Does the Precision Value Influence the Fusion Performance? A Method-Based Experimental Study

Abstract

Considering technically complex systems, the evaluation of situations or conditions is a challenging task. To ensure a high accuracy assignments from different classifiers can be fused. To define requirements for a good fusion performance and to evaluate the potential for higher accuracy, in this paper the idea of a fictional classifier is introduced. The precision values of one classifier denoted as fictional classifier are varied to demonstrate the influence on fused accuracy. From the results of this contribution some important challenges can be solved: Can the performance of one individual classifier improve the overall accuracy? Can the performance of the fused results be improved by changing performance measures of the fictional classifier? This allows the establishment of a supervised strategy to adapt precision values to get better fusion results. For illustrating the effectsfour benchmark examples are used. The introduced methods are applied to fault diagnosis of hot rolling mills. The results show that using a fictional classifier the overall accuracy can be outperformed depending on data sets.