To evaluate accurately working condition of guide, make maintenance strategy, and predict its residual life in the process of machining operation, a rolling guide rail condition monitoring system based on neural networks was constructed after key factors to guide life were investigated carefully. Eight B&K 4321 three-way vibration sensor were installed on slider surface to monitor the on-line condition of four guides and eight sliders. Vibration signals were processed by wavelet packet decomposition and the most sensitive features to guide life were selected by fuzzy clustering method. The relation between guide life and input vectors including vibration features and machining condition was built by radial basis probabilistic neural networks (RBPNN), which parameters were optimized by genetic algorithm. The experimental results show maximum forecast error is 360 hours and minimum forecast error is 63 hours.