Abstract: ATM networks oriented connections provide pure QoS (Quality of Service) for diversified services through a series of traffic management mechanisms, the ABR (Available Bit Rate) flow control is especially important among these approaches. In the binary flow control scheme, the cell rate and queue length may oscillate with great magnitude to reduce link utilization, so that EFCI (Explicit Forward Congestion Indication) algorithm is regard as ineffective, however its simplicity is attractive to high performance switch design. In this paper, the EFCI algorithm is analyzed based on the classical control theory. It is found that the nonlinear structure introduced as determining congestion a dominant reason that causes the oscillation. Then, a probability congestion detection approach, called p-EFCI, is put forward. Numerical results show that the improved algorithm deeply constrains the oscillation magnitude, and the queue length is controlled in limited scope to guarantee cell zero loss.