In order to avoid the accuracy deterioration or tool damage caused by milling chatter, it is necessary to have an efficient and reliable diagnosis system that can on-line predict/detect the occurrence of chatter. The diagnosis/predicting system proposed is to on-line process and analysis the vibration signals of the milling machine measured by accelerometers. According to the analysis results,the system will be able to detect/predict the occurrence of the chatter. The diagnosis algorithm is, first,collecting both the normal signals and chatter signals from milling processes, and then, converting the signals through wavelet transform and fast Fourier transform (FFT). Since the converted chatter signals exhibit different characteristics from the normal signals, through defining the characteristic values, such as root-mean-square value, max value, and ratio of peak value to root-mean-square value,etc, a diagnosis reference library that contains the distribution of these characteristic values is built for diagnosis. When a diagnosis is executing, the characteristic value of the measured signals is contrasted with the diagnosis reference. The approach index which shows the possibility of occurrence of milling chatter will, then, be calculated through the diagnosis system. Cutting experiments are conducted to verify the proposed diagnosis system. The results show the success of early chatter detecting for the system.