The objective of this research is to develop a robust online non-invasive arc duration measurement method in the substation, which is a critical factor for circuit breaker online contact erosion estimation and implementation of condition-based and reliability-centered maintenance practices. In this research, different arc duration measurement methods using various physical signals, including very-low-frequency and low-frequency magnetic field, vibration, and sound signals are proposed. Each method, however, has its limitations. For the method based on very-low-frequency and low-frequency magnetic field, although the waveform signatures can indicate the arc initiation and extinction time, their susceptibility to contact geometry changes caused by contact erosion restricts the application scenarios of this method. The domain-adaptive method based on vibration and sound signals, on the other hand, can achieve a high accuracy (below 0.2 ms) comparable to existing techniques and exhibit robustness against noise interference and domain shift, which existing methods lack. However, the method requires representative substation data to train neural networks. Eventually, a decision-level fusion strategy is proposed. This strategy can combine existing and proposed methods to address the limitations that cannot be solved by a single method individually and enhance the accuracy and robustness of arc duration measurement in the substation.