| With the continuous improvement of production level and automatic control technology in modern mineral processing industry,the process of grinding classification is developing towards automation and intelligence.It is developing in the direction of intelligence and complication.In order to ensure the safety,stability and higher economic efficiency of production,it is becoming more and more important to monitor the industrial process that may have failure.At present,the main method in the field of process monitoring is based on data-driven method.It is based on the collected process data and uses the classifier trained by the historical data of the system in normal and various fault situations to realize fault diagnosis.However,for a new production process or a process with different distribution of modeling data and test data,the traditional data-driven method is usually difficult to apply because of the lack of necessary operation experience and a large amount of running data.In this thesis,the process monitoring method based on model transfer learning is studied under the industrial background of grinding classification process.The main research contents are as follows.(1)First of all,summarize the process monitoring methods.Then,the research status of data driven process monitoring method is analyzed.Finally,through the existing process monitoring methods can not solve the problem leads to model transfer learning,and model transfer learning research status and related methods to do a detailed overview.(2)In the process flow of modern mineral processing,when it is necessary to adjust the operating conditions and process equipment in the process or to establish a new grinding and grading production line,if the new process has only a small number of marking samples,in order to train the new effective process monitoring model,It is still necessary to continue to manually annotate the data for the new process.This method not only reduces the efficiency of modeling,but also consumes a lot of manpower and material resources to mark samples.In order to solve the problem that there are only a few labeled samples in the new industrial process and the fault diagnosis model needs to be established quickly.In this thesis,a model transfer learning algorithm based on least squared twin K-class support vector classification is proposed.It transfers the old process monitoring model to the new process,and combines a small number of labeled samples of the new process to continue learning,so as to establish a monitoring model of the new process on the basis of the prior model of the old process.(3)In order to solve the problem that there are only unlabeled samples in the new industrial process and the fault diagnosis model needs to be established quickly.In this thesis,a neural network model migration method is proposed.While training the old process model,the model parameters are migrated to the new process.The Maximum Mean Discrepancy metric is used as regularization in the back propagation training to make the hidden layer distribution of the two process neural networks closer.Furthermore,the network depth is added to enhance the representation ability of the extracted features,and the convergence speed is trained by adding the batch standardized lifting model.(4)In the simulation experiment,aiming at the above problems encountered in the process of grinding classification production,this thesis first uses the data of TE simulation industrial process platform to verify the validity of the two methods mentioned above.Then,the practicability of the proposed method is verified by the real data of grinding classification production process. |
