Research And Application Of Multi-modal Unsupervised Domain Adaptive Soft Sensor

Diversification of production tasks makes industrial systems have multiple working conditions and multi-modal characteristics.It is difficult to realize real-time detection of key parameters of production processes in actual industry.The soft parameter sensor model of key parameters under knew working conditions are not suitable for changing working conditions.Unknown modality.Therefore,designing a suitable modeling strategy to adapt to the differences between different modes and the effects of modal changes and data drift is a difficult problem for multi-modal soft sensor.Traditional soft sensor schemes based on global model strategies and multi-model strategies are difficult to adapt to the effects of modal changes in working conditions.In the past,semi-supervised domain adaptation soft-sensing focused on the case where there were a small number of tags in the target domain,but it could not do anything in the face of the complex conditions of the target domain without tags.However,the existing unsupervised domain adaptation algorithm is limited by the inconsistency of different domain label categories,and it is difficult to complete the adaptation of the conditional distribution.Therefore,they mainly focus on the classification problem research of the adapted instance data,thus ignoring the importance of the conditional distribution adaptation of the label.However,the existing unsupervised domain adaptation algorithm is limited by the inconsistency of different domain label categories,and it is difficult to complete the adaptation of the conditional distribution of the labels,so that the soft measurement prediction based on them has problems such as negative transfer and under transfer.In order to solve the problems of under transfer and over-transfer in unsupervised domain adaptive soft-sampling,in this paper the soft classification strategy and similar domain selection are introduced to adapt the labels between different domains,so as to solve the problem of the performance degradation of transfer algorithms caused by the difference of labels between different domains in unsupervised regression transfer learning.This paper focuses on the unsupervised domain adaptive modeling method under the multi-case unknown mode,which can be summarized as follows:(1)For the traditional soft sensor modeling method in complex multi-cases,it is difficult to capture the effective feature representation.From the perspective of feature weight learning,it aims to find the effective feature representation of the source and target domains and reduce the domain offset.Fully exploit the information between different domains;for the existing feature matching domain adaptation method without considering the conditional distribution adaptation of the label,the joint subspace alignment strategy combined with the distribution alignment and subspace alignment method is used to reduce the data distribution difference between different domains.The distribution difference between the subspaces is reduced,and the data attributes in the original space are preserved.The joint distribution adaptation in the domain adaptation soft sensor cannot fully correct the problem of mismatch distribution of probability distribution between different domains.The feature mapping method re-weights the samples in the source domain to better match the distribution of the target domain,and reweights the data in the source domain to learn the target domain,and learns the domain invariant features in the manifold.And quantitatively evaluate the proportion of the aligned edge distribution and the conditional distribution at the time of feature mapping.(2)The existing unsupervised domain adaptation algorithm caused by the difference between the data and label of the target domain and the source domain cannot be used for conditional distribution adaptation.This paper introduces the soft classification strategy in a targeted manner to complete the regression label to the classification label through the classification strategy.The mapping,using the classification label for conditional distribution adaptation learning,to some extent solve the problem of the performance degradation of the transfer algorithm caused by the label difference between different domains in the unsupervised regression transfer learning.(3)For the problem of under-transfer,over-transfer and under-adaptation caused by the difference of distribution of target domain and source domain data and labels in regression transfer learning,the similar domain selection method is introduced on the basis of soft classification strategy,and the appropriate similar domain is used.The selection method selects the source domain most similar to the target domain from multiple known source domains,thereby improving the learning effect of unsupervised regression transfer learning to some extent.