in high dimensional data by unsupervised training process,Researchers have proposed a new depth limit learning machine algorithm(DELM),which uses multiple Autoencoders(AE)stacked layer by layer,using ELM algorithm for error reconstruction,And the output of the previous layer AE as the input of the next layer of AE.Finally,the more abstract feature representation of the data is obtained through the multi-layer AE.However,each layer of DELM network weights are calculated by ELM algorithm,the randomness of the ELM algorithm,resulting in network weights are random and are not optimal,the instability of the DELM model also will be affected.In the actual operation of the ball mill,there are many complex factors,such as time varying and operating mode migration.However,the traditional soft sensor modeling method is based on the existing offline data,once the model is established,it will not be able to track the current object for the newly created test set,which leads to the decline of model prediction performance.Therefore,the soft sensor model needs to be updated.Set membership estimation is a method to describe the set of feasible parameters under the condition of given data set,model structure and noise boundary.The parameters in the set can be considered as the effective parameters of model parameter identification.The optimal bounding ellipsoid algorithm(OBE)is one of the classical algorithms in set membership estimation theory,When the method is applied to the parameter optimization of soft sensor model,the model parameters can be constrained optimization under the condition of given error bounds,and it not only improves the robustness of the model,but also improves the prediction accuracy.Based on this,this paper mainly does the following research:(1)In the process of ball mill experiment,in order to predict the instability of material level in ball mill,this paper uses OBE to optimize the trained ELM network model under the condition of unknown but bounded error,and improve the prediction accuracy and stability of the model.The experimental results show that the method is effective.(2)In order to better extract the highest level of abstraction in the sample,the depth network to model the soft data of the ball mill data is used.Based on the idea of self-encoder reconstruction,OBE is used to learn the high-level representation of the input data.Finally,the ELM algorithm is used to obtain the relationship between the high-level features and the sample labels.And In order to verify the effectiveness of the proposed algorithm,the traditional UCI dataset and the actual ball mill data set are chosen as experimental data to verify that the algorithm has better prediction performance in regression and classification.(3)In order to solve the problem of time-varying and condition migration in ball mill material level,this paper proposes a dynamic soft-sensing model based on OBE-PLS.Firstly,offline measurement is used to train the soft-sensing model.When the new query sample arrives,the parameters of the model are dynamically adjusted based on the original model,so the real-time tracking of the model is realized.The effectiveness of the method is proved by experiments.