Congestion detection in transmission control protocol/active queue management networks remains a challenging problem in which the choosing of congestion signal is one of the most important factors. Exponentially weighted moving average of queue length, the most widely used congestion signal, is facing difficulties in detecting incipient congestion and quantifying the optimal forgetting factor. Aiming at these 2 disadvantages, we propose an average queue‐length‐difference–based congestion detection algorithm where exponentially weighted moving average of queue‐length difference is chosen as the congestion signal with the theoretical optimal forgetting factor deduced. First, by defining the queue‐length difference as the state, the corresponding state equation is derived from the fluid model. Second, we prove that the inflow traffic in state equation is a discrete‐time martingale, which can be transformed to …