When More Attention Makes Motor Diagnosis Worse
TL;DR for operators A motor can develop several faults simultaneously, while its vibration sensors record their combined effects. Diagnostic AI must distinguish the individual problems and their severity without losing information that the problems share. A study introducing RNDR-TFCA1 finds that combining two attention mechanisms improves diagnostic macro F1 from 0.877 to 0.898. Yet one mechanism used alone actually reduces performance. The improvement also requires only 0.57 million additional multiply-accumulate operations (MMACs), while substantially larger comparison models perform worse. ...