A time series data classification method, device, equipment and storage medium

By combining unsupervised concept drift detection and causal dilated convolutional neural networks, the problem of low classification accuracy of time series data in dynamic environments is solved, achieving efficient classification of non-stationary streaming time series data and improving classification accuracy. It is applicable to weather forecasting, medical monitoring and financial risk control.

CN122220397APending Publication Date: 2026-06-16STATE GRID ECONOMIC TECH RES INST CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
STATE GRID ECONOMIC TECH RES INST CO LTD
Filing Date
2026-03-13
Publication Date
2026-06-16

AI Technical Summary

Technical Problem

Existing technologies have poor accuracy in classifying and learning time series data in non-stationary dynamic environments, making it difficult to adapt to the dynamic changes of streaming time series data.

Method used

An unsupervised concept drift detection algorithm is used to segment time series data, select anchor samples, positive samples, and negative samples, and use a neural network structure based on causal dilated convolution for representation encoding. The encoder parameters are optimized through a loss function, and the classifier is trained to adapt to the dynamic changes of non-stationary streaming time series data.

Benefits of technology

It significantly improves the classification accuracy of complex time-series data, and is suitable for non-stationary time-series scenarios such as weather forecasting, medical monitoring, and financial risk control, providing efficient and reliable intelligent analysis and decision support.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122220397A_ABST
    Figure CN122220397A_ABST
Patent Text Reader

Abstract

This invention discloses a time-series data classification method, comprising: Step 1, segmenting the original time-series data using an unsupervised concept drift detection algorithm to obtain multiple subsequences; Step 2, selecting anchor samples, positive samples, and negative sample sets from the multiple subsequences; Step 3, inputting the samples into a time-series representation encoder to output a representation vector; Step 4, calculating the loss value based on the representation vector and a loss function, updating the encoder parameters, and repeating steps 2 to 4 until convergence is achieved to obtain a trained encoder; Step 5, using the trained encoder to extract representations from labeled historical time-series datasets, and training a classifier based on the extraction results; Step 6, for new time-series data, extracting its representation using the trained encoder and inputting it into the classifier to obtain the predicted category. This invention can adapt to the dynamic changes of non-stationary streaming time-series data and significantly improve the classification accuracy of complex time-series data.
Need to check novelty before this filing date? Find Prior Art