Research And Application Of Load Parameter Soft Sensor Based On Transfer Learning In Multi Working Conditions

Ball mill is a typical device which applied to the process industry of electric and chemical.The detection of ball mill’s load parameters plays important roles in optimizing production,running the equipment safely and saving energy.At present,the soft sensor method is used to predict the load parameters of ball mill.However,the data used to establish traditional soft sensor model has to with the same distribution.During the operation of ball mill,the working conditions and data distribution will change due to the variety of the set task and surrounding environment and the aging of the mechanical equipment,which may cause the problem of model mismatch.It is a key problem that how to get useful information from the features of different distribution data for predicting test data.Transfer learning relaxes the premise of data same-distribution.The existing knowledge of the source domain is transferred to solve the unknown but related learning problems with source domain by extracting "latent semantic" or "sharing knowledge structure" among different domains.According to the idea of feature mapping in transfer learning,a new method based on manifold regularization domain adaptation is researched for soft sensing of load parameters of wet ball mill.However,manifold regularization domain adaptation is an unsupervised feature mapping method,and only one historical working condition data is used in this method.We know that data from different working conditions have the characteristics of complementary information.And fusing label information into the process of feature mapping will improve the identifiability of data in public subspace.Therefore,a soft sensor of the ball mill load parameters based on integration of semi-supervised multi-source domain adaptation is researched.In practice running,we may get a small amount of labeled data in the current working condition.Because of the complexity of the operation environment,there will be strong uncertainty in vibration signals even in the case of the same load parameter.In view of the above situation,this paper studys an optimized semi-supervised domain adaptive fuzzy reasoning method.This method can make use of a small number of labeled data to modify the model and improve the generalization ability of the model.The specific research content is as follows:(1)Aiming at the challenging problems such as the measurement of key load parameters of ball mill under multi-operating conditions and less sample of test data,a soft sensor model based on manifold regularization domain adaptation for measuring wet ball mill load parameters is researched.Compared with the traditional soft sensor method and multi condition modeling method,the prediction accuracy of the manifold regularization domain adaptation is greatly improved,which verifies the effectiveness of the method.(2)Aiming at the model mismatching caused by data distribution mismatch between historical data and test data in changing of working condition and the problem of less sample of test data,a new semi-supervised domain adaptation method is researched in this paper.Compared with the manifold regularization domain adaptation method,the prediction ability and the model fitness of semi-supervised domain adaptation is further improved,and the results are more reliable.(3)The optimized semi-supervised domain adaptive fuzzy reasoning method is researched.After semi-supervised domain adapting fuzzy reasoning modeling,fuzzy rules are modified by using a small number of labeled data in the target domain.The experimental results show that the generalization ability of the model is further strengthened through a small number of labeled data modifications in the target domain.