Sintering Condition Recognition of Rotary Kiln Based on Multivariable Time Series Data with Long-tail Distribution

The classification of long-tailed multivariate time series data (LTMTS) is a common realistic challenge in pattern recognition tasks. LTMTS usually leads to poor feature separability and low tail pattern recognition accuracy. In this paper, a feature learning model with the goal of improving feature separability and a novel classification model based on dynamic kernel modification strategy are proposed for LTMTS classification. The sintering condition recognition experiment of rotary kiln using real thermal data shows that the proposed methods can distinguish various conditions better and improve the recognition accuracy of tail conditions effectively.