A New Approach to Predict the Mechanical Properties of Hot-Rolled Strip Based on Multi-Dimensional Feature Weighted Similarity and Integrated Just-In-Time Learning

Abstract

Predicting the mechanical properties of hot-rolled strip poses significant challenges due to the intricate interplay of multi-dimensional similarities within sample analysis and the time-varying characteristics of actual production data. Relying on single similarity metrics to select appropriate samples becomes inadequate, hindering timely and accurate predictions. To address these issues, in this article, a new approach is proposed to predict the mechanical properties of hot-rolled strip based on combining multi-dimensional-feature-weighted similarity (MDFWS) and integrated just-in-time learning (IJITL). First, the feature weights based on multiple methods are combined with the time weight as the similarity measure to select the relevant samples, respectively. Subsequently, the linear local models are constructed based on the selected relevant samples to predict the query data. Finally, the output of each local model is given differentiated weights based on its performance on the validation set, and the final prediction result is obtained through an integrated learning strategy. Furthermore, a cumulative similarity factor is introduced to screen the optimal dataset for local models, and a similarity threshold is set to reduce the frequency of model updates. In the experimental results, it is demonstrated that the proposed MDFWS-IJITL model excels in predicting the mechanical properties of hot-rolled strip, offering higher predictive accuracy and better adaptability compared to traditional global modeling methods and JITL models.