Research And Application Of Multi Working Conditions:Soft Sensor Based On Feature Transfer

Soft sensor modeling plays an important role in ensuring the efficient and stable operation of process industry and improving product quality.However,due to the change of process mechanism,raw materials and operation environment,it is easy to cause sudden changes of system working conditions,resulting in multi working conditions and multimode characteristics.The data distribution between different working conditions is different,and the traditional soft sensor model is inaccurate.Therefore,it is a key problem to find a suitable modeling method to adapt to the influence of the change of working conditions.1.Aiming at the problem that a small number of labeled sample information and the common information between different working conditions can’t be effectively used when the working condition changes suddenly,a soft sensor modeling method based on local linear embedding for parameter transfer is proposed.First,extract the common information of the known condition and the condition to be measured;then the common information of the two as the input of domain adaptation random weight neural network,make full use of the labeled sample information to realize migration modeling.2.In order to solve the problem that there is no labeled sample and the common information between different working conditions can’t be used effectively when the working conditions change abruptly,a multimode soft sensor modelling method based on local tangent space alignment with geodesic flow kernel(LTSA-GFK)is proposed.By extracting the common information of historical modeling conditions and real-time conditions to be measured,the common information of the two conditions is projected into a manifold subspace,and the feature transfer is completed in the subspace,so as to reduce the distribution difference between different conditions and realize an unsupervised soft sensor modeling method.3.In view of the problem that there is no labeled sample for the condition to be tested when the condition changes suddenly and it is difficult to effectively reduce the distribution difference among the conditions in the feature transfer.The maximum mean difference and covariance distribution difference are integrated to restrict the edge distribution,condition distribution and correlation of different conditions in the process of feature projection.Morever,the maximum variance is introduced to ensure the expression ability of different process parameters.The manifold regularization is introduced to keep the local structure information of the data of different working conditions in the projection process.And two modeling methods of subspace domain adaptation for geodesic flow kernel and domain adaptation with aligning subspace are proposed,which do not need the labeled sample to reduce the distribution difference between the known working conditions and the working conditions to be measured in the feature projection process.Aiming at the problem that the data distribution of different working conditions is still large after the projection to the subspace,the geodesic flow kernel and aligning subspace are used to transfer the feature information of the data of known and to be measured working conditions projected into the subspace.Reduce the distribution difference of the data of different working conditions in the subspace,so as to improve the soft sensor accuracy of the data construction model of known working conditions to the data of measured working conditions.