A multivariate time series missing value filling model based on multi-scale time sequence decoupling and semantic consistency alignment and a construction method thereof

By using a model that combines multi-scale temporal decoupling and semantic consistency alignment, we have solved the problems of robust aggregation and semantic constraints in imputing missing values ​​in multivariate time series, and achieved high stability and high accuracy in imputing missing values ​​under complex conditions.

CN122196364APending Publication Date: 2026-06-12ZHONGYUAN ENGINEERING COLLEGE
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
ZHONGYUAN ENGINEERING COLLEGE
Filing Date
2026-03-06
Publication Date
2026-06-12

AI Technical Summary

⚠Technical Problem

Existing technologies lack robust aggregation capabilities in imputing missing values ​​in multivariate time series, especially under complex missing and noisy conditions. The stability of the representation needs to be improved, and it is difficult for external semantic priors to form trainable deep alignment constraints with numerical representations.

⚗Method used

A multi-scale temporal decoupling and semantic consistency alignment model is adopted. Through missing value modeling and input preprocessing, multi-scale temporal decoupling module, median enhancement channel-time attention module and semantic consistency alignment module, a multivariate time series missing value imputation model is constructed to achieve robust representation and semantic constraints of numerical features.

🎯Benefits of technology

It improves the stability and accuracy of missing value imputation under conditions of high missing rate or noise interference, and enhances the model's anti-interference ability and the consistency of imputation results under complex conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122196364A_ABST
    Figure CN122196364A_ABST
Patent Text Reader

Abstract

The application provides a construction method of a multivariate time series missing value filling model based on multi-scale time sequence decoupling and semantic consistency alignment, comprising the following steps: S1, constructing a missing modeling and input preprocessing layer; S2, constructing a multi-scale time sequence decoupling module; S3, constructing a median enhancement channel-time attention module; S4, constructing a semantic consistency alignment module; and S5, constructing a missing value reconstruction module. The application improves the stability of multi-scale representation through multi-scale time sequence decoupling and a bidirectional cross-scale mixing mechanism, and reduces the interference of abnormal values and noise on feature aggregation through a median enhancement channel-time attention mechanism. At the same time, the consistency between the numerical representation and the semantic prior is enhanced through semantic consistency alignment constraints, thereby improving the missing value filling stability and filling precision under the condition of high missing rate or noise interference.
Need to check novelty before this filing date? Find Prior Art