In this article, the method extracting fault characteristics by utilizing the duality of current-time pairs is proposed to diagnose the single switch tube open-circuit (OC) fault for the inverter in the traction motor drive system. This method develops a neural-network-liked structure for OC fault diagnosis. The input layer is a fixed-period integral mean (FPIM) filter within one carrier cycle to smooth sampled discrete current-time pairs for effectively counteracting current PWM-notches and noise interference. The filtered current-time pairs are fed forward to the middle layer including functional activation functions. One of them is to normalize the current such that the diagnosis algorithm can be self-adaptive to different loads. Another one is to eliminate the singular problem near the current peak. Finally, the diagnostic functions are constructed by the dual-difference of discrete current-time pairs to quickly diagnose and locate the single-tube OC fault. It does not require additional sensors and other auxiliary detection circuits, and can accurately locate faults with few current signal nodes. It greatly reduces fault detection time and avoids triggering a vicious chain reaction due to detection delays. In addition, it also has excellent robustness under varying frequency conditions since it is only concerned with the local property. This algorithm can be embedded in the closed-loop system due to its simple structure and low computation and storage burden. The effectiveness of the proposed method is verified through experiments.